Initial version for riscv sifive x280

Makefile.install

@@ -8,6 +8,7 @@ PREFIX ?= /opt/OpenBLAS
 OPENBLAS_INCLUDE_DIR := $(PREFIX)/include
 OPENBLAS_LIBRARY_DIR := $(PREFIX)/lib
 OPENBLAS_BINARY_DIR := $(PREFIX)/bin
+OPENBLAS_RELEASE_DIR := $(PREFIX)/release
 OPENBLAS_BUILD_DIR := $(CURDIR)
 OPENBLAS_CMAKE_DIR := $(OPENBLAS_LIBRARY_DIR)/cmake/$(LIBSONAMEBASE)
 OPENBLAS_CMAKE_CONFIG := OpenBLASConfig.cmake
@@ -38,6 +39,7 @@ install : 	lib.grd
 	@-mkdir -p "$(DESTDIR)$(OPENBLAS_INCLUDE_DIR)"
 	@-mkdir -p "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@-mkdir -p "$(DESTDIR)$(OPENBLAS_BINARY_DIR)"
+	@-mkdir -p "$(DESTDIR)$(OPENBLAS_RELEASE_DIR)"
 	@-mkdir -p "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)"
 	@-mkdir -p "$(DESTDIR)$(OPENBLAS_PKGCONFIG_DIR)"
 	@echo Generating openblas_config.h in $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)
@@ -202,3 +204,8 @@ endif
 	@echo "  endif ()" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)"
 	@echo "endif ()" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG_VERSION)"
 	@echo Install OK!
+#Generating release tar
+	@echo Generating $(OPENBLAS_RELEASE_DIR)/$(basename $(LIBNAME)).tar.gz
+	@tar -cvz --file=$(OPENBLAS_RELEASE_DIR)/$(basename $(LIBNAME)).tar.gz --directory=$(PREFIX) --exclude=release .
+
+	

Makefile.prebuild

@@ -55,6 +55,14 @@ ifeq ($(TARGET), C910V)
 TARGET_FLAGS = -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d
 endif
 
+ifeq ($(TARGET), x280)
+TARGET_FLAGS = -march=rv64imafdcv_zba_zbb_zfh_xsfvqmaccqoq_xsfvfhbfmin -mabi=lp64d -mcpu=sifive-x280
+endif
+
+ifeq ($(TARGET), RISCV64_GENERIC)
+TARGET_FLAGS = -march=rv64imafdc -mabi=lp64d
+endif
+
 all: getarch_2nd
 	./getarch_2nd  0 >> $(TARGET_MAKE)
 	./getarch_2nd  1 >> $(TARGET_CONF)

Makefile.riscv64

@@ -2,3 +2,11 @@ ifeq ($(CORE), C910V)
 CCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920
 FCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920 -static
 endif
+ifeq ($(CORE), x280)
+CCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_xsfvqmaccqoq_xsfvfhbfmin -mabi=lp64d -menable-experimental-extensions -mllvm --riscv-v-vector-bits-min=512 -mcpu=sifive-x280 -ffast-math
+FCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_xsfvqmaccqoq_xsfvfhbfmin -mabi=lp64d -menable-experimental-extensions -static
+endif
+ifeq ($(CORE), RISCV64_GENERIC)
+CCOMMON_OPT += -march=rv64imafdc -mabi=lp64d
+FCOMMON_OPT += -march=rv64imafdc -mabi=lp64d -static
+endif

README.md

@@ -186,6 +186,11 @@ Please read `GotoBLAS_01Readme.txt` for older CPU models already supported by th
   ```
   (also known to work on C906)
 
+- **x280**: LLVM auto-vectorization using RISC-V Vector extension 1.0.
+  ```sh
+  make HOSTCC=gcc TARGET=x280 NUM_THREADS=8 CC=riscv64-unknown-linux-gnu-clang FC=riscv64-unknown-linux-gnu-gfortran
+  ```
+
 ### Support for multiple targets in a single library
 
 OpenBLAS can be built for multiple targets with runtime detection of the target cpu by specifiying `DYNAMIC_ARCH=1` in Makefile.rule, on the gmake command line or as `-DDYNAMIC_ARCH=TRUE` in cmake.

TargetList.txt

@@ -120,6 +120,7 @@ Z14
 10.RISC-V 64:
 RISCV64_GENERIC
 C910V
+x280
 
 11.LOONGARCH64:
 LOONGSONGENERIC

benchmark/Makefile

@@ -37,6 +37,12 @@ ESSL=/opt/ibm/lib
 #LIBESSL = -lesslsmp $(ESSL)/libxlomp_ser.so.1 $(ESSL)/libxlf90_r.so.1 $(ESSL)/libxlfmath.so.1 $(ESSL)/libxlsmp.so.1 /opt/ibm/xlC/13.1.3/lib/libxl.a
 LIBESSL = -lesslsmp  $(ESSL)/libxlf90_r.so.1 $(ESSL)/libxlfmath.so.1 $(ESSL)/libxlsmp.so.1 /opt/ibm/xlC/13.1.3/lib/libxl.a
 
+# x280 temporary workaround for gfortran
+ifeq ($(TARGET), x280)
+CCOMMON_OPT:=$(filter-out -mllvm --riscv-v-vector-bits-min=512,$(CCOMMON_OPT))
+endif
+
+
 ifneq ($(NO_LAPACK), 1)
 GOTO_LAPACK_TARGETS=slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
 		    scholesky.goto dcholesky.goto ccholesky.goto zcholesky.goto \

common_riscv64.h

@@ -92,6 +92,10 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define SEEK_ADDRESS
 
 #if defined(C910V)
+#include <riscv-vector.h>
+#endif
+
+#if defined(x280)
 #include <riscv_vector.h>
 #endif
 

cpuid_riscv64.c

@@ -72,10 +72,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define CPU_GENERIC   0
 #define CPU_C910V     1
+#define CPU_x280        2
 
 static char *cpuname[] = {
   "RISCV64_GENERIC",
   "C910V"
+  "x280"
 };
 
 int detect(void){

getarch.c

@@ -1677,6 +1677,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define LIBNAME   "c910v"
 #define CORENAME  "C910V"
 #endif
+#endif
+#ifdef FORCE_x280
+#define FORCE
+#define ARCHITECTURE    "RISCV64"
+#define SUBARCHITECTURE "x280"
+#define SUBDIRNAME      "riscv64"
+#define ARCHCONFIG   "-Dx280 " \
+       "-DL1_DATA_SIZE=64536 -DL1_DATA_LINESIZE=32 " \
+       "-DL2_SIZE=262144 -DL2_LINESIZE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 "
+#define LIBNAME   "x280"
+#define CORENAME  "x280"
 #else
 #endif
 

kernel/riscv64/KERNEL.x280

@@ -0,0 +1,267 @@
+# **********************************************************************************
+# Copyright (c) 2022, The OpenBLAS Project
+# All rights reserved.
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+# 1. Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+# 2. Redistributions in binary form must reproduce the above copyright
+# notice, this list of conditions and the following disclaimer in
+# the documentation and/or other materials provided with the
+# distribution.
+# 3. Neither the name of the OpenBLAS project nor the names of
+# its contributors may be used to endorse or promote products
+# derived from this software without specific prior written permission.
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+# USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+# **********************************************************************************
+
+SAMAXKERNEL  = amax_rvv.c
+DAMAXKERNEL  = amax_rvv.c
+CAMAXKERNEL  = zamax_rvv.c
+ZAMAXKERNEL  = zamax_rvv.c
+
+SAMINKERNEL  = amin_rvv.c
+DAMINKERNEL  = amin_rvv.c
+CAMINKERNEL  = zamin_rvv.c
+ZAMINKERNEL  = zamin_rvv.c
+
+SMAXKERNEL   = max_rvv.c
+DMAXKERNEL   = max_rvv.c
+
+SMINKERNEL   = min_rvv.c
+DMINKERNEL   = min_rvv.c
+
+ISAMAXKERNEL = iamax_rvv.c
+IDAMAXKERNEL = iamax_rvv.c
+ICAMAXKERNEL = izamax_rvv.c
+IZAMAXKERNEL = izamax_rvv.c
+
+ISAMINKERNEL = iamin_rvv.c
+IDAMINKERNEL = iamin_rvv.c
+ICAMINKERNEL = izamin_rvv.c
+IZAMINKERNEL = izamin_rvv.c
+
+ISMAXKERNEL  = imax_rvv.c
+IDMAXKERNEL  = imax_rvv.c
+
+ISMINKERNEL  = imin_rvv.c
+IDMINKERNEL  = imin_rvv.c
+
+SASUMKERNEL  = asum_rvv.c
+DASUMKERNEL  = asum_rvv.c
+CASUMKERNEL  = zasum_rvv.c
+ZASUMKERNEL  = zasum_rvv.c
+
+SSUMKERNEL  = sum_rvv.c
+DSUMKERNEL  = sum_rvv.c
+CSUMKERNEL  = zsum_rvv.c
+ZSUMKERNEL  = zsum_rvv.c
+
+SAXPYKERNEL  = axpy_rvv.c
+DAXPYKERNEL  = axpy_rvv.c
+CAXPYKERNEL  = zaxpy_rvv.c
+ZAXPYKERNEL  = zaxpy_rvv.c
+
+SAXPBYKERNEL  = axpby_rvv.c
+DAXPBYKERNEL  = axpby_rvv.c
+CAXPBYKERNEL  = zaxpby_rvv.c
+ZAXPBYKERNEL  = zaxpby_rvv.c
+
+SCOPYKERNEL  = copy_rvv.c
+DCOPYKERNEL  = copy_rvv.c
+CCOPYKERNEL  = zcopy_rvv.c
+ZCOPYKERNEL  = zcopy_rvv.c
+
+SDOTKERNEL   = dot_rvv.c
+DDOTKERNEL   = dot_rvv.c
+CDOTKERNEL   = zdot_rvv.c
+ZDOTKERNEL   = zdot_rvv.c
+DSDOTKERNEL  = dot_rvv.c
+
+SNRM2KERNEL  = nrm2_rvv.c
+DNRM2KERNEL  = nrm2_rvv.c
+CNRM2KERNEL  = znrm2_rvv.c
+ZNRM2KERNEL  = znrm2_rvv.c
+
+SROTKERNEL   = rot_rvv.c
+DROTKERNEL   = rot_rvv.c
+CROTKERNEL   = zrot_rvv.c
+ZROTKERNEL   = zrot_rvv.c
+
+SSCALKERNEL  = scal_rvv.c
+DSCALKERNEL  = scal_rvv.c
+CSCALKERNEL  = zscal_rvv.c
+ZSCALKERNEL  = zscal_rvv.c
+
+SSWAPKERNEL  = swap_rvv.c
+DSWAPKERNEL  = swap_rvv.c
+CSWAPKERNEL  = zswap_rvv.c
+ZSWAPKERNEL  = zswap_rvv.c
+
+SGEMVNKERNEL = gemv_n_rvv.c
+DGEMVNKERNEL = gemv_n_rvv.c
+CGEMVNKERNEL = zgemv_n_rvv.c
+ZGEMVNKERNEL = zgemv_n_rvv.c
+
+SGEMVTKERNEL = gemv_t_rvv.c
+DGEMVTKERNEL = gemv_t_rvv.c
+CGEMVTKERNEL = zgemv_t_rvv.c
+ZGEMVTKERNEL = zgemv_t_rvv.c
+
+CTRMMKERNEL	= ztrmmkernel_2x2_rvv.c
+ZTRMMKERNEL	= ztrmmkernel_2x2_rvv.c
+
+# SGEMM_UNROLL_N set in params.h
+ifeq ($(SGEMM_UNROLL_N), 2)
+SGEMMKERNEL    =  gemmkernel_2x2_rvv.c
+SGEMMONCOPY    =  gemm_ncopy_2_rvv.c
+SGEMMOTCOPY    =  gemm_tcopy_2_rvv.c
+SGEMMONCOPYOBJ =  sgemm_oncopy.o
+SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+
+STRMMKERNEL	= trmmkernel_2x2_rvv.c
+else ifeq ($(SGEMM_UNROLL_N), 4)
+SGEMMKERNEL    =  gemmkernel_4x4_rvv.c
+SGEMMONCOPY    =  gemm_ncopy_4_rvv.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+SGEMMONCOPYOBJ =  sgemm_oncopy.o
+SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+
+STRMMKERNEL	= trmmkernel_4x4_rvv.c
+else ifeq ($(SGEMM_UNROLL_N), 8)
+# UNROLL_M is VLMAX
+SGEMMKERNEL    =  gemmkernel_rvv_v1x8.c
+SGEMMINCOPY    =  gemm_ncopy_rvv_v1.c
+SGEMMITCOPY    =  gemm_tcopy_rvv_v1.c
+SGEMMONCOPY    =  gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
+SGEMMOTCOPY    =  gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+STRMMKERNEL	= trmmkernel_rvv_v1x8.c 
+
+STRMMUNCOPY_M  =  trmm_uncopy_rvv_v1.c
+STRMMLNCOPY_M  =  trmm_lncopy_rvv_v1.c
+STRMMUTCOPY_M  =  trmm_utcopy_rvv_v1.c
+STRMMLTCOPY_M  =  trmm_ltcopy_rvv_v1.c
+
+SSYMMUCOPY_M   =  symm_ucopy_rvv_v1.c 
+SSYMMLCOPY_M   =  symm_lcopy_rvv_v1.c
+endif
+
+# SGEMM_UNROLL_N set in params.h
+ifeq ($(DGEMM_UNROLL_N), 2)
+DGEMMKERNEL    = gemmkernel_2x2_rvv.c
+DGEMMONCOPY    = gemm_ncopy_2_rvv.c
+DGEMMOTCOPY    = gemm_tcopy_2_rvv.c
+DGEMMONCOPYOBJ = dgemm_oncopy.o
+DGEMMOTCOPYOBJ = dgemm_otcopy.o
+
+DTRMMKERNEL	= trmmkernel_2x2_rvv.c
+else ifeq ($(DGEMM_UNROLL_N), 4)
+DGEMMKERNEL    =  gemmkernel_4x4_rvv.c
+DGEMMONCOPY    =  gemm_ncopy_4_rvv.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMONCOPYOBJ =  dgemm_oncopy.o
+DGEMMOTCOPYOBJ =  dgemm_otcopy.o
+
+DTRMMKERNEL	= trmmkernel_4x4_rvv.c
+else ifeq ($(DGEMM_UNROLL_N), 8)
+# UNROLL_M is VLMAX
+DGEMMKERNEL    =  gemmkernel_rvv_v1x8.c
+DGEMMINCOPY    =  gemm_ncopy_rvv_v1.c
+DGEMMITCOPY    =  gemm_tcopy_rvv_v1.c
+DGEMMONCOPY    =  gemm_ncopy_$(DGEMM_UNROLL_N)_rvv.c
+DGEMMOTCOPY    =  gemm_tcopy_$(DGEMM_UNROLL_N)_rvv.c
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DTRMMKERNEL = trmmkernel_rvv_v1x8.c
+DTRMMUNCOPY_M  =  trmm_uncopy_rvv_v1.c
+DTRMMLNCOPY_M  =  trmm_lncopy_rvv_v1.c
+DTRMMUTCOPY_M  =  trmm_utcopy_rvv_v1.c
+DTRMMLTCOPY_M  =  trmm_ltcopy_rvv_v1.c
+
+DSYMMUCOPY_M   =  symm_ucopy_rvv_v1.c
+DSYMMLCOPY_M   =  symm_lcopy_rvv_v1.c
+endif
+
+CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+CGEMMONCOPYOBJ =  cgemm_oncopy.o
+CGEMMOTCOPYOBJ =  cgemm_otcopy.o
+
+ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMONCOPYOBJ =  zgemm_oncopy.o
+ZGEMMOTCOPYOBJ =  zgemm_otcopy.o
+
+STRSMKERNEL_LN	=  trsm_kernel_LN_rvv_v1.c
+STRSMKERNEL_LT	=  trsm_kernel_LT_rvv_v1.c
+STRSMKERNEL_RN	=  trsm_kernel_RN_rvv_v1.c
+STRSMKERNEL_RT	=  trsm_kernel_RT_rvv_v1.c
+
+DTRSMKERNEL_LN	=  trsm_kernel_LN_rvv_v1.c
+DTRSMKERNEL_LT	=  trsm_kernel_LT_rvv_v1.c
+DTRSMKERNEL_RN	=  trsm_kernel_RN_rvv_v1.c
+DTRSMKERNEL_RT	=  trsm_kernel_RT_rvv_v1.c
+
+CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+TRSMCOPYLN_M    = trsm_lncopy_rvv_v1.c
+TRSMCOPYLT_M    = trsm_ltcopy_rvv_v1.c
+TRSMCOPYUN_M    = trsm_uncopy_rvv_v1.c
+TRSMCOPYUT_M    = trsm_utcopy_rvv_v1.c
+
+SSYMV_U_KERNEL =  symv_U_rvv.c 
+SSYMV_L_KERNEL =  symv_L_rvv.c
+DSYMV_U_KERNEL =  symv_U_rvv.c 
+DSYMV_L_KERNEL =  symv_L_rvv.c
+CSYMV_U_KERNEL =  ../generic/zsymv_k.c
+CSYMV_L_KERNEL =  ../generic/zsymv_k.c
+ZSYMV_U_KERNEL =  ../generic/zsymv_k.c
+ZSYMV_L_KERNEL =  ../generic/zsymv_k.c
+
+
+LSAME_KERNEL = ../generic/lsame.c
+
+SCABS_KERNEL	= ../generic/cabs.c
+DCABS_KERNEL	= ../generic/cabs.c
+QCABS_KERNEL	= ../generic/cabs.c
+
+ifndef SGEMM_BETA
+SGEMM_BETA = gemm_beta_rvv.c
+endif
+ifndef DGEMM_BETA
+DGEMM_BETA = gemm_beta_rvv.c
+endif
+ifndef CGEMM_BETA
+CGEMM_BETA = zgemm_beta_rvv.c
+endif
+ifndef ZGEMM_BETA
+ZGEMM_BETA = zgemm_beta_rvv.c
+endif

kernel/riscv64/amax_rvv.c

@@ -0,0 +1,102 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMAXVV_FLOAT vfmax_vv_f32m8
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMAXVV_FLOAT vfmax_vv_f64m8
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT maxf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(maxf);
+
+    FLOAT_V_T vx, vmax;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmax = VFMVVF_FLOAT(0.0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vmax = VFMAXVV_FLOAT(vmax, vx, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vmax = VFMAXVV_FLOAT(vmax, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDMAXVS_FLOAT(v_res, vmax, v_res, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+
+    return(maxf);
+}

kernel/riscv64/amin_rvv.c

@@ -0,0 +1,102 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMINVV_FLOAT vfmin_vv_f32m8
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMINVV_FLOAT vfmin_vv_f64m8
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT minf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(minf);
+
+    FLOAT_V_T vx, vmin;
+    FLOAT_V_T_M1 v_res;
+    
+    v_res = VFMVVF_FLOAT_M1(FLT_MAX, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmin = VFMVVF_FLOAT(FLT_MAX, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vmin = VFMINVV_FLOAT(vmin, vx, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vmin = VFMINVV_FLOAT(vmin, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDMINVS_FLOAT(v_res, vmin, v_res, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+
+    return(minf);
+}

kernel/riscv64/asum_rvv.c

@@ -0,0 +1,99 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFADDVV_FLOAT vfadd_vv_f32m8
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m8_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFADDVV_FLOAT vfadd_vv_f64m8
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m8_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT asumf = 0.0;
+    if (n <= 0 || inc_x <= 0) return(asumf);
+
+    FLOAT_V_T vx, vsum;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vsum = VFMVVF_FLOAT(0.0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vsum = VFADDVV_FLOAT(vsum, vx, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+            vsum = VFADDVV_FLOAT(vsum, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDSUMVS_FLOAT(v_res, vsum, v_res, vlmax); 
+    asumf = VFMVFS_FLOAT_M1(v_res);
+    return(asumf);
+}

kernel/riscv64/axpby_rvv.c

@@ -0,0 +1,171 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMACCVF_FLOAT vfmacc_vf_f32m8
+#define VFMULVF_FLOAT vfmul_vf_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMACCVF_FLOAT vfmacc_vf_f64m8
+#define VFMULVF_FLOAT vfmul_vf_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#endif
+
+int CNAME(BLASLONG n, FLOAT alpha, FLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *y, BLASLONG inc_y)
+{
+    FLOAT_V_T vx, vy;
+
+    if ( n < 0     )  return(0);
+
+    if ( beta == 0.0 ) {
+        if ( alpha == 0.0 ) {
+            if (1 == inc_y) {
+                memset(&y[0], 0, n * sizeof(FLOAT));
+            } else {
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                size_t vl = VSETVL(n);
+                vy = VFMVVF_FLOAT(0.0, vl);
+                for ( ; n > 0; n -= vl, y += vl*stride_y) {
+                    vl = VSETVL(n);
+                    VSSEV_FLOAT(y, stride_y, vy, vl);
+                }
+            }
+
+        } else {
+            if ((1 == inc_x) && (1 == inc_y)) {
+                for (size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+                    vl = VSETVL(n);
+                    vx = VLEV_FLOAT(x, vl);
+                    vy = VFMULVF_FLOAT(vx, alpha, vl);
+                    VSEV_FLOAT (y, vy, vl);
+                }
+            } else if (1 == inc_x) {
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+                    vl = VSETVL(n);
+                    vx = VLEV_FLOAT(x, vl);
+                    vy = VFMULVF_FLOAT(vx, alpha, vl);
+                    VSSEV_FLOAT (y, stride_y, vy, vl);
+                }
+            } else if (1 == inc_y) {
+                BLASLONG stride_x = inc_x * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+                    vl = VSETVL(n);
+                    vx = VLSEV_FLOAT(x, stride_x, vl);
+                    vy = VFMULVF_FLOAT(vx, alpha, vl);
+                    VSEV_FLOAT (y, vy, vl);
+                }
+            } else {
+                BLASLONG stride_x = inc_x * sizeof(FLOAT);
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+                    vl = VSETVL(n);
+                    vx = VLSEV_FLOAT(x, stride_x, vl);
+                    vy = VFMULVF_FLOAT(vx, alpha, vl);
+                    VSSEV_FLOAT (y, stride_y, vy, vl);
+                }
+            }
+        }
+
+    } else {
+        if ( alpha == 0.0 ) {
+            if (1 == inc_y) {
+                for (size_t vl; n > 0; n -= vl, y += vl) {
+                    vl = VSETVL(n);
+                    vy = VLEV_FLOAT(y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    VSEV_FLOAT (y, vy, vl);
+                }
+            } else {
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, y += vl*inc_y) {
+                    vl = VSETVL(n);
+                    vy = VLSEV_FLOAT(y, stride_y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    VSSEV_FLOAT (y, stride_y, vy, vl);
+                }
+            }
+
+        } else {
+            if ((1 == inc_x) && (1 == inc_y)) {
+                for (size_t vl; n > 0; n -= vl, y += vl) {
+                    vl = VSETVL(n);
+                    vy = VLEV_FLOAT(y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    VSEV_FLOAT (y, vy, vl);
+                }
+            } else if (1 == inc_x) {
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+                    vl = VSETVL(n);
+                    vx = VLEV_FLOAT(x, vl);
+                    vy = VLSEV_FLOAT(y, stride_y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    vy = VFMACCVF_FLOAT(vy, alpha, vx, vl);
+                    VSSEV_FLOAT (y, stride_y, vy, vl);
+                }
+            } else if (1 == inc_y) {
+                BLASLONG stride_x = inc_x * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+                    vl = VSETVL(n);
+                    vx = VLSEV_FLOAT(x, stride_x, vl);
+                    vy = VLEV_FLOAT(y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    vy = VFMACCVF_FLOAT(vy, alpha, vx, vl);
+                    VSEV_FLOAT (y, vy, vl);
+                }
+            } else {
+                BLASLONG stride_x = inc_x * sizeof(FLOAT);
+                BLASLONG stride_y = inc_y * sizeof(FLOAT);
+                for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+                    vl = VSETVL(n);
+                    vx = VLSEV_FLOAT(x, stride_x, vl);
+                    vy = VLSEV_FLOAT(y, stride_y, vl);
+                    vy = VFMULVF_FLOAT(vy, beta, vl);
+                    vy = VFMACCVF_FLOAT(vy, alpha, vx, vl);
+                    VSSEV_FLOAT (y, stride_y, vy, vl);
+                }
+            }
+        }
+    }
+
+    return(0);
+}

kernel/riscv64/axpy_rvv.c

@@ -0,0 +1,109 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMACCVF_FLOAT vfmacc_vf_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMACCVF_FLOAT vfmacc_vf_f64m8
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+    if ( n <= 0    ) return(0);
+    if ( da == 0.0 ) return(0);
+
+    FLOAT_V_T vx, vy;
+
+    if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLEV_FLOAT(y, vl);
+            vy = VFMACCVF_FLOAT(vy, da, vx, vl);
+            VSEV_FLOAT (y, vy, vl);
+        }
+
+    } else if (1 == inc_y) {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLEV_FLOAT(y, vl);
+            vy = VFMACCVF_FLOAT(vy, da, vx, vl);
+            VSEV_FLOAT(y, vy, vl);
+        }
+
+    } else if (1 == inc_x) {
+
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+            vy = VFMACCVF_FLOAT(vy, da, vx, vl);
+            VSSEV_FLOAT(y, stride_y, vy, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+            vy = VFMACCVF_FLOAT(vy, da, vx, vl);
+            VSSEV_FLOAT(y, stride_y, vy, vl);
+        }
+
+    }
+
+    return(0);
+}

kernel/riscv64/copy_rvv.c

@@ -0,0 +1,94 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#endif
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+    if(n < 0)  return(0);
+
+    FLOAT_V_T v0;
+
+    if(inc_x == 1 && inc_y == 1) {
+
+        for(size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = VSETVL(n);
+            v0 = VLEV_FLOAT(x, vl);
+            VSEV_FLOAT(y, v0, vl);
+        }
+
+    } else if (inc_y == 1) {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+            vl = VSETVL(n);
+            v0 = VLSEV_FLOAT(x, stride_x, vl);
+            VSEV_FLOAT(y, v0, vl);
+        }
+
+    } else if(inc_x == 1) {
+
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+            vl = VSETVL(n);
+            v0 = VLEV_FLOAT(x, vl);
+            VSSEV_FLOAT(y, stride_y, v0, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+            vl = VSETVL(n);
+            v0 = VLSEV_FLOAT(x, stride_x, vl);
+            VSSEV_FLOAT(y, stride_y, v0, vl);
+        }
+
+    }
+
+    return(0);
+}

kernel/riscv64/dot_rvv.c

@@ -0,0 +1,126 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if defined(DSDOT)
+double CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+#else
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+#endif
+{
+    double dot = 0.0;
+
+    if ( n <= 0 ) return(dot);
+
+    size_t vlmax = vsetvlmax_e64m8();
+    vfloat64m8_t vr = vfmv_v_f_f64m8(0, vlmax);
+
+    if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = vsetvl_e64m8(n);
+
+#if !defined(DOUBLE)
+            vfloat32m4_t vx = vle32_v_f32m4(x, vl);
+            vfloat32m4_t vy = vle32_v_f32m4(y, vl);
+
+            vr = vfwmacc_vv_f64m8(vr, vx, vy, vl);
+#else
+            vfloat64m8_t vx = vle64_v_f64m8(x, vl);
+            vfloat64m8_t vy = vle64_v_f64m8(y, vl);
+
+            vr = vfmacc_vv_f64m8(vr, vx, vy, vl);
+#endif
+        }
+
+    } else if (1 == inc_x) {
+            
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+            vl = vsetvl_e64m8(n);
+
+#if !defined(DOUBLE)
+            vfloat32m4_t vx = vle32_v_f32m4(x, vl);
+            vfloat32m4_t vy = vlse32_v_f32m4(y, stride_y, vl);
+
+            vr = vfwmacc_vv_f64m8(vr, vx, vy, vl);
+#else
+            vfloat64m8_t vx = vle64_v_f64m8(x, vl);
+            vfloat64m8_t vy = vlse64_v_f64m8(y, stride_y, vl);
+
+            vr = vfmacc_vv_f64m8(vr, vx, vy, vl);
+#endif
+        }
+    } else if (1 == inc_y) {
+            
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+            vl = vsetvl_e64m8(n);
+
+#if !defined(DOUBLE)
+            vfloat32m4_t vx = vlse32_v_f32m4(x, stride_x, vl);
+            vfloat32m4_t vy = vle32_v_f32m4(y, vl);
+
+            vr = vfwmacc_vv_f64m8(vr, vx, vy, vl);
+#else
+            vfloat64m8_t vx = vlse64_v_f64m8(x, stride_x, vl);
+            vfloat64m8_t vy = vle64_v_f64m8(y, vl);
+
+            vr = vfmacc_vv_f64m8(vr, vx, vy, vl);
+#endif
+        }
+    } else {
+            
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+            vl = vsetvl_e64m8(n);
+
+#if !defined(DOUBLE)
+            vfloat32m4_t vx = vlse32_v_f32m4(x, stride_x, vl);
+            vfloat32m4_t vy = vlse32_v_f32m4(y, stride_y, vl);
+
+            vr = vfwmacc_vv_f64m8(vr, vx, vy, vl);
+#else
+            vfloat64m8_t vx = vlse64_v_f64m8(x, stride_x, vl);
+            vfloat64m8_t vy = vlse64_v_f64m8(y, stride_y, vl);
+
+            vr = vfmacc_vv_f64m8(vr, vx, vy, vl);
+#endif
+        }
+    }
+
+    vfloat64m1_t vec_zero = vfmv_v_f_f64m1(0, vlmax);
+    vfloat64m1_t vec_sum = vfredusum_vs_f64m8_f64m1(vec_zero, vr, vec_zero, vlmax);
+    dot = vfmv_f_s_f64m1_f64(vec_sum);
+
+    return(dot);
+}

kernel/riscv64/gemm_beta_rvv.c

@@ -0,0 +1,89 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMULVF_FLOAT vfmul_vf_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMULVF_FLOAT vfmul_vf_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#endif
+
+// Optimizes the implementation in ../generic/gemm_beta.c
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
+          IFLOAT *dummy2, BLASLONG dummy3, IFLOAT *dummy4, BLASLONG dummy5,
+          FLOAT *c, BLASLONG ldc)
+{
+    BLASLONG chunk;
+    FLOAT *c_offset;
+	size_t vl;
+    FLOAT_V_T vx;
+
+    if (beta == ZERO) {
+
+        vl = VSETVL(m);
+        vx = VFMVVF_FLOAT(0.0, vl);
+
+        for( ; n > 0; n--, c += ldc) {
+            c_offset = c;
+
+            for(chunk=m; chunk > 0; chunk -= vl, c_offset += vl) {
+                vl = VSETVL(chunk);
+
+                VSEV_FLOAT(c_offset, vx, vl);
+			}
+		}
+
+	} else {
+
+        for( ; n > 0; n--, c += ldc) {
+            c_offset = c;
+
+            for(chunk=m; chunk > 0; chunk -= vl, c_offset += vl) {
+                vl = VSETVL(chunk);
+
+                vx = VLEV_FLOAT(c_offset, vl);
+                vx = VFMULVF_FLOAT(vx, beta, vl);
+                VSEV_FLOAT(c_offset, vx, vl);
+			}
+		}
+
+	}
+
+    return 0;
+}

kernel/riscv64/gemm_ncopy_2_rvv.c

@@ -0,0 +1,92 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VSEV_FLOAT vse32_v_f32m4
+#define VSSEG2_FLOAT vsseg2e32_v_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VSEV_FLOAT vse64_v_f64m4
+#define VSSEG2_FLOAT vsseg2e64_v_f64m4
+#endif
+
+// Optimizes the implementation in ../generic/gemm_ncopy_2.c
+
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
+{
+    BLASLONG i, j;
+    IFLOAT *a_offset, *a_offset1, *a_offset2;
+    IFLOAT *b_offset;
+    FLOAT_V_T v1, v2;
+    size_t vl;
+
+    //fprintf(stderr, "gemm_ncopy_2 m=%ld n=%ld lda=%ld\n", m, n, lda); // KU
+
+    a_offset = a;
+    b_offset = b;
+
+    for(j = (n >> 1); j > 0; j--) {
+
+        a_offset1 = a_offset;
+        a_offset2 = a_offset + lda;
+        a_offset += 2 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+            VSSEG2_FLOAT(b_offset, v1, v2, vl);
+
+	        a_offset1 += vl;
+	        a_offset2 += vl;
+	        b_offset += vl*2;
+        }
+    }
+
+    if (n & 1) {
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset, vl);
+            VSEV_FLOAT(b_offset, v1, vl);
+
+	        a_offset += vl;
+	        b_offset += vl;
+        }
+	}
+
+    return 0;
+}

kernel/riscv64/gemm_ncopy_4_rvv.c

@@ -0,0 +1,123 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSSEG2_FLOAT vsseg2e32_v_f32m2
+#define VSSEG4_FLOAT vsseg4e32_v_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSSEG2_FLOAT vsseg2e64_v_f64m2
+#define VSSEG4_FLOAT vsseg4e64_v_f64m2
+#endif
+
+// Optimizes the implementation in ../generic/gemm_ncopy_4.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
+{
+    BLASLONG i, j;
+
+    FLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
+    FLOAT *b_offset;
+
+    FLOAT_V_T v1, v2, v3, v4;
+    size_t vl;
+
+    //fprintf(stderr, "gemm_ncopy_4 m=%ld n=%ld lda=%ld\n", m, n, lda);
+
+    a_offset = a;
+    b_offset = b;
+
+    for(j = (n >> 2); j > 0; j--) {
+        a_offset1  = a_offset;
+        a_offset2  = a_offset1 + lda;
+        a_offset3  = a_offset2 + lda;
+        a_offset4  = a_offset3 + lda;
+        a_offset += 4 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+            v3 = VLEV_FLOAT(a_offset3, vl);
+            v4 = VLEV_FLOAT(a_offset4, vl);
+
+            VSSEG4_FLOAT(b_offset, v1, v2, v3, v4, vl);
+
+            a_offset1 += vl;
+            a_offset2 += vl;
+            a_offset3 += vl;
+            a_offset4 += vl;
+            b_offset += vl*4;
+        }
+    }
+
+    if (n & 2) {
+        a_offset1  = a_offset;
+        a_offset2  = a_offset1 + lda;
+        a_offset += 2 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+
+            VSSEG2_FLOAT(b_offset, v1, v2, vl);
+
+            a_offset1 += vl;
+            a_offset2 += vl;
+            b_offset += vl*2;
+        }
+    }
+
+    if (n & 1) {
+        a_offset1  = a_offset;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+
+            VSEV_FLOAT(b_offset, v1, vl);
+
+            a_offset1 += vl;
+            b_offset += vl;
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/gemm_ncopy_8_rvv.c

@@ -0,0 +1,164 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m1(n)
+#define FLOAT_V_T vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m1
+#define VSEV_FLOAT vse32_v_f32m1
+#define VSSEG2_FLOAT vsseg2e32_v_f32m1
+#define VSSEG4_FLOAT vsseg4e32_v_f32m1
+#define VSSEG8_FLOAT vsseg8e32_v_f32m1
+#else
+#define VSETVL(n) vsetvl_e64m1(n)
+#define FLOAT_V_T vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m1
+#define VSEV_FLOAT vse64_v_f64m1
+#define VSSEG2_FLOAT vsseg2e64_v_f64m1
+#define VSSEG4_FLOAT vsseg4e64_v_f64m1
+#define VSSEG8_FLOAT vsseg8e64_v_f64m1
+#endif
+
+// Optimizes the implementation in ../generic/gemm_ncopy_8.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
+{
+    BLASLONG i, j;
+
+    FLOAT *a_offset;
+    FLOAT *a_offset1, *a_offset2, *a_offset3, *a_offset4;
+    FLOAT *a_offset5, *a_offset6, *a_offset7, *a_offset8;
+    FLOAT *b_offset;
+
+    FLOAT_V_T v1, v2, v3, v4, v5, v6, v7, v8;
+    size_t vl;
+
+    //fprintf(stderr, "gemm_ncopy_8 m=%ld n=%ld lda=%ld\n", m, n, lda);
+
+    a_offset = a;
+    b_offset = b;
+
+    for(j = (n >> 3); j > 0; j--) {
+        a_offset1  = a_offset;
+        a_offset2  = a_offset1 + lda;
+        a_offset3  = a_offset2 + lda;
+        a_offset4  = a_offset3 + lda;
+        a_offset5  = a_offset4 + lda;
+        a_offset6  = a_offset5 + lda;
+        a_offset7  = a_offset6 + lda;
+        a_offset8  = a_offset7 + lda;
+        a_offset += 8 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+            v3 = VLEV_FLOAT(a_offset3, vl);
+            v4 = VLEV_FLOAT(a_offset4, vl);
+            v5 = VLEV_FLOAT(a_offset5, vl);
+            v6 = VLEV_FLOAT(a_offset6, vl);
+            v7 = VLEV_FLOAT(a_offset7, vl);
+            v8 = VLEV_FLOAT(a_offset8, vl);
+
+            VSSEG8_FLOAT(b_offset, v1, v2, v3, v4, v5, v6, v7, v8, vl);
+
+            a_offset1 += vl;
+            a_offset2 += vl;
+            a_offset3 += vl;
+            a_offset4 += vl;
+            a_offset5 += vl;
+            a_offset6 += vl;
+            a_offset7 += vl;
+            a_offset8 += vl;
+            b_offset += vl*8;
+        }
+    }
+
+    if (n & 4) {
+        a_offset1  = a_offset;
+        a_offset2  = a_offset1 + lda;
+        a_offset3  = a_offset2 + lda;
+        a_offset4  = a_offset3 + lda;
+        a_offset += 4 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+            v3 = VLEV_FLOAT(a_offset3, vl);
+            v4 = VLEV_FLOAT(a_offset4, vl);
+
+            VSSEG4_FLOAT(b_offset, v1, v2, v3, v4, vl);
+
+            a_offset1 += vl;
+            a_offset2 += vl;
+            a_offset3 += vl;
+            a_offset4 += vl;
+            b_offset += vl*4;
+        }
+    }
+
+    if (n & 2) {
+        a_offset1  = a_offset;
+        a_offset2  = a_offset1 + lda;
+        a_offset += 2 * lda;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+            v2 = VLEV_FLOAT(a_offset2, vl);
+
+            VSSEG2_FLOAT(b_offset, v1, v2, vl);
+
+            a_offset1 += vl;
+            a_offset2 += vl;
+            b_offset += vl*2;
+        }
+    }
+
+    if (n & 1) {
+        a_offset1  = a_offset;
+
+        for(i = m; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            v1 = VLEV_FLOAT(a_offset1, vl);
+
+            VSEV_FLOAT(b_offset, v1, vl);
+
+            a_offset1 += vl;
+            b_offset += vl;
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/gemm_ncopy_rvv_v1.c

@@ -0,0 +1,76 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
+{
+    BLASLONG i, j;
+
+    FLOAT *a_offset;
+    FLOAT *a_offset1;
+    FLOAT *b_offset;
+
+    FLOAT_V_T v0;
+    size_t vl;
+
+    //fprintf(stderr, "%s, m=%ld n=%ld lda=%ld\n", __FUNCTION__, m, n, lda);
+
+    a_offset = a;
+    b_offset = b;
+
+    for(j = n; j > 0; j -= vl) {
+        vl = VSETVL(j);
+
+        a_offset1 = a_offset;
+        a_offset += vl * lda;
+
+        for(i = m; i > 0; i--) {
+            v0 = VLSEV_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
+            VSEV_FLOAT(b_offset, v0, vl);
+
+            a_offset1++;
+            b_offset += vl;
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/gemm_tcopy_2_rvv.c

@@ -0,0 +1,108 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSSSEG2_FLOAT vssseg2e32_v_f32m2
+#define VSSSEG4_FLOAT vssseg4e32_v_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSSSEG2_FLOAT vssseg2e64_v_f64m2
+#define VSSSEG4_FLOAT vssseg4e64_v_f64m2
+#endif
+
+// Optimizes the implementation in ../generic/gemm_tcopy_2.c
+
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
+{
+    BLASLONG i, j;
+    IFLOAT *a_offset, *a_offset1, *a_offset2;
+    IFLOAT *b_offset, *b_offset1, *b_offset2;
+    FLOAT_V_T v1a, v1b, v2a, v2b;
+    size_t vl;
+
+    //fprintf(stderr, "gemm_tcopy_2 m=%ld n=%ld lda=%ld\n", m, n, lda); // KU
+
+    a_offset = a;
+    b_offset = b;
+    b_offset2 = b + m * (n & ~1);
+
+    for(i = (m >> 1); i > 0; i--) {
+
+        a_offset1 = a_offset;
+        a_offset2 = a_offset + lda;
+        a_offset += 2 * lda;
+
+        b_offset1 = b_offset;
+        b_offset += 4;
+
+        for(j = (n >> 1); j > 0; j -= vl) {
+            vl = VSETVL(j);
+
+            VLSEG2_FLOAT(&v1a, &v1b, a_offset1, vl);
+            VLSEG2_FLOAT(&v2a, &v2b, a_offset2, vl);
+
+            VSSSEG4_FLOAT(b_offset1, m*2*sizeof(FLOAT), v1a, v1b, v2a, v2b, vl);
+
+            a_offset1 += vl * 2;
+	        a_offset2 += vl * 2;
+            b_offset1 += vl * m * 2;
+        }
+
+        if (n & 1) {
+	        *(b_offset2 + 0) = *(a_offset1 + 0);
+	        *(b_offset2 + 1) = *(a_offset2 + 0);
+	        b_offset2 += 2;
+        }
+    }
+
+    if (m & 1) {
+
+        for(j = (n >> 1); j > 0; j -= vl) {
+            vl = VSETVL(j);
+
+            VLSEG2_FLOAT(&v1a, &v1b, a_offset, vl);
+
+            VSSSEG2_FLOAT(b_offset, m*2*sizeof(FLOAT), v1a, v1b, vl);
+
+            a_offset += vl * 2;
+            b_offset += vl * m * 2;
+        }
+
+        if (n & 1){
+            *(b_offset2 + 0) = *(a_offset + 0);
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/gemm_tcopy_4_rvv.c

@@ -0,0 +1,236 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSSSEG2_FLOAT vssseg2e32_v_f32m2
+#define VSSSEG4_FLOAT vssseg4e32_v_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSSSEG2_FLOAT vssseg2e64_v_f64m2
+#define VSSSEG4_FLOAT vssseg4e64_v_f64m2
+#endif
+
+// Optimizes the implementation in ../generic/gemm_tcopy_4.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
+{
+    BLASLONG i, j;
+
+    FLOAT *a_offset, *a_offset1, *a_offset2, *a_offset3, *a_offset4;
+    FLOAT *b_offset, *b_offset1, *b_offset2, *b_offset3;
+    FLOAT  ctemp1,  ctemp2,  ctemp3,  ctemp4;
+    FLOAT  ctemp5,  ctemp6,  ctemp7,  ctemp8;
+    FLOAT  ctemp9, ctemp10, ctemp11, ctemp12;
+    FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+
+    //fprintf(stderr, "gemm_tcopy_4 m=%ld n=%ld lda=%ld\n", m, n, lda);
+
+    a_offset = a;
+    b_offset = b;
+
+    b_offset2 = b + m  * (n & ~3);
+    b_offset3 = b + m  * (n & ~1);
+
+    for(j = (m >> 2); j > 0; j--) {
+        a_offset1 = a_offset;
+        a_offset2 = a_offset1 + lda;
+        a_offset3 = a_offset2 + lda;
+        a_offset4 = a_offset3 + lda;
+        a_offset += 4 * lda;
+
+        b_offset1 = b_offset;
+        b_offset += 16;
+
+        for(i = (n >> 2); i > 0; i--) {
+            v1 = VLEV_FLOAT(a_offset1, 4);
+            v2 = VLEV_FLOAT(a_offset2, 4);
+            v3 = VLEV_FLOAT(a_offset3, 4);
+            v4 = VLEV_FLOAT(a_offset4, 4);
+
+            a_offset1 += 4;
+            a_offset2 += 4;
+            a_offset3 += 4;
+            a_offset4 += 4;
+
+            VSEV_FLOAT(b_offset1, v1, 4);
+            VSEV_FLOAT(b_offset2+4, v2, 4);
+            VSEV_FLOAT(b_offset2+8, v3, 4);
+            VSEV_FLOAT(b_offset2+12, v4, 4);
+
+            b_offset1 += m * 4;
+        }
+
+        if (n & 2) {
+            v1 = VLEV_FLOAT(a_offset1, 2);
+            v2 = VLEV_FLOAT(a_offset2, 2);
+            v3 = VLEV_FLOAT(a_offset3, 2);
+            v4 = VLEV_FLOAT(a_offset4, 2);
+
+            a_offset1 += 2;
+            a_offset2 += 2;
+            a_offset3 += 2;
+            a_offset4 += 2;
+
+            VSEV_FLOAT(b_offset2, v1, 2);
+            VSEV_FLOAT(b_offset2+2, v2, 2);
+            VSEV_FLOAT(b_offset2+4, v3, 2);
+            VSEV_FLOAT(b_offset2+6, v4, 2);
+
+            b_offset2 += 8;
+        }
+
+        if (n & 1) {
+            v1 = VLEV_FLOAT(a_offset1, 1);
+            v2 = VLEV_FLOAT(a_offset2, 1);
+            v3 = VLEV_FLOAT(a_offset3, 1);
+            v4 = VLEV_FLOAT(a_offset4, 1);
+
+            VSSEG4_FLOAT(b_offset3, v1, v2, v3, v4, 1);
+
+            b_offset3 += 4;
+        }
+
+    }
+
+// TODO cleanup
+
+  if (m & 2){
+    a_offset1  = a_offset;
+    a_offset2  = a_offset1 + lda;
+    a_offset  += 2 * lda;
+
+    b_offset1  = b_offset;
+    b_offset  += 8;
+
+    i = (n >> 2);
+    if (i > 0){
+      do{
+	ctemp1  = *(a_offset1 + 0);
+	ctemp2  = *(a_offset1 + 1);
+	ctemp3  = *(a_offset1 + 2);
+	ctemp4  = *(a_offset1 + 3);
+
+	ctemp5  = *(a_offset2 + 0);
+	ctemp6  = *(a_offset2 + 1);
+	ctemp7  = *(a_offset2 + 2);
+	ctemp8  = *(a_offset2 + 3);
+
+	a_offset1 += 4;
+	a_offset2 += 4;
+
+	*(b_offset1 +  0) = ctemp1;
+	*(b_offset1 +  1) = ctemp2;
+	*(b_offset1 +  2) = ctemp3;
+	*(b_offset1 +  3) = ctemp4;
+
+	*(b_offset1 +  4) = ctemp5;
+	*(b_offset1 +  5) = ctemp6;
+	*(b_offset1 +  6) = ctemp7;
+	*(b_offset1 +  7) = ctemp8;
+
+	b_offset1 += m * 4;
+	i --;
+      }while(i > 0);
+    }
+
+    if (n & 2) {
+      ctemp1  = *(a_offset1 + 0);
+      ctemp2  = *(a_offset1 + 1);
+
+      ctemp3  = *(a_offset2 + 0);
+      ctemp4  = *(a_offset2 + 1);
+
+      a_offset1 += 2;
+      a_offset2 += 2;
+
+      *(b_offset2 +  0) = ctemp1;
+      *(b_offset2 +  1) = ctemp2;
+      *(b_offset2 +  2) = ctemp3;
+      *(b_offset2 +  3) = ctemp4;
+
+      b_offset2 += 4;
+    }
+
+    if (n & 1) {
+      ctemp1  = *(a_offset1 + 0);
+      ctemp2  = *(a_offset2 + 0);
+
+      *(b_offset3 +  0) = ctemp1;
+      *(b_offset3 +  1) = ctemp2;
+      b_offset3 += 2;
+    }
+  }
+
+  if (m & 1){
+    a_offset1  = a_offset;
+    b_offset1  = b_offset;
+
+    i = (n >> 2);
+    if (i > 0){
+      do{
+	ctemp1  = *(a_offset1 + 0);
+	ctemp2  = *(a_offset1 + 1);
+	ctemp3  = *(a_offset1 + 2);
+	ctemp4  = *(a_offset1 + 3);
+
+	a_offset1 += 4;
+
+	*(b_offset1 +  0) = ctemp1;
+	*(b_offset1 +  1) = ctemp2;
+	*(b_offset1 +  2) = ctemp3;
+	*(b_offset1 +  3) = ctemp4;
+
+	b_offset1 += 4 * m;
+
+	i --;
+      }while(i > 0);
+    }
+
+    if (n & 2) {
+      ctemp1  = *(a_offset1 + 0);
+      ctemp2  = *(a_offset1 + 1);
+      a_offset1 += 2;
+
+      *(b_offset2 +  0) = ctemp1;
+      *(b_offset2 +  1) = ctemp2;
+    }
+
+    if (n & 1) {
+      ctemp1  = *(a_offset1 + 0);
+      *(b_offset3 +  0) = ctemp1;
+    }
+  }
+
+  return 0;
+}

kernel/riscv64/gemm_tcopy_8_rvv.c

@@ -0,0 +1,264 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m1(n)
+#define FLOAT_V_T vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m1
+#define VLSEV_FLOAT vlse32_v_f32m1
+#define VSEV_FLOAT vse32_v_f32m1
+#define VLSSEG2_FLOAT vlsseg2e32_v_f32m1
+#define VSSEG2_FLOAT vsseg2e32_v_f32m1
+#define VLSSEG4_FLOAT vlsseg4e32_v_f32m1
+#define VSSEG4_FLOAT vsseg4e32_v_f32m1
+#define VLSSEG8_FLOAT vlsseg8e32_v_f32m1
+#define VSSEG8_FLOAT vsseg8e32_v_f32m1
+#else
+#define VSETVL(n) vsetvl_e64m1(n)
+#define FLOAT_V_T vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m1
+#define VLSEV_FLOAT vlse64_v_f64m1
+#define VSEV_FLOAT vse64_v_f64m1
+#define VLSSEG2_FLOAT vlsseg2e64_v_f64m1
+#define VSSEG2_FLOAT vsseg2e64_v_f64m1
+#define VLSSEG4_FLOAT vlsseg4e64_v_f64m1
+#define VSSEG4_FLOAT vsseg4e64_v_f64m1
+#define VLSSEG8_FLOAT vlsseg8e64_v_f64m1
+#define VSSEG8_FLOAT vsseg8e64_v_f64m1
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
+{
+    BLASLONG i, j;
+
+    IFLOAT *aoffset;
+    IFLOAT *aoffset1;
+
+    IFLOAT *boffset, *boffset1, *boffset2, *boffset3, *boffset4;
+
+    FLOAT_V_T v0, v1, v2, v3, v4, v5, v6, v7;
+
+    // fprintf(stderr, "gemm_tcopy_8 m=%ld n=%ld lda=%ld\n", m, n, lda);
+
+    aoffset   = a;
+    boffset   = b;
+    boffset2  = b + m  * (n & ~7);
+    boffset3  = b + m  * (n & ~3);
+    boffset4  = b + m  * (n & ~1);
+
+    for(j = (m >> 3); j > 0; j--) {
+
+        aoffset1  = aoffset;
+        aoffset += 8 * lda;
+
+        boffset1  = boffset;
+        boffset  += 64;
+
+        for(i = (n >> 3); i > 0; i--) {
+            size_t vl = 8;
+
+            VLSSEG8_FLOAT(&v0, &v1, &v2, &v3, &v4, &v5, &v6, &v7, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG8_FLOAT(boffset1, v0, v1, v2, v3, v4, v5, v6, v7, vl);
+
+            aoffset1 += 8;
+            boffset1 += m * 8;
+        }
+
+        if (n & 4) {
+            size_t vl = 8;
+
+            VLSSEG4_FLOAT(&v0, &v1, &v2, &v3, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG4_FLOAT(boffset2, v0, v1, v2, v3, vl);
+
+            aoffset1 += 4;
+            boffset2 += 32;
+        }
+
+        if (n & 2) {
+            size_t vl = 8;
+
+            VLSSEG2_FLOAT(&v0, &v1, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG2_FLOAT(boffset3, v0, v1, vl);
+
+            aoffset1 += 2;
+            boffset3 += 16;
+        }
+
+        if (n & 1) {
+            size_t vl = 8;
+
+            v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
+            VSEV_FLOAT(boffset4, v0, vl);
+
+            aoffset1 += 1;
+            boffset4 += 8;
+        }
+
+    }
+
+    if (m & 4) {
+
+        aoffset1  = aoffset;
+        aoffset += 4 * lda;
+
+        boffset1  = boffset;
+        boffset  += 32;
+
+        for(i = (n >> 3); i > 0; i--) {
+            size_t vl = 4;
+
+            VLSSEG8_FLOAT(&v0, &v1, &v2, &v3, &v4, &v5, &v6, &v7, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG8_FLOAT(boffset1, v0, v1, v2, v3, v4, v5, v6, v7, vl);
+
+            aoffset1 += 8;
+            boffset1 += m * 8;
+        }
+
+        if (n & 4) {
+            size_t vl = 4;
+
+            VLSSEG4_FLOAT(&v0, &v1, &v2, &v3, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG4_FLOAT(boffset2, v0, v1, v2, v3, vl);
+
+            aoffset1 += 4;
+            boffset2 += 16;
+        }
+
+        if (n & 2) {
+            size_t vl = 4;
+
+            VLSSEG2_FLOAT(&v0, &v1, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG2_FLOAT(boffset3, v0, v1, vl);
+
+            aoffset1 += 2;
+            boffset3 += 8;
+	}
+
+        if (n & 1) {
+            size_t vl = 4;
+
+            v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
+            VSEV_FLOAT(boffset4, v0, vl);
+
+            aoffset1 += 1;
+            boffset4 += 4;
+        }
+    }
+
+    if (m & 2) {
+        aoffset1  = aoffset;
+        aoffset += 2 * lda;
+
+        boffset1  = boffset;
+        boffset  += 16;
+
+        for(i = (n >> 3); i > 0; i--) {
+            size_t vl = 2;
+
+            VLSSEG8_FLOAT(&v0, &v1, &v2, &v3, &v4, &v5, &v6, &v7, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG8_FLOAT(boffset1, v0, v1, v2, v3, v4, v5, v6, v7, vl);
+
+            aoffset1 += 8;
+            boffset1 += m * 8;
+        }
+
+        if (n & 4) {
+            size_t vl = 2;
+
+            VLSSEG4_FLOAT(&v0, &v1, &v2, &v3, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG4_FLOAT(boffset2, v0, v1, v2, v3, vl);
+
+            aoffset1 += 4;
+            boffset2 += 8;
+        }
+
+        if (n & 2) {
+            size_t vl = 2;
+
+            VLSSEG2_FLOAT(&v0, &v1, aoffset1, lda * sizeof(FLOAT), vl);
+            VSSEG2_FLOAT(boffset3, v0, v1, vl);
+
+            aoffset1 += 2;
+            boffset3 += 4;
+	}
+  
+        if (n & 1) {
+           size_t vl = 2;
+
+            v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
+            VSEV_FLOAT(boffset4, v0, vl);
+
+            aoffset1 += 1;
+            boffset4 += 2;
+        }
+    }
+
+    if (m & 1) {
+        aoffset1  = aoffset;
+        boffset1  = boffset;
+
+        for(i = (n >> 3); i > 0; i--) {
+            size_t vl = 8;
+
+            v0 = VLEV_FLOAT(aoffset1, vl);
+            VSEV_FLOAT(boffset1, v0, vl);
+
+            aoffset1 += 8;
+            boffset1 += 8 * m;
+        }
+
+        if (n & 4) {
+            size_t vl = 4;
+
+            v0 = VLEV_FLOAT(aoffset1, vl);
+            VSEV_FLOAT(boffset2, v0, vl);
+
+            aoffset1 += 4;
+            //boffset2 += 4;
+        }
+
+        if (n & 2) {
+            size_t vl = 2;
+
+            v0 = VLEV_FLOAT(aoffset1, vl);
+            VSEV_FLOAT(boffset3, v0, vl);
+
+            aoffset1 += 2;
+           // boffset3 += 2;
+        }
+
+        if (n & 1) {
+           *(boffset4) = *(aoffset1);
+           // aoffset1 ++;
+           // boffset4 ++;
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/gemm_tcopy_rvv_v1.c

@@ -0,0 +1,74 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
+{
+    BLASLONG i, j;
+
+    IFLOAT *aoffset;
+    IFLOAT *aoffset1;
+    IFLOAT *boffset;
+
+    FLOAT_V_T v0;
+    size_t vl;
+
+    //fprintf(stderr, "%s, m=%ld n=%ld lda=%ld\n", __FUNCTION__, m, n, lda);
+
+    aoffset = a;
+    boffset = b;
+
+    for(j = n; j > 0; j -= vl) {
+        vl = VSETVL(j);
+
+        aoffset1 = aoffset;
+        aoffset += vl;
+
+        for(i = m; i > 0; i--) {
+            v0 = VLEV_FLOAT(aoffset1, vl);
+            VSEV_FLOAT(boffset, v0, vl);
+
+            aoffset1 += lda;
+            boffset += vl;
+        }  
+    }
+
+    return 0;
+}

kernel/riscv64/gemmkernel_2x2_rvv.c

@@ -0,0 +1,214 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VLSEG2_FLOAT vlseg2e32_v_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFMACCVV_FLOAT vfmacc_vv_f32m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VLSEG2_FLOAT vlseg2e64_v_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFMACCVV_FLOAT vfmacc_vv_f64m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+// Optimizes the implementation in ../generic/gemm_kernel_2x2.c
+
+int CNAME(BLASLONG bm, BLASLONG bn, BLASLONG bk, FLOAT alpha, IFLOAT* ba, IFLOAT* bb, FLOAT* C, BLASLONG ldc
+#ifdef TRMMKERNEL
+		,BLASLONG offset
+#endif
+		)
+{
+    BLASLONG i,j,k;
+    FLOAT *C0,*C1;
+    IFLOAT *ptrba,*ptrbb;
+
+    //fprintf(stderr, "gemm_kernel_2x2 bm=%ld bn=%ld bk=%ld alpha=%f ldc=%ld\n", bm, bn, bk, alpha, ldc);
+
+    FLOAT_V_T va0, va1, vb0, vb1;
+    FLOAT_V_T vres0, vres1, vres2, vres3;
+    FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3;
+    FLOAT_V_T_M1 v_z0;
+
+    v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    size_t vl;
+
+    for (j = bn/2; j > 0; j--) {
+        C0 = C;
+        C1 = C0 + ldc;
+        ptrba = ba;
+
+        for (i = bm/2; i > 0; i--) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl); 
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl); 
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);
+
+                ptrba += vl*2;
+                ptrbb += vl*2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[1] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 2;
+            C1 += 2;
+        }
+
+        if(bm & 1) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                ptrba += vl;
+                ptrbb += vl*2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            C0 += 1;
+            C1 += 1;
+        }
+
+        bb += (bk<<1);
+        C += (ldc<<1);
+    }
+
+    if(bn & 1) {
+        C0 = C;
+        ptrba = ba;
+        for (i = bm/2; i > 0; i--) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+
+                ptrba += vl*2;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            C0 += 2;
+        }
+
+        if(bm & 1) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                va0 = VLEV_FLOAT(ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+
+                ptrba += vl;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+
+            C0 += 1;
+        }
+
+        bb += (bk<<0);
+        C += ldc;
+    }
+
+    return 0;
+}

kernel/riscv64/gemmkernel_4x4_rvv.c

@@ -0,0 +1,508 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m1(n)
+#define VSETVL_MAX vsetvlmax_e32m1()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m1_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m1
+#define VLSEG2_FLOAT vlseg2e32_v_f32m1
+#define VLSEG4_FLOAT vlseg4e32_v_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m1
+#define VFMACCVF_FLOAT vfmacc_vf_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m1
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m1_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m1(n)
+#define VSETVL_MAX vsetvlmax_e64m1()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m1_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m1
+#define VLSEG2_FLOAT vlseg2e64_v_f64m1
+#define VLSEG4_FLOAT vlseg4e64_v_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m1
+#define VFMACCVF_FLOAT vfmacc_vf_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m1
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m1_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+// Optimizes the implementation in ../generic/gemm_kernel_2x2.c
+
+int CNAME(BLASLONG bm, BLASLONG bn, BLASLONG bk, FLOAT alpha, IFLOAT* ba, IFLOAT* bb, FLOAT* C, BLASLONG ldc
+#ifdef TRMMKERNEL
+		,BLASLONG offset
+#endif
+		)
+{
+    BLASLONG i,j,k;
+    FLOAT *C0,*C1,*C2,*C3;
+    IFLOAT *ptrba,*ptrbb;
+
+    //fprintf(stderr, "gemm_kernel_4x4 bm=%ld bn=%ld bk=%ld alpha=%f ldc=%ld\n", bm, bn, bk, alpha, ldc); // KU
+
+    FLOAT_V_T va0, va1, va2, va3; 
+    FLOAT_V_T vb0, vb1, vb2, vb3;
+    FLOAT_V_T vres0, vres1, vres2, vres3, vres4, vres5, vres6, vres7;
+    FLOAT_V_T vres8, vres9, vres10, vres11, vres12, vres13, vres14, vres15;
+    FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3;
+    FLOAT_V_T_M1 v_z0;
+
+    v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    size_t vl;
+
+    for (j = bn/4; j > 0; j--) {
+        C0 = C;
+        C1 = C0 + ldc;
+        C2 = C1 + ldc;
+        C3 = C2 + ldc;
+        ptrba = ba;
+
+        for (i = bm/4; i > 0; i--) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+            vres4 = VFMVVF_FLOAT(0.0, vlmax);
+            vres5 = VFMVVF_FLOAT(0.0, vlmax);
+            vres6 = VFMVVF_FLOAT(0.0, vlmax);
+            vres7 = VFMVVF_FLOAT(0.0, vlmax);
+            vres8 = VFMVVF_FLOAT(0.0, vlmax);
+            vres9 = VFMVVF_FLOAT(0.0, vlmax);
+            vres10 = VFMVVF_FLOAT(0.0, vlmax);
+            vres11 = VFMVVF_FLOAT(0.0, vlmax);
+            vres12 = VFMVVF_FLOAT(0.0, vlmax);
+            vres13 = VFMVVF_FLOAT(0.0, vlmax);
+            vres14 = VFMVVF_FLOAT(0.0, vlmax);
+            vres15 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl); 
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl); 
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va0, vb3, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va1, vb3, vl);
+
+                vres8 = VFMACCVV_FLOAT(vres8, va2, vb0, vl);
+                vres9 = VFMACCVV_FLOAT(vres9, va3, vb0, vl);
+                vres10 = VFMACCVV_FLOAT(vres10, va2, vb1, vl);
+                vres11 = VFMACCVV_FLOAT(vres11, va3, vb1, vl);
+
+                vres12 = VFMACCVV_FLOAT(vres12, va2, vb2, vl);
+                vres13 = VFMACCVV_FLOAT(vres13, va3, vb2, vl);
+                vres14 = VFMACCVV_FLOAT(vres14, va2, vb3, vl);
+                vres15 = VFMACCVV_FLOAT(vres15, va3, vb3, vl);
+
+                ptrba += vl*4;
+                ptrbb += vl*4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres8, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres9, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres10, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres11, v_z0, vlmax);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres12, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres13, v_z0, vlmax);
+            C2[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C2[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C2[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C2[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres6, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres7, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres14, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres15, v_z0, vlmax);
+            C3[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C3[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C3[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C3[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 4;
+            C1 += 4;
+            C2 += 4;
+            C3 += 4;
+        }
+
+        if(bm & 2) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+            vres4 = VFMVVF_FLOAT(0.0, vlmax);
+            vres5 = VFMVVF_FLOAT(0.0, vlmax);
+            vres6 = VFMVVF_FLOAT(0.0, vlmax);
+            vres7 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl); 
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl); 
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va0, vb3, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va1, vb3, vl);
+
+                ptrba += vl*2;
+                ptrbb += vl*4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+ 
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3, v_z0, vlmax);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
+            C2[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C2[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres6, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres7, v_z0, vlmax);
+            C3[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C3[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            C0 += 2;
+            C1 += 2;
+            C2 += 2;
+            C3 += 2;
+        }
+
+        if(bm & 1) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl); 
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va0, vb3, vl);
+
+                ptrba += vl;
+                ptrbb += vl*4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C2[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C3[0] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 1;
+            C1 += 1;
+            C2 += 1;
+            C3 += 1;
+        }
+
+        bb += (bk<<2);
+        C += (ldc<<2);
+    }
+
+    if(bn & 2) {
+
+        C0 = C;
+        C1 = C0 + ldc;
+        ptrba = ba;
+
+        for (i = bm/4; i > 0; i--) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+            vres4 = VFMVVF_FLOAT(0.0, vlmax);
+            vres5 = VFMVVF_FLOAT(0.0, vlmax);
+            vres6 = VFMVVF_FLOAT(0.0, vlmax);
+            vres7 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb1, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va1, vb1, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va2, vb1, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va3, vb1, vl);
+
+                ptrba += vl*4;
+                ptrbb += vl*2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres4, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres5, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres6, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres7, v_z0, vlmax);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 4;
+            C1 += 4;
+        }
+
+        if(bm & 2) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);
+
+                ptrba += vl*2;
+                ptrbb += vl*2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[1] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 2;
+            C1 += 2;
+        }
+
+        if(bm & 1) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                ptrba += vl;
+                ptrbb += vl*2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[0] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            C0 += 1;
+            C1 += 1;
+        }
+
+        bb += (bk<<1);
+        C += (ldc<<1);
+    }
+
+    if(bn & 1) {
+        C0 = C;
+        ptrba = ba;
+        for (i = bm/4; i > 0; i--) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+
+                ptrba += vl*4;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] += alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] += alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            C0 += 4;
+        }
+
+        if(bm & 2) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+
+                ptrba += vl*2;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] += alpha * VFMVFS_FLOAT_M1(vsum1);
+
+            C0 += 2;
+        }
+
+        if(bm & 1) {
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = bk; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                va0 = VLEV_FLOAT(ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+
+                ptrba += vl;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0, v_z0, vlmax);
+            C0[0] += alpha * VFMVFS_FLOAT_M1(vsum0);
+
+            C0 += 1;
+        }
+
+        bb += (bk<<0);
+        C += ldc;
+    }
+
+    return 0;
+}

kernel/riscv64/gemmkernel_rvv_v1x8.c

@@ -0,0 +1,601 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VFMVVF_FLOAT vfmv_v_f_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VFMVVF_FLOAT vfmv_v_f_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#endif
+
+int CNAME(BLASLONG bm, BLASLONG bn, BLASLONG bk, FLOAT alpha, IFLOAT* ba, IFLOAT* bb, FLOAT* C, BLASLONG ldc
+#ifdef TRMMKERNEL
+		,BLASLONG offset
+#endif
+		)
+{
+    BLASLONG i,j,k;
+    FLOAT *C0,*C1,*C2,*C3,*C4,*C5,*C6,*C7;
+    IFLOAT *ptrba,*ptrbb;
+
+    //fprintf(stderr, "%s, bm=%ld bn=%ld bk=%ld alpha=%f ldc=%ld\n", __FUNCTION__, bm, bn, bk, alpha, ldc); // Debug
+
+    FLOAT_V_T va0, va1, va2, va3, va4, va5, va6, va7;
+    FLOAT_V_T vres0, vres1, vres2, vres3, vres4, vres5, vres6, vres7;
+    size_t vl;
+
+    // N:8
+    for (j = bn/8; j > 0; j--) {
+        C0 = C;
+        C1 = C0 + ldc;
+        C2 = C1 + ldc;
+        C3 = C2 + ldc;
+        C4 = C3 + ldc;
+        C5 = C4 + ldc;
+        C6 = C5 + ldc;
+        C7 = C6 + ldc;
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+            vres2 = VFMVVF_FLOAT(0.0, vl);
+            vres3 = VFMVVF_FLOAT(0.0, vl);
+            vres4 = VFMVVF_FLOAT(0.0, vl);
+            vres5 = VFMVVF_FLOAT(0.0, vl);
+            vres6 = VFMVVF_FLOAT(0.0, vl);
+            vres7 = VFMVVF_FLOAT(0.0, vl);
+#if 0
+            for (k = bk; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl); 
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va0, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va0, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va0, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va0, vl);
+
+                ptrba += vl;
+                ptrbb += 8;
+            }
+#else
+            // Unroll K
+            for (k = bk/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+ 
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va0, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va0, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va0, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va0, vl);
+                ptrbb += 8;
+  
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va1, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va1, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va1, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va1, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va1, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va1, vl);
+                ptrbb += 8;
+
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va2, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va2, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va2, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va2, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va2, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va2, vl);
+                ptrbb += 8;
+
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va3, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va3, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va3, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va3, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va3, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va3, vl);
+                ptrbb += 8;
+
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va4, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va4, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va4, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va4, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va4, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va4, vl);
+                ptrbb += 8;
+
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va5, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va5, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va5, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va5, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va5, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va5, vl);
+                ptrbb += 8;
+
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va6, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va6, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va6, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va6, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va6, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va6, vl);
+                ptrbb += 8;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va7, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va7, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va7, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va7, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va7, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va7, vl);
+                ptrbb += 8;
+            }
+
+            // K remainder
+            for (k = bk&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va0, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va0, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va0, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va0, vl);
+
+                ptrbb += 8;
+                ptrba += vl;
+            }
+#endif
+            va0 = VLEV_FLOAT(C0, vl);
+            va0 = VFMACCVF_FLOAT(va0, alpha, vres0, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VLEV_FLOAT(C1, vl);
+            va1 = VFMACCVF_FLOAT(va1, alpha, vres1, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+            va2 = VLEV_FLOAT(C2, vl);
+            va2 = VFMACCVF_FLOAT(va2, alpha, vres2, vl);
+            VSEV_FLOAT(C2, va2, vl);
+
+            va3 = VLEV_FLOAT(C3, vl);
+            va3 = VFMACCVF_FLOAT(va3, alpha, vres3, vl);
+            VSEV_FLOAT(C3, va3, vl);
+
+            va4 = VLEV_FLOAT(C4, vl);
+            va4 = VFMACCVF_FLOAT(va4, alpha, vres4, vl);
+            VSEV_FLOAT(C4, va4, vl);
+
+            va5 = VLEV_FLOAT(C5, vl);
+            va5 = VFMACCVF_FLOAT(va5, alpha, vres5, vl);
+            VSEV_FLOAT(C5, va5, vl);
+
+            va6 = VLEV_FLOAT(C6, vl);
+            va6 = VFMACCVF_FLOAT(va6, alpha, vres6, vl);
+            VSEV_FLOAT(C6, va6, vl);
+
+            va7 = VLEV_FLOAT(C7, vl);
+            va7 = VFMACCVF_FLOAT(va7, alpha, vres7, vl);
+            VSEV_FLOAT(C7, va7, vl);
+
+            C0 += vl;
+            C1 += vl;
+            C2 += vl;
+            C3 += vl;
+            C4 += vl;
+            C5 += vl;
+            C6 += vl;
+            C7 += vl;
+        }
+
+        bb += (bk<<3);
+        C += (ldc<<3);
+    }
+
+    // N:4
+    if (bn & 4) {
+        C0 = C;
+        C1 = C0 + ldc;
+        C2 = C1 + ldc;
+        C3 = C2 + ldc;
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+            vres2 = VFMVVF_FLOAT(0.0, vl);
+            vres3 = VFMVVF_FLOAT(0.0, vl);
+
+#if 0
+            for (k = bk; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl); 
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+
+                ptrba += vl;
+                ptrbb += 4;
+            }
+#else
+            // Unroll K
+            for (k = bk/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                ptrbb += 4;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+  
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va1, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va1, vl);
+                ptrbb += 4;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va2, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va2, vl);
+                ptrbb += 4;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va3, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va3, vl);
+                ptrbb += 4;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va4, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va4, vl);
+                ptrbb += 4;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va5, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va5, vl);
+                ptrbb += 4;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va6, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va6, vl);
+                ptrbb += 4;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va7, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va7, vl);
+                ptrbb += 4;
+            }
+
+            // K remainder
+            for (k = bk&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+
+                ptrbb += 4;
+                ptrba += vl;
+            }
+#endif
+            va0 = VLEV_FLOAT(C0, vl);
+            va0 = VFMACCVF_FLOAT(va0, alpha, vres0, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VLEV_FLOAT(C1, vl);
+            va1 = VFMACCVF_FLOAT(va1, alpha, vres1, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+            va2 = VLEV_FLOAT(C2, vl);
+            va2 = VFMACCVF_FLOAT(va2, alpha, vres2, vl);
+            VSEV_FLOAT(C2, va2, vl);
+
+            va3 = VLEV_FLOAT(C3, vl);
+            va3 = VFMACCVF_FLOAT(va3, alpha, vres3, vl);
+            VSEV_FLOAT(C3, va3, vl);
+
+            C0 += vl;
+            C1 += vl;
+            C2 += vl;
+            C3 += vl;
+        }
+
+        bb += (bk<<2);
+        C += (ldc<<2);
+    }
+
+    // N:2
+    if (bn & 2) {
+        C0 = C;
+        C1 = C0 + ldc;
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+#if 0
+            for (k = bk; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl); 
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+
+                ptrba += vl;
+                ptrbb += 2;
+            }
+#else
+            // Unroll K
+            for (k = bk/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                ptrbb += 2;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                ptrbb += 2;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                ptrbb += 2;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                ptrbb += 2;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                ptrbb += 2;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                ptrbb += 2;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl; 
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                ptrbb += 2;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                ptrbb += 2;
+            }
+
+            // K remainder
+            for (k = bk&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+
+                ptrbb += 2;
+                ptrba += vl;
+            }
+#endif
+            va0 = VLEV_FLOAT(C0, vl);
+            va0 = VFMACCVF_FLOAT(va0, alpha, vres0, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VLEV_FLOAT(C1, vl);
+            va1 = VFMACCVF_FLOAT(va1, alpha, vres1, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+            C0 += vl;
+            C1 += vl;
+        }
+
+        bb += (bk<<1);
+        C += (ldc<<1);
+    }
+
+    // N:1
+    if (bn & 1) {
+        C0 = C;
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl) {
+            vl = VSETVL(i);
+
+            ptrbb = bb;
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+#if 0
+            for (k = bk; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl); 
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+ 
+                ptrba += vl;
+                ptrbb += 1;
+            }
+#else
+            // Unroll K
+            for (k = bk/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                ptrbb += 1;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                ptrbb += 1;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                ptrbb += 1;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                ptrbb += 1;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                ptrbb += 1;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                ptrbb += 1;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                ptrbb += 1;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                ptrbb += 1;
+            }
+
+            // K remainder
+            for (k = bk&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+
+                ptrbb += 1;
+                ptrba += vl;
+            }
+#endif
+            va0 = VLEV_FLOAT(C0, vl);
+            va0 = VFMACCVF_FLOAT(va0, alpha, vres0, vl);
+            VSEV_FLOAT(C0, va0, vl);
+  
+            C0 += vl;
+        }
+
+        bb += (bk);
+        C += (ldc);
+    }
+
+    return 0;
+}

kernel/riscv64/gemv_n_rvv.c

@@ -0,0 +1,94 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMACCVF_FLOAT vfmacc_vf_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMACCVF_FLOAT vfmacc_vf_f64m8
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+    if(n < 0)  return(0);
+
+    FLOAT *a_ptr, *x_ptr;
+    BLASLONG i;
+    FLOAT_V_T va, vy;
+
+    if(inc_y == 1) {
+
+        for (size_t vl; m > 0; m -= vl, y += vl, a += vl) {
+            vl = VSETVL(m);
+            a_ptr = a;
+            x_ptr = x;
+            vy = VLEV_FLOAT(y, vl);
+            for(i = 0; i < n; i++) {
+                va = VLEV_FLOAT(a_ptr, vl);
+                vy = VFMACCVF_FLOAT(vy, (alpha * (*x_ptr)), va, vl);
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+            VSEV_FLOAT(y, vy, vl);
+        }
+ 
+    } else {
+
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; m > 0; m -= vl, y += vl*inc_y, a += vl) {
+            vl = VSETVL(m);
+            a_ptr = a;
+            x_ptr = x;
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+            for(i = 0; i < n; i++) {
+                va = VLEV_FLOAT(a_ptr, vl);
+                vy = VFMACCVF_FLOAT(vy, (alpha * (*x_ptr)), va, vl);
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+            VSSEV_FLOAT(y, stride_y, vy, vl);
+        }
+
+    }
+    return(0);
+}

kernel/riscv64/gemv_t_rvv.c

@@ -0,0 +1,119 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDSUM_FLOAT vfredusum_vs_f32m8_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDSUM_FLOAT vfredusum_vs_f64m8_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+    BLASLONG i, j;
+    FLOAT *a_ptr, *x_ptr;
+
+    FLOAT_V_T va, vx, vr;
+    FLOAT_V_T_M1 v_res, v_z0;
+    size_t vlmax = VSETVL_MAX_M1;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+    vlmax = VSETVL_MAX;
+
+    if(inc_x == 1) {
+
+        for(i = 0; i < n; i++) {
+            j = m;
+            a_ptr = a;
+            x_ptr = x;
+            vr = VFMVVF_FLOAT(0, vlmax);
+
+            for (size_t vl; j > 0; j -= vl, a_ptr += vl, x_ptr += vl) {
+                vl = VSETVL(j);
+
+                va = VLEV_FLOAT(a_ptr, vl);
+                vx = VLEV_FLOAT(x_ptr, vl);
+                vr = VFMACCVV_FLOAT(vr, va, vx, vl);
+            }
+
+            v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+            *y += alpha * VFMVFS_FLOAT_M1(v_res);
+            y += inc_y;
+            a += lda;
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+  
+        for(i = 0; i < n; i++) {
+            j = m;
+            a_ptr = a;
+            x_ptr = x;
+            vr = VFMVVF_FLOAT(0, vlmax);
+
+            for (size_t vl; j > 0; j -= vl, a_ptr += vl, x_ptr += vl*inc_x) {
+                vl = VSETVL(j);
+
+                va = VLEV_FLOAT(a_ptr, vl);
+                vx = VLSEV_FLOAT(x_ptr, stride_x, vl);
+                vr = VFMACCVV_FLOAT(vr, va, vx, vl);
+            }
+
+            v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+            *y += alpha * VFMVFS_FLOAT_M1(v_res);
+            y += inc_y;
+            a += lda;
+        }
+
+    }
+
+    return(0);
+}

kernel/riscv64/iamax_rvv.c

@@ -0,0 +1,150 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
+#define VMFLTVV_FLOAT vmflt_vv_f64m8_b8
+#define VMFGEVF_FLOAT vmfge_vf_f64m8_b8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define VFMAXVV_FLOAT vfmax_vv_f64m8
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint64m8_t
+#define VIDV_MASK_UINT vid_v_u64m8_m
+#define VIDV_UINT vid_v_u64m8
+#define VADDVX_MASK_UINT vadd_vx_u64m8_m
+#define VADDVX_UINT vadd_vx_u64m8
+#define VMVVX_UINT vmv_v_x_u64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m8
+#define VMVVXS_UINT vmv_x_s_u64m8_u64
+#else
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
+#define MASK_T vbool4_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
+#define VMFLTVV_FLOAT vmflt_vv_f32m8_b4
+#define VMFGEVF_FLOAT vmfge_vf_f32m8_b4
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define VFMAXVV_FLOAT vfmax_vv_f32m8
+#define VFIRSTM vfirst_m_b4
+#define UINT_V_T vuint32m8_t
+#define VIDV_MASK_UINT vid_v_u32m8_m
+#define VIDV_UINT vid_v_u32m8
+#define VADDVX_MASK_UINT vadd_vx_u32m8_m
+#define VADDVX_UINT vadd_vx_u32m8
+#define VMVVX_UINT vmv_v_x_u32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m8
+#define VMVVXS_UINT vmv_x_s_u32m8_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int max_index = 0;
+    if (n <= 0 || inc_x <= 0) return(max_index);
+
+    FLOAT_V_T vx, v_max;
+    UINT_V_T v_max_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_max_index = VMVVX_UINT(0, vlmax);
+    v_max = VFMVVF_FLOAT(-1, vlmax);
+    BLASLONG j=0;
+    FLOAT maxf=0.0;
+    
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+
+            //update v_max
+            v_max = VFMAXVV_FLOAT(v_max, vx, vl);
+        }
+
+    } else {
+  
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+
+            //update v_max
+            v_max = VFMAXVV_FLOAT(v_max, vx, vl);
+        }
+  
+    }
+
+    FLOAT_V_T_M1 v_res, v_z0;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+
+    v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFGEVF_FLOAT(v_max, maxf, vlmax);
+    max_index = VFIRSTM(mask, vlmax);
+    
+    v_max_index = VSLIDEDOWN_UINT(v_max_index, v_max_index, max_index, vlmax);
+    max_index = VMVVXS_UINT(v_max_index);
+
+    return(max_index+1);
+}

kernel/riscv64/iamin_rvv.c

@@ -0,0 +1,151 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
+#define VMFLTVV_FLOAT vmflt_vv_f64m8_b8
+#define VMFLEVF_FLOAT vmfle_vf_f64m8_b8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define VFMINVV_FLOAT vfmin_vv_f64m8
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint64m8_t
+#define VIDV_MASK_UINT vid_v_u64m8_m
+#define VIDV_UINT vid_v_u64m8
+#define VADDVX_MASK_UINT vadd_vx_u64m8_m
+#define VADDVX_UINT vadd_vx_u64m8
+#define VMVVX_UINT vmv_v_x_u64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m8
+#define VMVVXS_UINT vmv_x_s_u64m8_u64
+#else
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
+#define MASK_T vbool4_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
+#define VMFLTVV_FLOAT vmflt_vv_f32m8_b4
+#define VMFLEVF_FLOAT vmfle_vf_f32m8_b4
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define VFMINVV_FLOAT vfmin_vv_f32m8
+#define VFIRSTM vfirst_m_b4
+#define UINT_V_T vuint32m8_t
+#define VIDV_MASK_UINT vid_v_u32m8_m
+#define VIDV_UINT vid_v_u32m8
+#define VADDVX_MASK_UINT vadd_vx_u32m8_m
+#define VADDVX_UINT vadd_vx_u32m8
+#define VMVVX_UINT vmv_v_x_u32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m8
+#define VMVVXS_UINT vmv_x_s_u32m8_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int min_index = 0;
+    if (n <= 0 || inc_x <= 0) return(min_index);
+
+    FLOAT_V_T vx, v_min;
+    UINT_V_T v_min_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_min_index = VMVVX_UINT(0, vlmax);
+    v_min = VFMVVF_FLOAT(FLT_MAX, vlmax);
+    BLASLONG j=0;
+    FLOAT minf=0.0;
+    
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx, vl);
+        }
+
+    } else {
+  
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vx = VFABSV_FLOAT(vx, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx, vl);
+        }
+  
+    }
+
+    FLOAT_V_T_M1 v_res, v_max;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_max = VFMVVF_FLOAT_M1(FLT_MAX, vlmax);
+
+    v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFLEVF_FLOAT(v_min, minf, vlmax);
+    min_index = VFIRSTM(mask, vlmax);
+
+    v_min_index = VSLIDEDOWN_UINT(v_min_index, v_min_index, min_index, vlmax);
+    min_index = VMVVXS_UINT(v_min_index);
+
+    return(min_index+1);
+}

kernel/riscv64/imax_rvv.c

@@ -0,0 +1,147 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
+#define VMFLTVV_FLOAT vmflt_vv_f64m8_b8
+#define VMFGEVF_FLOAT vmfge_vf_f64m8_b8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMAXVV_FLOAT vfmax_vv_f64m8
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint64m8_t
+#define VIDV_MASK_UINT vid_v_u64m8_m
+#define VIDV_UINT vid_v_u64m8
+#define VADDVX_MASK_UINT vadd_vx_u64m8_m
+#define VADDVX_UINT vadd_vx_u64m8
+#define VMVVX_UINT vmv_v_x_u64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m8
+#define VMVVXS_UINT vmv_x_s_u64m8_u64
+#else
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
+#define MASK_T vbool4_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
+#define VMFLTVV_FLOAT vmflt_vv_f32m8_b4
+#define VMFGEVF_FLOAT vmfge_vf_f32m8_b4
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMAXVV_FLOAT vfmax_vv_f32m8
+#define VFIRSTM vfirst_m_b4
+#define UINT_V_T vuint32m8_t
+#define VIDV_MASK_UINT vid_v_u32m8_m
+#define VIDV_UINT vid_v_u32m8
+#define VADDVX_MASK_UINT vadd_vx_u32m8_m
+#define VADDVX_UINT vadd_vx_u32m8
+#define VMVVX_UINT vmv_v_x_u32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m8
+#define VMVVXS_UINT vmv_x_s_u32m8_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int max_index = 0;
+    if (n <= 0 || inc_x <= 0) return(max_index);
+
+    FLOAT_V_T vx, v_max;
+    UINT_V_T v_max_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_max_index = VMVVX_UINT(0, vlmax);
+    v_max = VFMVVF_FLOAT(-FLT_MAX, vlmax);
+    BLASLONG j=0;
+    FLOAT maxf=0.0;
+    
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+
+            //update v_max and start_index j
+            v_max = VFMAXVV_FLOAT(v_max, vx, vl);
+        }
+
+    } else {
+  
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+
+            //update v_max and start_index j
+            v_max = VFMAXVV_FLOAT(v_max, vx, vl);
+        }
+  
+    }
+
+    FLOAT_V_T_M1 v_res, v_min;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_min = VFMVVF_FLOAT_M1(-FLT_MAX, vlmax);
+
+    v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFGEVF_FLOAT(v_max, maxf, vlmax);
+    max_index = VFIRSTM(mask, vlmax);
+    
+    v_max_index = VSLIDEDOWN_UINT(v_max_index, v_max_index, max_index, vlmax);
+    max_index = VMVVXS_UINT(v_max_index);
+
+    return(max_index+1);
+}

kernel/riscv64/imin_rvv.c

@@ -0,0 +1,147 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
+#define VMFLTVV_FLOAT vmflt_vv_f64m8_b8
+#define VMFLEVF_FLOAT vmfle_vf_f64m8_b8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMINVV_FLOAT vfmin_vv_f64m8
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint64m8_t
+#define VIDV_MASK_UINT vid_v_u64m8_m
+#define VIDV_UINT vid_v_u64m8
+#define VADDVX_MASK_UINT vadd_vx_u64m8_m
+#define VADDVX_UINT vadd_vx_u64m8
+#define VMVVX_UINT vmv_v_x_u64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m8
+#define VMVVXS_UINT vmv_x_s_u64m8_u64
+#else
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
+#define MASK_T vbool4_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
+#define VMFLTVV_FLOAT vmflt_vv_f32m8_b4
+#define VMFLEVF_FLOAT vmfle_vf_f32m8_b4
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMINVV_FLOAT vfmin_vv_f32m8
+#define VFIRSTM vfirst_m_b4
+#define UINT_V_T vuint32m8_t
+#define VIDV_MASK_UINT vid_v_u32m8_m
+#define VIDV_UINT vid_v_u32m8
+#define VADDVX_MASK_UINT vadd_vx_u32m8_m
+#define VADDVX_UINT vadd_vx_u32m8
+#define VMVVX_UINT vmv_v_x_u32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m8
+#define VMVVXS_UINT vmv_x_s_u32m8_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int min_index = 0;
+    if (n <= 0 || inc_x <= 0) return(min_index);
+
+    FLOAT_V_T vx, v_min;
+    UINT_V_T v_min_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_min_index = VMVVX_UINT(0, vlmax);
+    v_min = VFMVVF_FLOAT(FLT_MAX, vlmax);
+    BLASLONG j=0;
+    FLOAT minf=0.0;
+    
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx, vl);
+        }
+
+    } else {
+  
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, j += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx, vl);
+        }
+  
+    }
+
+    FLOAT_V_T_M1 v_res, v_max;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_max = VFMVVF_FLOAT_M1(FLT_MAX, vlmax);
+
+    v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFLEVF_FLOAT(v_min, minf, vlmax);
+    min_index = VFIRSTM(mask, vlmax);
+
+    v_min_index = VSLIDEDOWN_UINT(v_min_index, v_min_index, min_index, vlmax);
+    min_index = VMVVXS_UINT(v_min_index);
+
+    return(min_index+1);
+}

kernel/riscv64/izamax_rvv.c

@@ -0,0 +1,162 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VLSEV_FLOAT vlse64_v_f64m4
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m4_f64m1
+#define MASK_T vbool16_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m4_b16
+#define VMFLTVV_FLOAT vmflt_vv_f64m4_b16
+#define VMFGEVF_FLOAT vmfge_vf_f64m4_b16
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFABSV_FLOAT vfabs_v_f64m4
+#define VFMAXVV_FLOAT vfmax_vv_f64m4
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#define VFIRSTM vfirst_m_b16
+#define UINT_V_T vuint64m4_t
+#define VIDV_MASK_UINT vid_v_u64m4_m
+#define VIDV_UINT vid_v_u64m4
+#define VADDVX_MASK_UINT vadd_vx_u64m4_m
+#define VADDVX_UINT vadd_vx_u64m4
+#define VMVVX_UINT vmv_v_x_u64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m4
+#define VMVVXS_UINT vmv_x_s_u64m4_u64
+#else
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VLSEV_FLOAT vlse32_v_f32m4
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m4_f32m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m4_b8
+#define VMFLTVV_FLOAT vmflt_vv_f32m4_b8
+#define VMFGEVF_FLOAT vmfge_vf_f32m4_b8
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFABSV_FLOAT vfabs_v_f32m4
+#define VFMAXVV_FLOAT vfmax_vv_f32m4
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint32m4_t
+#define VIDV_MASK_UINT vid_v_u32m4_m
+#define VIDV_UINT vid_v_u32m4
+#define VADDVX_MASK_UINT vadd_vx_u32m4_m
+#define VADDVX_UINT vadd_vx_u32m4
+#define VMVVX_UINT vmv_v_x_u32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m4
+#define VMVVXS_UINT vmv_x_s_u32m4_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int max_index = 0;
+    if (n <= 0 || inc_x <= 0) return(max_index);
+
+    FLOAT_V_T vx0, vx1, v_max;
+    UINT_V_T v_max_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_max_index = VMVVX_UINT(0, vlmax);
+    v_max = VFMVVF_FLOAT(-1, vlmax);
+    BLASLONG j=0;
+    FLOAT maxf=0.0;
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, j += vl) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+
+            vx0 = VFABSV_FLOAT(vx0, vl);
+            vx1 = VFABSV_FLOAT(vx1, vl);
+
+            vx0 = VFADDVV_FLOAT(vx0, vx1, vl);
+
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx0, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+
+            //update v_max and start_index j
+            v_max = VFMAXVV_FLOAT(v_max, vx0, vl);
+        }
+    }
+    else {
+        
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, j += vl) {
+            vl = VSETVL(n);
+        
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+        
+            vx0 = VFABSV_FLOAT(vx0, vl);
+            vx1 = VFABSV_FLOAT(vx1, vl);
+        
+            vx0 = VFADDVV_FLOAT(vx0, vx1, vl);
+        
+            //index where element greater than v_max
+            mask = VMFLTVV_FLOAT(v_max, vx0, vl);
+            v_max_index = VIDV_MASK_UINT(mask, v_max_index, vl);
+            v_max_index = VADDVX_MASK_UINT(mask, v_max_index, v_max_index, j, vl);
+        
+            //update v_max and start_index j
+            v_max = VFMAXVV_FLOAT(v_max, vx0, vl);
+        }
+
+    }
+    FLOAT_V_T_M1 v_res, v_z0;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+
+    v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFGEVF_FLOAT(v_max, maxf, vlmax);
+    max_index = VFIRSTM(mask, vlmax);
+    
+    v_max_index = VSLIDEDOWN_UINT(v_max_index, v_max_index, max_index, vlmax);
+    max_index = VMVVXS_UINT(v_max_index);
+
+    return(max_index+1);
+}

kernel/riscv64/izamin_rvv.c

@@ -0,0 +1,161 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if defined(DOUBLE)
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m4_f64m1
+#define MASK_T vbool16_t
+#define VMFLTVF_FLOAT vmflt_vf_f64m4_b16
+#define VMFLTVV_FLOAT vmflt_vv_f64m4_b16
+#define VMFLEVF_FLOAT vmfle_vf_f64m4_b16
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFABSV_FLOAT vfabs_v_f64m4
+#define VFMINVV_FLOAT vfmin_vv_f64m4
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#define VFIRSTM vfirst_m_b16
+#define UINT_V_T vuint64m4_t
+#define VIDV_MASK_UINT vid_v_u64m4_m
+#define VIDV_UINT vid_v_u64m4
+#define VADDVX_MASK_UINT vadd_vx_u64m4_m
+#define VADDVX_UINT vadd_vx_u64m4
+#define VMVVX_UINT vmv_v_x_u64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VSLIDEDOWN_UINT vslidedown_vx_u64m4
+#define VMVVXS_UINT vmv_x_s_u64m4_u64
+#else
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m4_f32m1
+#define MASK_T vbool8_t
+#define VMFLTVF_FLOAT vmflt_vf_f32m4_b8
+#define VMFLTVV_FLOAT vmflt_vv_f32m4_b8
+#define VMFLEVF_FLOAT vmfle_vf_f32m4_b8
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFABSV_FLOAT vfabs_v_f32m4
+#define VFMINVV_FLOAT vfmin_vv_f32m4
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#define VFIRSTM vfirst_m_b8
+#define UINT_V_T vuint32m4_t
+#define VIDV_MASK_UINT vid_v_u32m4_m
+#define VIDV_UINT vid_v_u32m4
+#define VADDVX_MASK_UINT vadd_vx_u32m4_m
+#define VADDVX_UINT vadd_vx_u32m4
+#define VMVVX_UINT vmv_v_x_u32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VSLIDEDOWN_UINT vslidedown_vx_u32m4
+#define VMVVXS_UINT vmv_x_s_u32m4_u32
+#endif
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    unsigned int min_index = 0;
+    if (n <= 0 || inc_x <= 0) return(min_index);
+
+    FLOAT_V_T vx0, vx1, v_min;
+    UINT_V_T v_min_index;
+    MASK_T mask;
+  
+    size_t vlmax = VSETVL_MAX;
+    v_min_index = VMVVX_UINT(0, vlmax);
+    v_min = VFMVVF_FLOAT(FLT_MAX, vlmax);
+    BLASLONG j=0;
+    FLOAT minf=0.0;
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, j += vl) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+
+            vx0 = VFABSV_FLOAT(vx0, vl);
+            vx1 = VFABSV_FLOAT(vx1, vl);
+
+            vx0 = VFADDVV_FLOAT(vx0, vx1, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx0, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx0, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, j += vl) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+
+            vx0 = VFABSV_FLOAT(vx0, vl);
+            vx1 = VFABSV_FLOAT(vx1, vl);
+
+            vx0 = VFADDVV_FLOAT(vx0, vx1, vl);
+
+            // index where element less than v_min
+            mask = VMFLTVV_FLOAT(vx0, v_min, vl);
+            v_min_index = VIDV_MASK_UINT(mask, v_min_index, vl);
+            v_min_index = VADDVX_MASK_UINT(mask, v_min_index, v_min_index, j, vl);
+
+            //update v_min and start_index j
+            v_min = VFMINVV_FLOAT(v_min, vx0, vl);
+        }
+
+    }
+
+    FLOAT_V_T_M1 v_res, v_max;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_max = VFMVVF_FLOAT_M1(FLT_MAX, vlmax);
+
+    v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+    mask = VMFLEVF_FLOAT(v_min, minf, vlmax);
+    min_index = VFIRSTM(mask, vlmax);
+
+    v_min_index = VSLIDEDOWN_UINT(v_min_index, v_min_index, min_index, vlmax);
+    min_index = VMVVXS_UINT(v_min_index);
+
+    return(min_index+1);
+}

kernel/riscv64/max_rvv.c

@@ -0,0 +1,98 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMAXVV_FLOAT vfmax_vv_f32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMAXVV_FLOAT vfmax_vv_f64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT maxf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(maxf);
+
+    FLOAT_V_T vx, vmax;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(-FLT_MAX, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmax = VFMVVF_FLOAT(-FLT_MAX, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vmax = VFMAXVV_FLOAT(vmax, vx, vl);
+       }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vmax = VFMAXVV_FLOAT(vmax, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDMAXVS_FLOAT(v_res, vmax, v_res, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+
+    return(maxf);
+}

kernel/riscv64/min_rvv.c

@@ -0,0 +1,98 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMINVV_FLOAT vfmin_vv_f32m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMINVV_FLOAT vfmin_vv_f64m8
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT minf = 0.0;
+
+    if (n <= 0 || inc_x <= 0) return(minf);
+
+    FLOAT_V_T vx, vmin;
+    FLOAT_V_T_M1 v_res;
+    
+    v_res = VFMVVF_FLOAT_M1(FLT_MAX, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmin = VFMVVF_FLOAT(FLT_MAX, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vmin = VFMINVV_FLOAT(vmin, vx, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vmin = VFMINVV_FLOAT(vmin, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDMINVS_FLOAT(v_res, vmin, v_res, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+
+    return(minf);
+}

kernel/riscv64/nrm2_rvv.c

@@ -0,0 +1,117 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <math.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDSUM_FLOAT vfredusum_vs_f32m8_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VFABSV_FLOAT vfabs_v_f32m8
+#define ABS fabsf
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDSUM_FLOAT vfredusum_vs_f64m8_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VFABSV_FLOAT vfabs_v_f64m8
+#define ABS fabs
+#endif
+
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+
+    if( n <= 0 ) return(0.0);
+    if(n == 1) return (ABS(x[0]));
+
+    FLOAT_V_T vr, v0;
+    FLOAT_V_T_M1 v_max, v_res;
+    FLOAT scale = 0.0, ssq = 0.0;
+
+    size_t vlmax = VSETVL_MAX;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_max = VFMVVF_FLOAT_M1(0, vlmax);
+
+    vr = VFMVVF_FLOAT(0, vlmax);
+ 
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            v0 = VLEV_FLOAT(x, vl);
+            v0 = VFABSV_FLOAT(v0, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v0, v_max, vl);
+
+            vr = VFMACCVV_FLOAT(vr, v0, v0, vl);
+        }
+
+    } else {
+ 
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl * inc_x) {
+            vl = VSETVL(n);
+
+            v0 = VLSEV_FLOAT(x, stride_x, vl);
+            v0 = VFABSV_FLOAT(v0, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v0, v_max, vl);
+
+            vr = VFMACCVV_FLOAT(vr, v0, v0, vl);
+        }
+
+    }
+
+    v_res = VFREDSUM_FLOAT(v_res, vr, v_res, vlmax);
+
+    ssq = VFMVFS_FLOAT_M1(v_res);
+    scale = VFMVFS_FLOAT_M1(v_max);
+    ssq = ssq / (scale*scale);
+
+    return(scale * sqrt(ssq));
+}

kernel/riscv64/rot_rvv.c

@@ -0,0 +1,149 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMACCVF_FLOAT vfmacc_vf_f32m8
+#define VFMULVF_FLOAT vfmul_vf_f32m8
+#define VFMSACVF_FLOAT vfmsac_vf_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMACCVF_FLOAT vfmacc_vf_f64m8
+#define VFMULVF_FLOAT vfmul_vf_f64m8
+#define VFMSACVF_FLOAT vfmsac_vf_f64m8
+#endif
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+
+    if(n <= 0) return(0);
+
+    FLOAT_V_T v0, v1, vx, vy;
+
+    if (inc_x == 0 || inc_y == 0) {
+        BLASLONG i=0;
+        BLASLONG ix=0,iy=0;
+        FLOAT temp;
+        while(i < n)
+        {
+            temp   = c*x[ix] + s*y[iy] ;
+            y[iy]  = c*y[iy] - s*x[ix] ;
+            x[ix]  = temp ;
+
+            ix += inc_x ;
+            iy += inc_y ;
+            i++ ;
+        }
+    }
+    else if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLEV_FLOAT(y, vl);
+
+            v0 = VFMULVF_FLOAT(vx, c, vl);
+            v0 = VFMACCVF_FLOAT(v0, s, vy, vl);
+            VSEV_FLOAT(x, v0, vl);
+
+            v1 = VFMULVF_FLOAT(vx, s, vl);
+            v1 = VFMSACVF_FLOAT(v1, c, vy, vl);
+            VSEV_FLOAT(y, v1, vl);
+        }
+
+    } else if(inc_y == 1) {
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLEV_FLOAT(y, vl);
+
+            v0 = VFMULVF_FLOAT(vx, c, vl);
+            v0 = VFMACCVF_FLOAT(v0, s, vy, vl);
+            VSSEV_FLOAT(x, stride_x, v0, vl);
+
+            v1 = VFMULVF_FLOAT(vx, s, vl);
+            v1 = VFMSACVF_FLOAT(v1, c, vy, vl);
+            VSEV_FLOAT(y, v1, vl);
+        }
+ 
+    } else if(inc_x == 1) {
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+
+            v0 = VFMULVF_FLOAT(vx, c, vl);
+            v0 = VFMACCVF_FLOAT(v0, s, vy, vl);
+            VSEV_FLOAT(x, v0, vl);
+
+            v1 = VFMULVF_FLOAT(vx, s, vl);
+            v1 = VFMSACVF_FLOAT(v1, c, vy, vl);
+            VSSEV_FLOAT(y, stride_y, v1, vl);
+        }
+
+    } else {
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+            vl = VSETVL(n);
+ 
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+
+            v0 = VFMULVF_FLOAT(vx, c, vl);
+            v0 = VFMACCVF_FLOAT(v0, s, vy, vl);
+            VSSEV_FLOAT(x, stride_x, v0, vl);
+
+            v1 = VFMULVF_FLOAT(vx, s, vl);
+            v1 = VFMSACVF_FLOAT(v1, c, vy, vl);
+            VSSEV_FLOAT(y, stride_y, v1, vl);
+        }
+
+    }
+    
+    return(0);
+}

kernel/riscv64/scal_rvv.c

@@ -0,0 +1,80 @@
+/***************************************************************************
+Copyright (c) 2020, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMULVF_FLOAT vfmul_vf_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMULVF_FLOAT vfmul_vf_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+    if ( (n <= 0) || (inc_x <= 0)) return(0);
+
+    FLOAT_V_T v0;
+ 
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            v0 = VLEV_FLOAT(x, vl);
+            v0 = VFMULVF_FLOAT(v0, da, vl);
+            VSEV_FLOAT(x, v0, vl);
+        }
+
+    } else { 
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            v0 = VLSEV_FLOAT(x, stride_x, vl);
+            v0 = VFMULVF_FLOAT(v0, da, vl);
+            VSSEV_FLOAT(x, stride_x, v0, vl);
+        }
+
+    }
+
+    return 0;
+}

kernel/riscv64/sum_rvv.c

@@ -0,0 +1,95 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFADDVV_FLOAT vfadd_vv_f32m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m8_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFADDVV_FLOAT vfadd_vv_f64m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m8_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT sumf = 0.0;
+    if (n <= 0 || inc_x <= 0) return(sumf);
+
+    FLOAT_V_T vx, vsum;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vsum = VFMVVF_FLOAT(0.0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vsum = VFADDVV_FLOAT(vsum, vx, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vsum = VFADDVV_FLOAT(vsum, vx, vl);
+        }
+
+    }
+
+    v_res = VFREDSUMVS_FLOAT(v_res, vsum, v_res, vlmax);
+    sumf = VFMVFS_FLOAT_M1(v_res);
+    return(sumf);
+}

kernel/riscv64/swap_rvv.c

@@ -0,0 +1,142 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VSEV_FLOAT vse32_v_f32m8
+#define VSSEV_FLOAT vsse32_v_f32m8
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VSEV_FLOAT vse64_v_f64m8
+#define VSSEV_FLOAT vsse64_v_f64m8
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+    BLASLONG stride_x, stride_y;
+    FLOAT_V_T vx, vy;
+
+    if (n <= 0) return(0);
+
+    if (inc_x == 0 && inc_y == 0) {
+        if (n & 1) {
+            FLOAT temp = x[0];
+            x[0] = y[0];
+            y[0] = temp;
+        }
+        else {
+            return 0;
+        }
+    }
+    else if(inc_x == 0) {
+        FLOAT temp = x[0];
+        x[0] = y[(n - 1) * inc_y];
+        FLOAT* ptr = y + (n - 1) * inc_y;   // start from the last one
+        stride_y = (0 - inc_y) * sizeof(FLOAT); // reverse
+        BLASLONG m = n - 1;
+        for (size_t vl; m > 0; m -= vl, ptr -= vl*inc_y) {
+            vl = VSETVL(m);
+            vy = VLSEV_FLOAT(ptr - 1, stride_y, vl);
+            VSSEV_FLOAT(ptr, stride_y, vy, vl);
+        }
+        y[0] = temp;
+    }
+    else if(inc_y == 0) {
+        FLOAT temp = y[0];
+        y[0] = x[(n - 1) * inc_x];
+        FLOAT* ptr = x + (n - 1) * inc_x;   // start from the last one
+        stride_x = (0 - inc_x) * sizeof(FLOAT); // reverse
+        BLASLONG m = n - 1;
+        for (size_t vl; m > 0; m -= vl, ptr -= vl*inc_x) {
+            vl = VSETVL(m);
+            vx = VLSEV_FLOAT(ptr - 1, stride_x, vl);
+            VSSEV_FLOAT(ptr, stride_x, vx, vl);
+        }
+        x[0] = temp;
+    }
+    else if(inc_x == 1 && inc_y == 1) {
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLEV_FLOAT(y, vl);
+            VSEV_FLOAT(y, vx, vl);
+            VSEV_FLOAT(x, vy, vl);
+        }
+  
+    } else if (inc_y == 1) {
+        stride_x = inc_x * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLEV_FLOAT(y, vl);
+            VSEV_FLOAT(y, vx, vl);
+            VSSEV_FLOAT(x, stride_x, vy, vl);
+        }
+ 
+    } else if(inc_x == 1) {
+        stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl, y += vl*inc_y) {
+            vl = VSETVL(n);
+
+            vx = VLEV_FLOAT(x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+            VSSEV_FLOAT(y, stride_y, vx, vl);
+            VSEV_FLOAT(x, vy, vl);
+        }
+ 
+    } else {
+        stride_x = inc_x * sizeof(FLOAT);
+        stride_y = inc_y * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x, y += vl*inc_y) {
+            vl = VSETVL(n);
+
+            vx = VLSEV_FLOAT(x, stride_x, vl);
+            vy = VLSEV_FLOAT(y, stride_y, vl);
+            VSSEV_FLOAT(y, stride_y, vx, vl);
+            VSSEV_FLOAT(x, stride_x, vy, vl);
+        }
+    }
+
+    return(0);
+}

kernel/riscv64/symm_lcopy_rvv_v1.c

@@ -0,0 +1,101 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define INT_V_T     vint32m2_t
+#define VID_V_INT   vid_v_i32m2
+#define VADD_VX_INT vadd_vx_i32m2
+#define VMSGT_VX_INT vmsgt_vx_i32m2_b16
+#define VBOOL_T     vbool16_t
+#define VMERGE_VVM_FLOAT  vmerge_vvm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define INT_V_T     vint64m2_t
+#define VID_V_INT   vid_v_i64m2
+#define VADD_VX_INT vadd_vx_i64m2
+#define VMSGT_VX_INT vmsgt_vx_i64m2_b32
+#define VBOOL_T     vbool32_t
+#define VMERGE_VVM_FLOAT  vmerge_vvm_f64m2
+#endif
+
+// Optimizes the implementation in ../generic/symm_lcopy_4.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b)
+{
+    BLASLONG i, js, offset;
+
+    FLOAT *ao1, *ao2;
+
+    BLASLONG stride_lda = sizeof(FLOAT)*lda;
+
+    FLOAT_V_T vb, va1, va2;
+    VBOOL_T vbool;
+    INT_V_T vindex_max, vindex;
+
+    size_t vl = VSETVL_MAX;
+    vindex_max   = VID_V_INT(vl);
+
+    for (js = n; js > 0; js -= vl, posX += vl) {
+        vl = VSETVL(js);
+        offset = posX - posY;
+
+        ao1 = a + posX + posY * lda;
+        ao2 = a + posY + (posX) * lda;
+
+        for (i = m; i > 0; i--, offset--) {
+            va2 = VLSEV_FLOAT(ao2, stride_lda, vl);
+            va1 = VLEV_FLOAT(ao1, vl);
+
+            // offset > (0 - vindex)   --->   (offset + vindex) > 0
+            vindex = VADD_VX_INT(vindex_max, offset, vl);
+            vbool  = VMSGT_VX_INT(vindex, 0, vl);
+
+            vb =  VMERGE_VVM_FLOAT(vbool, va2, va1, vl);
+            VSEV_FLOAT(b, vb, vl);
+
+            b += vl;
+            ao1 += lda;
+            ao2++;
+        }
+    }
+
+    return 0;
+}
+

kernel/riscv64/symm_ucopy_rvv_v1.c

@@ -0,0 +1,100 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define INT_V_T     vint32m2_t
+#define VID_V_INT   vid_v_i32m2
+#define VADD_VX_INT vadd_vx_i32m2
+#define VMSGT_VX_INT vmsgt_vx_i32m2_b16
+#define VBOOL_T     vbool16_t
+#define VMERGE_VVM_FLOAT  vmerge_vvm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define INT_V_T     vint64m2_t
+#define VID_V_INT   vid_v_i64m2
+#define VADD_VX_INT vadd_vx_i64m2
+#define VMSGT_VX_INT vmsgt_vx_i64m2_b32
+#define VBOOL_T     vbool32_t
+#define VMERGE_VVM_FLOAT  vmerge_vvm_f64m2
+#endif
+
+// Optimizes the implementation in ../generic/symm_ucopy_4.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b)
+{
+    BLASLONG i, js, offset;
+
+    FLOAT *ao1, *ao2;
+
+    BLASLONG stride_lda = sizeof(FLOAT)*lda;
+    
+    FLOAT_V_T vb, va1, va2;
+    VBOOL_T vbool;
+    INT_V_T vindex_max, vindex;
+
+    size_t vl = VSETVL_MAX;
+    vindex_max   = VID_V_INT(vl);
+
+    for (js = n; js > 0; js -= vl, posX += vl) {
+        vl = VSETVL(js);
+        offset = posX - posY;
+
+        ao1 = a + posY + (posX + 0) * lda;
+        ao2 = a + posX + 0 + posY * lda;
+
+        for (i = m; i > 0; i--, offset--) {
+            va1 = VLSEV_FLOAT(ao1, stride_lda, vl);
+            va2 = VLEV_FLOAT(ao2, vl);
+
+            // offset > (0 - vindex)   --->   (offset + vindex) > 0
+            vindex = VADD_VX_INT(vindex_max, offset, vl);
+            vbool  = VMSGT_VX_INT(vindex, 0, vl);
+
+            vb =  VMERGE_VVM_FLOAT(vbool, va2, va1, vl);
+            VSEV_FLOAT(b, vb, vl);
+
+            b += vl;
+            ao1++;
+            ao2 += lda;
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/symv_L_rvv.c

@@ -0,0 +1,224 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL_MAX_M1       vsetvlmax_e32m1()
+#define VSETVL(n)           vsetvl_e32m8(n)
+#define VSETVL_MAX          vsetvlmax_e32m8()
+#define FLOAT_V_T_M1        vfloat32m1_t
+#define FLOAT_V_T           vfloat32m8_t
+#define VLEV_FLOAT          vle32_v_f32m8
+#define VSEV_FLOAT          vse32_v_f32m8
+#define VLSEV_FLOAT         vlse32_v_f32m8
+#define VSSEV_FLOAT         vsse32_v_f32m8
+#define VFMACCVV_FLOAT      vfmacc_vv_f32m8
+#define VFMACCVF_FLOAT      vfmacc_vf_f32m8
+#define VFNMSACVF_FLOAT     vfnmsac_vf_f32m8
+#define VFMULVF_FLOAT       vfmul_vf_f32m8
+#define VFMVVF_FLOAT        vfmv_v_f_f32m8
+#define VFMSACVF_FLOAT      vfmsac_vf_f32m8
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f32m1
+#define VFREDSUM_FLOAT      vfredusum_vs_f32m8_f32m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f32m1_f32
+#else
+#define VSETVL_MAX_M1       vsetvlmax_e64m1()
+#define VSETVL(n)           vsetvl_e64m8(n)
+#define VSETVL_MAX          vsetvlmax_e64m8()
+#define FLOAT_V_T_M1        vfloat64m1_t
+#define FLOAT_V_T           vfloat64m8_t
+#define VLEV_FLOAT          vle64_v_f64m8
+#define VSEV_FLOAT          vse64_v_f64m8
+#define VLSEV_FLOAT         vlse64_v_f64m8
+#define VSSEV_FLOAT         vsse64_v_f64m8
+#define VFMACCVV_FLOAT      vfmacc_vv_f64m8
+#define VFMACCVF_FLOAT      vfmacc_vf_f64m8
+#define VFNMSACVF_FLOAT     vfnmsac_vf_f64m8
+#define VFMULVF_FLOAT       vfmul_vf_f64m8
+#define VFMVVF_FLOAT        vfmv_v_f_f64m8
+#define VFMSACVF_FLOAT      vfmsac_vf_f64m8
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f64m1
+#define VFREDSUM_FLOAT      vfredusum_vs_f64m8_f64m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f64m1_f64
+#endif
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+        BLASLONG i, j, k;
+        BLASLONG ix,iy;
+        BLASLONG jx,jy;
+        FLOAT temp1;
+        FLOAT *a_ptr = a;
+
+        FLOAT_V_T_M1 v_res, v_z0;
+        size_t vlmax = VSETVL_MAX_M1, vl;
+        v_res = VFMVVF_FLOAT_M1(0, vlmax);
+        v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+        vlmax = VSETVL_MAX;
+
+        FLOAT_V_T va, vx, vy, vr;
+        BLASLONG stride_x, stride_y, inc_xv, inc_yv;
+
+        if(inc_x == 1 && inc_y == 1)
+        {
+                for (j=0; j<offset; j++)
+                {
+                        temp1 = alpha * x[j];
+                        y[j] += temp1 * a_ptr[j];
+                        i = j + 1;
+                        vr = VFMVVF_FLOAT(0, vlmax);
+                        for (k = (m-i); k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                vr = VFMVVF_FLOAT(0, vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VLEV_FLOAT(&y[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSEV_FLOAT(&y[i], vy, vl);
+
+                                vx = VLEV_FLOAT(&x[i], vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+
+                        y[j] += alpha * VFMVFS_FLOAT_M1(v_res);
+                        a_ptr += lda;
+                }
+        }
+        else if(inc_x == 1)
+        {
+                jy = 0;
+                stride_y = inc_y * sizeof(FLOAT);
+                for (j=0; j<offset; j++)
+                {
+                        temp1 = alpha * x[j];
+                        y[jy] += temp1 * a_ptr[j];
+                        iy = jy + inc_y;
+                        i = j + 1;
+                        vr = VFMVVF_FLOAT(0, vlmax);
+                        for (k = (m-i); k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                inc_yv = inc_y * vl;
+                                vr = VFMVVF_FLOAT(0, vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VLSEV_FLOAT(&y[iy], stride_y, vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSSEV_FLOAT(&y[iy], stride_y, vy, vl);
+
+                                vx = VLEV_FLOAT(&x[i], vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                                iy += inc_yv;
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+
+                        y[jy] += alpha * VFMVFS_FLOAT_M1(v_res);
+                        jy    += inc_y;
+                        a_ptr += lda;
+                }
+        }
+        else if(inc_y == 1)
+        {
+                jx = 0;
+                stride_x = inc_x * sizeof(FLOAT);
+                for (j=0; j<offset; j++)
+                {
+                        temp1 = alpha * x[jx];
+                        y[j] += temp1 * a_ptr[j];
+                        ix = jx + inc_x;
+                        i = j + 1;
+                        vr = VFMVVF_FLOAT(0, vlmax);
+                        for (k = (m-i); k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                vr = VFMVVF_FLOAT(0, vl);
+                                inc_xv = inc_x * vl;
+
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VLEV_FLOAT(&y[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSEV_FLOAT(&y[i], vy, vl);
+
+                                vx = VLSEV_FLOAT(&x[ix], stride_x, vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                                ix += inc_xv;
+                        }
+
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+
+                        y[j] += alpha * VFMVFS_FLOAT_M1(v_res);
+                        jx    += inc_x;
+                        a_ptr += lda;
+                }
+        }
+        else
+        {
+                stride_x = inc_x * sizeof(FLOAT);
+                stride_y = inc_y * sizeof(FLOAT);
+                jx = 0;
+                jy = 0;
+                for (j=0; j<offset; j++)
+                {
+                        temp1 = alpha * x[jx];
+                        y[jy] += temp1 * a_ptr[j];
+                        ix = jx + inc_x;
+                        iy = jy + inc_y;
+                        i = j + 1;
+                        vr = VFMVVF_FLOAT(0, vlmax);
+                        for (k = (m-i); k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                inc_xv = inc_x * vl;
+                                inc_yv = inc_y * vl;
+                                vr = VFMVVF_FLOAT(0, vl);
+                                
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VLSEV_FLOAT(&y[iy], stride_y, vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSSEV_FLOAT(&y[iy], stride_y, vy, vl);
+
+                                vx = VLSEV_FLOAT(&x[ix], stride_x, vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                                ix += inc_xv;
+                                iy += inc_yv;
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+
+                        y[jy] += alpha * VFMVFS_FLOAT_M1(v_res);
+                        jx    += inc_x;
+                        jy    += inc_y;
+                        a_ptr += lda;
+                }
+        }
+        return(0);
+}
+

kernel/riscv64/symv_U_rvv.c

@@ -0,0 +1,221 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL_MAX_M1       vsetvlmax_e32m1()
+#define VSETVL(n)           vsetvl_e32m8(n)
+#define VSETVL_MAX          vsetvlmax_e32m8()
+#define FLOAT_V_T_M1        vfloat32m1_t
+#define FLOAT_V_T           vfloat32m8_t
+#define VLEV_FLOAT          vle32_v_f32m8
+#define VSEV_FLOAT          vse32_v_f32m8
+#define VLSEV_FLOAT         vlse32_v_f32m8
+#define VSSEV_FLOAT         vsse32_v_f32m8
+#define VFMACCVV_FLOAT      vfmacc_vv_f32m8
+#define VFMACCVF_FLOAT      vfmacc_vf_f32m8
+#define VFNMSACVF_FLOAT     vfnmsac_vf_f32m8
+#define VFMULVF_FLOAT       vfmul_vf_f32m8
+#define VFMVVF_FLOAT        vfmv_v_f_f32m8
+#define VFMSACVF_FLOAT      vfmsac_vf_f32m8
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f32m1
+#define VFREDSUM_FLOAT      vfredusum_vs_f32m8_f32m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f32m1_f32
+#else
+#define VSETVL_MAX_M1       vsetvlmax_e64m1()
+#define VSETVL(n)           vsetvl_e64m8(n)
+#define VSETVL_MAX          vsetvlmax_e64m8()
+#define FLOAT_V_T_M1        vfloat64m1_t
+#define FLOAT_V_T           vfloat64m8_t
+#define VLEV_FLOAT          vle64_v_f64m8
+#define VSEV_FLOAT          vse64_v_f64m8
+#define VLSEV_FLOAT         vlse64_v_f64m8
+#define VSSEV_FLOAT         vsse64_v_f64m8
+#define VFMACCVV_FLOAT      vfmacc_vv_f64m8
+#define VFMACCVF_FLOAT      vfmacc_vf_f64m8
+#define VFNMSACVF_FLOAT     vfnmsac_vf_f64m8
+#define VFMULVF_FLOAT       vfmul_vf_f64m8
+#define VFMVVF_FLOAT        vfmv_v_f_f64m8
+#define VFMSACVF_FLOAT      vfmsac_vf_f64m8
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f64m1
+#define VFREDSUM_FLOAT      vfredusum_vs_f64m8_f64m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f64m1_f64
+#endif
+
+int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+        BLASLONG i, j, k;
+        BLASLONG ix,iy;
+        BLASLONG jx,jy;
+        FLOAT temp1;
+        FLOAT *a_ptr = a;
+        FLOAT_V_T_M1 v_res, v_z0;
+        size_t vl_max = VSETVL_MAX_M1, vl;
+        v_res = VFMVVF_FLOAT_M1(0, vl_max);
+        v_z0 = VFMVVF_FLOAT_M1(0, vl_max);
+        vl_max = VSETVL_MAX;
+
+        FLOAT_V_T va, vx, vy, vr;
+        BLASLONG stride_x, stride_y, inc_xv, inc_yv;
+
+        BLASLONG m1 = m - offset;
+        if(inc_x == 1 && inc_y == 1)
+        {
+                a_ptr += m1 * lda;
+                for (j=m1; j<m; j++)
+                {
+                        temp1 = alpha * x[j];
+                        i = 0;
+                        vr = VFMVVF_FLOAT(0, vl_max);
+                        for (k = j; k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                vr = VFMVVF_FLOAT(0, vl);
+                                vy = VLEV_FLOAT(&y[i], vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSEV_FLOAT(&y[i], vy, vl);
+
+                                vx = VLEV_FLOAT(&x[i], vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vl_max);
+
+                        y[j] += temp1 * a_ptr[j] + alpha * VFMVFS_FLOAT_M1(v_res);
+                        a_ptr += lda;
+                }
+        }
+        else if(inc_x == 1)
+        {
+                jy = m1 * inc_y;
+                a_ptr += m1 * lda;
+                stride_y = inc_y * sizeof(FLOAT);
+                for (j=m1; j<m; j++)
+                {
+                        temp1 = alpha * x[j];
+                        iy = 0;
+                        i = 0;
+                        vr = VFMVVF_FLOAT(0, vl_max);
+                        for (k = j; k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                inc_yv = inc_y * vl;
+                                vr = VFMVVF_FLOAT(0, vl);
+                                vy = VLSEV_FLOAT(&y[iy], stride_y, vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSSEV_FLOAT(&y[iy], stride_y, vy, vl);
+
+                                vx = VLEV_FLOAT(&x[i], vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                                iy += inc_yv;
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vl_max);
+
+                        y[jy] += temp1 * a_ptr[j] + alpha * VFMVFS_FLOAT_M1(v_res);
+                        a_ptr += lda;
+                        jy    += inc_y;
+                }
+        }
+        else if(inc_y == 1)
+        {
+                jx = m1 * inc_x;
+                a_ptr += m1 * lda;
+                stride_x = inc_x * sizeof(FLOAT);
+                for (j=m1; j<m; j++)
+                {
+                        temp1 = alpha * x[jx];
+                        ix = 0;
+                        i = 0;
+                        vr = VFMVVF_FLOAT(0, vl_max);
+                        for (k = j; k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                inc_xv = inc_x * vl;
+                                vr = VFMVVF_FLOAT(0, vl);
+
+                                vy = VLEV_FLOAT(&y[i], vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSEV_FLOAT(&y[i], vy, vl);
+
+                                vx = VLSEV_FLOAT(&x[ix], stride_x, vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+
+                                ix += inc_xv;
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vl_max);
+
+                        y[j] += temp1 * a_ptr[j] + alpha * VFMVFS_FLOAT_M1(v_res);
+                        a_ptr += lda;
+                        jx    += inc_x;
+                }
+        }
+        else
+        {
+                jx = m1 * inc_x;
+                jy = m1 * inc_y;
+                a_ptr += m1 * lda;
+                stride_x = inc_x * sizeof(FLOAT);
+                stride_y = inc_y * sizeof(FLOAT);
+                for (j=m1; j<m; j++)
+                {
+                        temp1 = alpha * x[jx];
+
+                        ix = 0;
+                        iy = 0;
+                        i = 0;
+                        vr = VFMVVF_FLOAT(0, vl_max);
+                        for (k = j; k > 0; k -= vl, i += vl)
+                        {
+                                vl = VSETVL(k);
+                                inc_xv = inc_x * vl;
+                                inc_yv = inc_y * vl;
+                                vr = VFMVVF_FLOAT(0, vl);
+                                vy = VLSEV_FLOAT(&y[iy], stride_y, vl);
+                                va = VLEV_FLOAT(&a_ptr[i], vl);
+                                vy = VFMACCVF_FLOAT(vy, temp1, va, vl);
+                                VSSEV_FLOAT(&y[iy], stride_y, vy, vl);
+
+                                vx = VLSEV_FLOAT(&x[ix], stride_x, vl);
+                                vr = VFMACCVV_FLOAT(vr, vx, va, vl);
+                                ix += inc_xv;
+                                iy += inc_yv;
+                        }
+                        v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vl_max);
+
+                        y[jy] += temp1 * a_ptr[j] + alpha * VFMVFS_FLOAT_M1(v_res);
+                        a_ptr += lda;
+                        jx    += inc_x;
+                        jy    += inc_y;
+                }
+        }
+        return(0);
+}

kernel/riscv64/trmm_lncopy_rvv_v1.c

@@ -0,0 +1,138 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vint32m2_t
+#define VID_V_UINT vid_v_i32m2
+#define VMSGTU_VX_UINT vmsgt_vx_i32m2_b16
+#define VMSEQ_VX_UINT vmseq_vx_i32m2_b16
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u64m2_b32
+#define VMSEQ_VX_UINT vmseq_vx_u64m2_b32
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f64m2
+#endif
+
+// Optimizes the implementation in ../arm64/tmmm_lncopy_sve_v1.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
+
+    BLASLONG i, js, X;
+
+    FLOAT *ao;
+
+    BLASLONG stride_lda = sizeof(FLOAT)*lda;
+    
+    FLOAT_V_T vb, va1;
+
+    size_t vl;
+#ifdef UNIT
+    VBOOL_T vbool_eq;
+#endif
+
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        X = posX;
+
+        if (posX <= posY) 
+        {
+            ao = a + posY + posX * lda;
+        } 
+        else 
+        {
+            ao = a + posX + posY * lda;
+        }
+
+        i = 0;
+        do 
+        {
+            if (X > posY) 
+            {
+                va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                VSEV_FLOAT(b, va1, vl);
+
+                ao ++;
+                b += vl;
+                X ++;
+                i ++;
+            } 
+            else if (X < posY) 
+            {
+                ao += lda;
+                b += vl;
+                X ++;
+                i ++;
+            } 
+            else 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    vbool_cmp = VMSGTU_VX_UINT(vindex, j, vl);
+                    vb = VFMERGE_VFM_FLOAT(vbool_cmp, va1, ZERO, vl);
+#ifdef UNIT
+                    vbool_eq = VMSEQ_VX_UINT(vindex, j, vl);
+                    vb =  VFMERGE_VFM_FLOAT(vbool_eq, vb, ONE, vl);
+#endif
+                    VSEV_FLOAT(b, vb, vl);
+                    ao++;
+                    b += vl;
+                }
+
+                X += vl;
+                i += vl;
+            }
+        } while (i < m);
+
+        posY += vl;
+    }
+
+    return 0;
+}

kernel/riscv64/trmm_ltcopy_rvv_v1.c

@@ -0,0 +1,134 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSLTU_VX_UINT vmsltu_vx_u32m2_b16
+#define VMSEQ_VX_UINT vmseq_vx_u32m2_b16
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSLTU_VX_UINT vmsltu_vx_u64m2_b32
+#define VMSEQ_VX_UINT vmseq_vx_u64m2_b32
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f64m2
+#endif
+
+// Optimizes the implementation in ../arm64/tmmm_ltcopy_sve_v1.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
+
+    BLASLONG i, js, X;
+
+    FLOAT *ao;
+    
+    FLOAT_V_T vb, va1;
+    size_t vl;
+#ifdef UNIT
+    VBOOL_T vbool_eq;
+#endif
+
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        X = posX;
+
+        if (posX <= posY) 
+        {
+            ao = a + posY + posX * lda;
+        } 
+        else 
+        {
+            ao = a + posX + posY * lda;
+        }
+
+        i = 0;
+        do 
+        {
+            if (X > posY) 
+            {
+                ao ++;
+                b += vl;
+                X ++;
+                i ++;
+            } 
+            else if (X < posY) 
+            {
+                va1 = VLEV_FLOAT(ao, vl);
+                VSEV_FLOAT(b, va1, vl);
+
+                ao += lda;
+                b += vl;
+                X ++;
+                i ++;
+            }
+            else
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    va1 = VLEV_FLOAT(ao, vl);
+                    vbool_cmp = VMSLTU_VX_UINT(vindex, j, vl);
+                    vb = VFMERGE_VFM_FLOAT(vbool_cmp, va1, ZERO, vl);
+#ifdef UNIT
+                    vbool_eq = VMSEQ_VX_UINT(vindex, j, vl);
+                    vb =  VFMERGE_VFM_FLOAT(vbool_eq, vb, ONE, vl);
+#endif
+                    VSEV_FLOAT(b, vb, vl);
+                    ao += lda;
+                    b += vl;
+                }
+                X += vl;
+                i += vl;
+
+            }
+        } while (i < m);
+
+        posY += vl;
+    }
+
+    return 0;
+}
+

kernel/riscv64/trmm_uncopy_rvv_v1.c

@@ -0,0 +1,136 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSLTU_VX_UINT vmsltu_vx_u32m2_b16
+#define VMSEQ_VX_UINT vmseq_vx_u32m2_b16
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T    vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSLTU_VX_UINT vmsltu_vx_u64m2_b32
+#define VMSEQ_VX_UINT vmseq_vx_u64m2_b32
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f64m2
+#endif
+
+// Optimizes the implementation in ../arm64/tmmm_uncopy_sve_v1.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
+
+    BLASLONG i, js, X;
+    BLASLONG stride_lda = sizeof(FLOAT) * lda;
+    FLOAT *ao;
+
+    FLOAT_V_T vb, va1;
+    size_t vl;
+
+#ifdef UNIT
+    VBOOL_T vbool_eq;
+#endif
+
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        X = posX;
+
+        if (posX <= posY) 
+        {
+            ao = a + posX + posY * lda;
+        } 
+        else 
+        {
+            ao = a + posY + posX * lda;
+        }
+
+        i = 0;
+        do
+        {
+            if (X < posY) 
+            {
+                va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                VSEV_FLOAT(b, va1, vl);
+
+                ao ++;
+                b += vl;
+                X ++;
+                i ++;
+            } 
+            else if (X > posY) 
+            {
+                ao += lda;
+                b += vl;
+                X ++;
+                i ++;
+            } 
+            else 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    vbool_cmp = VMSLTU_VX_UINT(vindex, j, vl);
+                    vb = VFMERGE_VFM_FLOAT(vbool_cmp, va1, ZERO, vl);
+#ifdef UNIT
+                    vbool_eq = VMSEQ_VX_UINT(vindex, j, vl);
+                    vb =  VFMERGE_VFM_FLOAT(vbool_eq, vb, ONE, vl);
+#endif
+                    VSEV_FLOAT(b, vb, vl);
+                    ao++;
+                    b += vl;
+                }
+
+                X += vl;
+                i += vl;
+            }
+        }while (i < m);
+
+        posY += vl;
+    }
+
+    return 0;
+}

kernel/riscv64/trmm_utcopy_rvv_v1.c

@@ -0,0 +1,133 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u32m2_b16
+#define VMSEQ_VX_UINT vmseq_vx_u32m2_b16
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u64m2_b32
+#define VMSEQ_VX_UINT vmseq_vx_u64m2_b32
+#define VFMERGE_VFM_FLOAT  vfmerge_vfm_f64m2
+#endif
+
+// Optimizes the implementation in ../arm64/tmmm_utcopy_sve_v1.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLONG posY, FLOAT *b){
+
+    BLASLONG i, j, js, X;
+
+    FLOAT *ao;
+    FLOAT_V_T vb, va1;
+#ifdef UNIT
+    VBOOL_T vbool_eq;
+#endif
+
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    size_t vl;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+
+        X = posX;
+
+        if (posX <= posY) 
+        {
+            ao = a + posX + posY * lda;
+        } 
+        else 
+        {
+            ao = a + posY + posX * lda;
+        }
+
+        i = 0;
+        do
+        {
+            if (X < posY) 
+            {
+                ao ++;
+                b += vl;
+                X ++;
+                i++;
+            }
+            else if (X > posY)
+            {
+                va1 = VLEV_FLOAT(ao, vl);
+                VSEV_FLOAT(b, va1, vl);
+                ao += lda;
+                b += vl;
+                X++;
+                i++;
+            }
+            else
+            {
+                vindex  = VID_V_UINT(vl);
+                for (j = 0; j < vl; j++) 
+                {
+                    va1 = VLEV_FLOAT(ao, vl);
+                    vbool_cmp = VMSGTU_VX_UINT(vindex, j, vl);
+                    vb = VFMERGE_VFM_FLOAT(vbool_cmp, va1, ZERO, vl);
+#ifdef UNIT
+                    vbool_eq = VMSEQ_VX_UINT(vindex, j, vl);
+                    vb =  VFMERGE_VFM_FLOAT(vbool_eq, vb, ONE, vl);
+#endif
+                    VSEV_FLOAT(b, vb, vl);
+                    ao += lda;
+                    b += vl;
+                }
+                X += vl;
+                i += vl;
+            }
+        }while (i < m);
+        posY += vl;
+    }
+    return 0;
+}
+

kernel/riscv64/trmmkernel_2x2_rvv.c

@@ -0,0 +1,342 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFMACCVV_FLOAT vfmacc_vv_f32m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFMACCVV_FLOAT vfmacc_vv_f64m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+
+// Optimizes the implementation in ../generic/trmmkernel_2x2.c
+
+
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc
+#ifdef TRMMKERNEL
+        ,BLASLONG offset
+#endif
+        )
+{
+   BLASLONG i,j,k;
+   FLOAT *C0,*C1,*ptrba,*ptrbb;
+   BLASLONG off, temp;
+
+   FLOAT_V_T va0, va1, vb0, vb1;
+   FLOAT_V_T vres0, vres1, vres2, vres3;
+   FLOAT_V_T_M1 v_res, v_z0;
+   v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+   size_t vl;
+   size_t vlmax = VSETVL_MAX;
+   
+#if defined(TRMMKERNEL) && !defined(LEFT)
+   off = -offset;
+#else
+   off = 0;
+#endif
+
+   for (j = bn/2; j > 0; j--)
+   {
+        C0 = C;
+        C1 = C0+ldc;
+#if defined(TRMMKERNEL) && defined(LEFT)
+        off = offset;
+#endif
+        ptrba = ba;
+
+        for (i = bm/2; i > 0; i--)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || \
+              (!defined(LEFT) && !defined(TRANSA))
+             ptrbb = bb;
+#else
+              ptrba += off*2;
+              ptrbb = bb + off*2;
+#endif
+
+#if (defined(LEFT) && !defined(TRANSA)) || \
+             (!defined(LEFT) && defined(TRANSA))
+             temp = bk-off;
+#elif defined(LEFT)
+             temp = off+2;
+#else
+             temp = off+2;
+#endif
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb1, vl);
+
+                ptrba += vl * 2;
+                ptrbb += vl * 2;
+            }
+            v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUMVS_FLOAT(v_res, vres2, v_z0, vlmax);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUMVS_FLOAT(v_res, vres3, v_z0, vlmax);
+            C1[1] = alpha * VFMVFS_FLOAT_M1(v_res);
+
+#if ( defined(LEFT) && defined(TRANSA)) || \
+             (!defined(LEFT) && !defined(TRANSA))
+             temp = bk - off;
+#ifdef LEFT
+             temp -= 2;
+#else
+             temp -= 2;
+#endif
+             ptrba += temp*2;
+             ptrbb += temp*2;
+#endif
+#ifdef LEFT
+             off += 2;
+#endif
+             C0 = C0+2;
+             C1 = C1+2;
+        }
+
+        if (bm & 1)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) ||(!defined(LEFT) && !defined(TRANSA))
+             ptrbb = bb;
+#else
+             ptrba += off;
+             ptrbb = bb+off*2;
+#endif
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+             temp = bk-off;
+#elif defined(LEFT)
+             temp = off+1;
+#else
+             temp = off+2;
+#endif
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG_FLOAT(&vb0, &vb1, ptrbb, vl);
+                  
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                ptrba += vl;
+                ptrbb += vl * 2;
+
+            }
+            v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+             temp = bk-off;
+#ifdef LEFT
+             temp -= 1;
+#else
+             temp -= 2;
+#endif
+             ptrba += temp;
+             ptrbb += temp*2;
+#endif
+#ifdef LEFT
+             off += 1;
+#endif
+             C0 = C0+1;
+             C1 = C1+1;
+        }
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 2;
+#endif
+        k = (bk<<1);
+        bb = bb+k;
+        i = (ldc<<1);
+        C = C+i;
+   }
+
+   if (bn & 1)
+   {
+        C0 = C;
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+        off = offset;
+#endif
+        ptrba = ba;
+
+        for (i = bm/2; i > 0; i--)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || \
+              (!defined(LEFT) && !defined(TRANSA))
+              ptrbb = bb;
+#else
+              ptrba += off*2;
+              ptrbb = bb + off;
+#endif
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || \
+             (!defined(LEFT) && defined(TRANSA))
+             temp = bk-off;
+#elif defined(LEFT)
+             temp = off+2;
+#else
+             temp = off+1;
+#endif
+             vres0 = VFMVVF_FLOAT(0.0, vlmax);
+             vres1 = VFMVVF_FLOAT(0.0, vlmax);
+             
+             for (k = temp; k > 0; k -= vl)
+             {
+                 vl = VSETVL(k);
+                 vb0 = VLEV_FLOAT(ptrbb, vl);
+                 VLSEG_FLOAT(&va0, &va1, ptrba, vl);
+                   
+                 vres0 = VFMACCVV_FLOAT(vres0, vb0, va0, vl);
+                 vres1 = VFMACCVV_FLOAT(vres1, vb0, va1, vl);
+             
+                 ptrba += vl * 2;
+                 ptrbb += vl;
+             
+             }
+             v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
+             C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+             v_res = VFREDSUMVS_FLOAT(v_res, vres1, v_z0, vlmax);
+             C0[1] = alpha * VFMVFS_FLOAT_M1(v_res);
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || \
+             (!defined(LEFT) && !defined(TRANSA))
+             temp = bk - off;
+#ifdef LEFT
+             temp -= 2;
+#else
+             temp -= 1;
+#endif
+             ptrba += temp*2;
+             ptrbb += temp;
+#endif
+#ifdef LEFT
+             off += 2;
+#endif
+
+             C0 = C0+2;
+        }
+
+        if (bm & 1)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+             ptrbb = bb;
+#else
+             ptrba += off;
+             ptrbb = bb+off;
+#endif
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+             temp = bk-off;
+#elif defined(LEFT)
+             temp = off + 1;
+#else
+             temp = off + 1;
+#endif
+             vres0 = VFMVVF_FLOAT(0.0, vlmax);
+             
+             for (k = temp; k > 0; k -= vl)
+             {
+                 vl = VSETVL(k);
+                 va0 = VLEV_FLOAT(ptrba, vl);
+                 vb0 = VLEV_FLOAT(ptrbb, vl);
+                   
+                 vres0 = VFMACCVV_FLOAT(vres0, vb0, va0, vl);
+                 ptrba += vl;
+                 ptrbb += vl;
+             }
+             v_res = VFREDSUMVS_FLOAT(v_res, vres0, v_z0, vlmax);
+             C0[0] = alpha * VFMVFS_FLOAT_M1(v_res);
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+             temp = bk-off;
+#ifdef LEFT
+             temp -= 1;
+#else
+             temp -= 1;
+#endif
+             ptrba += temp;
+             ptrbb += temp;
+#endif
+#ifdef LEFT
+             off += 1;
+#endif
+             C0 = C0+1;
+        }
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 1;
+#endif
+        k = (bk<<0);
+        bb = bb+k;
+        C = C+ldc;
+   }
+   return 0;
+}
+

kernel/riscv64/trmmkernel_4x4_rvv.c

@@ -0,0 +1,881 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <stdbool.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n)           vsetvl_e32m2(n)
+#define VSETVL_MAX          vsetvlmax_e32m2()
+#define VSETVL_MAX_M1       vsetvlmax_e32m1()
+#define FLOAT_V_T           vfloat32m2_t
+#define FLOAT_V_T_M1        vfloat32m1_t
+#define VLEV_FLOAT          vle32_v_f32m2
+#define VLSEG4_FLOAT        vlseg4e32_v_f32m2
+#define VLSEG2_FLOAT        vlseg2e32_v_f32m2
+#define VFMVVF_FLOAT        vfmv_v_f_f32m2
+#define VFMUL_FLOAT         vfmul_vv_f32m2
+#define VFMACCVF_FLOAT      vfmacc_vf_f32m2
+#define VFMACCVV_FLOAT      vfmacc_vv_f32m2
+#define VFREDSUMVS_FLOAT    vfredusum_vs_f32m2_f32m1
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n)           vsetvl_e64m2(n)
+#define VSETVL_MAX          vsetvlmax_e64m2()
+#define VSETVL_MAX_M1       vsetvlmax_e64m1()
+#define FLOAT_V_T           vfloat64m2_t
+#define FLOAT_V_T_M1        vfloat64m1_t
+#define VLEV_FLOAT          vle64_v_f64m2
+#define VLSEG4_FLOAT        vlseg4e64_v_f64m2
+#define VLSEG2_FLOAT        vlseg2e64_v_f64m2
+#define VFMVVF_FLOAT        vfmv_v_f_f64m2
+#define VFMUL_FLOAT         vfmul_vv_f64m2
+#define VFMACCVF_FLOAT      vfmacc_vf_f64m2
+#define VFMACCVV_FLOAT      vfmacc_vv_f64m2
+#define VFREDSUMVS_FLOAT    vfredusum_vs_f64m2_f64m1
+#define VFMVVF_FLOAT_M1     vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1     vfmv_f_s_f64m1_f64
+#endif
+
+
+// Optimizes the implementation in ../generic/trmmkernel_4x4.c
+
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)
+{
+
+   BLASLONG i,j,k;
+   FLOAT *C0,*C1,*C2,*C3,*ptrba,*ptrbb;
+
+   FLOAT_V_T va0, va1, va2, va3, vb0, vb1, vb2, vb3;
+   FLOAT_V_T_M1 vsum0, vsum1, vsum2, vsum3, v_z0;
+   v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+   size_t vl;
+   size_t vlmax = VSETVL_MAX;
+
+   FLOAT_V_T vres0_0;
+   FLOAT_V_T vres0_1;
+   FLOAT_V_T vres0_2;
+   FLOAT_V_T vres0_3;
+
+   FLOAT_V_T vres1_0;
+   FLOAT_V_T vres1_1;
+   FLOAT_V_T vres1_2;
+   FLOAT_V_T vres1_3;
+
+   FLOAT_V_T vres2_0;
+   FLOAT_V_T vres2_1;
+   FLOAT_V_T vres2_2;
+   FLOAT_V_T vres2_3;
+
+   FLOAT_V_T vres3_0;
+   FLOAT_V_T vres3_1;
+   FLOAT_V_T vres3_2;
+   FLOAT_V_T vres3_3;
+
+   BLASLONG off, temp;
+
+   bool left;
+   bool transposed;
+   bool backwards;
+
+#ifdef LEFT
+   left = true;
+#else
+   left = false;
+#endif
+
+#ifdef TRANSA
+   transposed = true;
+#else
+   transposed = false;
+#endif
+
+   backwards = left != transposed;
+
+   if (!left) {
+      off = -offset;
+   }
+
+
+   for (j=0; j<bn/4; j+=1) // do blocks of the Mx4 loops 
+   {
+        C0 = C;
+        C1 = C0+ldc;
+        C2 = C1+ldc;
+        C3 = C2+ldc;
+
+
+        if (left) {
+            off = offset;
+        }
+
+        ptrba = ba;
+
+        for (i=0; i<bm/4; i+=1) // do blocks of 4x4
+        {
+
+            ptrbb = bb;
+            if (backwards)
+            {
+                ptrba += off*4; // number of values in A
+                ptrbb += off*4; // number of values in B
+            }
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+            vres0_2 = VFMVVF_FLOAT(0, vlmax);
+            vres0_3 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_1 = VFMVVF_FLOAT(0, vlmax);
+            vres1_2 = VFMVVF_FLOAT(0, vlmax);
+            vres1_3 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres2_0 = VFMVVF_FLOAT(0, vlmax);
+            vres2_1 = VFMVVF_FLOAT(0, vlmax);
+            vres2_2 = VFMVVF_FLOAT(0, vlmax);
+            vres2_3 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres3_0 = VFMVVF_FLOAT(0, vlmax);
+            vres3_1 = VFMVVF_FLOAT(0, vlmax);
+            vres3_2 = VFMVVF_FLOAT(0, vlmax);
+            vres3_3 = VFMVVF_FLOAT(0, vlmax);
+
+            temp = backwards ? bk-off :
+                         left ? off + 4 : // number of values in A
+                                off + 4;  // number of values in B
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+                vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
+                vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);
+            
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+                vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
+                vres2_1 = VFMACCVV_FLOAT(vres2_1, va1, vb2, vl);
+                vres3_1 = VFMACCVV_FLOAT(vres3_1, va1, vb3, vl);
+
+                vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
+                vres1_2 = VFMACCVV_FLOAT(vres1_2, va2, vb1, vl);
+                vres2_2 = VFMACCVV_FLOAT(vres2_2, va2, vb2, vl);
+                vres3_2 = VFMACCVV_FLOAT(vres3_2, va2, vb3, vl);
+
+                vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
+                vres1_3 = VFMACCVV_FLOAT(vres1_3, va3, vb1, vl);
+                vres2_3 = VFMACCVV_FLOAT(vres2_3, va3, vb2, vl);
+                vres3_3 = VFMACCVV_FLOAT(vres3_3, va3, vb3, vl);
+
+                ptrba += vl * 4;
+                ptrbb += vl * 4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres1_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_3, v_z0, vlmax);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres2_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres2_3, v_z0, vlmax);
+            C2[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C2[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C2[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C2[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres3_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres3_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres3_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_3, v_z0, vlmax);
+            C3[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C3[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C3[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C3[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            if (!backwards) {
+                temp = bk-off;
+                temp = left ? temp - 4 : // number of values in A
+                              temp - 4;  // number of values in B
+
+                ptrba += temp*4; // number of values in A
+                ptrbb += temp*4; // number of values in B
+            }
+#ifdef LEFT
+            off += 4; // number of values in A
+#endif
+
+            C0 = C0+4;
+            C1 = C1+4;
+            C2 = C2+4;
+            C3 = C3+4;
+
+        }
+
+        if ( bm & 2 ) // do any 2x4 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2;
+            ptrbb = bb + off*4;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_1 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres2_0 = VFMVVF_FLOAT(0, vlmax);
+            vres2_1 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres3_0 = VFMVVF_FLOAT(0, vlmax);
+            vres3_1 = VFMVVF_FLOAT(0, vlmax);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+2;   // number of values in A
+#else
+            temp = off+4;   // number of values in B
+#endif
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+                vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
+                vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);
+
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+                vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
+                vres2_1 = VFMACCVV_FLOAT(vres2_1, va1, vb2, vl);
+                vres3_1 = VFMACCVV_FLOAT(vres3_1, va1, vb3, vl);
+
+                ptrba += vl * 2;
+                ptrbb += vl * 4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_0, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_1, v_z0, vlmax);
+
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[1] = alpha * VFMVFS_FLOAT_M1(vsum3);
+            
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres2_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres2_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres3_0, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_1, v_z0, vlmax);
+
+            C2[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C2[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C3[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C3[1] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 2; // number of values in A
+#else
+            temp -= 4; // number of values in B
+#endif
+            ptrba += temp*2;
+            ptrbb += temp*4;
+#endif
+
+#ifdef LEFT
+            off += 2; // number of values in A
+#endif
+
+            C0 = C0+2;
+            C1 = C1+2;
+            C2 = C2+2;
+            C3 = C3+2;
+
+        }
+
+        if ( bm & 1 ) // do any 1x4 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*1;
+            ptrbb = bb + off*4;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+            vres2_0 = VFMVVF_FLOAT(0, vlmax);
+            vres3_0 = VFMVVF_FLOAT(0, vlmax);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+1;   // number of values in A
+#else
+            temp = off+4;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+                vres2_0 = VFMACCVV_FLOAT(vres2_0, va0, vb2, vl);
+                vres3_0 = VFMACCVV_FLOAT(vres3_0, va0, vb3, vl);
+
+                ptrba += vl;
+                ptrbb += vl * 4;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_0, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres2_0, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres3_0, v_z0, vlmax);
+
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C2[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C3[0] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 1; // number of values in A
+#else
+            temp -= 4; // number of values in B
+#endif
+            ptrba += temp*1;
+            ptrbb += temp*4;
+#endif
+
+#ifdef LEFT
+            off += 1; // number of values in A
+#endif
+
+            C0 = C0+1;
+            C1 = C1+1;
+            C2 = C2+1;
+            C3 = C3+1;
+
+        }
+
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 4;
+#endif
+
+        k = (bk<<2);
+        bb = bb+k;
+        i = (ldc<<2);
+        C = C+i;
+    }
+
+    for (j=0; j<(bn&2); j+=2) // do the Mx2 loops 
+    {
+        C0 = C;
+        C1 = C0+ldc;
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+        off = offset;
+#endif
+
+        ptrba = ba;
+
+        for (i=0; i<bm/4; i+=1) // do blocks of 4x2
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*4;
+            ptrbb = bb + off*2;
+#endif
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+            vres0_2 = VFMVVF_FLOAT(0, vlmax);
+            vres0_3 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_1 = VFMVVF_FLOAT(0, vlmax);
+            vres1_2 = VFMVVF_FLOAT(0, vlmax);
+            vres1_3 = VFMVVF_FLOAT(0, vlmax);
+        
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+4;   // number of values in A
+#else
+            temp = off+2;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+            
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+                vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
+            
+                vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
+                vres1_2 = VFMACCVV_FLOAT(vres1_2, va2, vb1, vl);
+            
+                vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
+                vres1_3 = VFMACCVV_FLOAT(vres1_3, va3, vb1, vl);
+            
+                ptrba += vl * 4;
+                ptrbb += vl * 2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres1_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_3, v_z0, vlmax);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 4; // number of values in A
+#else
+            temp -= 2; // number of values in B
+#endif
+            ptrba += temp*4;
+            ptrbb += temp*2;
+#endif
+
+#ifdef LEFT
+            off += 4; // number of values in A
+#endif
+
+            C0 = C0+4;
+            C1 = C1+4;
+
+        }
+
+        if ( bm & 2 ) // do any 2x2 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2;
+            ptrbb = bb + off*2;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+            
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_1 = VFMVVF_FLOAT(0, vlmax);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+2;   // number of values in A
+#else
+            temp = off+2;   // number of values in B
+#endif
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+            
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+                vres1_1 = VFMACCVV_FLOAT(vres1_1, va1, vb1, vl);
+            
+                ptrba += vl * 2;
+                ptrbb += vl * 2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres1_0, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres1_1, v_z0, vlmax);
+
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C1[1] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 2; // number of values in A
+#else
+            temp -= 2; // number of values in B
+#endif
+            ptrba += temp*2;
+            ptrbb += temp*2;
+#endif
+
+#ifdef LEFT
+            off += 2; // number of values in A
+#endif
+
+            C0 = C0+2;
+            C1 = C1+2;
+
+        }
+
+        if ( bm & 1 ) // do any 1x2 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*1;
+            ptrbb = bb + off*2;
+#endif
+
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres1_0 = VFMVVF_FLOAT(0, vlmax);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+1;   // number of values in A
+#else
+            temp = off+2;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                va0 = VLEV_FLOAT(ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                vres1_0 = VFMACCVV_FLOAT(vres1_0, va0, vb1, vl);
+                    
+                ptrba += vl;
+                ptrbb += vl * 2;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres1_0, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C1[0] = alpha * VFMVFS_FLOAT_M1(vsum1);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 1; // number of values in A
+#else
+            temp -= 2; // number of values in B
+#endif
+            ptrba += temp*1;
+            ptrbb += temp*2;
+#endif
+
+#ifdef LEFT
+            off += 1; // number of values in A
+#endif
+
+            C0 = C0+1;
+            C1 = C1+1;
+
+        }
+
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 2;
+#endif
+
+        k = (bk<<1);
+        bb = bb+k;
+        i = (ldc<<1);
+        C = C+i;
+    }
+
+    for (j=0; j<(bn&1); j+=1) // do the Mx1 loops
+    {
+        C0 = C;
+
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+    off = offset;
+#endif
+
+        ptrba = ba;
+
+        for (i=0; i<bm/4; i+=1) // do blocks of 4x1 loops
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*4;
+            ptrbb = bb + off*1;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+            vres0_2 = VFMVVF_FLOAT(0, vlmax);
+            vres0_3 = VFMVVF_FLOAT(0, vlmax);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+4;   // number of values in A
+#else
+            temp = off+1;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+           
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+            
+                vres0_2 = VFMACCVV_FLOAT(vres0_2, va2, vb0, vl);
+            
+                vres0_3 = VFMACCVV_FLOAT(vres0_3, va3, vb0, vl);
+            
+                ptrba += vl * 4;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            vsum2 = VFREDSUMVS_FLOAT(vsum2, vres0_2, v_z0, vlmax);
+            vsum3 = VFREDSUMVS_FLOAT(vsum3, vres0_3, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+            C0[2] = alpha * VFMVFS_FLOAT_M1(vsum2);
+            C0[3] = alpha * VFMVFS_FLOAT_M1(vsum3);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 4; // number of values in A
+#else
+            temp -= 1; // number of values in B
+#endif
+            ptrba += temp*4;
+            ptrbb += temp*1;
+#endif
+
+#ifdef LEFT
+            off += 4; // number of values in A
+#endif
+
+            C0 = C0+4;
+
+        }
+
+        if ( bm & 2 ) // do any 2x1 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2;
+            ptrbb = bb + off*1;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+            vres0_1 = VFMVVF_FLOAT(0, vlmax);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+2;   // number of values in A
+#else
+            temp = off+1;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+           
+                vres0_1 = VFMACCVV_FLOAT(vres0_1, va1, vb0, vl);
+           
+                ptrba += vl * 2;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            vsum1 = VFREDSUMVS_FLOAT(vsum1, vres0_1, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+            C0[1] = alpha * VFMVFS_FLOAT_M1(vsum1);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 2; // number of values in A
+#else
+            temp -= 1; // number of values in B
+#endif
+            ptrba += temp*2;
+            ptrbb += temp*1;
+#endif
+
+#ifdef LEFT
+            off += 2; // number of values in A
+#endif
+
+            C0 = C0+2;
+
+        }
+
+        if ( bm & 1 ) // do any 1x1 loop
+        {
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*1;
+            ptrbb = bb + off*1;
+#endif
+
+            vres0_0 = VFMVVF_FLOAT(0, vlmax);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+1;   // number of values in A
+#else
+            temp = off+1;   // number of values in B
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                va0 = VLEV_FLOAT(ptrba, vl);
+                vb0 = VLEV_FLOAT(ptrbb, vl);
+            
+                vres0_0 = VFMACCVV_FLOAT(vres0_0, va0, vb0, vl);
+                  
+                ptrba += vl;
+                ptrbb += vl;
+            }
+
+            vsum0 = VFREDSUMVS_FLOAT(vsum0, vres0_0, v_z0, vlmax);
+            C0[0] = alpha * VFMVFS_FLOAT_M1(vsum0);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 1; // number of values in A
+#else
+            temp -= 1; // number of values in B
+#endif
+            ptrba += temp*1;
+            ptrbb += temp*1;
+#endif
+
+#ifdef LEFT
+            off += 1; // number of values in A
+#endif
+
+            C0 = C0+1;
+
+        }
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 1;
+#endif
+
+        k = (bk<<0);
+        bb = bb+k;
+        C = C+ldc;
+    }
+    return 0;
+}

kernel/riscv64/trmmkernel_rvv_v1x8.c

@@ -0,0 +1,685 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VFMVVF_FLOAT vfmv_v_f_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFMULVF_FLOAT vfmul_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VFMVVF_FLOAT vfmv_v_f_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFMULVF_FLOAT vfmul_vf_f64m2
+#endif
+
+
+// Optimizes the implementation in ../generic/trmmkernel_8x8.c
+
+
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FLOAT* C,BLASLONG ldc ,BLASLONG offset)
+{
+    //fprintf(stderr, "%s, %s, bm=%4ld bn=%4ld bk=%4ld alpha=%f ldc=%ld\n", __FILE__, __FUNCTION__, bm, bn, bk, alpha, ldc);
+
+    BLASLONG i,j,k;
+    FLOAT *C0,*C1,*C2,*C3,*C4,*C5,*C6,*C7,*ptrba,*ptrbb;
+
+    FLOAT_V_T va0, va1, va2, va3, va4, va5, va6, va7;
+    FLOAT_V_T vres0, vres1, vres2, vres3, vres4, vres5, vres6, vres7;
+    size_t vl;
+
+    BLASLONG off, temp;
+
+#if !defined(LEFT)
+    off = -offset;
+#else
+    off = 0;
+#endif
+    for (j = bn/8; j > 0; j--)
+    {
+        C0 = C;
+        C1 = C0+ldc;
+        C2 = C1+ldc;
+        C3 = C2+ldc;
+        C4 = C3+ldc;
+        C5 = C4+ldc;
+        C6 = C5+ldc;
+        C7 = C6+ldc;
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+        off = offset;
+#endif
+
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl)
+        {
+            vl = VSETVL(i);
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*vl;
+            ptrbb = bb + off*8;
+#endif
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+            vres2 = VFMVVF_FLOAT(0.0, vl);
+            vres3 = VFMVVF_FLOAT(0.0, vl);
+            vres4 = VFMVVF_FLOAT(0.0, vl);
+            vres5 = VFMVVF_FLOAT(0.0, vl);
+            vres6 = VFMVVF_FLOAT(0.0, vl);
+            vres7 = VFMVVF_FLOAT(0.0, vl);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+vl;  // number of values in A
+#else
+            temp = off+8;   // number of values in B
+#endif
+
+            for (k = temp/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va0, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va0, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va0, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va0, vl);
+                ptrbb += 8;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va1, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va1, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va1, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va1, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va1, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va1, vl);
+                ptrbb += 8;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va2, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va2, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va2, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va2, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va2, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va2, vl);
+                ptrbb += 8;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va3, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va3, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va3, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va3, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va3, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va3, vl);
+                ptrbb += 8;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va4, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va4, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va4, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va4, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va4, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va4, vl);
+                ptrbb += 8;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va5, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va5, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va5, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va5, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va5, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va5, vl);
+                ptrbb += 8;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va6, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va6, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va6, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va6, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va6, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va6, vl);
+                ptrbb += 8;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va7, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va7, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va7, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va7, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va7, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va7, vl);
+                ptrbb += 8;
+            }
+
+            for (k = temp&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl); // M:8 (should be vlen);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                vres4 = VFMACCVF_FLOAT(vres4, *(ptrbb + 4), va0, vl);
+                vres5 = VFMACCVF_FLOAT(vres5, *(ptrbb + 5), va0, vl);
+                vres6 = VFMACCVF_FLOAT(vres6, *(ptrbb + 6), va0, vl);
+                vres7 = VFMACCVF_FLOAT(vres7, *(ptrbb + 7), va0, vl);
+
+                ptrbb += 8;
+                ptrba += vl;
+            }
+
+            va0 = VFMULVF_FLOAT(vres0, alpha, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VFMULVF_FLOAT(vres1, alpha, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+            va2 = VFMULVF_FLOAT(vres2, alpha, vl);
+            VSEV_FLOAT(C2, va2, vl);
+
+            va3 = VFMULVF_FLOAT(vres3, alpha, vl);
+            VSEV_FLOAT(C3, va3, vl);
+
+            va4 = VFMULVF_FLOAT(vres4, alpha, vl);
+            VSEV_FLOAT(C4, va4, vl);
+
+            va5 = VFMULVF_FLOAT(vres5, alpha, vl);
+            VSEV_FLOAT(C5, va5, vl);
+
+            va6 = VFMULVF_FLOAT(vres6, alpha, vl);
+            VSEV_FLOAT(C6, va6, vl);
+
+            va7 = VFMULVF_FLOAT(vres7, alpha, vl);
+            VSEV_FLOAT(C7, va7, vl);
+
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= vl; // number of values in A
+#else
+            temp -= 8; // number of values in B
+#endif
+            ptrba += temp*vl;
+            ptrbb += temp*8;
+#endif
+
+#ifdef LEFT
+            off += vl; // number of values in A
+#endif
+
+            C0 += vl;
+            C1 += vl;
+            C2 += vl;
+            C3 += vl;
+            C4 += vl;
+            C5 += vl;
+            C6 += vl;
+            C7 += vl;
+        }
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 8;
+#endif
+
+        bb += (bk<<3);
+        C += (ldc<<3);
+    }
+
+    if (bn & 4)
+    {
+        C0 = C;
+        C1 = C0+ldc;
+        C2 = C1+ldc;
+        C3 = C2+ldc;
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+        off = offset;
+#endif
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl)
+        {
+            vl = VSETVL(i);
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*vl;
+            ptrbb = bb + off*4;
+#endif
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+            vres2 = VFMVVF_FLOAT(0.0, vl);
+            vres3 = VFMVVF_FLOAT(0.0, vl);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+vl;  // number of values in A
+#else
+            temp = off+4;   // number of values in B
+#endif
+
+            for (k = temp/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+                ptrbb += 4;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va1, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va1, vl);
+                ptrbb += 4;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va2, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va2, vl);
+                ptrbb += 4;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va3, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va3, vl);
+                ptrbb += 4;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va4, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va4, vl);
+                ptrbb += 4;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va5, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va5, vl);
+                ptrbb += 4;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va6, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va6, vl);
+                ptrbb += 4;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va7, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va7, vl);
+                ptrbb += 4;
+            }
+
+            // K remainder
+            for (k = temp&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                vres2 = VFMACCVF_FLOAT(vres2, *(ptrbb + 2), va0, vl);
+                vres3 = VFMACCVF_FLOAT(vres3, *(ptrbb + 3), va0, vl);
+
+                ptrbb += 4;
+                ptrba += vl;
+            }
+
+            va0 = VFMULVF_FLOAT(vres0, alpha, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VFMULVF_FLOAT(vres1, alpha, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+            va2 = VFMULVF_FLOAT(vres2, alpha, vl);
+            VSEV_FLOAT(C2, va2, vl);
+
+            va3 = VFMULVF_FLOAT(vres3, alpha, vl);
+            VSEV_FLOAT(C3, va3, vl);
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= vl; // number of values in A
+#else
+            temp -= 4; // number of values in B
+#endif
+            ptrba += temp*vl;
+            ptrbb += temp*4;
+#endif
+
+#ifdef LEFT
+            off += vl; // number of values in A
+#endif
+
+            C0 += vl;
+            C1 += vl;
+            C2 += vl;
+            C3 += vl;
+        }
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 4;
+#endif
+
+        bb += (bk<<2);
+        C += (ldc<<2);
+    }
+
+    if (bn & 2)
+    {
+        C0 = C;
+        C1 = C0+ldc;
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+        off = offset;
+#endif
+
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl) 
+        {
+            vl = VSETVL(i);
+
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*vl;
+            ptrbb = bb + off*2;
+#endif
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+            vres1 = VFMVVF_FLOAT(0.0, vl);
+
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+vl;  // number of values in A
+#else
+            temp = off+2;   // number of values in B
+#endif
+
+            for (k = temp/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+                ptrbb += 2;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va1, vl);
+                ptrbb += 2;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va2, vl);
+                ptrbb += 2;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va3, vl);
+                ptrbb += 2;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va4, vl);
+                ptrbb += 2;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va5, vl);
+                ptrbb += 2;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va6, vl);
+                ptrbb += 2;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va7, vl);
+                ptrbb += 2;
+            }
+
+            // K remainder
+            for (k = temp&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                vres1 = VFMACCVF_FLOAT(vres1, *(ptrbb + 1), va0, vl);
+
+                ptrbb += 2;
+                ptrba += vl;
+            }
+            va0 = VFMULVF_FLOAT(vres0, alpha, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+            va1 = VFMULVF_FLOAT(vres1, alpha, vl);
+            VSEV_FLOAT(C1, va1, vl);
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= vl; // number of values in A
+#else
+            temp -= 2; // number of values in B
+#endif
+            ptrba += temp*vl;
+            ptrbb += temp*2;
+#endif
+
+#ifdef LEFT
+            off += vl; // number of values in A
+#endif
+
+            C0 += vl;
+            C1 += vl;
+        }
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 2;
+#endif
+
+        bb += (bk<<1);
+        C += (ldc<<1);
+    }
+
+    if (bn & 1)
+    {
+        C0 = C;
+
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+        off = offset;
+#endif
+
+        ptrba = ba;
+
+        for (i = bm; i > 0; i -= vl)
+        {
+            vl = VSETVL(i);
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*vl;
+            ptrbb = bb + off*1;
+#endif
+
+            vres0 = VFMVVF_FLOAT(0.0, vl);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off+vl;  // number of values in A
+#else
+            temp = off+1;   // number of values in B
+#endif
+
+            for (k = temp/8; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+                va1 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+                ptrbb += 1;
+                va2 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va1, vl);
+                ptrbb += 1;
+                va3 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va2, vl);
+                ptrbb += 1;
+                va4 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va3, vl);
+                ptrbb += 1;
+                va5 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va4, vl);
+                ptrbb += 1;
+                va6 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va5, vl);
+                ptrbb += 1;
+                va7 = VLEV_FLOAT(ptrba, vl);
+                ptrba += vl;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va6, vl);
+                ptrbb += 1;
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va7, vl);
+                ptrbb += 1;
+            }
+
+            // K remainder
+            for (k = temp&7; k > 0; k--) {
+                va0 = VLEV_FLOAT(ptrba, vl);
+
+                vres0 = VFMACCVF_FLOAT(vres0, *(ptrbb + 0), va0, vl);
+
+                ptrbb += 1;
+                ptrba += vl;
+            }
+            va0 = VFMULVF_FLOAT(vres0, alpha, vl);
+            VSEV_FLOAT(C0, va0, vl);
+
+#if ( defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= vl; // number of values in A
+#else
+            temp -= 1; // number of values in B
+#endif
+            ptrba += temp*vl;
+            ptrbb += temp*1;
+#endif
+
+#ifdef LEFT
+            off += vl; // number of values in A
+#endif
+
+            C0 += vl;
+        }
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 1;
+#endif
+
+        bb += (bk);
+        C += (ldc);
+    }
+    return 0;
+}
+

kernel/riscv64/trsm_kernel_LN_rvv_v1.c

@@ -0,0 +1,847 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSSEV_FLOAT vsse32_v_f32m2
+#define VSSEG2_FLOAT vsseg2e32_v_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFMULVF_FLOAT vfmul_vf_f32m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSSEV_FLOAT vsse64_v_f64m2
+#define VSSEG2_FLOAT vsseg2e64_v_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFMULVF_FLOAT vfmul_vf_f64m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m2
+#endif
+
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_kernel_LN_sve.c
+
+#ifndef COMPLEX
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa,  bb;
+    FLOAT *pa, *pc;
+
+    int i, j, k;
+    //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  offset = %4ld\n", __FILE__, __FUNCTION__, m, n, ldc); // Debug
+
+    size_t vl;
+    FLOAT_V_T va, vc;
+
+    a += (m - 1) * m;
+    b += (m - 1) * n;
+
+    for (i = m - 1; i >= 0; i--) 
+    {
+        aa = *(a + i);
+        for (j = 0; j < n; j ++) 
+        {
+            bb = *(c + i + j * ldc);
+            bb *= aa;
+            *b             = bb;
+            *(c + i + j * ldc) = bb;
+            b ++;
+
+            pa = a;
+            pc = c + j * ldc;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc = VLEV_FLOAT(pc, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc = VFNMSACVF_FLOAT(vc, bb, va, vl);
+                VSEV_FLOAT(pc, vc, vl);
+                pa += vl;
+                pc += vl;
+            }
+        }
+        a -= m;
+        b -= 2 * n;
+    }
+
+}
+#elif GEMM_DEFAULT_UNROLL_N == 2
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa,  bb0, bb1;
+    FLOAT *pa, *pc, *pc0, *pc1;
+    FLOAT *pb0, *pb1;
+
+    int i, j, k;
+    fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  offset = %4ld\n", __FILE__, __FUNCTION__, m, n, ldc); // Debug
+
+    size_t vl;
+    FLOAT_V_T va, vc0, vc1;
+
+    a += (m - 1) * m;
+    b += (m - 1) * n;
+
+    for (i = m - 1; i >= 0; i--) 
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/2; j ++) 
+        {
+            //bb = *(c + i + j * ldc);
+            pb0 = pc + j * ldc * 2;
+            pb1 = pb0 + ldc;
+            //bb *= aa;
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            //*b             = bb;
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *pb0    = bb0;
+            *pb1    = bb1;
+
+            //*(c + i + j * ldc) = bb;
+            //b ++;
+
+            b += 2;
+            //pa = a + i + 1;
+            pc0 = c + j * ldc * 2;
+            pc1 = pc0 + ldc;
+            pa = a;
+            //pc = c + j * ldc;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+        }
+        pc += ldc * (n/2) * 2;
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = (*pb0) * aa;
+            *b      = bb0;
+            *pb0    = bb0;
+            b += 1;
+
+            pc0 = pc - i;
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+
+                pa += vl;
+                pc0 += vl;
+            }
+        }
+
+        a -= m;
+        b -= 2 * n;
+    }
+
+}
+
+#elif GEMM_DEFAULT_UNROLL_N == 4
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa,  bb0, bb1, bb2, bb3;
+    FLOAT *pa, *pc, *pc0, *pc1, *pc2, *pc3;
+    FLOAT *pb0, *pb1, *pb2, *pb3;
+
+    int i, j, k;
+
+    size_t vl;
+    FLOAT_V_T va, vc0, vc1, vc2, vc3;
+
+    a += (m - 1) * m;
+    b += (m - 1) * n;
+
+    for (i = m - 1; i >= 0; i--) 
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/4; j ++) 
+        {
+            pb0 = pc + j * ldc * 4;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+
+            b += 4;
+
+            pc0 = c + j * ldc * 4;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+            }
+        }
+        pc += ldc * (n/4) * 4;
+
+        if (n & 2)
+        {
+            pb0 = pc + j * ldc * 2;
+            pb1 = pb0 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+
+            b += 2;
+
+            pc0 = c + j * ldc * 2;
+            pc1 = pc0 + ldc;
+
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+            pc += ldc * 2;
+        }
+
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = (*pb0) * aa;
+            *b      = bb0;
+            *pb0    = bb0;
+            b += 1;
+
+            pc0 = pc - i;
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+
+                pa += vl;
+                pc0 += vl;
+            }
+        }
+
+        a -= m;
+        b -= 2 * n;
+    }
+
+}
+#elif GEMM_DEFAULT_UNROLL_N == 8
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa,  bb0, bb1, bb2, bb3, bb4, bb5, bb6, bb7;
+    FLOAT *pa, *pc, *pc0, *pc1, *pc2, *pc3, *pc4, *pc5, *pc6, *pc7;
+    FLOAT *pb0, *pb1, *pb2, *pb3, *pb4, *pb5, *pb6, *pb7;
+
+    int i, j, k;
+
+    size_t vl;
+    FLOAT_V_T va, vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7;
+
+    a += (m - 1) * m;
+    b += (m - 1) * n;
+
+    for (i = m - 1; i >= 0; i--) 
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/8; j ++) 
+        {
+            pb0 = pc + j * ldc * 8;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+            pb4 = pb3 + ldc;
+            pb5 = pb4 + ldc;
+            pb6 = pb5 + ldc;
+            pb7 = pb6 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+            bb4 = (*pb4) * aa;
+            bb5 = (*pb5) * aa;
+            bb6 = (*pb6) * aa;
+            bb7 = (*pb7) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+            *(b+4)  = bb4;
+            *(b+5)  = bb5;
+            *(b+6)  = bb6;
+            *(b+7)  = bb7;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+            *pb4    = bb4;
+            *pb5    = bb5;
+            *pb6    = bb6;
+            *pb7    = bb7;
+
+            b += 8;
+
+            pc0 = c + j * ldc * 8;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+            pc4 = pc3 + ldc;
+            pc5 = pc4 + ldc;
+            pc6 = pc5 + ldc;
+            pc7 = pc6 + ldc;
+
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+                vc4 = VLEV_FLOAT(pc4, vl);
+                vc5 = VLEV_FLOAT(pc5, vl);
+                vc6 = VLEV_FLOAT(pc6, vl);
+                vc7 = VLEV_FLOAT(pc7, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+                vc4 = VFNMSACVF_FLOAT(vc4, bb4, va, vl);
+                vc5 = VFNMSACVF_FLOAT(vc5, bb5, va, vl);
+                vc6 = VFNMSACVF_FLOAT(vc6, bb6, va, vl);
+                vc7 = VFNMSACVF_FLOAT(vc7, bb7, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+                VSEV_FLOAT(pc4, vc4, vl);
+                VSEV_FLOAT(pc5, vc5, vl);
+                VSEV_FLOAT(pc6, vc6, vl);
+                VSEV_FLOAT(pc7, vc7, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+                pc4 += vl;
+                pc5 += vl;
+                pc6 += vl;
+                pc7 += vl;
+            }
+        }
+        pc += ldc * (n/8) * 8;
+
+        if (n & 4)
+        {
+            pb0 = pc + j * ldc * 4;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+
+            b += 4;
+
+            pc0 = c + j * ldc * 4;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+            }
+            pc += ldc * 4;
+        }
+
+        if (n & 2)
+        {
+            pb0 = pc + j * ldc * 2;
+            pb1 = pb0 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+
+            b += 2;
+
+            pc0 = c + j * ldc * 2;
+            pc1 = pc0 + ldc;
+
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+
+                pa += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+            pc += ldc * 2;
+        }
+
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = (*pb0) * aa;
+            *b      = bb0;
+            *pb0    = bb0;
+            b += 1;
+
+            pc0 = pc - i;
+            pa = a;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+
+                pa += vl;
+                pc0 += vl;
+            }
+        }
+
+        a -= m;
+        b -= 2 * n;
+    }
+
+}
+#else
+static inline void solve_generic(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = 0; k < i; k ++){
+        *(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#endif
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+        *(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+        *(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+        *(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+        *(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+    FLOAT dummy2,
+#endif
+    FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  
+  size_t vl = VSETVL_MAX;
+
+    //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  k = %4ld offset = %4ld\n", __FILE__, __FUNCTION__, m, n, k, offset); // Debug
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    i = m % vl;
+    if (i) {
+      aa = a + (m - i) * k * COMPSIZE;
+      cc = c + (m - i)     * COMPSIZE;
+
+      if (k - kk > 0) {
+        GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+            ZERO,
+#endif
+            aa + i             * kk * COMPSIZE,
+            b  + GEMM_UNROLL_N * kk * COMPSIZE,
+            cc,
+            ldc);
+      }
+
+      solve(i, GEMM_UNROLL_N,
+          aa + (kk - i) * i             * COMPSIZE,
+          b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+          cc, ldc);
+
+      kk -= i;
+
+    }
+
+    int mod = i;
+    i = vl;
+    if (i <= m) {
+      aa = a + (m - mod - vl) * k * COMPSIZE;
+      cc = c + (m - mod - vl)     * COMPSIZE;
+
+      do {
+        if (k - kk > 0) {
+          GEMM_KERNEL(vl, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+              ZERO,
+#endif
+              aa + vl * kk * COMPSIZE,
+              b +  GEMM_UNROLL_N * kk * COMPSIZE,
+              cc,
+              ldc);
+        }
+
+        solve(vl, GEMM_UNROLL_N,
+            aa + (kk - vl) * vl * COMPSIZE,
+            b  + (kk - vl) * GEMM_UNROLL_N * COMPSIZE,
+            cc, ldc);
+
+        aa -= vl * k * COMPSIZE;
+        cc -= vl     * COMPSIZE;
+        kk -= vl;
+
+        i += vl;
+      } while (i <= m);
+    }
+
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+        kk = m + offset;
+
+        i = m % vl;
+        if (i) {
+          aa = a + (m - i) * k * COMPSIZE;
+          cc = c + (m - i)     * COMPSIZE;
+
+          if (k - kk > 0) {
+            GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+                ZERO,
+#endif
+                aa + i * kk * COMPSIZE,
+                b  + j * kk * COMPSIZE,
+                cc, ldc);
+          }
+
+          solve(i, j,
+              aa + (kk - i) * i * COMPSIZE,
+              b  + (kk - i) * j * COMPSIZE,
+              cc, ldc);
+
+          kk -= i;
+
+        }
+
+        int mod = i;
+        i = vl;
+        if (i <= m) {
+          aa = a + (m - mod - vl) * k * COMPSIZE;
+          cc = c + (m - mod - vl)     * COMPSIZE;
+
+          do {
+            if (k - kk > 0) {
+              GEMM_KERNEL(vl, j, k - kk, dm1,
+#ifdef COMPLEX
+                  ZERO,
+#endif
+                  aa + vl * kk * COMPSIZE,
+                  b +  j             * kk * COMPSIZE,
+                  cc,
+                  ldc);
+            }
+
+            solve(vl, j,
+                aa + (kk - vl) * vl * COMPSIZE,
+                b  + (kk - vl) * j             * COMPSIZE,
+                cc, ldc);
+
+            aa -= vl * k * COMPSIZE;
+            cc -= vl     * COMPSIZE;
+            kk -= vl;
+
+            i += vl;
+          } while (i <= m);
+        }
+
+        b += j * k   * COMPSIZE;
+        c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/riscv64/trsm_kernel_LT_rvv_v1.c

@@ -0,0 +1,840 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSSEV_FLOAT vsse32_v_f32m2
+#define VSSEG2_FLOAT vsseg2e32_v_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFMULVF_FLOAT vfmul_vf_f32m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSSEV_FLOAT vsse64_v_f64m2
+#define VSSEG2_FLOAT vsseg2e64_v_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFMULVF_FLOAT vfmul_vf_f64m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m2
+#endif
+
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_kernel_LT_sve.c
+
+#ifndef COMPLEX
+#if GEMM_DEFAULT_UNROLL_N == 1
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) 
+{
+    FLOAT aa, bb;
+    FLOAT *pa, *pc;
+
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T va, vc;
+    for (i = 0; i < m; i++)
+    {
+        aa = *(a + i);
+        for (j = 0; j < n; j ++)
+        {
+            bb = *(c + i + j * ldc);
+            bb *= aa;
+            *b  = bb;
+            *(c + i + j * ldc) = bb;
+            b++;
+            pa = a + i + 1;
+            pc = c + j * ldc + i + 1;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc = VLEV_FLOAT(pc, vl);
+                va = VLEV_FLOAT(pa, vl);
+                vc = VFNMSACVF_FLOAT(vc, bb, va, vl);
+                VSEV_FLOAT(pc, vc, vl);
+                pa += vl;
+                pc += vl;
+            }
+        }
+        a += m;
+    }
+}
+#elif GEMM_DEFAULT_UNROLL_N == 2
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) 
+{
+
+    FLOAT aa, bb0, bb1;
+    FLOAT *pa, *pc, *pc0, *pc1;
+    FLOAT *pb0, *pb1;
+
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T va, vc0, vc1;
+    for (i = 0; i < m; i++)
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/2; j ++)
+        {
+            pb0 = pc + j * ldc * 2;
+            pb1 = pb0 + ldc;
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *pb0    = bb0;
+            *pb1    = bb1;
+            b += 2;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+        }
+        pc += ldc * (n/2) * 2;
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = *(pb0);
+            bb0 *= aa;
+            *b  = bb0;
+            *(c + i) = bb0;
+            b++;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                pa  += vl;
+                pc0 += vl;
+            }
+        }
+
+        a += m;
+    }
+}
+#elif GEMM_DEFAULT_UNROLL_N == 4
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) 
+{
+
+    FLOAT aa, bb0, bb1, bb2, bb3;
+    FLOAT *pa, *pc;
+    FLOAT *pc0, *pc1, *pc2, *pc3;
+    FLOAT *pb0, *pb1, *pb2, *pb3;
+
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T va;
+    FLOAT_V_T vc0, vc1, vc2, vc3;
+    for (i = 0; i < m; i++)
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/4; j ++)
+        {
+            pb0 = pc;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+            b += 4;
+
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+
+                va  = VLEV_FLOAT(pa, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+            }
+        }
+        pc += ldc * (n/4) * 4;
+
+        if (n & 2)
+        {
+            pb0 = pc;
+            pb1 = pb0 + ldc;
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *pb0    = bb0;
+            *pb1    = bb1;
+            b += 2;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+            pc += ldc * 2;
+        }
+
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = *(pb0);
+            bb0 *= aa;
+            *b  = bb0;
+            *(c + i) = bb0;
+            b++;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                pa  += vl;
+                pc0 += vl;
+            }
+        }
+
+        a += m;
+    }
+}
+#elif GEMM_DEFAULT_UNROLL_N == 8
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) 
+{
+
+    FLOAT aa, bb0, bb1, bb2, bb3, bb4, bb5, bb6, bb7;
+    FLOAT *pa, *pc;
+    FLOAT *pc0, *pc1, *pc2, *pc3, *pc4, *pc5, *pc6, *pc7;
+    FLOAT *pb0, *pb1, *pb2, *pb3, *pb4, *pb5, *pb6, *pb7;
+
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T va;
+    FLOAT_V_T vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7;
+    for (i = 0; i < m; i++)
+    {
+        aa = *(a + i);
+        pc = c + i;
+        for (j = 0; j < n/8; j ++)
+        {
+            pb0 = pc + j * ldc * 8;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+            pb4 = pb3 + ldc;
+            pb5 = pb4 + ldc;
+            pb6 = pb5 + ldc;
+            pb7 = pb6 + ldc;
+            
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+            bb4 = (*pb4) * aa;
+            bb5 = (*pb5) * aa;
+            bb6 = (*pb6) * aa;
+            bb7 = (*pb7) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+            *(b+4)  = bb4;
+            *(b+5)  = bb5;
+            *(b+6)  = bb6;
+            *(b+7)  = bb7;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+            *pb4    = bb4;
+            *pb5    = bb5;
+            *pb6    = bb6;
+            *pb7    = bb7;
+            b += 8;
+
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+            pc4 = pc3 + ldc;
+            pc5 = pc4 + ldc;
+            pc6 = pc5 + ldc;
+            pc7 = pc6 + ldc;
+
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+                vc4 = VLEV_FLOAT(pc4, vl);
+                vc5 = VLEV_FLOAT(pc5, vl);
+                vc6 = VLEV_FLOAT(pc6, vl);
+                vc7 = VLEV_FLOAT(pc7, vl);
+
+                va  = VLEV_FLOAT(pa, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+                vc4 = VFNMSACVF_FLOAT(vc4, bb4, va, vl);
+                vc5 = VFNMSACVF_FLOAT(vc5, bb5, va, vl);
+                vc6 = VFNMSACVF_FLOAT(vc6, bb6, va, vl);
+                vc7 = VFNMSACVF_FLOAT(vc7, bb7, va, vl);
+
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+                VSEV_FLOAT(pc4, vc4, vl);
+                VSEV_FLOAT(pc5, vc5, vl);
+                VSEV_FLOAT(pc6, vc6, vl);
+                VSEV_FLOAT(pc7, vc7, vl);
+
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+                pc4 += vl;
+                pc5 += vl;
+                pc6 += vl;
+                pc7 += vl;
+            }
+        }
+        pc += ldc * (n/8) * 8;
+
+        if (n & 4)
+        {
+            pb0 = pc;
+            pb1 = pb0 + ldc;
+            pb2 = pb1 + ldc;
+            pb3 = pb2 + ldc;
+
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            bb2 = (*pb2) * aa;
+            bb3 = (*pb3) * aa;
+
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *(b+2)  = bb2;
+            *(b+3)  = bb3;
+
+            *pb0    = bb0;
+            *pb1    = bb1;
+            *pb2    = bb2;
+            *pb3    = bb3;
+            b += 4;
+
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            pc2 = pc1 + ldc;
+            pc3 = pc2 + ldc;
+
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                vc2 = VLEV_FLOAT(pc2, vl);
+                vc3 = VLEV_FLOAT(pc3, vl);
+
+                va  = VLEV_FLOAT(pa, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, bb2, va, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, bb3, va, vl);
+
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                VSEV_FLOAT(pc2, vc2, vl);
+                VSEV_FLOAT(pc3, vc3, vl);
+
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+                pc2 += vl;
+                pc3 += vl;
+            }
+            pc += ldc * 4;
+        }
+
+        if (n & 2)
+        {
+            pb0 = pc;
+            pb1 = pb0 + ldc;
+            bb0 = (*pb0) * aa;
+            bb1 = (*pb1) * aa;
+            *b      = bb0;
+            *(b+1)  = bb1;
+            *pb0    = bb0;
+            *pb1    = bb1;
+            b += 2;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            pc1 = pc0 + ldc;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl  = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                vc1 = VLEV_FLOAT(pc1, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, bb1, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                VSEV_FLOAT(pc1, vc1, vl);
+                pa  += vl;
+                pc0 += vl;
+                pc1 += vl;
+            }
+            pc += ldc * 2;
+        }
+
+        if (n & 1)
+        {
+            pb0 = pc;
+            bb0 = *(pb0);
+            bb0 *= aa;
+            *b  = bb0;
+            *(c + i) = bb0;
+            b++;
+            pa = a + i + 1;
+            pc0 = pb0 + 1;
+            for (k = (m - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLEV_FLOAT(pc0, vl);
+                va  = VLEV_FLOAT(pa, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, bb0, va, vl);
+                VSEV_FLOAT(pc0, vc0, vl);
+                pa  += vl;
+                pc0 += vl;
+            }
+        }
+
+        a += m;
+    }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < m; i++) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = i + 1; k < m; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a += m;
+  }
+}
+
+#endif
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = i + 1; k < m; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= -cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a += m * 2;
+  }
+}
+
+
+static inline void solve_N1(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+  FLOAT *pa, *pc;
+
+  int i, j, k;
+
+  size_t vl;
+  FLOAT_V_T va0, va1, vc0, vc1;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+        pa = a + (i + 1) * 2;
+        pc = c + j * ldc + (i + 1) * 2;
+        for (k = (m - i - 1); k > 0; k -= vl)
+        {
+            vl = VSETVL(k);
+            VLSEG2_FLOAT(&va0, &va1, pa, vl);
+            VLSEG2_FLOAT(&vc0, &vc1, pc, vl);
+#ifndef CONJ
+            vc0 = VFNMSACVF_FLOAT(vc0, cc1, va0);
+            vc0 = VFMACCVF_FLOAT(vc0, cc2, va1);
+            vc1 = VFNMSACVF_FLOAT(vc1, cc1, va1);
+            vc1 = VFNMSACVF_FLOAT(vc1, cc2, va0);
+#else
+            vc0 = VFNMSACVF_FLOAT(vc0, cc1, va0);
+            vc0 = VFNMSACVF_FLOAT(vc0, cc2, va1);
+            vc1 = VFMACCVF_FLOAT(vc1, cc1, va1);
+            vc1 = VFNMSACVF_FLOAT(vc1, cc2, va0);
+#endif
+            VSSEG2_FLOAT(pc, vc0, vc1, vl);
+            pa += vl * 2;
+            pc += vl * 2;
+        }
+      }
+    }
+    a += m * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j;
+
+  size_t vl = VSETVL_MAX;
+
+  //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  k = %4ld offset = %4ld\n", __FILE__, __FUNCTION__, m, n, k, offset); // Debug
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = offset;
+    aa = a;
+    cc = c;
+
+    i = vl;
+
+    while (i <= m) {
+
+      if (kk > 0) {
+        GEMM_KERNEL(vl, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+            ZERO,
+#endif
+            aa, b, cc, ldc);
+      }
+
+      solve(vl, GEMM_UNROLL_N,
+          aa + kk * vl * COMPSIZE,
+          b  + kk * GEMM_UNROLL_N * COMPSIZE,
+          cc, ldc);
+
+      aa += vl * k * COMPSIZE;
+      cc += vl     * COMPSIZE;
+      kk += vl;
+      i += vl;
+    }
+
+    i = m % vl;
+    if (i) {
+      if (kk > 0) {
+        GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+            ZERO,
+#endif
+            aa, b, cc, ldc);
+      }
+      solve(i, GEMM_UNROLL_N,
+          aa + kk * i             * COMPSIZE,
+          b  + kk * GEMM_UNROLL_N * COMPSIZE,
+          cc, ldc);
+
+      aa += i * k * COMPSIZE;
+      cc += i     * COMPSIZE;
+      kk += i;
+
+    }
+
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+        kk = offset;
+        aa = a;
+        cc = c;
+
+        i = vl;
+
+        while (i <= m) {
+          if (kk > 0) {
+            GEMM_KERNEL(vl, j, kk, dm1,
+#ifdef COMPLEX
+                ZERO,
+#endif
+                aa,
+                b,
+                cc,
+                ldc);
+          }
+
+          solve(vl, j,
+              aa + kk * vl * COMPSIZE,
+              b  + kk * j             * COMPSIZE, cc, ldc);
+
+          aa += vl * k * COMPSIZE;
+          cc += vl     * COMPSIZE;
+          kk += vl;
+          i += vl;
+        }
+
+        i = m % vl;
+        if (i) {
+          if (kk > 0) {
+            GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+                ZERO,
+#endif
+                aa,
+                b,
+                cc,
+                ldc);
+          }
+
+          solve(i, j,
+              aa + kk * i * COMPSIZE,
+              b  + kk * j * COMPSIZE, cc, ldc);
+
+          aa += i * k * COMPSIZE;
+          cc += i     * COMPSIZE;
+          kk += i;
+
+        }
+
+        b += j * k   * COMPSIZE;
+        c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/riscv64/trsm_kernel_RN_rvv_v1.c

@@ -0,0 +1,792 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSSEV_FLOAT vsse32_v_f32m2
+#define VSSEG2_FLOAT vsseg2e32_v_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSSEV_FLOAT vsse64_v_f64m2
+#define VSSEG2_FLOAT vsseg2e64_v_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m2
+#endif
+
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_kernel_RN_sve.c
+
+#ifndef COMPLEX
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa, bb;
+    FLOAT *pb, *pc;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc;
+
+    for (i = 0; i < n; i++) 
+    {
+        bb = *(b + i);
+
+        for (j = 0; j < m; j ++) 
+        {
+            aa = *(c + j + i * ldc);
+            aa *= bb;
+            *a  = aa;
+            *(c + j + i * ldc) = aa;
+            a ++;
+
+            pb = b + i + 1;
+            pc = c + j + (i + 1) *ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc = VLSEV_FLOAT(pc, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc = VFNMSACVF_FLOAT(vc, aa, vb, vl);
+                VSSEV_FLOAT(pc, stride_ldc, vc, vl);
+                pb += vl;
+                pc ++;
+            }
+        }
+        b += n;
+    }
+}
+
+#elif GEMM_DEFAULT_UNROLL_N == 2
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa0, aa1, bb;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pc0, *pc1;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1;
+
+    for (i = 0; i < n; i++) 
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/2; j ++) 
+        {
+            pa0 = pc + j * 2;
+            pa1 = pc + j * 2 + 1;
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *a      = aa0;
+            *(a + 1)= aa1;
+            a  += 2;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+        }
+        pc += (m/2)*2;
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b += n;
+    }
+}
+
+#elif GEMM_DEFAULT_UNROLL_N == 4
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT bb;
+    FLOAT aa0, aa1, aa2, aa3;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pa2, *pa3;
+    FLOAT *pc0, *pc1, *pc2, *pc3;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1, vc2, vc3;
+
+    for (i = 0; i < n; i++) 
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/4; j ++) 
+        {
+            pa0 = pc + j * 4;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+
+            a  += 4;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            pc2 = pa2 + ldc;
+            pc3 = pa3 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+ 
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+            }
+        }
+        pc += (m/4)*4;
+
+        if (m & 2)
+        {
+            pa0 = pc;
+            pa1 = pa0 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+
+            a  += 2;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+ 
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+            pc += 2;
+        }
+
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b += n;
+    }
+}
+
+#elif GEMM_DEFAULT_UNROLL_N == 8
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT bb;
+    FLOAT aa0, aa1, aa2, aa3, aa4, aa5, aa6, aa7;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pa2, *pa3, *pa4, *pa5, *pa6, *pa7;
+    FLOAT *pc0, *pc1, *pc2, *pc3, *pc4, *pc5, *pc6, *pc7;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7;
+
+    for (i = 0; i < n; i++) 
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/8; j ++) 
+        {
+            pa0 = pc + j * 8;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            pa4 = pa3 + 1;
+            pa5 = pa4 + 1;
+            pa6 = pa5 + 1;
+            pa7 = pa6 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+            aa4 = *pa4 * bb;
+            aa5 = *pa5 * bb;
+            aa6 = *pa6 * bb;
+            aa7 = *pa7 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+            *pa4    = aa4;
+            *pa5    = aa5;
+            *pa6    = aa6;
+            *pa7    = aa7;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+            *(a + 4)= aa4;
+            *(a + 5)= aa5;
+            *(a + 6)= aa6;
+            *(a + 7)= aa7;
+
+            a  += 8;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            pc2 = pa2 + ldc;
+            pc3 = pa3 + ldc;
+            pc4 = pa4 + ldc;
+            pc5 = pa5 + ldc;
+            pc6 = pa6 + ldc;
+            pc7 = pa7 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vc4 = VLSEV_FLOAT(pc4, stride_ldc, vl);
+                vc5 = VLSEV_FLOAT(pc5, stride_ldc, vl);
+                vc6 = VLSEV_FLOAT(pc6, stride_ldc, vl);
+                vc7 = VLSEV_FLOAT(pc7, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+ 
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+                vc4 = VFNMSACVF_FLOAT(vc4, aa4, vb, vl);
+                vc5 = VFNMSACVF_FLOAT(vc5, aa5, vb, vl);
+                vc6 = VFNMSACVF_FLOAT(vc6, aa6, vb, vl);
+                vc7 = VFNMSACVF_FLOAT(vc7, aa7, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+                VSSEV_FLOAT(pc4, stride_ldc, vc4, vl);
+                VSSEV_FLOAT(pc5, stride_ldc, vc5, vl);
+                VSSEV_FLOAT(pc6, stride_ldc, vc6, vl);
+                VSSEV_FLOAT(pc7, stride_ldc, vc7, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+                pc4++;
+                pc5++;
+                pc6++;
+                pc7++;
+            }
+        }
+        pc += (m/8)*8;
+
+        if (m & 4)
+        {
+            pa0 = pc;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+
+            a  += 4;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            pc2 = pa2 + ldc;
+            pc3 = pa3 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+ 
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+            }
+            pc += 4;
+        }
+
+        if (m & 2)
+        {
+            pa0 = pc;
+            pa1 = pa0 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+
+            a  += 2;
+
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            pc1 = pa1 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+ 
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+            pc += 2;
+        }
+
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b + i + 1;
+            pc0 = pa0 + ldc;
+            for (k = (n - i - 1); k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b += n;
+    }
+}
+#else
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#endif
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j;
+
+  size_t vl = VSETVL_MAX;
+
+  //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  k = %4ld offset = %4ld\n", __FILE__, __FUNCTION__, m, n, k, offset); // Debug
+
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = vl;
+
+    if (i <= m) {
+      do {
+	if (kk > 0) {
+	  GEMM_KERNEL(vl, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa, b, cc, ldc);
+	}
+
+	solve(vl, GEMM_UNROLL_N,
+	      aa + kk * vl * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+
+	aa += vl * k * COMPSIZE;
+	cc += vl     * COMPSIZE;
+	i += vl;
+      } while (i <= m);
+    }
+
+
+    i = m % vl;
+    if (i) {
+      if (kk > 0) {
+        GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+            ZERO,
+#endif
+            aa, b, cc, ldc);
+      }
+      solve(i, GEMM_UNROLL_N,
+          aa + kk * i             * COMPSIZE,
+          b  + kk * GEMM_UNROLL_N * COMPSIZE,
+          cc, ldc);
+
+      aa += i * k * COMPSIZE;
+      cc += i     * COMPSIZE;
+
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+  i = vl;
+
+	while (i <= m) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(vl, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(vl, j,
+		aa + kk * vl * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += vl * k * COMPSIZE;
+	  cc += vl     * COMPSIZE;
+	  i += vl;
+	}
+
+  i = m % vl;
+  if (i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+  }
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/riscv64/trsm_kernel_RT_rvv_v1.c

@@ -0,0 +1,828 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSSEV_FLOAT vsse32_v_f32m2
+#define VSSEG2_FLOAT vsseg2e32_v_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m2
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSSEV_FLOAT vsse64_v_f64m2
+#define VSSEG2_FLOAT vsseg2e64_v_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m2
+#endif
+
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_kernel_RT_sve.c
+
+#ifndef COMPLEX
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+  FLOAT *pb, *pc;
+  BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+
+  int i, j, k;
+  size_t vl;
+  FLOAT_V_T vb, vc;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+        pb = b;
+        pc = c + j;
+        for (k = i; k > 0; k -= vl)
+        {
+            vl = VSETVL(k);
+            vc = VLSEV_FLOAT(pc, stride_ldc, vl);
+            vb = VLEV_FLOAT(pb, vl);
+            vc = VFNMSACVF_FLOAT(vc, aa, vb, vl);
+            VSSEV_FLOAT(pc, stride_ldc, vc, vl);
+            pb += vl;
+            pc++;
+        }
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+#elif GEMM_DEFAULT_UNROLL_N == 2
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa0, aa1, bb;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pc0, *pc1;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1;
+
+    a += (n - 1) * m;
+    b += (n - 1) * n;
+
+    for (i = n - 1; i >= 0; i--)
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/2; j ++) 
+        {
+            pa0 = pc + j * 2;
+            pa1 = pc + j * 2 + 1;
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *a      = aa0;
+            *(a + 1)= aa1;
+            a  += 2;
+
+            pb  = b;
+            pc0 = c + j * 2;
+            pc1 = pc0 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+        }
+        pc += (m/2)*2;
+
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b;
+            pc0 = pc - i * ldc;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b -= n;
+        a -= 2 * m;
+    }
+}
+
+#elif GEMM_DEFAULT_UNROLL_N == 4
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa0, aa1, aa2, aa3;
+    FLOAT bb;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pa2, *pa3;
+    FLOAT *pc0, *pc1, *pc2, *pc3;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1, vc2, vc3;
+
+    a += (n - 1) * m;
+    b += (n - 1) * n;
+
+    for (i = n - 1; i >= 0; i--)
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/4; j ++) 
+        {
+            pa0 = pc + j * 4;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+            a  += 4;
+
+            pb  = b;
+            pc0 = c + j * 4;
+            pc1 = pc0 + 1;
+            pc2 = pc1 + 1;
+            pc3 = pc2 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+            }
+        }
+        pc += (m/4)*4;
+
+        if (m & 2)
+        {
+            pa0 = pc + j * 2;
+            pa1 = pa0 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            a  += 2;
+
+            pb  = b;
+            pc0 = c + j * 4;
+            pc1 = pc0 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+            pc += 2;
+        }
+
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b;
+            pc0 = pc - i * ldc;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b -= n;
+        a -= 2 * m;
+    }
+}
+#elif GEMM_DEFAULT_UNROLL_N == 8
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+    FLOAT aa0, aa1, aa2, aa3, aa4, aa5, aa6, aa7;
+    FLOAT bb;
+    FLOAT *pb, *pc;
+    FLOAT *pa0, *pa1, *pa2, *pa3, *pa4, *pa5, *pa6, *pa7;
+    FLOAT *pc0, *pc1, *pc2, *pc3, *pc4, *pc5, *pc6, *pc7;
+    BLASLONG stride_ldc = sizeof(FLOAT) * ldc;
+    int i, j, k;
+    size_t vl;
+    FLOAT_V_T vb, vc0, vc1, vc2, vc3, vc4, vc5, vc6, vc7;
+
+    a += (n - 1) * m;
+    b += (n - 1) * n;
+
+    for (i = n - 1; i >= 0; i--)
+    {
+        bb = *(b + i);
+        pc = c + i * ldc;
+        for (j = 0; j < m/8; j ++) 
+        {
+            pa0 = pc + j * 8;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            pa4 = pa3 + 1;
+            pa5 = pa4 + 1;
+            pa6 = pa5 + 1;
+            pa7 = pa6 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+            aa4 = *pa4 * bb;
+            aa5 = *pa5 * bb;
+            aa6 = *pa6 * bb;
+            aa7 = *pa7 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+            *pa4    = aa4;
+            *pa5    = aa5;
+            *pa6    = aa6;
+            *pa7    = aa7;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+            *(a + 4)= aa4;
+            *(a + 5)= aa5;
+            *(a + 6)= aa6;
+            *(a + 7)= aa7;
+            a  += 8;
+
+            pb  = b;
+            pc0 = c + j * 8;
+            pc1 = pc0 + 1;
+            pc2 = pc1 + 1;
+            pc3 = pc2 + 1;
+            pc4 = pc3 + 1;
+            pc5 = pc4 + 1;
+            pc6 = pc5 + 1;
+            pc7 = pc6 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vc4 = VLSEV_FLOAT(pc4, stride_ldc, vl);
+                vc5 = VLSEV_FLOAT(pc5, stride_ldc, vl);
+                vc6 = VLSEV_FLOAT(pc6, stride_ldc, vl);
+                vc7 = VLSEV_FLOAT(pc7, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+                vc4 = VFNMSACVF_FLOAT(vc4, aa4, vb, vl);
+                vc5 = VFNMSACVF_FLOAT(vc5, aa5, vb, vl);
+                vc6 = VFNMSACVF_FLOAT(vc6, aa6, vb, vl);
+                vc7 = VFNMSACVF_FLOAT(vc7, aa7, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+                VSSEV_FLOAT(pc4, stride_ldc, vc4, vl);
+                VSSEV_FLOAT(pc5, stride_ldc, vc5, vl);
+                VSSEV_FLOAT(pc6, stride_ldc, vc6, vl);
+                VSSEV_FLOAT(pc7, stride_ldc, vc7, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+                pc4++;
+                pc5++;
+                pc6++;
+                pc7++;
+            }
+        }
+        pc += (m/8)*8;
+
+        if (m & 4)
+        {
+            pa0 = pc;
+            pa1 = pa0 + 1;
+            pa2 = pa1 + 1;
+            pa3 = pa2 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+            aa2 = *pa2 * bb;
+            aa3 = *pa3 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+            *pa2    = aa2;
+            *pa3    = aa3;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            *(a + 2)= aa2;
+            *(a + 3)= aa3;
+            a  += 4;
+
+            pb  = b;
+            pc0 = pc - i * ldc;
+            pc1 = pc0 + 1;
+            pc2 = pc1 + 1;
+            pc3 = pc2 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vc2 = VLSEV_FLOAT(pc2, stride_ldc, vl);
+                vc3 = VLSEV_FLOAT(pc3, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+                vc2 = VFNMSACVF_FLOAT(vc2, aa2, vb, vl);
+                vc3 = VFNMSACVF_FLOAT(vc3, aa3, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+                VSSEV_FLOAT(pc2, stride_ldc, vc2, vl);
+                VSSEV_FLOAT(pc3, stride_ldc, vc3, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+                pc2++;
+                pc3++;
+            }
+            pc += 4;
+        }
+
+        if (m & 2)
+        {
+            pa0 = pc;
+            pa1 = pa0 + 1;
+            
+            aa0 = *pa0 * bb;
+            aa1 = *pa1 * bb;
+
+            *pa0    = aa0;
+            *pa1    = aa1;
+
+            *a      = aa0;
+            *(a + 1)= aa1;
+            a  += 2;
+
+            pb  = b;
+            pc0 = pc - i * ldc;
+            pc1 = pc0 + 1;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vc1 = VLSEV_FLOAT(pc1, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                vc1 = VFNMSACVF_FLOAT(vc1, aa1, vb, vl);
+
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                VSSEV_FLOAT(pc1, stride_ldc, vc1, vl);
+
+                pb += vl;
+                pc0++;
+                pc1++;
+            }
+            pc += 2;
+        }
+
+        if (m & 1)
+        {
+            pa0 = pc;
+            aa0 = *pa0 * bb;
+            
+            *pa0    = aa0;
+            *a      = aa0;
+            a  += 1;
+           
+            pb = b;
+            pc0 = pc - i * ldc;
+            for (k = i; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                vc0 = VLSEV_FLOAT(pc0, stride_ldc, vl);
+                vb = VLEV_FLOAT(pb, vl);
+                vc0 = VFNMSACVF_FLOAT(vc0, aa0, vb, vl);
+                VSSEV_FLOAT(pc0, stride_ldc, vc0, vl);
+                pb += vl;
+                pc0++;
+            }
+        }
+        b -= n;
+        a -= 2 * m;
+    }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#endif
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+    FLOAT dummy2,
+#endif
+    FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+  size_t vl = VSETVL_MAX;
+
+  //fprintf(stderr, "%s , %s, m = %4ld  n = %4ld  k = %4ld offset = %4ld\n", __FILE__, __FUNCTION__, m, n, k, offset); // Debug
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+        aa  = a;
+        b -= j * k  * COMPSIZE;
+        c -= j * ldc* COMPSIZE;
+        cc  = c;
+
+        i = vl;
+        if (i <= m) {
+
+          do {
+            if (k - kk > 0) {
+              GEMM_KERNEL(vl, j, k - kk, dm1,
+#ifdef COMPLEX
+                  ZERO,
+#endif
+                  aa + vl * kk * COMPSIZE,
+                  b  +  j            * kk * COMPSIZE,
+                  cc,
+                  ldc);
+            }
+
+            solve(vl, j,
+                aa + (kk - j) * vl * COMPSIZE,
+                b  + (kk - j) * j             * COMPSIZE,
+                cc, ldc);
+
+            aa += vl * k * COMPSIZE;
+            cc += vl     * COMPSIZE;
+            i += vl;
+          } while (i <= m);
+        }
+
+        i = m % vl;
+        if (i) {
+          if (k - kk > 0) {
+            GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+                ZERO,
+#endif
+                aa + i * kk * COMPSIZE,
+                b  + j * kk * COMPSIZE,
+                cc, ldc);
+          }
+
+          solve(i, j,
+              aa + (kk - j) * i * COMPSIZE,
+              b  + (kk - j) * j * COMPSIZE,
+              cc, ldc);
+
+          aa += i * k * COMPSIZE;
+          cc += i     * COMPSIZE;
+
+        }
+        kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = vl;
+      if (i <= m) {
+	do {
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(vl, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + vl * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(vl, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * vl * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += vl * k * COMPSIZE;
+	  cc += vl     * COMPSIZE;
+	  i += vl;
+	} while (i <= m);
+      }
+
+      i = m % vl;
+      if (i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+

kernel/riscv64/trsm_lncopy_rvv_v1.c

@@ -0,0 +1,122 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSEV_FLOAT_M vse32_v_f32m2_m
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSLTU_VX_UINT vmsltu_vx_u32m2_b16
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSEV_FLOAT_M vse64_v_f64m2_m
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSLTU_VX_UINT vmsltu_vx_u64m2_b32
+
+#endif
+
+#ifndef UNIT
+#define INV(a) (ONE / (a))
+#else
+#define INV(a) (ONE)
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_lncopy_sve.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
+
+    BLASLONG i, ii, jj, js;
+
+    FLOAT *ao;
+
+    jj = offset;
+
+    BLASLONG stride_lda = sizeof(FLOAT)*lda;
+
+    FLOAT_V_T va1;
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+    size_t vl;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        ao = a;
+
+        ii = 0;
+        for (i = 0; i < m;)
+        {
+            if (ii == jj) 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    vbool_cmp = VMSLTU_VX_UINT(vindex, j, vl);
+                    VSEV_FLOAT_M(vbool_cmp, b, va1, vl);
+
+                    *(b + j) = INV(*(ao + j * lda));
+                    ao++;
+                    b += vl;
+                }
+                i += vl;
+                ii += vl;
+            }
+            else
+            {
+                if (ii > jj)
+                {
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    VSEV_FLOAT(b, va1, vl);
+                }
+                ao++;
+                b += vl;
+                i++;
+                ii++;
+            }
+        }
+
+        a += vl * lda;
+        jj += vl;
+    }
+
+    return 0;
+}

kernel/riscv64/trsm_ltcopy_rvv_v1.c

@@ -0,0 +1,122 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSEV_FLOAT_M vse32_v_f32m2_m
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u32m2_b16
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSEV_FLOAT_M vse64_v_f64m2_m
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u64m2_b32
+#endif
+
+#ifndef UNIT
+#define INV(a) (ONE / (a))
+#else
+#define INV(a) (ONE)
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_ltcopy_sve.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
+
+    BLASLONG i, ii, jj, js;
+
+    FLOAT *ao;
+
+    jj = offset;
+
+    FLOAT_V_T va1;
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    size_t vl;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        ao = a;
+
+        ii = 0;
+        for (i = 0; i < m;)
+        {
+
+            if (ii == jj) 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    *(b + j) = INV(*(ao + j));
+
+                    va1 = VLEV_FLOAT(ao, vl);
+                    vbool_cmp = VMSGTU_VX_UINT(vindex, j, vl);
+                    VSEV_FLOAT_M(vbool_cmp, b, va1, vl);
+
+                    b += vl;
+                    ao += lda;
+                }
+                i += vl;
+                ii += vl;
+            }
+            else 
+            {
+                if (ii < jj) 
+                {
+                    va1 = VLEV_FLOAT(ao, vl);
+                    VSEV_FLOAT(b, va1, vl);
+                }
+                ao += lda;
+                b += vl;
+                i ++;
+                ii ++;
+            }
+        }
+
+        a += vl;
+        jj += vl;
+    }
+    return 0;
+}
+

kernel/riscv64/trsm_uncopy_rvv_v1.c

@@ -0,0 +1,121 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSEV_FLOAT_M vse32_v_f32m2_m
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u32m2_b16
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSEV_FLOAT_M vse64_v_f64m2_m
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSGTU_VX_UINT vmsgtu_vx_u64m2_b32
+#endif
+
+
+#ifndef UNIT
+#define INV(a) (ONE / (a))
+#else
+#define INV(a) (ONE)
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_uncopy_sve.c
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
+
+    BLASLONG i, ii, jj, js;
+    BLASLONG stride_lda = sizeof(FLOAT)*lda;
+
+    FLOAT *ao;
+    jj = offset;
+
+    FLOAT_V_T va1;
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    size_t vl;
+
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        ao = a;
+
+        i = 0;
+        ii = 0;
+        for (i = 0; i < m;)
+        {
+            if (ii == jj) 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    *(b + j) = INV(*(ao + j * lda));
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    vbool_cmp = VMSGTU_VX_UINT(vindex, j, vl);
+                    VSEV_FLOAT_M(vbool_cmp, b, va1, vl);
+                    ao++;
+                    b += vl;
+                }
+                i += vl;
+                ii += vl;
+            } 
+            else
+            {
+                if (ii < jj) 
+                {
+                    va1 = VLSEV_FLOAT(ao, stride_lda, vl);
+                    VSEV_FLOAT(b, va1, vl);
+                }
+                ao++;
+                b += vl;
+                i++;
+                ii++;
+            }
+        } 
+
+        a += vl * lda;
+        jj += vl;
+    }
+    return 0;
+}

kernel/riscv64/trsm_utcopy_rvv_v1.c

@@ -0,0 +1,123 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define FLOAT_V_T vfloat32m2_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VSEV_FLOAT vse32_v_f32m2
+#define VSEV_FLOAT_M vse32_v_f32m2_m
+#define VLSEV_FLOAT vlse32_v_f32m2
+#define VBOOL_T vbool16_t
+#define UINT_V_T vuint32m2_t
+#define VID_V_UINT vid_v_u32m2
+#define VMSLTU_VX_UINT vmsltu_vx_u32m2_b16
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define FLOAT_V_T vfloat64m2_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VSEV_FLOAT vse64_v_f64m2
+#define VSEV_FLOAT_M vse64_v_f64m2_m
+#define VLSEV_FLOAT vlse64_v_f64m2
+#define VBOOL_T     vbool32_t
+#define UINT_V_T     vuint64m2_t
+#define VID_V_UINT   vid_v_u64m2
+#define VMSLTU_VX_UINT vmsltu_vx_u64m2_b32
+#endif
+
+
+#ifndef UNIT
+#define INV(a) (ONE / (a))
+#else
+#define INV(a) (ONE)
+#endif
+
+// Optimizes the implementation in ../arm64/trsm_utcopy_sve.c
+
+
+int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT *b){
+
+    BLASLONG i, ii, jj, js;
+
+    FLOAT *ao;
+
+    jj = offset;
+    FLOAT_V_T va1;
+
+    VBOOL_T vbool_cmp;
+    UINT_V_T vindex;
+
+    size_t vl;
+  
+    for (js = n; js > 0; js -= vl)
+    {
+        vl = VSETVL(js);
+        ao = a;
+
+        ii = 0;
+        for (i = 0; i < m;)
+        {
+
+            if (ii == jj) 
+            {
+                vindex  = VID_V_UINT(vl);
+                for (unsigned int j = 0; j < vl; j++) 
+                {
+                    va1 = VLEV_FLOAT(ao, vl);
+                    vbool_cmp = VMSLTU_VX_UINT(vindex, j, vl);
+                    VSEV_FLOAT_M(vbool_cmp, b, va1, vl);
+                    *(b + j) = INV(*(ao + j));
+
+                    ao += lda;
+                    b += vl;
+                }
+                i += vl;
+                ii += vl;
+            } 
+            else 
+            {
+                if (ii > jj) 
+                {
+                    va1 = VLEV_FLOAT(ao, vl);
+                    VSEV_FLOAT(b, va1, vl);
+                }
+                ao += lda;
+                b += vl;
+                i ++;
+                ii ++;
+            }
+        }
+
+        a += vl;
+        jj += vl;
+    }
+
+    return 0;
+}

kernel/riscv64/zamax_rvv.c

@@ -0,0 +1,113 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m4_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMAXVV_FLOAT vfmax_vv_f32m4
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#define VFABSV_FLOAT vfabs_v_f32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m4_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMAXVV_FLOAT vfmax_vv_f64m4
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#define VFABSV_FLOAT vfabs_v_f64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT maxf=0.0;
+
+    if (n <= 0 || inc_x <= 0) return(maxf);
+
+    FLOAT_V_T v0, v1, vmax;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmax = VFMVVF_FLOAT(0.0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&v0, &v1, x, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v0 = VFADDVV_FLOAT(v0, v1, vl);
+            vmax = VFMAXVV_FLOAT(vmax, v0, vl);
+  
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&v0, &v1, x, stride_x, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v0 = VFADDVV_FLOAT(v0, v1, vl);
+            vmax = VFMAXVV_FLOAT(vmax, v0, vl);
+        }
+
+    }
+
+    v_res = VFREDMAXVS_FLOAT(v_res, vmax, v_res, vlmax);
+    maxf = VFMVFS_FLOAT_M1(v_res);
+
+    return(maxf);
+}

kernel/riscv64/zamin_rvv.c

@@ -0,0 +1,112 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <float.h>
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDMINVS_FLOAT vfredmin_vs_f32m4_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMINVV_FLOAT vfmin_vv_f32m4
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#define VFABSV_FLOAT vfabs_v_f32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDMINVS_FLOAT vfredmin_vs_f64m4_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMINVV_FLOAT vfmin_vv_f64m4
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#define VFABSV_FLOAT vfabs_v_f64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT minf=0.0;
+
+    if (n <= 0 || inc_x <= 0) return(minf);
+
+    FLOAT_V_T v0, v1, vmin;
+    FLOAT_V_T_M1 v_res;
+
+    v_res = VFMVVF_FLOAT_M1(FLT_MAX, VSETVL_MAX_M1);
+    size_t vlmax = VSETVL_MAX;
+    vmin = VFMVVF_FLOAT(FLT_MAX, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&v0, &v1, x, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v0 = VFADDVV_FLOAT(v0, v1, vl);
+            vmin = VFMINVV_FLOAT(vmin, v0, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&v0, &v1, x, stride_x, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v0 = VFADDVV_FLOAT(v0, v1, vl);
+            vmin = VFMINVV_FLOAT(vmin, v0, vl);
+        }
+
+    }
+
+    v_res = VFREDMINVS_FLOAT(v_res, vmin, v_res, vlmax);
+    minf = VFMVFS_FLOAT_M1(v_res);
+
+    return(minf);
+}

kernel/riscv64/zasum_rvv.c

@@ -0,0 +1,108 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m8(n)
+#define VSETVL_MAX vsetvlmax_e32m8()
+#define FLOAT_V_T vfloat32m8_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m8
+#define VLSEV_FLOAT vlse32_v_f32m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m8_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VFADDVV_FLOAT vfadd_vv_f32m8
+#define VFABSV_FLOAT vfabs_v_f32m8
+#else
+#define VSETVL(n) vsetvl_e64m8(n)
+#define VSETVL_MAX vsetvlmax_e64m8()
+#define FLOAT_V_T vfloat64m8_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m8
+#define VLSEV_FLOAT vlse64_v_f64m8
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m8_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m8
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VFADDVV_FLOAT vfadd_vv_f64m8
+#define VFABSV_FLOAT vfabs_v_f64m8
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT asumf = 0.0;
+    if (n <= 0 || inc_x <= 0) return(asumf);
+
+    FLOAT_V_T v0, v1;
+    size_t vlmax = VSETVL_MAX; 
+    FLOAT_V_T v_sum = VFMVVF_FLOAT(0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2) {
+            vl = VSETVL(n);
+
+            v0 = VLEV_FLOAT(x, vl);
+            v1 = VLEV_FLOAT(x+vl, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v_sum = VFADDVV_FLOAT(v_sum, v0, vl);
+            v_sum = VFADDVV_FLOAT(v_sum, v1, vl);
+        }
+
+    }
+    else {
+
+        int stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+
+            v0 = VLSEV_FLOAT(x, stride_x, vl);
+            v1 = VLSEV_FLOAT(x+1, stride_x, vl);
+
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v_sum = VFADDVV_FLOAT(v_sum, v0, vl);
+            v_sum = VFADDVV_FLOAT(v_sum, v1, vl);
+        }
+
+    }
+
+    FLOAT_V_T_M1 v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+    FLOAT_V_T_M1 v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, vlmax);
+    asumf += VFMVFS_FLOAT_M1(v_res);
+
+    return(asumf);
+}

kernel/riscv64/zaxpby_rvv.c

@@ -0,0 +1,151 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+/***************************************************************************
+* 2014/06/07 Saar
+*
+***************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n)       vsetvl_e32m4(n)
+#define FLOAT_V_T       vfloat32m4_t
+#define VLSEV_FLOAT     vlse32_v_f32m4
+#define VSSEV_FLOAT     vsse32_v_f32m4
+#define VFMACCVF_FLOAT  vfmacc_vf_f32m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m4
+#define VFMVVF_FLOAT    vfmv_v_f_f32m4
+#define VFMULVF_FLOAT   vfmul_vf_f32m4
+#define VFMSACVF_FLOAT  vfmsac_vf_f32m4
+#define VLSEG_FLOAT     vlseg2e32_v_f32m4
+#define VSSEG_FLOAT     vsseg2e32_v_f32m4
+#define VLSSEG_FLOAT    vlsseg2e32_v_f32m4
+#define VSSSEG_FLOAT    vssseg2e32_v_f32m4
+#else
+#define VSETVL(n)       vsetvl_e64m4(n)
+#define FLOAT_V_T       vfloat64m4_t
+#define VLSEV_FLOAT     vlse64_v_f64m4
+#define VSSEV_FLOAT     vsse64_v_f64m4
+#define VFMACCVF_FLOAT  vfmacc_vf_f64m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m4
+#define VFMVVF_FLOAT    vfmv_v_f_f64m4
+#define VFMULVF_FLOAT   vfmul_vf_f64m4
+#define VFMSACVF_FLOAT  vfmsac_vf_f64m4
+#define VLSEG_FLOAT     vlseg2e64_v_f64m4
+#define VSSEG_FLOAT     vsseg2e64_v_f64m4
+#define VLSSEG_FLOAT    vlsseg2e64_v_f64m4
+#define VSSSEG_FLOAT    vssseg2e64_v_f64m4
+#endif
+
+int CNAME(BLASLONG n, FLOAT alpha_r, FLOAT alpha_i, FLOAT *x, BLASLONG inc_x, FLOAT beta_r, FLOAT beta_i,FLOAT *y, BLASLONG inc_y)
+{
+    BLASLONG inc_x2, inc_y2;
+
+    if ( n <= 0     )  return(0);
+
+    inc_x2 = 2 * inc_x;
+    inc_y2 = 2 * inc_y;
+    
+    BLASLONG stride_x = inc_x2 * sizeof(FLOAT);
+    BLASLONG stride_y = inc_y2 * sizeof(FLOAT);
+    FLOAT_V_T vx0, vx1, vy0, vy1;
+
+    if ( beta_r == 0.0 && beta_i == 0.0)
+    {
+        if ( alpha_r == 0.0 && alpha_i == 0.0 )
+        {
+            size_t vl = VSETVL(n);
+            FLOAT_V_T temp = VFMVVF_FLOAT(0.0, vl);
+            for ( ; n > 0; n -= vl, y += vl*stride_y)
+            {
+                vl = VSETVL(n);
+                VSSSEG_FLOAT(y, stride_y, temp, temp, vl);
+            }
+        }
+        else
+        {
+            for (size_t vl; n > 0; n -= vl, x += vl*inc_x2, y += vl*inc_y2) 
+            {
+                vl = VSETVL(n);
+                VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+                
+                vy0 = VFMULVF_FLOAT(vx1, alpha_i, vl);
+                vy0 = VFMSACVF_FLOAT(vy0, alpha_r, vx0, vl);
+
+                vy1 = VFMULVF_FLOAT(vx1, alpha_r, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, alpha_i, vx0, vl);
+
+                VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+            }
+        }
+    }
+    else
+    {
+        FLOAT_V_T v0, v1;
+
+        if ( alpha_r == 0.0 && alpha_i == 0.0 )
+        {
+            for (size_t vl; n > 0; n -= vl, y += vl*inc_y2) 
+            {
+                vl = VSETVL(n);
+                VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+                
+                v0 = VFMULVF_FLOAT(vy1, beta_i, vl);
+                v0 = VFMSACVF_FLOAT(v0, beta_r, vy0, vl);
+
+                v1 = VFMULVF_FLOAT(vy1, beta_r, vl);
+                v1 = VFMACCVF_FLOAT(v1, beta_i, vy0, vl);
+
+                VSSSEG_FLOAT(y, stride_y, v0, v1, vl);
+            }
+        }
+        else
+        {
+            for (size_t vl; n > 0; n -= vl, x += vl*inc_x2, y += vl*inc_y2) 
+            {
+                vl = VSETVL(n);
+                VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+                VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+                v0 = VFMULVF_FLOAT(vx0, alpha_r, vl);
+                v0 = VFNMSACVF_FLOAT(v0, alpha_i, vx1, vl);
+                v0 = VFMACCVF_FLOAT(v0, beta_r, vy0, vl);
+                v0 = VFNMSACVF_FLOAT(v0, beta_i, vy1, vl);
+                
+                v1 = VFMULVF_FLOAT(vx1, alpha_r, vl);
+                v1 = VFMACCVF_FLOAT(v1, alpha_i, vx0, vl);
+                v1 = VFMACCVF_FLOAT(v1, beta_r, vy1, vl);
+                v1 = VFMACCVF_FLOAT(v1, beta_i, vy0, vl);
+
+                VSSSEG_FLOAT(y, stride_y, v0, v1, vl);
+            }
+        }
+    }
+    return(0);
+
+}

kernel/riscv64/zaxpy_rvv.c

@@ -0,0 +1,154 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VSSSEG_FLOAT vssseg2e32_v_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VSSSEG_FLOAT vssseg2e64_v_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m4
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+    if(n < 0) return(0);
+    if(da_r == 0.0 && da_i == 0.0) return(0);
+ 
+    FLOAT_V_T vx0, vx1, vy0, vy1;
+
+    if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+        #if !defined(CONJ)
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #else
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFMACCVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #endif
+            VSSEG_FLOAT(y, vy0, vy1, vl);
+        }
+
+    } else if (inc_x == 1) {
+
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+        #if !defined(CONJ)
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #else
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFMACCVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #endif
+            VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+        }
+
+    } else if (inc_y == 1) {
+
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+        #if !defined(CONJ)
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #else
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFMACCVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #endif
+            VSSEG_FLOAT(y, vy0, vy1, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+        #if !defined(CONJ)
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #else
+            vy0 = VFMACCVF_FLOAT(vy0, da_r, vx0, vl);
+            vy0 = VFMACCVF_FLOAT(vy0, da_i, vx1, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, da_r, vx1, vl);
+            vy1 = VFMACCVF_FLOAT(vy1, da_i, vx0, vl);
+        #endif
+            VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+        }
+
+    }
+
+    return(0);
+}

kernel/riscv64/zcopy_rvv.c

@@ -0,0 +1,105 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL_M8(n) vsetvl_e32m8(n)
+#define FLOAT_V_T_M8 vfloat32m8_t
+#define VLEV_FLOAT_M8 vle32_v_f32m8
+#define VSEV_FLOAT_M8 vse32_v_f32m8
+
+#define VSETVL_M4(n) vsetvl_e32m4(n)
+#define FLOAT_V_T_M4 vfloat32m4_t
+#define VLSEG_FLOAT_M4 vlseg2e32_v_f32m4
+#define VSSEG_FLOAT_M4 vsseg2e32_v_f32m4
+#define VLSSEG_FLOAT_M4 vlsseg2e32_v_f32m4
+#define VSSSEG_FLOAT_M4 vssseg2e32_v_f32m4
+#else
+#define VSETVL_M8(n) vsetvl_e64m8(n)
+#define FLOAT_V_T_M8 vfloat64m8_t
+#define VLEV_FLOAT_M8 vle64_v_f64m8
+#define VSEV_FLOAT_M8 vse64_v_f64m8
+
+#define VSETVL_M4(n) vsetvl_e64m4(n)
+#define FLOAT_V_T_M4 vfloat64m4_t
+#define VLSEG_FLOAT_M4 vlseg2e64_v_f64m4
+#define VSSEG_FLOAT_M4 vsseg2e64_v_f64m4
+#define VLSSEG_FLOAT_M4 vlsseg2e64_v_f64m4
+#define VSSSEG_FLOAT_M4 vssseg2e64_v_f64m4
+#endif
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+    if(n < 0) return(0);
+
+    if(inc_x == 1 && inc_y == 1) {
+
+        FLOAT_V_T_M8 vx;
+        n *= 2; // convert to words
+
+        for(size_t vl; n > 0; n -= vl, x += vl, y += vl) {
+            vl = VSETVL_M8(n);
+            vx = VLEV_FLOAT_M8(x, vl);
+            VSEV_FLOAT_M8(y, vx, vl);
+        }
+
+    }else if (1 == inc_x) {
+
+        FLOAT_V_T_M4 vr, vi;
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL_M4(n);
+            VLSEG_FLOAT_M4(&vr, &vi, x, vl);
+            VSSSEG_FLOAT_M4(y, stride_y, vr, vi, vl);
+        }
+    } else if (1 == inc_y) {
+
+        FLOAT_V_T_M4 vr, vi;
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*2) {
+            vl = VSETVL_M4(n);
+            VLSSEG_FLOAT_M4(&vr, &vi, x, stride_x, vl);
+            VSSEG_FLOAT_M4(y, vr, vi, vl);
+        }
+    } else {
+
+        FLOAT_V_T_M4 vr, vi;
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for(size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*inc_y*2) {
+            vl = VSETVL_M4(n);
+            VLSSEG_FLOAT_M4(&vr, &vi, x, stride_x, vl);
+            VSSSEG_FLOAT_M4(y, stride_y, vr, vi, vl);
+        }
+    }
+
+    return(0);
+}

kernel/riscv64/zdot_rvv.c

@@ -0,0 +1,170 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMULVV_FLOAT vfmul_vv_f32m4
+#define VFMSACVV_FLOAT vfmsac_vv_f32m4
+#define VFNMSACVV_FLOAT vfnmsac_vv_f32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMULVV_FLOAT vfmul_vv_f64m4
+#define VFMSACVV_FLOAT vfmsac_vv_f64m4
+#define VFNMSACVV_FLOAT vfnmsac_vv_f64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+    OPENBLAS_COMPLEX_FLOAT result;
+    CREAL(result) = 0.0;
+    CIMAG(result) = 0.0;
+
+    if ( n <= 0 ) return(result);
+
+    FLOAT_V_T vr0, vr1, vx0, vx1, vy0, vy1;
+    FLOAT_V_T_M1 v_res, v_z0;
+    size_t vlmax_m1 = VSETVL_MAX_M1;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax_m1);
+    v_z0 = VFMVVF_FLOAT_M1(0, vlmax_m1);
+
+    size_t vlmax = VSETVL_MAX;
+    vr0 = VFMVVF_FLOAT(0, vlmax);
+    vr1 = VFMVVF_FLOAT(0, vlmax);
+ 
+    if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            vr0 = VFMACCVV_FLOAT(vr0, vx0, vy0, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx0, vy1, vl);
+        #if !defined(CONJ)
+            vr0 = VFNMSACVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx1, vy0, vl);
+        #else
+            vr0 = VFMACCVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFNMSACVV_FLOAT(vr1, vx1, vy0, vl);
+        #endif
+        }
+
+    }  else if (inc_x == 1){
+
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            vr0 = VFMACCVV_FLOAT(vr0, vx0, vy0, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx0, vy1, vl);
+        #if !defined(CONJ)
+            vr0 = VFNMSACVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx1, vy0, vl);
+        #else
+            vr0 = VFMACCVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFNMSACVV_FLOAT(vr1, vx1, vy0, vl);
+        #endif
+        }
+    } else if (inc_y == 1){
+
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            vr0 = VFMACCVV_FLOAT(vr0, vx0, vy0, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx0, vy1, vl);
+        #if !defined(CONJ)
+            vr0 = VFNMSACVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx1, vy0, vl);
+        #else
+            vr0 = VFMACCVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFNMSACVV_FLOAT(vr1, vx1, vy0, vl);
+        #endif
+        }
+    }else {
+
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            vr0 = VFMACCVV_FLOAT(vr0, vx0, vy0, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx0, vy1, vl);
+        #if !defined(CONJ)
+            vr0 = VFNMSACVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFMACCVV_FLOAT(vr1, vx1, vy0, vl);
+        #else
+            vr0 = VFMACCVV_FLOAT(vr0, vx1, vy1, vl);
+            vr1 = VFNMSACVV_FLOAT(vr1, vx1, vy0, vl);
+        #endif
+        }
+    }
+
+    v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, vlmax);
+    CREAL(result) = VFMVFS_FLOAT_M1(v_res);
+    v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, vlmax);
+    CIMAG(result) = VFMVFS_FLOAT_M1(v_res);
+ 
+   return(result);
+}

kernel/riscv64/zgemm_beta_rvv.c

@@ -0,0 +1,117 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMULVF_FLOAT vfmul_vf_f32m4
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#define VFSUBVV_FLOAT vfsub_vv_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMULVF_FLOAT vfmul_vf_f64m4
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#define VFSUBVV_FLOAT vfsub_vv_f64m4
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1,
+          FLOAT beta_r, FLOAT beta_i,
+          FLOAT *dummy2, BLASLONG dummy3,
+          FLOAT *dummy4, BLASLONG dummy5,
+          FLOAT *c, BLASLONG ldc)
+{
+    BLASLONG chunk;
+    FLOAT *c_offset;
+	size_t vl;
+    FLOAT_V_T vr, vi, v1, v2, v3, v4;
+
+    ldc *= 2;
+    c_offset = c;
+
+    if (beta_r == 0.0 && beta_i == 0.0) {
+
+        vl = VSETVL(m);
+        vr = VFMVVF_FLOAT(0.0, vl);
+        vi = VFMVVF_FLOAT(0.0, vl);
+
+        for( ; n > 0; n--, c += ldc) {
+            c_offset = c;
+
+            for(chunk=m; chunk > 0; chunk -= vl, c_offset += vl*2) {
+                vl = VSETVL(chunk);
+
+                VSSEG_FLOAT(c_offset, vr, vi, vl);
+			}
+		}
+
+    } else {
+
+        for( ; n > 0; n--, c += ldc) {
+            c_offset = c;
+
+            for(chunk=m; chunk > 0; chunk -= vl, c_offset += vl*2) {
+                vl = VSETVL(chunk);
+
+                VLSEG_FLOAT(&vr, &vi, c_offset, vl);
+
+                v1 = VFMULVF_FLOAT(vr, beta_r, vl);
+                v2 = VFMULVF_FLOAT(vi, beta_i, vl);
+
+                v3 = VFMULVF_FLOAT(vi, beta_r, vl);
+                v4 = VFMULVF_FLOAT(vr, beta_i, vl);
+
+				vr = VFSUBVV_FLOAT(v1, v2, vl);
+				vi = VFADDVV_FLOAT(v3, v4, vl);
+
+                VSSEG_FLOAT(c_offset, vr, vi, vl);
+			}
+		}
+
+	}
+
+    return 0;
+}

kernel/riscv64/zgemv_n_rvv.c

@@ -0,0 +1,170 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VLSEV_FLOAT vlse32_v_f32m4
+#define VSEV_FLOAT vse32_v_f32m4
+#define VSSEV_FLOAT vsse32_v_f32m4
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VSSSEG_FLOAT vssseg2e32_v_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VLSEV_FLOAT vlse64_v_f64m4
+#define VSEV_FLOAT vse64_v_f64m4
+#define VSSEV_FLOAT vsse64_v_f64m4
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VSSSEG_FLOAT vssseg2e64_v_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m4
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+    BLASLONG i;
+    BLASLONG ix;
+    FLOAT *a_ptr;
+    FLOAT temp_r, temp_i;
+    FLOAT_V_T va0, va1, vy0, vy1;
+
+    BLASLONG stride_y = inc_y * sizeof(FLOAT) * 2;
+
+    BLASLONG inc_x2 = inc_x * 2;
+    BLASLONG lda2 = lda * 2;
+    if (inc_y == 1)
+    {
+        for (size_t vl; m > 0; m -= vl, a += vl*2, y += vl*2) {
+            vl = VSETVL(m);
+            a_ptr = a;
+            ix = 0;
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            for(i = 0; i < n; i++){
+#if !defined(XCONJ)
+                temp_r = alpha_r * x[ix]   - alpha_i * x[ix+1];
+                temp_i = alpha_r * x[ix+1] + alpha_i * x[ix];
+#else
+                temp_r = alpha_r * x[ix]   + alpha_i * x[ix+1];
+                temp_i = alpha_r * x[ix+1] - alpha_i * x[ix];
+#endif
+
+                VLSEG_FLOAT(&va0, &va1, a_ptr, vl);
+#if !defined(CONJ)
+#if !defined(XCONJ)
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFNMSACVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_i, va0, vl);
+#else
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFMACCVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_i, va0, vl);
+#endif
+#else
+#if !defined(XCONJ)
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFMACCVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_i, va0, vl);
+#else
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFNMSACVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_i, va0, vl);
+#endif
+#endif
+                a_ptr += lda2;
+                ix += inc_x2;
+            }
+            VSSEG_FLOAT(y, vy0, vy1, vl);
+        }
+
+    }
+    else
+    {
+        for (size_t vl; m > 0; m -= vl, a += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL(m);
+            a_ptr = a;
+            ix = 0;
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            for(i = 0; i < n; i++){
+#if !defined(XCONJ)
+                temp_r = alpha_r * x[ix]   - alpha_i * x[ix+1];
+                temp_i = alpha_r * x[ix+1] + alpha_i * x[ix];
+#else
+                temp_r = alpha_r * x[ix]   + alpha_i * x[ix+1];
+                temp_i = alpha_r * x[ix+1] - alpha_i * x[ix];
+#endif
+
+                VLSEG_FLOAT(&va0, &va1, a_ptr, vl);
+#if !defined(CONJ)
+#if !defined(XCONJ)
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFNMSACVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_i, va0, vl);
+#else
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFMACCVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_i, va0, vl);
+#endif
+#else
+#if !defined(XCONJ)
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFMACCVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFMACCVF_FLOAT(vy1, temp_i, va0, vl);
+#else
+                vy0 = VFMACCVF_FLOAT(vy0, temp_r, va0, vl);
+                vy0 = VFNMSACVF_FLOAT(vy0, temp_i, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_r, va1, vl);
+                vy1 = VFNMSACVF_FLOAT(vy1, temp_i, va0, vl);
+#endif
+#endif
+                a_ptr += lda2;
+                ix += inc_x2;
+            }
+            VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+        }
+    }
+    return(0);
+}

kernel/riscv64/zgemv_t_rvv.c

@@ -0,0 +1,172 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m4
+#define VFNMSACVV_FLOAT vfnmsac_vv_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMULVV_FLOAT vfmul_vv_f32m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m4
+#define VFNMSACVV_FLOAT vfnmsac_vv_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMULVV_FLOAT vfmul_vv_f64m4
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+    BLASLONG i = 0, j = 0;
+    BLASLONG ix = 0, iy = 0;
+    FLOAT *a_ptr = a;
+    FLOAT temp_r, temp_i;
+
+    FLOAT_V_T va0, va1, vx0, vx1, vr, vi; 
+    FLOAT_V_T_M1 v_res, v_z0;
+
+    BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+    //BLASLONG stride_a = sizeof(FLOAT) * 2;
+    BLASLONG inc_y2 = inc_y * 2;
+    BLASLONG lda2 = lda * 2;
+
+    size_t vlmax = VSETVL_MAX_M1;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+    vlmax = VSETVL(m);
+
+    if (inc_x == 1)
+    {
+        for(i = 0; i < n; i++) {    
+            j = 0;
+            ix = 0;
+            vr = VFMVVF_FLOAT(0, vlmax);
+            vi = VFMVVF_FLOAT(0, vlmax);
+            for(size_t vl, k = m; k > 0; k -= vl) {
+                vl = VSETVL(k);
+
+                VLSEG_FLOAT(&va0, &va1, &a_ptr[j], vl);
+                VLSEG_FLOAT(&vx0, &vx1, &x[ix], vl);
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                vr = VFMACCVV_FLOAT(vr, va0, vx0, vl);
+                vr = VFNMSACVV_FLOAT(vr, va1, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va0, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va1, vx0, vl);
+#else
+                vr = VFMACCVV_FLOAT(vr, va0, vx0, vl);
+                vr = VFMACCVV_FLOAT(vr, va1, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va0, vx1, vl);
+                vi = VFNMSACVV_FLOAT(vi, va1, vx0, vl);
+#endif
+                j += vl * 2;
+                ix += vl * inc_x * 2;
+            }
+            
+            v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+            temp_r = VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUM_FLOAT(v_res, vi, v_z0, vlmax);
+            temp_i = VFMVFS_FLOAT_M1(v_res);
+
+#if !defined(XCONJ)
+            y[iy]   += alpha_r * temp_r - alpha_i * temp_i;
+            y[iy+1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y[iy]   += alpha_r * temp_r + alpha_i * temp_i;
+            y[iy+1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+            iy += inc_y2;
+            a_ptr += lda2;
+        }
+    }
+    else
+    {
+        for(i = 0; i < n; i++) {    
+            j = 0;
+            ix = 0;
+            vr = VFMVVF_FLOAT(0, vlmax);
+            vi = VFMVVF_FLOAT(0, vlmax);
+            for(size_t vl, k = m; k > 0; k -= vl) {
+                vl = VSETVL(k);
+    
+                VLSEG_FLOAT(&va0, &va1, &a_ptr[j], vl);
+                VLSSEG_FLOAT(&vx0, &vx1, &x[ix], stride_x, vl);
+    
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                vr = VFMACCVV_FLOAT(vr, va0, vx0, vl);
+                vr = VFNMSACVV_FLOAT(vr, va1, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va0, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va1, vx0, vl);
+#else
+                vr = VFMACCVV_FLOAT(vr, va0, vx0, vl);
+                vr = VFMACCVV_FLOAT(vr, va1, vx1, vl);
+                vi = VFMACCVV_FLOAT(vi, va0, vx1, vl);
+                vi = VFNMSACVV_FLOAT(vi, va1, vx0, vl);
+#endif
+                j += vl * 2;
+                ix += vl * inc_x * 2;
+            }
+            
+            v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, vlmax);
+            temp_r = VFMVFS_FLOAT_M1(v_res);
+            v_res = VFREDSUM_FLOAT(v_res, vi, v_z0, vlmax);
+            temp_i = VFMVFS_FLOAT_M1(v_res);
+    
+#if !defined(XCONJ)
+            y[iy]   += alpha_r * temp_r - alpha_i * temp_i;
+            y[iy+1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y[iy]   += alpha_r * temp_r + alpha_i * temp_i;
+            y[iy+1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+            iy += inc_y2;
+            a_ptr += lda2;
+        }
+
+    }
+
+
+    return(0);
+}

kernel/riscv64/znrm2_rvv.c

@@ -0,0 +1,122 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMACCVV_FLOAT vfmacc_vv_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFREDMAXVS_FLOAT vfredmax_vs_f32m4_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VFABSV_FLOAT vfabs_v_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMACCVV_FLOAT vfmacc_vv_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFREDMAXVS_FLOAT vfredmax_vs_f64m4_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VFABSV_FLOAT vfabs_v_f64m4
+#endif
+
+// TODO: Should single precision use the widening MAC, or perhaps all should be double? 
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+
+    if ( n <= 0 ) return(0.0);
+
+    FLOAT_V_T vr, v0, v1;
+    FLOAT_V_T_M1 v_max, v_res;
+    FLOAT scale = 0.0, ssq = 0.0;
+
+    size_t vlmax = VSETVL_MAX;
+    v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_max = VFMVVF_FLOAT_M1(0, vlmax);
+
+    vr = VFMVVF_FLOAT(0, vlmax);
+
+    if (inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&v0, &v1, x, vl);
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v0, v_max, vl);
+            vr = VFMACCVV_FLOAT(vr, v0, v0, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v1, v_max, vl);
+            vr = VFMACCVV_FLOAT(vr, v1, v1, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+ 
+            VLSSEG_FLOAT(&v0, &v1, x, stride_x, vl);
+            v0 = VFABSV_FLOAT(v0, vl);
+            v1 = VFABSV_FLOAT(v1, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v0, v_max, vl);
+            vr = VFMACCVV_FLOAT(vr, v0, v0, vl);
+
+            v_max = VFREDMAXVS_FLOAT(v_max, v1, v_max, vl);
+            vr = VFMACCVV_FLOAT(vr, v1, v1, vl);
+        }
+
+    }
+
+    v_res = VFREDSUM_FLOAT(v_res, vr, v_res, vlmax);
+
+    ssq = VFMVFS_FLOAT_M1(v_res);
+    scale = VFMVFS_FLOAT_M1(v_max);
+    ssq = ssq / (scale*scale);
+
+    return(scale * sqrt(ssq));
+}

kernel/riscv64/zrot_rvv.c

@@ -0,0 +1,181 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLEV_FLOAT vle32_v_f32m4
+#define VLSEV_FLOAT vlse32_v_f32m4
+#define VSEV_FLOAT vse32_v_f32m4
+#define VSSEV_FLOAT vsse32_v_f32m4
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VSSSEG_FLOAT vssseg2e32_v_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFMULVF_FLOAT vfmul_vf_f32m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLEV_FLOAT vle64_v_f64m4
+#define VLSEV_FLOAT vlse64_v_f64m4
+#define VSEV_FLOAT vse64_v_f64m4
+#define VSSEV_FLOAT vsse64_v_f64m4
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VSSSEG_FLOAT vssseg2e64_v_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFMULVF_FLOAT vfmul_vf_f64m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m4
+#endif
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+
+    if (n <= 0) return(0);
+
+    FLOAT_V_T vt0, vt1, vx0, vx1, vy0, vy1;
+
+    if (inc_x == 0 && inc_y == 0) {
+        BLASLONG i=0;
+        BLASLONG ix=0,iy=0;
+        FLOAT temp[2];
+        BLASLONG inc_x2;
+        BLASLONG inc_y2;
+
+        inc_x2 = 2 * inc_x ;
+        inc_y2 = 2 * inc_y ;
+
+        while(i < n)
+        {
+            temp[0]   = c*x[ix]   + s*y[iy] ;
+            temp[1]   = c*x[ix+1] + s*y[iy+1] ;
+            y[iy]     = c*y[iy]   - s*x[ix] ;
+            y[iy+1]   = c*y[iy+1] - s*x[ix+1] ;
+            x[ix]     = temp[0] ;
+            x[ix+1]   = temp[1] ;
+
+            ix += inc_x2 ;
+            iy += inc_y2 ;
+            i++ ;
+        }
+    }
+    else if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            vt0 = VFMULVF_FLOAT(vx0, c, vl);
+            vt0 = VFMACCVF_FLOAT(vt0, s, vy0, vl);
+            vt1 = VFMULVF_FLOAT(vx1, c, vl);
+            vt1 = VFMACCVF_FLOAT(vt1, s, vy1, vl);
+            vy0 = VFMULVF_FLOAT(vy0, c, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, s, vx0, vl);
+            vy1 = VFMULVF_FLOAT(vy1, c, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, s, vx1, vl);
+
+            VSSEG_FLOAT(x, vt0, vt1, vl);
+            VSSEG_FLOAT(y, vy0, vy1, vl);
+        }
+
+    } else if (inc_x == 1){
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            vt0 = VFMULVF_FLOAT(vx0, c, vl);
+            vt0 = VFMACCVF_FLOAT(vt0, s, vy0, vl);
+            vt1 = VFMULVF_FLOAT(vx1, c, vl);
+            vt1 = VFMACCVF_FLOAT(vt1, s, vy1, vl);
+            vy0 = VFMULVF_FLOAT(vy0, c, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, s, vx0, vl);
+            vy1 = VFMULVF_FLOAT(vy1, c, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, s, vx1, vl);
+
+            VSSEG_FLOAT(x, vt0, vt1, vl);
+            VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+        }
+
+    } else if (inc_y == 1){
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            vt0 = VFMULVF_FLOAT(vx0, c, vl);
+            vt0 = VFMACCVF_FLOAT(vt0, s, vy0, vl);
+            vt1 = VFMULVF_FLOAT(vx1, c, vl);
+            vt1 = VFMACCVF_FLOAT(vt1, s, vy1, vl);
+            vy0 = VFMULVF_FLOAT(vy0, c, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, s, vx0, vl);
+            vy1 = VFMULVF_FLOAT(vy1, c, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, s, vx1, vl);
+
+            VSSSEG_FLOAT(x, stride_x, vt0, vt1, vl);
+            VSSEG_FLOAT(y, vy0, vy1, vl);
+        }
+
+    } else {
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            vt0 = VFMULVF_FLOAT(vx0, c, vl);
+            vt0 = VFMACCVF_FLOAT(vt0, s, vy0, vl);
+            vt1 = VFMULVF_FLOAT(vx1, c, vl);
+            vt1 = VFMACCVF_FLOAT(vt1, s, vy1, vl);
+            vy0 = VFMULVF_FLOAT(vy0, c, vl);
+            vy0 = VFNMSACVF_FLOAT(vy0, s, vx0, vl);
+            vy1 = VFMULVF_FLOAT(vy1, c, vl);
+            vy1 = VFNMSACVF_FLOAT(vy1, s, vx1, vl);
+
+            VSSSEG_FLOAT(x, stride_x, vt0, vt1, vl);
+            VSSSEG_FLOAT(y, stride_y, vy0, vy1, vl);
+        }
+    }
+
+    return 0;
+}

kernel/riscv64/zscal_rvv.c

@@ -0,0 +1,148 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define FLOAT_V_T vfloat32m4_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VSSSEG_FLOAT vssseg2e32_v_f32m4
+#define VFMACCVF_FLOAT vfmacc_vf_f32m4
+#define VFMULVF_FLOAT vfmul_vf_f32m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f32m4
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define FLOAT_V_T vfloat64m4_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VSSSEG_FLOAT vssseg2e64_v_f64m4
+#define VFMACCVF_FLOAT vfmacc_vf_f64m4
+#define VFMULVF_FLOAT vfmul_vf_f64m4
+#define VFNMSACVF_FLOAT vfnmsac_vf_f64m4
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r,FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+
+    if((n <= 0) || (inc_x <= 0)) return(0);
+
+    FLOAT_V_T vt, vr, vi;
+    BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+    size_t vlmax = VSETVL_MAX;
+
+    if(da_r == 0.0 && da_i == 0.0) {
+
+        vr = VFMVVF_FLOAT(0.0, vlmax);
+        vi = VFMVVF_FLOAT(0.0, vlmax);
+
+        if(inc_x == 1) {
+
+            for (size_t vl; n > 0; n -= vl, x += vl*2) {
+                vl = VSETVL(n);
+
+                VSSEG_FLOAT(x, vr, vi, vl);
+            }
+
+        } else {
+
+            for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+                vl = VSETVL(n);
+
+                VSSSEG_FLOAT(x, stride_x, vr, vi, vl);
+            }
+        }
+
+    } else if(da_r == 0.0) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+            
+            VLSSEG_FLOAT(&vr, &vi, x, stride_x, vl);
+
+            vt = VFMULVF_FLOAT(vi, -da_i, vl);
+            vi = VFMULVF_FLOAT(vr, da_i, vl);
+
+            VSSSEG_FLOAT(x, stride_x, vt, vi, vl);
+        }
+
+    } else if(da_i == 0.0) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vr, &vi, x, stride_x, vl);
+
+            vr = VFMULVF_FLOAT(vr, da_r, vl);
+            vi = VFMULVF_FLOAT(vi, da_r, vl);
+
+            VSSSEG_FLOAT(x, stride_x, vr, vi, vl);
+        }
+
+    } else {
+
+        if(inc_x == 1) {
+
+            for (size_t vl; n > 0; n -= vl, x += vl*2) {
+                vl = VSETVL(n);
+
+                VLSEG_FLOAT(&vr, &vi, x, vl);
+
+                vt = VFMULVF_FLOAT(vr, da_r, vl);
+                vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl);
+                vi = VFMULVF_FLOAT(vi, da_r, vl);
+                vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
+
+                VSSEG_FLOAT(x, vt, vi, vl);
+            }
+
+        } else {
+
+            for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+                vl = VSETVL(n);
+
+                VLSSEG_FLOAT(&vr, &vi, x, stride_x, vl);
+
+                vt = VFMULVF_FLOAT(vr, da_r, vl);
+                vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl);
+                vi = VFMULVF_FLOAT(vi, da_r, vl);
+                vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
+
+                VSSSEG_FLOAT(x, stride_x, vt, vi, vl);
+            }
+        }
+    }
+
+    return(0);
+}

kernel/riscv64/zsum_rvv.c

@@ -0,0 +1,97 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define VSETVL_MAX vsetvlmax_e32m4()
+#define FLOAT_V_T vfloat32m4_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m4_f32m1
+#define VFMVVF_FLOAT vfmv_v_f_f32m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#define VFADDVV_FLOAT vfadd_vv_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define VSETVL_MAX vsetvlmax_e64m4()
+#define FLOAT_V_T vfloat64m4_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m4_f64m1
+#define VFMVVF_FLOAT vfmv_v_f_f64m4
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#define VFADDVV_FLOAT vfadd_vv_f64m4
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    FLOAT sumf = 0.0;
+    if (n <= 0 || inc_x <= 0) return(sumf);
+
+    FLOAT_V_T v0, v1;
+    size_t vlmax = VSETVL_MAX; 
+    FLOAT_V_T v_sum = VFMVVF_FLOAT(0, vlmax);
+
+    if(inc_x == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&v0, &v1, x, vl);
+
+            v_sum = VFADDVV_FLOAT(v_sum, v0, vl);
+            v_sum = VFADDVV_FLOAT(v_sum, v1, vl);
+        }
+
+    } else {
+
+        BLASLONG stride_x = inc_x * sizeof(FLOAT) * 2;
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&v0, &v1, x, stride_x, vl);
+
+            v_sum = VFADDVV_FLOAT(v_sum, v0, vl);
+            v_sum = VFADDVV_FLOAT(v_sum, v1, vl);
+        }
+
+    }
+
+    FLOAT_V_T_M1 v_z0 = VFMVVF_FLOAT_M1(0, vlmax);
+    FLOAT_V_T_M1 v_res = VFMVVF_FLOAT_M1(0, vlmax);
+    v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, vlmax);
+    sumf += VFMVFS_FLOAT_M1(v_res);
+
+    return(sumf);
+}

kernel/riscv64/zswap_rvv.c

@@ -0,0 +1,156 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m4(n)
+#define FLOAT_V_T vfloat32m4_t
+#define VLSEG_FLOAT vlseg2e32_v_f32m4
+#define VLSSEG_FLOAT vlsseg2e32_v_f32m4
+#define VSSEG_FLOAT vsseg2e32_v_f32m4
+#define VSSSEG_FLOAT vssseg2e32_v_f32m4
+#else
+#define VSETVL(n) vsetvl_e64m4(n)
+#define FLOAT_V_T vfloat64m4_t
+#define VLSEG_FLOAT vlseg2e64_v_f64m4
+#define VLSSEG_FLOAT vlsseg2e64_v_f64m4
+#define VSSEG_FLOAT vsseg2e64_v_f64m4
+#define VSSSEG_FLOAT vssseg2e64_v_f64m4
+#endif
+
+int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3, FLOAT dummy4, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
+{
+
+    if (n <= 0) return(0);
+
+    FLOAT_V_T vx0, vx1, vy0, vy1;
+
+    if (inc_x == 0 && inc_y == 0) {
+        if (n & 1) {
+            FLOAT temp[2];
+            temp[0] = x[0];
+            temp[1] = x[1];
+            x[0] = y[0];
+            x[1] = y[1];
+            y[0] = temp[0];
+            y[1] = temp[1];
+        }
+        else {
+            return 0;
+        }
+    }
+    else if(inc_x == 0) {
+        FLOAT temp[2];
+        temp[0] = x[0];
+        temp[1] = x[1];
+        x[0] = y[(n - 1) * inc_y * 2];
+        x[0] = y[(n - 1) * inc_y * 2 + 1];
+        FLOAT* ptr = y + (n - 1) * inc_y * 2;   // start from the last one
+        BLASLONG stride_y = (0 - inc_y) * sizeof(FLOAT) * 2; // reverse
+        BLASLONG m = n - 1;
+        for (size_t vl; m > 0; m -= vl * 2, ptr -= vl*inc_y * 2) {
+            vl = VSETVL(m);
+            VLSSEG_FLOAT(&vy0, &vy1, ptr - 2, stride_y, vl);
+            VSSSEG_FLOAT(ptr, stride_y, vy0, vy1, vl);
+        }
+        y[0] = temp[0];
+        y[1] = temp[1];
+    }
+    else if(inc_y == 0) {
+        FLOAT temp[2];
+        temp[0] = y[0];
+        temp[1] = y[1];
+        y[0] = x[(n - 1) * inc_x * 2];
+        y[0] = x[(n - 1) * inc_x * 2 + 1];
+        FLOAT* ptr = x + (n - 1) * inc_x * 2;   // start from the last one
+        BLASLONG stride_x = (0 - inc_x) * sizeof(FLOAT) * 2; // reverse
+        BLASLONG m = n - 1;
+        for (size_t vl; m > 0; m -= vl * 2, ptr -= vl*inc_x * 2) {
+            vl = VSETVL(m);
+            VLSSEG_FLOAT(&vx0, &vx1, ptr - 2, stride_x, vl);
+            VSSSEG_FLOAT(ptr, stride_x, vx0, vx1, vl);
+        }
+        x[0] = temp[0];
+        x[1] = temp[1];
+    }
+    else if(inc_x == 1 && inc_y == 1) {
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            VSSEG_FLOAT(y, vx0, vx1, vl);
+            VSSEG_FLOAT(x, vy0, vy1, vl);
+        }
+
+    } else if (inc_x == 1){
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSEG_FLOAT(&vx0, &vx1, x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            VSSSEG_FLOAT(y, stride_y, vx0, vx1, vl);
+            VSSEG_FLOAT(x, vy0, vy1, vl);
+        }
+
+    } else if (inc_y == 1){
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSEG_FLOAT(&vy0, &vy1, y, vl);
+
+            VSSEG_FLOAT(y, vx0, vx1, vl);
+            VSSSEG_FLOAT(x, stride_x, vy0, vy1, vl);
+        }
+
+    } else {
+        BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
+        BLASLONG stride_y = inc_y * 2 * sizeof(FLOAT);
+
+        for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2, y += vl*inc_y*2) {
+            vl = VSETVL(n);
+
+            VLSSEG_FLOAT(&vx0, &vx1, x, stride_x, vl);
+            VLSSEG_FLOAT(&vy0, &vy1, y, stride_y, vl);
+
+            VSSSEG_FLOAT(y, stride_y, vx0, vx1, vl);
+            VSSSEG_FLOAT(x, stride_x, vy0, vy1, vl);
+        }
+
+    }
+
+    return(0);
+}

kernel/riscv64/ztrmmkernel_2x2_rvv.c

@@ -0,0 +1,596 @@
+/***************************************************************************
+Copyright (c) 2022, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#if !defined(DOUBLE)
+#define VSETVL(n) vsetvl_e32m2(n)
+#define VSETVL_MAX vsetvlmax_e32m2()
+#define VSETVL_MAX_M1 vsetvlmax_e32m1()
+#define FLOAT_V_T vfloat32m2_t
+#define FLOAT_V_T_M1 vfloat32m1_t
+#define VLEV_FLOAT vle32_v_f32m2
+#define VLSEG4_FLOAT vlseg4e32_v_f32m2
+#define VLSEG2_FLOAT vlseg2e32_v_f32m2
+#define VFMVVF_FLOAT vfmv_v_f_f32m2
+#define VFMACCVF_FLOAT vfmacc_vf_f32m2
+#define VFMACCVV_FLOAT vfmacc_vv_f32m2
+#define VFNMSACVV_FLOAT vfnmsac_vv_f32m2
+#define VFREDSUMVS_FLOAT vfredusum_vs_f32m2_f32m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f32m1_f32
+#else
+#define VSETVL(n) vsetvl_e64m2(n)
+#define VSETVL_MAX vsetvlmax_e64m2()
+#define VSETVL_MAX_M1 vsetvlmax_e64m1()
+#define FLOAT_V_T vfloat64m2_t
+#define FLOAT_V_T_M1 vfloat64m1_t
+#define VLEV_FLOAT vle64_v_f64m2
+#define VLSEG4_FLOAT vlseg4e64_v_f64m2
+#define VLSEG2_FLOAT vlseg2e64_v_f64m2
+#define VFMVVF_FLOAT vfmv_v_f_f64m2
+#define VFMACCVF_FLOAT vfmacc_vf_f64m2
+#define VFMACCVV_FLOAT vfmacc_vv_f64m2
+#define VFNMSACVV_FLOAT vfnmsac_vv_f64m2
+#define VFREDSUMVS_FLOAT vfredusum_vs_f64m2_f64m1
+#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
+#define VFMVFS_FLOAT_M1 vfmv_f_s_f64m1_f64
+#endif
+
+// Optimizes the implementation in ../generic/ztrmmkernel_2x2.c
+
+
+/********************************
+  ADD1 a*c
+  ADD2 b*c
+  ADD3 a*d
+  ADD4 b*d
+ *********************************/
+int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alphar,FLOAT alphai,FLOAT* ba,FLOAT* bb,
+        FLOAT* C,BLASLONG ldc, BLASLONG offset)
+{
+    BLASLONG i,j,k;
+    FLOAT *C0,*C1,*ptrba,*ptrbb;
+    FLOAT res0,res1;
+    BLASLONG off, temp;
+    
+    FLOAT_V_T va0, va1, va2, va3, vb0, vb1, vb2, vb3;
+    FLOAT_V_T vres0, vres1, vres2, vres3, vres4, vres5, vres6, vres7;
+    FLOAT_V_T_M1 v_m1_res0, v_m1_res1;
+    FLOAT_V_T_M1 v_z0 = VFMVVF_FLOAT_M1(0, VSETVL_MAX_M1);
+
+    size_t vl;
+    size_t vlmax = VSETVL_MAX;
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+    off = -offset;
+#else
+    off = 0;
+#endif
+
+    for (j = bn/2; j > 0; j--)
+    {
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+        off = offset;
+#endif
+        C0 = C;
+        C1 = C0+2*ldc;
+        ptrba = ba;
+
+        for (i = bm/2; i > 0; i--)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2*2;
+            ptrbb = bb+off*2*2;
+#endif
+
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+            vres4 = VFMVVF_FLOAT(0.0, vlmax);
+            vres5 = VFMVVF_FLOAT(0.0, vlmax);
+            vres6 = VFMVVF_FLOAT(0.0, vlmax);
+            vres7 = VFMVVF_FLOAT(0.0, vlmax);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk - off;
+#elif defined(LEFT)
+            temp = off + 2;
+#else
+            temp = off + 2;
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
+
+#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFNMSACVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va2, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
+                vres4 = VFNMSACVV_FLOAT(vres4, va1, vb3, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va0, vb3, vl);
+
+                vres6 = VFMACCVV_FLOAT(vres6, va2, vb2, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va3, vb2, vl);
+                vres6 = VFNMSACVV_FLOAT(vres6, va3, vb3, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va2, vb3, vl);
+#endif
+#if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va2, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va1, vb2, vl);
+                vres4 = VFMACCVV_FLOAT(vres4, va1, vb3, vl);
+                vres5 = VFNMSACVV_FLOAT(vres5, va0, vb3, vl);
+
+                vres6 = VFMACCVV_FLOAT(vres6, va2, vb2, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va3, vb2, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va3, vb3, vl);
+                vres7 = VFNMSACVV_FLOAT(vres7, va2, vb3, vl);
+#endif
+#if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va2, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFNMSACVV_FLOAT(vres5, va1, vb2, vl);
+                vres4 = VFMACCVV_FLOAT(vres4, va1, vb3, vl);
+                vres5 = VFMACCVV_FLOAT(vres5, va0, vb3, vl);
+
+                vres6 = VFMACCVV_FLOAT(vres6, va2, vb2, vl);
+                vres7 = VFNMSACVV_FLOAT(vres7, va3, vb2, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va3, vb3, vl);
+                vres7 = VFMACCVV_FLOAT(vres7, va2, vb3, vl);
+#endif
+#if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va2, vb1, vl);
+
+                vres4 = VFMACCVV_FLOAT(vres4, va0, vb2, vl);
+                vres5 = VFNMSACVV_FLOAT(vres5, va1, vb2, vl);
+                vres4 = VFMACCVV_FLOAT(vres4, va1, vb3, vl);
+                vres5 = VFNMSACVV_FLOAT(vres5, va0, vb3, vl);
+
+                vres6 = VFMACCVV_FLOAT(vres6, va2, vb2, vl);
+                vres7 = VFNMSACVV_FLOAT(vres7, va3, vb2, vl);
+                vres6 = VFMACCVV_FLOAT(vres6, va3, vb3, vl);
+                vres7 = VFNMSACVV_FLOAT(vres7, va2, vb3, vl);
+
+#endif
+                ptrba += vl * 4;
+                ptrbb += vl * 4;
+            }
+            
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres0, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres1, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C0[0] = res0 * alphar - res1 * alphai;
+            C0[1] = res1 * alphar + res0 * alphai;
+
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres2, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres3, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C0[2] = res0 * alphar - res1 * alphai;
+            C0[3] = res1 * alphar + res0 * alphai;
+
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres4, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres5, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C1[0] = res0 * alphar - res1 * alphai;
+            C1[1] = res1 * alphar + res0 * alphai;
+
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres6, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres7, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C1[2] = res0 * alphar - res1 * alphai;
+            C1[3] = res1 * alphar + res0 * alphai;
+#if ( defined(LEFT) &&  defined(TRANSA)) || \
+            (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 2;
+#else
+            temp -= 2;
+#endif
+
+            ptrba += temp*2*2;
+            ptrbb += temp*2*2;
+
+#endif
+
+#ifdef LEFT
+            off += 2;
+#endif
+
+            C0 = C0+4;
+            C1 = C1+4;
+        }
+
+        if (bm & 1)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2;
+            ptrbb = bb + off*2*2;
+#endif
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk - off;
+#elif defined(LEFT)
+            temp = off+1;
+#else
+            temp = off+2;
+#endif
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG4_FLOAT(&vb0, &vb1, &vb2, &vb3, ptrbb, vl);
+
+#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb2, vl);
+                vres2 = VFNMSACVV_FLOAT(vres2, va1, vb3, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va0, vb3, vl);
+#endif
+#if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va1, vb2, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va1, vb3, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va0, vb3, vl);
+
+#endif
+#if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va1, vb2, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va1, vb3, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va0, vb3, vl);
+
+#endif
+#if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
+
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va0, vb2, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va1, vb2, vl);
+                vres2 = VFNMSACVV_FLOAT(vres2, va1, vb3, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va0, vb3, vl);
+
+#endif
+                ptrba += vl * 2;
+                ptrbb += vl * 4;
+            }
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres0, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres1, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C0[0] = res0 * alphar - res1 * alphai;
+            C0[1] = res1 * alphar + res0 * alphai;
+
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres2, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres3, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C1[0] = res0 * alphar - res1 * alphai;
+            C1[1] = res1 * alphar + res0 * alphai;
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || \
+            (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 1;
+#else
+            temp -= 2;
+#endif
+            ptrba += temp*2;
+            ptrbb += temp*2*2;
+#endif
+#ifdef LEFT
+            off += 1;
+#endif
+            C0 = C0+2;
+            C1 = C1+2;
+        }
+#if defined(TRMMKERNEL) && !defined(LEFT)
+        off += 2;
+#endif
+        k = (bk<<2);
+        bb = bb+k;
+        i = (ldc<<2);
+        C = C+i;
+    }
+
+    if (bn & 1)
+    {
+        C0 = C;
+#if defined(TRMMKERNEL) &&  defined(LEFT)
+        off = offset;
+#endif
+        ptrba = ba;
+
+        for (i = bm/2; i > 0; i--)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2*2;
+            ptrbb = bb+off*2;
+#endif
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+            vres2 = VFMVVF_FLOAT(0.0, vlmax);
+            vres3 = VFMVVF_FLOAT(0.0, vlmax);
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk - off;
+#elif defined(LEFT)
+            temp = off + 2;
+#else
+            temp = off + 1;
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG4_FLOAT(&va0, &va1, &va2, &va3, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFNMSACVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va2, vb1, vl);
+#endif
+#if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va2, vb1, vl);
+#endif
+#if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFMACCVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFMACCVV_FLOAT(vres3, va2, vb1, vl);
+
+#endif
+#if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+                vres2 = VFMACCVV_FLOAT(vres2, va2, vb0, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va3, vb0, vl);
+                vres2 = VFNMSACVV_FLOAT(vres2, va3, vb1, vl);
+                vres3 = VFNMSACVV_FLOAT(vres3, va2, vb1, vl);
+
+#endif
+                ptrba += vl * 4;
+                ptrbb += vl * 2;
+            }
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres0, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres1, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C0[0] = res0 * alphar - res1 * alphai;
+            C0[1] = res1 * alphar + res0 * alphai;
+
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres2, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres3, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            C0[2] = res0 * alphar - res1 * alphai;
+            C0[3] = res1 * alphar + res0 * alphai;
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || \
+            (!defined(LEFT) && !defined(TRANSA))
+            temp = bk-off;
+#ifdef LEFT
+            temp -= 2;
+#else
+            temp -= 1;
+#endif
+            ptrba += temp*2*2;
+            ptrbb += temp*2;
+#endif
+#ifdef LEFT
+            off += 2;
+#endif
+            C0 = C0+4;
+        }
+
+        if (bm & 1)
+        {
+#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
+            ptrbb = bb;
+#else
+            ptrba += off*2;
+            ptrbb = bb + off*2;
+#endif
+            vres0 = VFMVVF_FLOAT(0.0, vlmax);
+            vres1 = VFMVVF_FLOAT(0.0, vlmax);
+
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+            temp = bk-off;
+#elif defined(LEFT)
+            temp = off + 1;
+#else
+            temp = off + 1;
+#endif
+
+            for (k = temp; k > 0; k -= vl)
+            {
+                vl = VSETVL(k);
+                VLSEG2_FLOAT(&va0, &va1, ptrba, vl);
+                VLSEG2_FLOAT(&vb0, &vb1, ptrbb, vl);
+
+#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+#endif
+#if   defined(NR) || defined(NC) || defined(TR) || defined(TC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+#endif
+#if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFMACCVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFMACCVV_FLOAT(vres1, va0, vb1, vl);
+
+#endif
+#if   defined(RR) || defined(RC) || defined(CR) || defined(CC)
+                vres0 = VFMACCVV_FLOAT(vres0, va0, vb0, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va1, vb0, vl);
+                vres0 = VFNMSACVV_FLOAT(vres0, va1, vb1, vl);
+                vres1 = VFNMSACVV_FLOAT(vres1, va0, vb1, vl);
+
+#endif
+                ptrba += vl * 2;
+                ptrbb += vl * 2;
+                
+            }
+            
+            v_m1_res0 = VFREDSUMVS_FLOAT(v_m1_res0, vres0, v_z0, vlmax);
+            v_m1_res1 = VFREDSUMVS_FLOAT(v_m1_res1, vres1, v_z0, vlmax);
+            res0 = VFMVFS_FLOAT_M1(v_m1_res0);
+            res1 = VFMVFS_FLOAT_M1(v_m1_res1);
+            
+            C0[0] = res0 * alphar - res1 * alphai;
+            C0[1] = res1 * alphar + res0 * alphai;
+
+#if ( defined(LEFT) &&  defined(TRANSA)) || \
+            (!defined(LEFT) && !defined(TRANSA))
+            temp = bk - off;
+#ifdef LEFT
+            temp -= 1;
+#else
+            temp -= 1;
+#endif
+            ptrba += temp*2;
+            ptrbb += temp*2;
+            
+#endif
+#ifdef LEFT
+            off += 1;
+#endif
+            C0 = C0+2;
+        }
+        k = (bk<<1);
+        bb = bb+k;
+        i = (ldc<<1);
+        C = C+i;
+    }
+    return 0;
+}

param.h

@@ -3038,6 +3038,50 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #endif
 
+#if defined(x280)
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+#define GEMM_DEFAULT_ALIGN 0x03fffUL
+
+#define SGEMM_DEFAULT_UNROLL_M  16 // 4 // 16 // 2
+#define SGEMM_DEFAULT_UNROLL_N  8// 4 // 4 // 2
+
+/* SGEMM_UNROLL_MN is calculated as max(SGEMM_UNROLL_M, SGEMM_UNROLL_N)
+ * Since we don't define SGEMM_UNROLL_M correctly we have to manually set this macro.
+ * If VLMAX size is ever more than 1024, this should be increased also. */
+#define SGEMM_DEFAULT_UNROLL_MN  32
+
+#define DGEMM_DEFAULT_UNROLL_M  16 //2 // 8
+#define DGEMM_DEFAULT_UNROLL_N  8 //2 // 4
+#define DGEMM_DEFAULT_UNROLL_MN  32
+
+#define CGEMM_DEFAULT_UNROLL_M  2
+#define CGEMM_DEFAULT_UNROLL_N  2
+
+#define ZGEMM_DEFAULT_UNROLL_M  2
+#define ZGEMM_DEFAULT_UNROLL_N  2
+
+#define SGEMM_DEFAULT_P	160
+#define DGEMM_DEFAULT_P	160
+#define CGEMM_DEFAULT_P 96
+#define ZGEMM_DEFAULT_P 64
+
+#define SGEMM_DEFAULT_Q 240
+#define DGEMM_DEFAULT_Q 128
+#define CGEMM_DEFAULT_Q 120
+#define ZGEMM_DEFAULT_Q 120
+
+#define SGEMM_DEFAULT_R 12288
+#define DGEMM_DEFAULT_R 8192
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
+#define SYMV_P	16
+
+#define GEMM_DEFAULT_OFFSET_A 0
+#define GEMM_DEFAULT_OFFSET_B 0
+
+#endif
 #ifdef C910V
 #define GEMM_DEFAULT_OFFSET_A 0
 #define GEMM_DEFAULT_OFFSET_B 0