OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

[POWER] Add support for SMALL_MATRIX_OPT

tags/v0.3.19
Rafael Cardoso Fernandes Sousa 3 years ago
parent
454edd741c
commit
c78fdcc80d
11 changed files with 9008 additions and 0 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +2
    -0
      Makefile.system
  2. +20
    -0
      kernel/power/KERNEL.POWER10
  3. +923
    -0
      kernel/power/dgemm_small_kernel_nn_power10.c
  4. +581
    -0
      kernel/power/dgemm_small_kernel_nt_power10.c
  5. +882
    -0
      kernel/power/dgemm_small_kernel_tn_power10.c
  6. +829
    -0
      kernel/power/dgemm_small_kernel_tt_power10.c
  7. +84
    -0
      kernel/power/gemm_small_kernel_permit_power10.c
  8. +1563
    -0
      kernel/power/sgemm_small_kernel_nn_power10.c
  9. +887
    -0
      kernel/power/sgemm_small_kernel_nt_power10.c
  10. +1678
    -0
      kernel/power/sgemm_small_kernel_tn_power10.c
  11. +1559
    -0
      kernel/power/sgemm_small_kernel_tt_power10.c

+ 2
- 0
Makefile.system View File

@@ -258,6 +258,8 @@ endif
#For small matrix optimization #For small matrix optimization
ifeq ($(ARCH), x86_64) ifeq ($(ARCH), x86_64)
SMALL_MATRIX_OPT = 1 SMALL_MATRIX_OPT = 1
else ifeq ($(CORE), POWER10)
SMALL_MATRIX_OPT = 1
endif endif
ifeq ($(SMALL_MATRIX_OPT), 1) ifeq ($(SMALL_MATRIX_OPT), 1)
CCOMMON_OPT += -DSMALL_MATRIX_OPT CCOMMON_OPT += -DSMALL_MATRIX_OPT


+ 20
- 0
kernel/power/KERNEL.POWER10 View File

@@ -32,6 +32,16 @@ SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX) SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX) SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)


SGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_power10.c
SGEMM_SMALL_K_NN = sgemm_small_kernel_nn_power10.c
SGEMM_SMALL_K_B0_NN = sgemm_small_kernel_nn_power10.c
SGEMM_SMALL_K_NT = sgemm_small_kernel_nt_power10.c
SGEMM_SMALL_K_B0_NT = sgemm_small_kernel_nt_power10.c
SGEMM_SMALL_K_TN = sgemm_small_kernel_tn_power10.c
SGEMM_SMALL_K_B0_TN = sgemm_small_kernel_tn_power10.c
SGEMM_SMALL_K_TT = sgemm_small_kernel_tt_power10.c
SGEMM_SMALL_K_B0_TT = sgemm_small_kernel_tt_power10.c

DGEMMKERNEL = dgemm_kernel_power10.c DGEMMKERNEL = dgemm_kernel_power10.c
DGEMMINCOPY = DGEMMINCOPY =
DGEMMITCOPY = DGEMMITCOPY =
@@ -42,6 +52,16 @@ DGEMMITCOPYOBJ =
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)


DGEMM_SMALL_M_PERMIT = gemm_small_kernel_permit_power10.c
DGEMM_SMALL_K_NT = dgemm_small_kernel_nt_power10.c
DGEMM_SMALL_K_B0_NT = dgemm_small_kernel_nt_power10.c
DGEMM_SMALL_K_NN = dgemm_small_kernel_nn_power10.c
DGEMM_SMALL_K_B0_NN = dgemm_small_kernel_nn_power10.c
DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_power10.c
DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_power10.c
DGEMM_SMALL_K_TN = dgemm_small_kernel_tn_power10.c
DGEMM_SMALL_K_B0_TN = dgemm_small_kernel_tn_power10.c

CGEMMKERNEL = cgemm_kernel_power10.S CGEMMKERNEL = cgemm_kernel_power10.S
#CGEMMKERNEL = cgemm_kernel_8x4_power8.S #CGEMMKERNEL = cgemm_kernel_8x4_power8.S
CGEMMINCOPY = ../generic/zgemm_ncopy_8.c CGEMMINCOPY = ../generic/zgemm_ncopy_8.c


+ 923
- 0
kernel/power/dgemm_small_kernel_nn_power10.c View File

@@ -0,0 +1,923 @@
/***************************************************************************
Copyright (c) 2021, 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 <altivec.h>

typedef __vector unsigned char vec_t;

#if !__has_builtin(__builtin_vsx_assemble_pair)
#define __builtin_vsx_assemble_pair __builtin_mma_assemble_pair
#endif

#if !defined(B0)
#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
rc0 = vec_xl(0, C+(N+2)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
rc0 = vec_xl(0, C+(N+3)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl_len(C+(N+0)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+1)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+2)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+3)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_2x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl_len(C+(N+0)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+1)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8);

#define SAVE_1x4_VSR(result, N, M) \
rc0 = vec_xl(0, C+((N)*ldc)+M); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst(result, 0, C+((N)*ldc)+M);

#else

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
result[2] = vec_mul(result[2], valpha); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
result[3] = vec_mul(result[3], valpha); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
result[1] = vec_mul(result[1], valpha); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
result[2] = vec_mul(result[2], valpha); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
result[3] = vec_mul(result[3], valpha); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_2x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
result[1] = vec_mul(result[1], valpha); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8);

#define SAVE_1x4_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst(result, 0, C+((N)*ldc)+M);

#endif

#define INIT_8ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3); \
__builtin_mma_xxsetaccz(&acc4); \
__builtin_mma_xxsetaccz(&acc5); \
__builtin_mma_xxsetaccz(&acc6); \
__builtin_mma_xxsetaccz(&acc7);

#define INIT_4ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3);

#define INIT_2ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1);

#define INIT_1ACC() __builtin_mma_xxsetaccz(&acc0);

#if (defined(__GNUC__) && (__GNUC__ == 10))
#if defined(_AIX)
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_assemble_pair(&pair, (vec_t)v0, (vec_t)v1);
#else
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_assemble_pair(&pair, (vec_t)v1, (vec_t)v0);
#endif
#else
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_build_pair(&pair, (vec_t)v0, (vec_t)v1);
#endif

#define LOAD_A_1x8(K, M) \
ra0 = vec_xl(0, A+((K)*lda)+M+0); \
ra1 = vec_xl(0, A+((K)*lda)+M+2); \
ra2 = vec_xl(0, A+((K)*lda)+M+4); \
ra3 = vec_xl(0, A+((K)*lda)+M+6);

#define LOAD_A_1x4(K, M) \
ra0 = vec_xl(0, A+((K)*lda)+M+0); \
ra1 = vec_xl(0, A+((K)*lda)+M+2); \

#define LOAD_A_1x2(K, M) \
ra0 = vec_xl(0, A+((K)*lda)+M+0);

#define LOAD_A_1x1(K, M) \
ra0 = vec_splats(A[((K)*lda)+M+0]);

#define LOAD_BTP_8x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
rb2 = vec_xl(0, B+(N+2)*ldb+K+0); \
rb3 = vec_xl(0, B+(N+3)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
t1 = vec_mergeh(rb2, rb3); \
LOAD_PAIR(pb0, t0, t1); \
t0 = vec_mergel(rb0, rb1); \
t1 = vec_mergel(rb2, rb3); \
LOAD_PAIR(pb2, t0, t1); \
rb4 = vec_xl(0, B+(N+4)*ldb+K+0); \
rb5 = vec_xl(0, B+(N+5)*ldb+K+0); \
rb6 = vec_xl(0, B+(N+6)*ldb+K+0); \
rb7 = vec_xl(0, B+(N+7)*ldb+K+0); \
t0 = vec_mergeh(rb4, rb5); \
t1 = vec_mergeh(rb6, rb7); \
LOAD_PAIR(pb1, t0, t1); \
t0 = vec_mergel(rb4, rb5); \
t1 = vec_mergel(rb6, rb7); \
LOAD_PAIR(pb3, t0, t1);

#define LOAD_BTP_8x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
rb1 = vec_xor(rb1, rb1); \
rb1 = vec_insert(B[(N+2)*ldb+K], rb1, 0); \
rb1 = vec_insert(B[(N+3)*ldb+K], rb1, 1); \
LOAD_PAIR(pb0, rb0, rb1); \
rb2 = vec_xor(rb2, rb2); \
rb2 = vec_insert(B[(N+4)*ldb+K], rb2, 0); \
rb2 = vec_insert(B[(N+5)*ldb+K], rb2, 1); \
rb3 = vec_xor(rb3, rb3); \
rb3 = vec_insert(B[(N+6)*ldb+K], rb3, 0); \
rb3 = vec_insert(B[(N+7)*ldb+K], rb3, 1); \
LOAD_PAIR(pb1, rb2, rb3);

#define LOAD_BTP_4x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
rb2 = vec_xl(0, B+(N+2)*ldb+K+0); \
rb3 = vec_xl(0, B+(N+3)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
t1 = vec_mergeh(rb2, rb3); \
LOAD_PAIR(pb0, t0, t1); \
t0 = vec_mergel(rb0, rb1); \
t1 = vec_mergel(rb2, rb3); \
LOAD_PAIR(pb1, t0, t1);

#define LOAD_BTP_4x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
rb1 = vec_xor(rb1, rb1); \
rb1 = vec_insert(B[(N+2)*ldb+K], rb1, 0); \
rb1 = vec_insert(B[(N+3)*ldb+K], rb1, 1); \
LOAD_PAIR(pb0, rb0, rb1);

#define LOAD_BTP_2x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)t0, (vec_t)t0); \
t1 = vec_mergel(rb0, rb1); \
__builtin_vsx_assemble_pair(&pb1, (vec_t)t1, (vec_t)t1);

#define LOAD_BTP_2x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)rb0, (vec_t)rb0);

#define LOAD_B_1x1(N, K) \
rb0 = vec_splats(B[((N)*ldb)+K]);

#define KERNEL_MMA_8ACC(b0, b1, b2, b3, b4, b5, b6, b7, \
a0, a1, a2, a3, a4, a5, a6, a7) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3); \
__builtin_mma_xvf64gerpp(&acc4, b4, (vec_t)a4); \
__builtin_mma_xvf64gerpp(&acc5, b5, (vec_t)a5); \
__builtin_mma_xvf64gerpp(&acc6, b6, (vec_t)a6); \
__builtin_mma_xvf64gerpp(&acc7, b7, (vec_t)a7);

#define KERNEL_MMA_4ACC(b0, b1, b2, b3, a0, a1, a2, a3) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3);

#define KERNEL_MMA_2ACC(b0, b1, a0, a1) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1);

#define KERNEL_MMA_1ACC(b0, a0) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0);

#define KERNEL_VMADD_4VSR(a0, a1, a2, a3, b0, b1, b2, b3) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1); \
result2 = vec_madd(a2, b2, result2); \
result3 = vec_madd(a3, b3, result3);

#define KERNEL_VMADD_2VSR(a0, a1, b0, b1) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1);

#define KERNEL_VMADD_1VSR(a0, b0) \
result = vec_madd(a0, b0, result);

#define PACK_B(pb0, pb1, offset) \
*((__vector_pair *)(void *)(packB+(k*8)+0+offset)) = pb0; \
*((__vector_pair *)(void *)(packB+(k*8)+4+offset)) = pb1;

#define LOAD_PACKED_B(pb0, pb1, offset) \
pb0 = *((__vector_pair *)((void *)(packB+(k*8)+0+offset))); \
pb1 = *((__vector_pair *)((void *)(packB+(k*8)+4+offset)));

#ifdef B0
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
#else
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)
#endif
{
BLASLONG m, n, k;

BLASLONG m8 = M & ~7;
BLASLONG m4 = M & ~3;
BLASLONG m2 = M & ~1;

BLASLONG n8 = N & ~7;
BLASLONG n4 = N & ~3;
BLASLONG n2 = N & ~1;

BLASLONG k2 = K & ~1;

#if defined(__GNUC__) && !defined(__clang__)
int has_packing = (M >= 32 && N >= 32 && K >= 32) ? 1 : 0;
#else
int has_packing = 0;
#endif

double *packB;
if (has_packing) packB = (double *)malloc(K*8*sizeof(double));

vector double valpha = vec_splats(alpha);
#if !defined(B0)
vector double vbeta = vec_splats(beta);
#endif

for (n = 0; n < n8; n += 8) {
for (m = 0; m < m8; m += 8) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

if (has_packing) {
if (m == 0) {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x8(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
PACK_B(pb0, pb1, 0);
LOAD_A_1x8(k+1, m);
KERNEL_MMA_8ACC(pb2, pb3, pb2, pb3, pb2, pb3, pb2, pb3,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
PACK_B(pb2, pb3, 8);
}
for (; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
PACK_B(pb0, pb1, 0);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x8(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
LOAD_A_1x8(k+1, m);
LOAD_PACKED_B(pb2, pb3, 8);
KERNEL_MMA_8ACC(pb2, pb3, pb2, pb3, pb2, pb3, pb2, pb3,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
for (; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x8(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
LOAD_A_1x8(k+1, m);
KERNEL_MMA_8ACC(pb2, pb3, pb2, pb3, pb2, pb3, pb2, pb3,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
for (; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc4, n+0, m+4);
SAVE_4x2_ACC(&acc6, n+0, m+6);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
SAVE_4x2_ACC(&acc5, n+4, m+4);
SAVE_4x2_ACC(&acc7, n+4, m+6);
}

for (; m < m4; m += 4) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

if (!has_packing) {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x4(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
LOAD_A_1x4(k+1, m);
KERNEL_MMA_4ACC(pb2, pb3, pb2, pb3, ra0, ra0, ra1, ra1);
}
for (; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x4(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
LOAD_A_1x4(k+1, m);
LOAD_PACKED_B(pb2, pb3, 8);
KERNEL_MMA_4ACC(pb2, pb3, pb2, pb3, ra0, ra0, ra1, ra1);
}
for (; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
}

for (; m < m2; m += 2) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

if (!has_packing) {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x2(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_A_1x2(k+1, m);
KERNEL_MMA_2ACC(pb2, pb3, ra0, ra0);
}
for (; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x2(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_A_1x2(k+1, m);
LOAD_PACKED_B(pb2, pb3, 8);
KERNEL_MMA_2ACC(pb2, pb3, ra0, ra0);
}
for (; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
}

for (; m < M; m++) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

if (!has_packing) {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_2ACC(pb2, pb3, ra0, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_A_1x1(k+1, m);
LOAD_PACKED_B(pb2, pb3, 8);
KERNEL_MMA_2ACC(pb2, pb3, ra0, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_PACKED_B(pb0, pb1, 0);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x1_ACC(&acc0, n+0, m+0);
SAVE_4x1_ACC(&acc1, n+4, m+0);
}
}

for (; n < n4; n += 4) {
for (m = 0; m < m8; m += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x8(k, m);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
LOAD_A_1x8(k+1, m);
KERNEL_MMA_4ACC(pb1, pb1, pb1, pb1, ra0, ra1, ra2, ra3);
}
for (; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc2, n+0, m+4);
SAVE_4x2_ACC(&acc3, n+0, m+6);
}

for (; m < m4; m += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x4(k, m);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
LOAD_A_1x4(k+1, m);
KERNEL_MMA_2ACC(pb1, pb1, ra0, ra1);
}
for (; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
}

for (; m < m2; m += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x2(k, m);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x2(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n, m);
}

for (; m < M; m++) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x1_ACC(&acc0, n, m);
}
}

for (; n < n2; n += 2) {
for (m = 0; m < m8; m += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x8(k, m);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
LOAD_A_1x8(k+1, m);
KERNEL_MMA_4ACC(pb1, pb1, pb1, pb1, ra0, ra1, ra2, ra3);
}
for (; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
SAVE_2x2_ACC(&acc2, n+0, m+4);
SAVE_2x2_ACC(&acc3, n+0, m+6);
}

for (; m < m4; m += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x4(k, m);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
LOAD_A_1x4(k+1, m);
KERNEL_MMA_2ACC(pb1, pb1, ra0, ra1);
}
for (; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
}

for (; m < m2; m += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x2(k, m);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x2(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
}

for (; m < M; m++) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x1_ACC(&acc0, n+0, m+0);
}
}

for (; n < N; n++) {
for (m = 0; m < m8; m += 8) {
vector double result = ((vector double){0.,0.});
vector double result1 = ((vector double){0.,0.});
vector double result2 = ((vector double){0.,0.});
vector double result3 = ((vector double){0.,0.});

register vector double ra0, ra1, ra2, ra3;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x1(n, k);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
SAVE_1x4_VSR(result2, n, m+4);
SAVE_1x4_VSR(result3, n, m+6);
}

for (; m < m4; m += 4) {
vector double result = ((vector double){0.,0.});
vector double result1 = ((vector double){0.,0.});

register vector double ra0, ra1;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x1(n, k);
KERNEL_VMADD_2VSR(ra0, ra1, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
}

for (; m < m2; m += 2) {
vector double result = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x1(n, k);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
}

for (; m < M; m++) {
FLOAT result = 0.0;

for (k = 0; k < K; k++) {
result += A[m+k*lda] * B[n*ldb+k];
}
result = result * alpha;

#if !defined(B0)
C[n*ldc+m] = (C[n*ldc+m] * beta) + result;
#else
C[n*ldc+m] = result;
#endif
}
}

if (has_packing) free(packB);

return 0;
}

+ 581
- 0
kernel/power/dgemm_small_kernel_nt_power10.c View File

@@ -0,0 +1,581 @@
/***************************************************************************
Copyright (c) 2021, 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 <altivec.h>

typedef __vector unsigned char vec_t;

#if !__has_builtin(__builtin_vsx_assemble_pair)
#define __builtin_vsx_assemble_pair __builtin_mma_assemble_pair
#endif

#if !defined(B0)
#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
rc0 = vec_xl(0, C+(N+2)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
rc0 = vec_xl(0, C+(N+3)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
rc0 = vec_xl(0, C+((N)*ldc)+M); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = (C[(N+0)*ldc+M] * beta) + result[0]; \
C[(N+1)*ldc+M] = (C[(N+1)*ldc+M] * beta) + result[1];

#else

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
result[2] = vec_mul(result[2], valpha); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
result[3] = vec_mul(result[3], valpha); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = result[0]; \
C[(N+1)*ldc+M] = result[1];

#endif

#define INIT_8ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3); \
__builtin_mma_xxsetaccz(&acc4); \
__builtin_mma_xxsetaccz(&acc5); \
__builtin_mma_xxsetaccz(&acc6); \
__builtin_mma_xxsetaccz(&acc7);

#define INIT_4ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3);

#define INIT_2ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1);

#define INIT_1ACC() __builtin_mma_xxsetaccz(&acc0);

#define LOAD_A_1x8(K, M) \
ra0 = vec_xl(0, A+(K*lda)+M+0); \
ra1 = vec_xl(0, A+(K*lda)+M+2); \
ra2 = vec_xl(0, A+(K*lda)+M+4); \
ra3 = vec_xl(0, A+(K*lda)+M+6);

#define LOAD_A_1x4(K, M) \
ra0 = vec_xl(0, A+(K*lda)+M+0); \
ra1 = vec_xl(0, A+(K*lda)+M+2);

#define LOAD_A_1x2(K, M) ra0 = vec_xl(0, A+(K*lda)+M);

#define LOAD_A_1x1(K, M) ra0 = vec_splats(A[K*lda+M]);

#define LOAD_BP_1x8(K, N) \
pb0 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+0])); \
pb1 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+4]));

#define LOAD_BP_1x4(K, N) \
pb0 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+0]));

#define LOAD_BP_1x2(K, N) \
t0 = vec_xl(0, B+(K*ldb)+N); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)t0, (vec_t)t0);

#define LOAD_B_1x8(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+2); \
rb2 = vec_xl(0, B+(K*ldb)+N+4); \
rb3 = vec_xl(0, B+(K*ldb)+N+6); \

#define LOAD_B_1x4(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+2);

#define LOAD_B_1x2(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0);

#define LOAD_B_1x1(K, N) rb0 = vec_splats(B[K*ldb+N]);

#define KERNEL_MMA_8ACC(b0, b1, b2, b3, b4, b5, b6, b7, \
a0, a1, a2, a3, a4, a5, a6, a7) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3); \
__builtin_mma_xvf64gerpp(&acc4, b4, (vec_t)a4); \
__builtin_mma_xvf64gerpp(&acc5, b5, (vec_t)a5); \
__builtin_mma_xvf64gerpp(&acc6, b6, (vec_t)a6); \
__builtin_mma_xvf64gerpp(&acc7, b7, (vec_t)a7);

#define KERNEL_MMA_4ACC(b0, b1, b2, b3, a0, a1, a2, a3) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3);

#define KERNEL_MMA_2ACC(b0, b1, a0, a1) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1);

#define KERNEL_MMA_1ACC(b0, a0) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0);

#define KERNEL_VMADD_4VSR(a0, a1, a2, a3, b0, b1, b2, b3) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1); \
result2 = vec_madd(a2, b2, result2); \
result3 = vec_madd(a3, b3, result3);

#define KERNEL_VMADD_2VSR(a0, a1, b0, b1) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1);

#define KERNEL_VMADD_1VSR(a0, b0) \
result = vec_madd(a0, b0, result);

#ifdef B0
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
#else
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)
#endif
{
BLASLONG m, n, k;

BLASLONG m8 = M & ~7;
BLASLONG m4 = M & ~3;
BLASLONG m2 = M & ~1;

BLASLONG n8 = N & ~7;
BLASLONG n4 = N & ~3;
BLASLONG n2 = N & ~1;

vector double valpha = vec_splats(alpha);
#if !defined(B0)
vector double vbeta = vec_splats(beta);
#endif

for (m = 0; m < m8; m += 8) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector double ra0, ra1, ra2, ra3;
__vector_pair pb0, pb1;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc4, n+0, m+4);
SAVE_4x2_ACC(&acc6, n+0, m+6);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
SAVE_4x2_ACC(&acc5, n+4, m+4);
SAVE_4x2_ACC(&acc7, n+4, m+6);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BP_1x4(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc2, n+0, m+4);
SAVE_4x2_ACC(&acc3, n+0, m+6);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double t0;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_BP_1x2(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
SAVE_2x2_ACC(&acc2, n+0, m+4);
SAVE_2x2_ACC(&acc3, n+0, m+6);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});
register vector double result2 = ((vector double){0.,0.});
register vector double result3 = ((vector double){0.,0.});

register vector double ra0, ra1, ra2, ra3;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
SAVE_1x4_VSR(result2, n, m+4);
SAVE_1x4_VSR(result3, n, m+6);
}
}

for (; m < m4; m += 4) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1;
__vector_pair pb0, pb1;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BP_1x8(k, n);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BP_1x4(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
register vector double t0;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_BP_1x2(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});

register vector double ra0, ra1;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_2VSR(ra0, ra1, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
}
}

for (; m < m2; m += 2) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0;
__vector_pair pb0, pb1;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BP_1x8(k, n);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BP_1x4(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n, m);
}

for (; n < n2; n += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double t0;
__vector_pair pb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_BP_1x2(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n, m);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
}
}

for (; m < M; m++) {
for (n = 0; n < n8; n += 8) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});
register vector double result2 = ((vector double){0.,0.});
register vector double result3 = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x8(k, n);
KERNEL_VMADD_4VSR(ra0, ra0, ra0, ra0, rb0, rb1, rb2, rb3);
}

SAVE_4x1_VSR(result, n, m);
SAVE_4x1_VSR(result1, n+2, m);
SAVE_4x1_VSR(result2, n+4, m);
SAVE_4x1_VSR(result3, n+6, m);
}

for (; n < n4; n += 4) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x4(k, n);
KERNEL_VMADD_2VSR(ra0, ra0, rb0, rb1);
}

SAVE_4x1_VSR(result, n, m);
SAVE_4x1_VSR(result1, n+2, m);
}

for (; n < n2; n += 2) {
register vector double result = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x2(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

SAVE_4x1_VSR(result, n, m);
}

for (; n < N; n++) {
FLOAT result = 0.0;

for (k = 0; k < K; k++) {
result += A[k*lda+m] * B[k*ldb+n];
}
result = result * alpha;

#if !defined(B0)
C[n*ldc+m] = (C[n*ldc+m] * beta) + result;
#else
C[n*ldc+m] = result;
#endif
}
}

return 0;
}

+ 882
- 0
kernel/power/dgemm_small_kernel_tn_power10.c View File

@@ -0,0 +1,882 @@
/***************************************************************************
Copyright (c) 2021, 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 <altivec.h>

typedef __vector unsigned char vec_t;

#if !__has_builtin(__builtin_vsx_assemble_pair)
#define __builtin_vsx_assemble_pair __builtin_mma_assemble_pair
#endif

#if !defined(B0)
#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
rc0 = vec_xl(0, C+(N+2)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
rc0 = vec_xl(0, C+(N+3)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl_len(C+(N+0)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+1)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+2)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+3)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_2x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl_len(C+(N+0)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+1)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8);

#define SAVE_1x4_VSR(result, N, M) \
rc0 = vec_xl(0, C+((N)*ldc)+M); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst(result, 0, C+((N)*ldc)+M);

#else

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
result[2] = vec_mul(result[2], valpha); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
result[3] = vec_mul(result[3], valpha); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
result[1] = vec_mul(result[1], valpha); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
result[2] = vec_mul(result[2], valpha); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
result[3] = vec_mul(result[3], valpha); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_2x1_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
result[1] = vec_mul(result[1], valpha); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8);

#define SAVE_1x4_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst(result, 0, C+((N)*ldc)+M);

#endif

#define INIT_8ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3); \
__builtin_mma_xxsetaccz(&acc4); \
__builtin_mma_xxsetaccz(&acc5); \
__builtin_mma_xxsetaccz(&acc6); \
__builtin_mma_xxsetaccz(&acc7);

#define INIT_4ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3);

#define INIT_2ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1);

#define INIT_1ACC() __builtin_mma_xxsetaccz(&acc0);

#if (defined(__GNUC__) && (__GNUC__ == 10))
#if defined(_AIX)
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_assemble_pair(&pair, (vec_t)v0, (vec_t)v1);
#else
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_assemble_pair(&pair, (vec_t)v1, (vec_t)v0);
#endif
#else
#define LOAD_PAIR(pair, v0, v1) \
__builtin_vsx_build_pair(&pair, (vec_t)v0, (vec_t)v1);
#endif

#define LOAD_AT_8x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergel(ra0, ra1); \
ra0 = t0; \
ra1 = t1; \
ra2 = vec_xl(0, A+(M+2)*lda+K+0); \
ra3 = vec_xl(0, A+(M+3)*lda+K+0); \
t0 = vec_mergeh(ra2, ra3); \
t1 = vec_mergel(ra2, ra3); \
ra2 = t0; \
ra3 = t1; \
ra4 = vec_xl(0, A+(M+4)*lda+K+0); \
ra5 = vec_xl(0, A+(M+5)*lda+K+0); \
t0 = vec_mergeh(ra4, ra5); \
t1 = vec_mergel(ra4, ra5); \
ra4 = t0; \
ra5 = t1; \
ra6 = vec_xl(0, A+(M+6)*lda+K+0); \
ra7 = vec_xl(0, A+(M+7)*lda+K+0); \
t0 = vec_mergeh(ra6, ra7); \
t1 = vec_mergel(ra6, ra7); \
ra6 = t0; \
ra7 = t1;

#define LOAD_AT_8x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1); \
ra1 = vec_xor(ra1, ra1); \
ra1 = vec_insert(A[(M+2)*lda+K], ra1, 0); \
ra1 = vec_insert(A[(M+3)*lda+K], ra1, 1); \
ra2 = vec_xor(ra2, ra2); \
ra2 = vec_insert(A[(M+4)*lda+K], ra2, 0); \
ra2 = vec_insert(A[(M+5)*lda+K], ra2, 1); \
ra3 = vec_xor(ra3, ra3); \
ra3 = vec_insert(A[(M+6)*lda+K], ra3, 0); \
ra3 = vec_insert(A[(M+7)*lda+K], ra3, 1); \

#define LOAD_AT_4x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
ra2 = vec_xl(0, A+(M+2)*lda+K+0); \
ra3 = vec_xl(0, A+(M+3)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergeh(ra2, ra3); \
t2 = vec_mergel(ra0, ra1); \
t3 = vec_mergel(ra2, ra3); \
ra0 = t0; \
ra1 = t2; \
ra2 = t1; \
ra3 = t3;

#define LOAD_AT_4x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1); \
ra1 = vec_xor(ra1, ra1); \
ra1 = vec_insert(A[(M+2)*lda+K], ra1, 0); \
ra1 = vec_insert(A[(M+3)*lda+K], ra1, 1); \

#define LOAD_AT_2x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergel(ra0, ra1); \
ra0 = t0; \
ra1 = t1;

#define LOAD_AT_2x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1);

#define LOAD_A_1x1(K, M) \
ra0 = vec_splats(A[((M+0)*lda)+K+0]);

#define LOAD_BTP_8x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
rb2 = vec_xl(0, B+(N+2)*ldb+K+0); \
rb3 = vec_xl(0, B+(N+3)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
t1 = vec_mergeh(rb2, rb3); \
LOAD_PAIR(pb0, t0, t1); \
t0 = vec_mergel(rb0, rb1); \
t1 = vec_mergel(rb2, rb3); \
LOAD_PAIR(pb2, t0, t1); \
rb4 = vec_xl(0, B+(N+4)*ldb+K+0); \
rb5 = vec_xl(0, B+(N+5)*ldb+K+0); \
rb6 = vec_xl(0, B+(N+6)*ldb+K+0); \
rb7 = vec_xl(0, B+(N+7)*ldb+K+0); \
t0 = vec_mergeh(rb4, rb5); \
t1 = vec_mergeh(rb6, rb7); \
LOAD_PAIR(pb1, t0, t1); \
t0 = vec_mergel(rb4, rb5); \
t1 = vec_mergel(rb6, rb7); \
LOAD_PAIR(pb3, t0, t1);

#define LOAD_BTP_8x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
rb1 = vec_xor(rb1, rb1); \
rb1 = vec_insert(B[(N+2)*ldb+K], rb1, 0); \
rb1 = vec_insert(B[(N+3)*ldb+K], rb1, 1); \
LOAD_PAIR(pb0, rb0, rb1); \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+4)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+5)*ldb+K], rb0, 1); \
rb1 = vec_xor(rb1, rb1); \
rb1 = vec_insert(B[(N+6)*ldb+K], rb1, 0); \
rb1 = vec_insert(B[(N+7)*ldb+K], rb1, 1); \
LOAD_PAIR(pb1, rb0, rb1);

#define LOAD_BTP_4x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
rb2 = vec_xl(0, B+(N+2)*ldb+K+0); \
rb3 = vec_xl(0, B+(N+3)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
t1 = vec_mergeh(rb2, rb3); \
LOAD_PAIR(pb0, t0, t1); \
t0 = vec_mergel(rb0, rb1); \
t1 = vec_mergel(rb2, rb3); \
LOAD_PAIR(pb1, t0, t1);

#define LOAD_BTP_4x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
rb1 = vec_xor(rb1, rb1); \
rb1 = vec_insert(B[(N+2)*ldb+K], rb1, 0); \
rb1 = vec_insert(B[(N+3)*ldb+K], rb1, 1); \
LOAD_PAIR(pb0, rb0, rb1);

#define LOAD_BTP_2x2(N, K) \
rb0 = vec_xl(0, B+(N+0)*ldb+K+0); \
rb1 = vec_xl(0, B+(N+1)*ldb+K+0); \
t0 = vec_mergeh(rb0, rb1); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)t0, (vec_t)t0); \
t1 = vec_mergel(rb0, rb1); \
__builtin_vsx_assemble_pair(&pb1, (vec_t)t1, (vec_t)t1);

#define LOAD_BTP_2x1(N, K) \
rb0 = vec_xor(rb0, rb0); \
rb0 = vec_insert(B[(N+0)*ldb+K], rb0, 0); \
rb0 = vec_insert(B[(N+1)*ldb+K], rb0, 1); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)rb0, (vec_t)rb0);

#define LOAD_B_1x1(N, K) rb0 = vec_splats(B[((N)*ldb)+K]);

#define KERNEL_MMA_8ACC(b0, b1, b2, b3, b4, b5, b6, b7, \
a0, a1, a2, a3, a4, a5, a6, a7) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3); \
__builtin_mma_xvf64gerpp(&acc4, b4, (vec_t)a4); \
__builtin_mma_xvf64gerpp(&acc5, b5, (vec_t)a5); \
__builtin_mma_xvf64gerpp(&acc6, b6, (vec_t)a6); \
__builtin_mma_xvf64gerpp(&acc7, b7, (vec_t)a7);

#define KERNEL_MMA_4ACC(b0, b1, b2, b3, a0, a1, a2, a3) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3);

#define KERNEL_MMA_2ACC(b0, b1, a0, a1) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1);

#define KERNEL_MMA_1ACC(b0, a0) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0);

#define KERNEL_MMA_1ACC_(acc, b0, a0) \
__builtin_mma_xvf64gerpp(&acc, b0, (vec_t)a0);

#define KERNEL_VMADD_4VSR(a0, a1, a2, a3, b0, b1, b2, b3) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1); \
result2 = vec_madd(a2, b2, result2); \
result3 = vec_madd(a3, b3, result3);

#define KERNEL_VMADD_2VSR(a0, a1, b0, b1) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1);

#define KERNEL_VMADD_1VSR(a0, b0) \
result = vec_madd(a0, b0, result);

#ifdef B0
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
#else
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)
#endif
{
BLASLONG m, n, k;

BLASLONG m8 = M & ~7;
BLASLONG m4 = M & ~3;
BLASLONG m2 = M & ~1;

BLASLONG n8 = N & ~7;
BLASLONG n4 = N & ~3;
BLASLONG n2 = N & ~1;

BLASLONG k2 = K & ~1;

vector double valpha = vec_splats(alpha);
#if !defined(B0)
vector double vbeta = vec_splats(beta);
#endif

for (m = 0; m < m8; m += 8) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_8ACC(pb0, pb0, pb0, pb0, pb1, pb1, pb1, pb1,
ra0, ra2, ra4, ra6, ra0, ra2, ra4, ra6);
KERNEL_MMA_8ACC(pb2, pb2, pb2, pb2, pb3, pb3, pb3, pb3,
ra1, ra3, ra5, ra7, ra1, ra3, ra5, ra7);
}
// workaround to avoid register spilling
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_1ACC_(acc0, pb0, ra0);
KERNEL_MMA_1ACC_(acc1, pb0, ra1);
LOAD_AT_4x1(m+4, k);
KERNEL_MMA_1ACC_(acc2, pb0, ra0);
KERNEL_MMA_1ACC_(acc3, pb0, ra1);
LOAD_AT_4x1(m, k);
LOAD_BTP_4x1(n+4, k);
KERNEL_MMA_1ACC_(acc4, pb0, ra0);
KERNEL_MMA_1ACC_(acc5, pb0, ra1);
LOAD_AT_4x1(m+4, k);
KERNEL_MMA_1ACC_(acc6, pb0, ra0);
KERNEL_MMA_1ACC_(acc7, pb0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+4);
SAVE_4x2_ACC(&acc4, n+4, m+0);
SAVE_4x2_ACC(&acc6, n+4, m+4);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc3, n+0, m+6);
SAVE_4x2_ACC(&acc5, n+4, m+2);
SAVE_4x2_ACC(&acc7, n+4, m+6);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
KERNEL_MMA_4ACC(pb1, pb1, pb1, pb1, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+4);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc3, n+0, m+6);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
KERNEL_MMA_4ACC(pb1, pb1, pb1, pb1, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc2, n+0, m+4);
SAVE_2x2_ACC(&acc1, n+0, m+2);
SAVE_2x2_ACC(&acc3, n+0, m+6);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});
register vector double result2 = ((vector double){0.,0.});
register vector double result3 = ((vector double){0.,0.});

register vector double ra0, ra1, ra2, ra3;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_B_1x1(n, k);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
SAVE_1x4_VSR(result2, n, m+4);
SAVE_1x4_VSR(result3, n, m+6);
}
}

for (; m < m4; m += 4) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1, t2, t3;

__vector_pair pb0, pb1, pb2, pb3;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb1, pb1, ra0, ra2, ra0, ra2);
KERNEL_MMA_4ACC(pb2, pb2, pb3, pb3, ra1, ra3, ra1, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_4ACC(pb0, pb0, pb1, pb1, ra0, ra1, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc2, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1, t2, t3;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra2);
KERNEL_MMA_2ACC(pb1, pb1, ra1, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double rb0, rb1;
register vector double t0, t1, t2, t3;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra2);
KERNEL_MMA_2ACC(pb1, pb1, ra1, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});

register vector double ra0, ra1;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_B_1x1(n, k);
KERNEL_VMADD_2VSR(ra0, ra1, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
}
}

for (; m < m2; m += 2) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
KERNEL_MMA_2ACC(pb2, pb3, ra1, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0, ra1;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
KERNEL_MMA_1ACC(pb1, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n, m);
}

for (; n < n2; n += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0, ra1;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
KERNEL_MMA_1ACC(pb1, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n, m);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});

register vector double ra0, ra1;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_B_1x1(n, k);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
}
}

for (; m < M; m++) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3, rb4, rb5, rb6, rb7;
register vector double t0, t1;

__vector_pair pb0, pb1, pb2, pb3;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_8x2(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_2ACC(pb2, pb3, ra0, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_8x1(n, k);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x1_ACC(&acc0, n+0, m+0);
SAVE_4x1_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1, rb2, rb3;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_4x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_4x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x1_ACC(&acc0, n, m);
}

for (; n < n2; n += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0;
register vector double rb0, rb1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_A_1x1(k, m);
LOAD_BTP_2x2(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_A_1x1(k+1, m);
KERNEL_MMA_1ACC(pb1, ra0);
}
for (; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_BTP_2x1(n, k);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x1_ACC(&acc0, n+0, m+0);
}

for (; n < N; n++) {
FLOAT result = 0.0;

for (k = 0; k < K; k++) {
result += A[m*lda+k] * B[n*ldb+k];
}
result = result * alpha;

#if !defined(B0)
C[n*ldc+m] = (C[n*ldc+m] * beta) + result;
#else
C[n*ldc+m] = result;
#endif
}
}

return 0;
}

+ 829
- 0
kernel/power/dgemm_small_kernel_tt_power10.c View File

@@ -0,0 +1,829 @@
/***************************************************************************
Copyright (c) 2021, 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 <altivec.h>

typedef __vector unsigned char vec_t;

#if !__has_builtin(__builtin_vsx_assemble_pair)
#define __builtin_vsx_assemble_pair __builtin_mma_assemble_pair
#endif

#if !defined(B0)
#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
rc0 = vec_xl(0, C+(N+2)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
rc0 = vec_xl(0, C+(N+3)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
rc0 = vec_xl(0, C+((N)*ldc)+M); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = (C[(N+0)*ldc+M] * beta) + result[0]; \
C[(N+1)*ldc+M] = (C[(N+1)*ldc+M] * beta) + result[1];

#else

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
result[2] = vec_mul(result[2], valpha); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
result[3] = vec_mul(result[3], valpha); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_2x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = result[0]; \
C[(N+1)*ldc+M] = result[1];

#endif

#define INIT_8ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3); \
__builtin_mma_xxsetaccz(&acc4); \
__builtin_mma_xxsetaccz(&acc5); \
__builtin_mma_xxsetaccz(&acc6); \
__builtin_mma_xxsetaccz(&acc7);

#define INIT_4ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3);

#define INIT_2ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1);

#define INIT_1ACC() __builtin_mma_xxsetaccz(&acc0);

#define LOAD_AT_8x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
ra2 = vec_xl(0, A+(M+2)*lda+K+0); \
ra3 = vec_xl(0, A+(M+3)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergeh(ra2, ra3); \
t2 = vec_mergel(ra0, ra1); \
t3 = vec_mergel(ra2, ra3); \
ra0 = t0; \
ra1 = t2; \
ra2 = t1; \
ra3 = t3; \
ra4 = vec_xl(0, A+(M+4)*lda+K+0); \
ra5 = vec_xl(0, A+(M+5)*lda+K+0); \
ra6 = vec_xl(0, A+(M+6)*lda+K+0); \
ra7 = vec_xl(0, A+(M+7)*lda+K+0); \
t0 = vec_mergeh(ra4, ra5); \
t1 = vec_mergeh(ra6, ra7); \
t2 = vec_mergel(ra4, ra5); \
t3 = vec_mergel(ra6, ra7); \
ra4 = t0; \
ra5 = t2; \
ra6 = t1; \
ra7 = t3;

#define LOAD_AT_8x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1); \
ra1 = vec_xor(ra1, ra1); \
ra1 = vec_insert(A[(M+2)*lda+K], ra1, 0); \
ra1 = vec_insert(A[(M+3)*lda+K], ra1, 1); \
ra2 = vec_xor(ra2, ra2); \
ra2 = vec_insert(A[(M+4)*lda+K], ra2, 0); \
ra2 = vec_insert(A[(M+5)*lda+K], ra2, 1); \
ra3 = vec_xor(ra3, ra3); \
ra3 = vec_insert(A[(M+6)*lda+K], ra3, 0); \
ra3 = vec_insert(A[(M+7)*lda+K], ra3, 1); \

#define LOAD_AT_4x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
ra2 = vec_xl(0, A+(M+2)*lda+K+0); \
ra3 = vec_xl(0, A+(M+3)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergeh(ra2, ra3); \
t2 = vec_mergel(ra0, ra1); \
t3 = vec_mergel(ra2, ra3); \
ra0 = t0; \
ra1 = t2; \
ra2 = t1; \
ra3 = t3;

#define LOAD_AT_4x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1); \
ra1 = vec_xor(ra1, ra1); \
ra1 = vec_insert(A[(M+2)*lda+K], ra1, 0); \
ra1 = vec_insert(A[(M+3)*lda+K], ra1, 1); \

#define LOAD_AT_2x2(M, K) \
ra0 = vec_xl(0, A+(M+0)*lda+K+0); \
ra1 = vec_xl(0, A+(M+1)*lda+K+0); \
t0 = vec_mergeh(ra0, ra1); \
t1 = vec_mergel(ra0, ra1); \
ra0 = t0; \
ra1 = t1;

#define LOAD_AT_2x1(M, K) \
ra0 = vec_xor(ra0, ra0); \
ra0 = vec_insert(A[(M+0)*lda+K], ra0, 0); \
ra0 = vec_insert(A[(M+1)*lda+K], ra0, 1);

#define LOAD_A_1x1(M, K) ra0 = vec_splats(A[(M)*lda+K]);

#define LOAD_BP_1x8(K, N) \
pb0 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+0])); \
pb1 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+4]));

#define LOAD_BP_1x4(K, N) \
pb0 = *((__vector_pair *)((void *)&B[((K)*ldb)+N+0]));

#define LOAD_BP_1x2(K, N) \
t0 = vec_xl(0, B+((K)*ldb)+N); \
__builtin_vsx_assemble_pair(&pb0, (vec_t)t0, (vec_t)t0);

#define LOAD_B_1x8(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+2); \
rb2 = vec_xl(0, B+(K*ldb)+N+4); \
rb3 = vec_xl(0, B+(K*ldb)+N+6); \

#define LOAD_B_1x4(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+2);

#define LOAD_B_1x2(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0);

#define LOAD_B_1x1(K, N) rb0 = vec_splats(B[(K)*ldb+N]);

#define KERNEL_MMA_8ACC(b0, b1, b2, b3, b4, b5, b6, b7, \
a0, a1, a2, a3, a4, a5, a6, a7) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3); \
__builtin_mma_xvf64gerpp(&acc4, b4, (vec_t)a4); \
__builtin_mma_xvf64gerpp(&acc5, b5, (vec_t)a5); \
__builtin_mma_xvf64gerpp(&acc6, b6, (vec_t)a6); \
__builtin_mma_xvf64gerpp(&acc7, b7, (vec_t)a7);

#define KERNEL_MMA_4ACC(b0, b1, b2, b3, a0, a1, a2, a3) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1); \
__builtin_mma_xvf64gerpp(&acc2, b2, (vec_t)a2); \
__builtin_mma_xvf64gerpp(&acc3, b3, (vec_t)a3);

#define KERNEL_MMA_2ACC(b0, b1, a0, a1) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0); \
__builtin_mma_xvf64gerpp(&acc1, b1, (vec_t)a1);

#define KERNEL_MMA_1ACC(b0, a0) \
__builtin_mma_xvf64gerpp(&acc0, b0, (vec_t)a0);

#define KERNEL_VMADD_4VSR(a0, a1, a2, a3, b0, b1, b2, b3) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1); \
result2 = vec_madd(a2, b2, result2); \
result3 = vec_madd(a3, b3, result3);

#define KERNEL_VMADD_2VSR(a0, a1, b0, b1) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1);

#define KERNEL_VMADD_1VSR(a0, b0) \
result = vec_madd(a0, b0, result);

#define PACK_A(ra0, ra1, ra2, ra3, offset) \
vec_xst(ra0, 0, packA+(k*8)+0+offset); \
vec_xst(ra1, 0, packA+(k*8)+2+offset); \
vec_xst(ra2, 0, packA+(k*8)+4+offset); \
vec_xst(ra3, 0, packA+(k*8)+6+offset);

#define LOAD_PACKED_A(ra0, ra1, ra2, ra3, offset) \
ra0 = vec_xl(0, packA+(k*8)+0+offset); \
ra1 = vec_xl(0, packA+(k*8)+2+offset); \
ra2 = vec_xl(0, packA+(k*8)+4+offset); \
ra3 = vec_xl(0, packA+(k*8)+6+offset);

#ifdef B0
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
#else
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)
#endif
{
BLASLONG m, n, k;

BLASLONG m8 = M & ~7;
BLASLONG m4 = M & ~3;
BLASLONG m2 = M & ~1;

BLASLONG n8 = N & ~7;
BLASLONG n4 = N & ~3;
BLASLONG n2 = N & ~1;

BLASLONG k2 = K & ~1;

#if defined(__GNUC__) && !defined(__clang__)
int has_packing = (M >= 32 && N >= 32 && K >= 32) ? 1 : 0;
#else
int has_packing = 0;
#endif

double *packA;
if (has_packing) packA = (double *)malloc(K*8*sizeof(double));

vector double valpha = vec_splats(alpha);
#if !defined(B0)
vector double vbeta = vec_splats(beta);
#endif

for (m = 0; m < m8; m += 8) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double t0, t1, t2, t3;

__vector_pair pb0, pb1;

if (has_packing) {
if (n == 0) {
for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra2, ra2, ra4, ra4, ra6, ra6);
PACK_A(ra0, ra2, ra4, ra6, 0);
LOAD_BP_1x8(k+1, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra1, ra1, ra3, ra3, ra5, ra5, ra7, ra7);
PACK_A(ra1, ra3, ra5, ra7, 8);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
PACK_A(ra0, ra1, ra2, ra3, 0);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_PACKED_A(ra0, ra2, ra4, ra6, 0);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra2, ra2, ra4, ra4, ra6, ra6);
LOAD_PACKED_A(ra1, ra3, ra5, ra7, 8);
LOAD_BP_1x8(k+1, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra1, ra1, ra3, ra3, ra5, ra5, ra7, ra7);
}
for (; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra2, ra2, ra4, ra4, ra6, ra6);
LOAD_BP_1x8(k+1, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra1, ra1, ra3, ra3, ra5, ra5, ra7, ra7);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_8ACC(pb0, pb1, pb0, pb1, pb0, pb1, pb0, pb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc4, n+0, m+4);
SAVE_4x2_ACC(&acc6, n+0, m+6);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
SAVE_4x2_ACC(&acc5, n+4, m+4);
SAVE_4x2_ACC(&acc7, n+4, m+6);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double t0, t1, t2, t3;

__vector_pair pb0;

if (!has_packing) {
for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
LOAD_BP_1x4(k+1, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_PACKED_A(ra0, ra2, ra4, ra6, 0);
LOAD_BP_1x4(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
LOAD_PACKED_A(ra1, ra3, ra5, ra7, 8);
LOAD_BP_1x4(k+1, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_BP_1x4(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
SAVE_4x2_ACC(&acc2, n+0, m+4);
SAVE_4x2_ACC(&acc3, n+0, m+6);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3, ra4, ra5, ra6, ra7;
register vector double t0, t1, t2, t3;

__vector_pair pb0;

if (!has_packing) {
for (k = 0; k < k2; k += 2) {
LOAD_AT_8x2(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
LOAD_BP_1x2(k+1, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}
} else {
for (k = 0; k < k2; k += 2) {
LOAD_PACKED_A(ra0, ra2, ra4, ra6, 0);
LOAD_BP_1x2(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra2, ra4, ra6);
LOAD_PACKED_A(ra1, ra3, ra5, ra7, 8);
LOAD_BP_1x2(k+1, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra1, ra3, ra5, ra7);
}
for (; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_BP_1x2(k, n);
KERNEL_MMA_4ACC(pb0, pb0, pb0, pb0, ra0, ra1, ra2, ra3);
}
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
SAVE_2x2_ACC(&acc2, n+0, m+4);
SAVE_2x2_ACC(&acc3, n+0, m+6);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});
register vector double result2 = ((vector double){0.,0.});
register vector double result3 = ((vector double){0.,0.});

register vector double ra0, ra1, ra2, ra3;
register vector double rb0;

if (!has_packing) {
for (k = 0; k < K; k++) {
LOAD_AT_8x1(m, k);
LOAD_B_1x1(k, n);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}
} else {
for (k = 0; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_B_1x1(k, n);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
SAVE_1x4_VSR(result2, n, m+4);
SAVE_1x4_VSR(result3, n, m+6);
}
}

for (; m < m4; m += 4) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double t0, t1, t2, t3;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra2, ra2);
LOAD_BP_1x8(k+1, n);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra1, ra1, ra3, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_4ACC(pb0, pb1, pb0, pb1, ra0, ra0, ra1, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc2, n+0, m+2);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc3, n+4, m+2);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double t0, t1, t2, t3;

__vector_pair pb0;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra2);
LOAD_BP_1x4(k+1, n);
KERNEL_MMA_2ACC(pb0, pb0, ra1, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+0, m+2);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1, ra2, ra3;
register vector double t0, t1, t2, t3;

__vector_pair pb0;

for (k = 0; k < k2; k += 2) {
LOAD_AT_4x2(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra2);
LOAD_BP_1x2(k+1, n);
KERNEL_MMA_2ACC(pb0, pb0, ra1, ra3);
}
for (; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_2ACC(pb0, pb0, ra0, ra1);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n+0, m+0);
SAVE_2x2_ACC(&acc1, n+0, m+2);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});

register vector double ra0, ra1;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_AT_4x1(m, k);
LOAD_B_1x1(k, n);
KERNEL_VMADD_2VSR(ra0, ra1, rb0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
SAVE_1x4_VSR(result1, n, m+2);
}
}

for (; m < m2; m += 2) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector double ra0, ra1;
register vector double t0, t1;

__vector_pair pb0, pb1;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
LOAD_BP_1x8(k+1, n);
KERNEL_MMA_2ACC(pb0, pb1, ra1, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BP_1x8(k, n);
KERNEL_MMA_2ACC(pb0, pb1, ra0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0, ra1;
register vector double t0, t1;

__vector_pair pb0;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_BP_1x4(k+1, n);
KERNEL_MMA_1ACC(pb0, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BP_1x4(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_4x2_ACC(&acc0, n, m);
}

for (; n < n2; n += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector double ra0, ra1;
register vector double t0, t1;

__vector_pair pb0;

for (k = 0; k < k2; k += 2) {
LOAD_AT_2x2(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
LOAD_BP_1x2(k+1, n);
KERNEL_MMA_1ACC(pb0, ra1);
}
for (; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_BP_1x2(k, n);
KERNEL_MMA_1ACC(pb0, ra0);
}

#if !defined(B0)
register vector double rc0;
#endif
vector double result[4];
SAVE_2x2_ACC(&acc0, n, m);
}

for (; n < N; n++) {
register vector double result = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_AT_2x1(m, k);
LOAD_B_1x1(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector double rc0;
#endif
SAVE_1x4_VSR(result, n, m+0);
}
}

for (; m < M; m++) {
for (n = 0; n < n8; n += 8) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});
register vector double result2 = ((vector double){0.,0.});
register vector double result3 = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x1(m, k);
LOAD_B_1x8(k, n);
KERNEL_VMADD_4VSR(ra0, ra0, ra0, ra0, rb0, rb1, rb2, rb3);
}

SAVE_4x1_VSR(result, n, m);
SAVE_4x1_VSR(result1, n+2, m);
SAVE_4x1_VSR(result2, n+4, m);
SAVE_4x1_VSR(result3, n+6, m);
}

for (; n < n4; n += 4) {
register vector double result = ((vector double){0.,0.});
register vector double result1 = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x1(m, k);
LOAD_B_1x4(k, n);
KERNEL_VMADD_2VSR(ra0, ra0, rb0, rb1);
}

SAVE_4x1_VSR(result, n, m);
SAVE_4x1_VSR(result1, n+2, m);
}

for (; n < n2; n += 2) {
register vector double result = ((vector double){0.,0.});

register vector double ra0;
register vector double rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x1(m, k);
LOAD_B_1x2(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

SAVE_4x1_VSR(result, n, m);
}

for (; n < N; n++) {
FLOAT result = 0.0;

for (k = 0; k < K; k++) {
result += A[m*lda+k] * B[k*ldb+n];
}
result = result * alpha;

#if !defined(B0)
C[n*ldc+m] = (C[n*ldc+m] * beta) + result;
#else
C[n*ldc+m] = result;
#endif
}
}

if(has_packing) free(packA);

return 0;
}

+ 84
- 0
kernel/power/gemm_small_kernel_permit_power10.c View File

@@ -0,0 +1,84 @@
/***************************************************************************
Copyright (c) 2021, 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"

int CNAME(int transa, int transb, BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, FLOAT beta)
{
double MNK = (double) M * (double) N * (double) K;

#if defined(DOUBLE) // dgemm

// gcc11 (minor <= 2) has an issue when multiple assemble_pairs are used. This
// issue affects both dgemm_nn and dgemm_tn.
#if (defined(__GNUC__) && (__GNUC__ == 11 && __GNUC_MINOR__ <= 2))
if (!transb)
return 0;
#endif

if (MNK <= 54.0*54.0*54.0)
return 1;

#else // sgemm

#if defined(__GNUC__) && defined(__clang__)
// clang generates code with register spilling for the region of code with
// packing, thus, we had to disable this optimization for clang. Given that
// the packing on-demand used in this work is one of the reasons that lead the
// small kernels to outperform the normal flow (when MNK increases), with it
// disabled we had to reduce the MNK inputs used by the code generated by clang.
if (MNK > 84.0*84.0*84.0)
return 0;

if (transa && !transb) {
// sgemm_tn works better when packing on-demand is used
if (MNK <= 64.0*64.0*64.0 && K >= 4)
return 1;
else
return 0;
}
#else // gcc

if (MNK > 100.0*100.0*100.0)
return 0;

#endif

// Multi-threading execution outperforms (or approaches) the execution of the
// small kernel.
if (num_cpu_avail(3) > 1) {
if (MNK <= 64.0*64.0*64.0)
return 1;
} else {
return 1;
}

#endif

return 0;
}

+ 1563
- 0
kernel/power/sgemm_small_kernel_nn_power10.c
File diff suppressed because it is too large
View File


+ 887
- 0
kernel/power/sgemm_small_kernel_nt_power10.c View File

@@ -0,0 +1,887 @@
/***************************************************************************
Copyright (c) 2021, 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 <altivec.h>

typedef __vector unsigned char vec_t;

#if !defined(B0)
#define SAVE_4x4_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
rc0 = vec_xl(0, C+(N+2)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
rc0 = vec_xl(0, C+(N+3)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl_len(C+(N+0)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+1)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+2)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[2] = vec_madd(result[2], valpha, rc0); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
rc0 = vec_xl_len(C+(N+3)*ldc+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result[3] = vec_madd(result[3], valpha, rc0); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x4_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
rc0 = vec_xl(0, C+(N+0)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[0] = vec_madd(result[0], valpha, rc0); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
rc0 = vec_xl(0, C+(N+1)*ldc+M); \
rc0 = vec_mul(rc0, vbeta); \
result[1] = vec_madd(result[1], valpha, rc0); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
rc0 = vec_xl(0, C+((N)*ldc)+M); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_2x2_VSR(result, N, M) \
rc0 = vec_xl_len(C+(N*ldc)+M, 8); \
rc0 = vec_insert(C[(N+1)*ldc+M+0], rc0, 2); \
rc0 = vec_insert(C[(N+1)*ldc+M+1], rc0, 3); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst_len(result, C+(N*ldc)+M, 8); \
C[(N+1)*ldc+M+0] = result[2]; \
C[(N+1)*ldc+M+1] = result[3];

#define SAVE_1x2_VSR(result, N, M) \
rc0 = vec_xl_len(C+(N*ldc)+M, 8); \
rc0 = vec_mul(rc0, vbeta); \
result = vec_madd(result, valpha, rc0); \
vec_xst_len(result, C+(N*ldc)+M, 8);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = (C[(N+0)*ldc+M] * beta) + result[0]; \
C[(N+1)*ldc+M] = (C[(N+1)*ldc+M] * beta) + result[1]; \
C[(N+2)*ldc+M] = (C[(N+2)*ldc+M] * beta) + result[2]; \
C[(N+3)*ldc+M] = (C[(N+3)*ldc+M] * beta) + result[3];

#define SAVE_2x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = (C[(N+0)*ldc+M] * beta) + result[0]; \
C[(N+1)*ldc+M] = (C[(N+1)*ldc+M] * beta) + result[1];

#else

#define SAVE_4x4_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M); \
result[2] = vec_mul(result[2], valpha); \
vec_xst(result[2], 0, C+(N+2)*ldc+M); \
result[3] = vec_mul(result[3], valpha); \
vec_xst(result[3], 0, C+(N+3)*ldc+M);

#define SAVE_4x2_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst_len(result[0], C+(N+0)*ldc+M, 8); \
result[1] = vec_mul(result[1], valpha); \
vec_xst_len(result[1], C+(N+1)*ldc+M, 8); \
result[2] = vec_mul(result[2], valpha); \
vec_xst_len(result[2], C+(N+2)*ldc+M, 8); \
result[3] = vec_mul(result[3], valpha); \
vec_xst_len(result[3], C+(N+3)*ldc+M, 8);

#define SAVE_2x4_ACC(ACC, N, M) \
__builtin_mma_disassemble_acc((void *)result, ACC); \
result[0] = vec_mul(result[0], valpha); \
vec_xst(result[0], 0, C+(N+0)*ldc+M); \
result[1] = vec_mul(result[1], valpha); \
vec_xst(result[1], 0, C+(N+1)*ldc+M);

#define SAVE_1x4_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst(result, 0, C+((N)*ldc)+M);

#define SAVE_2x2_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst_len(result, C+(N*ldc)+M, 8); \
C[(N+1)*ldc+M+0] = result[2]; \
C[(N+1)*ldc+M+1] = result[3];

#define SAVE_1x2_VSR(result, N, M) \
result = vec_mul(result, valpha); \
vec_xst_len(result, C+(N*ldc)+M, 8);

#define SAVE_4x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = result[0]; \
C[(N+1)*ldc+M] = result[1]; \
C[(N+2)*ldc+M] = result[2]; \
C[(N+3)*ldc+M] = result[3];

#define SAVE_2x1_VSR(result, N, M) \
result = vec_mul(result, valpha); \
C[(N+0)*ldc+M] = result[0]; \
C[(N+1)*ldc+M] = result[1];

#endif

#define INIT_8ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3); \
__builtin_mma_xxsetaccz(&acc4); \
__builtin_mma_xxsetaccz(&acc5); \
__builtin_mma_xxsetaccz(&acc6); \
__builtin_mma_xxsetaccz(&acc7);

#define INIT_4ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1); \
__builtin_mma_xxsetaccz(&acc2); \
__builtin_mma_xxsetaccz(&acc3);

#define INIT_2ACCS() \
__builtin_mma_xxsetaccz(&acc0); \
__builtin_mma_xxsetaccz(&acc1);

#define INIT_1ACC() __builtin_mma_xxsetaccz(&acc0);

#define LOAD_A_1x16(K, M) \
ra0 = vec_xl(0, A+(K*lda)+M+0); \
ra1 = vec_xl(0, A+(K*lda)+M+4); \
ra2 = vec_xl(0, A+(K*lda)+M+8); \
ra3 = vec_xl(0, A+(K*lda)+M+12);

#define LOAD_A_1x8(K, M) \
ra0 = vec_xl(0, A+(K*lda)+M+0); \
ra1 = vec_xl(0, A+(K*lda)+M+4);

#define LOAD_A_1x4(K, M) ra0 = vec_xl(0, A+(K*lda)+M);

#define LOAD_A_2x2(K, M) \
ra0 = vec_splats(A[K*lda+M+0]); \
ra0 = vec_insert(A[K*lda+M+1], ra0, 1); \
ra0 = vec_insert(A[K*lda+M+1], ra0, 3);

#define LOAD_A_1x2(K, M) ra0 = vec_xl_len(A+(K*lda)+M, 8);

#define LOAD_A_1x1(K, M) ra0 = vec_splats(A[K*lda+M+0]);

#define LOAD_B_1x16(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+4); \
rb2 = vec_xl(0, B+(K*ldb)+N+8); \
rb3 = vec_xl(0, B+(K*ldb)+N+12);

#define LOAD_B_1x8(K, N) \
rb0 = vec_xl(0, B+(K*ldb)+N+0); \
rb1 = vec_xl(0, B+(K*ldb)+N+4);

#define LOAD_B_1x4(K, N) rb0 = vec_xl(0, B+(K*ldb)+N);

#define LOAD_B_2x2(K, N) \
rb0 = vec_splats(B[K*ldb+N]); \
rb0 = vec_insert(B[K*ldb+N+1], rb0, 2); \
rb0 = vec_insert(B[K*ldb+N+1], rb0, 3);

#define LOAD_B_1x2(K, N) rb0 = vec_xl_len(B+(K*ldb)+N, 8);

#define LOAD_B_1x1(K, N) rb0 = vec_splats(B[K*ldb+N]);

#define KERNEL_MMA_8ACC(b0, b1, b2, b3, b4, b5, b6, b7, \
a0, a1, a2, a3, a4, a5, a6, a7) \
__builtin_mma_xvf32gerpp(&acc0, (vec_t)b0, (vec_t)a0); \
__builtin_mma_xvf32gerpp(&acc1, (vec_t)b1, (vec_t)a1); \
__builtin_mma_xvf32gerpp(&acc2, (vec_t)b2, (vec_t)a2); \
__builtin_mma_xvf32gerpp(&acc3, (vec_t)b3, (vec_t)a3); \
__builtin_mma_xvf32gerpp(&acc4, (vec_t)b4, (vec_t)a4); \
__builtin_mma_xvf32gerpp(&acc5, (vec_t)b5, (vec_t)a5); \
__builtin_mma_xvf32gerpp(&acc6, (vec_t)b6, (vec_t)a6); \
__builtin_mma_xvf32gerpp(&acc7, (vec_t)b7, (vec_t)a7);

#define KERNEL_MMA_4ACC(b0, b1, b2, b3, a0, a1, a2, a3) \
__builtin_mma_xvf32gerpp(&acc0, (vec_t)b0, (vec_t)a0); \
__builtin_mma_xvf32gerpp(&acc1, (vec_t)b1, (vec_t)a1); \
__builtin_mma_xvf32gerpp(&acc2, (vec_t)b2, (vec_t)a2); \
__builtin_mma_xvf32gerpp(&acc3, (vec_t)b3, (vec_t)a3);

#define KERNEL_MMA_2ACC(b0, b1, a0, a1) \
__builtin_mma_xvf32gerpp(&acc0, (vec_t)b0, (vec_t)a0); \
__builtin_mma_xvf32gerpp(&acc1, (vec_t)b1, (vec_t)a1);

#define KERNEL_MMA_1ACC(b0, a0) \
__builtin_mma_xvf32gerpp(&acc0, (vec_t)b0, (vec_t)a0);

#define KERNEL_VMADD_4VSR(a0, a1, a2, a3, b0, b1, b2, b3) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1); \
result2 = vec_madd(a2, b2, result2); \
result3 = vec_madd(a3, b3, result3);

#define KERNEL_VMADD_2VSR(a0, a1, b0, b1) \
result = vec_madd(a0, b0, result); \
result1 = vec_madd(a1, b1, result1);

#define KERNEL_VMADD_1VSR(a0, b0) \
result = vec_madd(a0, b0, result);

#define PACK_A(ra0, ra1, ra2, ra3, offset) \
vec_xst(ra0, 0, packA+(k*16)+0+offset); \
vec_xst(ra1, 0, packA+(k*16)+4+offset); \
vec_xst(ra2, 0, packA+(k*16)+8+offset); \
vec_xst(ra3, 0, packA+(k*16)+12+offset);

#define LOAD_PACKED_A(ra0, ra1, ra2, ra3, offset) \
ra0 = vec_xl(0, packA+(k*16)+0+offset); \
ra1 = vec_xl(0, packA+(k*16)+4+offset); \
ra2 = vec_xl(0, packA+(k*16)+8+offset); \
ra3 = vec_xl(0, packA+(k*16)+12+offset);

#ifdef B0
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
#else
int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, IFLOAT * A, BLASLONG lda, FLOAT alpha, IFLOAT * B, BLASLONG ldb, FLOAT beta, FLOAT * C, BLASLONG ldc)
#endif
{
BLASLONG m, n, k;

BLASLONG m16 = M & ~15;
BLASLONG m8 = M & ~7;
BLASLONG m4 = M & ~3;
BLASLONG m2 = M & ~1;

BLASLONG n16 = N & ~15;
BLASLONG n8 = N & ~7;
BLASLONG n4 = N & ~3;
BLASLONG n2 = N & ~1;

vector float valpha = vec_splats(alpha);
#if !defined(B0)
vector float vbeta = vec_splats(beta);
#endif

#if defined(__GNUC__) && !defined(__clang__)
int has_packing = (M >= 40 && N >= 40 && K >= 40) ? 1 : 0;
#else
int has_packing = 0;
#endif

float *packA;
if (has_packing) packA = (float *)malloc(K*16*sizeof(float));

for (m = 0; m < m16; m += 16) {
for (n = 0; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector float ra0, ra1, ra2, ra3;
register vector float rb0, rb1;

if (has_packing) {
if (n == 0) {
for (k = 0; k < K; k++) {
LOAD_A_1x16(k, m);
LOAD_B_1x8(k, n);
KERNEL_MMA_8ACC(rb0, rb1, rb0, rb1, rb0, rb1, rb0, rb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
PACK_A(ra0, ra1, ra2, ra3, 0);
}
} else {
for (k = 0; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_B_1x8(k, n);
KERNEL_MMA_8ACC(rb0, rb1, rb0, rb1, rb0, rb1, rb0, rb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}
} else {
for (k = 0; k < K; k++) {
LOAD_A_1x16(k, m);
LOAD_B_1x8(k, n);
KERNEL_MMA_8ACC(rb0, rb1, rb0, rb1, rb0, rb1, rb0, rb1,
ra0, ra0, ra1, ra1, ra2, ra2, ra3, ra3);
}
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc2, n+0, m+4);
SAVE_4x4_ACC(&acc4, n+0, m+8);
SAVE_4x4_ACC(&acc6, n+0, m+12);
SAVE_4x4_ACC(&acc1, n+4, m+0);
SAVE_4x4_ACC(&acc3, n+4, m+4);
SAVE_4x4_ACC(&acc5, n+4, m+8);
SAVE_4x4_ACC(&acc7, n+4, m+12);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector float ra0, ra1, ra2, ra3;
register vector float rb0;

if (!has_packing) {
for (k = 0; k < K; k++) {
LOAD_A_1x16(k, m);
LOAD_B_1x4(k, n);
KERNEL_MMA_4ACC(rb0, rb0, rb0, rb0, ra0, ra1, ra2, ra3);
}
} else {
for (k = 0; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_B_1x4(k, n);
KERNEL_MMA_4ACC(rb0, rb0, rb0, rb0, ra0, ra1, ra2, ra3);
}
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc1, n+0, m+4);
SAVE_4x4_ACC(&acc2, n+0, m+8);
SAVE_4x4_ACC(&acc3, n+0, m+12);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector float ra0, ra1, ra2, ra3;
register vector float rb0;

if (!has_packing) {
for (k = 0; k < K; k++) {
LOAD_A_1x16(k, m);
LOAD_B_1x2(k, n);
KERNEL_MMA_4ACC(rb0, rb0, rb0, rb0, ra0, ra1, ra2, ra3);
}
} else {
for (k = 0; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_B_1x2(k, n);
KERNEL_MMA_4ACC(rb0, rb0, rb0, rb0, ra0, ra1, ra2, ra3);
}
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_2x4_ACC(&acc0, n, m+0);
SAVE_2x4_ACC(&acc1, n, m+4);
SAVE_2x4_ACC(&acc2, n, m+8);
SAVE_2x4_ACC(&acc3, n, m+12);
}

for (; n < N; n++) {
vector float result = ((vector float){0., 0., 0., 0.});
vector float result1 = ((vector float){0., 0., 0., 0.});
vector float result2 = ((vector float){0., 0., 0., 0.});
vector float result3 = ((vector float){0., 0., 0., 0.});

register vector float ra0, ra1, ra2, ra3;
register vector float rb0;

if (!has_packing) {
for (k = 0; k < K; k++) {
LOAD_A_1x16(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}
} else {
for (k = 0; k < K; k++) {
LOAD_PACKED_A(ra0, ra1, ra2, ra3, 0);
LOAD_B_1x1(k, n);
KERNEL_VMADD_4VSR(ra0, ra1, ra2, ra3, rb0, rb0, rb0, rb0);
}
}

#if !defined(B0)
register vector float rc0;
#endif
SAVE_1x4_VSR(result, n, m);
SAVE_1x4_VSR(result1, n, m+4);
SAVE_1x4_VSR(result2, n, m+8);
SAVE_1x4_VSR(result3, n, m+12);
}
}

for (; m < m8; m += 8) {
for (n = 0; n < n16; n += 16) {
__vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;

INIT_8ACCS();

register vector float ra0, ra1;
register vector float rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x16(k, n);
KERNEL_MMA_8ACC(rb0, rb1, rb2, rb3, rb0, rb1, rb2, rb3,
ra0, ra0, ra0, ra0, ra1, ra1, ra1, ra1);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc4, n+0, m+4);
SAVE_4x4_ACC(&acc1, n+4, m+0);
SAVE_4x4_ACC(&acc5, n+4, m+4);
SAVE_4x4_ACC(&acc2, n+8, m+0);
SAVE_4x4_ACC(&acc6, n+8, m+4);
SAVE_4x4_ACC(&acc3, n+12, m+0);
SAVE_4x4_ACC(&acc7, n+12, m+4);
}

for (; n < n8; n += 8) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector float ra0, ra1;
register vector float rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x8(k, n);
KERNEL_MMA_4ACC(rb0, rb1, rb0, rb1, ra0, ra0, ra1, ra1);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc2, n+0, m+4);
SAVE_4x4_ACC(&acc1, n+4, m+0);
SAVE_4x4_ACC(&acc3, n+4, m+4);
}

for (; n < n4; n += 4) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector float ra0, ra1;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x4(k, n);
KERNEL_MMA_2ACC(rb0, rb0, ra0, ra1);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc1, n+0, m+4);
}

for (; n < n2; n += 2) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector float ra0, ra1;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x2(k, n);
KERNEL_MMA_2ACC(rb0, rb0, ra0, ra1);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_2x4_ACC(&acc0, n, m+0);
SAVE_2x4_ACC(&acc1, n, m+4);
}

for (; n < N; n++) {
vector float result = ((vector float){0.,0.,0.,0.});
vector float result1 = ((vector float){0.,0.,0.,0.});

register vector float ra0, ra1;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x8(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_2VSR(ra0, ra1, rb0, rb0);
}

#if !defined(B0)
register vector float rc0;
#endif
SAVE_1x4_VSR(result, n, m);
SAVE_1x4_VSR(result1, n, m+4);
}
}

for (; m < m4; m += 4) {
for (n = 0; n < n16; n += 16) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector float ra0;
register vector float rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x16(k, n);
KERNEL_MMA_4ACC(rb0, rb1, rb2, rb3, ra0, ra0, ra0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc1, n+4, m+0);
SAVE_4x4_ACC(&acc2, n+8, m+0);
SAVE_4x4_ACC(&acc3, n+12, m+0);
}

for (; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector float ra0;
register vector float rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x8(k, n);
KERNEL_MMA_2ACC(rb0, rb1, ra0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
SAVE_4x4_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x4(k, n);
KERNEL_MMA_1ACC(rb0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x4_ACC(&acc0, n+0, m+0);
}

for (; n < n2; n += 2) {
__vector_quad acc0;

INIT_1ACC();

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x2(k, n);
KERNEL_MMA_1ACC(rb0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_2x4_ACC(&acc0, n, m);
}

for (; n < N; n++) {
vector float result = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x4(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector float rc0;
#endif
SAVE_1x4_VSR(result, n, m);
}
}

for (; m < m2; m += 2) {
for (n = 0; n < n16; n += 16) {
__vector_quad acc0, acc1, acc2, acc3;

INIT_4ACCS();

register vector float ra0;
register vector float rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x16(k, n);
KERNEL_MMA_4ACC(rb0, rb1, rb2, rb3, ra0, ra0, ra0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
SAVE_4x2_ACC(&acc2, n+8, m+0);
SAVE_4x2_ACC(&acc3, n+12, m+0);
}

for (; n < n8; n += 8) {
__vector_quad acc0, acc1;

INIT_2ACCS();

register vector float ra0;
register vector float rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x8(k, n);
KERNEL_MMA_2ACC(rb0, rb1, ra0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
SAVE_4x2_ACC(&acc1, n+4, m+0);
}

for (; n < n4; n += 4) {
__vector_quad acc0;

INIT_1ACC();

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x4(k, n);
KERNEL_MMA_1ACC(rb0, ra0);
}

#if !defined(B0)
register vector float rc0;
#endif
vector float result[4];
SAVE_4x2_ACC(&acc0, n+0, m+0);
}

for (; n < n2; n += 2) {
vector float result = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_2x2(k, m);
LOAD_B_2x2(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector float rc0;
#endif
SAVE_2x2_VSR(result, n, m);
}

for (; n < N; n++) {
vector float result = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x2(k, m);
LOAD_B_1x1(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

#if !defined(B0)
register vector float rc0;
#endif
SAVE_1x2_VSR(result, n, m);
}
}

for (; m < M; m++) {
for (n = 0; n < n16; n += 16) {
vector float result = ((vector float){0.,0.,0.,0.});
vector float result1 = ((vector float){0.,0.,0.,0.});
vector float result2 = ((vector float){0.,0.,0.,0.});
vector float result3 = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0, rb1, rb2, rb3;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x16(k, n);
KERNEL_VMADD_4VSR(ra0, ra0, ra0, ra0, rb0, rb1, rb2, rb3);
}

SAVE_4x1_VSR(result, n+0, m);
SAVE_4x1_VSR(result1, n+4, m);
SAVE_4x1_VSR(result2, n+8, m);
SAVE_4x1_VSR(result3, n+12, m);
}

for (; n < n8; n += 8) {
vector float result = ((vector float){0.,0.,0.,0.});
vector float result1 = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0, rb1;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x8(k, n);
KERNEL_VMADD_2VSR(ra0, ra0, rb0, rb1);
}

SAVE_4x1_VSR(result, n+0, m);
SAVE_4x1_VSR(result1, n+4, m);
}

for (; n < n4; n += 4) {
vector float result = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x4(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

SAVE_4x1_VSR(result, n+0, m);
}

for (; n < n2; n += 2) {
vector float result = ((vector float){0.,0.,0.,0.});

register vector float ra0;
register vector float rb0;

for (k = 0; k < K; k++) {
LOAD_A_1x1(k, m);
LOAD_B_1x2(k, n);
KERNEL_VMADD_1VSR(ra0, rb0);
}

SAVE_2x1_VSR(result, n+0, m);
}

for (; n < N; n++) {
FLOAT result = 0.0f;

for (k = 0; k < K; k++) {
result += A[k*lda+m] * B[k*ldb+n];
}
result = result * alpha;

#if !defined(B0)
C[n*ldc+m] = (C[n*ldc+m] * beta) + result;
#else
C[n*ldc+m] = result;
#endif
}
}

if (has_packing) free (packA);

return 0;
}

+ 1678
- 0
kernel/power/sgemm_small_kernel_tn_power10.c
File diff suppressed because it is too large
View File


+ 1559
- 0
kernel/power/sgemm_small_kernel_tt_power10.c
File diff suppressed because it is too large
View File


Write Preview
Loading…
Cancel
Save