Merge pull request #1 from xianyi/develop

rebase

Makefile.x86_64

@@ -9,6 +9,7 @@ endif
 endif
 
 ifeq ($(CORE), SKYLAKEX)
+ifndef DYNAMIC_ARCH
 ifndef NO_AVX512
 CCOMMON_OPT += -march=skylake-avx512
 FCOMMON_OPT += -march=skylake-avx512
@@ -22,6 +23,7 @@ endif
 endif
 endif
 endif
+endif
 
 ifeq ($(OSNAME), Interix)
 ARFLAGS		= -m x64

cmake/arch.cmake

@@ -44,6 +44,10 @@ endif ()
 
 
 if (DYNAMIC_ARCH)
+  if (ARM64)
+    set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99)
+  endif ()
+  
   if (X86)
     set(DYNAMIC_CORE KATMAI COPPERMINE NORTHWOOD PRESCOTT BANIAS CORE2 PENRYN DUNNINGTON NEHALEM ATHLON OPTERON OPTERON_SSE3 BARCELONA BOBCAT ATOM NANO)
   endif ()

common_level3.h

@@ -47,6 +47,14 @@ __global__ void cuda_dgemm_kernel(int, int, int, double *, double *, double *);
 extern "C" {
 #endif
 
+extern void sgemm_kernel_direct(BLASLONG M, BLASLONG N, BLASLONG K,
+	float * A, BLASLONG strideA,
+	float * B, BLASLONG strideB,
+	float * R, BLASLONG strideR);
+
+extern int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
+
+
 int sgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,
 	       float  *, BLASLONG, float   *, BLASLONG, float  *, BLASLONG);
 int dgemm_beta(BLASLONG, BLASLONG, BLASLONG, double,

cpuid_power.c

@@ -136,7 +136,7 @@ int detect(void){
   char buffer[512], *p;
 
   p = (char *)NULL;
-  infile = popen("prtconf|grep 'Processor Type'");
+  infile = popen("prtconf|grep 'Processor Type'", "r");
   while (fgets(buffer, sizeof(buffer), infile)){
     if (!strncmp("Pro", buffer, 3)){
 	p = strchr(buffer, ':') + 2;

driver/others/CMakeLists.txt

@@ -47,7 +47,11 @@ GenerateNamedObjects("abs.c" "DOUBLE" "z_abs" 0 "" "" 1)
 GenerateNamedObjects("openblas_get_config.c;openblas_get_parallel.c" "" "" 0 "" "" 1)
 
 if (DYNAMIC_ARCH)
-  list(APPEND COMMON_SOURCES dynamic.c)
+  if (ARM64)
+    list(APPEND COMMON_SOURCES dynamic_arm64.c)
+  else ()  
+    list(APPEND COMMON_SOURCES dynamic.c)
+  endif ()  
 else ()
   list(APPEND COMMON_SOURCES parameter.c)
 endif ()

interface/gemm.c

@@ -271,6 +271,14 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
   PRINT_DEBUG_CNAME;
 
+#if !defined(COMPLEX) && !defined(DOUBLE) && defined(USE_SGEMM_KERNEL_DIRECT)
+  if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && sgemm_kernel_direct_performant(m,n,k)) {
+	sgemm_kernel_direct(m, n, k, a, lda, b, ldb, c, ldc);
+	return;
+  }
+
+#endif
+
 #ifndef COMPLEX
   args.alpha = (void *)α
   args.beta  = (void *)β

kernel/CMakeLists.txt

@@ -125,10 +125,13 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
       set(USE_TRMM true)
     endif ()
 
-    foreach (float_type ${FLOAT_TYPES})
+    foreach (float_type SINGLE DOUBLE)
       string(SUBSTRING ${float_type} 0 1 float_char)
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type})
+    endforeach()
 
+    foreach (float_type ${FLOAT_TYPES})
+      string(SUBSTRING ${float_type} 0 1 float_char)
       if (${float_char}GEMMINCOPY)
         GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMINCOPY}" "${float_type}" "${${float_char}GEMMINCOPYOBJ}" false "" "" true ${float_type})
       endif ()

kernel/Makefile

@@ -5,8 +5,43 @@ endif
 TOPDIR	= ..
 include $(TOPDIR)/Makefile.system
 
+AVX2OPT = 
+ifeq ($(C_COMPILER), GCC)
+# AVX2 support was added in 4.7.0
+  GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
+  GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
+  ifeq ($(GCCVERSIONGTEQ4)$(GCCMINORVERSIONGTEQ7), 11)
+   AVX2OPT = -mavx2
+  endif
+endif
+ifeq ($(C_COMPILER), CLANG)
+# Any clang posing as gcc 4.2 should be new enough (3.4 or later)
+  GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
+  GCCMINORVERSIONGTEQ2 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 2)
+  ifeq ($(GCCVERSIONGTEQ4)$(GCCMINORVERSIONGTEQ2), 11)
+   AVX2OPT = -mavx2
+  endif
+endif
+ifdef NO_AVX2
+ AVX2OPT=
+endif
+
 ifdef TARGET_CORE
-override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
+ifeq ($(TARGET_CORE), SKYLAKEX)
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) -march=skylake-avx512
+ ifeq ($(OSNAME), CYGWIN_NT)
+  override CFLAGS += -fno-asynchronous-unwind-tables
+ endif
+ ifeq ($(OSNAME), WINNT)
+  ifeq ($(C_COMPILER), GCC)
+   override CFLAGS += -fno-asynchronous-unwind-tables
+  endif
+ endif
+else ifeq ($(TARGET_CORE), HASWELL)
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE) $(AVX2OPT)
+else
+ override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
+endif
 BUILD_KERNEL = 1
 KDIR =
 TSUFFIX = _$(TARGET_CORE)

kernel/arm64/KERNEL.ARMV8

@@ -104,8 +104,38 @@ CDOTKERNEL     = zdot.S
 ZDOTKERNEL     = zdot.S
 DSDOTKERNEL    = dot.S
 
-ifneq ($(OS_DARWIN)$(CROSS),11)
+ifeq ($(OS_DARWIN)$(CROSS),11)
+
+STRMMKERNEL	= ../generic/trmmkernel_2x2.c
+DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
+CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
+
+SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
+DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
+DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
+DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
+DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
+ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
+ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+else
 SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
 STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
 ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
@@ -173,35 +203,4 @@ ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
 ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
 ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-else
-
-STRMMKERNEL	= ../generic/trmmkernel_2x2.c
-DTRMMKERNEL	= ../generic/trmmkernel_2x2.c
-CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-
-SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
-SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
-SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
-SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-DGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-DGEMMONCOPY    = ../generic/gemm_ncopy_2.c
-DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
-DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
-DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
-CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
-ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
-ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
-
 endif

kernel/x86_64/KERNEL.HASWELL

@@ -33,9 +33,10 @@ ZAXPYKERNEL = zaxpy.c
 
 STRMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
+SGEMM_BETA     =  sgemm_beta_skylakex.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
-SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+SGEMMONCOPY    =  sgemm_ncopy_4_skylakex.c
 SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
 SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
 SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
@@ -44,9 +45,10 @@ SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
 DTRMMKERNEL    =  dtrmm_kernel_4x8_haswell.c
 DGEMMKERNEL    =  dgemm_kernel_4x8_haswell.S
+DGEMM_BETA     =  dgemm_beta_skylakex.c
 DGEMMINCOPY    =  ../generic/gemm_ncopy_4.c
 DGEMMITCOPY    =  ../generic/gemm_tcopy_4.c
-DGEMMONCOPY    =  ../generic/gemm_ncopy_8.c
+DGEMMONCOPY    =  dgemm_ncopy_8_skylakex.c
 DGEMMOTCOPY    =  ../generic/gemm_tcopy_8.c
 DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
 DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)

kernel/x86_64/dgemm_beta_skylakex.c

@@ -61,17 +61,17 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   c_offset = c;
 
   if (beta == ZERO){
-    __m512d z_zero;
 
-    z_zero = _mm512_setzero_pd();
     j = n;
     do {
       c_offset1 = c_offset;
       c_offset += ldc;
 
       i = m;
-
+#ifdef __AVX2__
+#ifdef __AVX512CD__
       while (i >= 32) {
+	  __m512d z_zero = _mm512_setzero_pd();
 	  _mm512_storeu_pd(c_offset1, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 8, z_zero);
 	  _mm512_storeu_pd(c_offset1 + 16, z_zero);
@@ -79,12 +79,20 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 	  c_offset1 += 32;
 	  i -= 32;
       }
+#endif
       while (i >= 8) {
+#ifdef __AVX512CD__
+	  __m512d z_zero = _mm512_setzero_pd();
 	  _mm512_storeu_pd(c_offset1, z_zero);
+#else
+	 __m256d y_zero = _mm256_setzero_pd();
+	 _mm256_storeu_pd(c_offset1, y_zero);
+	 _mm256_storeu_pd(c_offset1 + 4, y_zero);
+#endif
 	  c_offset1 += 8;
 	  i -= 8;
       }
-
+#endif
       while (i > 0) {
 	  *c_offset1 = ZERO;
 	  c_offset1 ++;

kernel/x86_64/sgemm_beta_skylakex.c

@@ -61,30 +61,36 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   c_offset = c;
 
   if (beta == ZERO){
-    __m512 z_zero;
-    __m256 y_zero;
 
-    z_zero = _mm512_setzero_ps();
-    y_zero = _mm256_setzero_ps();
     j = n;
     do {
       c_offset1 = c_offset;
       c_offset += ldc;
 
       i = m;
-
+#ifdef __AVX2__
       while (i >= 32) {
+#ifdef __AVX512CD__
+	  __m512 z_zero = _mm512_setzero_ps();
 	  _mm512_storeu_ps(c_offset1, z_zero);
 	  _mm512_storeu_ps(c_offset1 + 16, z_zero);
+#else
+	  __m256 y_zero = _mm256_setzero_ps();
+	  _mm256_storeu_ps(c_offset1, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 8, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 16, y_zero);
+	  _mm256_storeu_ps(c_offset1 + 24, y_zero);
+#endif
 	  c_offset1 += 32;
 	  i -= 32;
       }
       while (i >= 8) {
+	    __m256 y_zero = _mm256_setzero_ps();
 	  _mm256_storeu_ps(c_offset1, y_zero);
 	  c_offset1 += 8;
 	  i -= 8;
       }
-
+#endif
       while (i > 0) {
 	  *c_offset1 = ZERO;
 	  c_offset1 ++;

kernel/x86_64/sgemm_kernel_16x4_skylakex.c

@@ -760,7 +760,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *************************************************************************************/
 
 int __attribute__ ((noinline))
-CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG ldc)
+CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict A, float * __restrict B, float * __restrict C, BLASLONG ldc)
 {
 	unsigned long M = m, N = n, K = k;
 	if (M == 0)
@@ -1175,3 +1175,468 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f
 
 	return 0;
 }
+
+
+/*
+ * "Direct sgemm" code. This code operates directly on the inputs and outputs
+ * of the sgemm call, avoiding the copies, memory realignments and threading,
+ * and only supports alpha = 1 and beta = 0.
+ * This is a common case and provides value for relatively small matrixes.
+ * For larger matrixes the "regular" sgemm code is superior, there the cost of
+ * copying/shuffling the B matrix really pays off.
+ */
+
+
+
+#define DECLARE_RESULT_512(N,M) __m512 result##N##M = _mm512_setzero_ps()
+#define BROADCAST_LOAD_A_512(N,M) __m512 Aval##M = _mm512_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_512(N,M)  __m512 Bval##N = _mm512_loadu_ps(&B[strideB * k + j + (N*16)])
+#define MATMUL_512(N,M)  result##N##M = _mm512_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_512(N,M) _mm512_storeu_ps(&R[(i+M) * strideR + j+(N*16)], result##N##M)
+
+
+#define DECLARE_RESULT_256(N,M) __m256 result##N##M = _mm256_setzero_ps()
+#define BROADCAST_LOAD_A_256(N,M) __m256 Aval##M = _mm256_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_256(N,M)  __m256 Bval##N = _mm256_loadu_ps(&B[strideB * k + j + (N*8)])
+#define MATMUL_256(N,M)  result##N##M = _mm256_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_256(N,M) _mm256_storeu_ps(&R[(i+M) * strideR + j+(N*8)], result##N##M)
+
+#define DECLARE_RESULT_128(N,M) __m128 result##N##M = _mm_setzero_ps()
+#define BROADCAST_LOAD_A_128(N,M) __m128 Aval##M = _mm_broadcastss_ps(_mm_load_ss(&A[k  + strideA * (i+M)]))
+#define LOAD_B_128(N,M)  __m128 Bval##N = _mm_loadu_ps(&B[strideB * k + j + (N*4)])
+#define MATMUL_128(N,M)  result##N##M = _mm_fmadd_ps(Aval##M, Bval##N , result##N##M)
+#define STORE_128(N,M) _mm_storeu_ps(&R[(i+M) * strideR + j+(N*4)], result##N##M)
+
+#define DECLARE_RESULT_SCALAR(N,M) float result##N##M = 0;
+#define BROADCAST_LOAD_A_SCALAR(N,M) float Aval##M = A[k + strideA * (i + M)];
+#define LOAD_B_SCALAR(N,M)  float Bval##N  = B[k * strideB + j + N];
+#define MATMUL_SCALAR(N,M) result##N##M +=  Aval##M * Bval##N;
+#define STORE_SCALAR(N,M)  R[(i+M) * strideR + j + N] = result##N##M;
+
+int sgemm_kernel_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K)
+{
+	int mnk = M * N * K;
+	/* large matrixes -> not performant */
+	if (mnk >= 28 * 512 * 512)
+		return 0;
+
+	/*
+	 * if the B matrix is not a nice multiple if 4 we get many unaligned accesses,
+	 * and the regular sgemm copy/realignment of data pays off much quicker
+	 */
+	if ((N & 3) != 0 && (mnk >= 8 * 512 * 512))
+		return 0;
+
+#ifdef SMP
+	/* if we can run multithreaded, the threading changes the based threshold */
+	if (mnk > 2 * 350 * 512 && num_cpu_avail(3)> 1)
+		return 0;
+#endif
+
+	return 1;
+}
+
+
+
+void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A, BLASLONG strideA, float * __restrict B, BLASLONG strideB , float * __restrict R, BLASLONG strideR)
+{
+	int i, j, k;
+
+        int m4 = M & ~3;
+	int m2 = M & ~1;
+
+	int n64 = N & ~63;
+	int n32 = N & ~31;
+	int n16 = N & ~15;
+	int n8 = N & ~7;
+	int n4 = N & ~3;
+	int n2 = N & ~1;
+
+	i = 0;
+
+	for (i = 0; i < m4; i+=4) {
+
+		for (j = 0; j < n64; j+= 64) {
+			k = 0;
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);    			DECLARE_RESULT_512(2, 1);    DECLARE_RESULT_512(3, 1);
+			DECLARE_RESULT_512(0, 2);    DECLARE_RESULT_512(1, 2);    			DECLARE_RESULT_512(2, 2);    DECLARE_RESULT_512(3, 2);
+			DECLARE_RESULT_512(0, 3);    DECLARE_RESULT_512(1, 3);    			DECLARE_RESULT_512(2, 3);    DECLARE_RESULT_512(3, 3);
+
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);			MATMUL_512(2, 1);		MATMUL_512(3, 1);
+				MATMUL_512(0, 2);		MATMUL_512(1, 2);			MATMUL_512(2, 2);		MATMUL_512(3, 2);
+				MATMUL_512(0, 3);		MATMUL_512(1, 3);			MATMUL_512(2, 3);		MATMUL_512(3, 3);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);			STORE_512(2, 1);		STORE_512(3, 1);
+			STORE_512(0, 2);		STORE_512(1, 2);			STORE_512(2, 2);		STORE_512(3, 2);
+			STORE_512(0, 3);		STORE_512(1, 3);			STORE_512(2, 3);		STORE_512(3, 3);
+		}
+
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);
+			DECLARE_RESULT_512(0, 2);    DECLARE_RESULT_512(1, 2);
+			DECLARE_RESULT_512(0, 3);    DECLARE_RESULT_512(1, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);
+				MATMUL_512(0, 2);		MATMUL_512(1, 2);
+				MATMUL_512(0, 3);		MATMUL_512(1, 3);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);
+			STORE_512(0, 2);		STORE_512(1, 2);
+			STORE_512(0, 3);		STORE_512(1, 3);
+		}
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+			DECLARE_RESULT_512(0, 1);
+			DECLARE_RESULT_512(0, 2);
+			DECLARE_RESULT_512(0, 3);
+
+		 	for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+				BROADCAST_LOAD_A_512(x, 2);
+				BROADCAST_LOAD_A_512(x, 3);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+				MATMUL_512(0, 1);
+				MATMUL_512(0, 2);
+				MATMUL_512(0, 3);
+			}
+			STORE_512(0, 0);
+			STORE_512(0, 1);
+			STORE_512(0, 2);
+			STORE_512(0, 3);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+			DECLARE_RESULT_256(0, 1);
+			DECLARE_RESULT_256(0, 2);
+			DECLARE_RESULT_256(0, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				BROADCAST_LOAD_A_256(x, 1);
+				BROADCAST_LOAD_A_256(x, 2);
+				BROADCAST_LOAD_A_256(x, 3);
+
+				LOAD_B_256(0, x);
+
+				MATMUL_256(0, 0);
+				MATMUL_256(0, 1);
+				MATMUL_256(0, 2);
+				MATMUL_256(0, 3);
+			}
+			STORE_256(0, 0);
+			STORE_256(0, 1);
+			STORE_256(0, 2);
+			STORE_256(0, 3);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+			DECLARE_RESULT_128(0, 1);
+			DECLARE_RESULT_128(0, 2);
+			DECLARE_RESULT_128(0, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				BROADCAST_LOAD_A_128(x, 1);
+				BROADCAST_LOAD_A_128(x, 2);
+				BROADCAST_LOAD_A_128(x, 3);
+
+				LOAD_B_128(0, x);
+
+				MATMUL_128(0, 0);
+				MATMUL_128(0, 1);
+				MATMUL_128(0, 2);
+				MATMUL_128(0, 3);
+			}
+			STORE_128(0, 0);
+			STORE_128(0, 1);
+			STORE_128(0, 2);
+			STORE_128(0, 3);
+		}
+
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+			DECLARE_RESULT_SCALAR(0, 1);	DECLARE_RESULT_SCALAR(1, 1);
+			DECLARE_RESULT_SCALAR(0, 2);	DECLARE_RESULT_SCALAR(1, 2);
+			DECLARE_RESULT_SCALAR(0, 3);	DECLARE_RESULT_SCALAR(1, 3);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				BROADCAST_LOAD_A_SCALAR(x, 1);
+				BROADCAST_LOAD_A_SCALAR(x, 2);
+				BROADCAST_LOAD_A_SCALAR(x, 3);
+
+				LOAD_B_SCALAR(0, x);	LOAD_B_SCALAR(1, x);
+
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 1);	MATMUL_SCALAR(1, 1);
+				MATMUL_SCALAR(0, 2);	MATMUL_SCALAR(1, 2);
+				MATMUL_SCALAR(0, 3);	MATMUL_SCALAR(1, 3);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+			STORE_SCALAR(0, 1);	STORE_SCALAR(1, 1);
+			STORE_SCALAR(0, 2);	STORE_SCALAR(1, 2);
+			STORE_SCALAR(0, 3);	STORE_SCALAR(1, 3);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0)
+			DECLARE_RESULT_SCALAR(0, 1)
+			DECLARE_RESULT_SCALAR(0, 2)
+			DECLARE_RESULT_SCALAR(0, 3)
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				BROADCAST_LOAD_A_SCALAR(0, 1);
+				BROADCAST_LOAD_A_SCALAR(0, 2);
+				BROADCAST_LOAD_A_SCALAR(0, 3);
+
+				LOAD_B_SCALAR(0, 0);
+
+				MATMUL_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 1);
+				MATMUL_SCALAR(0, 2);
+				MATMUL_SCALAR(0, 3);
+			}
+			STORE_SCALAR(0, 0);
+			STORE_SCALAR(0, 1);
+			STORE_SCALAR(0, 2);
+			STORE_SCALAR(0, 3);
+		}
+	}
+
+	for (; i < m2; i+=2) {
+		j = 0;
+
+		for (; j < n64; j+= 64) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);    			DECLARE_RESULT_512(2, 1);    DECLARE_RESULT_512(3, 1);
+
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);			MATMUL_512(2, 1);		MATMUL_512(3, 1);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);			STORE_512(2, 1);		STORE_512(3, 1);
+		}
+
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+			DECLARE_RESULT_512(0, 1);    DECLARE_RESULT_512(1, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+				MATMUL_512(0, 1);		MATMUL_512(1, 1);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+			STORE_512(0, 1);		STORE_512(1, 1);
+		}
+
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+			DECLARE_RESULT_512(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				BROADCAST_LOAD_A_512(x, 1);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+				MATMUL_512(0, 1);
+			}
+			STORE_512(0, 0);
+			STORE_512(0, 1);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+			DECLARE_RESULT_256(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				BROADCAST_LOAD_A_256(x, 1);
+
+				LOAD_B_256(0, x);
+
+				MATMUL_256(0, 0);
+				MATMUL_256(0, 1);
+			}
+			STORE_256(0, 0);
+			STORE_256(0, 1);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+			DECLARE_RESULT_128(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				BROADCAST_LOAD_A_128(x, 1);
+
+				LOAD_B_128(0, x);
+
+				MATMUL_128(0, 0);
+				MATMUL_128(0, 1);
+			}
+			STORE_128(0, 0);
+			STORE_128(0, 1);
+		}
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+			DECLARE_RESULT_SCALAR(0, 1);	DECLARE_RESULT_SCALAR(1, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				BROADCAST_LOAD_A_SCALAR(x, 1);
+
+				LOAD_B_SCALAR(0, x);	LOAD_B_SCALAR(1, x);
+
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 1);	MATMUL_SCALAR(1, 1);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+			STORE_SCALAR(0, 1);	STORE_SCALAR(1, 1);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0);
+			DECLARE_RESULT_SCALAR(0, 1);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				BROADCAST_LOAD_A_SCALAR(0, 1);
+
+				LOAD_B_SCALAR(0, 0);
+
+				MATMUL_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 1);
+			}
+			STORE_SCALAR(0, 0);
+			STORE_SCALAR(0, 1);
+		}
+	}
+
+	for (; i < M; i+=1) {
+		j = 0;
+		for (; j < n64; j+= 64) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);    			DECLARE_RESULT_512(2, 0);    DECLARE_RESULT_512(3, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);			LOAD_B_512(2, x);		LOAD_B_512(3, x);
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);			MATMUL_512(2, 0);		MATMUL_512(3, 0);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);			STORE_512(2, 0);		STORE_512(3, 0);
+		}
+		for (; j < n32; j+= 32) {
+			DECLARE_RESULT_512(0, 0);    DECLARE_RESULT_512(1, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+				LOAD_B_512(0, x);		LOAD_B_512(1, x);
+				MATMUL_512(0, 0);		MATMUL_512(1, 0);
+			}
+			STORE_512(0, 0);		STORE_512(1, 0);
+		}
+
+
+		for (; j < n16; j+= 16) {
+			DECLARE_RESULT_512(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_512(x, 0);
+
+				LOAD_B_512(0, x);
+
+				MATMUL_512(0, 0);
+			}
+			STORE_512(0, 0);
+		}
+
+		for (; j < n8; j+= 8) {
+			DECLARE_RESULT_256(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_256(x, 0);
+				LOAD_B_256(0, x);
+				MATMUL_256(0, 0);
+			}
+			STORE_256(0, 0);
+		}
+
+		for (; j < n4; j+= 4) {
+			DECLARE_RESULT_128(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_128(x, 0);
+				LOAD_B_128(0, x);
+				MATMUL_128(0, 0);
+			}
+			STORE_128(0, 0);
+		}
+
+		for (; j < n2; j+= 2) {
+			DECLARE_RESULT_SCALAR(0, 0);	DECLARE_RESULT_SCALAR(1, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(x, 0);
+				LOAD_B_SCALAR(0, 0);	LOAD_B_SCALAR(1, 0);
+				MATMUL_SCALAR(0, 0);	MATMUL_SCALAR(1, 0);
+			}
+			STORE_SCALAR(0, 0);	STORE_SCALAR(1, 0);
+		}
+
+		for (; j < N; j++) {
+			DECLARE_RESULT_SCALAR(0, 0);
+
+			for (k = 0; k < K; k++) {
+				BROADCAST_LOAD_A_SCALAR(0, 0);
+				LOAD_B_SCALAR(0, 0);
+				MATMUL_SCALAR(0, 0);
+			}
+			STORE_SCALAR(0, 0);
+		}
+	}
+}

kernel/x86_64/sgemm_ncopy_4_skylakex.c

@@ -49,8 +49,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT * __restrict a, BLASLONG lda, FLOAT * __
   FLOAT *b_offset;
   FLOAT  ctemp1,  ctemp2,  ctemp3,  ctemp4;
   FLOAT  ctemp5,  ctemp6,  ctemp7,  ctemp8;
-  FLOAT  ctemp9, ctemp10, ctemp11, ctemp12;
-  FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
+  FLOAT  ctemp9,  ctemp13;
 
   a_offset = a;
   b_offset = b;

param.h

@@ -1508,6 +1508,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P  8
 
 #define SWITCH_RATIO	32
+#define GEMM_PREFERED_SIZE	16
 
 #ifdef ARCH_X86
 
@@ -1628,6 +1629,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define SWITCH_RATIO	32
 #define GEMM_PREFERED_SIZE	32
+#define USE_SGEMM_KERNEL_DIRECT 1
 
 #ifdef ARCH_X86