#define DGEMM_SMALL_KERNEL_TN dgemm_small_kernel_tn
#define DGEMM_SMALL_KERNEL_TT dgemm_small_kernel_tt
-#define DGEMM_SMALL_KERNEL_A1B0_NN dgemm_small_kernel_a1b0_nn
-#define DGEMM_SMALL_KERNEL_A1B0_NT dgemm_small_kernel_a1b0_nt
-#define DGEMM_SMALL_KERNEL_A1B0_TN dgemm_small_kernel_a1b0_tn
-#define DGEMM_SMALL_KERNEL_A1B0_TT dgemm_small_kernel_a1b0_tt
+#define DGEMM_SMALL_KERNEL_B0_NN dgemm_small_kernel_b0_nn
+#define DGEMM_SMALL_KERNEL_B0_NT dgemm_small_kernel_b0_nt
+#define DGEMM_SMALL_KERNEL_B0_TN dgemm_small_kernel_b0_tn
+#define DGEMM_SMALL_KERNEL_B0_TT dgemm_small_kernel_b0_tt
#else
int dgemm_small_kernel_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
int dgemm_small_kernel_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double beta, double * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-int sgemm_small_kernel_a1b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
-
-int dgemm_small_kernel_a1b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
-int dgemm_small_kernel_a1b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+int sgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+
+int dgemm_small_kernel_b0_nn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_nt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_tn(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+int dgemm_small_kernel_b0_tt(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
#endif
#define GEMM_SMALL_KERNEL_NT DGEMM_SMALL_KERNEL_NT
#define GEMM_SMALL_KERNEL_TN DGEMM_SMALL_KERNEL_TN
#define GEMM_SMALL_KERNEL_TT DGEMM_SMALL_KERNEL_TT
-#define GEMM_SMALL_KERNEL_A1B0_NN DGEMM_SMALL_KERNEL_A1B0_NN
-#define GEMM_SMALL_KERNEL_A1B0_NT DGEMM_SMALL_KERNEL_A1B0_NT
-#define GEMM_SMALL_KERNEL_A1B0_TN DGEMM_SMALL_KERNEL_A1B0_TN
-#define GEMM_SMALL_KERNEL_A1B0_TT DGEMM_SMALL_KERNEL_A1B0_TT
+#define GEMM_SMALL_KERNEL_B0_NN DGEMM_SMALL_KERNEL_B0_NN
+#define GEMM_SMALL_KERNEL_B0_NT DGEMM_SMALL_KERNEL_B0_NT
+#define GEMM_SMALL_KERNEL_B0_TN DGEMM_SMALL_KERNEL_B0_TN
+#define GEMM_SMALL_KERNEL_B0_TT DGEMM_SMALL_KERNEL_B0_TT
#elif defined(BFLOAT16)
#define GEMM_SMALL_KERNEL_TN SGEMM_SMALL_KERNEL_TN
#define GEMM_SMALL_KERNEL_TT SGEMM_SMALL_KERNEL_TT
-#define GEMM_SMALL_KERNEL_A1B0_NN SGEMM_SMALL_KERNEL_A1B0_NN
-#define GEMM_SMALL_KERNEL_A1B0_NT SGEMM_SMALL_KERNEL_A1B0_NT
-#define GEMM_SMALL_KERNEL_A1B0_TN SGEMM_SMALL_KERNEL_A1B0_TN
-#define GEMM_SMALL_KERNEL_A1B0_TT SGEMM_SMALL_KERNEL_A1B0_TT
+#define GEMM_SMALL_KERNEL_B0_NN SGEMM_SMALL_KERNEL_B0_NN
+#define GEMM_SMALL_KERNEL_B0_NT SGEMM_SMALL_KERNEL_B0_NT
+#define GEMM_SMALL_KERNEL_B0_TN SGEMM_SMALL_KERNEL_B0_TN
+#define GEMM_SMALL_KERNEL_B0_TT SGEMM_SMALL_KERNEL_B0_TT
#endif
#define GEMM_SMALL_KERNEL_TN SGEMM_SMALL_KERNEL_TN
#define GEMM_SMALL_KERNEL_TT SGEMM_SMALL_KERNEL_TT
-#define GEMM_SMALL_KERNEL_A1B0_NN SGEMM_SMALL_KERNEL_A1B0_NN
-#define GEMM_SMALL_KERNEL_A1B0_NT SGEMM_SMALL_KERNEL_A1B0_NT
-#define GEMM_SMALL_KERNEL_A1B0_TN SGEMM_SMALL_KERNEL_A1B0_TN
-#define GEMM_SMALL_KERNEL_A1B0_TT SGEMM_SMALL_KERNEL_A1B0_TT
+#define GEMM_SMALL_KERNEL_B0_NN SGEMM_SMALL_KERNEL_B0_NN
+#define GEMM_SMALL_KERNEL_B0_NT SGEMM_SMALL_KERNEL_B0_NT
+#define GEMM_SMALL_KERNEL_B0_TN SGEMM_SMALL_KERNEL_B0_TN
+#define GEMM_SMALL_KERNEL_B0_TT SGEMM_SMALL_KERNEL_B0_TT
#endif
#else
#define SGEMM_SMALL_KERNEL_TN sgemm_small_kernel_tn
#define SGEMM_SMALL_KERNEL_TT sgemm_small_kernel_tt
-#define SGEMM_SMALL_KERNEL_A1B0_NN sgemm_small_kernel_a1b0_nn
-#define SGEMM_SMALL_KERNEL_A1B0_NT sgemm_small_kernel_a1b0_nt
-#define SGEMM_SMALL_KERNEL_A1B0_TN sgemm_small_kernel_a1b0_tn
-#define SGEMM_SMALL_KERNEL_A1B0_TT sgemm_small_kernel_a1b0_tt
+#define SGEMM_SMALL_KERNEL_B0_NN sgemm_small_kernel_b0_nn
+#define SGEMM_SMALL_KERNEL_B0_NT sgemm_small_kernel_b0_nt
+#define SGEMM_SMALL_KERNEL_B0_TN sgemm_small_kernel_b0_tn
+#define SGEMM_SMALL_KERNEL_B0_TT sgemm_small_kernel_b0_tt
#else
#endif
};
-static int (*gemm_small_kernel_a1b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
+static int (*gemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
#ifndef GEMM3M
#ifndef COMPLEX
- GEMM_SMALL_KERNEL_A1B0_NN, GEMM_SMALL_KERNEL_A1B0_TN, NULL, NULL,
- GEMM_SMALL_KERNEL_A1B0_NT, GEMM_SMALL_KERNEL_A1B0_TT, NULL, NULL,
+ GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, NULL, NULL,
+ GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, NULL, NULL,
#endif
#endif
};
//need to tune small matrices cases.
if(MNK <= 100.0*100.0*100.0){
- if(*(FLOAT *)(args.alpha) == 1.0 && *(FLOAT *)(args.beta) == 0.0){
- (gemm_small_kernel_a1b0[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda,args.b, args.ldb, args.c, args.ldc);
+ if(*(FLOAT *)(args.beta) == 0.0){
+ (gemm_small_kernel_b0[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, *(FLOAT *)(args.alpha), args.b, args.ldb, args.c, args.ldc);
}else{
(gemm_small_kernel[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, *(FLOAT *)(args.alpha), args.b, args.ldb, *(FLOAT *)(args.beta), args.c, args.ldc);
}
SBLASOBJS += \
sgemm_small_kernel_nn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_nt$(TSUFFIX).$(SUFFIX) \
sgemm_small_kernel_tn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_tt$(TSUFFIX).$(SUFFIX) \
- sgemm_small_kernel_a1b0_nn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_a1b0_nt$(TSUFFIX).$(SUFFIX) \
- sgemm_small_kernel_a1b0_tn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_a1b0_tt$(TSUFFIX).$(SUFFIX)
+ sgemm_small_kernel_b0_nn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_b0_nt$(TSUFFIX).$(SUFFIX) \
+ sgemm_small_kernel_b0_tn$(TSUFFIX).$(SUFFIX) sgemm_small_kernel_b0_tt$(TSUFFIX).$(SUFFIX)
DBLASOBJS += \
dgemm_small_kernel_nn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_nt$(TSUFFIX).$(SUFFIX) \
dgemm_small_kernel_tn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_tt$(TSUFFIX).$(SUFFIX) \
- dgemm_small_kernel_a1b0_nn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_a1b0_nt$(TSUFFIX).$(SUFFIX) \
- dgemm_small_kernel_a1b0_tn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_a1b0_tt$(TSUFFIX).$(SUFFIX)
+ dgemm_small_kernel_b0_nn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_b0_nt$(TSUFFIX).$(SUFFIX) \
+ dgemm_small_kernel_b0_tn$(TSUFFIX).$(SUFFIX) dgemm_small_kernel_b0_tt$(TSUFFIX).$(SUFFIX)
endif
$(KDIR)dgemm_small_kernel_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_TT)
$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
-ifndef DGEMM_SMALL_K_A1B0_NN
-DGEMM_SMALL_K_A1B0_NN = ../generic/gemm_small_matrix_kernel_a1b0_nn.c
+ifndef DGEMM_SMALL_K_B0_NN
+DGEMM_SMALL_K_B0_NN = ../generic/gemm_small_matrix_kernel_b0_nn.c
endif
-ifndef DGEMM_SMALL_K_A1B0_NT
-DGEMM_SMALL_K_A1B0_NT = ../generic/gemm_small_matrix_kernel_a1b0_nt.c
+ifndef DGEMM_SMALL_K_B0_NT
+DGEMM_SMALL_K_B0_NT = ../generic/gemm_small_matrix_kernel_b0_nt.c
endif
-ifndef DGEMM_SMALL_K_A1B0_TN
-DGEMM_SMALL_K_A1B0_TN = ../generic/gemm_small_matrix_kernel_a1b0_tn.c
+ifndef DGEMM_SMALL_K_B0_TN
+DGEMM_SMALL_K_B0_TN = ../generic/gemm_small_matrix_kernel_b0_tn.c
endif
-ifndef DGEMM_SMALL_K_A1B0_TT
-DGEMM_SMALL_K_A1B0_TT = ../generic/gemm_small_matrix_kernel_a1b0_tt.c
+ifndef DGEMM_SMALL_K_B0_TT
+DGEMM_SMALL_K_B0_TT = ../generic/gemm_small_matrix_kernel_b0_tt.c
endif
-$(KDIR)dgemm_small_kernel_a1b0_nn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_A1B0_NN)
+$(KDIR)dgemm_small_kernel_b0_nn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_B0_NN)
$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)dgemm_small_kernel_a1b0_nt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_A1B0_NT)
+$(KDIR)dgemm_small_kernel_b0_nt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_B0_NT)
$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)dgemm_small_kernel_a1b0_tn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_A1B0_TN)
+$(KDIR)dgemm_small_kernel_b0_tn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_B0_TN)
$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)dgemm_small_kernel_a1b0_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_A1B0_TT)
+$(KDIR)dgemm_small_kernel_b0_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMM_SMALL_K_B0_TT)
$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
$(KDIR)sgemm_small_kernel_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_TT)
$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
-ifndef SGEMM_SMALL_K_A1B0_NN
-SGEMM_SMALL_K_A1B0_NN = ../generic/gemm_small_matrix_kernel_a1b0_nn.c
+ifndef SGEMM_SMALL_K_B0_NN
+SGEMM_SMALL_K_B0_NN = ../generic/gemm_small_matrix_kernel_b0_nn.c
endif
-ifndef SGEMM_SMALL_K_A1B0_NT
-SGEMM_SMALL_K_A1B0_NT = ../generic/gemm_small_matrix_kernel_a1b0_nt.c
+ifndef SGEMM_SMALL_K_B0_NT
+SGEMM_SMALL_K_B0_NT = ../generic/gemm_small_matrix_kernel_b0_nt.c
endif
-ifndef SGEMM_SMALL_K_A1B0_TN
-SGEMM_SMALL_K_A1B0_TN = ../generic/gemm_small_matrix_kernel_a1b0_tn.c
+ifndef SGEMM_SMALL_K_B0_TN
+SGEMM_SMALL_K_B0_TN = ../generic/gemm_small_matrix_kernel_b0_tn.c
endif
-ifndef SGEMM_SMALL_K_A1B0_TT
-SGEMM_SMALL_K_A1B0_TT = ../generic/gemm_small_matrix_kernel_a1b0_tt.c
+ifndef SGEMM_SMALL_K_B0_TT
+SGEMM_SMALL_K_B0_TT = ../generic/gemm_small_matrix_kernel_b0_tt.c
endif
-$(KDIR)sgemm_small_kernel_a1b0_nn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_A1B0_NN)
+$(KDIR)sgemm_small_kernel_b0_nn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_B0_NN)
$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)sgemm_small_kernel_a1b0_nt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_A1B0_NT)
+$(KDIR)sgemm_small_kernel_b0_nt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_B0_NT)
$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)sgemm_small_kernel_a1b0_tn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_A1B0_TN)
+$(KDIR)sgemm_small_kernel_b0_tn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_B0_TN)
$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
-$(KDIR)sgemm_small_kernel_a1b0_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_A1B0_TT)
+$(KDIR)sgemm_small_kernel_b0_tt$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_SMALL_K_B0_TT)
$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
#include "common.h"
-int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT * B, BLASLONG ldb,FLOAT * C, BLASLONG ldc)
+int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb,FLOAT * C, BLASLONG ldc)
{
//naive implemtation
//Column major
for(k=0; k<K; k++){
result += A[i+k*lda] * B[k+j*ldb];
}
- C[i+j*ldc]=result;
+ C[i+j*ldc]=alpha * result;
}
}
#include "common.h"
-int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
+int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
//naive implemtation
//Column major
for(k=0; k<K; k++){
result += A[i+k*lda] * B[k*ldb+j];
}
- C[i+j*ldc]=result;
+ C[i+j*ldc]=alpha * result;
}
}
#include "common.h"
-int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda,FLOAT * B, BLASLONG ldb,FLOAT * C, BLASLONG ldc)
+int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb,FLOAT * C, BLASLONG ldc)
{
//naive implemtation
//Column major
for(k=0; k<K; k++){
result += A[i*lda+k] * B[k+j*ldb];
}
- C[i+j*ldc]=result;
+ C[i+j*ldc]=alpha * result;
}
}
#include "common.h"
-int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
+int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT * A, BLASLONG lda, FLOAT alpha, FLOAT * B, BLASLONG ldb, FLOAT * C, BLASLONG ldc)
{
//naive implemtation
//Column major
for(k=0; k<K; k++){
result += A[i*lda+k] * B[k*ldb+j];
}
- C[i+j*ldc]=result;
+ C[i+j*ldc]=alpha * result;
}
}