int (*sgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
#endif
#ifdef BUILD_SINGLE
+#ifdef SMALL_MATRIX_OPT
+ int (*sgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float beta);
+
+ int (*sgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
+ int (*sgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
+ int (*sgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * C, BLASLONG ldc);
+ int (*sgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha, float * B, BLASLONG ldb, float beta, float * 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);
+#endif
+
int (*strsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*strsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*strsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*dgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
#endif
#ifdef BUILD_DOUBLE
+#ifdef SMALL_MATRIX_OPT
+ int (*dgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha, double beta);
+
+ int (*dgemm_small_kernel_nn )(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_nt )(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_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 (*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
int (*dtrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
int (*dtrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
int (*dtrsm_kernel_RN)(BLASLONG, BLASLONG, BLASLONG, double, double *, double *, double *, BLASLONG, BLASLONG);
int (*cgemm_oncopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
int (*cgemm_otcopy )(BLASLONG, BLASLONG, float *, BLASLONG, float *);
+#ifdef SMALL_MATRIX_OPT
+ int (*cgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, float alpha0, float alpha1, float beta0, float beta1);
+
+ int (*cgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float beta0, float beta1, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+
+ int (*cgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+ int (*cgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, float * A, BLASLONG lda, float alpha0, float alpha1, float * B, BLASLONG ldb, float * C, BLASLONG ldc);
+#endif
+
int (*ctrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*ctrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*ctrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, float, float, float *, float *, float *, BLASLONG, BLASLONG);
int (*zgemm_oncopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
int (*zgemm_otcopy )(BLASLONG, BLASLONG, double *, BLASLONG, double *);
+#ifdef SMALL_MATRIX_OPT
+ int (*zgemm_small_matrix_permit)(int transa, int transb, BLASLONG m, BLASLONG n, BLASLONG k, double alpha0, double alpha1, double beta0, double beta1);
+
+ int (*zgemm_small_kernel_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double beta0, double beta1, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_b0_nn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_nt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_nr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_nc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_b0_tn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_tt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_tr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_tc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_b0_rn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_rt )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_rr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_rc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+
+ int (*zgemm_small_kernel_b0_cn )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_ct )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_cr )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+ int (*zgemm_small_kernel_b0_cc )(BLASLONG m, BLASLONG n, BLASLONG k, double * A, BLASLONG lda, double alpha0, double alpha1, double * B, BLASLONG ldb, double * C, BLASLONG ldc);
+#endif
+
int (*ztrsm_kernel_LN)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
int (*ztrsm_kernel_LT)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
int (*ztrsm_kernel_LR)(BLASLONG, BLASLONG, BLASLONG, double, double, double *, double *, double *, BLASLONG, BLASLONG);
extern gotoblas_t *gotoblas;
+#define FUNC_OFFSET(func) (size_t)(&((gotoblas_t *)NULL)->func)
+
#define DTB_ENTRIES gotoblas -> dtb_entries
#define GEMM_OFFSET_A gotoblas -> offsetA
#define GEMM_OFFSET_B gotoblas -> offsetB
#else
+#define FUNC_OFFSET(func) (size_t)(func)
+
#define DTB_ENTRIES DTB_DEFAULT_ENTRIES
#define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
};
#ifdef SMALL_MATRIX_OPT
+#ifndef DYNAMIC_ARCH
+#define SMALL_KERNEL_ADDR(table, idx) ((void *)(table[idx]))
+#else
+#define SMALL_KERNEL_ADDR(table, idx) ((void *)(*(uintptr_t *)((char *)gotoblas + (size_t)(table[idx]))))
+#endif
+
#ifndef COMPLEX
-static int (*gemm_small_kernel[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT ,FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG) = {
+static size_t gemm_small_kernel[] = {
#ifndef GEMM3M
- GEMM_SMALL_KERNEL_NN, GEMM_SMALL_KERNEL_TN, NULL, NULL,
- GEMM_SMALL_KERNEL_NT, GEMM_SMALL_KERNEL_TT, NULL, NULL,
+ GEMM_SMALL_KERNEL_NN, GEMM_SMALL_KERNEL_TN, 0, 0,
+ GEMM_SMALL_KERNEL_NT, GEMM_SMALL_KERNEL_TT, 0, 0,
#endif
};
-static int (*gemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
+
+static size_t gemm_small_kernel_b0[] = {
#ifndef GEMM3M
- GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, NULL, NULL,
- GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, NULL, NULL,
+ GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, 0, 0,
+ GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, 0, 0,
#endif
};
+#define GEMM_SMALL_KERNEL_B0(idx) (int (*)(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG)) SMALL_KERNEL_ADDR(gemm_small_kernel_b0, (idx))
+#define GEMM_SMALL_KERNEL(idx) (int (*)(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT ,FLOAT *, BLASLONG, FLOAT, FLOAT *, BLASLONG)) SMALL_KERNEL_ADDR(gemm_small_kernel, (idx))
#else
-static int (*zgemm_small_kernel[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG) = {
+static size_t zgemm_small_kernel[] = {
#ifndef GEMM3M
GEMM_SMALL_KERNEL_NN, GEMM_SMALL_KERNEL_TN, GEMM_SMALL_KERNEL_RN, GEMM_SMALL_KERNEL_CN,
GEMM_SMALL_KERNEL_NT, GEMM_SMALL_KERNEL_TT, GEMM_SMALL_KERNEL_RT, GEMM_SMALL_KERNEL_CT,
#endif
};
-static int (*zgemm_small_kernel_b0[])(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG) = {
+static size_t zgemm_small_kernel_b0[] = {
#ifndef GEMM3M
GEMM_SMALL_KERNEL_B0_NN, GEMM_SMALL_KERNEL_B0_TN, GEMM_SMALL_KERNEL_B0_RN, GEMM_SMALL_KERNEL_B0_CN,
GEMM_SMALL_KERNEL_B0_NT, GEMM_SMALL_KERNEL_B0_TT, GEMM_SMALL_KERNEL_B0_RT, GEMM_SMALL_KERNEL_B0_CT,
GEMM_SMALL_KERNEL_B0_NC, GEMM_SMALL_KERNEL_B0_TC, GEMM_SMALL_KERNEL_B0_RC, GEMM_SMALL_KERNEL_B0_CC,
#endif
};
+
+#define ZGEMM_SMALL_KERNEL(idx) (int (*)(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG)) SMALL_KERNEL_ADDR(zgemm_small_kernel, (idx))
+#define ZGEMM_SMALL_KERNEL_B0(idx) (int (*)(BLASLONG, BLASLONG, BLASLONG, FLOAT *, BLASLONG, FLOAT , FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG)) SMALL_KERNEL_ADDR(zgemm_small_kernel_b0, (idx))
#endif
#endif
IFLOAT *buffer;
IFLOAT *sa, *sb;
-#if defined (SMP) || defined(SMALL_MATRIX_OPT)
+#ifdef SMP
double MNK;
#if defined(USE_SIMPLE_THREADED_LEVEL3) || !defined(NO_AFFINITY)
#ifndef COMPLEX
XFLOAT *buffer;
XFLOAT *sa, *sb;
-#if defined (SMP) || defined(SMALL_MATRIX_OPT)
- double MNK;
-#endif
-
#ifdef SMP
+ double MNK;
#if defined(USE_SIMPLE_THREADED_LEVEL3) || !defined(NO_AFFINITY)
#ifndef COMPLEX
#ifdef XDOUBLE
FUNCTION_PROFILE_START();
-#if defined(SMP) || defined(SMALL_MATRIX_OPT)
- MNK = (double) args.m * (double) args.n * (double) args.k;
-#endif
-
#ifdef SMALL_MATRIX_OPT
#if !defined(COMPLEX)
if(GEMM_SMALL_MATRIX_PERMIT(transa, transb, args.m, args.n, args.k, *(FLOAT *)(args.alpha), *(FLOAT *)(args.beta))){
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);
+ (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);
+ (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);
}
return;
}
#else
if(GEMM_SMALL_MATRIX_PERMIT(transa, transb, args.m, args.n, args.k, alpha[0], alpha[1], beta[0], beta[1])){
if(beta[0] == 0.0 && beta[1] == 0.0){
- (zgemm_small_kernel_b0[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, args.c, args.ldc);
+ (ZGEMM_SMALL_KERNEL_B0((transb << 2) | transa))(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, args.c, args.ldc);
}else{
- (zgemm_small_kernel[(transb << 2) | transa])(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, beta[0], beta[1], args.c, args.ldc);
+ (ZGEMM_SMALL_KERNEL((transb << 2) | transa))(args.m, args.n, args.k, args.a, args.lda, alpha[0], alpha[1], args.b, args.ldb, beta[0], beta[1], args.c, args.ldc);
}
return;
}
#endif
#endif
-
buffer = (XFLOAT *)blas_memory_alloc(0);
mode |= (transb << BLAS_TRANSB_SHIFT);
#endif
-
+ MNK = (double) args.m * (double) args.n * (double) args.k;
if ( MNK <= (SMP_THRESHOLD_MIN * (double) GEMM_MULTITHREAD_THRESHOLD) )
args.nthreads = 1;
else