add cblas_?gemm_batch

cblas.h

@@ -416,6 +416,18 @@ void cblas_cgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint
 void cblas_zgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double *calpha, double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double *cbeta, 
 		  double *c, OPENBLAS_CONST blasint cldc); 
 
+void cblas_sgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,
+		       OPENBLAS_CONST float * alpha_array, OPENBLAS_CONST float ** A_array, OPENBLAS_CONST blasint * lda_array, OPENBLAS_CONST float ** B_array, OPENBLAS_CONST blasint * ldb_array, OPENBLAS_CONST float * beta_array, float ** C_array, OPENBLAS_CONST blasint * ldc_array, OPENBLAS_CONST blasint group_count, OPENBLAS_CONST blasint * group_size);
+
+void cblas_dgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,
+		       OPENBLAS_CONST double * alpha_array, OPENBLAS_CONST double ** A_array, OPENBLAS_CONST blasint * lda_array, OPENBLAS_CONST double ** B_array, OPENBLAS_CONST blasint * ldb_array, OPENBLAS_CONST double * beta_array, double ** C_array, OPENBLAS_CONST blasint * ldc_array, OPENBLAS_CONST blasint group_count, OPENBLAS_CONST blasint * group_size);
+
+void cblas_cgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,
+		       OPENBLAS_CONST void * alpha_array, OPENBLAS_CONST void ** A_array, OPENBLAS_CONST blasint * lda_array, OPENBLAS_CONST void ** B_array, OPENBLAS_CONST blasint * ldb_array, OPENBLAS_CONST void * beta_array, void ** C_array, OPENBLAS_CONST blasint * ldc_array, OPENBLAS_CONST blasint group_count, OPENBLAS_CONST blasint * group_size);
+
+void cblas_zgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,
+		       OPENBLAS_CONST void * alpha_array, OPENBLAS_CONST void ** A_array, OPENBLAS_CONST blasint * lda_array, OPENBLAS_CONST void ** B_array, OPENBLAS_CONST blasint * ldb_array, OPENBLAS_CONST void * beta_array, void ** C_array, OPENBLAS_CONST blasint * ldc_array, OPENBLAS_CONST blasint group_count, OPENBLAS_CONST blasint * group_size);
+
 /*** BFLOAT16 and INT8 extensions ***/
 /* convert float array to BFLOAT16 array by rounding */
 void   cblas_sbstobf16(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *in, OPENBLAS_CONST blasint incin, bfloat16 *out, OPENBLAS_CONST blasint incout);
@@ -431,6 +443,9 @@ void   cblas_sbgemv(OPENBLAS_CONST enum CBLAS_ORDER order,  OPENBLAS_CONST enum
 
 void   cblas_sbgemm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint N, OPENBLAS_CONST blasint K,
 		    OPENBLAS_CONST float alpha, OPENBLAS_CONST bfloat16 *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST bfloat16 *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST blasint ldc);
+void cblas_sbgemm_batch(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransA_array, OPENBLAS_CONST enum CBLAS_TRANSPOSE * TransB_array, OPENBLAS_CONST blasint * M_array, OPENBLAS_CONST blasint * N_array, OPENBLAS_CONST blasint * K_array,
+		       OPENBLAS_CONST float * alpha_array, OPENBLAS_CONST bfloat16 ** A_array, OPENBLAS_CONST blasint * lda_array, OPENBLAS_CONST bfloat16 ** B_array, OPENBLAS_CONST blasint * ldb_array, OPENBLAS_CONST float * beta_array, float ** C_array, OPENBLAS_CONST blasint * ldc_array, OPENBLAS_CONST blasint group_count, OPENBLAS_CONST blasint * group_size);
+
 #ifdef __cplusplus
 }
 #endif  /* __cplusplus */

common_level3.h

@@ -1937,8 +1937,13 @@ int zimatcopy_k_rtc(BLASLONG, BLASLONG,  double, double, double *, BLASLONG);
 int sgeadd_k(BLASLONG, BLASLONG, float, float*, BLASLONG, float, float *, BLASLONG); 
 int dgeadd_k(BLASLONG, BLASLONG, double, double*, BLASLONG, double, double *, BLASLONG); 
 int cgeadd_k(BLASLONG, BLASLONG, float, float, float*, BLASLONG, float, float, float *, BLASLONG); 
-int zgeadd_k(BLASLONG, BLASLONG, double,double, double*, BLASLONG, double, double, double *, BLASLONG); 
+int zgeadd_k(BLASLONG, BLASLONG, double,double, double*, BLASLONG, double, double, double *, BLASLONG);
 
+int sgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
+int dgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
+int cgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
+int zgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
+int sbgemm_batch_thread(blas_arg_t * queue, BLASLONG nums);
 
 #ifdef __CUDACC__
 }

common_macro.h

@@ -2655,9 +2655,20 @@ typedef struct {
   BLASLONG prea, preb, prec, pred;
 #endif
 
+
+  //for gemm_batch
+  void * routine;
+  int routine_mode;
+
 } blas_arg_t;
 #endif
 
+#ifdef SMALL_MATRIX_OPT
+#define BLAS_SMALL_OPT  0x10000U
+#define BLAS_SMALL_B0_OPT  0x30000U
+#endif
+
+
 #ifdef XDOUBLE
 
 #define TRSV_NUU qtrsv_NUU