add CBLAS interface for s/d/c/zimatcopy

cblas.h

@@ -327,6 +327,14 @@ void cblas_comatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum
 void cblas_zomatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double* calpha, OPENBLAS_CONST double* a, 
 		     OPENBLAS_CONST blasint clda,  double *b, OPENBLAS_CONST blasint cldb); 
 
+void cblas_simatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float calpha, float *a, 
+		     OPENBLAS_CONST blasint clda, OPENBLAS_CONST blasint cldb); 
+void cblas_dimatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double calpha, double *a,
+		     OPENBLAS_CONST blasint clda, OPENBLAS_CONST blasint cldb); 
+void cblas_cimatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST float* calpha, float* a, 
+		     OPENBLAS_CONST blasint clda, OPENBLAS_CONST blasint cldb); 
+void cblas_zimatcopy(OPENBLAS_CONST enum CBLAS_ORDER CORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE CTRANS, OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double* calpha, double* a, 
+		     OPENBLAS_CONST blasint clda, OPENBLAS_CONST blasint cldb); 
 
 #ifdef __cplusplus
 }

cblas_noconst.h

@@ -315,6 +315,14 @@ void cblas_comatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  bl
 void cblas_zomatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  void* calpha,  void* a, 
 		      blasint clda,  void *b,  blasint cldb); 
 
+void cblas_simatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  float calpha, float *a, 
+		      blasint clda,  blasint cldb); 
+void cblas_dimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  double calpha, double *a,
+		      blasint clda,  blasint cldb); 
+void cblas_cimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  float* calpha, float* a, 
+		      blasint clda,  blasint cldb); 
+void cblas_zimatcopy( enum CBLAS_ORDER CORDER,  enum CBLAS_TRANSPOSE CTRANS,  blasint crows,  blasint ccols,  double* calpha, double* a, 
+		      blasint clda,  blasint cldb); 
 #ifdef __cplusplus
 }
 #endif  /* __cplusplus */

exports/gensymbol

@@ -53,7 +53,8 @@
 	       cblas_ztbmv, cblas_ztbsv, cblas_ztpmv, cblas_ztpsv, cblas_ztrmm, cblas_ztrmv, cblas_ztrsm,
 	       cblas_ztrsv, cblas_cdotc_sub, cblas_cdotu_sub, cblas_zdotc_sub, cblas_zdotu_sub,
                cblas_saxpby,cblas_daxpby,cblas_caxpby,cblas_zaxpby,
-	       cblas_somatcopy, cblas_domatcopy, cblas_comatcopy, cblas_zomatcopy
+	       cblas_somatcopy, cblas_domatcopy, cblas_comatcopy, cblas_zomatcopy, 
+	       cblas_simatcopy, cblas_dimatcopy, cblas_cimatcopy, cblas_zimatcopy
                );
 
 @exblasobjs = (

interface/Makefile

@@ -267,7 +267,7 @@ CSBLAS2OBJS   = \
 
 CSBLAS3OBJS   = \
 	cblas_sgemm.$(SUFFIX) cblas_ssymm.$(SUFFIX) cblas_strmm.$(SUFFIX) cblas_strsm.$(SUFFIX) \
-	cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX) 
+	cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX)  cblas_simatcopy.$(SUFFIX) 
 
 CDBLAS1OBJS   = \
 	cblas_idamax.$(SUFFIX) cblas_dasum.$(SUFFIX) cblas_daxpy.$(SUFFIX) \
@@ -283,7 +283,7 @@ CDBLAS2OBJS   = \
 
 CDBLAS3OBJS   += \
 	cblas_dgemm.$(SUFFIX) cblas_dsymm.$(SUFFIX) cblas_dtrmm.$(SUFFIX) cblas_dtrsm.$(SUFFIX) \
-	cblas_dsyrk.$(SUFFIX) cblas_dsyr2k.$(SUFFIX) cblas_domatcopy.$(SUFFIX) 
+	cblas_dsyrk.$(SUFFIX) cblas_dsyr2k.$(SUFFIX) cblas_domatcopy.$(SUFFIX)  cblas_dimatcopy.$(SUFFIX) 
 
 CCBLAS1OBJS   = \
 	cblas_icamax.$(SUFFIX) cblas_scasum.$(SUFFIX)  cblas_caxpy.$(SUFFIX) \
@@ -306,7 +306,8 @@ CCBLAS3OBJS   = \
 	cblas_cgemm.$(SUFFIX) cblas_csymm.$(SUFFIX) cblas_ctrmm.$(SUFFIX) cblas_ctrsm.$(SUFFIX) \
 	cblas_csyrk.$(SUFFIX) cblas_csyr2k.$(SUFFIX) \
 	cblas_chemm.$(SUFFIX) cblas_cherk.$(SUFFIX) cblas_cher2k.$(SUFFIX) \
-	cblas_comatcopy.$(SUFFIX)
+	cblas_comatcopy.$(SUFFIX) cblas_cimatcopy.$(SUFFIX) 
+
 
 CZBLAS1OBJS   = \
 	cblas_izamax.$(SUFFIX) cblas_dzasum.$(SUFFIX)  cblas_zaxpy.$(SUFFIX) \
@@ -329,7 +330,8 @@ CZBLAS3OBJS   = \
 	cblas_zgemm.$(SUFFIX) cblas_zsymm.$(SUFFIX) cblas_ztrmm.$(SUFFIX) cblas_ztrsm.$(SUFFIX) \
 	cblas_zsyrk.$(SUFFIX) cblas_zsyr2k.$(SUFFIX) \
 	cblas_zhemm.$(SUFFIX) cblas_zherk.$(SUFFIX) cblas_zher2k.$(SUFFIX)\
-	cblas_zomatcopy.$(SUFFIX) 
+	cblas_zomatcopy.$(SUFFIX) cblas_zimatcopy.$(SUFFIX) 
+
 
 ifndef NO_CBLAS
 
@@ -2061,13 +2063,25 @@ cblas_zomatcopy.$(SUFFIX) cblas_zomatcopy.$(PSUFFIX) : zomatcopy.c
 dimatcopy.$(SUFFIX) dimatcopy.$(PSUFFIX) : imatcopy.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+cblas_dimatcopy.$(SUFFIX) cblas_dimatcopy.$(PSUFFIX) : imatcopy.c
+	$(CC) -c $(CFLAGS) -DCBLAS $< -o $(@F)
+
 simatcopy.$(SUFFIX) simatcopy.$(PSUFFIX) : imatcopy.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+cblas_simatcopy.$(SUFFIX) cblas_simatcopy.$(PSUFFIX) : imatcopy.c
+	$(CC) -c $(CFLAGS) -DCBLAS $< -o $(@F)
+
 cimatcopy.$(SUFFIX) cimatcopy.$(PSUFFIX) : zimatcopy.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+cblas_cimatcopy.$(SUFFIX) cblas_cimatcopy.$(PSUFFIX) : zimatcopy.c
+	$(CC) -c $(CFLAGS) -DCBLAS $< -o $(@F)
+
 zimatcopy.$(SUFFIX) zimatcopy.$(PSUFFIX) : zimatcopy.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+cblas_zimatcopy.$(SUFFIX) cblas_zimatcopy.$(PSUFFIX) : zimatcopy.c
+	$(CC) -c $(CFLAGS) -DCBLAS $< -o $(@F)
+
 

interface/imatcopy.c

@@ -50,6 +50,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #undef malloc
 #undef free
 
+#ifndef CBLAS
 void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, FLOAT *a, blasint *lda, blasint *ldb)
 {
 
@@ -71,6 +72,28 @@ void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha,
 	if ( Trans == 'R' ) trans = BlasNoTrans;
 	if ( Trans == 'T' ) trans = BlasTrans;
 	if ( Trans == 'C' ) trans = BlasTrans;
+#else 
+void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows, blasint ccols, FLOAT calpha, FLOAT *a, blasint clda, blasint cldb)
+{
+	char Order, Trans;
+	int order=-1,trans=-1;
+	blasint info = -1;
+	FLOAT *b;
+	size_t msize;
+	blasint *lda, *ldb, *rows, *cols; 
+	FLOAT *alpha; 
+
+	if ( CORDER == CblasColMajor) order = BlasColMajor; 
+	if ( CORDER == CblasRowMajor) order = BlasRowMajor; 
+	if ( CTRANS == CblasNoTrans || CTRANS == CblasConjNoTrans) trans = BlasNoTrans; 
+	if ( CTRANS == CblasTrans   || CTRANS == CblasConjTrans  ) trans = BlasTrans; 
+
+	rows = &crows; 
+	cols = &ccols; 
+	alpha = &calpha; 
+	lda = &clda; 
+	ldb = &cldb; 	
+#endif 
 
 	if ( order == BlasColMajor)
 	{

interface/zimatcopy.c

@@ -49,6 +49,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define BlasTransConj    2
 #define BlasConj         3
 
+
+#ifndef CBLAS
 void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, FLOAT *a, blasint *lda, blasint *ldb)
 {
 
@@ -71,6 +73,30 @@ void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha,
 	if ( Trans == 'C' ) trans = BlasTransConj;
 	if ( Trans == 'R' ) trans = BlasConj;
 
+#else 
+void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows, blasint ccols, FLOAT *alpha, FLOAT *a, blasint clda, blasint cldb)
+{
+
+	blasint *rows, *cols, *lda, *ldb; 
+	int order=-1,trans=-1;
+	blasint info = -1;
+	FLOAT *b;
+	size_t msize;
+
+	if ( CORDER == CblasColMajor ) order = BlasColMajor; 
+	if ( CORDER == CblasRowMajor ) order = BlasRowMajor; 
+
+	if ( CTRANS == CblasNoTrans) trans = BlasNoTrans; 
+	if ( CTRANS == CblasConjNoTrans ) trans = BlasConj; 
+	if ( CTRANS == CblasTrans) trans = BlasTrans; 
+	if ( CTRANS == CblasConjTrans) trans = BlasTransConj; 
+
+	rows = &crows; 
+	cols = &ccols; 
+	lda  = &clda; 
+	ldb  = &cldb; 
+#endif
+
 	if ( order == BlasColMajor)
 	{
         	if ( trans == BlasNoTrans      &&  *ldb < *rows ) info = 9;