Merge pull request #4452 from kseniyazaytseva/riscv-generic

Fix BLAS, BLAS-like functions and Generic RISC-V kernels

cblas.h

@@ -289,6 +289,14 @@ void cblas_zgemm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLA
 void cblas_zgemm3m(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 void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
 
+void cblas_sgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST float alpha, OPENBLAS_CONST float *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST float *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST blasint ldc);
+void cblas_dgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST double alpha, OPENBLAS_CONST double *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST double *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST double beta, double *C, OPENBLAS_CONST blasint ldc);
+void cblas_cgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
+void cblas_zgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
 
 void cblas_ssymm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_SIDE Side, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint N,
                  OPENBLAS_CONST float alpha, OPENBLAS_CONST float *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST float *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST blasint ldc);

common_interface.h

@@ -498,6 +498,15 @@ void BLASFUNC(zgemm3m)(char *, char *, blasint *, blasint *, blasint *, double *
 void BLASFUNC(xgemm3m)(char *, char *, blasint *, blasint *, blasint *, xdouble *,
 	   xdouble *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *);
 
+void BLASFUNC(sgemmt)(char*, char *, char *, blasint *, blasint *, float *,
+	   float  *, blasint *, float  *, blasint *, float  *, float  *, blasint *);
+void BLASFUNC(dgemmt)(char*, char *, char *, blasint *, blasint *, double *,
+	   double *, blasint *, double *, blasint *, double *, double *, blasint *);
+void BLASFUNC(cgemmt)(char*, char *, char *, blasint *, blasint *, float *,
+	   float  *, blasint *, float  *, blasint *, float  *, float  *, blasint *);
+void BLASFUNC(zgemmt)(char*, char *, char *, blasint *, blasint *, double *,
+	   double *, blasint *, double *, blasint *, double *, double *, blasint *);
+
 int BLASFUNC(sge2mm)(char *, char *, char *, blasint *, blasint *,
 		     float *, float  *, blasint *, float  *, blasint *,
 		     float *, float  *, blasint *);

ctest/c_cblat1.f

@@ -96,7 +96,7 @@
       INTEGER           ICAMAXTEST
       EXTERNAL          SCASUMTEST, SCNRM2TEST, ICAMAXTEST
 *     .. External Subroutines ..
-      EXTERNAL          CSCAL, CSSCALTEST, CTEST, ITEST1, STEST1
+      EXTERNAL          CSCALTEST, CSSCALTEST, CTEST, ITEST1, STEST1
 *     .. Intrinsic Functions ..
       INTRINSIC         MAX
 *     .. Common blocks ..
@@ -214,8 +214,8 @@
                CALL STEST1(SCASUMTEST(N,CX,INCX),STRUE4(NP1),
      +                     STRUE4(NP1),SFAC)
             ELSE IF (ICASE.EQ.8) THEN
-*              .. CSCAL ..
-               CALL CSCAL(N,CA,CX,INCX)
+*              .. CSCALTEST ..
+               CALL CSCALTEST(N,CA,CX,INCX)
                CALL CTEST(LEN,CX,CTRUE5(1,NP1,INCX),CTRUE5(1,NP1,INCX),
      +                    SFAC)
             ELSE IF (ICASE.EQ.9) THEN
@@ -236,14 +236,14 @@
 *
       INCX = 1
       IF (ICASE.EQ.8) THEN
-*        CSCAL
+*        CSCALTEST
 *        Add a test for alpha equal to zero.
          CA = (0.0E0,0.0E0)
          DO 80 I = 1, 5
             MWPCT(I) = (0.0E0,0.0E0)
             MWPCS(I) = (1.0E0,1.0E0)
    80    CONTINUE
-         CALL CSCAL(5,CA,CX,INCX)
+         CALL CSCALTEST(5,CA,CX,INCX)
          CALL CTEST(5,CX,MWPCT,MWPCS,SFAC)
       ELSE IF (ICASE.EQ.9) THEN
 *        CSSCALTEST

ctest/c_cblat1c.c

@@ -685,7 +685,7 @@ real *sfac;
     static integer i__;
     extern /* Subroutine */ int ctest_();
     static complex mwpcs[5], mwpct[5];
-    extern /* Subroutine */ int itest1_(), stest1_();
+    extern /* Subroutine */ int cscaltest_(), itest1_(), stest1_();
     static complex cx[8];
     extern real scnrm2test_();
     static integer np1;
@@ -727,7 +727,7 @@ real *sfac;
 		stest1_(&r__1, &strue4[np1 - 1], &strue4[np1 - 1], sfac);
 	    } else if (combla_1.icase == 8) {
 /*              .. CSCAL .. */
-		cscal_(&combla_1.n, &ca, cx, &combla_1.incx);
+		cscaltest_(&combla_1.n, &ca, cx, &combla_1.incx);
 		ctest_(&len, cx, &ctrue5[(np1 + combla_1.incx * 5 << 3) - 48],
 			 &ctrue5[(np1 + combla_1.incx * 5 << 3) - 48], sfac);
 	    } else if (combla_1.icase == 9) {
@@ -761,7 +761,7 @@ real *sfac;
 	    mwpcs[i__1].r = (float)1., mwpcs[i__1].i = (float)1.;
 /* L80: */
 	}
-	cscal_(&c__5, &ca, cx, &combla_1.incx);
+	cscaltest_(&c__5, &ca, cx, &combla_1.incx);
 	ctest_(&c__5, cx, mwpct, mwpcs, sfac);
     } else if (combla_1.icase == 9) {
 /*        CSSCALTEST */

interface/gemmt.c

@@ -35,29 +35,26 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include "common.h"
-#ifdef FUNCTION_PROFILE
-#include "functable.h"
-#endif
 
 #ifndef COMPLEX
 #define SMP_THRESHOLD_MIN 65536.0
 #ifdef XDOUBLE
-#define ERROR_NAME "QGEMT "
+#define ERROR_NAME "QGEMMT "
 #elif defined(DOUBLE)
-#define ERROR_NAME "DGEMT "
+#define ERROR_NAME "DGEMMT "
 #elif defined(BFLOAT16)
-#define ERROR_NAME "SBGEMT "
+#define ERROR_NAME "SBGEMMT "
 #else
-#define ERROR_NAME "SGEMT "
+#define ERROR_NAME "SGEMMT "
 #endif
 #else
 #define SMP_THRESHOLD_MIN 8192.0
 #ifdef XDOUBLE
-#define ERROR_NAME "XGEMT "
+#define ERROR_NAME "XGEMMT "
 #elif defined(DOUBLE)
-#define ERROR_NAME "ZGEMT "
+#define ERROR_NAME "ZGEMMT "
 #else
-#define ERROR_NAME "CGEMT "
+#define ERROR_NAME "CGEMMT "
 #endif
 #endif
 
@@ -68,18 +65,22 @@
 #ifndef CBLAS
 
 void NAME(char *UPLO, char *TRANSA, char *TRANSB,
-	  blasint * M, blasint * N, blasint * K,
+	  blasint * M, blasint * K,
 	  FLOAT * Alpha,
 	  IFLOAT * a, blasint * ldA,
 	  IFLOAT * b, blasint * ldB, FLOAT * Beta, FLOAT * c, blasint * ldC)
 {
 
-	blasint m, n, k;
+	blasint m, k;
 	blasint lda, ldb, ldc;
 	int transa, transb, uplo;
 	blasint info;
 
 	char transA, transB, Uplo;
+	blasint nrowa, nrowb;
+#if defined(COMPLEX)
+	blasint ncolb;
+#endif
 	IFLOAT *buffer;
 	IFLOAT *aa, *bb;
 	FLOAT *cc;
@@ -92,7 +93,6 @@ void NAME(char *UPLO, char *TRANSA, char *TRANSB,
 	PRINT_DEBUG_NAME;
 
 	m = *M;
-	n = *N;
 	k = *K;
 
 #if defined(COMPLEX)
@@ -159,32 +159,47 @@ void NAME(char *UPLO, char *TRANSA, char *TRANSB,
 	if (Uplo == 'L')
 		uplo = 1;
 
+	nrowa = m;
+	if (transa & 1) nrowa = k;
+	nrowb = k;
+#if defined(COMPLEX)
+	ncolb = m;
+#endif
+	if (transb & 1) {
+		nrowb = m;
+#if defined(COMPLEX)
+		ncolb = k;
+#endif
+	}
+
 	info = 0;
 
-	if (uplo < 0)
-		info = 14;
-	if (ldc < m)
+	if (ldc < MAX(1, m))
 		info = 13;
+	if (ldb < MAX(1, nrowb))
+		info = 10;
+	if (lda < MAX(1, nrowa))
+		info = 8;
 	if (k < 0)
 		info = 5;
-	if (n < 0)
-		info = 4;
 	if (m < 0)
-		info = 3;
+		info = 4;
 	if (transb < 0)
-		info = 2;
+		info = 3;
 	if (transa < 0)
+		info = 2;
+	if (uplo < 0)
 		info = 1;
 
-	if (info) {
+	if (info != 0) {
 		BLASFUNC(xerbla) (ERROR_NAME, &info, sizeof(ERROR_NAME));
 		return;
 	}
 #else
 
 void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
-	   enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint M,
-	   blasint N, blasint k,
+	   enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, blasint m,
+	   blasint k,
 #ifndef COMPLEX
 	   FLOAT alpha,
 	   IFLOAT * A, blasint LDA,
@@ -205,17 +220,23 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
 	int transa, transb, uplo;
 	blasint info;
-	blasint m, n, lda, ldb;
+	blasint lda, ldb;
 	FLOAT *a, *b;
+#if defined(COMPLEX)
+	blasint nrowb, ncolb;
+#endif
 	XFLOAT *buffer;
 
 	PRINT_DEBUG_CNAME;
 
+	uplo = -1;
 	transa = -1;
 	transb = -1;
 	info = 0;
 
 	if (order == CblasColMajor) {
+		if (Uplo == CblasUpper) uplo = 0;
+		if (Uplo == CblasLower) uplo = 1;
 
 		if (TransA == CblasNoTrans)
 			transa = 0;
@@ -248,9 +269,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 			transb = 3;
 #endif
 
-		m = M;
-		n = N;
-
 		a = (void *)A;
 		b = (void *)B;
 		lda = LDA;
@@ -258,23 +276,42 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
 		info = -1;
 
-		if (ldc < m)
+		blasint nrowa;
+#if !defined(COMPLEX)
+		blasint nrowb;
+#endif
+		nrowa = m;
+		if (transa & 1) nrowa = k;
+		nrowb = k;
+#if defined(COMPLEX)
+		ncolb = m;
+#endif
+		if (transb & 1) {
+			nrowb = m;
+#if defined(COMPLEX)
+			ncolb = k;
+#endif
+		}
+
+		if (ldc < MAX(1, m))
 			info = 13;
+		if (ldb < MAX(1, nrowb))
+			info = 10;
+		if (lda < MAX(1, nrowa))
+			info = 8;
 		if (k < 0)
 			info = 5;
-		if (n < 0)
-			info = 4;
 		if (m < 0)
-			info = 3;
+			info = 4;
 		if (transb < 0)
-			info = 2;
+			info = 3;
 		if (transa < 0)
+			info = 2;
+		if (uplo < 0)
 			info = 1;
 	}
 
 	if (order == CblasRowMajor) {
-		m = N;
-		n = M;
 
 		a = (void *)B;
 		b = (void *)A;
@@ -282,6 +319,9 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 		lda = LDB;
 		ldb = LDA;
 
+		if (Uplo == CblasUpper) uplo = 0;
+		if (Uplo == CblasLower) uplo = 1;
+
 		if (TransB == CblasNoTrans)
 			transa = 0;
 		if (TransB == CblasTrans)
@@ -315,29 +355,42 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
 		info = -1;
 
-		if (ldc < m)
+		blasint ncola; 
+#if !defined(COMPLEX)
+		blasint ncolb;
+#endif
+		ncola = m;
+		if (transa & 1) ncola = k;
+		ncolb = k;
+#if defined(COMPLEX)
+		nrowb = m;
+#endif
+
+		if (transb & 1) {
+#if defined(COMPLEX)
+			nrowb = k;
+#endif
+			ncolb = m;
+		}
+
+		if (ldc < MAX(1,m))
 			info = 13;
+		if (ldb < MAX(1, ncolb))
+			info = 8;
+		if (lda < MAX(1, ncola))
+			info = 10;
 		if (k < 0)
 			info = 5;
-		if (n < 0)
-			info = 4;
 		if (m < 0)
-			info = 3;
+			info = 4;
 		if (transb < 0)
 			info = 2;
 		if (transa < 0)
+			info = 3;
+		if (uplo < 0)
 			info = 1;
-
 	}
 
-	uplo = -1;
-	if (Uplo == CblasUpper)
-		uplo = 0;
-	if (Uplo == CblasLower)
-		uplo = 1;
-	if (uplo < 0)
-		info = 14;
-
 	if (info >= 0) {
 		BLASFUNC(xerbla) (ERROR_NAME, &info, sizeof(ERROR_NAME));
 		return;
@@ -407,37 +460,48 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
 #endif
 
-	if ((m == 0) || (n == 0))
+	if (m == 0)
 		return;
 
 	IDEBUG_START;
 
-	FUNCTION_PROFILE_START();
+#if defined(COMPLEX)
+	if (transb > 1){
+#ifndef CBLAS
+		IMATCOPY_K_CNC(nrowb, ncolb, (FLOAT)(1.0), (FLOAT)(0.0), b, ldb);
+#else
+		if (order == CblasColMajor)
+			IMATCOPY_K_CNC(nrowb, ncolb, (FLOAT)(1.0), (FLOAT)(0.0), b, ldb);
+		if (order == CblasRowMajor)
+			IMATCOPY_K_RNC(nrowb, ncolb, (FLOAT)(1.0), (FLOAT)(0.0), b, ldb);
+#endif
+	}
+#endif
 
-	const blasint incb = (transb == 0) ? 1 : ldb;
+	const blasint incb = ((transb & 1) == 0) ? 1 : ldb;
 
 	if (uplo == 1) {
-		for (i = 0; i < n; i++) {
-			j = n - i;
+		for (i = 0; i < m; i++) {
+			j = m - i;
 
 			l = j;
 #if defined(COMPLEX)
 			aa = a + i * 2;
 			bb = b + i * ldb * 2;
-			if (transa) {
-				l = k;
+			if (transa & 1) {
 				aa = a + lda * i * 2;
-				bb = b + i * 2;
 			}
+			if (transb & 1)
+				bb = b + i * 2;
 			cc = c + i * 2 * ldc + i * 2;
 #else
 			aa = a + i;
 			bb = b + i * ldb;
-			if (transa) {
-				l = k;
+			if (transa & 1) {
 				aa = a + lda * i;
-				bb = b + i;
 			}
+			if (transb & 1)
+				bb = b + i;
 			cc = c + i * ldc + i;
 #endif
 
@@ -447,7 +511,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 				       NULL, 0);
 
 			if (alpha_r == ZERO && alpha_i == ZERO)
-				return;
+				continue;
 #else
 			if (beta != ONE)
 				SCAL_K(l, 0, 0, beta, cc, 1, NULL, 0, NULL, 0);
@@ -458,8 +522,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
 			IDEBUG_START;
 
-			FUNCTION_PROFILE_START();
-
 			buffer_size = j + k + 128 / sizeof(FLOAT);
 #ifdef WINDOWS_ABI
 			buffer_size += 160 / sizeof(FLOAT);
@@ -479,20 +541,34 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 #endif
 
 #if defined(COMPLEX)
+				if (!(transa & 1))
 				(gemv[(int)transa]) (j, k, 0, alpha_r, alpha_i,
 						     aa, lda, bb, incb, cc, 1,
 						     buffer);
+				else
+				(gemv[(int)transa]) (k, j, 0, alpha_r, alpha_i,
+						     aa, lda, bb, incb, cc, 1,
+						     buffer);
 #else
+				if (!(transa & 1))
 				(gemv[(int)transa]) (j, k, 0, alpha, aa, lda,
 						     bb, incb, cc, 1, buffer);
+				else
+				(gemv[(int)transa]) (k, j, 0, alpha, aa, lda,
+						     bb, incb, cc, 1, buffer);
 #endif
 #ifdef SMP
 			} else {
-
+				if (!(transa & 1))
 				(gemv_thread[(int)transa]) (j, k, alpha, aa,
 							    lda, bb, incb, cc,
 							    1, buffer,
 							    nthreads);
+				else
+				(gemv_thread[(int)transa]) (k, j, alpha, aa,
+							    lda, bb, incb, cc,
+							    1, buffer,
+							    nthreads);
 
 			}
 #endif
@@ -501,21 +577,19 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 		}
 	} else {
 
-		for (i = 0; i < n; i++) {
+		for (i = 0; i < m; i++) {
 			j = i + 1;
 
 			l = j;
 #if defined COMPLEX
 			bb = b + i * ldb * 2;
-			if (transa) {
-				l = k;
+			if (transb & 1) {
 				bb = b + i * 2;
 			}
 			cc = c + i * 2 * ldc;
 #else
 			bb = b + i * ldb;
-			if (transa) {
-				l = k;
+			if (transb & 1) {
 				bb = b + i;
 			}
 			cc = c + i * ldc;
@@ -527,7 +601,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 				       NULL, 0);
 
 			if (alpha_r == ZERO && alpha_i == ZERO)
-				return;
+				continue;
 #else
 			if (beta != ONE)
 				SCAL_K(l, 0, 0, beta, cc, 1, NULL, 0, NULL, 0);
@@ -537,8 +611,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 #endif
 			IDEBUG_START;
 
-			FUNCTION_PROFILE_START();
-
 			buffer_size = j + k + 128 / sizeof(FLOAT);
 #ifdef WINDOWS_ABI
 			buffer_size += 160 / sizeof(FLOAT);
@@ -558,32 +630,41 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 #endif
 
 #if defined(COMPLEX)
+				if (!(transa & 1))
 				(gemv[(int)transa]) (j, k, 0, alpha_r, alpha_i,
 						     a, lda, bb, incb, cc, 1,
 						     buffer);
+				else
+				(gemv[(int)transa]) (k, j, 0, alpha_r, alpha_i,
+						     a, lda, bb, incb, cc, 1,
+						     buffer);
 #else
+				if (!(transa & 1))
 				(gemv[(int)transa]) (j, k, 0, alpha, a, lda, bb,
 						     incb, cc, 1, buffer);
+				else
+				(gemv[(int)transa]) (k, j, 0, alpha, a, lda, bb,
+						     incb, cc, 1, buffer);
 #endif
 
 #ifdef SMP
 			} else {
-
+				if (!(transa & 1))
 				(gemv_thread[(int)transa]) (j, k, alpha, a, lda,
 							    bb, incb, cc, 1,
 							    buffer, nthreads);
-
+				else
+				(gemv_thread[(int)transa]) (k, j, alpha, a, lda,
+							    bb, incb, cc, 1,
+							    buffer, nthreads);
 			}
 #endif
 
 			STACK_FREE(buffer);
 		}
 	}
-	FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE,
-			     args.m * args.k + args.k * args.n +
-			     args.m * args.n, 2 * args.m * args.n * args.k);
 
 	IDEBUG_END;
 
 	return;
-}
+}

interface/imatcopy.c

@@ -149,10 +149,10 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
 
 #endif
 
-	if ( *lda >  *ldb )
-		msize = (size_t)(*lda) * (*ldb)  * sizeof(FLOAT);
-	else
-		msize = (size_t)(*ldb) * (*ldb)  * sizeof(FLOAT);
+	if ( *rows >  *cols )
+            msize = (size_t)(*rows) * (*ldb)  * sizeof(FLOAT);
+    else
+            msize = (size_t)(*cols) * (*ldb)  * sizeof(FLOAT);
 
 	b = malloc(msize);
 	if ( b == NULL )

interface/rotmg.c

@@ -96,12 +96,6 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 	else
 	{
 		dp2 = *dd2 * dy1;
-		if(dp2 == ZERO)
-		{
-			dflag = -TWO;
-			dparam[0] = dflag;
-			return;
-		}
 		dp1 = *dd1 * *dx1;
 		dq2 =  dp2 * dy1;
 		dq1 =  dp1 * *dx1;
@@ -113,24 +107,10 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 			dh12 =    dp2 /  dp1;
 
 			du   = ONE - dh12 * dh21;
-			if(du > ZERO)
-			{
-				dflag = ZERO;
-				*dd1  = *dd1 / du;
-				*dd2  = *dd2 / du;
-				*dx1  = *dx1 * du;
-			} else {
-				dflag = -ONE;
-
-				dh11  = ZERO;
-				dh12  = ZERO;
-				dh21  = ZERO;
-				dh22  = ZERO;
-
-				*dd1  = ZERO;
-				*dd2  = ZERO;
-				*dx1  = ZERO;
-			}
+			dflag = ZERO;
+			*dd1  = *dd1 / du;
+			*dd2  = *dd2 / du;
+			*dx1  = *dx1 * du;
 			
 		}
 		else

interface/zimatcopy.c

@@ -171,10 +171,10 @@ void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows,
     }
 #endif
 
-	if ( *lda >  *ldb )
-                msize = (size_t)(*lda) * (*ldb)  * sizeof(FLOAT) * 2;
+		if ( *rows >  *cols )
+                msize = (size_t)(*rows) * (*ldb)  * sizeof(FLOAT) * 2;
         else
-                msize = (size_t)(*ldb) * (*ldb)  * sizeof(FLOAT) * 2;
+                msize = (size_t)(*cols) * (*ldb)  * sizeof(FLOAT) * 2;
 
         b = malloc(msize);
         if ( b == NULL )

kernel/generic/zimatcopy_cnc.c

@@ -40,7 +40,6 @@ int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a,
 
 	if ( rows <= 0     )  return(0);
 	if ( cols <= 0     )  return(0);
-    if ( alpha_r == 1.0 && alpha_i == 0.0 ) return (0); 
 
 	aptr = a;
 	lda *= 2;

kernel/riscv64/KERNEL.RISCV64_GENERIC

@@ -45,6 +45,11 @@ DAXPYKERNEL  = ../riscv64/axpy.c
 CAXPYKERNEL  = ../riscv64/zaxpy.c
 ZAXPYKERNEL  = ../riscv64/zaxpy.c
 
+SAXPBYKERNEL  = ../riscv64/axpby.c
+DAXPBYKERNEL  = ../riscv64/axpby.c
+CAXPBYKERNEL  = ../riscv64/zaxpby.c
+ZAXPBYKERNEL  = ../riscv64/zaxpby.c
+
 SCOPYKERNEL  = ../riscv64/copy.c
 DCOPYKERNEL  = ../riscv64/copy.c
 CCOPYKERNEL  = ../riscv64/zcopy.c

kernel/riscv64/axpby.c

@@ -33,7 +33,7 @@ int CNAME(BLASLONG n, FLOAT alpha, FLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *
 	BLASLONG i=0;
 	BLASLONG ix,iy;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	ix = 0;
 	iy = 0;

kernel/riscv64/axpy.c

@@ -42,7 +42,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
 	BLASLONG i=0;
 	BLASLONG ix,iy;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 	if ( da == 0.0 ) return(0);
 
 	ix = 0;

kernel/riscv64/copy.c

@@ -41,7 +41,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	BLASLONG i=0;
 	BLASLONG ix=0,iy=0;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	while(i < n)
 	{

kernel/riscv64/dot.c

@@ -46,7 +46,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	BLASLONG ix=0,iy=0;
 	double dot = 0.0 ;
 
-	if ( n < 0 )  return(dot);
+	if ( n < 1 )  return(dot);
 
 	while(i < n)
 	{

kernel/riscv64/swap.c

@@ -41,7 +41,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3, FLOAT *x,
 	BLASLONG ix=0,iy=0;
 	FLOAT temp;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	while(i < n)
 	{

kernel/riscv64/zaxpy.c

@@ -44,7 +44,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i,
 	BLASLONG inc_x2;
 	BLASLONG inc_y2;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 	if ( da_r == 0.0 && da_i == 0.0 ) return(0);
 
 	ix = 0;

kernel/riscv64/zcopy.c

@@ -43,7 +43,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	BLASLONG inc_x2;
 	BLASLONG inc_y2;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	inc_x2 = 2 * inc_x;
 	inc_y2 = 2 * inc_y;

kernel/riscv64/zswap.c

@@ -45,7 +45,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3, FLOAT dumm
 	BLASLONG inc_x2;
 	BLASLONG inc_y2;
 
-	if ( n < 0     )  return(0);
+	if ( n <= 0     )  return(0);
 
 	inc_x2 = 2 * inc_x;
 	inc_y2 = 2 * inc_y;