Merge pull request #4929 from martin-frbg/issue4905

Fix CBLAS_?GEMMT filling in the wrong triangle for Row-Major

interface/gemmt.c

@@ -319,8 +319,8 @@ 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 (Uplo == CblasUpper) uplo = 1;
+		if (Uplo == CblasLower) uplo = 0;
 
 		if (TransB == CblasNoTrans)
 			transa = 0;

utest/test_extensions/test_cgemmt.c

@@ -81,6 +81,28 @@ static void cgemmt_trusted(char api, enum CBLAS_ORDER order, char uplo, char tra
 
     ldc *= 2;
 
+#ifndef NO_CBLAS
+    if (order == CblasRowMajor) {
+    if (uplo == 'U' || uplo == CblasUpper)
+    {
+        for (i = 0; i < m; i++)
+            for (j = i * 2; j < m * 2; j+=2){
+                data_cgemmt.c_verify[i * ldc + j] =
+                    data_cgemmt.c_gemm[i * ldc + j];
+                data_cgemmt.c_verify[i * ldc + j + 1] =
+                    data_cgemmt.c_gemm[i * ldc + j + 1];
+            }
+    } else {
+        for (i = 0; i < m; i++)
+            for (j = 0; j <= i * 2; j+=2){
+                data_cgemmt.c_verify[i * ldc + j] =
+                    data_cgemmt.c_gemm[i * ldc + j];
+                data_cgemmt.c_verify[i * ldc + j + 1] =
+                    data_cgemmt.c_gemm[i * ldc + j + 1];
+            }
+    }
+    } else 
+#endif
     if (uplo == 'L' || uplo == CblasLower)
     {
         for (i = 0; i < m; i++)

utest/test_extensions/test_dgemmt.c

@@ -77,6 +77,21 @@ static void dgemmt_trusted(char api, enum CBLAS_ORDER order, char uplo, char tra
     else
         cblas_dgemm(order, transa, transb, m, m, k, alpha, data_dgemmt.a_test, lda,
                 data_dgemmt.b_test, ldb, beta, data_dgemmt.c_gemm, ldc);
+
+	if (order == CblasRowMajor) {
+    if (uplo == 'U' || uplo == CblasUpper)
+    {
+        for (i = 0; i < m; i++)
+            for (j = i; j < m; j++)
+                data_dgemmt.c_verify[i * ldc + j] =
+                    data_dgemmt.c_gemm[i * ldc + j];
+    } else {
+        for (i = 0; i < m; i++)
+            for (j = 0; j <= i; j++)
+                data_dgemmt.c_verify[i * ldc + j] =
+                    data_dgemmt.c_gemm[i * ldc + j];
+    }
+	}else
 #endif
 
     if (uplo == 'L' || uplo == CblasLower)

utest/test_extensions/test_sgemmt.c

@@ -77,6 +77,21 @@ static void sgemmt_trusted(char api, enum CBLAS_ORDER order, char uplo, char tra
     else
         cblas_sgemm(order, transa, transb, m, m, k, alpha, data_sgemmt.a_test, lda,
                 data_sgemmt.b_test, ldb, beta, data_sgemmt.c_gemm, ldc);
+	if (order == CblasRowMajor) {
+    if (uplo == 'U' || uplo == CblasUpper)
+    {
+        for (i = 0; i < m; i++)
+            for (j = i; j < m; j++)
+                data_sgemmt.c_verify[i * ldc + j] =
+                    data_sgemmt.c_gemm[i * ldc + j];
+    } else {
+        for (i = 0; i < m; i++)
+            for (j = 0; j <= i; j++)
+                data_sgemmt.c_verify[i * ldc + j] =
+                    data_sgemmt.c_gemm[i * ldc + j];
+    }
+
+	} else
 #endif
 
     if (uplo == 'L' || uplo == CblasLower)

utest/test_extensions/test_zgemmt.c

@@ -80,7 +80,28 @@ static void zgemmt_trusted(char api, enum CBLAS_ORDER order, char uplo, char tra
 #endif
 
     ldc *= 2;
-
+#ifndef NO_CBLAS
+    if (order == CblasRowMajor) {
+    if (uplo == 'U' || uplo == CblasUpper)
+    {
+        for (i = 0; i < m; i++)
+            for (j = i * 2; j < m * 2; j+=2){
+                data_zgemmt.c_verify[i * ldc + j] =
+                    data_zgemmt.c_gemm[i * ldc + j];
+                data_zgemmt.c_verify[i * ldc + j + 1] =
+                    data_zgemmt.c_gemm[i * ldc + j + 1];
+            }
+    } else {
+        for (i = 0; i < m; i++)
+            for (j = 0; j <= i * 2; j+=2){
+                data_zgemmt.c_verify[i * ldc + j] =
+                    data_zgemmt.c_gemm[i * ldc + j];
+                data_zgemmt.c_verify[i * ldc + j + 1] =
+                    data_zgemmt.c_gemm[i * ldc + j + 1];
+	    }
+    	}
+    }else
+#endif	    
     if (uplo == 'L' || uplo == CblasLower)
     {
         for (i = 0; i < m; i++)