gemmt fills output matrix in wrong corner for upper row-major

[david-cortes]

In function cblas_?gemmt, if I pass CblasRowMajor + CblasUplo to MKL's function with that name, the output ends up being filled in the upper corner (in row-major order), but in OpenBLAS's version of cblas_?gemmt, the output ends up being filled in the lower corner instead.

Example code:

#include <iostream>
#include "cblas.h"

extern "C" 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);


int main()
{
    const double A[] = {
        2., 3.,
        0., 4.
    };
    double B[4] = {0.};

    cblas_dgemmt(
        CblasRowMajor, CblasUpper, CblasTrans, CblasNoTrans,
        2, 2,
        1., A, 2,
        A, 2,
        0., B, 2
    );

    std::cout << "B:" << std::endl;
    std::cout << "[ " << B[0] << ", " << B[1] << "]" << std::endl;
    std::cout << "[ " << B[2] << ", " << B[3] << "]" << std::endl;

    return 0;
}

If I run the example with OpenBLAS, the output is:

B:
[ 4, 0]
[ 6, 25]

Whereas with MKL, the output is:

B:
[ 4, 6]
[ 0, 25]

Version: OpenBLAS 0.3.28, OpenMP variant, installed from source.

[XiWeiGu]

Yes, I encountered the same issue. The reason might be here：

OpenBLAS/interface/gemmt.c

Lines 314 to 323 in 73c1882

	if (order == CblasRowMajor) {
	a = (void *)B;
	b = (void *)A;
	lda = LDB;
	ldb = LDA;
	if (Uplo == CblasUpper) uplo = 0;
	if (Uplo == CblasLower) uplo = 1;

Make the following revisions to the code on lines 322 and 323 (opposite to CblasColMajor instead of consistent)：

 	if (Uplo == CblasUpper) uplo = 1; 
 	if (Uplo == CblasLower) uplo = 0; 

[martin-frbg]

Indeed this is a copy/paste blunder, thanks