Half-Precision SGEMV() and SGEMM() not support CblasColMajor

[djeong20]

Currently, the implementation of half-precision SGEMV() and SGEMM() does not support CblasColMajor for the order parameter (CBLAS_ORDER).

/** CBLAS_ORDER order parameter is not used. */
static void sgemv_FP16(CBLAS_ORDER order, CBLAS_TRANSPOSE TransA,
                       const unsigned int M, const unsigned int N,
                       const float alpha, const _FP16 *A,
                       const unsigned int lda, const _FP16 *X, const int incX,
                       const float beta, _FP16 *Y, const int incY) {

  unsigned int incy = abs(incY);
  unsigned int incx = abs(incX);

  if (TransA == CblasTrans) {
#if (defined USE__FP16 && USE_NEON)
    nntrainer::neon::sgemv_transpose_neon_fp16(A, X, Y, M, N, alpha, beta);
#else
    sgemv_loop_fp16(i, j, N, M);
#endif
  } else {
#if (defined USE__FP16 && USE_NEON)
    nntrainer::neon::sgemv_neon_fp16(A, X, Y, M, N, alpha, beta);
#else
    sgemv_loop_fp16(j, i, M, N);
#endif
  }
}

This limitation results in issues when trying to use these functions with column-major data layouts.
Such support for CblasColMajor in the half-precision implementations of SGEMV() and SGEMM() is needed.

[taos-ci]

cibot: Thank you for posting issue #2399. The person in charge will reply soon.