Row Major gemm for complex double throws oneapi::math::unimplemented

[LukeTheWalker]

Summary

I'm trying to use the library for a simple example, I'd like to use the row_major interface, already tried the column major one and it works as expected-

Version

Latest commit is:
Commit af66e5e

Environment

oneMath works with multiple HW and backend libraries and also depends on the
compiler and build environment. Include
the following information to help reproduce the issue:

• NVIDIA A100
• CUDA 12.4
• RedHat
• icpx 2025.0.4

Steps to reproduce

cpx A[4] = {1, 2, 3, 4};
cpx B[4] = {5, 6, 7, 8};
cpx C[4] = {0};

sycl::buffer<cpx> buf_A(A, sycl::range<1>(size * size));
sycl::buffer<cpx> buf_B(B, sycl::range<1>(size * size));
sycl::buffer<cpx> buf_C(C, sycl::range<1>(size * size));

oneapi::math::blas::row_major::gemm(queue_, 
            oneapi::math::transpose::nontrans,  oneapi::math::transpose::nontrans,
            size, size, size, 
            cpx(1.0, 0), buf_A, size, buf_B, size, cpx(0.0, 0.0), buf_C, size);

Observed behavior

When I run the code I get:

terminate called after throwing an instance of 'oneapi::math::unimplemented'
  what():  oneMath: blas/gemm: function is not implemented for row_major layout

Expected behavior

I would expect this to just work as it is a quite standard blas routine and to me it seems to not be advertised anywhere that it should not work so I assume it is actually an issue

[s-Nick]

Hi @LukeTheWalker Can you also add to how you compiled the library and the example please?
I suppose you are using cuBLAS backend, so please keep in mind that oneMath doesn't offer more functionality than underlying libraries. e.g. cuBLAS doesn't have row_major support for gemm operator therefore neither oneMath supports row_major for cuBLAS backend.

[sknepper]

Please note that for many standard BLAS functions, it is possible to rearrange the parameters and call a column-major function to compute when the matrices are in row-major layout. E.g., see Netlib cblas_zgemm.c: https://github.com/Reference-LAPACK/lapack/blob/master/CBLAS/src/cblas_zgemm.c
We'd be happy to get such a contribution!

[LukeTheWalker]

I see, I still think it should be provided as an interface for completeness as suggested in #588