Does sycl can support cuda tensorcore invoking？

[wangzy0327]

Does sycl can support cuda tensorcore invoking？
If support，how to invoke cuda tensorcore?

[JackAKirk]

Yes, DPC++ supports the same Tensor Cores capabilities as CUDA Runtime API (minus experimental ones). The full list is here https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_matrix/sycl_ext_oneapi_matrix.asciidoc#nvidia-tensor-cores-supported-combinations
and that document tells you how to use them within the joint_matrix dpc++ extension.

[JackAKirk]

Also basic tests for this functionality are available in this test file (for Tensor Cores operations supported by sm80 and later): https://github.com/intel/llvm/blob/sycl/sycl/test-e2e/Matrix/joint_matrix_tensorcores_sm80.cpp
See the joint_matrix_tensorcores_sm72.cpp and joint_matrix_tensorcores_sm70.cpp files for Tensor Cores operations supported by earlier sm versions.

[wangzy0327]

@JackAKirk Does AMD matrix core support at present？

[JackAKirk]

No but I expect the PR to add it to be merged very soon: #11485 , although note this is only for gfx90a gpus.