Can inline assembly in sycl like cuda does?

[alanzhai219]

You know, developers can inline PTX in the cuda kernel, which is much more efficient for expert developers. That is much more important for extreme optimization.
Can DPC++ or sycl support inline assembly?
If yes, please provide a sample.
If not, is any plan for such feature.
Thank you.

[JackAKirk]

Yes it is exactly as in the nvcc compiler (see https://docs.nvidia.com/cuda/inline-ptx-assembly/index.html) except you have to wrap ptx code in the following MACRO:

if defined(__NVPTX__)
your_ptx_asm_here
#endif

In a parallel_for.

[alanzhai219]

@JackAKirk Maybe my description is unclear. I mean SYCL can support inline assembly now?

[AlexeySachkov]

@alanzhai219, yes, this functionality is supported already, you can find some examples (but with Intel GPUs ASM) here: https://github.com/intel/llvm/tree/sycl/sycl/test-e2e/InlineAsm

[JackAKirk]

Yes, I mean to do

if defined(__NVPTX__)
your_ptx_asm_here
#endif

inside any SYCL parallel_for. e.g. queue::parallel_for.

[alanzhai219]

@AlexeySachkov Thanks, that is what I want. BTW, is there any such asm spec for intel GPU?

[AlexeySachkov]

@alanzhai219, the only document I'm aware of is https://github.com/intel/intel-graphics-compiler/blob/master/documentation/visa/1_introduction.md

[alanzhai219]

@AlexeySachkov Thanks a lot.