Possibility of using RDNA3 Matrix Multiplication Instructions to Speed up on AMD GPU

[fancyIX]

AMD's RDNA 3 architecture introduces matrix cores with specialized instructions for accelerating matrix multiplication.

Details here: https://gpuopen.com/learn/wmma_on_rdna3/

[fancyIX]

Wave Matrix Multiply-Accumulate (WMMA) instructions provide acceleration for common matrix arithmetic
operations. The instructions are encoded using the VOP3P encoding.

[fancyIX]

I tried the clang that comes with amd official driver. Seems like it can compile opencl code which calls __builtin_amdgcn_wmma_f32_16x16x16_f16_w32 etc.

[CNugteren]

Sounds interesting. I myself (the maintainer of this project) am not able to work on this project, but I'm happy to help people that are willing to in the right direction, and review & accept pull requests.