CheckerboardAMDGPU
A version of Checkerboard.jl created for AMDGPU.
GPU kernel calling is based on this simple-gemm example.
Benchmarking shows that the time taken to run the kernel is dominated by the kernel launch latency:
$ for n in 16 32 64 128; do julia -O2 --project=. example.jl $n 128 1000; done
Size 128x256, time per M per step (us) 5.8708339589589595
Size 128x1024, time per M per step (us) 5.999556650400401
Size 128x4096, time per M per step (us) 6.348324832645146
Size 128x16384, time per M per step (us) 7.90716812124625
$ for n in 16 32 64 128; do julia -O2 --project=. example.jl $n 64 1000; done
Size 64x256, time per M per step (us) 11.6896200731982
Size 64x1024, time per M per step (us) 11.771267282907909
Size 64x4096, time per M per step (us) 12.628145442317313
Size 64x16384, time per M per step (us) 14.428985047547549