JuliaGPU/AMDGPU.jl

Investigate using MIOpen's immediate mode for conv algorithm search

ToucheSir opened this issue · 0 comments

As described in https://rocm.docs.amd.com/projects/MIOpen/en/latest/find_and_immediate.html#immediate-mode-api. The main benefits would be enabling persistent on-disk caching for users while removing the somewhat tricky algorithm cache handling code on the Julia side. The main drawback would be not being able to control how the caching is done, but on balance I think the benefits outweigh this bit of flexibility. Notably, cuDNN does not support this kind of transparent caching (ref. pytorch/pytorch#88410) and currently requires manual caching on the Julia side (ref. JuliaGPU/CUDA.jl#1947).