[Perf] Streams 2: Add AMDGPU/HIP stream support#408

Draft

hughperkins wants to merge 8 commits intohp/streams-quadrantsic-1-cuda-streamsfrom

hp/streams-quadrantsic-2-amdgpu-cpu

Commits on Mar 11, 2026

Add AMDGPU/HIP stream support and async memory operations
hughperkins
committed

Commits on Mar 12, 2026

Merge branch 'hp/streams-quadrantsic-1-cuda-streams' into hp/streams-quadrantsic-2-amdgpu-cpu
hughperkins
committed
Move AMDGPU mem_free_async before transfers sync to match CUDA ordering
hughperkins
committed
Convert AMDGPU sync memcpy_host_to_device to async on active_stream
hughperkins
committed
Document ROCm >= 5.4 requirement for hipMallocAsync/hipFreeAsync
hughperkins
committed
Relax concurrency test threshold and log timings
hughperkins
committed
Add handle==0 guard to AMDGPU stream_synchronize and make stream_ thread_local
hughperkins
committed
Revert "Relax concurrency test threshold and log timings"
hughperkins
committed