ArtHIP — standalone HIP Winograd WMMA inference for AMD GPUs

Runs fp16 conv-chain ONNX models (ArtCNN-style: 3x3 convs + DepthToSpace) directly on AMD GPUs through hand-written gfx1100 Winograd F(2x2,3x3) WMMA kernels. Currently beats the MIGraphX reference plugin (~23.0 vs ~15.7 fps @1080p, ArtCNN R8F64).

Layout

• src/common/ — hip_kernels.h, the gfx1100 Winograd WMMA kernels shared by both backends below.
• src/onnxruntime-hip/ — the ONNX Runtime execution provider (hip_graph.cc, hip_execution_provider.cc, hip_provider_factory.cc) and its build.sh.
• src/vapoursynth/ — the standalone VapourSynth plugin (vs_hip.cpp
  • hip_engine.cc/h, no ORT involved), shared ONNX helpers (onnx_utils, convert_float_to_float16), and build_hip.sh.
• tests/ — fixtures/ (ONNX models + reference images), tools/ (manual benches and comparison scripts), correctness/ (the pytest gate). See tests/README.md.
• NOTES.md — experiment log; read this before touching kernels.
• third_party/onnxruntime/ — the ORT 1.29.0 source tree the EP build links against (git-ignored, not committed).

Build / Test

• Plugin: cd src/vapoursynth && ./build_hip.sh (hipcc --offload-arch=gfx1100), then copy the result into your VapourSynth plugins dir, e.g. cp build/libhip.so /usr/lib/python3.14/site-packages/vapoursynth/plugins/libhip.so
• Provider (ORT EP): cd src/onnxruntime-hip && ./build.sh
• VapourSynth comparison: VS_BACKEND=<hip|migx> vspipe -p tests/tools/vs_test.py -- (500 blank 1920x1080 frames)
• EP accuracy/speed: python tests/tools/multires.py 1920

Notes

• Compute is fp16; clip IO follows the model like the MIGX plugin: GRAYS in → GRAYS out, GRAYH/int in → GRAYH out.