znsoftm 5df381d4a6 Add a native Metal backend for macOS (backend=metal) ...

src/metal_driver.mm implements the compute subset of the Vulkan entry
points on Metal, so the existing GPU graph, fusions and shaders run
unchanged. Shaders go GLSL -> SPIR-V -> MSL (spirv-cross) at build time
and compile at runtime with fast math off; per-mode pipelines are
specialized on the low 16 mode bits. Weights and activations live in
shared-storage MTLBuffers; each queue submit is one command buffer.

- CMake: PPOCR_ENABLE_METAL (default ON on Apple), works with Vulkan OFF.
- API: Backend::metal / MetalAvailable(), PPOCR_BACKEND_METAL /
  ppocr_metal_available(), Python backend "metal" / use_metal. Missing
  Metal falls back to CPU (PPOCR_DISABLE_METAL=1 forces that).
- ppocr_metal_check, tests/run_metal_macos.sh, python/tests/test_metal.py
  and a macos-14 CI job.
- README / README-CN: Metal usage and M4 numbers.

M4, idle: page 538 ms (MoltenVK 706 ms, NEON 209 ms); hello 76 ms.
Text and boxes equal to CPU on hello and page (IoU 1), 4-thread
determinism, 0 errors under MTL_DEBUG_LAYER=1, pytest passes.

2026-10-13 14:49:49 +08:00

..

export.h

feat: add C ABI, Python package, and CPU ISA paths

2026-10-04 23:40:06 +00:00

ppocr.h

Add a native Metal backend for macOS (backend=metal)

2026-10-13 14:49:49 +08:00

ppocr.hpp

Add a native Metal backend for macOS (backend=metal)

2026-10-13 14:49:49 +08:00