Private AMD byte copy #

copy.cl derives from ROCm clr's MIT-licensed __amd_rocclr_copyBuffer in rocclr/device/blitcl.cpp, commit 2007bc5482af4d67b402016d13a9c98b4170a881. LICENSE preserves its license. It takes destination, source, and byte count. Aligned addresses use 16-byte vectors and a byte tail; other addresses use byte loads and stores.

Regenerate ../copy_kernel.ml with Clang and AMDGPU LLD 22:

uv run python generate.py --clang /path/to/clang --lld /path/to/ld.lld

It refuses an object whose copy_bytes is not what the driver dispatches: no private or group memory, 24 bytes of arguments, groups of 256 work-items.

The checked-in objects were compiled by Homebrew Clang 22.1.7 and linked by Triton's LLD 22.0.0, LLVM revision f58b06dce1f9c15707c5f808fd002e18c2accf7e. They use code object V6. The runtime uses its existing ELF reader and launch descriptors. It has no compiler or linker dependency.

Target	ELF bytes	Image bytes	SHA256
gfx9-4-generic	6152	14336	62ca0208baa1276b8821811de8c241b14c3e3bc583768aecdde4aad22f504f30
gfx11-generic	5816	10240	1e3d8c55025d3cdbf1024a5c4229c18d279fdbac8af7363ed69c54f735e38f90
gfx12-generic	5816	10240	7fa73234c9da7ed91211dfd830982d0dd611aa9a164c3beaa073790fb361e0ed

All three objects have zero private and group segments. Only gfx12-generic has run on hardware: gfx1201, through ordinary amdgpu/KFD with PM4, including Pinned instruction storage and the complete 256 MiB byte oracle. The other targets, AQL, and driverless paths remain untested.