zeekay
0b96ebc51a
dlopens every lux_<name>.plugin and runs the same op (out=a+b) through each backend's shading language, asserting bit-identical results vs a CPU reference. One-line-per-backend registry (kBackends[]). Verified metal+vulkan MATCH live on Apple M1 Max; cuda/hip/webgpu reported as needing their runtime/hardware.
1.7 KiB
lux-gpu-conformance — cross-backend kernel-parity harness
A single self-contained runner (conformance.cpp) that proves kernel parity
across every lux-accel backend plugin: it dlopens each lux_<name>.plugin, runs
the SAME logical op (out = a + b, N=4096) through each backend's own shading
language, and asserts every backend's output is bit-identical to a CPU reference.
One-line registration
The registry is the kBackends[] array in conformance.cpp. Each backend is
ONE row: { type, name, plugin-filename, kernel-source, entry }. Today it
carries metal (MSL), webgpu (WGSL), cuda (CUDA C), hip (CUDA C), vulkan (GLSL).
This is the structure meant to drop straight into the lux-accel/gpu parity
harness when it exists — adding a backend stays one line.
Plugin discovery
In order: "<name>=<path>" argv overrides → $LUX_PLUGIN_DIR → sibling build
trees (../lux-<name>/build/lux_<name>.plugin or ../../... from a build/
subdir). Backends that can't be located or aren't available are reported, not
failed; the harness exits non-zero only on an actual cross-backend mismatch.
Build + run (macOS, Metal + Vulkan live)
cmake -B build -DCMAKE_BUILD_TYPE=Release \
-DLUXACCEL_INCLUDE_DIR=/path/to/lux-accel/include
cmake --build build -j
export VK_ICD_FILENAMES=/opt/homebrew/opt/molten-vk/etc/vulkan/icd.d/MoltenVK_icd.json
export DYLD_LIBRARY_PATH=/opt/homebrew/opt/molten-vk/lib:/opt/homebrew/opt/vulkan-loader/lib
./build/lux_gpu_conformance
Verified on Apple M1 Max: metal=MATCH, vulkan=MATCH (2 live backends,
bit-identical). cuda/hip report "unavailable (no runtime/hardware here)";
webgpu needs Dawn/wgpu-native. Run cuda/hip on NVIDIA/AMD hosts (evo/spark) to
extend the proven set to all five.