spectrometerHBH opened a new pull request, #20223: URL: https://github.com/apache/tvm/pull/20223
This adds an exact CUDA launch contract for TIRx kernels that need PTX `.reqntid` rather than advisory launch bounds. - preserve `tirx.required_block_size` through host/device splitting - emit CUDA 13 `__block_size__` with static thread and cluster dimensions - pass the required-block flag through runtime metadata without adding a packed operand - launch with `CU_LAUNCH_KERNEL_REQUIRED_BLOCK_DIM` and avoid duplicate cluster attributes - reject incompatible launch-bound and max-register controls Validation: - `cmake --build build --parallel` - `python -m pytest -q tests/python/tirx-transform/test_tir_transform_split_host_device.py tests/python/tirx/codegen/test_codegen_cuda.py` (225 passed) - `pre-commit run --files include/tvm/tirx/function.h src/backend/cuda/codegen/codegen_cuda.cc src/backend/cuda/runtime/cuda_module.cc src/runtime/metadata.h src/runtime/thread_storage_scope.h src/tirx/transform/split_host_device.cc tests/python/tirx-transform/test_tir_transform_split_host_device.py tests/python/tirx/codegen/test_codegen_cuda.py` - CUDA 13.2 SM100 runtime launch through a TIRx kernel with generated PTX `.reqntid 128,1,1` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
