spectrometerHBH opened a new pull request, #19896:
URL: https://github.com/apache/tvm/pull/19896
## Summary
This bundles the 16 commits currently carried in `spectrometerHBH/tvm` on
top of `apache/tvm:main`.
Major areas:
- Extend CUDA TIRx tile primitives and op dispatch paths, including vector
PTX ld/st, shared-memory copy paths, TMA/tcgen05 descriptor handling, dense
FP8/TF32 `gemm_async`, and CUDA elementwise tile dispatch.
- Add support utilities for benchmark timing, CUDA ptxas option plumbing,
and TMA/TFLOAT32 descriptors.
- Fix unsigned integer floormod/floordiv simplification rewrites without
overflow.
- Update TIRx dtype handling for upstream `PrimType` compatibility.
- Add and update TIRx CUDA/operator tests for copy, elementwise, permute
layout, and `gemm_async` behavior.
## Validation
- `git diff --check apache/main..HEAD`
- `python -m tirx_kernels.tir_bench --check-imports`
- `python -m tirx_kernels.registry --cc 10 --strict`
- `python -m pytest tests/python/tirx/ -n 16`
- `2033 passed, 39 skipped, 3 xpassed`
- Focused TIRx regression tests after formatting:
- `test_cast_vec2_packed_dispatch`
- `test_cast_warpgroup_src_layout_to_flat_uses_vec2_intrinsic`
- `test_gemm_tcgen05_cta_group_1[task0]`
- Full `tir-bench --impls all` sweep through `tirx-kernels-staging`: 256/256
workloads completed successfully.
--
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]
