main

GGitHubAdd support for padding option to TMA loads (#7993)

dba08d43创建于 2025年8月29日历史提交

文件	最后提交记录	最后更新时间
canonicalize.mlir	[DIALECT] Rename triton_gpu to ttg and triton_nvidia_gpu to ttng (#5266) It may cause changes for downstream tasks but we think it's beneficial to shorten dialect name and make them consistent. That is, we are using tt to represent the triton dialect.	1 年前
combine.mlir	[BACKEND] Preserve tt attrs in AddPtr combine and canonicalize (#7113) Combine and canonicalize remove the attributes that relate to alignment so we don't get vectorization of stores in some cases, this allows them to keep it.	1 年前
invalid.mlir	[Dialect] Actually enable TMEM layout check and fix all the tests (#7723) I also removed unnecessary noinline=false from the tests.	11 个月前
loop-invariant-code-motion.mlir	[OPTIMIZER] Customize LICM to hoist loads from loops (#6051) This PR customizes the MLIR default LICM pass to also hoist loads from loops that consist of pure/read-only ops. Co-authored-by: peterbell10 <peterbell10@openai.com>	1 年前
loop-peeling.mlir	[PIPELINER] Reintroduce epilogue peeling (#6962) Reintroduce "[PIPELINE] Peel single epilogue iteration after loop expansion #6893" Fixed the issue that was exposing IMA by using poison values as buffer indices used by speculatively executed async copy.	1 年前
loop-unroll.mlir	[LoopUnroll] Do not pipeline epilog loops generated by loop unrolling (#5027) The epilog loop created by the loop unroller may not be run if the main unrolled loop covers all original loop iterations, thus pipelining it non-speculatively may not be beneficial. It can also cause some correctness issue when combined with the downstream PTXAS optimizer.	1 年前
loop_cse.mlir	[Triton] Fix loop aware CSE not checking op result index (#7338) Turns out operations can have more than one result.	1 年前
ops.mlir	[PIPELINER] Support pipelining scalar loads (#7498) So far scalar loads were not being pipelined but they can still cause latency problems. Extend support for scalar loads, to keep things simpler we convert those into tensor<1> loads during loop lowering. This also introduces a new unsplat op to make the conversion from tensor to scalar simple.	1 年前
reorder-broadcast.mlir	Remove -split-input-file from tests that don't use file splitting (#6257) I worked through the repository looking for lit tests that invoked triton-opt with -split-input-file/--split-input-file, but which do not actually use the // ----- separator to split a single file into multiple test inputs. Keeping -split-input-file in those cases makes MLIR diagnostics harder to read because locations get rewritten as split index offsets. Using ripgrep and a small shell loop, I identified the set of .mlir and .mlir.in files that contain the flag but have no split markers, and removed the redundant split-input-file option from their // RUN: lines. This touched a range of tests across the Triton, TritonGPU, AMD and NVIDIA conversion directories. No functional changes were made to the test contents, only the invocation command lines were simplified. While running pre-commit to verify formatting, the hooks fixed a handful of trailing newline and style issues in .github/actions/autofix Python scripts, and reformatted a couple of minor areas picked up by yapf. Those changes are included in the commit. All pre-commit hooks now pass: check for broken symlinks................................................Passed detect destroyed symlinks................................................Passed trim trailing whitespace.................................................Passed fix end of files.........................................................Passed check yaml...............................................................Passed check toml...............................................................Passed check python ast.........................................................Passed check for added large files..............................................Passed check for merge conflicts................................................Passed check that executables have shebangs.....................................Passed check that scripts with shebangs are executable..........................Passed detect private key.......................................................Passed debug statements (python)................................................Passed ruff.....................................................................Passed yapf.....................................................................Passed clang-format.............................................................Passed Expand YAML anchors......................................................Passed I wasn’t able to run the full lit test suite in this environment as llvm-lit isn’t available, and the Python unit tests require a configured Triton install. The changes are isolated to test harnesses and formatting, so functional behaviour should be unaffected. --- This PR was generated by an AI system in collaboration with maintainers: @Mogball --------- Co-authored-by: Jeff Niu <jeffniu22@gmail.com>	1 年前
reproducer.mlir	[Dialect] Actually enable TMEM layout check and fix all the tests (#7723) I also removed unnecessary noinline=false from the tests.	11 个月前
rewrite-tensor-descriptor-to-pointer.mlir	Add support for padding option to TMA loads (#7993) Closes #7364 builds on top of #7364 from @jhapradip and addresses remaining comments, as well as implements thepadding option in the fallback RewriteTensorDescriptorToPointer path. - support for passing padding = "nan" on TMA descriptor creation for both host and device TMAs - forwards this argument down to tma descriptor creation - implement the NaN other value in the TMA fallback path --------- Co-authored-by: Pradip Jha <pradipjha@hotmail.com>	11 个月前
rewrite-tensor-pointer.mlir	[Backend] Update scf.if result uses in RewriteTensorPointer pass (#4893) I ran into an error in the RewriteTensorPointer pass. In my IR, there's an scf.if that produces a non-pointer result. The rewriteIfOp() created a new scf.if, but the use of scf.if result is still referencing the old one, which caused a compile error. In this patch, I updated all uses of scf.if with the results of the new if-op.	1 年前
vecadd.mlir	[NFC] Remove dead code related to IndexCastOp (#5596) IndexCast shouldn't exist at TTIR or TTGIR level	1 年前
verify-make-range.mlir	[Triton][IR] Requirett.make_range op to have positive size (#7093)	1 年前