64 Pull requests merged by 23 people

• [mlir][dialect] Refactor DotLike trait into a DotOpInterface + Enable verification of scaled_dot

  #5796 merged Feb 5, 2025

• Bump actions/checkout from 3 to 4

  #5714 merged Feb 5, 2025

• Fix default num_stages values mismatch between Python frontend and MLIR.

  #5804 merged Feb 5, 2025

• [Blackwell] Enable MMA pipelining for scaled dot when TMEM copy is used

  #5812 merged Feb 5, 2025

• [Blackwell] Add support for mixed precision scaled dot

  #5799 merged Feb 4, 2025

• [LAYOUTS] Choose between wgmma RS and SS within AccelerateMatmul

  #5798 merged Feb 4, 2025

• [TUTORIAL] Apply remat perf fix to non-TMA persistent matmul

  #5811 merged Feb 4, 2025

• [INTERPRETER] Support tuple arguments in interpreter

  #5790 merged Feb 4, 2025

• [TEST] Use a fresh triton cache dir for warning tests

  #5809 merged Feb 4, 2025

• [LAYOUTS] Fix TransOp::fold

  #5807 merged Feb 4, 2025

• [PIPELINE] Remove outer loop pipelining transformation

  #5766 merged Feb 4, 2025

• [PROTON] Fix incorrect tmp_path initialization

  #5803 merged Feb 4, 2025

• [BACKEND] Disable ldmatrix.trans for fp8

  #5800 merged Feb 4, 2025

• [LAYOUTS] Don't hoist into ifs outside of loops

  #5801 merged Feb 4, 2025

• [LAYOUTS] Remove HoistLayoutConversion in favour of backwardsRemat

  #5788 merged Feb 3, 2025

• [NFC] Move element bit width into NVMMASharedEncoding

  #5794 merged Feb 3, 2025

• [INTERP] Support tensor descriptor ops

  #5795 merged Feb 3, 2025

• [PROTON] Parallelize proton tests

  #5792 merged Feb 3, 2025

• [PROTON-DEV] Improve profile interface

  #5793 merged Feb 2, 2025

• [AMD] NFC: Drop unused constructors in elementwise patterns

  #5789 merged Feb 2, 2025

• [BACKEND] Refactor shared memory layout representation

  #5786 merged Feb 1, 2025

• Use os.sep in filter_traceback function

  #5781 merged Feb 1, 2025

• Tutorial 09 Descriptor Kernel

  #5779 merged Feb 1, 2025

• [Proton] Fixed pc sampling error

  #5787 merged Feb 1, 2025

• [BC Breaking] Add output dtype to tl.sum with default

  #5763 merged Feb 1, 2025

• [Pipeliner] Fix condition for pipelining loads

  #5780 merged Feb 1, 2025

• [AMD] Add GFX950 fp32 to bf16 Conversion Ops

  #5782 merged Feb 1, 2025

• [BACKEND] bump llvm to ffe3129e9bdc146ee4d91e849173d1c64b1ae974

  #5784 merged Feb 1, 2025

• [Layouts] Remove sketchy remat condition

  #5783 merged Feb 1, 2025

• Do not reorder transpose of dot operand that is used in ops other than dotOp

  #5686 merged Jan 31, 2025

• [DEV] Don't use .ONESHELL in Makefile

  #5775 merged Jan 31, 2025

• [AMD] Emit AMD specific intrinsics for dot

  #4594 merged Jan 31, 2025

• [AMD] Rewrite canonicalize pointers to use 1:N conversion

  #5329 merged Jan 31, 2025

• [PROTON] Skip warnings caused by legacy clang compilers

  #5778 merged Jan 31, 2025

• Revert "[LAYOUTS] Generalise HoistLayoutConversion to work with arbit…

  #5776 merged Jan 31, 2025

• [TOOLS] Fixed bug in AOT compiler

  #5771 merged Jan 31, 2025

• Fix __builtin_clz implementation on Windows

  #5774 merged Jan 31, 2025

• [LAYOUTS] Generalise HoistLayoutConversion to work with arbitrary layouts and chains of ops

  #5673 merged Jan 31, 2025

• [AMD][BACKEND] Bugfix to small tile pingpong

  #5759 merged Jan 31, 2025

• [PROTON] Explicitly list all cpp files

  #5756 merged Jan 31, 2025

• [ANALYSIS][DEBUG] Output theoretical vs actual peak memory allocation size

  #5658 merged Jan 31, 2025

• [DRIVER] Pass correct SM and PTX versions to llvm

  #5770 merged Jan 31, 2025

• [Triton] Change xor_sum to use @jit (NFC)

  #5769 merged Jan 31, 2025

• [DOC] Update core maintainers list

  #5767 merged Jan 30, 2025

• Use env builtin implementation from LLVM's lit utility for platform independence

  #5762 merged Jan 30, 2025

• [PROTON] Add the -diff option to proton-viewer

  #5740 merged Jan 30, 2025

• [BACKEND] Canonicalize ReshapeOp even if not allowing reorder

  #5752 merged Jan 30, 2025

• [DEV] Unify Makefile and cuda CI commands

  #5753 merged Jan 30, 2025

• [PIPELINE] Limit number of buffers for register operands

  #5755 merged Jan 30, 2025

• Reapply "[Layouts] Propagate layouts into conditionals (#5610)"

  #5725 merged Jan 30, 2025

• [Proton][Dialect] Middle-end Proton operator definitions

  #5754 merged Jan 30, 2025

• Improve thread locality for reduction ops (#5671)

  #5757 merged Jan 30, 2025

• [PROTON] Reworked the mechanism for finding libraries for profiling backends.

  #5751 merged Jan 30, 2025

• [Frontend][Diagnostics] Improve emitting diagnostic information

  #5581 merged Jan 30, 2025

• [LAYOUTS] Create a trait that implements Layout equality by comparing the LLs

  #5747 merged Jan 29, 2025

• [BACKEND] Limit vector size to scratch size for convert_layout

  #5746 merged Jan 29, 2025

• [backend] NFC: Split architecture dependant and independant parts of FMA dot conversion

  #5655 merged Jan 29, 2025

• [BACKEND] bump to llvm/llvm-project@c118864223c6

  #5684 merged Jan 29, 2025

• Optimize reduce(reshape_1D)

  #5748 merged Jan 29, 2025

• [AMD][BACKEND] Disable pingpong with non-local_load input.

  #5718 merged Jan 29, 2025

• Revert "[PROTON] Prefer the default library path when loading profiler backends"

  #5749 merged Jan 29, 2025

• Revert "[Coalesce] Fix the default order to be row major "

  #5744 merged Jan 29, 2025

• [NVIDIA] Use correct commit type for TMA

  #5738 merged Jan 29, 2025

• [BACKEND] Deprecate SharedToDotOperandMMAv2OrV3.cpp

  #5734 merged Jan 29, 2025

13 Pull requests opened by 11 people

• [AMD] Initial support for LDS transpose load instructions

  #5750 opened Jan 29, 2025

• [AMD] Added `ConcatOp` to AMDGPU Dialect

  #5760 opened Jan 30, 2025

• [release/3.2.x] Get proper PTX version for CUDA >= 12.6

  #5765 opened Jan 30, 2025

• [OPTIMIZER] Fix insertion location in HoistLayoutConversion pattern

  #5772 opened Jan 31, 2025

• [WIP] [AMD] Specific swizzling pattern for TN GEMMs

  #5797 opened Feb 3, 2025

• [WIP][DNR][Pipeliner] Pipeline prologue/epilogue loads

  #5802 opened Feb 4, 2025

• [AMD] Moved membar analysis to its dedicated pass

  #5805 opened Feb 4, 2025

• [AMD] refactored instruction `sched.hint`

  #5808 opened Feb 4, 2025

• [PIPELINE] Relax requirements for wgmma operand register pipelining

  #5810 opened Feb 4, 2025

• [Blackwell][TUTORIALS] Add tutorial 10-block-scaled-matmul.py

  #5813 opened Feb 5, 2025

• [BACKEND] bump to llvm/llvm-project@ffe3129e9bdc

  #5814 opened Feb 5, 2025

• Copy local files if variable is specified

  #5815 opened Feb 5, 2025

• [Blackwell][Clean up] Remove use of SharedMemoryObject on TMEM

  #5817 opened Feb 5, 2025

10 Issues closed by 5 people

• histogram allows the creation of non-power-of-two tensor sizes

  #4826 closed Feb 4, 2025

• Assertion error from linear layouts

  #4727 closed Feb 4, 2025

• Where can i get the triton cp38 win version?

  #4732 closed Feb 4, 2025

• Blocksparse.matmul result does not align with torch

  #4709 closed Feb 4, 2025

• Assertion failure in LinearLayoutConversions on H100s when num_warps=8

  #5609 closed Feb 4, 2025

• jit issue when INTERPRETER=1

  #5056 closed Feb 2, 2025

• import te raise error

  #5722 closed Jan 31, 2025

• Assertion error when lowering a reduce->reshape->reshape->broadcast pattern to LLIR

  #5745 closed Jan 29, 2025

• Triton does not really enable -ftz

  #5735 closed Jan 29, 2025

• Not able to install dependencies file. : triton=3.0.0

  #5741 closed Jan 29, 2025

4 Issues opened by 4 people

• Potential Bug in **_attn_fwd_tma** Function

  #5816 opened Feb 5, 2025

• Triton interpreter cannot handle parameters that alias

  #5791 opened Feb 2, 2025

• "ImportError: cannot import name 'backends' from 'triton.backends' (unknown location)" for triton installed from source

  #5773 opened Jan 31, 2025

• Misreport "Cannot have `return` statements inside `while` or `for`" if values returned by function are disregarded

  #5768 opened Jan 30, 2025

18 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [AMD] AsyncCopyGlobalToLocal lowering to global.load.lds

  #5729 commented on Feb 3, 2025 • 22 new comments

• Fix assertion in ScanLowering for num_ctas>1

  #5680 commented on Feb 5, 2025 • 3 new comments

• [AMD] Enable pingpong scheduling by default

  #5696 commented on Feb 1, 2025 • 2 new comments

• [Proton][Dialect] Add Proton Device Memory Buffer Init and Allocate Pass

  #5606 commented on Feb 5, 2025 • 2 new comments

• [WIP][Pipeliner] Enable automatic loop fusion

  #5726 commented on Feb 4, 2025 • 0 new comments

• [WIP] Support shared encoding defined with linear layout

  #5720 commented on Feb 1, 2025 • 0 new comments

• [WIP][AMD] Add MFMA and WMMA layouts to LinearEncodingTest

  #5698 commented on Jan 29, 2025 • 0 new comments

• [Proton][Dialect] Middle-end support of the Proton Dialect and the frontend Python package

  #5677 commented on Feb 5, 2025 • 0 new comments

• [WIP] [AMD] Remove "remove unsupported conversions" pass

  #5674 commented on Feb 3, 2025 • 0 new comments

• Ensured that dtype subclasses are hashable

  #5657 commented on Feb 4, 2025 • 0 new comments

• [AMD] refactor convert buffer ops

  #5563 commented on Feb 5, 2025 • 0 new comments

• [MXFP] Implement SW emulation of dot_scale as a decomposition

  #5475 commented on Feb 4, 2025 • 0 new comments

• [WIP][SWP] Print recurring dependencies when reporting scheduling conflicts

  #5375 commented on Feb 4, 2025 • 0 new comments

• [AMD-Pipeline] Add multi-stage global/local prefetch

  #5353 commented on Feb 5, 2025 • 0 new comments

• Why is the documentation not versioned like other Read The Docs sites?

  #4454 commented on Feb 4, 2025 • 0 new comments

• [RFC] Improve performance for layer-norm in turtorial

  #5712 commented on Feb 3, 2025 • 0 new comments

• Is Triton unable to install in python 3.10 versions?

  #1057 commented on Jan 31, 2025 • 0 new comments

• [3.2.x] `ptx_get_version` cannot handle CUDA>12.6

  #5737 commented on Jan 30, 2025 • 0 new comments