PR Digest Iteration 52

PR Summary

Period: 2026-03-01 to 2026-03-14 (Iteration 52) | Total PRs: 80 (from Xe2/Xe3/Xe3P) | Lines changed: +14,752 / -7,160

---

Triton XPU BE (Xe2/Xe3/Xe3P) (80 PRs, +14,752 / -7,160)

The team extended tensor descriptor and block I/O lowering infrastructure, streamlined the compiler pipeline with pass reordering and layout optimization improvements, fixed correctness issues in stride analysis and error handling. 8 upstream synchronizations were completed with a 97.53%–97.91% test pass rate.

Key accomplishments:

• Advanced the tensor descriptor native lowering path with store support, transpose fusion, and block pointer pass reordering
• Simplified the compiler pipeline by adopting the common Coalesce pass and refactoring layout propagation
• Fixed correctness bugs in stride analysis and barrier removal that could produce wrong results
• Maintained upstream alignment with 8 OpenAI Triton merges (97.53%–97.91%)
• Reduced CI execution time through test shape reductions and suite rebalancing

Memory Access & Lowering

Tensor descriptor lowering, 2D block I/O, and block pointer infrastructure.

• #6333 Revert "Revert "[LoadStoreOpToLLVM] Enable block io for all layout by default. (#5641) (#6044)" — (+10/-6, @chengjunlu)
• #6349 Add TRITON_INTEL_ENABLE_BLOCK_IO_ALL_LAYOUTS checks in descriptor-store-block-2d.mlir — (+9/-5, @whitneywhtsang)
• #6342 Add descriptor_prefetch operation. — (+59/0, @anthonycanino)
• #6268 Add DescriptorStoreOp to block io lowering — (+693/-5, @anthonycanino)
• #6343 Add TRITON_INTEL_ENABLE_BLOCK_IO_ALL_LAYOUTS checks in descriptor-load-block-2d.mlir — (+4/-2, @whitneywhtsang)
• #6327 Fix test_host_tensor_descriptor_in_tuple for Intel XPU — (+40/-61, @etiotto)
• #6238 Fuse transpose into descriptor load for dot operands — (+465/-20, @etiotto)
• #6286 Move after AccelerateMatmul pass — (+18/-12, @whitneywhtsang)
• #6293 Reduce tensor shape in test_tensor_descriptor — (+9/-1, @whitneywhtsang)
• #6260 : Fix rematerialization for atomic operations — (+136/0, @etiotto)
• #6275 Move after RemoveLayoutConversions pass — (+1/-1, @whitneywhtsang)
• #6231 Intel RemoveLayoutConversions pass support for tensor descriptors — (+123/-1, @mieshkiwrk)
• #6253 : Lower tt.descriptor_load to efficient 2D block loads — (+443/-7, @etiotto)
• #6267 Move after coalesce pass — (+1/-1, @whitneywhtsang)
• #6257 Support tensor rank > layout rank — (+80/-21, @whitneywhtsang)
• #6252 : Refactor LoadOpToBlockIOConversion to extract shared helpers — (+364/-274, @etiotto)
• #6246 : Support tt.descriptor_load in remove layout backward propagate. — (+62/-12, @chengjunlu)
• #6227 Fix remainder op stride analysis to prevent incorrect 2D block I/O — (+93/-7, @dev-tomek)

Compiler Pipeline

Pass ordering, layout conversion removal, and IR transformation improvements.

• #6308 [2/3] Update Intel passes to use StrideInfo analysis — (+115/-342, @mieshkiwrk)
• #6188 Reapply "[BACKEND] Refactor forward layout propagation (#9463)" — (+215/-427, @slawblauciak)
• #6334 Enable test_reduce_funky_layout on XPU — (+0/-3, @slawblauciak)
• #6296 Reduce test_pipeline_matmul tensor shape — (+3/-1, @whitneywhtsang)
• #6287 Use common Coalesce pass — (+32/-1170, @whitneywhtsang)
• #6071 Enhance remove layout for issue 5763 (#5948) — (+497/-56, @whitneywhtsang)
• #6217 [1/3] Add standalone StrideInfo analysis with lit tests — (+1354/-1, @mieshkiwrk)

Analysis & Correctness

Bug fixes in stride analysis, error handling, and barrier removal.

• #6330 Fix barrier removal for non-relaxed atomics with unused results — (+44/-12, @etiotto)
• #6303 Fix issue that breaks running the reproducer on CRI. — (+7/-12, @chengjunlu)
• #6306 Fix build failure from StrideInfo — (+3/-3, @mieshkiwrk)
• #6295 [TEST] Use tl.constexpr for strides in test_dot[3d] — (+91/-6, @whitneywhtsang)
• #6282 Fix XPUUtils warning — (+1/0, @etiotto)
• #6263 Fix backward profiling event mismatch in benchmark — (+3/-1, @mieshkiwrk)
• #6258 Fix getStride of mul and div — (+73/-15, @whitneywhtsang)

Test & CI Reliability

Test parameter tuning, CI configuration, and skip list maintenance.

• #6358 [CI] Update DLE upto 2025.3.2 — (+4/-4, @anmyachev)
• #6352 test classes — (+73/-26, @vlad-penkin)
• #6348 Add test_block_load_tdesc.py to skiplist — (+1/0, @anthonycanino)
• #6339 Replace hardcoded skip with programmatic SLM check in test_batched_mxfp — (+7/-2, @etiotto)
• #6337 Enable test_module_load_unload on XPU — (+1/-3, @slawblauciak)
• #6265 Change vllm benchmarks configuration for CRI requirements — (+166/-118, @Egor-Krivov)
• #6314 Only trigger build-benchmarks-wheel on main branch — (+2/0, @whitneywhtsang)
• #6301 Unskip 01-vector-add for B580 — (+0/-2, @kwasd)
• #6310 [SCRIPT] Move run_line_info_tests before use — (+12/-12, @whitneywhtsang)
• #6131 Build benchmarks wheel as GitHub Actions artifact — (+62/0, @januszjah)
• #6294 [NFC] Create utility to run language tests — (+25/-13, @whitneywhtsang)
• #6289 Install cmake from pip in auto-update-translator-cid.yml the same way we do it for benchmarks — (+4/0, @anmyachev)
• #6274 Run matmul tests separately from other language tests — (+69/-67, @whitneywhtsang)
• #6280 Separated vllm into it's own CI and folder — (+313/-125, @Egor-Krivov)
• #6278 Use the most current pip version in setup-pyenv-python action — (+9/-3, @anmyachev)
• #6277 Rebalance test suites — (+2/-4, @whitneywhtsang)
• #6237 Implement both benchmarks as diff rather than reimplemented kernels for easier pin updates — (+1069/-1501, @Egor-Krivov)
• #6264 Build benchmarks before running microbench tests in test-triton.sh — (+2/0, @anmyachev)
• #6270 [TEST] Reduce number of iterations in test_load_store_same_ptr — (+1/-1, @whitneywhtsang)
• #6269 Reduce number of iterations in test_divide — (+7/-8, @whitneywhtsang)
• #6241 Remove third party workflow usage — (+42/-14, @whitneywhtsang)
• #6200 Implement batched moe benchmark as diff rather than reimplemented kernels for easier pin updates — (+677/-689, @Egor-Krivov)

Infrastructure & Tooling

Developer tooling, dependency management, and build infrastructure.

• #6357 Update PyTorch pin — (+1/-16, @anmyachev)
• #6305 [XPU] Use event-less SYCL APIs whenever possible — (+33/-6, @uditagarwal97)
• #6350 Add G31 support (#6133) — (+19/-2, @whitneywhtsang)
• #6335 Don't let tutorials influence each other through environment variables — (+1/0, @anmyachev)
• #6332 Update spirv-llvm-translator.conf — (+1/-1, @glados-intel)
• #6290 Dump all launch kernels commands — (+47/0, @anmyachev)
• #6328 Revert KhronosGroup/SPIRV-LLVM-Translator#3609 — (+254/-72, @whitneywhtsang)
• #6313 Extract TritonGENMemorySpace enum into its own header — (+26/-18, @etiotto)
• #6316 Reland upstream commit 6915ba7 — (+68/-2, @mieshkiwrk)
• #6319 Add --artifact-pattern flag to filter downloaded artifacts — (+69/-1, @vlad-penkin)
• #6307 [XPU] don't perform automatic status checking for viewGetNextRecord function, as we handle it manually — (+1/-1, @anmyachev)
• #6304 Update spirv-llvm-translator.conf — (+1/-1, @glados-intel)
• #6281 Disable example plugins on Windows — (+22/-2, @quinnlp)
• #6284 Remove static from const bool env var declaration — (+2/-2, @whitneywhtsang)
• #6272 Remove transitive TritonGENDialect.h include from TritonIntelGPU/IR/Dialect.h — (+9/-1, @etiotto)
• #6235 [triton_utils] Print long names — (+132/-13, @vlad-penkin)
• #6240 Bump actions/upload-artifact from 6 to 7 — (+26/-26, @dependabot)
• #6211 Limit to row-major — (+61/-19, @whitneywhtsang)

Upstream Alignment

8 upstream merges from OpenAI Triton (commits 6844269, 52e69ac, d28db13, 9a39c7c, 58bc6d3, 4473e29, f7abce7, 9b61f2a) — pass rate 97.53%–97.91%.