What's Changed

• [DT] Fuse encoding ops more aggressively for multi-use, gather, and slices ops. by @hanhanW in #21830
• [Codegen][Tuner]: improve python binding to query target info by @bangtianliu in #21812
• [Codegen][Tuner] retire the C/Python binding for querying mma intrinsic. NFC. by @bangtianliu in #21816
• [Integrate] Drop llvm/llvm-project@b4c31dc revert. by @hanhanW in #21851
• [Encoding] Support SetEncoding on scaled contraction ops by @Max191 in #21825
• [Test] Add onnx_ops test suites with O2/O3 optimization level. by @hanhanW in #21838
• [CodeGen] Do not fuse parallel ops if they directly write to destination. by @hanhanW in #21837
• [GPU] Add pattern to fold fill into pad ops by @nirvedhmeshram in #21864
• [Codegen][IGEMM] Do not pre-pad convs with CHW layout or small input channel size by @yzhang93 in #21839
• [GPU] Remove reshape by expansion in workgroup scope of combine layout pass by @nirvedhmeshram in #21869
• [CPU] Remove passing tests from expected_compile_failures list. by @hanhanW in #21871
• [GPU] Use Affine map for size calculations of alloca's in fission pass by @nirvedhmeshram in #21870
• [Codegen][AMDGPU] Fix matmul miscompile on RDNA4 by @kuhar in #21873
• [NFC] Code Quality changes by @Muzammiluddin-Syed-ECE in #21876
• Avoid needles isa checks. NFC. by @kuhar in #21885
• [VectorDistribute] Refactor layout configuration to a simpler logic by @Groverkss in #21883
• [StableHLO][CHLO]Refactor CHLO decompositions to follow upstream StableHLO by @LekkalaSravya3 in #21682
• Revert "[VectorDistribute] Refactor layout configuration to a simpler logic" by @Groverkss in #21887
• [docs] Clarify compiler coding standards by @kuhar in #21886
• Upgrade Preprocessing and Modules to free create functions. NFC. by @kuhar in #21877
• [Codegen] Upgrade Common, SPIRV, VMVX to free create functions. NFC. by @kuhar in #21879
• [Codegen] Upgrade LLVMCPU and LLVMGPU to free create functions. NFC. by @kuhar in #21880
• [Codegen] Upgrade Dialect and Interfaces to free create functions. NFC. by @kuhar in #21881
• Add gfx950 ukernel patterns by @sebvince in #21856
• Bump version to 3.8.0 after 3.7.0 release. by @sa-faizal in #21852
• [docs] Update the file config file for running ONNX operator tests on CPU. by @hanhanW in #21892
• Upgrade GlobalOpt, InputConversion, ExternalInterfacess to free create function. NFC. by @kuhar in #21878
• [Codegen] Upgrade Transforms and Utils to free create functions. NFC. by @kuhar in #21882
• [ROCM] Update Ukernel infra to handle InnerTiledOp/Multi_MMA_MFMA by @Abhishek-Varma in #21759
• Reland "[VectorDistribute] Refactor layout configuration to a simpler logic" by @Groverkss in #21895
• Upgrade IREE plugins to free create functions. NFC. by @kuhar in #21896
• [GPU] Remove MMAScheduleAttr by @Groverkss in #21884
• [LLVMCPU] Respect dominance when doing replacement of tile and fused values by @MaheshRavishankar in #21901
• [Codegen] Upgrade iree dialects to free create functions. NFC. by @kuhar in #21898
• Integrate LLVM at llvm-project/llvm@daf8f9fc1ccc6c5679bc89058fd66d8ea4da9d59 by @rkayaith in #21893
• Upgrade all remaining code to free create functions. NFC. by @kuhar in #21902
• [LLVMGPU] Move LLVMGPUVectorLowering after OptimizeIntArithmetic by @Max191 in #21597
• [Codegen] Promote scales to LDS by @Muzammiluddin-Syed-ECE in #21767
• Bump the github-actions group with 2 updates by @dependabot[bot] in #21897
• Integrate llvm/llvm-project@31bee3421ba4 by @rkayaith in #21905
• [CPU] Tile all the ops to target vector sizes before vectorization. by @hanhanW in #21900
• [LinalgExt] Fold subview ops into map_scatter output before decomposing by @Max191 in #21891
• [GPU] Do not do c promotion for unaligned (I)GEMMs by @nirvedhmeshram in #21823
• [Codegen][ROCm] Add repro instructions for .rocmasm files by @kuhar in #21874
• [LinalgExt] Fix FoldWithProducerReshapeByExpansion for >1 dyn dim by @IanWood1 in #21894
• At the beginning of emulate narrow type, flatten incoming memrefs by @lialan in #21910
• Revert "Disable failing ARM-SME tests. (#21715)" by @banach-space in #21860
• [Codegen][AMDGPU] Drop backend reverts, emergency RDNA4 lowering fix by @krzysz00 in #21906
• [codegen] more consumer fusion by @jtuyls in #21848
• [CPU][DT] Add codegen support for broadcast/dequant -> matmul dispatch. by @hanhanW in #21911
• [Codegen][IGEMM] Set convolution pre-padding as default by @yzhang93 in #21899
• [Codegen][GenericVectorization] Fix incorrect usage of std::accumulation that led to overflow by @mshockwave in #21920
• Integrate llvm/llvm-project@e92cbfbe3087 by @rkayaith in #21917
• [Codegen][Cleanup] Always enable vectorization for padding and gather. by @hanhanW in #21924
• [Test] Disable AMDGPU onnx_ops test suite (O0) job. by @hanhanW in #21929
• Integrate llvm/torch-mlir@7000187b by @rkayaith in #21918
• Bump nanobind version by @Hardcode84 in #21926
• [iree][codegen] Add #iree_codegen.denormal_fp_math to set denormals behavior by @fabianmcg in #21840
• [ROCM] Add back specialization pattern tests by @jtuyls in #21939
• Fix --iree-hip-target validation by @bjacob in #21909
• Integrate LLVM at llvm/llvm-project@b22f94dcc58e by @rkayaith in #21943
• [GPU] Propagate extract slice when doing convolution padding by @nirvedhmeshram in #21948
• [CPU] Adjust tile sizes for mmt4d dispatches that have relayout ops. by @hanhanW in #21934
• Fixing CSE of hoisted encoding ops. by @benvanik in #21921
• Adding util.list.construct pseudo-op. by @benvanik in #21950
• [Dispatch Creation] Don't fuse no input producer with reduction by @IanWood1 in #21930
• Revert "[LinalgExt] Fix FoldWithProducerReshapeByExpansion for >1 dyn dim" by @IanWood1 in #21947
• [DispatchCreation]: Add FormSplitReductionDispatchesPass support for ArgCompare op by @bangtianliu in #21903
• [CPU] Add an experimental flag to disable linalg.conv generalization. by @hanhanW in #21953
• [GPU][DT] Add pingpong ukernels for data tiling (f8 and f16) by @Yu-Zhewen in #21919
• [ROCM][DT] Add encoding specialization infra for data-tiled ukernels by @jtuyls in #21914
• [GPU] Use UkernelDescriptor and deprecate UkernelConfigAttr and GPULowerToUkernelsPass by @Abhishek-Varma in #21766
• [docs] Update docs on sdxl golden output by @efric in #21936
• Fix Dispatch Creation TransformOptions by @IanWood1 in #21964
• Integrate LLVM at llvm/llvm-project@ed1f1b8 by @rkayaith in #21963
• [docs] Add a blog post for data-tiling introduction. by @hanhanW in #21774
• Avoid needless isa checks. NFC. by @bangtianliu in #21968
• using Base::Base in tablegen passes. by @benvanik in #21969
• [iree][codegen] Set #iree_codegen.denormal_fp_math in attention dispatches by @fabianmcg in #21940
• [compiler][NFC] Update remaining code to free create functions. by @hanhanW in #21972
• [plugins][NFC] Upgrade plugins/ to free create functions. by @hanhanW in #21973
• [GPU][DT] Update data layout strategy for pingpong ukernels by @Yu-Zhewen in #21957
• Using explicit operation types in passes. by @benvanik in #21971
• Converting compiler/Bindings/ to tablegen Passes.td. by @benvanik in #21974
• [Codegen] Unroll instead of linearize vector.to_elements. by @amd-eochoalo in #21959
• [Codegen] Added erf ; FastMath rewrite for vector types. by @keshavvinayak01 in #21849
• Adding hal.executable lazy flag. by @benvanik in #21966
• Don't inline immutable globals with non-util dialect attrs. by @benvanik in #21986
• [Codegen][RISCV] Do not lower vector.gather to branches in the presence of RVV by @mshockwave in #21927
• [GPU] Only combine complex relayout chains in GPUCombineLayoutTransformation by @Max191 in #21985
• [LLVMGPU] Move masked load optimizations after vector lowering by @Max191 in #21962
• [iree-test-suites] Update golden benchmark numbers by @Max191 in #21980
• [Encoding] Deprecate MatmulKAttr encoding attribute. by @hanhanW in #21976
• [Codegen] Make collapse_shape hoisting pattern work with store_to_buffer by @Max191 in #21999
• [LinalgExt] Add canonicalization to convert identity map_scatter to copy by @Max191 in #21998
• [GPU][DT] Add benchmark files for llama_8b_f16 with data-tiling. by @hanhanW in #21975
• [Encoding] set default option for scaled matmul encodings to false by @Muzammiluddin-Syed-ECE in #21994
• Marking stream.async.dispatch as pure. by @benvanik in #21989
• [Dispatch Creation] Allow fusing pad with split reduction dispatch by @IanWood1 in #21987
• Fix data race in GPU C ukernels caching of shared memory size by @bjacob in #22004
• [LinalgExt][NFC] Remove unused code in TransposeFusion by @IanWood1 in #22006
• [Dispatch Creation] Rework dispatch formation logic by @IanWood1 in #21854
• [TensorExt] Fix dynamic dim canonicalization in bitcast folder by @jtuyls in #21997
• [CPU] Switch matmul config to use linalg::LinalgOp interface. by @hanhanW in #21954
• Bump the github-actions group with 2 updates by @dependabot[bot] in #21992
• Integrate LLVM at llvm/llvm-project@0648c5183f32 by @qedawkins in #22003
• [VectorDistribute] Use subgroup_basis instead of subgroup_m/n_count by @Groverkss in #21912
• Splitting hoisted async constant lifetime. by @benvanik in #21995
• Adding iree_hal_executable_export_info_t and queries. by @benvanik in #21754
• Respect user FILECHECK_OPTS/LIT_OPTS environment variables when running through ctest by @rkayaith in #22019
• [PassUtils] Allow passing overload constructors to addPredicatedPass by @rkayaith in #22021
• [NFC] remove unused header files by @bangtianliu in #21977
• [Codegen][GPU] Enable TileAndFuse for matmul by default by @jerryyin in #21834
• [CPU] Populate to_elements unrolling patterns in LLVM conversion. by @hanhanW in #22010
• Fix mi308 Pkgci failures by @IanWood1 in #22028
• [Dispatch Creation] Fuse bcast with attention instead of producer by @IanWood1 in #22008
• [CPU] Add precondition to kernel dispatch method selection for gemm. by @hanhanW in #22031
• [Codegen][Tuner] update lowering config binding for subgroup basis by @bangtianliu in #22027
• Fix indices in scaled matmul rank assert by @jtuyls in #22016
• [GPU] Introduce Intentional Padded Configurations for (I)GEMM by @nirvedhmeshram in #21931
• [CI] Disabling WebGPU build due to CI failures. by @MaheshRavishankar in #22030
• [DispatchCreation] Add option to infer split-reduction sizes for batchnorm by @rkayaith in #21731
• Implement iree_gpu.coalesced_gather_dma op by @lialan in #21846
• [LinalgExt] Support map_scatter decomposition with strided memrefs by @Max191 in #21952
• [Codegen] Tile map_scatter op for large vector sizes by @Max191 in #22035
• [DispatchCreation] Fix iree-compile split-reduction flag name by @rkayaith in #22038
• Integrate LLVM at llvm/llvm-project@dffd7f3d9a3 by @qedawkins in #22023
• [NFC][ROCM] Refactor bitcode ukernel to a separate file by @Abhishek-Varma in #21983
• [VectorDistribute] Allow distributing subgroups on multiple m dimensions by @Groverkss in #22000
• [LLVMGPU] Vectorize map_scatter in LLVMGPUTileAndFuse pipeline by @Max191 in #21890
• [Codegen] Push up memref reshapes before empty tensor elimination by @Max191 in #22045
• [LLVMGPU] Add support for direct convolution in tile and fuse pipeline by @yzhang93 in #22033
• LLVM-Integrate: Drop revert for f645d209d by @qedawkins in #22044
• Integrate llvm/llvm-project@50ef746a12 by @qedawkins in #22046
• [Encoding] Propagate layout encodings through tensor.cast ops by @Max191 in #21970
• [python] Expose python bindings for nvvm in iree.compiler.dialects by @saladpalad in #21993
• Revert "[Dispatch Creation] Rework dispatch formation logic (#21854)" by @IanWood1 in #22058
• [Codegen] Add transform ops for matching contraction ops by @bangtianliu in #21981
• Integrate llvm/llvm-project@1ee18959bcdf by @efric in #22062
• Disable data-tiling flag by default and refresh the CPU docs. by @hanhanW in #21935
• [DispatchCreation] Propagate and fuse encodings in default path by @Max191 in #22063
• [LinalgExt][NFC] Remove unused VectorOps include by @hanhanW in #22066
• Integrate llvm/llvm-project@9d48df7a92e7 by @efric in #22064
• [LLVMGPU] Enable iree-llvmgpu-test-combine-layout-transformation by default by @Max191 in #21979
• Adding interface support for stream.async.transfer result placement. by @benvanik in #22048
• Marking stream.tensor.dispatch pure. by @benvanik in #22075
• [GPU][DT] Add data-tiling resolver by default. by @hanhanW in #22074
• [Util] Allow varying types in optimization barrier by @qedawkins in #22076
• [ROCm] Add an experimental target for gfx1250 by @kuhar in #22077
• Remove e2e matmul tests with explicit compilation-info by @bjacob in #22085
• [NFC] Improving consistency of Util/Transforms/Passes.h. by @benvanik in #22078
• e2e matmul tests covering vector-distribution by @bjacob in #22086
• Reapply "[LinalgExt] Fix FoldWithProducerReshapeByExpansion for >1 … by @IanWood1 in #22088
• [Test] Trim data-tiling compile flags from tests. by @hanhanW in #22092
• [Codegen] Support scaled matmul in ConvertAccGEMMToGEMM by @Max191 in #22093
• [GPU] Fix bug in shared memory computation for scaled intrinsics by @Max191 in #22095
• Integrate llvm/llvm-project@876296e9b7f0 by @efric in #22097
• [Codegen] Add transform op for matching dimension sizes. by @bangtianliu in #22040
• Revert e2e matmul tests changes by @bjacob in #22111
• [mlir][amdgpu] Replaced nullopt with target arch chipset in populateGpuPromoteShuffleToAMDGPUPatterns pass by @xintin in #21799
• Display a warning when we spill SGPRs or VGPRs by @sebvince in #21863
• [Codegen][GPU] Fix MMA Intrinsics Sorting by @bangtianliu in #22090
• [Codegen][GPU][NFC] Fix mma sort follow up by @bangtianliu in #22122
• [DT] Add support for materializing func.func args with encodings. by @hanhanW in #22115
• Break generate_e2e_matmul_test.py into multiple files by @bjacob in #22120
• NFC: Simplify generation of e2e matmul test functions. by @bjacob in #22123
• [ROCm] Fix up gfx1250 definitions by @kuhar in #22131
• Clean up KnownTargets.cpp. NFC. by @kuhar in #22133
• Fix the Windows build: portably set environment variable PYTHONPATH. by @bjacob in #22136
• [Codegen][ROCm] Attempt to fix MMA sorting CI failures by @kuhar in #22141
• [NFC][GlobalOpt] Update function names in LIT by @AGindinson in #22083
• [VectorDistribute] Flush denormals for attention reduction config by @Groverkss in #22041
• Simplify op conversion pattern inheriting constructor definitions. NFC. by @kuhar in #22143
• Simplify op rewrite pattern inheriting constructor definitions. NFC. by @kuhar in #22142
• [LLVMGPU] Don't use DMA for scaled matmul by @Max191 in #22094
• [Codegen] Add bufferization support for new iree_gpu.coalesced_gather_dma op by @lialan in #22049
• [GPU] Support iree_codegen.load_from_buffer in GPUBubbleResourceCasts by @Max191 in #22140
• [Preprocessing] Add pass to sink transpose through pad by @IanWood1 in #22106
• [LLVMGPU] Unroll elementwise operations by @Groverkss in #21665
• Bump actions/cache from 4.2.4 to 4.3.0 in the github-actions group by @dependabot[bot] in #22152
• Increase golden time. by @amd-eochoalo in #22159
• [Codegen][AMDGPU] Fix incorrect canonical map for MXFP RHS scales by @krzysz00 in #22162
• [Preprocessing] Transpose conv filter layout from CHWF to FHWC by @yzhang93 in #22100
• Integrate llvm/llvm-project@7af31bf by @amd-eochoalo in #22148
• [LLVMGPU][Codegen] Increase parallel rows read for matvec by @efric in #22163
• [Codegen] support matching any values for dims_equal transform op by @bangtianliu in #22149
• Integrate llvm/llvm-project@a33544b by @amd-eochoalo in #22167
• E2E MXFP4 matmul tests by @bjacob in #22170
• [Test][NFC] Drop input_type from e2e tests because IREE can infer the input type. by @hanhanW in #22014
• [DispatchCreation] infer split-reduction sizes for ArgCompare by @bangtianliu in #22154
• [Codegen][AMDGPU] Tile and convert gather to coalesced DMA by @lialan in #22157
• Revert "[LLVMGPU] Unroll elementwise operations (#21665)" by @MaheshRavishankar in #22186
• Integrate llvm/llvm-project@0cb9d40 by @amd-eochoalo in #22182
• Port e2e matmul tests from gfx942 to gfx950 by @bjacob in #22191
• [E2E-Matmul] Remove redundant flag from scaled matmul e2e test by @Max191 in #22190
• [Codegen][AMDGPU] Enable gpu.printf patterns by @krzysz00 in #22192
• [GPU] Add thread tile size inference for map_scatter op by @Abhishek-Varma in #22179
• [DT][ROCM] Fix inner_tiled bitcode ukernel lowering with instrinsicsM(N) = 1 by @Yu-Zhewen in #22184
• [Codegen] Add ResolveShapedTypeResultDimsPass pass to GPU vector distribute by @fabianmcg in #22196
• [LinalgExt] Introduce linalg_ext.exp_reduction by @hhkit in #21761
• Integrate llvm/llvm-project@4845b3e by @amd-eochoalo in #22200
• [GPU] Allow multi result and indexing compute generic ops in TilleAndFuse pipeline by @nirvedhmeshram in #22195
• [Codegen][GPU] Fix IGEMM pre-padding and fusion patterns by @yzhang93 in #22197
• [DataTiling] Introduce DataTiledMMAInterfaceAttr by @Max191 in #22098
• [Codegen] Follow-up Fix for MatchContractionOp by @bangtianliu in #22201
• Removing virtual MMAs from vector distribute matmul/conv pipeline by @jerryyin in #22202
• [Codegen] Add transform op for matching convolution ops by @bangtianliu in #22194
• Revert "[GPU] Allow multi result and indexing compute generic ops in TilleAndFuse pipeline" by @IanWood1 in #22205
• [Codegen] Fix premature return in iree_codegen.inner_tiled verifier by @Max191 in #22183
• [Codegen][LLVMGPU] Later scf-to-cf to support math.erf by @newling in #21817
• [CI] Add dummy torch pkgci by @Groverkss in #22203
• [DataTiling][GPU] Introduce DataTiledScaledMMAAttr by @Max191 in #22176
• [Preprocessing] Reorder the iterator dims to match forward NHWC-FHWC convs by @yzhang93 in #22208
• Decrease llama 8b_f16_decode golden time by @efric in #22220
• [DataTiling][GPU] Implement scaled matmul data tiling materialization by @Max191 in #22189
• Integrate llvm/llvm-project@95e0ae9f by @newling in #22214
• Control const expr hoisting in Dispatch Creation by @IanWood1 in #22164
• [codegen] Fix test after PR 22196 by @fabianmcg in #22218
• [codegen][gpu] GPUApplyPaddingLevel: fold case where no padding by @newling in #22193
• [GlobalOpt] Use Option<> for TransformOptions by @IanWood1 in #22222
• [ROCm] Enable e2e stablehlo tests by @kuhar in #22224
• Adding a LiftCFGToSCFPass. by @benvanik in #22101
• Improving support for unreachable control flow in both CFG and SCF. by @benvanik in #22102
• Adding VerifyStructuredControlFlowPass. by @benvanik in #22110
• Fix g++ warning -Werror=parentheses by @IanWood1 in #22225
• Update CODEOWNERS to include new tests and dialect owners by @Groverkss in #22213
• [CI] Add clip and llama torch_models tests by @Groverkss in #22212
• [samples] Update PyTorch JIT notebook for Python 3.12 by @HeatCrab in #22209
• [Codegen] add transform op for matching attention op by @bangtianliu in #22199
• Fix linking MSVC error from forward declaration used as templated type by @Max191 in #22233
• [PkgCI] Use urllib instead of github cli in pkgci artifact_run by @Groverkss in #22211
• Transposed Workgroup Reordering for large rectangular matmuls by @sebvince in #22165
• Fix typo ConditionalTranspose attribute description by @sebvince in #22238
• [docs] Add LLVM debugging and some AMDGPU-specific tips by @krzysz00 in #22146
• Integrate llvm/llvm-project@7546bd3 by @newling in #22234
• [CI][iree-test-suites] Add random weight 8b_fp8 and 8b_fp16 benchmarks by @Groverkss in #22239
• Integrate llvm/llvm-project@327a89c by @newling in #22255
• [CI][iree-test-suites] Upload json summary for torch_models CI by @Groverkss in #22253
• [CI][iree-test-suites] Update ref for iree-test-suites by @Groverkss in #22263
• [build flags] prepare to enable more warnings in compile flags (#21996) by @schuermans-roofline in #22252
• [Codegen] Update the assembly formats and corresponding tests for matcher ops by @bangtianliu in #22270