8 Releases published by 1 person

• candidate-20241011.1043 iree candidate candidate-20241011.1043

  published Oct 11, 2024

• candidate-20241012.1044 iree candidate candidate-20241012.1044

  published Oct 12, 2024

• candidate-20241013.1045 iree candidate candidate-20241013.1045

  published Oct 13, 2024

• candidate-20241014.1046 iree candidate candidate-20241014.1046

  published Oct 14, 2024

• candidate-20241015.1047 iree candidate candidate-20241015.1047

  published Oct 15, 2024

• candidate-20241016.1048 iree candidate candidate-20241016.1048

  published Oct 16, 2024

• candidate-20241016.21 iree candidate candidate-20241016.21

  published Oct 16, 2024

• candidate-20241017.1049 iree candidate candidate-20241017.1049

  published Oct 17, 2024

40 Pull requests merged by 18 people

• [Codegen] Drop TransformStrategies

  #18820 merged Oct 18, 2024

• Skip ROCM/test/opt_pass_plugin on Windows while broken.

  #18823 merged Oct 17, 2024

• iree_gpu Python bindings (GPUPipelineOptionsAttr)

  #18804 merged Oct 17, 2024

• [TileSwizzle] Make the dump and variable name match. (NFC)

  #18821 merged Oct 17, 2024

• Warn when --iree-llvmcpu-target-cpu defaults to "generic".

  #18682 merged Oct 17, 2024

• [Flow] Fix FoldSplatReshapeIntoSplat pattern

  #18818 merged Oct 17, 2024

• Opt into free-threaded Python

  #18770 merged Oct 17, 2024

• Enable arithmetic optimizations as part of the stream simplification pipeline.

  #18806 merged Oct 17, 2024

• Fixes a range inference overflow with util.align.

  #18808 merged Oct 17, 2024

• Add util.assume.int folder.

  #18805 merged Oct 17, 2024

• Various tweaks to numeric optimizations found while looking at programs.

  #18765 merged Oct 17, 2024

• Run ONNX model tests as part of pkgci_test_onnx.

  #18795 merged Oct 17, 2024

• Integrates LLVM @ 36d936a2d057ddbd7822614edf01e39a0c21d654

  #18801 merged Oct 16, 2024

• Produce releases for Python 3.13.

  #18799 merged Oct 16, 2024

• Bump to latest stablehlo commit, including MSVC build fix.

  #18797 merged Oct 16, 2024

• Add region to linalg_ext.attention

  #18728 merged Oct 16, 2024

• [GPU] Adding support for opt pass plugins during AMDGPU executable serialization

  #18347 merged Oct 16, 2024

• [DispatchCreation] Extend multi-use producer fusion

  #18551 merged Oct 16, 2024

• [docs] Update and harmonize guids for deployment

  #18762 merged Oct 16, 2024

• Integrates LLVM @ a758bcdbd92efb64a3482eb95d2769d74e33f5bb

  #18783 merged Oct 16, 2024

• Revert tensor.cast to flow reshape conversion

  #18772 merged Oct 16, 2024

• [CMake] Don't update compile definitions for imported targets for MSCV

  #18766 merged Oct 15, 2024

• Bump torch-mlir to 45bb17e

  #18782 merged Oct 15, 2024

• [Codegen] Replace LICM with a version that checks trip count

  #18679 merged Oct 15, 2024

• [ROCM] Fix feature flags for gfx1100 and improve flag handling

  #18781 merged Oct 15, 2024

• [Codegen][GPU] Add tiling cleanup pattern to fuse pad without zero gaurd

  #18748 merged Oct 15, 2024

• Integrate LLVM @ 7900daaa7ba57b5f9729bbbdb54f4e0599a45cd7

  #18773 merged Oct 15, 2024

• Add a default lowering config setting for custom_op.

  #18737 merged Oct 14, 2024

• [LLVMGPU] Add configuration tests for IGEMM, fix NCHW case

  #18734 merged Oct 14, 2024

• [IGEMM] Generate matmuls with expanded H and W dims

  #18735 merged Oct 14, 2024

• [Flow] Fold flow reshape with mismatching dyn dims

  #18680 merged Oct 14, 2024

• [LLVMCPU] Enable tileDispatchUsingForall for multiTilingExpert

  #18730 merged Oct 14, 2024

• [LinalgExt] Remove default implementation for getStaticLoopRanges

  #18745 merged Oct 14, 2024

• [Codegen][GPU] Add pass for fallback distribution

  #18726 merged Oct 13, 2024

• Bump torch-mlir to ab62f35373c3944b68e564214fd04fff39dd92fc

  #18763 merged Oct 11, 2024

• Use integer range and divisibility analysis to propagate int assumptions into dispatch executables.

  #18755 merged Oct 11, 2024

• Adding iree_vm_context_fork to fork a context.

  #18751 merged Oct 11, 2024

• [Codegen] Remove unused arguements from attention op

  #18743 merged Oct 11, 2024

• [Codegen] Remove memref optimizations from OptimizeTensorInsertExtractSlices

  #18732 merged Oct 11, 2024

• Enable analysis based integer optimizations.

  #18756 merged Oct 11, 2024

16 Pull requests opened by 11 people

• [Codegen] Remove wrong usages of OptimizeVectorTransfer

  #18757 opened Oct 11, 2024

• [Codegen] Add pass to convert splat constants to fills

  #18758 opened Oct 11, 2024

• Add end-to-end tests for `iree_linalg_ext.custom_op`.

  #18764 opened Oct 11, 2024

• [LLVMGPU] Combine parallel and reduction padding in LLVMGPUPadAndVectorDistribute

  #18771 opened Oct 14, 2024

• [LLVMCPU] Enable tileDispatchUsingForall as default

  #18777 opened Oct 15, 2024

• [LinalgExt] Generalize attribute setting for attention decomposition

  #18780 opened Oct 15, 2024

• [VectorDistribution] Add scalar support for distributing multi-dim reduction (1/4)

  #18784 opened Oct 15, 2024

• [hip] Added hip_device_group_device to the runtime.

  #18790 opened Oct 16, 2024

• [GPU][WIP] Use TileAndFuse pipeline for non-intrinsic multiple batch matmuls

  #18791 opened Oct 16, 2024

• [GPU] Prefer TileAndFuse pipeline over SIMT pipeline

  #18793 opened Oct 16, 2024

• [Codegen][LLVMGPU] Use `scf.forall` for workgroup distribution in LLVMGPU

  #18796 opened Oct 16, 2024

• [VectorDistribution] Add layout analysis for distributing multi-dim reduction (2/4)

  #18800 opened Oct 16, 2024

• Add generalize matmul pass to sdxl fp16 benchmarks

  #18816 opened Oct 17, 2024

• Document new external ONNX model and linalg operator test suites.

  #18819 opened Oct 17, 2024

• [VectorDistribution] Plumb the VectorDistribute pipeline to support reduction operations (3/4)

  #18822 opened Oct 17, 2024

• [ROCM] OPT Pass plugin test refactoring and organization

  #18824 opened Oct 18, 2024

16 Issues closed by 11 people

• error: 'memref.alloca' op expected no unbounded stack allocations

  #18810 closed Oct 17, 2024

• crash: verifyInvariants failed

  #18809 closed Oct 17, 2024

• Operand `#1` does not dominate this use

  #18815 closed Oct 17, 2024

• error: 'func.func' op exceeded stack allocation limit of 32768 bytes for function. Got 1204288 bytes

  #18776 closed Oct 17, 2024

• Hit an assertion in several onnx models (likely a bug in frontend shape inference)

  #18741 closed Oct 17, 2024

• Windows compiler builds broken with "number of sections exceeded object file format limit"

  #18785 closed Oct 16, 2024

• Does the front-end support onnx-mlir? Why not use Onnx-mlir directly when importing onnx, but use Torch-mlir to convert the onnx file into the torch dialect?

  #18788 closed Oct 16, 2024

• Why is the size of the exported mnist_test.bin file 0 bytes?

  #18787 closed Oct 16, 2024

• Matmul codegen regression compared to MLPerf branch on main.

  #18786 closed Oct 15, 2024

• Deprecate Attention scripts and make C++ pipeline functional

  #18025 closed Oct 15, 2024

• Land Horizontal contraction fusion changes into IREE main

  #18009 closed Oct 15, 2024

• [HIP] f8e4m3fnuz matmul compile failure for RDNA3

  #18769 closed Oct 14, 2024

• canonicalize dropping the `lowering_config` attached to an operation.

  #18697 closed Oct 14, 2024

• Support different access pattern for iree_linalg_ext.attention

  #18768 closed Oct 14, 2024

• Add `iree_vm_context_fork` to support forking contexts while reusing resources.

  #18747 closed Oct 11, 2024

• Numerical inaccuracies in multi-device sharded toy-sized Llama

  #18687 closed Oct 11, 2024

17 Issues opened by 12 people

• Add Apple GPU runners and run Metal tests again

  #18817 opened Oct 17, 2024

• Add NVIDIA GPU runners and run CUDA tests again

  #18814 opened Oct 17, 2024

• Run Windows build/test workflows more regularly

  #18813 opened Oct 17, 2024

• Does IREE support native cuda buffer VM invocation or is there any chance to improve the performance of result torch tensors construction?

  #18811 opened Oct 17, 2024

• Does IREE currently support large language models?

  #18807 opened Oct 17, 2024

• Rework iree-compile pipeline for ONNX

  #18803 opened Oct 17, 2024

• In-place argument updates with device affinities fails

  #18802 opened Oct 16, 2024

• [Codegen][AMDGPU Backend] Correctness issue for conv_2d_ngchw_gfchw

  #18798 opened Oct 16, 2024

• Rename `iree-hip-` compiler flags to `iree-rocm-` when they apply to codegen.

  #18792 opened Oct 16, 2024

• [GPU] : HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION in GPU while passing inference in CPU

  #18789 opened Oct 16, 2024

• HoistIntoGlobals fails to hoist constantOp

  #18779 opened Oct 15, 2024

• Bufferization needs tensor.extract_slice to bufferize in place

  #18778 opened Oct 15, 2024

• When importing onnx model, do I convert onnx to torch dialect?

  #18775 opened Oct 15, 2024

• CUDA out of memory due to huge memory allocation request

  #18767 opened Oct 13, 2024

• Some friction when using Rye/uv for package management.

  #18761 opened Oct 11, 2024

• [ROCM] Evaluate whether we can attach `amdgpu-no-implicitarg-ptr` to our generated functions.

  #18760 opened Oct 11, 2024

• Robustify upstream LICM for zero-trip count loops and other loop kinds

  #18759 opened Oct 11, 2024

20 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.

• [VectorDistribution]reduction support along LLVMGPUVectorDistribute pipeline

  #18519 commented on Oct 15, 2024 • 11 new comments

• [compiler][flow] Move cast, reshape and bitcast after transfer op

  #18742 commented on Oct 15, 2024 • 10 new comments

• [LLVMGPU] Use forall workgroup distribution in TileAndFuse pipeline

  #18565 commented on Oct 17, 2024 • 5 new comments

• Add conversions for 1x1 conv_2d to matmul

  #18736 commented on Oct 17, 2024 • 4 new comments

• [Codegen] Replace LLVMGPUVectorize with LLVMGPUTileAndFuse for supported cases

  #18474 commented on Oct 15, 2024 • 2 new comments

• Build and publish Python 3.13 and 3.13t wheels

  #18652 commented on Oct 15, 2024 • 0 new comments

• [gpu] 'func.func' op uses 401920 bytes of shared memory; exceeded the limit of 65536 bytes

  #18603 commented on Oct 15, 2024 • 0 new comments

• 'util.initializer' op failed to inline into combined initializer

  #18386 commented on Oct 15, 2024 • 0 new comments

• Migrate jobs off current GCP GHA runner cluster

  #18238 commented on Oct 17, 2024 • 0 new comments

• Release tracker - 2024/09

  #18432 commented on Oct 18, 2024 • 0 new comments

• stablehlo.sort not working correctly on Metal

  #18698 commented on Oct 15, 2024 • 0 new comments

• [Compiler] Rename compiler ROCM target to HIP

  #18477 commented on Oct 17, 2024 • 0 new comments

• onnx.Pad with mode 'reflect' fails to compile

  #18695 commented on Oct 15, 2024 • 0 new comments

• [codegen] [gpu]: SD3 MMDiT attention dispatch fails on LinalgExtToLoops for amdgpu targets

  #18629 commented on Oct 14, 2024 • 0 new comments

• [compiler] strip execution context affinities in const eval

  #18663 commented on Oct 16, 2024 • 0 new comments

• Support i1 datatype

  #18713 commented on Oct 15, 2024 • 0 new comments

• [GPU] Support multiple contraction dims in MmaSchedules

  #18720 commented on Oct 17, 2024 • 0 new comments

• [DT] Plans for the buffer allocation in data-tiling

  #17924 commented on Oct 11, 2024 • 0 new comments

• DNS - check bazel deps

  #18738 commented on Oct 12, 2024 • 0 new comments

• [GlobalOptimization] 1x1 filter convolutions not converted to matmul

  #18710 commented on Oct 11, 2024 • 0 new comments