Skip to content

Releases: Xilinx/mlir-aie

v1.2.0

23 Jan 17:41
fd4ed98

Choose a tag to compare

Fixup a bug where the i8 matmul was using opcode 8 which means unsign…

v1.2.0-alpha0

19 Jan 18:53

Choose a tag to compare

more stringent parsing

v1.1.4

03 Dec 17:35
84c0522

Choose a tag to compare

Bump llvm version to 49d5bb0a (#2746)

v1.1.3

21 Nov 23:37
19cce64

Choose a tag to compare

python wheels: fix author field (#2728)

v1.1.2

14 Nov 16:24
f718101

Choose a tag to compare

What's Changed

  • AIE Vectorization Updates: Updated lowering strategy of aievec.srs to llvm.
  • AIE Kernel Management: Enabled aie_kernels includes and cleaned up CMake scripts.
  • Driver and Toolchain Updates: Updated drivers to version 1.6 and made XRT binaries available in PATH for LIT testing.
  • ELF Support Enhancements: Added full ELF flow support with load_pdi instruction and usage of ELF for instructions in some programming examples.
  • AIE Vector Operations: Added support for AIE2P vector operations in -convert-aievec-to-llvm and AIE2P floating point and integer matmuls.
  • Control Packet Lowering: Implemented control packet lowering for aiex.configure and created control packet ops from transaction ops as a pass.
  • Code Refactoring: Refactored detection and common logic into helper utilities.
  • Bug Fixes: Fixed issues related to bf16 gemm implementations, adjacent tile allocation, and getLowerBoundValue method.
  • New Features and Updates: Added support for Vitis 2025.1, updated Memcopy Exercise, and introduced Unplaced GEMM for Large Workloads.
  • Bug Fixes: Fixed case bug and vector_scalar_mul placed lit test.
  • Feature Enhancements: Added support for reading AIE registers via ctrl packet configuration and SplitVectorLoadUpsChains pattern for optimizing wide vector operations.
  • Code Improvements: Updated env_setup.sh, integrated python test lit config, enabled DataLayoutOpInterface, and switched to using bundled convert-to-llvm pass.
  • Vectorization Optimization: Added AIEHoistVectorTransferPointersPass and AIEVectorToPointerLoops to optimize vector reads/writes and lower vector.load/store on loop-carried indices.
  • Code Optimization: Added llvm-loop-opt pass to optimize generated LLVM IR and enabled recent passes in aiecc.py to optimize vectorized code’s schedule.
  • Python Runtime Refinement: Refined default Python runtime sequence behavior and removed experimental code.

New Contributors

Full Changelog: v1.1.1...v1.1.2

v1.1.1

06 Oct 20:27
f4b1d8c

Choose a tag to compare

What's Changed

  • AIE Runtime Removal: Removed Vitis aie-rt usage and xbutil in lit script.
  • Project Renaming and Documentation: Renamed mlir_tutorials to mlir_exercises and added getting started programming examples.
  • Code Improvements: Fixed various bugs, including tag issues, AIEPut/GetCascade casting, and object fifo lowering crash.
  • AIE Vector Optimization: Further optimized vector version for back to back Macs.
  • AIE Wheel Update: Add parse_trace.py to mlir-aie wheel.
  • AIE Driver Patch: Patch xdna-driver name in build_drivers.sh script.
  • AIE Compiler Enhancements: Added support for vector.reduction, multiple devices in one design, and device configuration.
  • Bug Fixes and Optimizations: Fixed issues related to arith.subf, vector.reduction, and router non-determinism.
  • Documentation and Examples: Updated README.md, added examples, and documented JIT vs. non-JIT.

Full Changelog: v1.1.0...v1.1.1


For a detailed view of all recent changes, you can explore the mlir-aie GitHub repository.


Utilizes:

latest-wheels-no-rtti

09 Sep 04:18
faa0c07

Choose a tag to compare

Fix tag to not have slash (#2569)

v1.1.0

08 Sep 15:35
1ad71ae

Choose a tag to compare

What's Changed

  • Bug Fixes: Several bug fixes were implemented, including correcting data layout translation, unplaced vector reduce max designs, and missing template arguments in programming examples.
  • Library Enhancements: Added multi-core RMSnorm, rotary positional embedding, Layernorm, and external C functions support for JIT. Performance improvements were made to col maj gemm and multicore SILU/GELU.
  • Dependency Improvements: Separated development-related packages, updated xdna-driver/xrt, and added python host code for elf test. Removed PATH and PYTHONPATH manipulation from scripts and wheel installation.
  • Python Version Addition: Added Python 3.11 to the wheels action.
  • Neighboring MemTile Feature: Added a neighboring MemTile DMA test with object fifo of depth=3.
  • Support for Preemption: Preemption operation to transaction sequence.
  • AIE-RT Installation Directory: Allow specifying the installation directory of AIE-RT in the build-mlir-aie.sh script.
  • Tests for XRT Runlist: Added two kernel tests for the xrt::runlist functionality.
  • ObjectFIFO Pass Update: @AndraBisca implemented packet flow emission in ObjectFIFO pass (#2420).
  • Packet ID Bug Fix: Fixed a bug where the packet ID was incorrectly set to 5 bits.
  • LLVM Version Update: @erwei-xilinx updated LLVM to 064f02da (#2560).

New Contributors

Full Changelog: v1.0.1...v1.1.0


For a detailed view of all recent changes, you can explore the mlir-aie GitHub repository.


Utilizes:

v1.0.1 - Incremental release

18 Jul 15:52
3b4ecb7

Choose a tag to compare

What's Changed

  • Package Management: Add dependency tracking to wheels.
  • Documentation Update: Update programming guide and more examples for NPU2 Strix, Strix Halo and Krackan devices.
  • Pragma Unification: Unify pragmas to support Chess and Peano.
  • Tracing Enhancements: Added event options to trace, updated trace instructions in the programming guide, and fixed a trace parser bug.
  • Compiler Modifications: Removed uncaptured dependence in Conversion passes, replaced GlobalBuffer with LocalBuffer, moved BD address calculation to AIETargetModel, extended BFP16 support, and renamed ObjectFifo default_depth to depth.
  • Programming Examples: Addition of a new transpose example with support for varying sizes and data types, and vector reduce max single column designs showcasing multiple data movement patterns.
  • Windows Support Enhancement: Updated Windows support instructions, bumped MLIR wheel runners to Windows 2022 and Python 3.12, and updated Windows workflows to version windows-2022.
  • Testing Improvement: Flag slow tests in test names and remove -j1 flag from lit opts.
  • Device Driver Updates: Removed the Five-Column Phoenix Device.
  • Tutorial Slides: Added IPDPS and ISCA 2025 presentations.

New Contributors

Full Changelog: v1.0...v1.0.1


Utilizes:

latest-wheels-2

17 Jul 17:21
c5b3be6

Choose a tag to compare

Use a new release for nightly wheels (#2437)