v21.0.0

[alexreinking]

Release highlights

We have deliberately skipped version 20.0.0 to align with the LLVM version we are now using. Note that LLVM 21.1.1 or higher is required as LLVM 21.1.0 has a major bug in the NVPTX backend.

Major changes

• The rfactor scheduling directive was rewritten and enhanced. It is now compatible with autoschedulers.
• The Mullapudi2016 autoscheduler now supports experimental GPU scheduling.
• The Python bindings have been substantially improved, with many missing bindings filled in.
• HL_DEBUG_CODEGEN gained a new filtering mode. Debug levels can now be set on a per-file/per-function basis.
• Support was added for AMD Zen5 and the iOS Simulator.
• The strict_float feature has been reimplemented and should be much more reliable.
• Lots of bugfixes, performance improvements, and build system improvements. We spent a lot of time fixing issues with our testing infrastructure and are looking forward to implementing a more stable contribution experience going forward.

Deprecations

• LLVM 19 and below are no longer supported, in keeping with our support policy.
• Halide_BUNDLE_STATIC will be removed in the next release. If you are using it, please migrate to the shared library instead.
• Support for Python 3.8 has been dropped.

Changelog

Scheduling

• The rfactor scheduling directive was rewritten and enhanced.
  • Rewrite the rfactor scheduling directive #8490
  • Dequalify names when constructing RVars in rfactor #8560
  • Add promise_clamped in rfactor #8608
  • Add rfactor patterns for NaN-propagating min/max #8587
• The Mullapudi2016 autoscheduler now supports experimental GPU scheduling.
  • GPU autoscheduling with Mullapudi2016: the reference implementation #7787
  • Mullapudi2016-GPU: Reorder to avoid for-loops to be sandwiched between gpu_blocks. #8647
  • Enable experimental Mullapudi2016 GPU scheduler for test-bench #8650
  • Highlight Metal GPU code in stmt_html #8659
  • Always ensure gpu_threads count >= warp size of 32 #8656
• Fix incorrect natural vector size on Zen4 #8570
• Make it an error to use a device extern stage without target support #8794
• Add support for adding tuple outputs in the configure() method #8649

Python

• Fix argument order in __rpow__ #8677
• Drop support for Python 3.8 #8678
• Fix segfault in RDom's operator<< #8679
• Use ruff to format and lint Python code #8684
• Get raw Runtime::Buffer from Buffer in Python rather than use PyBuffer #8682
• Bind in-place update operators (e.g. +=) in Python #8683
• Clean up Python dependencies; document uv usage #8694
• Fix several printing segfaults. #8700
• Add Python bindings for serialization #8718
• Add all remaining IROperator ops to Python bindings #8771
• Fix up memoize; bind to Python #8778
• Fix invalid Python type annotation and return types (#8772) #8773
• Expose Runtime::Buffer::cropped to C++ and Python Buffer #8787

Debugging

• New feature flag to allow for stack backtrace/unwind #8703
• Add filtering capabilities to HL_DEBUG_CODEGEN #8627
• Adding worker_thread_idle() for more informative profiling #8719
• Color IR output in cout and cerr.  #8635
• Improve output format for lowering passes timing. #8749
• fix(stmt-html): Fix embedded Buffer processing performance issue. #8748
• Use AArch64 assembly syntax on macOS with LLVM<22 #8710

CodeGen

• Mark our PTX kernels as kernels, to stop them from being stripped #8571
• Math functions renaming table for GPU backends to support vectorized evaluation of math functions. #8595
• Apply version constraints to iOS objects #8546
• Redirect bitwise ops to logical ops in case the arguments are bool. #8597
• scalarize select condition for LLVM where possible #8575
• Add missing addition simplifier rules #8630
• Bounds and alignment analysis through bitwise ops #8574
• Make the vld2 pattern more obviously profitable #8765
• [vulkan] Fix vector shuffle for Vulkan CodeGen #8621
• Suppress warning on Windows for duplicate constant symbols. #8555
• Use lossless_cast for saturating casts from unsigned to signed on x86 #8527
• AMD Zen5 support #8612

Compiler

• Rework strict_float to use individual op intrinsics instead #8641
• Don't cache mutations of Exprs that have only one reference to them #8518
• Only use the nodes-visited set for nodes with multiple refs #8547
• In graph_equal(), call the correct implementation for comparing equalities between statements and expressions #8611

Runtime

• Support copying the overlapping region from one buffer to another. #8463
• Add (iOS) simulator target feature. #8623
• Opt out of JIT exceptions #8615
• Experimental: support removing unused runtime functions via HL_RUNTIME_DROP_FUNCS environment variable.
  • PoC feature: drop functions from the runtime #8653

Apps

• The onnx app now builds with CMake:
  • Add CMake for onnx app #8707
  • Fix halide_as_onnx_backend_test #8784

Documentation

• Add note about relative paths to readme #8613

Bugfixes

• Fix #8534 #8535
• Fix CUDA HTML code printing bug. #8558
• [bug] Fix halide_get_cpu_features() linkage to avoid name mangling issues #8573
• Fix for #8578 #8579
• Fix shuffle bug in CodeGen C. #8567
• Check if expression is defined before trying to compute its constant_integer_bounds #8599
• Drop invalid "in-bounds" GEP for constant offsets #8768
• Record trace_loads directly on ImageParam. #8803
• [PTX] RewriteLoadsAs32Bit should use the mutated index #8581
• Set any_strict_float for wrapper module if target has strict_flag feature #8681
• Fix wrong type of the bound #8781
• Fix UB-introducing rewrite in FindIntrinsics #8539
• Fix rewrite that doesn't preserve type #8674
• Fix nested select handling in remove_undef #8669
• Add an underlying type to the halide_buffer_flags to prevent UB in C++ #8690

Testing / CI

• Limit depth more strictly in CSE fuzz test #8512
• Skip fast exp/log/pow/sin/cosine tests without sse 4.1 #8541
• Hopefully fix flaky mullapudi reorder test #8542
• Skip test when code could be using x87 #8537
• [vulkan] Fix stale GPU lifetime management tests for Vulkan. #8601
• Upgrade runner for cmake_cmake_file_lists job #8609
• Buildbot fixes #8706
• Fix the pip packaging workflow #8708
• Fix complexity of bounds of nested pure intrinsics #8689
• Skip two sub-tests on llvm 21.1 #8782
• Speed up simd_op_check_wasm #8780
• Reduce the beam size in the adams2019 apps test to avoid timeouts #8786
• Workaround llvm slow compile time bug in Mullapudi overlap test #8793
• Restore concurrent behavior to gpu_allocation_cache test #8792
• Revert "Skip two sub-tests on llvm 21.1" #8806
• Fix WASM splat op check test. #8705

Build

• Fix workflow for next release #8514
• Fix Debian packaging #8524
• Remove llvm version check from Makefile #8533
• Drop deprecated / unsupported setups for Halide 20 #8508
• Fix check for Windows never having aligned_alloc available. #8551
• Don't include CMAKE_INSTALL_PREFIX when LIBDIR is absolute #8552
• Add target-nvptx to target-all in vcpkg.json #8562
• Fix top of LLVM, and remove upper limit of LLVM version from CMakeLists. #8568
• build_halide_h asserts that every header it slurps in is one of the args #8559
• Upgrade pybind11 to 2.11.1 #8616
• Drop check for LLVM_LIBCXX in FindHalide_LLVM.cmake #8617
• Fix finding LLD on Homebrew when multiple versions are installed. #8619
• Fix build on GCC 15 (Comes with Fedora 42). #8626
• Constrain Clang and LLD searches to LLVM version #8634
• Disallow empty CMAKE_BUILD_TYPE on single-config generators #8651
• Add missing outputs to add_halide_library; fix advice in Lesson 21. #8660
• Bump the LLVM version in the pip package to 20.1.8 #8698
• Prefer to build against libjpeg-turbo and document this. #8775
• Add C++17 requirement to RunGenMain CMake target #8795
• Allow llvm-ar in BundleStatic.cmake #8799
• Fix dubious find_package logic in test/generator #8804
• Warning when extra-output is requested w/o filename #8671
• Makefile linker flag fixes and cleanups #8764

Ongoing maintenance

• Fix clang-tidy-19 errors #8509
• Remove unused function in HexagonOptimize #8511
• Fix two non-idiomatic uses of node_type #8520
• Handle some misc TODOs #8528
• Use a consistent idiom for visit_let #8540
• Upgrade to clang-format 19 #8543
• Suppress clang-tidy warning for make_with_shape_of() #8545
• Remove debugging print left in #8572
• Fixes for llvm trunk #8590
• Another fix for llvm trunk #8591
• Our internal error macros were redesigned:
  • Accurately annotate Error system with [[noreturn]] #8564
  • Teach compilers that internal_error does not return. #8807
  • Use a new macro trick to avoid throwing in destructors. #8774
• Move to opaque llvm pointers #8614
• Avoid throwing from a destructor in PartitionLoops.cpp #8767
• Bump version to 21.0.0 #8810
• remove superfluous overload that causes compile errors #8654
• Remove obsolete WasmExecutor specific debug macro. #8670
• Add missing header #8680
• Attempted fix for LLVM change #8642
• Fix top LLVM: renamed NVPTX barrier intrinsics. #8631

New Contributors

• @changhoon-sung made their first contribution in AMD Zen5 support #8612
• @vawale made their first contribution in Add CMake for onnx app #8707
• @rtzam made their first contribution in Fix invalid Python type annotation and return types (#8772) #8773

Full Changelog: v19.0.0...v21.0.0

---

This discussion was created from the release v21.0.0.