Halide 19.0.0

[alexreinking]

Major improvements

• Halide is now available for both C++ and Python usage via Pip. Try pip install halide today!
• The Vulkan backend has matured substantially.
• The HTML "conceptual statement" output now supports dark mode viewing.
• For developers, CMake 3.28 is now required and we no longer require an internet connection during the build.
• Thread pool improvements mean that workloads that do a small number of small tasks in parallel (e.g. a cheap operation applied to a small image) are up to 3x faster. If you have schedules that do not use parallelism for small inputs because you found it didn't provide any speedup, you may want to re-benchmark.
• You can now query properties of the compiled-for target as Exprs, simplifying helper code that wants to do different things depending on the target architecture. Example: f(x) = select(target_arch_is(Target::ARM), 3, 7). Helpers include target_arch_is, target_os_is, target_has_feature, target_bits, and target_natural_vector_size. These are resolved to constants at compile-time and simplified away. Use with care, as this (intentionally) results in different behavior on different platforms.

Breaking changes

• We now distribute libGenGen.a rather than GenGen.cpp.
  • Downstream users should link to this library with /WHOLEARCHIVE: or -Wl,--whole-archive rather than build GenGen.cpp themselves.
  • Users of the CMake package should be unaffected.
• In keeping with our LLVM support policy, support for LLVM 16 has been removed.
• We no longer use the le64/le32 generic targets for compiling runtime modules to LLVM. These targets were removed in LLVM upstream.

What's Changed

Apps and tests

• Reschedule the matrix multiply performance app #8418 by @abadams
• Update lesson_22_jit_performance.cpp #8438 by @abadams
• Add threadpool performance test #8447 by @abadams
• Don't allow internal_error to pass an error test #8458 by @alexreinking
• Get more consistent distributions in parallel scenarios test #8451 by @abadams

Autoschedulers

• Consider *all* Exprs a func uses, not just the RHS, in Li2018 #8326 by @abadams

Build system

• Python_bindings-test-as-installed #8355 by @LebedevRI
• Bump Halide version to 19 in main branch #8357 by @steven-johnson
• Remove warning for unsupported compilers #8362 by @alexreinking
• Bump CMake minimum version to 3.28 #8363 by @alexreinking
• Quick CMake fixes enabled by 3.28 #8365 by @alexreinking
• Distribute GenGen as a static library #8367 by @alexreinking
• Clean up serialization build code #8369 by @alexreinking
• List headers with target_sources FILE_SETS #8370 by @alexreinking
• Clean up autoscheduler dependencies #8372 by @alexreinking
• Use a Find module for V8 #8373 by @alexreinking
• Use a Find module for NodeJS #8374 by @alexreinking
• Move dependencies/wasm to use sites #8377 by @alexreinking
• Replace FetchContent with a custom dependency provider #8378 by @alexreinking
• Two more build fixes #8371 by @LebedevRI
• Rework LLVM into Find module and enact new component policy. #8379 by @alexreinking
• Reflow src/CMakeLists.txt in logical groups #8383 by @alexreinking
• Introduce HalideFeatures system for optional components #8384 by @alexreinking
• Scan generated export files to determine dependencies. #8385 by @alexreinking
• Rewrite bundle_static to be much more efficient. #8386 by @alexreinking
• Support using vcpkg to build dependencies on all platforms #8387 by @alexreinking
• Fix bundling error on buildbots #8392 by @alexreinking
• Support CMAKE_OSX_ARCHITECTURES #8390 by @alexreinking
• Fix Homebrew LLVM 19 #8431 by @alexreinking
• Fix CPack package naming when cross-compiling #8492 by @alexreinking
• Fix Apple libtool detection in bundle_static #8495 by @alexreinking

CodeGen

• Select condition vector lanes must match the true and false value #8465 by @abadams
• Emit vscale_range() fn attribute in correct syntax #8457 by @steven-johnson
• Fix #8455 (in combination with #8457) #8456 by @steven-johnson
• Fix bonehead mistake in get_md_bool() #8469 by @steven-johnson
• Propagate some facts about inequalities with min/max #8475 by @shoaibkamil
  • This fixed an issue where Specialize does not remove select conditions as expected #8443

Debugging

• Add LLDB pretty-printing #8460 by @alexreinking
• Print constants in scientific precision #8506 by @antonysigma
• Adaptive Dark colorscheme for Stmt HTML. Ability to programmatically export conceptual stmt files. #8327 by @mcourteaux

Documentation

• Update README.md #8404 by @abadams
• Big documentation update #8410 by @alexreinking
  • Document how to find Halide from a pip installation #8411 by @alexreinking
  • Link to PyPI from Doxygen index.html #8415 by @alexreinking
  • Include our Markdown documentation in the Doxygen site. #8417 by @alexreinking
  • Add missing backslash #8419 by @abadams

Frontend

• Don't let users disguise RVars as Vars #8441 by @abadams
• Add helper functions to query properties of the lowered Target (#8192) #8359 by @steven-johnson

Hardware backends

• Fix injection of GPU buffers that do not go by a Func name (i.e. alloc groups). #8333 by @mcourteaux
• Remove vestigial AMDGPU backend #8382 by @alexreinking
• Add ARMv8.x feature flags #4489 by @steven-johnson
• [vulkan] Fixes to address outstanding validation failures  #8448 by @derek-gerstmann
• [vulkan] Reduce descriptor sets, use official headers, improve allocator, remove module destructor #8452 by @derek-gerstmann
• [vulkan] Skip async_copy_chain and gpu_allocation_cache correctness tests on Windows  #8503 by @derek-gerstmann

LLVM

• Don't use le32/le64 #8344 by @steven-johnson
• Fix for the removed DataLayout constructor. #8391 by @mcourteaux
• Drop support for LLVM 16 in main #8358 by @steven-johnson
• Allow LLVM 20 #8352 by @steven-johnson
• Fix for top-of-tree LLVM #8421 by @steven-johnson
• Fix for top-of-tree LLVM #8425 by @steven-johnson
• Fix for top-of-tree LLVM #8442 by @steven-johnson
• Fix datalayout for osx-arm-64 #8449 by @abadams
• Fix top of LLVM. #8454 by @mcourteaux
• Replace all use of getPointerTo() with PointerType::get() #8473 by @steven-johnson

Python

• Fix Numpy 2.0 compatibility bug in lesson 10 #8381 by @alexreinking
• Pip packaging at last! #8405 by @alexreinking
  • Update pip package metadata #8412 by @alexreinking
  • Fix classifier spelling #8413 by @alexreinking
  • Upgrade LLVM to 19.1.0 in pip package #8423 by @alexreinking
  • Update PIP LLVM to 19.1.4 #8488 by @alexreinking
• PythonExtensionGen: ~PyHalideBuffer should call device_free() (#8399) #8439 by @steven-johnson

Runtime

• Fix profiler to report time spent on GPU kernels again instead of on 'wait for parallel tasks'. #8453 by @mcourteaux
• Don't spin on the main mutex while waiting for new work #8433 by @abadams

Minor bugfixes / other cleanup

• Remove remaining dregs of tuple_select (oops) #8329 by @steven-johnson
• Fix incorrect output in Python tutorial, lesson 5 #8331 by @qqaatw
• Make pybind11 minimum version check compatible with pybind11 v3. #8366 by @rwgk
• Partially apply clang-tidy fixes we don't enforce yet #8376 by @abadams
• Fix incorrect std::array sizes in Target.cpp #8396 by @steven-johnson
• Fix _Float16 detection on ARM64 GCC<13 #8401 by @alexreinking
• Make run-clang-tidy.sh work on macOS #8416 by @alexreinking
• Some minor fixes for C++23 compilation errors. #8422 by @zvookin
• Fix two warnings on GCC 14.2.1. #8430 by @mcourteaux
• Fix typos #8459 by @alexreinking
• Fix two trivial build errors #8467 by @steven-johnson
• Fix #8470: fuzz_bounds should use select() rather than Select::make() #8471 by @steven-johnson
• Add missing #include #8476 by @steven-johnson
• Fix heap-use-after-free error in find_best_fit() #8483 by @steven-johnson
• Fix typos in comments #8485 by @alexreinking
• Remove unused is_update argument #8487 by @alexreinking
• Backport reverse_view to clean up some code #8486 by @alexreinking
• Remove type inspection helpers from ApplySplitResult and Split #8489 by @alexreinking
• Use std::optional to clean up some code and prevent use-after-free bugs #8484 by @abadams
• Fix comment for Buffer::copy() (Fixes #8498) #8500 by @steven-johnson
• Make the default constructor for ConstantInterval inlinable #8505 by @abadams
• Remove two unused functions #8501 by @steven-johnson
• Move some large stack frames off recursive paths. #8507 by @abadams

New Contributors

• @qqaatw made their first contribution in Fix incorrect output in Python tutorial, lesson 5 #8331
• @rwgk made their first contribution in Make pybind11 minimum version check compatible with pybind11 v3. #8366

Full Changelog: v18.0.0...v19.0.0