94ad542 | Steven Johnson | 13 March 2023, 18:16:04 UTC | Two tweaks to runtime headers - printer.h uses `uint64_t`, so it needs to include something that ensures that type is defined. `HalideRuntime.h` is probably the right choice (since it always transitively includes `runtime_internal.h` when compiling runtime. - HalideBuffer.h should completely elide itself when `COMPILING_HALIDE_RUNTIME` is defined, for a subtle reason: some compilation environments require that all .h files can be compiled 'standalone' -- ie, they include all prerequisites and can be included in any order. As it turns out, HalideBuffer.h has a previously unnoticed glitch that is now being caught by this: - It includes both `<cstring>` and `HalideRuntime.h` - but when `COMPILING_HALIDE_RUNTIME` is defined, `HalideRuntime.h` includes `runtime_internal.h`, which assumes that no std headers are included - as it happens, `runtime_internal.h` defines `strstr()` and `strchr()` in a way that doesn't match the std headers (they both return non-const `char *`, perversely enough) and we get a compile error for mismatched function prototypes I think the neatest solution here is to just skip the entire contents of `HalideBuffer.h` in this situation, since none of its contents should ever be used inside the Halide runtime. (I could work around this situation on the google side by loosening the 'must be able to compile on its own' requirement in this case, but IMHO it's a good check and one that is worthy of keeping.) The fact that our function prototypes are a mismatch for the 'correct' ones could be debated; IMHO our 'wrong' definitions are safer that the std and should be kept. | 13 March 2023, 18:16:04 UTC |
78097a7 | Steven Johnson | 13 March 2023, 17:01:35 UTC | Dont use negative values for exit() (#7405) * Dont use negative values for exit() A program that terminates via `exit(-1)` leaves `$? = 255` (per the exit(3) manpage), as the value is chopped with `& 0377`. This makes it hard to use `git bisect` to track down bugs, as considers any exit code > 127 (or = 15, oddly enough) equivalent to an `abort()` and terminates the bisect. IMHO this is perverse behavior on the part of `git bisect`, but it is what it is, so let's revamp our tests to avoid calling `exit(-1);`. * trigger buildbots | 13 March 2023, 17:01:35 UTC |
822f5da | Steven Johnson | 13 March 2023, 17:00:50 UTC | Don't return negative values from main() (#7406) * Don't return negative values from main() A program that returns `-1` from `main()` leaves $? = 255 (per the exit(3) manpage), as the value is chopped with & 0377. This makes it hard to use git bisect to track down bugs, as considers any exit code > 127 (or = 15, oddly enough) equivalent to an abort() and terminates the bisect. IMHO this is perverse behavior on the part of git bisect, but it is what it is, so let's revamp our tests to avoid returning `-1` from main to indicate failure, and return `1` instead. * trigger buildbots * validate_gpu_object_lifetime() handling | 13 March 2023, 17:00:50 UTC |
f76bcc8 | Steven Johnson | 12 March 2023, 17:49:46 UTC | Cleanup of src/runtime/internal and test/runtime (#7399) * Cleanup of src/runtime/internal and test/runtime - Don't include .cpp files. - Don't use header-only "libraries" that rely on include order or being included only once (to wit: test/runtime/common.h -> common.cpp) - All files should explicitly #include what they need, even if they think it's already included (ie, order of include files should not matter) - In src/runtime/internal, change all `halide_abort_if_false` -> `halide_debug_assert` - in test/runtime, add HALIDE_CHECK to common.h and use it for tests instead of `halide_abort_if_false` * 32-bit fixes | 12 March 2023, 17:49:46 UTC |
0beb081 | Ttayu | 11 March 2023, 20:45:17 UTC | docs: Use halide.imageio instead of imageio (#7409) * docs: Use halide.imageio instead of imageio * trigger buildbots --------- Co-authored-by: Steven Johnson <srj@google.com> | 11 March 2023, 20:45:17 UTC |
4bf660e | Andrew Adams | 11 March 2023, 00:53:44 UTC | Fix overflow in x86 absd lowering (#7407) * Fix overflow in x86 absd lowering * Fix default lowering too | 11 March 2023, 00:53:44 UTC |
ead83e6 | Ttayu | 10 March 2023, 18:23:21 UTC | Docs/update readme python (#7402) * docs: Update simple usage in README_python.md * docs: Add imports | 10 March 2023, 18:23:21 UTC |
b02de4b | Steven Johnson | 09 March 2023, 17:46:43 UTC | Update all use of actions/checkout to v3 (#7400) | 09 March 2023, 17:46:43 UTC |
6a6c842 | Steven Johnson | 07 March 2023, 23:56:03 UTC | Backport changes to CodegenC from the xtensa branch (#7396) Backported with some trivial style changes (virtual method instead of flag, since this should never changes for a given instance) | 07 March 2023, 23:56:03 UTC |
a2df5bd | Steven Johnson | 07 March 2023, 19:29:08 UTC | Attempt to give proper labels to pip dev builds (v2) (#7395) A much-simpler alternative to #7394 | 07 March 2023, 19:29:08 UTC |
4a80251 | Andrew Adams | 03 March 2023, 20:57:47 UTC | Destringify CanonicalizeGPUVars (#7386) * Destringify CanonicalizeGPUVars This new implementation takes the high-water marks of each type of GPU loop, instead of filtering using the prefix of the loop name. * Better comments * fix typo | 03 March 2023, 20:57:47 UTC |
aa8fcad | Terry Heo | 03 March 2023, 20:56:52 UTC | hannk: Provide weak symbol functions to use op profiling (#7388) * hannk: Provide weak symbol functions to use op profiling You can add your own profiler with strong symbols. * hannk: Guard profiler feature with HANNK_PROFILER | 03 March 2023, 20:56:52 UTC |
387a19c | Xuanda Yang | 03 March 2023, 01:10:50 UTC | Make README consistent on supported LLVM versions (#7390) Make README consistent with https://github.com/halide/Halide/pull/7093 | 03 March 2023, 01:10:50 UTC |
122b5b6 | Terry Heo | 01 March 2023, 22:22:20 UTC | hannk: Add device_sync method hannk::Tensor class (#7387) hannk: Add devic_sync method hannk::Tensor class This method is useful to debug performance with synchronous execution. | 01 March 2023, 22:22:20 UTC |
303a90c | Steven Johnson | 01 March 2023, 17:21:48 UTC | Minor cleanup of GPUCompilationCache (#7376) * Minor cleanup of GPUCompilationCache While tracking down an apparently-unrelated threading bug in the webgpu backend, I made some tweaks to this code that I think are worth keeping. The main one of importance is that `release_hold()` and `release_context()` really should acquire the mutex -- they weren't before -- so now all public methods are properly mutexed. The other changes are mostly cosmetic: - Moved helper methods to be private rather than public - Changed the id value size to be `uintptr_t` rather than `uint32_t`; the space allocated for them is sizeof(void*). (Not sure this moves the needle but it felt right.) - Removed unused ctor for CachedCompilation Also, drive-by change in printer.h to capture some logging improvements. * Fix deadlock release_all() and release_context() were contending for the mutex | 01 March 2023, 17:21:48 UTC |
fba892a | Volodymyr Kysenko | 01 March 2023, 00:29:42 UTC | Specify a full type for llvm::IRBuilder (#7384) Specify full type for llvm::IRBuilder | 01 March 2023, 00:29:42 UTC |
5c02ae2 | Adrian Lebioda | 28 February 2023, 02:14:41 UTC | Bounds visitor for div was missing single_point mutated case (#7379) * Bounds visitor for div was missing single_point mutated case Signed-off-by: Adrian Lebioda <adrian.lebioda@hexagon.com> * Add test Signed-off-by: Adrian Lebioda <adrian.lebioda@hexagon.com> --------- Signed-off-by: Adrian Lebioda <adrian.lebioda@hexagon.com> Co-authored-by: Adrian Lebioda <adrian.lebioda@hexagon.com> | 28 February 2023, 02:14:41 UTC |
bdba694 | Steven Johnson | 28 February 2023, 02:14:24 UTC | Add Callable default ctor + `defined()` method (#7380) * Add Callable default ctor + `defined()` method This allows it to behave like * Add user_assert + test | 28 February 2023, 02:14:24 UTC |
c42a5b2 | Andrew Adams | 27 February 2023, 16:46:41 UTC | Remove a gross hack from gpu_only_aottest (#7378) * Remove a gross hack from gpu_only_aottest Also add metal support * Add missing include | 27 February 2023, 16:46:41 UTC |
09400f6 | Andrew Adams | 25 February 2023, 17:16:32 UTC | Bounds visitors for min/max were missing single_point mutated case (#7377) * Bounds visitors for min/max were missing single_point mutated case Partially fixes #7374 * Add test | 25 February 2023, 17:16:32 UTC |
b6a18b8 | Steven Johnson | 23 February 2023, 19:36:32 UTC | Update WABT to 1.0.32; Increase stack size for WASM AOT apps (#7373) | 23 February 2023, 19:36:32 UTC |
144c1a4 | Steven Johnson | 23 February 2023, 17:59:39 UTC | correctness_round should use Target::supports_type() (#7372) This gives it proper support for new GPU backends | 23 February 2023, 17:59:39 UTC |
b17806d | Steven Johnson | 23 February 2023, 17:59:13 UTC | Use HalideFreeHelper for the register_destructor (#7371) Slightly cleaner code. Also, drive-by change of NULL -> nullptr | 23 February 2023, 17:59:13 UTC |
629da52 | Steven Johnson | 23 February 2023, 17:58:37 UTC | Use single-char form of `unique_name` for semaphores (#7370) The multi-char form of `unique_name` will append a `$` to the identifier, e.g. `sema$4`. This isn't really legal for a C/C++ identifier. | 23 February 2023, 17:58:37 UTC |
386a2d1 | Steven Johnson | 22 February 2023, 23:39:25 UTC | Clean Up HalideFreeHelper code (main) (#7369) * Clean Up HalideFreeHelper code (main) - Revise HalideFreeHelper to be a templated struct, to save the unnecessary stack storage for the function - Add emit_halide_free_helper() method to consolidate usage * Update CodeGen_C.cpp | 22 February 2023, 23:39:25 UTC |
3246844 | Andrew Adams | 21 February 2023, 21:18:30 UTC | Use a std::unique_ptr for the IR Builder (#7356) * Use a std::unique_ptr for the IR Builder instead of a raw owning pointer * Use make_unique | 21 February 2023, 21:18:30 UTC |
e19a036 | Andrew Adams | 21 February 2023, 17:28:43 UTC | Overflow on casts is fine for ints < 32 bits (#7366) | 21 February 2023, 17:28:43 UTC |
ccc085a | Steven Johnson | 16 February 2023, 19:54:27 UTC | Update CMakePresets.json to use VS2022 instead of VS2019 (main) (#7359) | 16 February 2023, 19:54:27 UTC |
b65ea62 | Steven Johnson | 15 February 2023, 18:25:04 UTC | Remove unused code in VectorizeLoops (#7354) | 15 February 2023, 18:25:04 UTC |
18eb7d8 | Andrew Adams | 15 February 2023, 17:18:56 UTC | Permit vectorization of non-recursive atomic operations (#7346) * Vectorization of non-recursive atomic operations * Remove dead Vars | 15 February 2023, 17:18:56 UTC |
e5ed226 | Steven Johnson | 15 February 2023, 17:03:41 UTC | Fix Python error handling (#7352) * Fix Python error handling Error handling in the Python bindings wasn't quite right for JIT: We previously replaced halide_error() to throw a C++ exception. Sounds good, but unfortunately, doesn't work reliably: if called from jitted code (which doesn't know about C++ exceptions), the throw statement may be unable to find the enclosing try block (which is outside jitted code), meaning it will call std::terminate. Now, instead, we just leave the JIT error handler unset, and call with an explicit JITUserContext with a custom print handler; in theory, this meant that the code in JITFuncCallContext::finalize() would check for an error after the call into jitted code, and call `halide_runtime_error` if so (which would then trigger an all-in-C++-exception). Unfortunately... (2) JITFuncCallContext is broken by design; it mutates the input JITUserContext, so that trying to use the same JITUserContext for two calls in a row leaves you with a JITUserContext with (at least) the error_handler set. Since at least one of the realize() calls does this twice (once for bounds query, once for execution), this means that an error in the second call would never be seen, since finalize() only reported errors if there wasn't a custom error handler on input. Per @abadams suggestion, we work around this by treating 'JITErrorBuffer::handler' as 'no custom error handler', which is mostly true. (But really, JITFuncCallContext and JITUserContext are a hard-to-reason-about mess and arguably need to rethought entirely.) (3) Removed entirely-unnecessary overrides of runtime print and error handlers from PyStubImpl; despite the comments, this code is unnecessary. * format | 15 February 2023, 17:03:41 UTC |
7963cd4 | Steven Johnson | 14 February 2023, 17:14:05 UTC | Change early-bound default args in Python bindings to late-bound (#7347) In PyBind11, if you specify a default argument for a method, it is evaluated when the Python module is initialized, *not* when the method is called (as you might expect in C++). For defaults that are just constants/literals, this is no big deal, but when calling get_*_target_from_environment, this means it is called at module init time -- also normally not a big deal (since the values ~never change at runtime anyway), with one big exception (no pun intended): if the function throws an exception (e.g. via calling user_assert() or similar), that exception is thrown at Module-initialization time, which is a much more inscrutable crash, and one that is very hard to recover from. This may seem unlikely, but can happen pretty easily if you set (say) HL_JIT_TARGET=host-cuda (or other gpu) and the given GPU runtime isn't present on the given system; the current behavior is basically "make if impossible for the libHalidePython bindings to run", whereas what we want is "runtime exception thrown when you call the method". This changes the relevant methods to use `Target()` as the default value, and inside the method wrapper, if the value passed equals `Target()`, it replaces the value with the righ `get_*_target_from_environment()` call. (This turned up while doing some testing of https://github.com/halide/Halide/pull/6924 on a system without Vulkan available) | 14 February 2023, 17:14:05 UTC |
8bd07fb | Andrew Adams | 14 February 2023, 01:52:53 UTC | Fix tuple output bounds checks (#7345) Fix #7343 Tuple outputs weren't getting appropriate bounds checks due to overzealous culling of uninteresting code in the add_image_checks pass. | 14 February 2023, 01:52:53 UTC |
22aed20 | Steven Johnson | 11 February 2023, 02:42:05 UTC | Devirtualize the protected compile() methods in Codegen_C (#7341) With the addition of `preprocess_function_body()`, neither of these need to be virtual, and devirtualizing them avoid `hidden overloaded virtual function` warnings in subclasses that don't override them | 11 February 2023, 02:42:05 UTC |
6c5ca8e | Steven Johnson | 11 February 2023, 00:34:41 UTC | Tiny improvements in codegen in C backend (#7337) * Tiny improvements in codegen in C backend (1) Emit `true` or `false` instead of `(bool)(0ull)` etc for bool literals (2) Avoid redundant temporaries in print_cast_expr(), which occur in a small but nonzero number of cases Basically this means that code currently like ``` bool _523 = (bool)(0ull); bool _524 = (bool)(_523); ... foo(_524); ``` becomes ``` foo(false); ``` ...I'm sure this has no output on final object code, but it makes the generated C code less weird to read. * Also avoid extra intermediates for typed nullptr * Also use std::isnan() and std::isinf() * Update CodeGen_C.cpp | 11 February 2023, 00:34:41 UTC |
a6c5be7 | Steven Johnson | 10 February 2023, 21:28:51 UTC | Add a hook to Codegen_C::compile() (#7335) At least one subclass of Codegen_C currently has to replicate ~all of the compile(LoweredFunc) method, with the result that it has often gone stale (and still is stale) wrt changes in the base; this adds an optional method to allow some modifications to the function body just before it is printed, to avoid redundant code. | 10 February 2023, 21:28:51 UTC |
88d40c2 | Steve Suzuki | 10 February 2023, 18:07:56 UTC | Fix issue in find_package in cross-compilation for no OS (#7282) When using toolchain where Threads libs are not available, which is the case in baremetal target cross-compilation, we were not able to load even HalideHelpers pacakge. Co-authored-by: Alex Reinking <alex.reinking@gmail.com> | 10 February 2023, 18:07:56 UTC |
35322c3 | Steven Johnson | 10 February 2023, 00:22:10 UTC | Fix a subtle uninitialized-memory-read in Buffer::for_each_value() (#7330) * Fix a subtle uninitialized-memory-read in Buffer::for_each_value() When we flattened dimensions in for_each_value_prep(), we would copy from one past the end, meaning the last element contained uninitialized garbage. (This wasn't noticed as an out-of-bounds read because we overallocated in structure in for_each_value_impl()). This garbage stride was later used to advance ptrs in for_each_value_helper()... but only on the final iteration, so even if the ptr was wrong, it didn't matter, as the ptr was never used again. Under certain MSAN configurations, though, the read would be (correctly) flagged as uninitialized. This fixes the MSAN bug, and also (slightly) improves the efficiency by returning the post-flattened number of dimensions, potentially reducing the number of iterations f for_each_value_helper() needed. * Oopsie * Update HalideBuffer.h * Update HalideBuffer.h | 10 February 2023, 00:22:10 UTC |
ae3f401 | Steven Johnson | 09 February 2023, 23:08:37 UTC | Explicitly remove -D_GLIBCXX_ASSERTIONS from LLVM definitions (#7332) Explicitly remove -D_GLIBCXX_ASSERTIONS from LLVM definitions as a workaround for https://reviews.llvm.org/D142279 | 09 February 2023, 23:08:37 UTC |
4156c5a | Steven Johnson | 09 February 2023, 21:51:22 UTC | Allow _Float16 as alias for float16_t in halide_type_of<>() (#7325) (#7326) | 09 February 2023, 21:51:22 UTC |
734e34a | Steven Johnson | 09 February 2023, 17:59:12 UTC | Remove deprecated `HVX_shared_object` feature (#7331) This has been marked 'deprecated' for quite a while, and has no affect on codegen or, well, anything else. Let's remove it. | 09 February 2023, 17:59:12 UTC |
0f6003e | Terry Heo | 08 February 2023, 20:26:31 UTC | Float16: Remove unused header dependency (#7324) IRMutator.h is not needed for the Float16.h. | 08 February 2023, 20:26:31 UTC |
c3f3318 | Steven Johnson | 08 February 2023, 20:25:37 UTC | Fixes for top-of-tree LLVM (#7329) * Fixes for top-of-tree LLVM * fix * times ten * Update LLVM_Output.cpp | 08 February 2023, 20:25:37 UTC |
ddb515a | Steve Suzuki | 07 February 2023, 18:41:04 UTC | Improve support for Arm baremetal compilation and runtime (#7286) * Improve support for Arm baremetal compilation and runtime - Add Target feature "semihosting" mode for baremetal runtime - Fix error of aligned_alloc() when compiled by Arm GNU toolchain * Modify comments for Target feature semihosting * Add an example app to guide cross-compilation for baremetal target * Update build steps in HelloBaremetal * Fix line-ending * Set CMake variable BAREMETAL in toolchain file | 07 February 2023, 18:41:04 UTC |
34d256f | Steve Suzuki | 07 February 2023, 18:40:22 UTC | Make auto scheduler libs available in HalideHelpers package (#7285) * Make auto scheduler libs available in HalideHelpers package find_package(HalideHelpers) allows us to use add_halide_library(). But auto scheduler libs are not available unless they are in Halide-Interfaces.cmake. Note: Those libraries are not actually linked to the target application, but need to be available for add_custom_command call. | 07 February 2023, 18:40:22 UTC |
0c7722f | Terry Heo | 07 February 2023, 17:37:09 UTC | Add buffer sync methods hannk::Tensor class (#7323) Add few methods for GPU memory interaction. | 07 February 2023, 17:37:09 UTC |
0b7379f | Steve Suzuki | 07 February 2023, 17:08:32 UTC | Warn emulated float16 equivalent is generated (#7307) * Warn emulated float16 equivalent is generated | 07 February 2023, 17:08:32 UTC |
a55a09a | Dmitry Kurtaev | 07 February 2023, 14:17:58 UTC | Fix Halide cross-compilation (#7073) Use CMAKE_CROSSCOMPILING_EMULATOR for llvm-as and clang imported targets | 07 February 2023, 14:17:58 UTC |
1ad328a | Alex Reinking | 07 February 2023, 01:18:21 UTC | Fix LLVM 17+ build integration on 32-bit systems (#7322) * Fix LLVM 17+ build integration on 32-bit systems Fixes #7319 * add detail and precision to comment | 07 February 2023, 01:18:21 UTC |
91f3ac0 | Steve Suzuki | 06 February 2023, 22:23:47 UTC | Fix segfault by nonconstant bound in Adams2019 (#7321) Fix segmentation fault in Adams2019 in case the estimate or bound of Func is set to nonconstant Expr. | 06 February 2023, 22:23:47 UTC |
01f9e2d | Andrew Adams | 06 February 2023, 19:05:01 UTC | Replace some push_backs with emplace_back (#7317) | 06 February 2023, 19:05:01 UTC |
e9aecee | Steven Johnson | 03 February 2023, 21:17:47 UTC | Make visit_leaf() public in hannk/ops.h (#7318) * Make visit_leaf() public in hannk/ops.h This makes it easier for downstream code to experiment with adding ops * Update ops.h | 03 February 2023, 21:17:47 UTC |
0782d80 | Tom Westerhout | 01 February 2023, 18:24:01 UTC | Make Callable::call_argv_fast public (#7315) * Make Callable::call_argv_fast public * Add rough specification of the calling convention * Fix a typo | 01 February 2023, 18:24:01 UTC |
beba53a | Steven Johnson | 31 January 2023, 21:08:15 UTC | halide_popcount<uint64_t> is broken (#7313) Would not compile for Win32 or any other compiler without __builtin_popcountll available. (How did this get checked in without being tested on MSVC?) | 31 January 2023, 21:08:15 UTC |
fe76ab2 | Steven Johnson | 30 January 2023, 22:01:42 UTC | Minimal updates to allow Halide building with LLVM17 (#7309) * Minimal updates to allow Halide building with LLVM17 (Opening as draft initially until Buildbots build the new LLVM versions) * trigger buildbots | 30 January 2023, 22:01:42 UTC |
dd973f4 | Mikhail Usvyatsov | 25 January 2023, 21:40:54 UTC | Improved halide_popcount (#7225) * Improved halide_popcount * reused popcount64 from Utils.cpp in CodeGen_C * Fixed comment for popcount | 25 January 2023, 21:40:54 UTC |
810bd0b | Andrew Adams | 21 January 2023, 22:08:30 UTC | Hoist vector slices using rewrite rules (#7243) * Hoist slices using rewrite rules This lets us add associative variants more easily, which are helpful in the work on staging strided loads. * Don't hoist extract_element shuffles The Shuffle visitor wants to sink them * Add some static asserts * Add explanatory comment on shuffle hoisting * Fix comment * add lanes predicate to slice hoisting * add vector slice hoisting test cases Co-authored-by: Steven Johnson <srj@google.com> Co-authored-by: Alexander <ajroot@stanford.edu> | 21 January 2023, 22:08:30 UTC |
bafd60f | Alexander Root | 20 January 2023, 18:03:25 UTC | [x86 & wasm] Split up double saturating-narrows from i32 (#7280) * better x86 double sat-cast + add test * fix wasm too + test Co-authored-by: Steven Johnson <srj@google.com> | 20 January 2023, 18:03:25 UTC |
c601e4e | Steven Johnson | 20 January 2023, 17:43:56 UTC | Add workaround for the const-or-not user_context issue (#635) (#7291) Add a workaround for the const-or-not user_context issue (https://github.com/halide/Halide/issues/635) | 20 January 2023, 17:43:56 UTC |
2cc0468 | Steve Suzuki | 20 January 2023, 17:39:41 UTC | Fix issue in add_halide_runtime in cross-compilation (#7284) * Fix issue in add_halide_runtime in cross-compilation add_halide_runtime() tries to build generator executable, but it fails if we are working with cross-compiler toolchain. By using existing generator set as "FROM", we can work around this. | 20 January 2023, 17:39:41 UTC |
d44e99d | Steve Suzuki | 20 January 2023, 13:12:30 UTC | Fix error of add_halide_generator in cross-compilation (#7283) In case the project name is CamelCase, add_halide_generator() was not able to find the generator package, because CMake searches <name>Config.cmake or <lower-case-name>-config.cmake | 20 January 2023, 13:12:30 UTC |
147ff48 | Alex Reinking | 20 January 2023, 12:54:34 UTC | Remove dependency on platform threads library (#7297) * Refactor internal ThreadPool.h into halide_thread_pool.h tool * Drop dependency of libHalide on threads library * Remove other redundant uses of Threads::Threads * Update CMake documentation. | 20 January 2023, 12:54:34 UTC |
314b2fd | Alexander Root | 20 January 2023, 00:35:14 UTC | [HVX] Fix EliminateInterleaves (#7279) * fix EliminateInterleaves Co-authored-by: Steven Johnson <srj@google.com> | 20 January 2023, 00:35:14 UTC |
c9f3602 | Steven Johnson | 19 January 2023, 23:48:26 UTC | Remove the watchdog timer from generator_main(). It was intended to k… (#7295) Remove the watchdog timer from generator_main(). It was intended to kill pathologically slow builds, but in the environment it was added for (Google build servers), it ended up being redundant to existing mechanisms, and removing it allows us to remove a dependency on threading libraries in libHalide. | 19 January 2023, 23:48:26 UTC |
51a4f6c | Steven Johnson | 19 January 2023, 23:36:47 UTC | Emit prototypes for destructor functions in C Backend (#7296) We gathered up the destructors, but only emitted the prototypes if there was at least one non-C++ function declaration needed -- so if you built with cpp_name_mangling enabled, you might omit the right prototype. Fixed and added the right flag to a Generator test to tickle this behavior. | 19 January 2023, 23:36:47 UTC |
e8e1481 | Steven Johnson | 18 January 2023, 21:56:14 UTC | Drop support for MIPS (#7287) (#7289) * Drop support for MIPS (#7287) * Update Target.cpp | 18 January 2023, 21:56:14 UTC |
888c41c | Steven Johnson | 18 January 2023, 00:47:52 UTC | Add CMake support for C++ backend in test/generator (#7274) * Add support for C++ backend in test/generator When the CMake rules were rewritten a while back, the support for building/testing generators with the C++ backend (instead of the standard LLVM, etc) got lost. This adds it back in. Also made some drive-by fixes to the Makefile to enable some tests there that work correctly now. Also made a drive-by fix in in Codegen_C to fix allocation nodes that were just wrappers around buffer_get_host -- this prevented the cleanup_on_error test from building with the C++ backend. | 18 January 2023, 00:47:52 UTC |
0d43318 | Steven Johnson | 10 January 2023, 19:23:58 UTC | Optimize Module::compile() for some edge cases (#7269) * Optimize Module::compile() for some edge cases Avoid redundant `compile_to_buffer()` calls for output requests that can't possibly ever need them. * Avoid mutation | 10 January 2023, 19:23:58 UTC |
a8d88bb | Steven Johnson | 10 January 2023, 17:51:38 UTC | Use ::aligned_alloc() instead of std::aligned_alloc() in HalideBuffer.h (#7268) | 10 January 2023, 17:51:38 UTC |
eea7696 | Steven Johnson | 09 January 2023, 17:58:04 UTC | Update README_python.md (#7266) | 09 January 2023, 17:58:04 UTC |
c070bb8 | Alina Sbirlea | 06 January 2023, 00:00:34 UTC | Update change following LLVM WASM change f841ad30d77eeb4c51663e68efefdb734c7a3d07 (#7264) * Update change following LLVM WASM change https://github.com/llvm/llvm-project/commit/f841ad30d77eeb4c51663e68efefdb734c7a3d07 * Update checks conditional on LLVM version. | 06 January 2023, 00:00:34 UTC |
4b74049 | Andrew Adams | 05 January 2023, 21:12:50 UTC | Inline into extern function args during bounds inference (#7261) * Inline into extern function args during bounds inference Fixes #7260 * Run CSE once at the end * Actually recursively inline * clang-tidy * trigger buildbots * Make test invariant to the number of times the warning is printed as long as it's at least once Co-authored-by: Steven Johnson <srj@google.com> | 05 January 2023, 21:12:50 UTC |
04bb986 | Steven Johnson | 28 December 2022, 17:31:01 UTC | Conditional allocations shouldn't fail for size=0 in C++ backend (#7255) (#7256) * Conditional allocations shouldn't fail for size=0 in C++ backend (#7255) Allocations can be conditional; if the condition evaluates to false, we end up calling `halide_malloc(0)` (or `halide_tcm_malloc(0)` in the xtensa branch). Since it's legal via spec for `malloc(0)` to return nullptr, we need to be cautious here: if we are compiling with assertions enabled, *and* have a malloc() (etc) implementation that returns nullptr for alloc(0), we need to skip the assertion check, since we know the result won't be used. Note: a similar check will be inserted in the xtensa branch separately. Note 2: LLVM backend already has this check via Codegen_Posix.cpp * Update CodeGen_C.cpp | 28 December 2022, 17:31:01 UTC |
ade8b56 | Steven Johnson | 20 December 2022, 20:05:19 UTC | Remove deprecated halide_target_feature_disable_llvm_loop_opt (#7247) * Remove deprecated halide_target_feature_disable_llvm_loop_opt Was deprecated in Halide 15; let's remove in Halide 16 * trigger buildbots * trigger buildbots * Update CodeGen_LLVM.cpp | 20 December 2022, 20:05:19 UTC |
10345d4 | Andrew Adams | 16 December 2022, 17:56:08 UTC | Explicitly stage strided loads (#7230) * Add a pass to do explicit densification of strided loads * densify more types of strided load * Reorder downsample in local laplacian for slightly better performance * Move allocation padding into the IR. Still WIP. * Simplify concat_bits handling * Use evidence from parent scopes to densify * Disallow padding allocations with custom new expressions * Add test for parent scopes * Remove debugging prints. Avoid nested ramps. * Avoid parent scope loops * Update cmakefiles * Fix for large_buffers * Pad stack allocations too * Restore vld2/3/4 generation on non-Apple ARM chips * Appease clang-format and clang-tidy * Silence clang-tidy * Better comments * Comment improvements * Nuke code that reads out of bounds * Fix stage_strided_loads test * Change strategy for loads from external buffers Some backends don't like non-power-of-two vectors. Do two overlapping half-sized loads and shuffle instead of one funny-sized load. * Add explanatory comment to ARM backend * Fix cpp backend shuffling * Fix missing msan annotations * Magnify heap cost effect in stack_vs_heap performance test * Address review comments * clang-tidy * Fix for when same load node occurs in two different allocate nodes | 16 December 2022, 17:56:08 UTC |
382f813 | Steven Johnson | 16 December 2022, 17:54:21 UTC | Fix "may be used uninitialized" warnings in Codegen_C::print_scalarized_expr() (#7244) | 16 December 2022, 17:54:21 UTC |
da6746e | Steven Johnson | 14 December 2022, 05:29:45 UTC | correctness/exception.cpp needs to check HALIDE_WITH_EXCEPTIONS (fixes #7240) (#7241) correctness/exception.cpp needs to check HALIDE_WITH_EXCEPTIONS | 14 December 2022, 05:29:45 UTC |
1a4a469 | Andrew Adams | 13 December 2022, 16:11:54 UTC | Fix some sources of signed integer overflow in the compiler (#7231) * Fix some sources of signed integer overflow in the compiler Also, use compiler intrinsics when possible to handle overflow, as it generates faster code. * Fix msvc macro * Must use result * Actually perform the requested operation | 13 December 2022, 16:11:54 UTC |
533e6e5 | Steven Johnson | 12 December 2022, 20:54:46 UTC | Remove rogue string suffix in simd_op_check_arm.cpp (#7227) * Remove rogue string suffix in simd_op_check_arm.cpp Interestingly, it compiles here, but in some compilers it will fail with "unexpected token". * Update simd_op_check_arm.cpp | 12 December 2022, 20:54:46 UTC |
6ecdcbd | Steven Johnson | 11 December 2022, 18:05:55 UTC | Tighten alignment promises for halide_malloc() (#7222) This makes a couple of changes to the behavior/implementation of `halide_malloc()`: * Currently, halide_malloc must return a pointer aligned to the maximum meaningful alignment for the platform for the purpose of vector loads and stores. This PR also adds the requirement that the memory returned must be legal to access in an integral multple of alignment >= the requested size (in other words: you should be able to do vector load/stores "off the end" without causing any faults). * Currently, the `halide_malloc_alignment()` function is used to determine the default alignment; this cannot be overridden by user code (well, it can be, but the override will have no useful effect). It is intended to be "internal only" but is used in at least one place outside the runtime (apps/hannk). This change removes the call entirely, in favor of a call that is harder to access from outside the runtime and much less likely for end users to attempt to call. (It also changes apps/hannk to stop using it.) | 11 December 2022, 18:05:55 UTC |
16421a7 | Steven Johnson | 09 December 2022, 17:21:30 UTC | Revise simd_op_check tests to ignore HL_TARGET (#7207) (#7216) * Revise simd_op_check tests to ignore HL_TARGET (#7207) The simd_op_check tests have historically only run using the value of HL_TARGET, which mean that the coverage they had was low (since HL_TARGET is only set to values that are runnable on at least one buildbot). This change completely disconnects these tests from HL_TARGET; instead, each test now tests for a range of targets appropriate to the architecture being tested. On all platforms, they still compile to assembly and verify that the correct instructions are generated; additionally, if the host platform can JIT for the given target, it verifies that the results are as expected. * Update simd_op_check_riscv.cpp * Update simd_op_check_x86.cpp * Update simd_op_check_x86.cpp * Update simd_op_check_arm.cpp * Add more features that must match; re-enable the bfloat instructions * Update simd_op_check_x86.cpp * Update simd_op_check_riscv.cpp * trigger buildbots * Fix simd_op_check_wasm | 09 December 2022, 17:21:30 UTC |
ba31688 | Steven Johnson | 09 December 2022, 01:22:54 UTC | Increase __clang_major__ check in Float16.h to 16 (#7224) | 09 December 2022, 01:22:54 UTC |
066559b | Steven Johnson | 08 December 2022, 23:35:01 UTC | Remove check_jit_user_context() from V8 bindings (#7220) Obsolete code from early V8 work, it can trigger inappropriately in some corner-case scenarios. Remove it entirely to avoid false errors. | 08 December 2022, 23:35:01 UTC |
8fa8221 | Steven Johnson | 08 December 2022, 04:34:13 UTC | Fix bonehead version-checking test in HalideBuffer.h for Apple (#7218) | 08 December 2022, 04:34:13 UTC |
e8615bb | Steven Johnson | 08 December 2022, 01:22:38 UTC | clang-tidy: add [[maybe-unused]] to the DECLARE_NO_INITMOD stubs. (#7215) | 08 December 2022, 01:22:38 UTC |
a7fa32e | Steven Johnson | 07 December 2022, 17:31:01 UTC | Use aligned_alloc() as default allocator for HalideBuffer.h on most platforms (#7190) Use aligned_alloc() as default allocator for HalideBuffer.h on most platforms (See also https://github.com/halide/Halide/pull/7189) Modify H::R::Buffer to default to using `aligned_alloc()` instead of `malloc()`, except: - If user code passes a non-null `allocate_fn` or `deallocate_fn`, we always use those (and/or malloc/free) - If the code is compiling under MSVC, never use `aligned_alloc` (Windows doesn't support it) - If HALIDE_RUNTIME_BUFFER_USE_ALIGNED_ALLOC is defined to be 0, never use `aligned_alloc` (this is to allow for usage on e.g. older Android and OSX versions which don't provide `aligned_alloc()` in the stdlib, regardless of C++ versions.) Also, as with #7189, this ensures that the allocated space has the start of the host data as 128-aligned, and also now ensures that the size allocated 128-aligned (rounding up as needed). | 07 December 2022, 17:31:01 UTC |
8ce1212 | Steven Johnson | 07 December 2022, 17:29:19 UTC | Fix bitrot in PowerPC testing (#7211) * Fix bitrot in PowerPC testing (See #7208) - DataLayout was wrong (and has been for a long time) - simd_op_check_powerpc had errors. Some were easy to fix; the rest I commented out with a TODO since this backend doesn't appear to be in active use. (Want to fix this in preparation for fixing #7207) * Move x86 absd tests to the right place Co-authored-by: Andrew Adams <andrew.b.adams@gmail.com> | 07 December 2022, 17:29:19 UTC |
35020c5 | Zalman Stern | 07 December 2022, 07:15:46 UTC | Extend LLVM IR type mangling to handle scalars. (#7212) Extend LLVM IR type mangling to handle scalars and use this in vector predication intrinsic codegen. Fixes an error denerating vector predicated strided stores. | 07 December 2022, 07:15:46 UTC |
d4b4c50 | Zalman Stern | 07 December 2022, 07:15:10 UTC | Add RISC V zvl flag for LLVM version 16 or greater. (#7209) | 07 December 2022, 07:15:10 UTC |
e0d1e15 | Zalman Stern | 07 December 2022, 00:58:28 UTC | Fix issue with vector predicated comparison and select instructions. (#7205) Fix invalid LLVM IR issues with vector predicated comparison and select instructions. Add start of RISC V simd_op_check test. | 07 December 2022, 00:58:28 UTC |
59f5412 | Zalman Stern | 06 December 2022, 22:57:24 UTC | Add bridging for clang _Float16 type. (#7201) Add type bridging between Halide::float16_t and _Float16 if the compiler supports the latter. Testing is done using clang specific logic and may need to be extended for other compilers. I chose not to add support for __fp16 and __bf16 right now as __fp16 is less useful in being storage only and __bf16 also only supports a subset of operations and was running into undefined symbols during compilation that did not look promising. Co-authored-by: Steven Johnson <srj@google.com> | 06 December 2022, 22:57:24 UTC |
90459b0 | Steven Johnson | 06 December 2022, 00:53:05 UTC | Revert "Fix for top-of-tree LLVM" (#7200) Revert "Fix for top-of-tree LLVM (#7194)" This reverts commit a9ea9b565018774e52bb4028cbc91e14cb86959e. | 06 December 2022, 00:53:05 UTC |
345cf18 | Steven Johnson | 02 December 2022, 20:48:57 UTC | Don't attempt to use makecontext()/swapcontext() on Android (#7196) Despite being 'posixy', it doesn't actually implement these calls. | 02 December 2022, 20:48:57 UTC |
a9ea9b5 | Steven Johnson | 02 December 2022, 00:17:48 UTC | Fix for top-of-tree LLVM (#7194) | 02 December 2022, 00:17:48 UTC |
43911f4 | Steven Johnson | 01 December 2022, 18:20:03 UTC | Add a -v flag to generator_main() (#7193) This is a simple thing that just logs the path to all generated file(s) to stdout if `-v=1` is specified. It's intended for people running Generators directly from the commandline, and is intended as a more user-friendly alternative to HL_DEBUG_CODEGEN=1. No makefiles, etc specify it at present, but I anticipate using it in some tooling in the future. Example usage: ``` $ resize_image_bilinear.generator_binary -v 1 -o /tmp -g resize_image_bilinear -n resize_image_bilinear_uint16 -f resize_image_bilinear_uint16 -e assembly,c_header,llvm_assembly,registration,static_library,stmt 'target=arm-64-android' 'input.type=uint16' 'output.type=uint16' Generated file: /tmp/resize_image_bilinear_uint16.s Generated file: /tmp/resize_image_bilinear_uint16.h Generated file: /tmp/resize_image_bilinear_uint16.ll Generated file: /tmp/resize_image_bilinear_uint16.registration.cpp Generated file: /tmp/resize_image_bilinear_uint16.a Generated file: /tmp/resize_image_bilinear_uint16.stmt ``` | 01 December 2022, 18:20:03 UTC |
5a8c324 | Steven Johnson | 30 November 2022, 01:44:30 UTC | Fix metadata generation for multitarget Generators (#7181) Fix metadata generation for multitarget Generators We had a mechanism in place to ensure that Outputs that got renamed during lowering still emitted the proper names in the metadata... but this didn't work reliably for Multitarget generation. Now it does. | 30 November 2022, 01:44:30 UTC |
caf4b71 | Steven Johnson | 29 November 2022, 17:45:49 UTC | Disable unreachable-code clang-tidy warnings (#7182) Some configurations of clang-tidy will (correctly) complain that the code inside the `if` clauses here will never be executed, since it ends up as something like `if (strcmp("foo", "foo"))`... but for testing purposes, we want to keep it, for obvious reasons. It's hard to construct a string-compare here as constexpr, so I'm just going to NOLINT it. Also changed the `count_buffers()` check to a static_assert for simplicity. | 29 November 2022, 17:45:49 UTC |
2cfc315 | Steven Johnson | 28 November 2022, 22:11:35 UTC | Tweak the import paths in Python apps & tests (#7179) * Tweak the import paths in Python apps & tests This change makes it a bit easier for me to transform the import paths when merging into Google: we can't set PYTHONPATH, and calling `sys.path.append()` is frowned upon. This should have no effect on the GitHub repo but will make my life easier downstream. * More tweaks * force builds * Update Generator.cpp | 28 November 2022, 22:11:35 UTC |
73c61c3 | Steven Johnson | 28 November 2022, 21:52:51 UTC | Add optional "function_info" header output (#7170) Add optional "function_info" header output At first glance, this looks like a subset of what is already provided by the `_metadata()` functionality: describing the argument attributes of an AOT-generated Halide function. However, _metadata() is suboptimal for some use cases: Because it's expressed as ordinary data, we can only process it at runtime; the new fuctionality is expressed as a `constexpr` data structure, meaning we can process it at *compile* time if we so choose. (This is quite useful for producing automatic call wrappers, etc). At first I considered adding this to the normal `.h` file, but moving it into a new file is cleaner in a few ways: - It maintains the 'C-only' nature of the existing .h files (adding this would have imposed a C++17-only section on them) - Splitting into a new file means no existing users are affected by this change at all Note also that this is deliberately not replicating all of the existing `_metadata()` functionality (it's just the argument signature, but no e.g. estimates or default values, etc). This approach means that it is probably more sensible to add several separate constexpr "getters" to this file, rather than trying to mash everything together into one clumsy structure. (With _metadata(), there was an incentive to keep the surface area of the API small, even if that meant combining somewhat-unrelated concerns; there is no such incentive here.) | 28 November 2022, 21:52:51 UTC |
3ff9e66 | Dmitry Kurtaev | 28 November 2022, 20:00:58 UTC | Use n32:64 in RISC-V data layout (#7175) * Use n32:64 in RISC-V data layout * Remove unused LLVM header | 28 November 2022, 20:00:58 UTC |