Skip to main content

Browse the archive

https://github.com/halide/Halide
19 April 2024, 08:20:39 UTC
sort by:
Revision Author Date Message Commit Date
37a0d77 Andrew Adams 06 May 2021, 21:16:35 UTC Use a VectorReduce not to determine if any lanes are true in Hexagon backend 06 May 2021, 21:16:35 UTC
2ebaedd Alex Reinking 06 May 2021, 17:06:32 UTC Don't add Halide DLL to PATH on Windows. (#5973) This conflicts with vcpkg's own binary copying on Windows and makes cross compiling more difficult. It also runs into issues with excessively long commands when the user's PATH variable is very long. 06 May 2021, 17:06:32 UTC
ed989db Dillon Sharlet 06 May 2021, 16:52:19 UTC Remove some more old codegen workarounds and cleanups (#5932) * Remove old codegen workarounds * Pre-AVX2 codegen still needs this :( * trigger buildbots Co-authored-by: Steven Johnson <srj@google.com> 06 May 2021, 16:52:19 UTC
813180b Steven Johnson 06 May 2021, 16:38:22 UTC HalideBuffer should use D=max_rank instead of D=4 (#5971) This prevents mallocs for some degenerate cases where we need buffers with > 4 dimensions. 06 May 2021, 16:38:22 UTC
ab3670d Alex Reinking 16 March 2021, 11:37:38 UTC Clean up CMake helpers. 1. Add HEADER output to add_halide_library. 2. Use $<BUILD_INTERFACE:...> in generated target include paths. 3. Clean up logic (reduce nesting). 4. Use lower-case names for local variables. 5. Print paths to detected Clang and LLD config scripts. 6. Honor normal variable overrides for Halide_TARGET. 06 May 2021, 05:52:34 UTC
0b77168 Alex Reinking 08 April 2021, 18:20:37 UTC Consistently use Halide_* prefixes in CMake. 06 May 2021, 05:52:34 UTC
0c0b117 Steven Johnson 05 May 2021, 23:41:34 UTC Upgrade hannk's TFLite version to 2.5.0-rc3 (#5970) * Upgrade hannk's TFLite version to 2.5.0-rc3 * Drive-by cleanup of test names 05 May 2021, 23:41:34 UTC
438ddf8 Dillon Sharlet 05 May 2021, 23:39:44 UTC Optimize depthwise convolution (#5964) * Try factoring depthwise reduction. * Revert unnecessary change. Co-authored-by: Steven Johnson <srj@google.com> 05 May 2021, 23:39:44 UTC
93dad62 Alex Reinking 05 May 2021, 20:23:56 UTC Fix tutorial 15 test (#5966) 05 May 2021, 20:23:56 UTC
4a489e6 Steven Johnson 05 May 2021, 19:55:01 UTC Ensure our local flatbuffers.h is included in preference to system variants (#5962) * Ensure our local flatbuffers.h is included before system variants The local version might be too old for TFLite * Update CMakeLists.txt * Update CMakeLists.txt * Silence the noise noise noise noise NOISE * fix policy 05 May 2021, 19:55:01 UTC
95047ca Dillon Sharlet 05 May 2021, 16:55:59 UTC Optimize pooling ops (#5963) * Avoid padding for pool ops. * Use reciprocal to implement division. 05 May 2021, 16:55:59 UTC
115e597 Steven Johnson 05 May 2021, 01:28:21 UTC Fix FetchContent for hannk (#5960) Apparently using SOURCE_SUBDIR + FetchContent_MakeAvailable() doesn't work on the buildbots. This is equivalent and does work. ¯\_(ツ)_/¯ 05 May 2021, 01:28:21 UTC
675748a Steven Johnson 05 May 2021, 00:02:02 UTC Add apps/hannk to apps/CMakeLists.txt (#5958) ...but with an option for disabling it, for the buildbots 05 May 2021, 00:02:02 UTC
abefa3c Dillon Sharlet 04 May 2021, 23:45:21 UTC Remove nn_ops app in favor of hannk app. (#5893) 04 May 2021, 23:45:21 UTC
7b79dca Dillon Sharlet 04 May 2021, 23:44:53 UTC Add HANNK app (#5891) * More accurate approx_log2/exp2. * Add tests from inception_v4 * Improve precision of log2/exp2 related functions. * Add tanh and clean up generators. * Add version-checking to compare_vs_tflite and issue a warning if major and minor versions mismatch * Restore inadvertent @ removal * Add build_hannk/test_hannk targets to Makefile, to make specialized testing on select buildbots easier for now * More hacky padding for depthwise. * Add TODO * trigger buildbots * Add mean op, enable resnet50 to work. * Fix build failure on ARM. * Grammar * Remove stale TODO. * Make tensors shared_ptr. * Fuse double paddings. * Reduce padding for ARM. * Enable DimMap to express alignment. * Remove crops from execute. * Model -> OpGroup refactor. * Add DimMap::align. * Add proper alignment to DimMap. * Recursively transform. * Use cubic polynomials to approximate log2 and exp2. * Add --use_hannk option * Add mul op support. * Add TODO * Add disambiguating parens * Fix boneheaded broadcasting bug. * Less aggressive broadcasting. * Inline basic arithmetic. * Remove unnecessary using directives. * Fix stray unique_ptr<Tensor> * Implement space to depth and depth to space. * Enable scalar boolean comparison ops. * Support ReLUx as unary ops. * CHECK(false) -> LOG(FATAL) * Naming consistency. * More precise mul * Add some easy ops (NEG and SQUARE) * Fix asserts. * Don't segfault if interpreter can't be created * Add comment. * Remove dead file. * Fix excessive precision in softmax. * Lazy-init seeds in compare_vs_tflite, in case use_hannk=0 * Add TODO * Remove scalpel left in patient * Update model.h * Allow broadcasting of c of input 2 * Remove now-pointless specialization helper. * Put the common case specialization first. * Move pooling ops to the same generator file. * Fix softmax correctness issues * Don't benchmark when testing. * Rearrange input parameters. * Remove multiply_quantized helper. * kTfLiteError -> kTfLiteDelegateError * Remove unnecessary check for log2(0) * Fix details of ReshapeOp to match tflite's impl * Add Shape op. * Generically handle elementwise operations of any rank. * Some of these aren't elementwise. * Minor cleanups * Minor cleanup in ReshapeOp::execute() * Remove unused functions * Add Greater, GreaterEqual to delegate * clang-format * Update normalizations_generator.cpp * Avoid horrific clang-format suggestion. * clang-format * Fix common_halide test. * Fix typo. * Fix asserts. * clang-format * Save compare_vs_tflite outputs from first run (not post-benchmark) * Enable approx_exp2 for int16 results without overflow. * Clean up precision of transcendentals * Fix accidental widening of shift by a constant. * Move elementwise generators to the same file. * Report profiler after each test * Optimize fully connected a lot * Add elementwise program interpreter * Add elementwise program interpreter * clang-format * WIP LSTM * Fix Interpreter::inputs and outputs. * Fix some precision and scheduling issues of LSTM * Fix LSTM op * Fix build breakage. * Fix comments. * clang-format * Add wrapper for constructing elementwise programs. * clang-format * Use ElementwiseProgram to implement LstmElementwise * Compress programs and instructions by storing them in int16 and more CISC * Reduce verbose repetitive declarations. * Optimize constant zeros. * Use a named constant for the size of each instruction. * Remove unnecessary const instruction. * Optimize and clean up elementwise programs * Reduce overhead from H::R::B * More H::R::B overhead cleanup. * Add missing include. * Various fixes and improvements. * Add support for LSTM to the hannk delegate (#5943) * Add support for LSTM to the hannk delegate * clang-format * Add support for dynamic tensors to hannk (#5942) * Initial support for Dynamic Tensors in hannk * Update hannk_delegate.cpp * Fixes * Smarten Tensor::resize() * More H::R::B overhead cleanup * Refactor IsNodeSupported * Minor fixes * Fix member name style * Fix is_alias * Add is_no_op for some ops * Log reasons for node rejection if verbosity >= 1 * clang-format * Fix Concat handling for delegate * Properly parse split op * Add SPLIT_V * Fix regression in PadForOps * Add asserts * Revise ReshapeOp to just use a shape tensor * Update ops.cpp * clang-format * Scale tolerance with the data type. * Compress disassembly a bit * Fix bug with aliasing. * Add --use_tflite flag to compare_vs_tflite Allows disabling the reference run, for running just our delegate alone * Clean up hannk makefile * Regularize all of hannk's own include paths to be relative to apps/hannk; this simplifies things and will allow removing some hacks in Blaze/Bazel and also the upcoming CMake support * Remove some gratuitous uses of std::vector. * Remove unused function. * Remove more instances of std::vector * More SmallVector usage. * Avoid vector in can_use_elementwise_program. * Minor drive by fixes. * Remove some more H::R::B copies * Add broadcast support to elementwise programs * Also tweak the generated schema file include path * clang-format * clang-format * Stale TODO * Upgrade hannk to tf2.5 + more (#5948) * Upgrade hannk to tf2.5 + more - upgrade default TFLite to 2.5.0-rc2 - Revise build instructions & assumptions for TFlite (use CMake for it now instead of Bazel) - Revised Android build instructions (now assumes that tflite is built locally rather than pulled from a prebuilt) - Remove the need for flatc/flatbuffers - Minor fixes to the run-on-device scripts * Update Makefile * Update Makefile * Fix some harmless errors related to input slots * TFlite is too sloppy with dimensions. * Intervals are min, max, not min, extent. * Refactor compare_vs_tflite Lots of internal code motion to clean things up and prepare for adding an internal-delegate code path. Immediate change is just `--enable [h][t][x]` instead of the old "use" flags, and reducing the default max-num-of-diffs to 8 instead of 32. * Add an alias for OpPtr * Don't alias inputs that might be used elsewhere. * Use the right tensors when parsing. * Don't schedule sum_filter separately, and avoid 8-bit multiplies for x86 * Improve aliasing logic * Small optimizations. * Add no_bounds_query to elementwise pipelines. * Clean up std:;shared_ptr/H::R::B overhead. * More cvt refactoring, plus clang-format * Fix asserts. * Remove unused trace_ member * Remove unnecessary argument. * Fix x86 * Update compare_vs_tflite.cpp * Add internal-delegate option to CVT * Fix run_compare_on_device for recent changes * Revert possibly broken space to depth optimization. * Add gather and more binary op support. * Update compare_vs_tflite.cpp * Default external-delegate to disabled * clang-format * Better implementation of SpaceToDepth/DepthToSpace. * clang-format * Reduce buffer copy overhead. * Remove unnecessary types. * Consistent multiplication order. * Pad to at least FnRank * clang-format * De-inline two operator<<()'s * compare_vs_tflite error handling - If any of the comparisons fail, exit with a nonzero error code - add `--tolerance` flag to allow tweaking allowable tolerance on a per-pipeline basis * Add CMake build rules to apps/hannk (#5955) * Add CMake build rules to hannk * Update ops.cpp * Fix features * Update CMakeLists.txt * Add tests * Fix cmake issues * Use CMAKE_GENERATOR * Update configure_cmake.sh * Delegate Fixes * Update flag handling in compare_vs_tflite This is gratuitous but was bugging me: - Flags now understand both `--flag value` or `--flag=value` - Unknown flags now fail hard instead of being ignored Co-authored-by: Steven Johnson <srj@google.com> 04 May 2021, 23:44:53 UTC
f45d323 Andrew Adams 04 May 2021, 16:32:56 UTC Non-widening lowering of rounding shifts (#5956) This version lowers it without needing to widen, which is a large win on x86 for 16 and 32-bit types (3.8x faster and 2.8x faster respectively). It's a very slight slowdown for 8-bit because x86 doesn't have 8-bit shift instructions. Also drive-by typo fix. 04 May 2021, 16:32:56 UTC
94c0eca Dillon Sharlet 04 May 2021, 00:17:39 UTC Use dot products for sums. (#5954) 04 May 2021, 00:17:39 UTC
5a0d1e5 Volodymyr Kysenko 03 May 2021, 16:34:37 UTC Support VectorReduce in CodeGen_C (#5952) 03 May 2021, 16:34:37 UTC
8b9deea Dillon Sharlet 30 April 2021, 20:56:30 UTC Fix bugs when D != 4 (#5951) * Fix bugs when D != 4 * clang-format 30 April 2021, 20:56:30 UTC
093e8df Fangrui Song 29 April 2021, 22:58:43 UTC Replace llvm::sys::fs::F_None with llvm::sys::fs::OF_None (#5946) The former is deprecated. 29 April 2021, 22:58:43 UTC
fcbd2ee Dillon Sharlet 27 April 2021, 23:52:57 UTC Fix build issue in runtime. (#5944) 27 April 2021, 23:52:57 UTC
a391e9a AbdouTlili 27 April 2021, 23:13:06 UTC adding a note in the README.md to use -j option in make --build (#5938) * adding a note in the README.md to use -j option in make --build * wrapped the added section to 80 column 27 April 2021, 23:13:06 UTC
5a69e9f Dillon Sharlet 26 April 2021, 21:12:53 UTC Fix flattening of ramps involving 64-bit mins (#5940) * Fix flattening of ramps involving 64-bit mins. * Use make_const instead of cast. 26 April 2021, 21:12:53 UTC
91e42f4 Steven Johnson 26 April 2021, 20:10:21 UTC Don't use as_const_int() on temporaries (#5939) Sometimes we get lucky and it's still valid, but it's always wrong. 26 April 2021, 20:10:21 UTC
1b3cbcb aankit-ca 26 April 2021, 17:55:12 UTC [Hexagon] Try vdelta/vrdelta before vlut for some shuffles. (#5935) The patch tries to generate vdelta/vrdelta instructions for non-ramp shuffles. Eg: shuffle(lut_expr, < 0, 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14, 15, 16, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 33, 34, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 48, 49, 50, 51, 52, 54, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 66, 67, 68, 69, 70>) can be generated using vrdelta. The patch also fixes a bug where we bitcast vdelta/vrdelta with 16/32 bits elements to wrong type. User would see the below error: llvm-project/llvm/lib/IR/Instructions.cpp:2905: static llvm::CastInst *llvm::CastInst::Create(Instruction::CastOps, llvm::Value *, llvm::Type *, const llvm::Twine &, llvm::Instruction *): Assertion `castIsValid(op, S, Ty) && "Invalid cast!"' failed. Co-authored-by: Ankit Aggarwal <aankit@quicinc.com> 26 April 2021, 17:55:12 UTC
ba89623 Shivam Gupta 23 April 2021, 16:19:40 UTC Small Typo fix in lesson 06 (#5936) Signed-off-by: xgupta <shivam98.tkg@rediffmail.com> 23 April 2021, 16:19:40 UTC
a407acd Steven Johnson 22 April 2021, 16:29:01 UTC Revert "Temporarily disable hanging test (#5925)" (#5933) This reverts commit 62505857694ab8af2a88a22edf291e630c8c0cfd. 22 April 2021, 16:29:01 UTC
fb13fb0 Dillon Sharlet 21 April 2021, 22:10:32 UTC Add mul_shift_right intrinsic and related improvements (#5916) * Add multiply_quantized intrinsic * clang-format * Fix build on some compilers. * Fix incorrect saturating_pmulhrs * multiply_quantized -> mul_shift_right * Remove workaround and just cast shift amounts. * Fix error message * Fix declaration of mul_shift_right. 21 April 2021, 22:10:32 UTC
6867005 Shoaib Kamil 21 April 2021, 19:06:50 UTC Suppress Metal unused function warning (#5913) Co-authored-by: Steven Johnson <srj@google.com> 21 April 2021, 19:06:50 UTC
5dd85ae Andrew Adams 21 April 2021, 16:50:56 UTC Let the user pass the Func to use to the reduction helpers (#5929) * Let the user pass the Func to use to the reduction helpers * Pass Funcs by const ref 21 April 2021, 16:50:56 UTC
17d4771 Dillon Sharlet 21 April 2021, 16:04:27 UTC Update test to reflect behavior we expect. (#5928) 21 April 2021, 16:04:27 UTC
087567f Dillon Sharlet 21 April 2021, 16:04:09 UTC Remove old codegen. LLVM rewrites this back to a multiply anyways. (#5930) 21 April 2021, 16:04:09 UTC
6250585 Steven Johnson 20 April 2021, 21:23:26 UTC Temporarily disable hanging test (#5925) * Temporarily disable hanging test LLVM13 is causing vector_reductions to hang (https://reviews.llvm.org/D100099 appears to be the injection point). Disabling this test to unbreak the buildbots. * Update vector_reductions.cpp 20 April 2021, 21:23:26 UTC
c1de142 Alexander Root 20 April 2021, 21:21:33 UTC [adams2019] Add caching to autoscheduler (#5697) * add feature caching and block caching to adams2019 autoscheduler * added caching verification for feautures * add caching docstrings 20 April 2021, 21:21:33 UTC
ac23987 Dillon Sharlet 20 April 2021, 15:02:14 UTC Speed up simd_op_check by only compiling one pipeline per op (#5918) * Speed up simd_op_check and compute_with * Dense vector loads can be written many different ways. 20 April 2021, 15:02:14 UTC
6963673 Dillon Sharlet 20 April 2021, 00:24:06 UTC Add Target::ARMv81a and improve shift instruction selection (#5917) * Add Target::ARMv81a and improve shift instruction selection. * Remove merge mistake. * Don't use ARM intrinsic on arm32, it seems to be missing sometimes. 20 April 2021, 00:24:06 UTC
493dbd4 Steven Johnson 17 April 2021, 17:46:20 UTC Comment out specialiations for f64x2.convert_low_i32x4_s/u (#5914) LLVM removed the primitives we need (so our code can't be used), but it also doesn't seem to be generating the expected instructions directly (as claimed). Commenting out to un-break tests; issue has been reported to wasm/llvm team. 17 April 2021, 17:46:20 UTC
9cdb4aa Andrew Adams 16 April 2021, 22:23:30 UTC Simplify and improve cuda_mat_mul schedule (#5909) * Simplify and improve cuda_mat_mul schedule 16 April 2021, 22:23:30 UTC
a41cce7 Volodymyr Kysenko 16 April 2021, 20:47:16 UTC Basic support of predicated loads/stores in C++ backend (#5908) * Basic support of predicated load/stores in C++ backend * Fix formatting and maybe build * Fix * trigger buildbots Co-authored-by: Steven Johnson <srj@google.com> 16 April 2021, 20:47:16 UTC
3531167 Steven Johnson 15 April 2021, 18:34:37 UTC Drop LLVM10 support from master (#5740) * Drop LLVM10 support from master Update build files to require LLVM11+ in master branch. (Since we only regularly test master with 12 and 13 this is conservative.) Remove all code that is specialized for LLVM < 11.0. * Update CodeGen_ARM.cpp * Update CodeGen_LLVM.cpp 15 April 2021, 18:34:37 UTC
780ebd2 Zalman Stern 15 April 2021, 16:19:21 UTC Add an error for realize with a different number of outputs than defined for pipeline. (#5906) * Add an error for calling realize with a different number of outputs than the pipeline was compiled with. * Forgot to add test. * A readability scarifice to the clang deity. * Add CMake file. * Minor change to error text. * Fix logic to handle Funcs returning Tuples. * Formatting. 15 April 2021, 16:19:21 UTC
da02c0d Jiawen (Kevin) Chen 14 April 2021, 22:17:01 UTC Add missing "struct" before halide_type_t. (#5904) This allows it to compile as pure C instead of C++. Co-authored-by: Jiawen Chen <jiawen@adobe.com> 14 April 2021, 22:17:01 UTC
ccde965 Steven Johnson 14 April 2021, 21:32:29 UTC Enable some more wasm simd tests that are now working with top-of-tree LLVM. (#5903) 14 April 2021, 21:32:29 UTC
3ac277b Dillon Sharlet 14 April 2021, 17:23:02 UTC Rewrite double and triple narrowing on ARM (#5896) * Rewrite double and triple narrowing on ARM. * clang-format. Co-authored-by: Steven Johnson <srj@google.com> 14 April 2021, 17:23:02 UTC
ce9b324 Steven Johnson 14 April 2021, 16:18:45 UTC Fix UB in halide_buffer_t::size_in_bytes (#5898) Just a port of https://github.com/halide/Halide/pull/4389 to the equivalent methods in HalideRuntime.h, since offset-from-a-null-pointer is UB in C++. 14 April 2021, 16:18:45 UTC
1ff3e3f Mario Emmenlauer 12 April 2021, 20:50:02 UTC CMake build: Add more user control (#5859) * packaging/CMakeLists.txt: Allow users to override RPATH (i.e. for packaging Halide) * CMakeLists.txt: Allow users to override the C++ standard 12 April 2021, 20:50:02 UTC
9cc17b4 Alexander Root 12 April 2021, 17:23:57 UTC Add fuzzer to bounds_of_expr_in_scope + fix discovered overflow bugs (#5895) * add interval bounds fuzzer * correct overflow checks in bounds inference * catch uint32->int32 overflow in simplifier and revert bounds change 12 April 2021, 17:23:57 UTC
687c7d8 Andrew Adams 09 April 2021, 05:04:11 UTC Use guarded versions of vars if they exist in bounds inference (#5890) 09 April 2021, 05:04:11 UTC
cf40bc8 Steven Johnson 08 April 2021, 17:14:26 UTC Improve wasm_threads documentation (#5843) * Improve wasm_threads documentation * Update HalideRuntime.h 08 April 2021, 17:14:26 UTC
71b895e Alex Reinking 18 February 2021, 21:44:02 UTC Fix existing presets (remove -O2 stuff, typos) 07 April 2021, 22:51:53 UTC
bd16b37 Alex Reinking 18 February 2021, 21:42:47 UTC Add shebang line to autotune_loop.sh 07 April 2021, 22:51:53 UTC
efea7a2 Alex Reinking 18 February 2021, 21:42:22 UTC Remove WITH_APPS from README_cmake.md 07 April 2021, 22:51:53 UTC
69011bc Alex Reinking 17 February 2021, 06:56:20 UTC Fix spelling mistakes and Doxygen references 07 April 2021, 22:51:53 UTC
b9cd9f2 Alex Reinking 07 April 2021, 22:37:12 UTC Require LLD_DIR in zip/package.bat (#5887) 07 April 2021, 22:37:12 UTC
79fd0c9 Alex Reinking 07 April 2021, 19:12:23 UTC [cmake] Fix and reorganize warnings for building Halide (#5885) 07 April 2021, 19:12:23 UTC
7473402 Steven Johnson 07 April 2021, 02:23:25 UTC error_run_with_large_stack_throws should compile without exceptions (#5884) (1) Some downstream environments compile C++ without exceptions by default; this won't compile on those. (2) We should check that assert-fail also errors out as expected. 07 April 2021, 02:23:25 UTC
ea76214 Alex Reinking 07 April 2021, 00:53:11 UTC Improve ClangCL support by disabling, fixing warnings (#5876) Co-authored-by: Mario Emmenlauer <memmenlauer@biodataanalysis.de> 07 April 2021, 00:53:11 UTC
85816e4 Andrew Adams 06 April 2021, 20:56:47 UTC Add explicit cast to remove ambiguous operator== (Fixes #5329) (#5879) 06 April 2021, 20:56:47 UTC
e877e5b Steven Johnson 06 April 2021, 16:27:09 UTC Fix natural_vector_size for wasm 64-bit types (#5880) In the original spec, wasm-simd128 didn't have int64 or float64; the final spec adds these types, so this bit of code is outdated and incorrect. 06 April 2021, 16:27:09 UTC
7825d48 Steven Johnson 06 April 2021, 16:25:21 UTC Enable i64x2 comparisons in simd_op_check (#5881) * Enable i64x2 comparisons in simd_op_check * More drive-by fixes 06 April 2021, 16:25:21 UTC
9944dda Ming Yan 06 April 2021, 16:21:21 UTC Fix typos in tutorial lesson_08 (#5875) 06 April 2021, 16:21:21 UTC
525e246 Volodymyr Kysenko 03 April 2021, 01:29:55 UTC Try to vectorize inner statement of else branch of likely (#5874) * Try to vectorize inner statement of scalarize * Extend test to check for other scalarized loop * Add more details to the comment * make format * Remove note 03 April 2021, 01:29:55 UTC
59a04e4 Alex Reinking 02 April 2021, 17:53:49 UTC Use fibers to guarantee stack size on Windows (#5873) * Use fibers for lowering. * Move fibers to Util * Wrap compile_func call in call_with_stack_requirement * Rename call_with_stack_requirement -> run_with_large_stack * Appease clang_format * Add exception handling to run_with_large_stack * clang-format * Fix 32-bit? * Fix error wording for Makefile * Improve naming in run_with_large_stack 02 April 2021, 17:53:49 UTC
42092e3 Zalman Stern 01 April 2021, 19:13:01 UTC Fix an issue in Halide's float16 compilation support. Add tests. (#5872) In EmulateFloat16Math.cpp, conversion from 32-bit float to 16-bit float could produce a NaN value when and infinity is correct. This is because for numbers larger than the exact infinity value, the mantissa could be non zero. Add tests to cover this case, and float16 infinities in general. Couple small style/comment cleanups. 01 April 2021, 19:13:01 UTC
cb78a6b Steven Johnson 31 March 2021, 22:30:55 UTC Don't strip strict_float() from lets (#5871) * Don't strip strict_float() from lets Bug injected in #5856: the change in Simplify_Let.cpp was inadvertently stripping `strict_float()` calls that wrapped the RHS of a Let-expr, which can change results nontrivially in some cases. I don't think a new test for this fix is practical -- it would be a little fragile, as it would rely on the specifics of simplification that could change over time. As a drive-by, also added an explicit rule to Simplify_Call to ensure that strict_float(strict_float(x)) -> strict_float(x) in *all* cases. (The existing rule didn't do this in all cases.) 31 March 2021, 22:30:55 UTC
896b260 Dillon Sharlet 31 March 2021, 15:08:33 UTC Add some not rules. (#5870) 31 March 2021, 15:08:33 UTC
3e59294 Dillon Sharlet 30 March 2021, 23:26:22 UTC Add TailStrategy::Predicate (#5856) * Add TailStrategy::Predicate * Add some tests for TailStrategy::Predicate. * Fix missing override. * Fix comment. * Tweak target behavior. * Remove all heuristics * clang-format. * clang-tidy. * TailStrategy::GuardWithIf isn't always faster than scalar code :( * Use TailStrategy::Predicate in the predicated store/load test. * What is this test * Fix test bug. * Revert x86 behavior. * Move predicate to Internal namespace. * Recursively strip tags. * trigger buildbots * strip_tags -> unwrap_tags * Fix comment. Co-authored-by: Steven Johnson <srj@google.com> 30 March 2021, 23:26:22 UTC
7bbe2fd Steven Johnson 30 March 2021, 23:00:23 UTC Add wasm support for int32->f64 and f32->f64 simd ops (#5863) * Add wasm support for int32->f64 and f32->f64 simd ops At top-of-tree LLVM, the wasm backend never seems to emit the vector version of these ops; pattern-match to target them specifically. 30 March 2021, 23:00:23 UTC
e7eec5c Steven Johnson 30 March 2021, 22:45:32 UTC Add support for wasm dot-product instruction (#5861) * Add support for wasm dot-product instruction 30 March 2021, 22:45:32 UTC
f2143bf Steven Johnson 30 March 2021, 19:29:32 UTC Add a way to set a GeneratorInput's type in code (#5868) * Add a way to set a GeneratorInput's type in code Currently, if you want to vary the type of a Generator's inputs or outputs, you have to specify the types in the makefile. This can be awkward for things with complex logic. This PR proposes adding a way to do this: a new `set_type()` method which can only be called from the rarely-used Generator::configure() method. It only allows setting the type for an input or output that has no type specified. I'm not 100% sure if this is a good idea, but for certain rare corner cases, it may be quite handy. (Note that extending this to allow specifying dimensions and/or array size in the same way might be handy, but is omitted from this PR.) * Update Generator.h * Also add set_dimensions, set_array_size 30 March 2021, 19:29:32 UTC
2dd7a6b Thales Sabino 30 March 2021, 19:00:21 UTC Add support for AVX-512 VNNI saturating dot products (#5807) * Add support for AVX-512 VNNI saturating dot products This commit adds support to Intel VNNI saturating dot product instructions vpdpbuds and vpdpwssd This was accomplished by adding a new VectorReduce operation to perform the saturating_add and exposing a new inline reduction saturaring_sum. Users can then write RDom r(0, 4); f(x) = saturating_sum(i32(0), i16(i8(g(x + r)) * u8(h(x + r)))) bool override_associativity_test = true; int vector_width = 4; Var xo, xi; f.update() .split(x, xo, xi, vector_width) .atomic(override_associativity_test) .vectorize(r) .vectorize(xi); To lower the expression into a call to vpdpbuds. Note that override_associativity_test is set to true or halide will fail to prove the associativity of the saturating_add operation Add support for VectorReduce::SaturatingAdd in CodeGen_LLVM Code is correctly generated when no intrinsic is available to perform a saturating dot product. Add vpdpbusds,vpdpwssd tests to simd_op_check Test if the saturating dot product instructions are being generated for AVX512_SapphireRapids targets * Improve code according to report from clang-tidy * Make init_val a const ref since it only used that way inside saturating_sum * clang-format * Revert removal of clang-format tag in CodeGen_X86.cpp * Add SaturatingAdd case Monotonic VectorReduce visit * Bail out in Bounds when dealing with a SaturatingAdd VectorReduce * Move saturating_mul to Simplify_Internal.h so it can be used in Simplify_Exprs.cpp * Remove init_val from the saturating_sum inline reduction * Unconditionally override the associativity test in the simd_op_check tests * Remove annonymous namespace from saturating_mul utility Co-authored-by: Thales Sabino <thales@codeplay.com> 30 March 2021, 19:00:21 UTC
5b238e7 Steven Johnson 30 March 2021, 17:18:45 UTC Use a varying seed for random test data in simd_op_check (#5864) * Use a varying seed for random test data in simd_op_check We currently use `123` as a hardcoded seed, so we may sometimes be getting lucky with test patterns that happen to match scalar and vector. Let's vary the seed in the same way we do for (eg) fuzz_simplify to slightly broaden test coverage. 30 March 2021, 17:18:45 UTC
602cbac Shivam Gupta 30 March 2021, 16:56:13 UTC [NFC] LLVM trunk is now called main (#5866) Reference - https://foundation.llvm.org/docs/branch-rename/ 30 March 2021, 16:56:13 UTC
b7bc8e2 Steven Johnson 30 March 2021, 16:48:14 UTC Add support for wasm-simd saturating-narrow ops. (#5854) * Add support for wasm-simd saturating-narrow ops. 30 March 2021, 16:48:14 UTC
e0461e9 Steven Johnson 30 March 2021, 03:01:20 UTC Add support for i16x8.q15mulr_sat_s in wasm (#5853) * Add support for i16x8.q15mulr_sat_s in wasm Also, some drive-by clarifications to other wasm-simd instructions in simd_op_check -- some of the yet-to-be-implemented ones are of dubious use in Halide and may not be worth implementing. 30 March 2021, 03:01:20 UTC
07f880e Steven Johnson 29 March 2021, 23:37:23 UTC Add support for pairwise_widening_add in wasm (#5850) * Add support for widening_mul in wasm 29 March 2021, 23:37:23 UTC
e8085bd Dillon Sharlet 29 March 2021, 17:09:15 UTC Add simplifier rules helpful for specialization (#5836) * Add simplifier rules helpful for specialization. * clang-format * Revert sketchy select simplification. * Add different rules. 29 March 2021, 17:09:15 UTC
4f152f3 Steven Johnson 27 March 2021, 23:16:45 UTC Add align_extent(), to align extent but not min (#5829) * Allow align_bounds() to align extent but not min This can be handy when you have an intermediate Func that is being tiled inside an outer Func and you want to ensure that it fits an exact multiple of tiles. * Add separate align_extent() method 27 March 2021, 23:16:45 UTC
bc42da9 Steven Johnson 25 March 2021, 16:40:28 UTC Add support for widening_mul in wasm (#5849) * Add support for widening_mul in wasm 25 March 2021, 16:40:28 UTC
9a8ddf7 Andrew Adams 24 March 2021, 18:52:17 UTC Remove buggy deinterleave misfeature (#5844) 24 March 2021, 18:52:17 UTC
92dfc82 Dillon Sharlet 24 March 2021, 17:56:58 UTC Enable sliding window in registers (#5815) * Sliding in registers * Fix some failure cases. * Handle if_then_else in loop partitioning. * Add rebase_loops_to_zero pass. * Use select instead of if_then_else. * Add select comparison simplifications. * Don't rewrite lets * Rebase producer loops of register slides to 0, and don't overwrite realization bounds. * Add rules for ramp < broadcast * Put the likely on the old value instead of the new value. * New rules for comparing ramps and broadcasts * Switch back to if_then_else * Update comments. * Don't try to fold dimensions with a constant min or max. * More comments. * Make the vectorized register sliding window test tighter. * Remove debug helper. * Fix tests broken by loop rebasing. * Move rebasing after loop partitioning * clang-format * clang-tidy * Also put MemoryType::Register on the stack. * Expand arg before substitute. Co-authored-by: Andrew Adams <andrew.b.adams@gmail.com> Co-authored-by: Steven Johnson <srj@google.com> 24 March 2021, 17:56:58 UTC
28dd74f Steven Johnson 24 March 2021, 16:30:16 UTC Upgrade WABT to 1.0.22 and do some minimal update to codegen and simd_op_check for LLVM13. (#5841) * Upgrade WABT to 1.0.22 and do some minimal update to codegen and simd_op_check for LLVM13. * tickle * trigger buildbots 24 March 2021, 16:30:16 UTC
602d15a Dillon Sharlet 22 March 2021, 21:39:49 UTC Fix missed pattern matching due to cast optimization. (#5834) 22 March 2021, 21:39:49 UTC
8aa778c Dillon Sharlet 22 March 2021, 19:19:44 UTC Enable targeting broadcasting dot products on ARM (#5833) * Enable targeting broadcasting dot products on ARM. * Add to comment. 22 March 2021, 19:19:44 UTC
80307ec Andrew Adams 21 March 2021, 23:07:23 UTC Update README.md (#5831) 21 March 2021, 23:07:23 UTC
e92f05d Steven Johnson 18 March 2021, 16:58:27 UTC Upgrade pybind11 to 2.6.2 (#5821) * Upgrade pybind11 to 2.6.2 Released January 2021 with various bug fixes, see https://github.com/pybind/pybind11/releases/tag/v2.6.2 * Update requirements.txt 18 March 2021, 16:58:27 UTC
8d5c7cd Andrew Adams 18 March 2021, 02:06:03 UTC In fast_sine_cosine, fix message, relax threshold (#5820) 18 March 2021, 02:06:03 UTC
1a61c4b Volodymyr Kysenko 17 March 2021, 23:42:41 UTC Use print_type function when printing vector type definitions, which can (#5819) be overridden in derived classes if needed. 17 March 2021, 23:42:41 UTC
b549d18 Steven Johnson 16 March 2021, 21:05:33 UTC Grab-bag of minor fixes (#5813) * All TraceViz scripts should set pipefail (at least) * Use ".ldscript" for linker version scripts * More TraceViz fixes 16 March 2021, 21:05:33 UTC
e1c0fd3 Alex Reinking 16 March 2021, 20:19:05 UTC Conform to CMP0116 when using CMake 3.20+ (#5810) 16 March 2021, 20:19:05 UTC
5aa1a65 Steven Johnson 16 March 2021, 03:16:33 UTC Add Halide_CCACHE_BUILD option for CMake (#5804) * Add Halide_CCACHE_BUILD option for CMake This replicates the approach LLVM already has of baking a use-cache option directly into the CMake options, without having to futz with changing CC/CXX/etc. * Update CMakeLists.txt * Add CCACHE_SLOPPINESS=pch_defines,time_macros * Update CMakeLists.txt 16 March 2021, 03:16:33 UTC
b061a32 Dillon Sharlet 15 March 2021, 21:58:31 UTC Avoid printing the statement when it is unchanged (#5809) * Only print statement if it changed. * Move simplification after unroll from lower to after unrolling a loop. * Avoid making new IR when not necessary. * Use a small helper class instead. * clang-format. Co-authored-by: Steven Johnson <srj@google.com> 15 March 2021, 21:58:31 UTC
c49b4af Mike Woodworth 15 March 2021, 20:12:57 UTC support building for m1 with only AARCH target (#5802) move test for ARM + metal to AARCH64 Co-authored-by: Steven Johnson <srj@google.com> 15 March 2021, 20:12:57 UTC
28c742f Steven Johnson 15 March 2021, 19:21:36 UTC Set CMAKE_CROSSCOMPILING_EMULATOR in toolchain.linux-arm32.cmake (#5808) Also add comments 15 March 2021, 19:21:36 UTC
d523b83 Steven Johnson 13 March 2021, 00:54:23 UTC Fix deprecation warnings for trunk LLVM (#5803) (1) Both the 'deprecated' and new, non-deprecated variants existed back to at least LLVM10, and the deprecated variant was commented as deprecated at that point as well; the change in LLVM13 is that they are now annotated with LLVM_ATTRIBUTE_DEPRECATED so we get compiler warnings (and thus errors). (2) The fixes are simply replicating what the old, deprecated methods did internally. 13 March 2021, 00:54:23 UTC
c3882a5 Dillon Sharlet 11 March 2021, 23:14:36 UTC Change warmup strategy for sliding window (#5755) * Implement sliding window warmups by backing up the loop min. * Fix indirect sliding windows. * Improve is_monotonic. * Small cleanups. * Avoid generating vector valued bounds. * Fix build error on some compilers. * Fix loop bounds. * Don't try to slide things that should just be compute_at the store_at location. * Print condition when printing boxes. * Less things broken. * Add/fix comments. * Comments * Fix async by moving if inside consume (and so inside acquires). * Fix division. * This doesn't work on master either. * Add TODO * Acquire is not a no-op. * Add comment about unfortunate simplification. * Remove debug(0) * Add simplification of for { acquire { noop } } * Fix folding factors finally! * Update storage_folding test. * Fix bug when cloning a semaphore used more than once. * Disable failing test. * Work around bad complexity in is_monotonic. * Fix sub bug * Significantly faster schedule for blur. * Update tracing test. * New simplifications that help with upsampled and downsampled sliding windows. * This doesn't need explicit folding any more. * Fix new simplifier rules. * Fix simplifier div rule * Remove ancient brittle test. * Fix simplify rule again * More LT -> EQ rules for mod * Fix nested sliding windows with upsamples. * Replace hack with better solution. * Add missing override * Don't rewrite loop variable if the min doesn't change. * Refactor sliding window lowering. * Fixed bounds growing redundantly for independent producers. * Don't take the union unless possibly needed. * Respect conditional provide/required. * Add missing overrides * Much better schedule. * Use a smaller image for blur benchmarking so that different schedules have different perf * Replace Interval with ConstantInterval for is_monotonic. * Don't try to handle unsigned deltas. * Add failing test. * Remove unused new code. * Remove weird debugging code. * Avoid expanding bounds of split producers * Remove stray likely_if_innermost. * Remove old autotune tests. * Update test for guarded producers. * Reenable test. * Update trace for guarding producers. * Don't overwrite required.used * Handle LE/LT in bounds of lanes in vectorize * Fix acquire and release of warmups * Earlier fix for multiply cloned acquires was wrong. * Handle nested vectorization. * clang-format * Remove autotune_bug_* tests * Fix shadowing error on some compilers. * Appease overzealous clang-tidy warning. * clang-format * Don't use silly hack. * clang-tidy... * It's no longer safe to assume monotonic means bounds_of_expr_in_scope is exact * Address review comments * Add comment * Add missing override. * Fix constant interval issues. * Revert and remove empty interval * Fix multiply!? * Reduce need for simplifications. * Simplifications from dsharletg/sliding-window branch * Don't learn likely(x) and x. * Add comment * Add some min/max rules. * Also substitute facts from asserts * Remove is_empty from header too. * More rules * Add double stairstep rule. * Disable rule that uncovers bugs. * Consider anded expressions as if they were independent nested ifs. * Add promise_clamped to producer guards. * Revert "Consider anded expressions as if they were independent nested ifs." This reverts commit 03efb3f784b3078b64961c98edde383f4de04fb4. * Don't combine ifs, split them instead. * Update trace * clang-tidy/clang-format * Remove splitting of ifs, it breaks brittle tests. * Safer check on old conditions. * Fix producer guard condition. * Interval fixes. * Handle sliding backwards * Handle transitive dependencies. * Backport abadams' fix from abadams/slide_over_split_loop * Fix select visitor. * More simplifier rules. * Bring back old logic as a fallback. * Avoid specializations corrupting sliding * Fix boneheaded rule errors. * Fix slightly conservative bounds at the max for split case. * This pattern is too sensitive to the simplifier. In a real use case, it's just a sum, and the result can be subtracted after doing a reduction. * Add missing clamp rule * Don't count unlikely loops as inner loops for likely_if_innermost * Use <= instead of == to solve for the new loop min Useful when the warmup is a partial vector or something * Verify simplifier changes and add variants as suggested by synthesizer * Make implicit assumption explicit, for clarity * Use find_constant_bounds * Guard against expanded bounds more effectively. * Update tracing test * Small cleanup. * Don't simplify/prove using lets that might change value. * Stronger solving without expanding lets. * New simplifier rule for alignment * Fix case where no warmup needed * Add some useful rules. * Add safety check on when we can use the new loop min. * Better proof to avoid hacky condition that is hard to prove. * Small cleanup and use the nice new folding factors. * Bring back unrolled producer test. * clang-format * Expand comment. * Fix sliding backwards condition. * min(new_loop_min, loop_min) isn't needed any more. * We need that min, but we can be more conservative about it. * Stronger handling of previous loop mins. * Remove unused is_monotonic_strong. * Remove ConstantInterval::make_intersection. * Avoid need to handle uint specially. * Add cache for depends_on. * Reduce unnecessarily large cache scope * The first part of the key is always the same Co-authored-by: Andrew Adams <andrew.b.adams@gmail.com> 11 March 2021, 23:14:36 UTC
c2a0db1 Steven Johnson 10 March 2021, 23:39:27 UTC Fix correctness_memoize on arm32 (#5799) Conversion from a pointer to an integer is "implementation defined"; in general, conversion to `uintptr_t` is reliable, but other conversions aren't. It turns out that for our arm32 compiler, casting a pointer to a `uint64_t` sign-extends, rather than zero-extends, causing unexpected behavior in memoize cache eviction *if* the pointer is in the top half of memory. The fix here is simple (cast to `uintptr_t` directly), but it brings up the question of where else in our codebase we might be doing direct conversions elsewhere in our code without noticing the potential UB. 10 March 2021, 23:39:27 UTC
7a8c771 Steven Johnson 10 March 2021, 22:53:40 UTC Fix wasm-interp ucon error (#5797) This is a subtle error that only shows up in builds with assertions enabled in wabt; the issue is that we ask for an int64 value from a wabt `Value` struct, but that struct only has an int32, so an assertion can fail. (Note that there is always storage for both, and the values are constrained and unused inside the JIT anyway, so this is mostly just a cosmetic fix.) Also added a .gitignore entry. 10 March 2021, 22:53:40 UTC
34c402f Dillon Sharlet 10 March 2021, 21:54:51 UTC Fix bug found by asan from #5784 (#5798) 10 March 2021, 21:54:51 UTC
c67f486 John Lawson 10 March 2021, 18:51:33 UTC Move where intrinsic function attributes are set (#5795) * Make declare_intrin_overload return LLVM function * Make names same as elsewhere * Remove unneeded enum name * Set moved attributes in Hexagon backend * Use declare_intrin_overload for ARM vabdl * Fix ARM vabdl intrinsic types * Format and clang-tidy * Rename intrinsic to widening_absd 10 March 2021, 18:51:33 UTC
back to top