04ec9c0 | Steven Johnson | 14 November 2019, 20:17:17 UTC | Merge branch 'master' into pdb_mul_div_mod_multi_thread | 14 November 2019, 20:17:17 UTC |
08bd367 | Steven Johnson | 14 November 2019, 17:49:04 UTC | Merge pull request #4405 from halide/srj-llvmfix Update for trunk LLVM API changes | 14 November 2019, 17:49:04 UTC |
64c21d8 | Steven Johnson | 14 November 2019, 02:30:15 UTC | Merge branch 'master' into srj-llvmfix | 14 November 2019, 02:30:15 UTC |
58277ba | Steven Johnson | 14 November 2019, 02:29:50 UTC | Merge pull request #4400 from halide/srj-llvm-git LLVM fixes (draft) | 14 November 2019, 02:29:50 UTC |
da732f4 | Steven Johnson | 14 November 2019, 01:39:43 UTC | Update for trunk LLVM API changes | 14 November 2019, 01:39:43 UTC |
7f73523 | Steven Johnson | 13 November 2019, 22:54:48 UTC | Merge branch 'master' into pdb_mul_div_mod_multi_thread | 13 November 2019, 22:54:48 UTC |
4e4bb48 | Steven Johnson | 13 November 2019, 22:44:03 UTC | Merge pull request #4395 from halide/srj-llvm Update README for LLVM build instructions (Issue #4394) | 13 November 2019, 22:44:03 UTC |
b3631fd | Steven Johnson | 13 November 2019, 22:43:30 UTC | Merge branch 'master' into srj-llvm-git | 13 November 2019, 22:43:30 UTC |
afd9f4d | Steven Johnson | 13 November 2019, 22:18:52 UTC | Merge pull request #4391 from halide/srj-onnx-cast avoid reinterpret_cast in onnx_converter | 13 November 2019, 22:18:52 UTC |
ee4fbb9 | Steven Johnson | 13 November 2019, 18:22:08 UTC | Merge branch 'master' into pdb_mul_div_mod_multi_thread | 13 November 2019, 18:22:08 UTC |
59b616f | Steven Johnson | 13 November 2019, 18:18:18 UTC | Update README.md | 13 November 2019, 18:18:18 UTC |
72994e0 | Steven Johnson | 13 November 2019, 18:13:23 UTC | Merge pull request #4401 from vksnk/fix-doc Fix comment on how to build tutorial_lesson_05 with make | 13 November 2019, 18:13:23 UTC |
291bc72 | Steven Johnson | 13 November 2019, 01:20:42 UTC | Merge branch 'master' into pdb_mul_div_mod_multi_thread | 13 November 2019, 01:20:42 UTC |
64ffb4e | Volodymyr Kysenko | 13 November 2019, 00:57:11 UTC | Fix comment | 13 November 2019, 00:57:11 UTC |
a5621ec | Steven Johnson | 12 November 2019, 23:34:39 UTC | Merge pull request #4398 from halide/srj-simd Fix set_host_alignment() in simd_op_check | 12 November 2019, 23:34:39 UTC |
b962486 | Steven Johnson | 12 November 2019, 23:27:23 UTC | Update Makefile | 12 November 2019, 23:27:23 UTC |
e70e688 | Steven Johnson | 12 November 2019, 19:31:55 UTC | Update simd_op_check.h | 12 November 2019, 19:31:55 UTC |
47d70d1 | Steven Johnson | 12 November 2019, 19:18:05 UTC | Update simd_op_check_hvx.cpp | 12 November 2019, 19:18:05 UTC |
8777450 | Steven Johnson | 12 November 2019, 18:54:41 UTC | Fix set_host_alignment() in simd_op_check We have been calling set_host_alignment(128) on our input params for ~ever, but this has always been wrong; as the argument is in bytes (not bits), and most of our targets provide only a 32-byte guarantee at the memory allocator level (x86 does 64-bytes, see LLVM_Runtime_Linker.cpp) (Note that I'm not sure if the _hvx variant really should be 128 bytes or not, please review carefully.) | 12 November 2019, 18:54:41 UTC |
a47d58e | Steven Johnson | 12 November 2019, 01:59:27 UTC | Merge pull request #4389 from halide/srj-ub Fix UB in Halide::Runtime::Buffer<>::size_in_bytes() | 12 November 2019, 01:59:27 UTC |
fd3da4d | Steven Johnson | 12 November 2019, 01:50:58 UTC | Update README for LLVM build instructions (Issue #4394) | 12 November 2019, 01:50:58 UTC |
53bb8ef | Steven Johnson | 11 November 2019, 23:24:33 UTC | avoid reinterpret_cast in onnx_converter | 11 November 2019, 23:26:21 UTC |
568ddfc | Steven Johnson | 11 November 2019, 21:48:42 UTC | Fix UB in Halide::Runtime::Buffer<>::size_in_bytes() | 11 November 2019, 21:48:42 UTC |
f60b8d2 | Steven Johnson | 06 November 2019, 22:02:38 UTC | Merge pull request #4383 from halide/srj-pt Misc minor PyTorch cleanups | 06 November 2019, 22:02:38 UTC |
50f6f92 | Steven Johnson | 06 November 2019, 18:14:37 UTC | Merge branch 'master' into pdb_mul_div_mod_multi_thread | 06 November 2019, 18:14:37 UTC |
d4aace9 | Steven Johnson | 06 November 2019, 18:13:08 UTC | Merge pull request #4380 from aankit-ca/offload_autosched_results Fix offload mode autoscheduler results. | 06 November 2019, 18:13:08 UTC |
aab0cd3 | Shubham Pawar | 06 November 2019, 16:11:05 UTC | Add null pointer check for results Check if autoscheduler results are present before copying to the lowered module after resolving submodules. | 06 November 2019, 16:11:05 UTC |
1f0179b | Pranav Bhandarkar | 06 November 2019, 00:33:02 UTC | Revert "Add default constexpr constructor to ImplicitVar because Clang complaints about it" This reverts commit 9ff00ffe4ed0f2d476ba974ce1bad15bd70d9390. | 06 November 2019, 00:33:02 UTC |
d76fa09 | Pranav Bhandarkar | 06 November 2019, 00:21:08 UTC | Revert PR #4336 as underlying issue in LLVM has been addressed by LLVM commit 041f35c Link - https://github.com/llvm/llvm-project/commit/041f35c468088d315bae6c2a71ec901a12cca1b5 | 06 November 2019, 00:27:22 UTC |
6104128 | Steven Johnson | 06 November 2019, 00:12:31 UTC | Misc minor PyTorch cleanups - remove no-longer-needed .gitignore - trim all trailing blank space - add estimates to add_generator.cpp - Codegen_PyTorch declares the `_th_` functions as `inline` | 06 November 2019, 00:12:31 UTC |
befed33 | Pranav Bhandarkar | 06 November 2019, 00:04:24 UTC | Merge branch 'master' of https://github.com/halide/Halide | 06 November 2019, 00:04:24 UTC |
f2062aa | Shubham Pawar | 05 November 2019, 22:30:58 UTC | Fix offload mode autoscheduler results. Currently, when autoscheduler is invoked in offload mode, the the emit options {featurization and schedule} are not working as expected and bailing out due to auto_scheduler_results not present in the lowered module, resulting in empty feature and schedule files produced. This happens because, when offloaded submodules are recursively resolved, the autoscheduler results are not copied back into the lowered module. This change fixes this issue by setting back the auto_scheduler_results in the ModuleContents of the module that is lowered after resolving submodules. | 05 November 2019, 22:58:31 UTC |
d10ca19 | Steven Johnson | 05 November 2019, 21:37:47 UTC | Merge pull request #4373 from halide/srj-pytorch Fix OSX+Homebrew+Python issue for HelloPyTorch | 05 November 2019, 21:37:47 UTC |
ce5b510 | Steven Johnson | 04 November 2019, 17:06:56 UTC | Merge pull request #4375 from halide/srj-fix ScheduleFeatures::num_features() and PipelineFeatures::num_features()… | 04 November 2019, 17:06:56 UTC |
e3055e9 | Steven Johnson | 01 November 2019, 23:06:46 UTC | ScheduleFeatures::num_features() and PipelineFeatures::num_features() should return size_t (Issue #4371) | 01 November 2019, 23:06:46 UTC |
893d986 | Steven Johnson | 01 November 2019, 22:21:04 UTC | Merge pull request #4374 from halide/srj-mangle Fix name-mangling [u]int64 on IOS | 01 November 2019, 22:21:04 UTC |
0f32df5 | Steven Johnson | 01 November 2019, 22:20:49 UTC | Add iOS test cases for mangling | 01 November 2019, 22:20:49 UTC |
c2a46b7 | Steven Johnson | 01 November 2019, 20:23:47 UTC | Fix name-mangling [u]int64 on IOS The iOS SDK defines int64 as `long long` and uint64 as `unsigned long long`, but we are currently mangling them as `long` and `unsigned long`, so we will get linker errors for compiled filters that use one of these types as an input (assuming c_plus_plus_name_mangling is enabled). | 01 November 2019, 20:23:47 UTC |
38bde2e | Steven Johnson | 01 November 2019, 00:39:40 UTC | Makefile reformatting | 01 November 2019, 00:39:40 UTC |
cf4d8d0 | Steven Johnson | 01 November 2019, 00:27:14 UTC | Fix OSX+Homebrew+Python issue for HelloPyTorch Some combinations have a known issue and the `pip install` part of the Makefile will fail with `DistutilsOptionError: must supply either home or prefix/exec-prefix — not both` Adding this dummy setup.cfg is (apparently) a reasonably safe solution, and works on my setup. | 01 November 2019, 00:27:14 UTC |
6e25c57 | Steven Johnson | 31 October 2019, 17:18:14 UTC | Merge pull request #4253 from halide/srj-passman Re-Enable NewPassManager, plus ASAN fixes | 31 October 2019, 17:18:14 UTC |
c9cc634 | Steven Johnson | 31 October 2019, 16:56:59 UTC | Merge pull request #4369 from halide/srj-plugin Add Halide::load_plugin | 31 October 2019, 16:56:59 UTC |
2bd118b | Steven Johnson | 31 October 2019, 16:56:51 UTC | Merge pull request #4370 from halide/srj-dbg Reduce binary size of of libauto_schedule.so | 31 October 2019, 16:56:51 UTC |
a01242e | Steven Johnson | 31 October 2019, 00:53:10 UTC | Update Makefile | 31 October 2019, 00:53:10 UTC |
5d51cde | Steven Johnson | 30 October 2019, 23:43:58 UTC | Update Makefile | 30 October 2019, 23:43:58 UTC |
41733d1 | Steven Johnson | 30 October 2019, 22:45:51 UTC | Reduce binary size of of libauto_schedule.so - Don't include debug symbols (`-g`) by default (>10x size savings); when necessary for debugging, the preferred fix here it override the value of the OPTIMIZE flag - Add `-fvisibility=hidden -fvisibility-inlines-hidden` to save another 30k or so | 30 October 2019, 22:45:51 UTC |
6a32ce7 | Steven Johnson | 30 October 2019, 22:33:25 UTC | fix compiler flags | 30 October 2019, 22:33:25 UTC |
6995a48 | Steven Johnson | 30 October 2019, 22:27:01 UTC | various fixes | 30 October 2019, 22:27:01 UTC |
5a68a20 | Steven Johnson | 30 October 2019, 21:49:50 UTC | Merge branch 'master' into srj-passman | 30 October 2019, 21:49:50 UTC |
99d33fb | Dillon Sharlet | 30 October 2019, 18:13:08 UTC | Merge pull request #4357 from kasper93/vtcm Hexagon: Fix VTCM support on targets > v65 | 30 October 2019, 18:13:08 UTC |
87cfbba | Steven Johnson | 30 October 2019, 17:41:49 UTC | Update test.cpp | 30 October 2019, 17:41:49 UTC |
8c3f233 | Steven Johnson | 30 October 2019, 17:36:32 UTC | Add Halide::load_plugin Promote this to a first-class Halide API, along with Python wrapper. Remove the python-related code from apps/gradient_autoscheduler and use this instead. Also drive-by cleanup in gradient_autoscheduler to run all the tests properly. | 30 October 2019, 17:36:32 UTC |
2fcbe76 | Steven Johnson | 30 October 2019, 16:56:24 UTC | Merge pull request #4367 from halide/srj-schednames Allow multiple autoschedulers to be registered (Issue #4057) | 30 October 2019, 16:56:24 UTC |
0b12ef8 | Andrew Adams | 30 October 2019, 16:37:29 UTC | Merge pull request #4338 from halide/tzumao-autodiff-remove-boundary Remove boundary conditions in autodiff | 30 October 2019, 16:37:29 UTC |
20494f2 | Steven Johnson | 30 October 2019, 16:27:38 UTC | Merge pull request #4366 from halide/tzumao-python-undef-buffer Add Buffer.defined() checks in Python bindings when printing buffers | 30 October 2019, 16:27:38 UTC |
91e5c73 | kasper93 | 28 October 2019, 17:58:31 UTC | Hexagon: Fix VTCM support on targets > v65 | 30 October 2019, 10:09:59 UTC |
cda6dd1 | Steven Johnson | 30 October 2019, 00:59:15 UTC | Update GradientAutoscheduler.cpp | 30 October 2019, 00:59:15 UTC |
95682f0 | Steven Johnson | 30 October 2019, 00:58:12 UTC | Update CMakeLists.txt | 30 October 2019, 00:58:12 UTC |
1ac841e | Steven Johnson | 30 October 2019, 00:56:03 UTC | Merge branch 'master' into srj-schednames | 30 October 2019, 00:56:03 UTC |
be254dd | Steven Johnson | 30 October 2019, 00:54:53 UTC | Merge pull request #4365 from halide/srj-grad gradient_autoscheduler fixes | 30 October 2019, 00:54:53 UTC |
f189a8f | Steven Johnson | 29 October 2019, 23:38:28 UTC | Merge pull request #4352 from benoitsteiner/master Improved the onnx converter | 29 October 2019, 23:38:28 UTC |
2c1da11 | Steven Johnson | 29 October 2019, 23:32:48 UTC | Various fixes | 29 October 2019, 23:32:48 UTC |
f0931eb | Steven Johnson | 29 October 2019, 22:46:46 UTC | Merge branch 'master' into srj-schednames | 29 October 2019, 22:46:46 UTC |
9fc80be | Steven Johnson | 29 October 2019, 22:45:22 UTC | Move plugin-loading to Util.cpp | 29 October 2019, 22:45:22 UTC |
d8099e7 | Steven Johnson | 29 October 2019, 22:32:10 UTC | Update Generator.cpp | 29 October 2019, 22:32:10 UTC |
2bcdd6a | Steven Johnson | 29 October 2019, 22:27:57 UTC | Allow multiple autoschedulers to be registered (Issue #4057) Currently, the plugin autoschedulers take over all autoschedule hooks once loaded; there's no way to mix and match between them, and no API-level way to control which one you get. This reworks this nontrivially: - Names for the autoschedulers are revised per the comments in #4057 - We now keep a global map of all known schedulers, with the 'classic' (aka 'Mullapudi2016') always available for now - When you load a plugin (either explicitly via dlopen, or manually via the `-p` option to Generator_main), it does *not* change the default scheduler, it just adds itself to the map. - There is a new API to set the default autoscheduler to use (`Pipeline::set_default_autoscheduler_name`) - There is also a new generator_main flag (`-s`) to set the default autoscheduler to use - The existing call to `Pipeline::auto_schedule()` uses the default autoscheduler name; there is also a new variant in which you can pass the autoscheduler name explicitly. Stuff to consider: - I contemplated adding the `autoscheduler_name` as a new field to MachineParams, but decided that was a bad fit. - I contemplated avoiding the `-s` argument for Generators entirely and instead somehow overloading `auto_schedule=` instead, but this seems fraught with weirdness to make legacy code work without changes. (It still feels like something that might make more sense as a GeneratorParam, but not sure yet.) - It might make sense to (someday) have plugin loading be done implicitly when `-s` is specified, but definitely not yet. | 29 October 2019, 22:27:57 UTC |
ac60839 | Steven Johnson | 29 October 2019, 22:15:04 UTC | Merge pull request #4362 from halide/srj-samp Fix apps/autoscheduler out-of-tree builds (Issue #4355) | 29 October 2019, 22:15:04 UTC |
a34157f | Tzu-Mao Li | 29 October 2019, 21:15:25 UTC | Add Buffer.defined() checks in Python bindings when printing buffers | 29 October 2019, 21:15:25 UTC |
17efee4 | Steven Johnson | 29 October 2019, 20:52:11 UTC | gradient_autoscheduler fixes - Migrate the aslog() utility from apps/autoscheduler and use aslog(1) instead of std::cerr (schedulers need to be *absolutely silent* by default to avoid build-system spam) - Add gradient_autoscheduler to 'make test_apps' - build with -fvisibility=hidden to save code size - default WITH_PYTHON to 0... as written, this requires every piece of code using to link to Python, which is definitely not what we want. (The code should be modified to either have a separate library for Python usage, or to find a way to safely lazy-load the Python extension library.) - Use `$(PYTHON)-config --cflags` to ensure path to Python.h can be found (if WITH_PYTHON=1 is set) | 29 October 2019, 20:52:11 UTC |
488f315 | Tzu-Mao Li | 29 October 2019, 18:49:30 UTC | Merge branch 'master' into tzumao-autodiff-remove-boundary | 29 October 2019, 18:49:30 UTC |
667d4b1 | Benoit Steiner | 29 October 2019, 18:42:56 UTC | Added comment | 29 October 2019, 18:42:56 UTC |
499b8f5 | Tzu-Mao Li | 29 October 2019, 18:03:02 UTC | Merge pull request #4340 from halide/tzumao-gradient-autoscheduler Gradient autoscheduler | 29 October 2019, 18:03:02 UTC |
96badf9 | Steven Johnson | 29 October 2019, 17:10:33 UTC | Merge branch 'master' into srj-samp | 29 October 2019, 17:10:33 UTC |
7d7701e | Steven Johnson | 29 October 2019, 17:09:48 UTC | Merge pull request #4361 from halide/srj-strict-2 Make is_nan/is_inf/is_finite enforce strict_float() (Issue #4281) | 29 October 2019, 17:09:48 UTC |
7d47d4f | Benoit Steiner | 29 October 2019, 16:36:27 UTC | Improved comment | 29 October 2019, 16:36:27 UTC |
1b487da | Benoit Steiner | 29 October 2019, 16:29:44 UTC | Don't import from __future__ since Halide switched to python v3 | 29 October 2019, 16:29:44 UTC |
15ebae0 | Benoit Steiner | 29 October 2019, 16:28:58 UTC | Apply clang-format | 29 October 2019, 16:28:58 UTC |
71126ad | Tzu-Mao Li | 29 October 2019, 02:30:09 UTC | remove Derivative.funcs in python bindings | 29 October 2019, 02:30:09 UTC |
08f2168 | Tzu-Mao Li | 28 October 2019, 22:28:40 UTC | Fix fit function | 28 October 2019, 22:28:40 UTC |
739fa6d | Steven Johnson | 28 October 2019, 22:25:23 UTC | Update IROperator.cpp | 28 October 2019, 22:25:23 UTC |
2a03a4d | Tzu-Mao Li | 28 October 2019, 21:59:39 UTC | Set WITH_PYTHON's default to 1 | 28 October 2019, 21:59:39 UTC |
01978f3 | Steven Johnson | 28 October 2019, 21:58:52 UTC | Fix apps/autoscheduler out-of-tree builds (Issue #4355) - `make test` should ensure that the `samples` output dir is a subdir of `$(BIN)`, so that it doesn't go in-source-tree and fill up the buildbots - autoscheduler.inc should default $(AUTOSCHED_BIN) to match $(BIN) so that the toplevel Makefile can override it properly (previously we ended up with some build products in-tree and some in-build-products-dir) | 28 October 2019, 21:58:52 UTC |
40af63e | Tzu-Mao Li | 28 October 2019, 21:48:48 UTC | style: << in front of the line instead of at the end | 28 October 2019, 21:48:48 UTC |
dcc91ea | Tzu-Mao Li | 28 October 2019, 21:43:59 UTC | wording | 28 October 2019, 21:43:59 UTC |
0f08ca3 | Tzu-Mao Li | 28 October 2019, 21:35:23 UTC | properly schedule the derivatives | 28 October 2019, 21:35:23 UTC |
bee033f | Tzu-Mao Li | 28 October 2019, 21:02:50 UTC | remove funcs() | 28 October 2019, 21:02:50 UTC |
8fa7c46 | Steven Johnson | 28 October 2019, 21:00:50 UTC | Make is_nan/is_inf/is_finite enforce strict_float() (Issue #4281) Quietly wrap the arg to these functions in strict_float(), to ensure the results aren't meaningless. (Also: drive-by cleanup of code in StrictifyFloat that was a holdover from when it was possible to have contexts in which strict-float was not allowed) | 28 October 2019, 21:00:50 UTC |
972d521 | Tzu-Mao Li | 28 October 2019, 20:47:45 UTC | Remove Derivative::funcs() | 28 October 2019, 20:47:45 UTC |
5df73bd | Steven Johnson | 28 October 2019, 15:54:16 UTC | Merge pull request #4356 from halide/tzumao-python-func-in Change Func.in to Func.inside in Python | 28 October 2019, 15:54:16 UTC |
12aaefe | Steven Johnson | 28 October 2019, 15:49:13 UTC | Merge pull request #4342 from kasper93/generator_outputs Generator outputs | 28 October 2019, 15:49:13 UTC |
51945f0 | Tzu-Mao Li | 28 October 2019, 15:41:23 UTC | Merge branch 'master' into tzumao-gradient-autoscheduler | 28 October 2019, 15:41:23 UTC |
20ef6cc | Tzu-Mao Li | 28 October 2019, 15:41:08 UTC | Merge branch 'master' into tzumao-autodiff-remove-boundary | 28 October 2019, 15:41:08 UTC |
89136a0 | Tzu-Mao Li | 28 October 2019, 15:28:19 UTC | Merge branch 'master' into tzumao-autodiff-remove-boundary | 28 October 2019, 15:28:19 UTC |
c967e9e | Tzu-Mao Li | 28 October 2019, 15:18:29 UTC | Merge branch 'tzumao-gradient-autoscheduler' of https://github.com/halide/Halide into tzumao-gradient-autoscheduler | 28 October 2019, 15:18:29 UTC |
6947f58 | Tzu-Mao Li | 28 October 2019, 15:18:19 UTC | Merge branch 'master' into tzumao-gradient-autoscheduler | 28 October 2019, 15:18:19 UTC |
778d4b2 | kasper93 | 25 October 2019, 16:02:10 UTC | CMake: Fix output file names for generator Multiple entries were missing and some others had incorrect extension, resulting in wrong dependency handling by CMake. | 28 October 2019, 11:43:54 UTC |
579a739 | kasper93 | 25 October 2019, 16:02:05 UTC | Do not print legacy aliases in generator's supported outputs list Also skip last space when printing list | 28 October 2019, 11:38:38 UTC |
c796f50 | Tzu-Mao Li | 28 October 2019, 05:11:36 UTC | Merge pull request #4350 from halide/tzumao-allocation_cache Let halide_reuse_device_allocations defaults to be true | 28 October 2019, 05:11:36 UTC |
b754470 | Tzu-Mao Li | 28 October 2019, 02:27:40 UTC | inside to in_ | 28 October 2019, 02:27:40 UTC |
daa434a | Tzu-Mao Li | 28 October 2019, 01:36:21 UTC | Comments | 28 October 2019, 01:36:21 UTC |