| 56e2041 | Steven Johnson | 19 August 2019, 17:24:53 UTC | Merge pull request #4146 from halide/srj-tsan-features Make can_use_target TSAN-safe (Issue #4140) | 19 August 2019, 17:24:53 UTC |
| c8af918 | Steven Johnson | 17 August 2019, 00:44:04 UTC | Only test autoscheduler on Linux | 17 August 2019, 00:44:04 UTC |
| 34364d7 | Steven Johnson | 16 August 2019, 23:57:47 UTC | Avoid unused-var warnings, grr | 16 August 2019, 23:57:47 UTC |
| 3cc46cb | Andrew Adams | 16 August 2019, 23:42:11 UTC | Merge remote-tracking branch 'origin/master' into bfloat16 | 16 August 2019, 23:42:11 UTC |
| 8c7338a | Steven Johnson | 16 August 2019, 23:22:24 UTC | Make can_use_target TSAN-safe (Issue #4140) | 16 August 2019, 23:22:24 UTC |
| 37edd63 | Steven Johnson | 16 August 2019, 22:48:44 UTC | Update autotune_loop.sh | 16 August 2019, 22:48:44 UTC |
| 7871e70 | Steven Johnson | 16 August 2019, 22:15:04 UTC | Update retrain_cost_model.cpp | 16 August 2019, 22:15:04 UTC |
| b9eaa49 | Steven Johnson | 16 August 2019, 21:28:51 UTC | Update retrain_cost_model.cpp | 16 August 2019, 21:28:51 UTC |
| 9d36f67 | Steven Johnson | 16 August 2019, 21:24:03 UTC | --weights_out -> required; --weights -> --initial_weights (optional) | 16 August 2019, 21:24:03 UTC |
| dc07855 | Steven Johnson | 16 August 2019, 21:01:11 UTC | Update autotune_loop.sh | 16 August 2019, 21:01:11 UTC |
| 692be4f | Steven Johnson | 16 August 2019, 19:56:44 UTC | Rename | 16 August 2019, 19:56:44 UTC |
| e400274 | Steven Johnson | 16 August 2019, 19:55:14 UTC | Improve train_cost_model tool - Rename to retrain_cost_model, since (1) that's what it does, and (2) this avoids confusion with the Generator-produces filter of the same name - Remove all use of env vars for inputs; explicit commandline flags are now required instead. Update autotune_loop.sh accordingly. - add an option to copy the best .schedule.h file somewhere at the end (note that the previous HL_BEST_SCHEDULE_FILE env var actually output benchmark results, not the schedule file) - Check the return value of calls to train_cost_model() and cost_model(), just in case someone uses a nonstandard override of halide_error. - Tighten some verbosity of output (display just filenames instead of pathnames in some cases) - Some minor drive-by style cleanups. | 16 August 2019, 19:55:14 UTC |
| 872c56d | Michael Gharbi | 16 August 2019, 19:22:57 UTC | fixes Make dependencies | 16 August 2019, 19:22:57 UTC |
| bfc184d | Michael Gharbi | 16 August 2019, 19:20:36 UTC | fixes python indent | 16 August 2019, 19:20:36 UTC |
| e206cc8 | Michael Gharbi | 16 August 2019, 19:16:11 UTC | removes WEAK qualifier for cuda hooks | 16 August 2019, 19:16:11 UTC |
| cb23358 | Michael Gharbi | 16 August 2019, 19:04:33 UTC | fixes cpp indent | 16 August 2019, 19:04:33 UTC |
| b39793d | Michael Gharbi | 16 August 2019, 18:25:43 UTC | fixes typos in readme | 16 August 2019, 18:25:43 UTC |
| 60f1c5c | Steven Johnson | 16 August 2019, 17:28:17 UTC | Merge pull request #4143 from halide/fix_simplifier_bounds_overflow Avoid 64-bit overflow when tracking constant bounds in simplifier | 16 August 2019, 17:28:17 UTC |
| 83ab2c9 | Michael Gharbi | 16 August 2019, 01:21:28 UTC | rewrite adjoint using Derivative.cpp | 16 August 2019, 01:21:28 UTC |
| 27e9440 | Michael Gharbi | 16 August 2019, 00:53:08 UTC | adds a gradient check | 16 August 2019, 00:53:08 UTC |
| 538a804 | Michael Gharbi | 16 August 2019, 00:47:31 UTC | adds note on test | 16 August 2019, 00:47:31 UTC |
| a7a9574 | Michael Gharbi | 16 August 2019, 00:32:20 UTC | adds readme | 16 August 2019, 00:32:20 UTC |
| c93ea38 | Michael Gharbi | 16 August 2019, 00:32:11 UTC | fix output | 16 August 2019, 00:32:11 UTC |
| 7ab366c | Michael Gharbi | 16 August 2019, 00:06:32 UTC | merge | 16 August 2019, 00:14:21 UTC |
| 7fa1028 | Andrew Adams | 15 August 2019, 20:38:15 UTC | Avoid 64-bit overflow when tracking constant bounds in simplifier These are no-overflow types, so treat such cases not as things that overflow (in which an overflow in the upper bound would affect the lower bound), but rather as things which still have upper/lower bounds that we can no longer represent. | 15 August 2019, 20:38:15 UTC |
| 4628a24 | Steven Johnson | 15 August 2019, 18:58:21 UTC | Merge pull request #4137 from halide/srj-has-feat Allow Target::has_feature() to operator on halide_target_feature_t | 15 August 2019, 18:58:21 UTC |
| 1d49dbd | Steven Johnson | 15 August 2019, 00:36:51 UTC | Allow Target::has_feature() to operator on halide_target_feature_t This is really a workaround to allow generated .schedule.h files to compile happily in all cases. A better long-term answer would be for those files to use the Target::Feature names (rather than the halide_target_feature_t names), but doing so will require an additional value->name map (which must be maintained and could get out of sync), so I'm reluctant to do that without a little more thought. | 15 August 2019, 17:38:11 UTC |
| 07c3c27 | Steven Johnson | 15 August 2019, 17:07:02 UTC | Merge pull request #4138 from halide/srj-llvm-fix Fix for https://reviews.llvm.org/rL368647, from alinas@ | 15 August 2019, 17:07:02 UTC |
| 70485a7 | Steven Johnson | 15 August 2019, 17:03:16 UTC | Update CodeGen_LLVM.cpp | 15 August 2019, 17:03:16 UTC |
| 3affe21 | Andrew Adams | 15 August 2019, 16:38:35 UTC | Merge remote-tracking branch 'origin/master' into bfloat16 | 15 August 2019, 16:38:35 UTC |
| 98991c0 | Steven Johnson | 15 August 2019, 01:16:00 UTC | Fix for https://reviews.llvm.org/rL368647, from alinas@ | 15 August 2019, 01:16:00 UTC |
| 7eb5b24 | Steven Johnson | 14 August 2019, 22:35:38 UTC | Merge pull request #4131 from halide/srj-tic Change halide_toc_impl to use debug(1) instead of debug(0) | 14 August 2019, 22:35:38 UTC |
| 3d29b66 | Andrew Adams | 14 August 2019, 22:16:24 UTC | Avoid half types in the x86 backend | 14 August 2019, 22:16:24 UTC |
| 8766bc4 | Steven Johnson | 14 August 2019, 21:13:43 UTC | Merge pull request #4133 from halide/srj-aslog2 Clean up default autotuner output | 14 August 2019, 21:13:43 UTC |
| 5afe043 | Steven Johnson | 14 August 2019, 18:45:51 UTC | Merge pull request #4129 from benoitsteiner/master Improved onnx support | 14 August 2019, 18:45:51 UTC |
| 2576f3d | Steven Johnson | 14 August 2019, 18:09:11 UTC | Merge pull request #4134 from halide/srj-llvm-fix Fix for Trunk LLVM | 14 August 2019, 18:09:11 UTC |
| 19964e9 | Steven Johnson | 14 August 2019, 16:58:31 UTC | Fix for Trunk LLVM | 14 August 2019, 17:03:19 UTC |
| 7c78788 | Steven Johnson | 14 August 2019, 16:44:04 UTC | Update DefaultCostModel.cpp | 14 August 2019, 16:44:04 UTC |
| cdec241 | Steven Johnson | 13 August 2019, 23:48:05 UTC | Back to dump | 13 August 2019, 23:48:05 UTC |
| 5f56d69 | Steven Johnson | 13 August 2019, 23:37:29 UTC | Update AutoSchedule.cpp | 13 August 2019, 23:37:29 UTC |
| a6599e7 | Steven Johnson | 13 August 2019, 23:20:43 UTC | Update AutoSchedule.cpp | 13 August 2019, 23:20:43 UTC |
| d99d640 | Steven Johnson | 13 August 2019, 23:19:40 UTC | Also emit the target being used | 13 August 2019, 23:19:40 UTC |
| 9808edc | Steven Johnson | 13 August 2019, 23:15:43 UTC | Simplify further | 13 August 2019, 23:15:43 UTC |
| 6dbfc57 | Steven Johnson | 13 August 2019, 22:51:46 UTC | Merge pull request #4132 from halide/srj-mod2 Tweak schedule.h output | 13 August 2019, 22:51:46 UTC |
| ee60804 | Steven Johnson | 13 August 2019, 22:51:33 UTC | Fix indentation | 13 August 2019, 22:51:33 UTC |
| 90c52f8 | Steven Johnson | 13 August 2019, 22:48:47 UTC | Clean up default autotuner output This restructures so that the default output of the autoscheduler (at HL_DEBUG_AUTOSCHEDULE=0) is much smaller, with just some output for each pass. (All the previous output is still present for HL_DEBUG_AUTOSCHEDULE=1.) Not entire sure if this is the best "per-pass" summary; definitely open for suggestions here (even shorter/terser would be even nicer here IMHO). Also, a little drive-by cleanup to the progress bar (to avoid leaving them dangling) and to give slightly better feedback during the autotune script. | 13 August 2019, 22:48:47 UTC |
| 9a2224b | Pranav Bhandarkar | 13 August 2019, 21:53:31 UTC | Merge branch 'master' of https://github.com/halide/Halide | 13 August 2019, 21:53:31 UTC |
| ed4add8 | Steven Johnson | 13 August 2019, 21:13:50 UTC | Tweak schedule.h output - avoid unnecessary blank lines - avoid lines that are whitespace-only | 13 August 2019, 21:13:50 UTC |
| 50f0861 | Steven Johnson | 13 August 2019, 18:04:21 UTC | Change halide_toc_impl to use debug(1) instead of debug(0) While handy for debugging, this means the autoscheduler currently spams stdout with timing info by default, which is puzzling to people not actually working on the autoscheduler. (alternately: add debug-level option to HALIDE_TIC, or just comment out the calls to it in the autoscheduler) | 13 August 2019, 18:04:21 UTC |
| 67b9a52 | Andrew Adams | 13 August 2019, 17:22:47 UTC | Fix unhandled f16 cast in arm backend | 13 August 2019, 17:22:47 UTC |
| 5700dbb | Steven Johnson | 12 August 2019, 23:46:03 UTC | Merge pull request #4128 from halide/srj-rg2 RunGen: bounds-query failure shouldn't matter if we use estimates anyway | 12 August 2019, 23:46:03 UTC |
| fe9d62b | Andrew Adams | 12 August 2019, 23:36:50 UTC | Merge remote-tracking branch 'origin/master' into bfloat16 | 12 August 2019, 23:36:50 UTC |
| 14f9fb2 | Benoit Steiner | 12 August 2019, 23:15:00 UTC | Improved conversion of onnx to halide: * avoid the creation of identity functions that slow down the scheduling. * recusrively inline and simplify shape expressions to make them more amenable to analysis during scheduling | 12 August 2019, 23:15:00 UTC |
| eda5da8 | Benoit Steiner | 12 August 2019, 23:08:09 UTC | Fixed the code that converts numpy arrays to Halide input buffers | 12 August 2019, 23:08:09 UTC |
| 2a4f364 | Steven Johnson | 12 August 2019, 21:09:20 UTC | RunGen: bounds-query failure shouldn't matter if we use estimates anyway bounds_query_input_shapes() agressively fails if we can't complete the bounds-query, which can happen if the constraints on inputs are nontrivial. While we can (and should) improve the bounds-query logic to make this more robust, we shouldn't aggressively fail here in the first place, as the bounds-query shape(s) end up unused if we have estimates for the inputs (which we usually do). This just adds a ShapePromise type that wraps access to the resulting Shapes inside a function; if we never need the shape, the failure doesn't matter and never happens. | 12 August 2019, 21:09:20 UTC |
| a8128b1 | Steven Johnson | 09 August 2019, 22:49:37 UTC | Merge pull request #4123 from halide/srj-weights Convert autoscheduler weights to a single file | 09 August 2019, 22:49:37 UTC |
| d4ec333 | Steven Johnson | 09 August 2019, 18:13:22 UTC | Update Weights.cpp | 09 August 2019, 18:13:22 UTC |
| 94dd5d7 | Steven Johnson | 09 August 2019, 18:11:21 UTC | Add versions to the weights file | 09 August 2019, 18:11:21 UTC |
| 0f24963 | Steven Johnson | 09 August 2019, 01:02:37 UTC | Convert autoscheduler weights to a single file The main gist of this change is to put all of the weights used by the autoscheduler into a single file (instead of a directory of multiple files); the goal is to make saving 'overnight' tuned weights for a pipeline simpler, via having only a single `.weights` file rather than a directory of things. There are various drive-by fixes, of course: - The 'Weights' struct moved into its own file, and now manages its own loading/saving, in both the classic dir-of-files format, and new, single-file format. (The single-file format is very simpleminded but likely more than adequate; it has a trivial signature at the front so some smart upgrading over time should be easy to do.) - DefaultCostModel.cpp got its own .h file, instead of being incestuously connected to the ABC it implements - existing `weights` folder was deleted; the equivalent data is now in `baseline.weights` - added the `weightsdir_to_weightsfile` utility for any downstream consumers that have custom weights in the 'old' format that they need to update to the 'new' format. - Removed some unused code here and there. Note that HL_WIEGHTS_DIR (and HL_WEIGHTS_OUT_DIR) are still used, but are 'smart' for now; if you pass a pathname that ends in `.weights` it loads as a single file, otherwise it assumes an old-style directory. (I didn't try to smarten the env var names because we really need to eradicate the use of env vars for this purpose anyway. but that's a job for another day.) | 09 August 2019, 01:02:37 UTC |
| 6a36290 | Steven Johnson | 08 August 2019, 21:41:42 UTC | Merge pull request #4120 from halide/srj-autotune Pass $HALIDE_DISTRIB_PATH to autotune_loop.sh | 08 August 2019, 21:41:42 UTC |
| 9d72f4a | Pranav Bhandarkar | 08 August 2019, 21:27:13 UTC | Merge branch 'master' of https://github.com/halide/Halide | 08 August 2019, 21:27:13 UTC |
| d89c94b | Steven Johnson | 08 August 2019, 21:25:20 UTC | Update autotune_loop.sh | 08 August 2019, 21:25:20 UTC |
| aa03f99 | Steven Johnson | 08 August 2019, 21:23:11 UTC | Pass $HALIDE_DISTRIB_PATH to autotune_loop.sh | 08 August 2019, 21:23:11 UTC |
| 3b518a6 | Steven Johnson | 08 August 2019, 21:20:15 UTC | Merge pull request #4119 from halide/srj-dag Avoid signed/unsigned warnings in FunctionDAG.h | 08 August 2019, 21:20:15 UTC |
| c7fd30c | Steven Johnson | 08 August 2019, 21:15:07 UTC | Avoid signed/unsigned warnings in FunctionDAG.h | 08 August 2019, 21:15:07 UTC |
| ac66a97 | Steven Johnson | 08 August 2019, 21:04:56 UTC | Merge pull request #4112 from halide/srj-dag Avoid crash in FunctionDAG.h (Issue #4110) | 08 August 2019, 21:04:56 UTC |
| 6292d7a | Andrew Adams | 08 August 2019, 20:51:05 UTC | Don't rely on opencl supporting half types | 08 August 2019, 20:51:05 UTC |
| bf6cf87 | Andrew Adams | 08 August 2019, 20:43:09 UTC | Merge remote-tracking branch 'origin/master' into bfloat16 | 08 August 2019, 20:43:09 UTC |
| bf74fba | Steven Johnson | 08 August 2019, 20:41:11 UTC | Merge pull request #4114 from halide/srj-loopt Add EnableLLVMLoopOpt and DisableLLVMLoopOpt (Issue #4113) | 08 August 2019, 20:41:11 UTC |
| a8e1303 | Steven Johnson | 08 August 2019, 20:17:49 UTC | Merge branch 'master' into srj-dag | 08 August 2019, 20:17:49 UTC |
| 0d50bc9 | Steven Johnson | 08 August 2019, 20:16:16 UTC | Merge pull request #4117 from halide/srj-incl PerfectHashMap.h should be standalone | 08 August 2019, 20:16:16 UTC |
| 7857ea6 | Steven Johnson | 08 August 2019, 20:15:33 UTC | Merge pull request #4102 from halide/srj-autoclean Strip 'irrelevant' features from the target comment when generating schedule.h files | 08 August 2019, 20:15:33 UTC |
| adb5bde | Steven Johnson | 08 August 2019, 18:58:31 UTC | Update PerfectHashMap.h | 08 August 2019, 18:58:31 UTC |
| 01ecadc | Steven Johnson | 08 August 2019, 18:55:12 UTC | PerfectHashMap.h should be standalone We don't want it to require libHalide, but including Errors.h implicitly does that. Move the 'standalone' asserter into the template class and out of the test, so that the .h file will compile standalone (regardless of what is #defined before it). | 08 August 2019, 18:55:12 UTC |
| b69fcc9 | Steven Johnson | 08 August 2019, 17:58:13 UTC | Add enable_llvm_loop_opt | 08 August 2019, 17:58:13 UTC |
| a0d5f28 | Steven Johnson | 08 August 2019, 17:46:43 UTC | Combine disable-llvm-loop features into one | 08 August 2019, 17:46:43 UTC |
| 081141e | Steven Johnson | 08 August 2019, 17:18:37 UTC | Update Module.cpp | 08 August 2019, 17:18:37 UTC |
| 2558fa6 | Steven Johnson | 08 August 2019, 17:16:04 UTC | Merge pull request #4106 from halide/distinct_wrapper_names Wrappers need distinct names | 08 August 2019, 17:16:04 UTC |
| 39682df | Steven Johnson | 08 August 2019, 17:14:26 UTC | Avoid crash in FunctionDAG.h | 08 August 2019, 17:14:26 UTC |
| 02e43c6 | Steven Johnson | 08 August 2019, 17:08:28 UTC | Merge pull request #4063 from halide/srj-autosched-test Add apps/autoscheduler to test_apps target | 08 August 2019, 17:08:28 UTC |
| 1606bac | Steven Johnson | 08 August 2019, 03:31:49 UTC | Merge pull request #4109 from halide/srj-incl Add missing includes in apps/autoscheduler | 08 August 2019, 03:31:49 UTC |
| 2255a92 | Steven Johnson | 08 August 2019, 01:34:18 UTC | Merge pull request #4105 from halide/introspection_64_bit_offsets Fix for llvm trunk | 08 August 2019, 01:34:18 UTC |
| e2b2524 | Steven Johnson | 08 August 2019, 00:45:16 UTC | Add missing includes in apps/autoscheduler also, drive-by sorting of includes | 08 August 2019, 00:45:16 UTC |
| bfe3471 | Steven Johnson | 07 August 2019, 22:38:38 UTC | Merge branch 'introspection_64_bit_offsets' into srj-autosched-test | 07 August 2019, 22:38:38 UTC |
| 2bcfba3 | Andrew Adams | 07 August 2019, 22:28:57 UTC | Merge remote-tracking branch 'origin/introspection_64_bit_offsets' into bfloat16 | 07 August 2019, 22:28:57 UTC |
| a785848 | Andrew Adams | 07 August 2019, 22:27:31 UTC | Use better typedef name from #4107 | 07 August 2019, 22:27:31 UTC |
| 0468fc7 | Andrew Adams | 07 August 2019, 22:15:19 UTC | Wrappers need distinct names | 07 August 2019, 22:15:19 UTC |
| ae74331 | Andrew Adams | 07 August 2019, 22:06:57 UTC | Fix for llvm trunk | 07 August 2019, 22:06:57 UTC |
| d7126d8 | Steven Johnson | 07 August 2019, 21:51:07 UTC | Update autoscheduler.inc | 07 August 2019, 21:51:07 UTC |
| 9616d2f | Andrew Adams | 07 August 2019, 21:46:00 UTC | Fix incorrect type on prefetch call | 07 August 2019, 21:46:00 UTC |
| dec777c | Benoit Steiner | 07 August 2019, 18:47:23 UTC | Merge pull request #4097 from benoitsteiner/master Improved onnx support | 07 August 2019, 18:47:23 UTC |
| f8af1cb | Andrew Adams | 07 August 2019, 18:44:58 UTC | Enable half types for cuda. Add half test for cuda/metal. | 07 August 2019, 18:44:58 UTC |
| 31bb617 | Steven Johnson | 07 August 2019, 18:41:37 UTC | Add missing includes | 07 August 2019, 18:41:37 UTC |
| a33cd9b | Steven Johnson | 07 August 2019, 18:39:54 UTC | Merge branch 'master' into srj-autosched-test | 07 August 2019, 18:39:54 UTC |
| defb6ad | Steven Johnson | 07 August 2019, 18:37:13 UTC | Merge pull request #4099 from halide/srj-schar Add `signed char` handling to RunGen.h (Issue #3909) | 07 August 2019, 18:37:13 UTC |
| 32bcdef | Andrew Adams | 07 August 2019, 18:15:20 UTC | Delete dead code | 07 August 2019, 18:15:20 UTC |
| b2e08df | Andrew Adams | 07 August 2019, 18:13:16 UTC | Delete dead code | 07 August 2019, 18:13:16 UTC |
| 81b02fa | Andrew Adams | 07 August 2019, 18:12:30 UTC | Merge pull request #3678 from halide/cuda_sticky_allocator Cuda sticky allocator | 07 August 2019, 18:12:30 UTC |
| f82e7c4 | Andrew Adams | 07 August 2019, 18:11:09 UTC | Handle float16 in codegen | 07 August 2019, 18:11:09 UTC |
| c8933e7 | Tiago De Gaspari | 07 August 2019, 12:51:02 UTC | Fix comment on FFT App Generator file Fix the comment that documents the direction of the FFT on generator file. | 07 August 2019, 12:51:02 UTC |
