05a487c | Steven Johnson | 21 December 2018, 23:03:00 UTC | Merge branch 'standalone_autoscheduler' into srj-auto-multi | 21 December 2018, 23:03:00 UTC |
0b459b5 | Steven Johnson | 21 December 2018, 21:33:40 UTC | Merge branch 'master' into standalone_autoscheduler | 21 December 2018, 21:33:40 UTC |
6b1c98d | Andrew Adams | 21 December 2018, 20:06:51 UTC | Merge pull request #3530 from benoitsteiner/standalone_autoscheduler Added a constructor for the pipeline feature class. I… | 21 December 2018, 20:06:51 UTC |
2c461d4 | Benoit Steiner | 21 December 2018, 19:50:14 UTC | ScheduleFeatures don't need a default constructor since each member is already zero initialized. | 21 December 2018, 19:50:14 UTC |
a83205a | Benoit Steiner | 21 December 2018, 19:09:11 UTC | Added a constructor for both schedule and pipeline feature classes. I'm getting some NaN when training a cost model for a long time. It's not clear yet if the NaNs come from the inputs or from the network itself, so I'm just making sure that the inputs are always initialized to rule this out. | 21 December 2018, 19:09:11 UTC |
c0c87c5 | Andrew Adams | 21 December 2018, 18:57:30 UTC | Merge pull request #3529 from benoitsteiner/standalone_autoscheduler Moved ScheduleFeatures and PipelineFeatures back into the Halide::Int… | 21 December 2018, 18:57:30 UTC |
10324a5 | Benoit Steiner | 21 December 2018, 18:16:36 UTC | Moved ScheduleFeatures and PipelineFeatures back into the Halide::Internal namespace | 21 December 2018, 18:16:36 UTC |
f61e648 | Steven Johnson | 21 December 2018, 18:11:05 UTC | Minor cleanup | 21 December 2018, 18:11:05 UTC |
7d2ee03 | Andrew Adams | 21 December 2018, 18:03:02 UTC | Merge pull request #3527 from benoitsteiner/standalone_autoscheduler Store all the schedule features using an int64. This ensures that the… | 21 December 2018, 18:03:02 UTC |
f5e7664 | Steven Johnson | 21 December 2018, 17:50:09 UTC | Merge remote-tracking branch 'origin/standalone_autoscheduler' into srj-auto-multi | 21 December 2018, 17:50:09 UTC |
0f1f3fd | Steven Johnson | 21 December 2018, 17:50:00 UTC | Merge branch 'master' into srj-auto-multi | 21 December 2018, 17:50:00 UTC |
0500f6f | Steven Johnson | 21 December 2018, 17:44:22 UTC | Merge pull request #3521 from halide/srj-regmulti Rework RunGen registration model | 21 December 2018, 17:44:22 UTC |
fce881b | Benoit Steiner | 21 December 2018, 17:40:17 UTC | Store all the schedule features using an int64. This ensures that the number of features can still be correctly computed by dividing sizeof(ScheduleFeatures) by sizeof(int64). | 21 December 2018, 17:40:17 UTC |
c24f924 | Andrew Adams | 21 December 2018, 00:42:26 UTC | Merge pull request #3525 from halide/abadams-patch-1 Update README.md | 21 December 2018, 00:42:26 UTC |
8de8386 | Andrew Adams | 21 December 2018, 00:37:37 UTC | Update README.md | 21 December 2018, 00:37:37 UTC |
ea64ad8 | Andrew Adams | 20 December 2018, 22:08:54 UTC | Clean up cost model | 20 December 2018, 22:08:54 UTC |
a1909fe | Andrew Adams | 20 December 2018, 21:52:21 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/halide/Halide into standalone_autoscheduler | 20 December 2018, 21:52:21 UTC |
996e0a7 | Andrew Adams | 20 December 2018, 21:52:19 UTC | Handle some corner cases | 20 December 2018, 21:52:19 UTC |
e61e987 | Andrew Adams | 20 December 2018, 21:51:00 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler | 20 December 2018, 21:51:00 UTC |
bc2a96d | Andrew Adams | 20 December 2018, 21:50:23 UTC | New weights | 20 December 2018, 21:50:23 UTC |
a4b1601 | Andrew Adams | 20 December 2018, 20:52:34 UTC | Merge pull request #3522 from LazyDodo/master Fix HalideTraceViz on windows. | 20 December 2018, 20:52:34 UTC |
d1ee266 | Steven Johnson | 20 December 2018, 19:55:54 UTC | Add extra 'using', because google | 20 December 2018, 19:55:54 UTC |
5a6459d | Ray Molenkamp | 20 December 2018, 19:48:22 UTC | fix whitespace | 20 December 2018, 19:48:22 UTC |
b7ddb59 | Ray Molenkamp | 20 December 2018, 19:46:36 UTC | Fix HalideTraceViz on windows. stdin and stdout are by default opened as text and need switching to binary before reading/writing binary data. | 20 December 2018, 19:46:36 UTC |
a49ba81 | Steven Johnson | 20 December 2018, 18:35:43 UTC | Removed unused var | 20 December 2018, 18:35:43 UTC |
b809460 | Steven Johnson | 20 December 2018, 17:52:02 UTC | Merge branch 'standalone_autoscheduler' into srj-auto-multi | 20 December 2018, 17:52:02 UTC |
3c6b5cf | Steven Johnson | 20 December 2018, 17:51:53 UTC | Merge branch 'srj-regmulti' into srj-auto-multi | 20 December 2018, 17:51:53 UTC |
d25cd87 | Steven Johnson | 20 December 2018, 17:51:37 UTC | Fix | 20 December 2018, 17:51:37 UTC |
c82514a | Andrew Adams | 20 December 2018, 17:18:14 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/halide/Halide into standalone_autoscheduler | 20 December 2018, 17:18:14 UTC |
459256c | Andrew Adams | 20 December 2018, 17:18:07 UTC | Tweak cost model | 20 December 2018, 17:18:07 UTC |
a82d590 | Andrew Adams | 20 December 2018, 02:46:35 UTC | Merge pull request #3520 from halide/srj-getenv Don't crash when env vars aren't set | 20 December 2018, 02:46:35 UTC |
56e234c | Steven Johnson | 20 December 2018, 02:12:48 UTC | Don't crash when env vars aren't set creating a std::string from a NULL ptr tends to be crashy | 20 December 2018, 02:12:48 UTC |
dff852b | Steven Johnson | 20 December 2018, 01:52:59 UTC | WIP | 20 December 2018, 01:52:59 UTC |
ea74cf2 | Steven Johnson | 20 December 2018, 01:20:24 UTC | Merge branch 'srj-regmulti' into srj-auto-multi | 20 December 2018, 01:20:24 UTC |
168b325 | Steven Johnson | 20 December 2018, 01:19:40 UTC | Merge branch 'master' into standalone_autoscheduler | 20 December 2018, 01:19:40 UTC |
640dad7 | Steven Johnson | 20 December 2018, 01:18:29 UTC | Change link order | 20 December 2018, 01:18:29 UTC |
f2e5213 | Steven Johnson | 20 December 2018, 01:12:21 UTC | Tweak CMake rules | 20 December 2018, 01:12:21 UTC |
9203b1a | Steven Johnson | 20 December 2018, 01:03:03 UTC | Rework RunGen registration model Expand to allow for multiple filters to link against RunGen at once, selecting via a flag; this is of limited use for routine purposes, but should give us greater flexibility when experimenting with multiple auto-scheduler models. | 20 December 2018, 01:03:03 UTC |
60ed6c4 | Andrew Adams | 20 December 2018, 00:03:46 UTC | Throw a sigmoid around the pipeline weights. Seems to prevent overfitting | 20 December 2018, 00:03:46 UTC |
ad5bb1f | Andrew Adams | 19 December 2018, 23:44:03 UTC | Comment out robust fitting hack | 19 December 2018, 23:44:03 UTC |
557599e | Andrew Adams | 19 December 2018, 23:38:40 UTC | Use a hold-out set in the cost model trainer | 19 December 2018, 23:38:40 UTC |
43689f7 | Andrew Adams | 19 December 2018, 22:18:21 UTC | Isolate use of environment variables | 19 December 2018, 22:18:21 UTC |
33ee4fc | Andrew Adams | 19 December 2018, 22:02:12 UTC | Virtualize the cost model | 19 December 2018, 22:02:12 UTC |
0602b74 | Steven Johnson | 19 December 2018, 21:52:23 UTC | Merge pull request #3517 from minrk/python-cxxflags add $CXXFLAGS to $CCFLAGS in Python bindings | 19 December 2018, 21:52:23 UTC |
bf16b0a | Andrew Adams | 19 December 2018, 21:40:45 UTC | Crudely slice things up into multiple files | 19 December 2018, 21:40:45 UTC |
cc1454c | Andrew Adams | 19 December 2018, 20:21:17 UTC | Can't move a std::ostringstream | 19 December 2018, 20:21:17 UTC |
dc7e5b7 | Andrew Adams | 19 December 2018, 19:59:37 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/halide/Halide into standalone_autoscheduler | 19 December 2018, 19:59:37 UTC |
683e13e | Andrew Adams | 19 December 2018, 19:59:14 UTC | Turn sliding back on | 19 December 2018, 19:59:14 UTC |
57d72e1 | Min RK | 19 December 2018, 14:36:57 UTC | respect $CXXFLAGS in Python bindings | 19 December 2018, 14:36:57 UTC |
8e86529 | Andrew Adams | 19 December 2018, 06:09:23 UTC | Randomly initialize weights if weights are missing | 19 December 2018, 06:09:23 UTC |
1f86716 | Andrew Adams | 19 December 2018, 06:04:25 UTC | Don't truncate the weights before reading them | 19 December 2018, 06:04:25 UTC |
4d15db4 | Andrew Adams | 19 December 2018, 03:59:31 UTC | Remove a failure assert | 19 December 2018, 03:59:31 UTC |
6e7f859 | Andrew Adams | 19 December 2018, 03:38:43 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler | 19 December 2018, 03:38:43 UTC |
5fd7faf | Andrew Adams | 19 December 2018, 03:36:58 UTC | Merge pull request #3516 from halide/srj-autosched-2 Add error checking for the ifstream / ofstream calls in autoscheduler | 19 December 2018, 03:36:58 UTC |
a9510df | Andrew Adams | 19 December 2018, 03:35:20 UTC | Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler | 19 December 2018, 03:35:20 UTC |
78f3069 | Andrew Adams | 19 December 2018, 03:35:09 UTC | Don't print a worst schedule if there isn't one | 19 December 2018, 03:35:09 UTC |
c6c4d71 | Andrew Adams | 19 December 2018, 03:34:43 UTC | Fix autotune makefile | 19 December 2018, 03:34:43 UTC |
c879542 | Andrew Adams | 19 December 2018, 03:33:50 UTC | Add more load cost terms | 19 December 2018, 03:33:50 UTC |
d569aad | Steven Johnson | 19 December 2018, 02:18:23 UTC | Add error checking for the ifstream / ofstream calls in autoscheduler (Note that buffer_from_file() in TPP.h is commented out, as some of the weights/ files are the wrong size. Not sure of the right fix there; we should uncomment once the file is fixed.) | 19 December 2018, 02:18:23 UTC |
9867b14 | Andrew Adams | 19 December 2018, 01:17:27 UTC | Merge pull request #3515 from halide/srj-autosched-nuke Remove apparently-dead code | 19 December 2018, 01:17:27 UTC |
6912cb4 | Steven Johnson | 19 December 2018, 01:10:29 UTC | Remove apparently-dead code This code appears to be totally unused -- if it is, let's remove it to avoid confusion (we can always retrieve it from git history if we need it). If it *isn't* unused, then let's add some comment about how/why that is the case. | 19 December 2018, 01:10:29 UTC |
25f6802 | Andrew Adams | 19 December 2018, 00:08:38 UTC | Correct scalar_loads_per_scalar features | 19 December 2018, 00:08:38 UTC |
8a1cc09 | Steven Johnson | 18 December 2018, 23:35:08 UTC | Merge pull request #3512 from halide/srj-bufconst Improve Buffer conversion checks when assertions are disabled | 18 December 2018, 23:35:08 UTC |
efe3a07 | Andrew Adams | 18 December 2018, 20:19:19 UTC | Track sites per stage, not per func | 18 December 2018, 20:19:19 UTC |
172078f | Steven Johnson | 18 December 2018, 19:28:56 UTC | Improve Buffer conversion checks when assertions are disabled assert_can_convert_from() was a complete no-op for builds that disable assertions, but there are meaningful compile-time checks that can (and should) be done even in these situations. | 18 December 2018, 19:28:56 UTC |
cd78609 | Andrew Adams | 18 December 2018, 19:28:00 UTC | Fix plugin filename | 18 December 2018, 19:28:00 UTC |
b497cf1 | Andrew Adams | 18 December 2018, 19:03:41 UTC | Enable tests | 18 December 2018, 19:03:41 UTC |
062b6b3 | Andrew Adams | 18 December 2018, 19:03:30 UTC | Remove wrappers from conv+relu | 18 December 2018, 19:03:30 UTC |
cbf16e3 | Andrew Adams | 18 December 2018, 18:59:45 UTC | Merge remote-tracking branch 'origin/standalone_autoscheduler' into add_vectorization_to_search_space | 18 December 2018, 18:59:45 UTC |
c02ad0c | Andrew Adams | 18 December 2018, 18:26:09 UTC | Turn off llvm optimization | 18 December 2018, 18:26:09 UTC |
d056db8 | Andrew Adams | 18 December 2018, 18:25:52 UTC | Add some pruning | 18 December 2018, 18:25:52 UTC |
8c80ea6 | Andrew Adams | 18 December 2018, 17:34:43 UTC | Merge branch 'add_vectorization_to_search_space' of https://github.com/halide/Halide into add_vectorization_to_search_space | 18 December 2018, 17:34:43 UTC |
abd4e14 | Andrew Adams | 18 December 2018, 16:51:41 UTC | Save weights in the autotuner loop | 18 December 2018, 16:51:41 UTC |
74260b3 | Andrew Adams | 18 December 2018, 16:49:52 UTC | Remove possible sources of nan | 18 December 2018, 16:49:52 UTC |
1820007 | Andrew Adams | 18 December 2018, 16:39:44 UTC | Fix innermost_bytes_at_task computation | 18 December 2018, 16:39:44 UTC |
815e30c | Andrew Adams | 18 December 2018, 05:04:38 UTC | Merge branch 'add_vectorization_to_search_space' of https://github.com/halide/Halide into add_vectorization_to_search_space | 18 December 2018, 05:04:38 UTC |
4b7ade2 | Andrew Adams | 18 December 2018, 05:04:35 UTC | Delete schedules that survived due to randomness | 18 December 2018, 05:04:35 UTC |
8839162 | Andrew Adams | 18 December 2018, 04:53:13 UTC | Update scripts | 18 December 2018, 04:53:13 UTC |
36eb74b | Andrew Adams | 18 December 2018, 04:52:04 UTC | Switch to L2 on throughput | 18 December 2018, 04:52:04 UTC |
d193a20 | Andrew Adams | 18 December 2018, 04:50:49 UTC | Train in terms of batches, not epochs | 18 December 2018, 04:50:49 UTC |
a291f62 | Andrew Adams | 18 December 2018, 04:49:52 UTC | Better handling of parallelism and vectorization | 18 December 2018, 04:49:52 UTC |
e4a18a4 | Steven Johnson | 18 December 2018, 01:20:11 UTC | Merge branch 'master' into standalone_autoscheduler | 18 December 2018, 01:20:11 UTC |
64b6fc3 | Andrew Adams | 18 December 2018, 00:48:37 UTC | Reduce timeout on watchdog_bench | 18 December 2018, 00:48:37 UTC |
075294c | Andrew Adams | 17 December 2018, 23:34:13 UTC | Expand featurization, and account for page fault behavior | 17 December 2018, 23:34:13 UTC |
c3c34a0 | Andrew Adams | 17 December 2018, 23:29:32 UTC | Merge pull request #3505 from halide/bug_device_free Fix an issue that device frees are not injected. | 17 December 2018, 23:29:32 UTC |
76a8c27 | Jing Pu | 17 December 2018, 21:46:11 UTC | force a builder rerun. | 17 December 2018, 21:46:11 UTC |
3889b5b | Andrew Adams | 17 December 2018, 00:12:02 UTC | Temporarily turn off all sliding | 17 December 2018, 00:12:02 UTC |
7fd86fc | Andrew Adams | 16 December 2018, 23:09:32 UTC | Report worst failure | 16 December 2018, 23:09:32 UTC |
7b975c0 | Andrew Adams | 16 December 2018, 23:09:12 UTC | Tweak cost model | 16 December 2018, 23:09:12 UTC |
0b4e930 | Andrew Adams | 16 December 2018, 23:06:31 UTC | Also tile update stages | 16 December 2018, 23:06:31 UTC |
4519897 | Andrew Adams | 16 December 2018, 01:02:51 UTC | Fix bug with excessive inner loop parallelism | 16 December 2018, 01:02:51 UTC |
6339015 | Andrew Adams | 16 December 2018, 00:57:14 UTC | Update cost model to include false sharing penalties | 16 December 2018, 00:57:14 UTC |
a002ed1 | Andrew Adams | 15 December 2018, 00:39:46 UTC | Just switch back to L2 | 15 December 2018, 00:39:46 UTC |
0c9abb6 | Andrew Adams | 15 December 2018, 00:20:33 UTC | Fix input load detection | 15 December 2018, 00:20:33 UTC |
32317f4 | Andrew Adams | 15 December 2018, 00:20:17 UTC | Don't penalize memory bandwidth for idle cores | 15 December 2018, 00:20:17 UTC |
2cf9da3 | Andrew Adams | 14 December 2018, 23:03:33 UTC | Tweak random pipeline scripts | 14 December 2018, 23:03:33 UTC |
7a22427 | Andrew Adams | 14 December 2018, 23:02:50 UTC | Tweak cost model and lowering | 14 December 2018, 23:02:50 UTC |
3f26cbc | Jing Pu | 14 December 2018, 21:46:57 UTC | Merge branch 'master' into bug_device_free | 14 December 2018, 21:46:57 UTC |
d46b547 | Steven Johnson | 14 December 2018, 21:11:02 UTC | Merge pull request #3502 from halide/fix_rtti_flag Fix no-rtti flag detection | 14 December 2018, 21:11:02 UTC |
2bb0577 | Andrew Adams | 14 December 2018, 17:31:52 UTC | Merge pull request #3500 from halide/fix_sgemm_schedule Reschedule test/performance/matrix_multiplication | 14 December 2018, 17:31:52 UTC |