https://github.com/halide/Halide

sort by:
Revision Author Date Message Commit Date
53435b9 Add autotune_classic.sh and support This is a trivial script to try running 'classic' autoscheduler with the multiple variants of MachineParams used in the original paper; all it does is disable benchmarking of non-classic variants, and run 'make test' in the current directory repeatedly with different values (removing any classic artifacts from bin to ensure a rebuild occurs). 10 January 2019, 21:05:59 UTC
0305d46 Force-inline pointwise stages before boundary conditions 10 January 2019, 19:49:43 UTC
0eed105 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 10 January 2019, 18:14:11 UTC
4752f41 Fresh weights 10 January 2019, 18:13:57 UTC
b6b7b28 Remove explicit setting of HL_MACHINE_PARAMS 10 January 2019, 09:08:03 UTC
24e0f8f Physical cores instead of logical cores 10 January 2019, 09:04:37 UTC
2b48014 Fix erroneous comparison 10 January 2019, 09:03:24 UTC
b82ebe4 Extra simplification pass before unrolling 10 January 2019, 05:52:58 UTC
b52bf57 Make it possible to autoschedule the cost model itself 10 January 2019, 05:52:41 UTC
ad2a0cd Use proper variable to point to weights dir in resnet makefile 10 January 2019, 05:51:36 UTC
e537ca4 Increase benchmarking time in random pipeline generator 10 January 2019, 05:51:03 UTC
49fab26 Fix possible factorization simplification in random pipeline generator 10 January 2019, 05:50:45 UTC
7570aea Use explicit bounds in camera pipe to make autotuner work 10 January 2019, 05:49:20 UTC
9084f2c Use more consistent benchmarking methodology 10 January 2019, 05:49:06 UTC
24db4f2 Always use large buffers for lens blur 10 January 2019, 05:47:25 UTC
e5f3d23 Use explicit bounds in resnet 10 January 2019, 05:46:26 UTC
3eb429e Add disabled code to turn estimates into bound directives 10 January 2019, 05:45:42 UTC
92c4996 autotune_loop does one iteration. Run it in a loop if you want more 10 January 2019, 05:45:21 UTC
99d67f9 Change the meaning of dropout to be normalized across the number of decisions to be made 10 January 2019, 05:45:03 UTC
a0fa880 Fix wrong estimate on harris 10 January 2019, 03:36:11 UTC
85484e7 remove oopsie 10 January 2019, 00:22:09 UTC
c18a851 Add apps/fit_and_slice_3x4 10 January 2019, 00:21:36 UTC
ea61b9a nuke useless files 09 January 2019, 23:43:02 UTC
0364432 Merge branch 'standalone_autoscheduler' of https://github.com/halide/Halide into standalone_autoscheduler 09 January 2019, 23:38:07 UTC
2252bf1 Add apps/iir_blur_generator This is basically a clone of the IIR Blur generator from apps/HelloMatlab, with appropriate machinery to do autoscheduler testing and tuning. 09 January 2019, 23:37:47 UTC
c69b33a Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 09 January 2019, 23:32:50 UTC
77c8965 Fresh weights 09 January 2019, 23:32:43 UTC
e48909e Make resnet parallel 09 January 2019, 23:25:43 UTC
ad77bd9 Fix namespace nonsense 09 January 2019, 23:16:24 UTC
74b16d8 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 09 January 2019, 23:01:15 UTC
9d96eba Inline things that should trivially be inlined 09 January 2019, 23:00:30 UTC
31c207f Merge pull request #3567 from halide/srj-resnet Add classic auto-schedule testing to apps/resnet 09 January 2019, 22:59:32 UTC
b73863a Add classic auto-schedule testing to apps/resnet Note that (unlike most other apps) this doesn't attempt to run all three in the same run; it selects one of the options via command-line flag. 'make test' runs all three in sequence. Note also that the 'manual' schedule is known to be awful (it's just enough compute-roots added to make the runtime tractable); improvement is left as an exercise to the reader. 09 January 2019, 22:52:58 UTC
934281c Merge pull request #3566 from benoitsteiner/standalone_autoscheduler autoscheduler improvements 09 January 2019, 22:27:34 UTC
0835680 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 09 January 2019, 22:25:24 UTC
335f2f6 Fresh weights 09 January 2019, 22:25:12 UTC
2ba4544 Fixed the spelling of halide 09 January 2019, 21:36:21 UTC
f5a8c10 Added an API to be able to call the autoscheduler code from c++ 09 January 2019, 21:35:54 UTC
872b556 Track versions of the featurization. This simplifies the lifecycle management features extracted ahead of time, for example in order to build a training set for the cost model. 09 January 2019, 21:35:09 UTC
bc0df86 Code cleanup Added the definition of the new autoscheduler API 09 January 2019, 21:35:09 UTC
57f9320 Moved the cost model and network size code into the halide namespace 09 January 2019, 21:35:09 UTC
a6b0d07 Added an API to be able to call the autoscheduler code from c++ 09 January 2019, 21:34:36 UTC
f7c5c85 Change default beam size 09 January 2019, 20:26:33 UTC
2dce11d Fresh weights 09 January 2019, 20:25:34 UTC
7018c14 Network size to match latest weights 09 January 2019, 16:05:41 UTC
be98fd7 Fresh weights 09 January 2019, 16:03:56 UTC
b13af96 Fresh weights 09 January 2019, 05:05:36 UTC
7a17a95 Make cost model training faster 09 January 2019, 03:54:46 UTC
5205159 New schedule for cost model 09 January 2019, 02:29:44 UTC
f8548a0 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 09 January 2019, 02:09:22 UTC
9e9e416 Fresh weights 09 January 2019, 02:09:15 UTC
7f1b30a Merge branch 'master' into standalone_autoscheduler 09 January 2019, 01:48:27 UTC
402831d Merge pull request #3564 from halide/srj-llvm-fix Update for trunk LLVM 09 January 2019, 01:47:43 UTC
e312ad1 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 09 January 2019, 01:16:08 UTC
9fd8c76 Try 24 hidden nodes 09 January 2019, 01:16:01 UTC
dc713bc Merge pull request #3560 from halide/srj-resnet Resnet50 cleanups 09 January 2019, 01:04:22 UTC
edd98b8 Merge branch 'standalone_autoscheduler' into srj-resnet 09 January 2019, 01:03:35 UTC
8d0db32 Try 24 hidden nodes 09 January 2019, 00:51:15 UTC
a9592e9 Merge pull request #3563 from halide/srj-resnet2 Add 'make autotune' to apps/resnet50 09 January 2019, 00:48:11 UTC
04cc5f7 Merge branch 'srj-resnet' into srj-resnet2 09 January 2019, 00:21:21 UTC
78b3ef4 Constrain to 224x224x3 09 January 2019, 00:04:38 UTC
cf6c438 Fresh weights 08 January 2019, 23:43:02 UTC
6e66b66 Update for trunk LLVM 08 January 2019, 23:41:50 UTC
f0ab2f1 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 08 January 2019, 23:29:56 UTC
0faadcf Fresh weights 08 January 2019, 23:29:50 UTC
6ab15d7 Remove scalpel left in patient 08 January 2019, 23:24:51 UTC
fca97ed Remove dangling comment 08 January 2019, 23:16:38 UTC
efe5916 Merge branch 'srj-resnet' into srj-resnet2 08 January 2019, 23:15:00 UTC
ca62966 Merge branch 'standalone_autoscheduler' into srj-resnet 08 January 2019, 23:14:40 UTC
9965b75 Add 'make autotune' to apps/resnet50 08 January 2019, 23:14:29 UTC
db857d8 Merge branch 'standalone_autoscheduler' into srj-resnet2 08 January 2019, 23:11:30 UTC
2e02b4b remove rogue line from Makefile 08 January 2019, 23:08:49 UTC
a1c6bd6 Fresh weights 08 January 2019, 22:56:14 UTC
5dcfef0 Try L2 on throughput again 08 January 2019, 22:42:11 UTC
11c5dbd Cleanup 08 January 2019, 22:11:36 UTC
1526c25 Fix image loading 08 January 2019, 21:22:15 UTC
3af18d4 Resnet50 cleanups Some cleanups to smooth the way for adding autotuning: -- remove unused generator and cmake file -- parameterize load_weights and process to be more configurable from the command line as to files, directories, etc -- pytorch weights are now downloaded as part of the make process and stored in a bin subfolder, so 'make test' should work as-is (assuming pytorch is installed) -- misc cleanup in process.cpp, including a halfhearted attempt to make loading nonprocessed images work. (doesn't work yet, left in place with a TODO for future investigation.) 08 January 2019, 20:34:51 UTC
8dc813d Try a 1,4 setup 08 January 2019, 20:33:16 UTC
b773f18 Fresh weights 08 January 2019, 20:06:18 UTC
80847bf Weights for shrunken model 08 January 2019, 19:52:24 UTC
c3d6c6c Aggressively ablate model 08 January 2019, 19:51:44 UTC
22ce212 Fresh weights 08 January 2019, 18:58:17 UTC
6acda50 Fresh weights 08 January 2019, 18:37:03 UTC
93bb008 Fresh weights 08 January 2019, 17:57:59 UTC
7999956 Network size to match latest weights 08 January 2019, 17:55:54 UTC
d9fffd8 Fresh weights 08 January 2019, 17:55:38 UTC
1dac40f Fresh weights 08 January 2019, 17:41:11 UTC
7d07aae Fresh weights 08 January 2019, 17:28:06 UTC
28afc47 Fresh weights 08 January 2019, 16:42:32 UTC
02753c0 Add network size change to match fresh weights 08 January 2019, 07:09:27 UTC
5d2466f Fresh weights 08 January 2019, 07:08:47 UTC
7224944 Fresh weights 08 January 2019, 06:38:02 UTC
ae7b835 Don't double-bill for off-core loads 08 January 2019, 05:35:46 UTC
c91b755 More stable learning rate for autotuning 08 January 2019, 05:35:04 UTC
119ec0a Fresh weights 08 January 2019, 03:51:49 UTC
d84f312 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 08 January 2019, 02:05:26 UTC
e055978 Fresh weights 08 January 2019, 02:05:23 UTC
50eed0d Fix bug in parallel tile selection 08 January 2019, 01:38:59 UTC
0d0702b Fresh weights 08 January 2019, 01:10:03 UTC
0e7bd35 Merge branch 'standalone_autoscheduler' of https://github.com/Halide/Halide into standalone_autoscheduler 07 January 2019, 22:47:17 UTC
back to top