Граф коммитов

373 Коммитов

Автор SHA1 Сообщение Дата
Yuwei HU b5c6b99314 update depthwise convolution api (#344) 2017-08-17 21:35:41 -07:00
Haichen Shen 8ca12d8708 Add tutorial for convolution in CUDA (#343) 2017-08-17 15:11:38 -07:00
Tianqi Chen d2a98a059c [DOC] Add link to release blog (#342) 2017-08-17 12:55:30 -07:00
Tianqi Chen 8b247607ad [SUBMODULE] switch to https (#341) 2017-08-17 11:51:31 -07:00
Tianqi Chen 7ffcff2d67 [DOC] Release note (#340) 2017-08-17 11:42:14 -07:00
William Moses ae6abe827e Fix CUDA library search (#339) 2017-08-17 09:06:42 -07:00
William Moses 48ec5445c3 Allow install-dev to include all necessary header files (#338) 2017-08-17 08:29:25 -07:00
Tianqi Chen 090468aa53 [PASS] RewriteUnsafeSelect lowers unsafe select to condition expr (#335) 2017-08-15 22:32:43 -07:00
ziheng 25ded693ec [NNPack] Support for threadpool (#334)
* [NNPack] Support for threadpool

* fix lint

* fix lint

* Use static class function
2017-08-15 19:39:48 -07:00
wetliu 989e99e693 [WIP] [TOPI] Depth wise Conv for NHWC (#325)
* rename the nchw and pass the unit test; going to do it for nhwc depthwise

* bug with fusion

* nchw works fine; nhwc float32 problem remains

* still cannot bind them together

* fusion works

* syntax fix

* all bugs fixed; test cases pass

* minor fix on nn.h
2017-08-15 17:10:05 -07:00
ziheng 64870ffb86 [Contrib] CuDNN v7 Support (#311)
* [Contrib] CuDNN v7 Support

* Add test
2017-08-15 12:37:10 -07:00
Tianqi Chen 0ccc281d16 [BUILD] Enable cudnn in gpu build (#333) 2017-08-15 11:42:09 -07:00
Tianqi Chen 7196c791cb [TOPI] Isolate padding option, improve decl of depthwise/conv2d/pool (#332) 2017-08-15 11:24:07 -07:00
Tianqi Chen abccd9cda4 [TOPI] Improve dilate (#330) 2017-08-15 09:18:56 -07:00
Leyuan Wang 9ac46beafa [TOPI] Fix conv2d for small input channels (#331)
* __init__ updated

* pull request updated

* build_module added

* typo fixed

* another typo fixed

* conv2d gpu scheduler for two layouts moved to tvm

* changes made according to CR

* conv2d_nchw formating updated, conv2d_hwcn tests updated

* lint error fixed

* element wise operator schedule fusing fixed for conv2d

* conv2d_nchw topi test added, all resnet workloads now pass

* conv compute lint error fixed

* fixed python 3 compatibility problem

* conv2d tensor input support added, test typo fixed, ir_pass.Simplify changed to util.get_const_int

* fixed channel numer < 4 error, also made sure other splitting factor woudn't be 0
2017-08-15 09:18:47 -07:00
ziheng 0ad590c00f [TOPI] Add ops compute (#323)
* [TOPI] Add ops compute

Remove 'compute' and add assert for safety

Add document

fix lint

fix softmax

* fix batch norm
2017-08-14 22:13:51 -07:00
Tianqi Chen ce18b56543 [DOC] Document update (#329) 2017-08-14 22:13:28 -07:00
Yuwei HU 07e56b9ad4 update depthwise_conv2d schedule and testing (#328) 2017-08-14 21:09:25 -07:00
Tianqi Chen 8edd047b6e [TOPI] Move ewise.h -> elemwise.h (#327)
* [TOPI] Move ewise.h -> elemwise.h

* fix test
2017-08-14 20:00:30 -07:00
Xingjian Shi 760475f9a7 [TOPI] Add broadcast and reduce operators (#267)
[TOPI] Add broadcast and reduce operators
2017-08-14 19:35:40 -07:00
Tianqi Chen a59774e361 [BUILD] Simplify build process (#326) 2017-08-14 19:35:09 -07:00
Nicolas Vasilache cbdd14f106 [TOPI] C++ doc (#320) 2017-08-14 15:23:20 -07:00
Yuwei HU b0c42f3b2e [TOPI] add dilation operators (#316)
* add dilation operators

* fix pylint

* dilate testcases success

* n-D tensor dilation

* support arbitrary dimension
2017-08-14 10:31:42 -07:00
Tianqi Chen ba6664a32a [DOC] Include TOPI in doxygen (#321)
* [DOC] Include TOPI in doxygen

* update
2017-08-14 10:28:36 -07:00
Leyuan Wang cbff637f53 [TOPI] conv2d nchw gpu scheduler (#315)
* __init__ updated

* pull request updated

* build_module added

* typo fixed

* another typo fixed

* conv2d gpu scheduler for two layouts moved to tvm

* changes made according to CR

* conv2d_nchw formating updated, conv2d_hwcn tests updated

* lint error fixed

* element wise operator schedule fusing fixed for conv2d

* conv2d_nchw topi test added, all resnet workloads now pass

* conv compute lint error fixed

* fixed python 3 compatibility problem

* conv2d tensor input support added, test typo fixed, ir_pass.Simplify changed to util.get_const_int
2017-08-13 20:57:37 -07:00
Tianqi Chen d76712d121 [TOPI] Move topi.nn.util to topi.util (#319)
* [TOPI] Move topi.nn.util to topi.util

* update the path
2017-08-13 20:22:51 -07:00
Nicolas Vasilache f08de2b690 [WIP] C++ topi contributions (#312)
* [WIP] C++ topi contributions

Summary:
This diff implements C++ topi contributions for:
  - relu with parametrix threshold
  - pad with generic padBefore / padAfter specification
  - matmult with transposes
  - conv2d_nchw, conv2d_hwcn with runtime constant padding and strides
  - depthwise_conv2d_nchw with runtime constant padding and strides
  - group_conv2d_ngchw with runtime constant padding and strides
  - broadcast_to a broadcastable shape
  - broadcast_bop where bop is an usual binary op (+ - * / %)

Convolution padding is implemented using the pad operation.
To avoid extra memory consumption, it is generally recommended to inline the padding with the autoinliner.
Unfortunately in its current form the elemwise checks are too restrictive to allow inlining.
So this diff also proposes an extension to LHS injective (i.e. no reduction axis in the current IR design)

Test Plan:
Tested in C++ testsuite in a separate repository, I am looking for suggestions to quickly spin up some tests for tvm.

Reviewers: tqchen

Subscribers:

Tasks:

Tags:

Blame Revision:

* Review + Lint + GSG C++
2017-08-13 15:50:25 -07:00
Tianqi Chen a3776ba5ae [PASS][PRAGMA] Allow pragma debug_skip_region to skip region of computation (#318) 2017-08-13 15:35:38 -07:00
Tianqi Chen 79e482bc5a [PASS] Memory barrier detection, storage access lower. (#317) 2017-08-13 11:35:50 -07:00
Tianqi Chen afa20869b4 [PASS] More improvement of canonical (#314) 2017-08-11 18:17:08 -07:00
Yuwei HU 3c2569a0ab minor fix (#313) 2017-08-11 12:48:54 -07:00
Tianqi Chen e4b500b608 [PASS][FIX] Fix LiftAttrScope with if (#309)
* [PASS][FIX] Fix LiftAttrScope with if

* [PASS] Fix on proc sync

* fix
2017-08-10 18:38:09 -07:00
Tianqi Chen 19381b51ff [RPC] Enable shutdown hook (#308) 2017-08-10 09:36:42 -07:00
Yizhi Liu 48fd9e461e [tvm4j] disable proxy test for now (#307) 2017-08-10 09:36:22 -07:00
Tianqi Chen 47eb85de3e [TEST] Upgrade gpu docker to cudnn7 (#306)
* [TEST] Upgrade gpu docker to cudnn7

* fx
2017-08-09 18:08:23 -07:00
Tianqi Chen 29338ea428 [PASS] Allow allocation in parallel scope (#305) 2017-08-09 11:21:50 -07:00
ziheng 11328f64f9 [FIX] Pass the attributes of master node (#304) 2017-08-09 08:32:37 -07:00
Tianqi Chen b40d43c479 [PASS][RUNTIME] Support attr scope lift and runonce (#303) 2017-08-08 15:21:37 -07:00
Yizhi Liu 7d67e473d6 [tvm4j] RPC Server (#268)
* [tvm4j] RPC Server

* [tvm4j] fix recursively function calling; connect to proxy server; osx rename .so to .dylib

* [tvm4j] test case for proxy connection; thread pool for serving
2017-08-08 10:23:17 -07:00
Tianqi Chen 1146495f3e [RUNTIME][PASS] Allow declare vector type array (#302)
* [RUNTIME][PASS] Allow declare vector type array

* fix bcast

* [BUFFER] Enable vload/store function in buffer

* ok
2017-08-07 17:51:33 -07:00
ziheng 1e48b02f8c [NNPACK] Add nnpack.convolution (#301)
* [NNPACK] Add nnpack.convolution

* Add instrinsic

* Fix lint
2017-08-07 08:29:37 -07:00
ziheng 1389d20888 [EXECUTOR] Split graph_executor to header file and (runtime) source file (#300)
* [EXECUTOR] Split graph_executor to header file and (runtime) source file

* Fix
2017-08-05 20:30:40 -07:00
Tianqi Chen 41768cf918 [SCHEDULE][RUNIME] Introduce pragma for additional extension hint, threadpool runtime. (#299) 2017-08-05 15:37:22 -07:00
Tianqi Chen fd96d28534 [PASS] More storage sync. (#297) 2017-08-03 22:24:51 -07:00
ziheng 581be16556 [PASS] Enhance LayoutTransform pass (#293)
* [PASS] Enhance LayoutTransform pass

* Fix

* Fix Compilation

* Refactor

* Refactor

* doc

* fix

* add file
2017-08-03 21:48:57 -07:00
Tianqi Chen eefcfe1985 [PASS] Refactor thread storage sync to a common visitor (#296)
* [PASS] Refactor thread storage sync to a common visitor

* Fix the sync scope check behavior
2017-08-03 13:21:49 -07:00
Tianqi Chen 6bc0ae12ca [ARITH] Refactor intset eval with functor (#295) 2017-08-02 23:28:28 -07:00
Edward Z. Yang 10bc2fdfe0 Conda build recipe (#288)
* Typofix.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Probe for nvrtc in lib directory as well.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Conda build recipe for TVM.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>
2017-08-01 15:23:56 -07:00
ziheng fe3b885755 [FIX] Fix doc_string of reducer (#292) 2017-08-01 10:49:59 -07:00
Tianqi Chen 989dda89e1 [PASS] Simplify dependency of StorageRewrite (#291) 2017-08-01 10:12:18 -07:00