Porting schedules (except convolutions) to C++ #763

alex-weaver · 2018-01-08T17:51:26Z

Operators

Generic schedules

injective
_default_schedule
extern

CUDA schedules

Other schedules

x86: binarize_pack, binary_dense
rocm: dense
x86: injective, default from [TOPI] Basic x86 schedules #775

joyalbin · 2018-01-09T10:18:21Z

How about "power" operator in elementwise category?

alex-weaver · 2018-01-09T10:20:51Z

power would definitely be useful. Would it make sense to define two?

pow(Tensor, Expr)
pow(Tensor, Tensor) - where this one follows broadcasting rules

joyalbin · 2018-01-09T14:02:14Z

yes, two versions of pow make sense.
I am working on implementation of basic version first, will post the progress here later.

Fixed issue where TVM_DECLARE_INTRIN_UNARY used the PureExtern flag instead of PureIntrinsic. Added softmax CUDA schedule.

global_pool and pool ops. Extended pad to allow specifying pad_value. Fixed issue where pad would throw if padding was zero in all dimensions.

…s and schedules. Added test of broadcast ops

…ixed corner case in default schedules of attempting to fuse 0 length axes. Added tests for reduce ops.

…dded split_sections op to cover the other mode of the python split op. Added tests.

tqchen · 2018-01-22T18:02:50Z

topi/include/topi/reduction.h

+Tensor CommReduceIdx(const Tensor& data,
+                     const std::vector<int>& axis,
+                     FCommReduce func,
+                     bool keepdims = false) {


any update wrt this comment?

tqchen · 2018-01-22T18:04:45Z

topi/include/topi/nn/dilate.h

+* \return The output tensor.
+*/
+inline Tensor dilate(const Tensor& x,
+                     std::vector<int> strides,


Change most of the std::vector arguments to Array

How should an array of ints be represented? Should it be an Array<Expr> filled with constants? (Array<int> fails to compile)

Array of note that strides can be passed in as symbolic expressions, although not frequently used

Sorry array of what? Github may have broken your comment. Are you meaning Array of Expr?

I see that strides could be expressions, but nn/dilate.h:69 relies on being able to evaluate strides[i]. This and surrounding lines are:

if (strides[i] != 1) { index_tuple.push_back(indices[i] / strides[i]); not_zero.push_back((indices[i] % strides[i]) == 0); } else { index_tuple.push_back(indices[i]); }

Would a correct implementation of this with changing the signature to Array<Expr> strides be this?

if (IsConstInt(strides[i]) && GetConstInt(strides[i]) == 1) { index_tuple.push_back(indices[i]); } else { index_tuple.push_back(indices[i] / strides[i]); not_zero.push_back((indices[i] % strides[i]) == 0); }

yes, let us do that

tqchen · 2018-01-26T03:44:24Z

topi/include/topi/detail/extern.h

+using FExtern = std::function<Expr(Array<Buffer>, Array<Buffer>)>;
+
+/*! \brief Create tensors representing the result of invoking an external function */
+Array<Tensor> make_extern(const Array<Array<Expr>>& out_shapes,


>> -> > > note the space, for backward compatibility for MSVC

tqchen · 2018-01-26T03:46:35Z

topi/include/topi/detail/extern.h

+}
+
+/*! \brief Pack a buffer object to be used as an argument to a PackedFunc */
+Expr pack_buffer(Buffer buf) {


this belongs to a detail namespace. Add more comments to argument and return value.

This functions is used to create a DLTensor structure on heap to be able to pass a symbolic buffer as arguments to TVM PackedFunc

tqchen · 2018-01-26T03:46:48Z

topi/include/topi/detail/fuse.h

+namespace topi {
+using namespace tvm;
+
+/*! \brief Fuse all of the given args */


document all arguments

tqchen · 2018-01-26T03:47:10Z

topi/include/topi/detail/pad_utils.h

+namespace topi {
+using namespace tvm;
+
+/*! \brief Get padding size for each side given padding height and width */


document all arguments, and what is expected in return

tqchen · 2018-01-26T03:51:57Z

Thanks for all the effort to bring this set of changes! I like the structure overall and the code looks good as I skim through.

Since there are a lot of changes. Please take a pass over the code to make sure we have good readability of the code, put in comments delibrately in places that you think requires explaination so it is easier for others to understand and improve upon this.

alex-weaver · 2018-01-26T23:55:57Z

Ok I've fixed up the comments on some of the more obscure bits of the code. Let me know if there's anything else that needs sorting. I'd like to get some kind of schedule registry working like there is in the python code but I think that's best left for a separate PR ;)

tqchen · 2018-01-28T05:57:19Z

Thanks! this is merged

* Ported injective schedules to C++. Added some elementwise ops. * Fix lint errors * Added reduction ops and schedules * Fix lint errors * Fix lint errors * Fix lint errors * Added transform ops * Fix lint errors * Fix lint errors * Added softmax, log_softmax, leaky_relu and flatten ops. Fixed issue where TVM_DECLARE_INTRIN_UNARY used the PureExtern flag instead of PureIntrinsic. Added softmax CUDA schedule. * Fix lint * Fix lint * Added binary_dense, batch_norm_inference, dense, dilate, scale_shift_*, global_pool and pool ops. Extended pad to allow specifying pad_value. Fixed issue where pad would throw if padding was zero in all dimensions. * Fix lint * Fix lint * Added CUDA schedules for dense, pool and global_pool * Added extern schedules for generic and CUDA * Fix lint * Added x86 binary schedules * Fix lint * Added rocm dense schedule. Added rocBLAS and cuBLAS support to dense ops * Added pow ops. Added x86 default and injective schedules * Fix lint * Fix lint * Fix lint * Fix lint * Fix lint * Fix indent * Removed schedules directory * Changed left_shift, right_shift to operators. Changed pad_value in pad() to remove pointer usage * Fixed usage of pad in nn/pooling.h. Fixed declaration of operator>> * Fixed comments for shift operators * Added comments to utility functions * Added TOPI C++ library, exporting broadcast_add op * Fix lint * Share libinfo.py with TVM * Fix lint * Add other broadcast ops * Fix lint * Fix imports in topi * Fix lib names * Fixed build issue where windows builds don't apply correct definitions * Removed TVM_EXPORTS from topi library * Attempted CI build fix * Add topi lib to tvm_multilib * Fix Jenkinsfile * Added TOPI build target to Makefile * Fix nn op namespaces. * Fix lint * Renamed TOPI lib to libtvm_topi * Removed _ffi/base.py * Remove _ffi from topi, now shared with tvm. * Make libtvm_topi loading optional * Fix compiler warnings * Fix lint * Fix lint * Fix lint * Fix build error by making new libs argument to Target optional * Added C++ Target type interop. Added registration of remaining C++ ops and schedules. Added test of broadcast ops * Fix lint * Fix lint * Fix compile error * Fix compiler warnings * Fix compiler warnings * Fixed int vector interop. Fixed argmin incorrectly invoking argmax. Fixed corner case in default schedules of attempting to fuse 0 length axes. Added tests for reduce ops. * Refactored reduce builders * Fixed typos in topi.cc. Added basic test. * Fixed padding size error. Added dense, dilate, pooling tests * Fixed issue where clip would output a different dtype to the input. Added split_sections op to cover the other mode of the python split op. Added tests. * Changed extension type numbers to avoid clash with NNVM * Fix lint * Fix compiler warnings * Removed use of std::vector from the public TOPI API * Fix lint * Add TOPI C++ tests to CI * Fixed detail namespacing. Improved comments.

alex-weaver added 2 commits January 8, 2018 17:42

Ported injective schedules to C++. Added some elementwise ops.

0534f82

Fix lint errors

1ae471f

alex-weaver added 25 commits January 10, 2018 12:12

Added reduction ops and schedules

762bf43

Fix lint errors

fa126f6

Fix lint errors

08901a3

Fix lint errors

deb80ce

Added transform ops

a574ee6

Fix lint errors

6217d8c

Fix lint errors

8f9f77e

Added softmax, log_softmax, leaky_relu and flatten ops.

a5f35e9

Fixed issue where TVM_DECLARE_INTRIN_UNARY used the PureExtern flag instead of PureIntrinsic. Added softmax CUDA schedule.

Fix lint

66ad2a6

Fix lint

687b432

Added binary_dense, batch_norm_inference, dense, dilate, scale_shift_*,

9bcf5d5

global_pool and pool ops. Extended pad to allow specifying pad_value. Fixed issue where pad would throw if padding was zero in all dimensions.

Fix lint

0245df5

Fix lint

2db5040

Added CUDA schedules for dense, pool and global_pool

321dbb2

Added extern schedules for generic and CUDA

1b538a0

Fix lint

98e33a2

Added x86 binary schedules

1ff94d3

Fix lint

77c598a

Added rocm dense schedule. Added rocBLAS and cuBLAS support to dense ops

c7c39dc

Added pow ops. Added x86 default and injective schedules

f44cb7f

Fix lint

5a5e220

Fix lint

350a874

Fix lint

3558769

Fix lint

40cb234

Fix lint

14c7cdf

alex-weaver added 15 commits January 19, 2018 19:43

Fix build error by making new libs argument to Target optional

99b7477

Added C++ Target type interop. Added registration of remaining C++ op…

920bd65

…s and schedules. Added test of broadcast ops

Fix lint

3f255d0

Fix lint

5968acd

Fix compile error

6059254

Fix compiler warnings

78fabd8

Fix compiler warnings

faf59b5

Fixed int vector interop. Fixed argmin incorrectly invoking argmax. F…

076aca1

…ixed corner case in default schedules of attempting to fuse 0 length axes. Added tests for reduce ops.

Refactored reduce builders

b4eee54

Fixed typos in topi.cc. Added basic test.

2fa15c3

Fixed padding size error. Added dense, dilate, pooling tests

f505d04

Fixed issue where clip would output a different dtype to the input. A…

8653a0f

…dded split_sections op to cover the other mode of the python split op. Added tests.

Changed extension type numbers to avoid clash with NNVM

050a261

Fix lint

082691c

Fix compiler warnings

11f764e

tqchen requested changes Jan 22, 2018

View reviewed changes

alex-weaver added 3 commits January 24, 2018 11:15

Removed use of std::vector from the public TOPI API

a507aaf

Fix lint

83fe698

Add TOPI C++ tests to CI

408617d

tqchen requested changes Jan 26, 2018

View reviewed changes

Fixed detail namespacing. Improved comments.

0dc1af5

tqchen changed the title ~~[WIP] Porting schedules (except convolutions) to C++~~ Porting schedules (except convolutions) to C++ Jan 28, 2018

tqchen approved these changes Jan 28, 2018

View reviewed changes

tqchen merged commit fda2fa1 into apache:master Jan 28, 2018

tqchen mentioned this pull request May 29, 2018

C++ Compiler API #1184

Closed

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Porting schedules (except convolutions) to C++ #763

Porting schedules (except convolutions) to C++ #763

alex-weaver commented Jan 8, 2018 •

edited

Loading

joyalbin commented Jan 9, 2018

alex-weaver commented Jan 9, 2018

joyalbin commented Jan 9, 2018

tqchen Jan 22, 2018

tqchen Jan 22, 2018

alex-weaver Jan 22, 2018 •

edited

Loading

tqchen Jan 23, 2018

alex-weaver Jan 23, 2018 •

edited

Loading

tqchen Jan 23, 2018

tqchen Jan 26, 2018

tqchen Jan 26, 2018

tqchen Jan 26, 2018

tqchen Jan 26, 2018

tqchen commented Jan 26, 2018

alex-weaver commented Jan 26, 2018

tqchen commented Jan 28, 2018

Porting schedules (except convolutions) to C++ #763

Porting schedules (except convolutions) to C++ #763

Conversation

alex-weaver commented Jan 8, 2018 • edited Loading

Operators

Generic schedules

CUDA schedules

Other schedules

joyalbin commented Jan 9, 2018

alex-weaver commented Jan 9, 2018

joyalbin commented Jan 9, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alex-weaver Jan 22, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alex-weaver Jan 23, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tqchen commented Jan 26, 2018

alex-weaver commented Jan 26, 2018

tqchen commented Jan 28, 2018

alex-weaver commented Jan 8, 2018 •

edited

Loading

alex-weaver Jan 22, 2018 •

edited

Loading

alex-weaver Jan 23, 2018 •

edited

Loading