Change Dispatch.h to use ScalarType over Type #17527

li-roy · 2019-02-27T00:58:55Z

Stack:
    :black_circle: #17527 Change Dispatch.h to use ScalarType over Type  💚
    :white_circle: #17529 Remove some simple use cases of Type::ScalarType()  💚
    :white_circle: #17530 Small clean up of aten_op  💚
    :white_circle: #17601 Store ScalarType and Backend instead of Type in TensorIterator  💚
    :white_circle: #17785 Remove Type::elementSizeInBytes  💚
    :white_circle: #17723 Store python default type as PyTensorType instead of at::Type  💚
    :white_circle: #17786 Pass ScalarType separately from Type in python constructors  💚
    :white_circle: #17792 Remove Type::ScalarType()  💚
    :white_circle: #17603 Remove conversion operator from Type to TensorOptions  💛
    :white_circle: #17787 Add ScalarType arg to Type::options()  💛

Differential Revision: D14235395

Differential Revision: D14233250 Differential Version: 73474786

Differential Revision: D14233250 Differential Version: 73486177

Differential Revision: D14235395 Differential Version: 73486176

Differential Revision: D14235395 Differential Version: 73589963

Differential Revision: D14235395 Differential Version: 74395307

Differential Revision: D14233250 Differential Version: 74395306

Differential Revision: D14235395 Differential Version: 74488005

Differential Revision: D14233250 Differential Version: 74537076

Differential Revision: D14235395 Differential Version: 74537072

gchanan

This made a number of error messages worse.

For example, let's say I have a CPU and CUDA implementation of the same function, fn, which I give in both cases to the dispatch macro as "fn". And as with most functions, I implement the half variant on CUDA, but not CPU.

Then, the user tries to call "fn" with CPU half, and gets the message "fn is not implemented for Half" which is wrong.

Can you just go through and find the cases where this exists (if such a case exists, note it will exist if you shepherd TH functions through this macro).

Differential Revision: D14233250 Differential Version: 74694645

Differential Revision: D14235395 Differential Version: 74694653

Differential Revision: D14235395 Differential Version: 74711130

Differential Revision: D14235395 Differential Version: 74714416

ezyang · 2019-03-08T19:09:24Z

aten/src/ATen/native/BatchLinearAlgebra.cpp

  Tensor self_c = inplace ? self : self.contiguous();
  Tensor result = inplace ? self : at::empty_like(self);
-  AT_DISPATCH_ALL_TYPES(self.type(), "tril", [&]{
+  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "tril", [&]{


tril_cpu here

ezyang · 2019-03-08T19:09:31Z

aten/src/ATen/native/BatchLinearAlgebra.cpp

  }
  Tensor self_c = checkTrilTriuBatchContiguous(self) ? self : self.contiguous();
-  AT_DISPATCH_ALL_TYPES(self.type(), "tril", [&]{
+  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "tril", [&]{


ezyang · 2019-03-08T19:14:03Z

aten/src/ATen/native/BatchLinearAlgebra.cpp

  Tensor self_c = inplace ? self : self.contiguous();
  Tensor result = inplace ? self : at::empty_like(self);
-  AT_DISPATCH_ALL_TYPES(self.type(), "triu", [&]{
+  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "triu", [&]{


ezyang · 2019-03-08T19:15:41Z

aten/src/ATen/native/BatchLinearAlgebra.cpp

  }
  Tensor self_c = checkTrilTriuBatchContiguous(self) ? self : self.contiguous();
-  AT_DISPATCH_ALL_TYPES(self.type(), "triu", [&]{
+  AT_DISPATCH_ALL_TYPES(self.scalar_type(), "triu", [&]{


ezyang · 2019-03-08T19:15:55Z

aten/src/ATen/native/Copy.cpp

      if (!in_parallel_region()) {
        AT_DISPATCH_ALL_TYPES_AND(
-          at::ScalarType::Half, self.type(), "_copy_same_type_", [&]() {
+          at::ScalarType::Half, self.scalar_type(), "_copy_same_type_", [&]() {


ezyang · 2019-03-08T19:15:59Z

aten/src/ATen/native/Copy.cpp

  if (serial_path) {
    AT_DISPATCH_ALL_TYPES_AND(
-      at::ScalarType::Half, self.type(), "_copy_same_type_", [&]() {
+      at::ScalarType::Half, self.scalar_type(), "_copy_same_type_", [&]() {


ezyang · 2019-03-08T19:16:15Z

aten/src/ATen/native/FractionalMaxPool2d.cpp


-  AT_DISPATCH_FLOATING_TYPES(input.type(),
+  AT_DISPATCH_FLOATING_TYPES(input.scalar_type(),
  "fractional_max_pool2d_out_frame", [&] {


ezyang · 2019-03-08T19:16:19Z

aten/src/ATen/native/FractionalMaxPool2d.cpp

  /* backprop */
  AT_DISPATCH_FLOATING_TYPES(
-    input.type(), "fractional_max_pool2d_backward_out_frame", [&] {
+    input.scalar_type(), "fractional_max_pool2d_backward_out_frame", [&] {


ezyang · 2019-03-08T19:16:27Z

aten/src/ATen/native/FractionalMaxPool3d.cpp

  AT_DISPATCH_FLOATING_TYPES(
-    input.type(),
+    input.scalar_type(),
    "fractional_max_pool3d_out_frame",


ezyang

I stopped something like 30% of the way through the patch, but it seems there are still a lot of missing sites

li-roy · 2019-03-08T19:18:42Z

@ezyang I only changed the strings that are used on both cpu and gpu kernels.

Differential Revision: D14233250 Differential Version: 74781242

Differential Revision: D14235395 Differential Version: 74781247

Differential Revision: D14235395 Differential Version: 74811455

Differential Revision: D14235395 Differential Version: 74818459

Summary: Pull Request resolved: pytorch/pytorch#17527 Reviewed By: zou3519 Differential Revision: D14235395 fbshipit-source-id: 3f53e33f6794f1f14c2edf79014b8ef8397822c5

Changes: 1) pytorch#17527 changed dispatch macros to be ScalarType based instead of at::Type based. This broke cpp extensions that relied on dispatch macros. Since IMO these should be ScalarType based, we allow either at::Type or at::ScalarType to be passed, but passing at::Type will result in a deprecated warning. 2) Reintroduce macros that were deleted (AT_DISPATCH_ALL_TYPES_AND_HALF, AT_DISPATCH_COMPLEX_TYPES, AT_DISPATCH_ALL_TYPES_AND_HALF_AND_COMPLEX, AT_DISPATCH_ALL_TYPES_AND_COMPLEX); the AND_HALF ones now give a deprecated warning because there are more extensible macros that were introduced in their place. 3) Makes AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND into a ScalarType based macro (and updates usages). This was the result of a logical merge conflicts.

Summary: Changes: 1) pytorch/pytorch#17527 changed dispatch macros to be ScalarType based instead of at::Type based. This broke cpp extensions that relied on dispatch macros. Since IMO these should be ScalarType based (and some extensions have already updated), we allow either at::Type or at::ScalarType to be passed, but passing at::Type will result in a deprecated warning. 2) Reintroduce macros that were deleted (AT_DISPATCH_ALL_TYPES_AND_HALF, AT_DISPATCH_COMPLEX_TYPES, AT_DISPATCH_ALL_TYPES_AND_HALF_AND_COMPLEX, AT_DISPATCH_ALL_TYPES_AND_COMPLEX); the AND_HALF ones now give a deprecated warning because there are more extensible macros that were introduced in their place. 3) Makes AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND into a ScalarType based macro (and updates usages). This was the result of a logical merge conflicts. 4) Adds a new macro, C10_DEPRECATED_MESSAGE for passing a deprecated message to the compiler. I didn't spend much time seeing if this can be enabled for versions before C++14. Pull Request resolved: pytorch/pytorch#17996 Reviewed By: ezyang Differential Revision: D14446203 Pulled By: gchanan fbshipit-source-id: 1da56e2e9c15aa8f913ebbf6bf1110c5b6dc375e

royboy added 3 commits February 26, 2019 14:15

V1: Initial commit

029ba61

Differential Revision: D14233250 Differential Version: 73474786

V2: (no description)

c789040

Differential Revision: D14233250 Differential Version: 73486177

V1: Initial commit

aa6f43d

Differential Revision: D14235395 Differential Version: 73486176

This was referenced Feb 27, 2019

Replace tensor.type().scalarType() calls with tensor.scalar_type() #17515

Closed

Remove some simple use cases of Type::ScalarType() #17529

Closed

Small clean up of aten_op #17530

Closed

V2: (no description)

df84312

Differential Revision: D14235395 Differential Version: 73589963

This was referenced Mar 1, 2019

Store ScalarType and Backend instead of Type in TensorIterator #17601

Closed

Remove conversion operator from Type to TensorOptions #17603

Closed

li-roy requested a review from gchanan March 1, 2019 23:21

royboy added 3 commits March 5, 2019 19:12

V3: Merge with parent diff changes

8634053

Differential Revision: D14235395 Differential Version: 74395307

V3: Merge with parent diff changes

88179bc

Differential Revision: D14233250 Differential Version: 74395306

V4: (no description)

65f6158

Differential Revision: D14235395 Differential Version: 74488005

li-roy mentioned this pull request Mar 6, 2019

Store python default type as PyTensorType instead of at::Type #17723

Closed

royboy added 2 commits March 6, 2019 16:33

V4: Merge with parent diff changes

d5d2774

Differential Revision: D14233250 Differential Version: 74537076

V5: Merge with parent diff changes

cc0ff69

Differential Revision: D14235395 Differential Version: 74537072

gchanan requested changes Mar 7, 2019

View reviewed changes

royboy added 2 commits March 7, 2019 16:27

V5: Merge with parent diff changes

c536d29

Differential Revision: D14233250 Differential Version: 74694645

V6: Merge with parent diff changes

0ad5534

Differential Revision: D14235395 Differential Version: 74694653

This was referenced Mar 8, 2019

Remove Type::elementSizeInBytes #17785

Closed

Pass ScalarType separately from Type in python constructors #17786

Closed

Add ScalarType arg to Type::options() #17787

Closed

royboy added 2 commits March 7, 2019 18:19

V7: (no description)

0c329b5

Differential Revision: D14235395 Differential Version: 74711130

V8: (no description)

1f2c6b6

Differential Revision: D14235395 Differential Version: 74714416

li-roy mentioned this pull request Mar 8, 2019

Remove Type::ScalarType() #17792

Closed

ezyang reviewed Mar 8, 2019

View reviewed changes

ezyang requested changes Mar 8, 2019

View reviewed changes

gchanan approved these changes Mar 8, 2019

View reviewed changes

royboy added 4 commits March 8, 2019 13:10

V6: Merge with parent diff changes

4fe3ca5

Differential Revision: D14233250 Differential Version: 74781242

V9: Merge with parent diff changes

37f1cdc

Differential Revision: D14235395 Differential Version: 74781247

V10: (no description)

8b00594

Differential Revision: D14235395 Differential Version: 74811455

V11: Merge with parent diff changes

9810bf6

Differential Revision: D14235395 Differential Version: 74818459

li-roy changed the base branch from export-D14233250 to master March 8, 2019 22:18

facebook-github-bot closed this in 3aeb780 Mar 9, 2019

pytorchbot added the merged label Mar 9, 2019

neerajprad mentioned this pull request Mar 12, 2019

Fix underflow issue with dirichlet sample #17488

Closed

gchanan mentioned this pull request Mar 13, 2019

Add backwards compatibility and other fixes to Dispatch macros. #17996

Closed

ezyang deleted the export-D14235395 branch May 30, 2019 15:46

Change Dispatch.h to use ScalarType over Type #17527

Change Dispatch.h to use ScalarType over Type #17527

Uh oh!

Conversation

li-roy commented Feb 27, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gchanan left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ezyang left a comment

Choose a reason for hiding this comment

Uh oh!

li-roy commented Mar 8, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

li-roy commented Feb 27, 2019 •

edited

Loading