Skip to content

[prototype] Clean up and port the resize kernel in V2#6892

Merged
datumbox merged 6 commits intopytorch:mainfrom
datumbox:prototype/port_resize
Nov 3, 2022
Merged

[prototype] Clean up and port the resize kernel in V2#6892
datumbox merged 6 commits intopytorch:mainfrom
datumbox:prototype/port_resize

Conversation

@datumbox
Copy link
Contributor

@datumbox datumbox commented Nov 2, 2022 •
edited
Loading

This PR:

  • Moves the implementation of resize into V2
  • Removes unnecessary logic related to reshaping from the previous implementation and performs equivalent simplifications
  • Does a few minor perf optimzations such inplace clamp/round etc

No speed regression, small speed improvements due to simplified logic:

InterpolationMode.BICUBIC

[----------------- resize cpu torch.float32 ----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |    1 (+-  0) ms  |    1 (+-  0) ms
      (16, 3, 400, 500)  |   22 (+-  0) ms  |   22 (+-  0) ms
6 threads: --------------------------------------------------
      (3, 400, 500)      |    1 (+-  0) ms  |    1 (+-  0) ms
      (16, 3, 400, 500)  |   22 (+-  0) ms  |   22 (+-  0) ms

Times are in milliseconds (ms).

[---------------- resize cuda torch.float32 ----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |   23 (+-  0) us  |   22 (+-  0) us
      (16, 3, 400, 500)  |  104 (+-  1) us  |  104 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |   23 (+-  0) us  |   22 (+-  0) us
      (16, 3, 400, 500)  |  104 (+-  0) us  |  104 (+-  0) us

Times are in microseconds (us).

[------------------ resize cpu torch.uint8 -----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |    2 (+-  0) ms  |    2 (+-  0) ms
      (16, 3, 400, 500)  |   45 (+-  0) ms  |   42 (+-  0) ms
6 threads: --------------------------------------------------
      (3, 400, 500)      |    2 (+-  0) ms  |    2 (+-  0) ms
      (16, 3, 400, 500)  |   47 (+-  0) ms  |   43 (+-  1) ms

Times are in milliseconds (ms).

[----------------- resize cuda torch.uint8 -----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |   60 (+-  0) us  |   56 (+-  0) us
      (16, 3, 400, 500)  |  190 (+-  0) us  |  188 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |   61 (+-  1) us  |   56 (+-  0) us
      (16, 3, 400, 500)  |  190 (+-  1) us  |  188 (+-  0) us
Times are in microseconds (us).

InterpolationMode.BILINEAR:
[------------------ resize cpu torch.float32 -----------------]
                         |         old       |         new     
1 threads: ----------------------------------------------------
      (3, 400, 500)      |   320 (+-  1) us  |   319 (+-  1) us
      (16, 3, 400, 500)  |  6120 (+-100) us  |  6037 (+- 37) us
6 threads: ----------------------------------------------------
      (3, 400, 500)      |   372 (+-  4) us  |   370 (+-  8) us
      (16, 3, 400, 500)  |  6111 (+-174) us  |  6113 (+-109) us

Times are in microseconds (us).

[---------------- resize cuda torch.float32 ----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |   17 (+-  0) us  |   16 (+-  0) us
      (16, 3, 400, 500)  |   53 (+-  0) us  |   53 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |   17 (+-  0) us  |   16 (+-  0) us
      (16, 3, 400, 500)  |   53 (+-  0) us  |   53 (+-  0) us

Times are in microseconds (us).

[-------------------- resize cpu torch.uint8 -------------------]
                         |         old        |         new      
1 threads: ------------------------------------------------------
      (3, 400, 500)      |   624 (+-  1) us   |   591 (+-  1) us 
      (16, 3, 400, 500)  |  27336 (+- 47) us  |  25671 (+-182) us
6 threads: ------------------------------------------------------
      (3, 400, 500)      |   797 (+- 18) us   |   777 (+- 12) us 
      (16, 3, 400, 500)  |  28049 (+-321) us  |  26405 (+-237) us

Times are in microseconds (us).

[----------------- resize cuda torch.uint8 -----------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |   44 (+-  0) us  |   41 (+-  0) us
      (16, 3, 400, 500)  |  131 (+-  0) us  |  130 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |   44 (+-  0) us  |   41 (+-  1) us
      (16, 3, 400, 500)  |  131 (+-  1) us  |  130 (+-  2) us

Times are in microseconds (us).

cc @vfdev-5 @bjuncek @pmeier

@datumbox datumbox requested review from pmeier and vfdev-5 November 2, 2022 18:08
datumbox
datumbox commented Nov 2, 2022
View reviewed changes
pmeier
pmeier approved these changes Nov 2, 2022
View reviewed changes
Copy link
Contributor

@pmeier pmeier left a comment •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One question, otherwise LGTM if CI is green! Thanks Vasilis.

torchvision/prototype/transforms/functional/_geometry.py
Comment on lines +146 to +148
if interpolation == InterpolationMode.BICUBIC and dtype == torch.uint8:
image = image.clamp_(min=0, max=255)
image = image.round_().to(dtype=dtype)
Copy link
Contributor

@pmeier pmeier Nov 2, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reminder to self that I need to address this in #6830.

@datumbox datumbox merged commit d8cec34 into pytorch:main Nov 3, 2022
@datumbox datumbox deleted the prototype/port_resize branch November 3, 2022 09:22
facebook-github-bot pushed a commit that referenced this pull request Nov 4, 2022
@NicolasHug @pmeier @facebook-github-bot
[fbsync] [prototype] Clean up and port the resize kernel in V2 (#6892)
06681ef 
Summary:
* Ported `resize`

* Align with previous behaviour

* Update torchvision/prototype/transforms/functional/_geometry.py

* Moving input verification on top of method.

Reviewed By: datumbox

Differential Revision: D41020541

fbshipit-source-id: ef55147eb263fc530e2068b914d1e2b539415260

Co-authored-by: Philip Meier <github.pmeier@posteo.de>
Co-authored-by: Philip Meier <github.pmeier@posteo.de>
@datumbox datumbox mentioned this pull request Nov 8, 2022
Closed
31 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@vfdev-5 vfdev-5 Awaiting requested review from vfdev-5

1 more reviewer

@pmeier pmeier pmeier approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

cla signed code quality module: transforms prototype

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@datumbox @pmeier @facebook-github-bot

Comments