Remaining BBox kernel perf optimizations #6896
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
datumbox
commented
Nov 3, 2022
Comment on lines
+184
to
+188
| w_ratio = new_width / old_width | ||
| h_ratio = new_height / old_height | ||
| ratios = torch.tensor([w_ratio, h_ratio, w_ratio, h_ratio], device=bounding_box.device) | ||
| return ( | ||
| bounding_box.reshape(-1, 2, 2).mul(ratios).to(bounding_box.dtype).reshape(bounding_box.shape), | ||
| bounding_box.mul(ratios).to(bounding_box.dtype), |
There was a problem hiding this comment.
Improvement:
[------------ resize cpu torch.float32 ------------]
| old | new
1 threads: -----------------------------------------
(128, 4) | 13 (+- 0) us | 8 (+- 0) us
6 threads: -----------------------------------------
(128, 4) | 13 (+- 0) us | 8 (+- 0) us
Times are in microseconds (us).
[----------- resize cuda torch.float32 ------------]
| old | new
1 threads: -----------------------------------------
(128, 4) | 37 (+- 0) us | 31 (+- 0) us
6 threads: -----------------------------------------
(128, 4) | 37 (+- 0) us | 31 (+- 0) us
Times are in microseconds (us).
[------------- resize cpu torch.uint8 -------------]
| old | new
1 threads: -----------------------------------------
(128, 4) | 19 (+- 0) us | 13 (+- 0) us
6 threads: -----------------------------------------
(128, 4) | 19 (+- 0) us | 13 (+- 0) us
Times are in microseconds (us).
[------------ resize cuda torch.uint8 -------------]
| old | new
1 threads: -----------------------------------------
(128, 4) | 45 (+- 0) us | 39 (+- 0) us
6 threads: -----------------------------------------
(128, 4) | 45 (+- 0) us | 39 (+- 1) us
Times are in microseconds (us).
pmeier
reviewed
Nov 3, 2022
vfdev-5
reviewed
Nov 3, 2022
|
Maybe, we can merge this after #6879 |
…astic_bounding_box`.
|
@vfdev-5 I just pushed a couple of untested opts. Could you check again which you think are safe? I'll do benchmarks after we confirm which ones we want in. |
datumbox
commented
Nov 3, 2022
|
I'll cherry pick those for elastic those which makes sense. Thanks for pointers! |
datumbox
commented
Nov 3, 2022
| @@ -388,8 +389,7 @@ def _affine_bounding_box_xyxy( | |||
| new_points = torch.matmul(points, transposed_affine_matrix) | |||
| tr, _ = torch.min(new_points, dim=0, keepdim=True) | |||
| # Translate bounding boxes | |||
| out_bboxes[:, 0::2] = out_bboxes[:, 0::2] - tr[:, 0] | |||
| out_bboxes[:, 1::2] = out_bboxes[:, 1::2] - tr[:, 1] | |||
| out_bboxes.sub_(tr.repeat((1, 2))) | |||
There was a problem hiding this comment.
Improvement for both changes:
[-------------------- bbox_rotate cpu -------------------]
| False | True
1 threads: ----------------------------------------------
torch.float32 | 265 (+- 40) us | 225 (+- 2) us
torch.float64 | 261 (+- 1) us | 241 (+- 1) us
torch.int32 | 258 (+- 1) us | 239 (+- 2) us
torch.int64 | 260 (+- 1) us | 239 (+- 1) us
6 threads: ----------------------------------------------
torch.float32 | 466 (+- 10) us | 405 (+- 20) us
torch.float64 | 483 (+- 10) us | 422 (+- 55) us
torch.int32 | 479 (+- 10) us | 420 (+- 10) us
torch.int64 | 482 (+- 18) us | 422 (+- 10) us
Times are in microseconds (us).
[-------------------- bbox_rotate cpu -------------------]
| False | True
1 threads: ----------------------------------------------
torch.float32 | 498 (+- 46) us | 432 (+- 0) us
torch.float64 | 489 (+- 1) us | 446 (+- 0) us
torch.int32 | 503 (+- 0) us | 459 (+- 3) us
torch.int64 | 504 (+- 3) us | 458 (+- 0) us
6 threads: ----------------------------------------------
torch.float32 | 573 (+- 2) us | 530 (+- 0) us
torch.float64 | 600 (+- 20) us | 554 (+- 20) us
torch.int32 | 609 (+- 20) us | 560 (+- 10) us
torch.int64 | 598 (+- 58) us | 563 (+- 10) us
Times are in microseconds (us).
datumbox
changed the title
vfdev-5
approved these changes
Nov 3, 2022
facebook-github-bot
pushed a commit
that referenced
this pull request
Nov 4, 2022
Summary: * Bbox resize optimization * Other (untested) optimizations on `_affine_bounding_box_xyxy` and `elastic_bounding_box`. * fix conflict * Reverting changes on elastic * revert one more change * Further improvement Reviewed By: datumbox Differential Revision: D41020550 fbshipit-source-id: dfd1f2d91490b45176f1976bcec1fc99248f8587
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Some of the opts highlighted at #6872
cc @vfdev-5 @bjuncek @pmeier