Skip to content

[gfx90a] Winograd f3x2 fp32 support#1158

Merged
junliume merged 1 commit into
developfrom
winograd-f3x2-gfx90a-fp32
Sep 16, 2021
Merged

[gfx90a] Winograd f3x2 fp32 support#1158
junliume merged 1 commit into
developfrom
winograd-f3x2-gfx90a-fp32

Conversation

@Slimakanzer
Copy link
Copy Markdown
Contributor

@Slimakanzer Slimakanzer commented Sep 15, 2021

This PR adds Winograd f3x2 fp32 support for gfx90a. Partially resolves issue https://github.com/ROCmSoftwarePlatform/MIOpen-internal/issues/2.

Local verification

  • gfx90a, rocm 4.3.1
Test cases
-W -H -c -n -k -x -y -q -p -u -v dir
17 17 128 32 128 7 1 3 0 1 1 fwd/bwd/wrw
17 17 128 32 192 7 1 3 0 1 1 fwd/bwd/wrw
17 17 160 32 160 7 1 3 0 1 1 fwd/bwd/wrw
17 17 160 32 192 7 1 3 0 1 1 fwd/bwd/wrw
17 17 192 32 192 7 1 3 0 1 1 fwd/bwd/wrw
17 17 192 32 224 7 1 3 0 1 1 fwd/bwd/wrw
17 17 224 32 224 7 1 3 0 1 1 fwd/bwd/wrw
17 17 256 32 256 7 1 3 0 1 1 fwd/bwd/wrw
73 73 64 32 64 7 1 3 0 1 1 fwd/bwd/wrw
7 7 32 32 128 5 5 2 2 1 1 fwd/bwd/wrw
7 7 48 32 128 5 5 2 2 1 1 fwd/bwd/wrw
7 7 832 1 128 5 5 2 2 1 1 fwd/bwd/wrw
7 7 832 2 128 5 5 2 2 1 1 fwd/bwd/wrw
7 7 832 16 128 5 5 2 2 1 1 fwd/bwd/wrw
14 14 16 32 48 5 5 2 2 1 1 fwd/bwd/wrw
14 14 24 32 64 5 5 2 2 1 1 fwd/bwd/wrw
14 14 32 32 64 5 5 2 2 1 1 fwd/bwd/wrw
14 14 32 32 128 5 5 2 2 1 1 fwd/bwd/wrw
14 14 512 1 48 5 5 2 2 1 1 fwd/bwd/wrw
14 14 512 2 48 5 5 2 2 1 1 fwd/bwd/wrw
14 14 512 16 48 5 5 2 2 1 1 fwd/bwd/wrw
27 27 64 128 192 5 5 2 2 1 1 fwd/bwd/wrw
27 27 96 50 256 5 5 2 2 1 1 fwd/bwd/wrw
27 27 96 64 256 5 5 2 2 1 1 fwd/bwd/wrw
27 27 96 256 256 5 5 2 2 1 1 fwd/bwd/wrw
28 28 16 32 32 5 5 2 2 1 1 fwd/bwd/wrw
28 28 32 32 96 5 5 2 2 1 1 fwd/bwd/wrw
28 28 192 1 32 5 5 2 2 1 1 fwd/bwd/wrw
28 28 192 2 32 5 5 2 2 1 1 fwd/bwd/wrw
28 28 192 16 32 5 5 2 2 1 1 fwd/bwd/wrw
35 35 48 32 64 5 5 2 2 1 1 fwd/bwd/wrw
128 64 16 1 48 5 5 2 2 1 1 fwd/bwd/wrw
128 64 24 1 64 5 5 2 2 1 1 fwd/bwd/wrw
128 64 32 1 64 5 5 2 2 1 1 fwd/bwd/wrw
128 64 32 1 128 5 5 2 2 1 1 fwd/bwd/wrw
128 64 48 1 128 5 5 2 2 1 1 fwd/bwd/wrw
256 128 16 1 32 5 5 2 2 1 1 fwd/bwd/wrw
256 128 32 1 96 5 5 2 2 1 1 fwd/bwd/wrw
17 17 64 32 64 7 1 3 0 1 1 fwd/bwd/wrw
73 73 64 32 64 7 1 3 0 1 1 fwd/bwd/wrw
7 7 32 32 128 5 5 2 2 1 1 fwd/bwd/wrw
14 14 16 32 48 5 5 2 2 1 1 fwd/bwd/wrw
14 14 24 32 64 5 5 2 2 1 1 fwd/bwd/wrw
13 13 64 16 384 3 3 1 1 1 1 fwd/bwd/wrw
28 28 128 16 128 3 3 1 1 1 1 fwd/bwd/wrw

@junliume
Copy link
Copy Markdown
Contributor

@Slimakanzer test_conv2d has failed because:

FAILED: filesystem::recursive_directory_iterator increment error: No such file or directory

It looks familiar ... I'll start the job again.

@codecov

This comment has been minimized.

Copy link
Copy Markdown
Contributor

@atamazov atamazov left a comment

Choose a reason for hiding this comment

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

LGTM!

@atamazov
Copy link
Copy Markdown
Contributor

@junliume I see that you've launched another CI job, so I leave the merging to you.

@junliume
Copy link
Copy Markdown
Contributor

@junliume I see that you've launched another CI job, so I leave the merging to you.

Yes that one has passed too, will merge asap. Thanks!

@junliume junliume merged commit db72b7d into develop Sep 16, 2021
@atamazov atamazov deleted the winograd-f3x2-gfx90a-fp32 branch October 8, 2021 13:29
assistant-librarian Bot pushed a commit that referenced this pull request Aug 13, 2025
[ROCm][Windows] Add missing MIOPEN_INTERNALS_EXPORT to
 convolution ocl functions (#1158)

This PR solves an issue which is same as
https://github.com/ROCm/rocm-libraries/pull/877/files#r2261316493 but in
different functions.
Missing MIOPEN_INTERNALS_EXPORT causes Windows build failure with
undefined symbol error, so adding MIOPEN_INTERNALS_EXPORT before
functions (just like in this
[PR](ROCm/rocm-libraries#1110)) solves the
issue.

Co-authored-by: BrianHarrisonAMD <169072757+BrianHarrisonAMD@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants