Skip to content

[IE CLDNN] Perf improvements for global avg pooling in fsv16 format#666

Merged
vladimir-paramuzov merged 1 commit intoopenvinotoolkit:releases/2020/4from
vladimir-paramuzov:private/vparamuz/pooling_acc
Jun 11, 2020
Merged

[IE CLDNN] Perf improvements for global avg pooling in fsv16 format#666
vladimir-paramuzov merged 1 commit intoopenvinotoolkit:releases/2020/4from
vladimir-paramuzov:private/vparamuz/pooling_acc

Conversation

@vladimir-paramuzov
Copy link
Copy Markdown

@vladimir-paramuzov vladimir-paramuzov commented May 28, 2020

  • FP32 accumulator usage in fsv16 pooling led to registers usage increase and spills in some cases, thus significant perf regressions on some models (e.g. efficientnet-b7). Added second branch in the code without src line preload and corresponding function for registers usage estimation. Also reduced simd size to 8 when we have small output tensor to have more threads.

@vladimir-paramuzov vladimir-paramuzov added the category: GPU OpenVINO GPU plugin label May 28, 2020
@vladimir-paramuzov vladimir-paramuzov force-pushed the private/vparamuz/pooling_acc branch 3 times, most recently from 52388f0 to c0af371 Compare June 8, 2020 11:08
@vladimir-paramuzov vladimir-paramuzov force-pushed the private/vparamuz/pooling_acc branch from c0af371 to fb2b55e Compare June 8, 2020 19:34
@vladimir-paramuzov vladimir-paramuzov changed the title [IE CLDNN] fsv16 pooling float accumulator for avg case [IE CLDNN] Perf improvement for global avg pooling in fsv16 format Jun 10, 2020
@vladimir-paramuzov vladimir-paramuzov changed the title [IE CLDNN] Perf improvement for global avg pooling in fsv16 format [IE CLDNN] Perf improvements for global avg pooling in fsv16 format Jun 10, 2020
@vladimir-paramuzov vladimir-paramuzov changed the base branch from master to releases/2020/4 June 10, 2020 08:57
@vladimir-paramuzov vladimir-paramuzov force-pushed the private/vparamuz/pooling_acc branch from fb2b55e to b5819fa Compare June 10, 2020 08:58
@vladimir-paramuzov vladimir-paramuzov marked this pull request as ready for review June 10, 2020 08:59
@vladimir-paramuzov vladimir-paramuzov requested review from a team as code owners June 10, 2020 08:59
Copy link
Copy Markdown
Contributor

@sshlyapn sshlyapn left a comment

Choose a reason for hiding this comment

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

LGTM

@vladimir-paramuzov vladimir-paramuzov force-pushed the private/vparamuz/pooling_acc branch from b5819fa to aa4b3e4 Compare June 10, 2020 21:25
@vladimir-paramuzov vladimir-paramuzov added this to the 2020.4 milestone Jun 10, 2020
@vladimir-paramuzov vladimir-paramuzov merged commit 30f8af7 into openvinotoolkit:releases/2020/4 Jun 11, 2020
@vladimir-paramuzov vladimir-paramuzov deleted the private/vparamuz/pooling_acc branch June 11, 2020 17:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

category: GPU OpenVINO GPU plugin

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants