Skip to content

Reduce LLC loads, stores and multiplies on MulTransposed - 8% faster#16375

Merged
alalek merged 4 commits intoopencv:3.4from
ChipKerchner:vectorizeMultTranspose
Jan 24, 2020
Merged

Reduce LLC loads, stores and multiplies on MulTransposed - 8% faster#16375
alalek merged 4 commits intoopencv:3.4from
ChipKerchner:vectorizeMultTranspose

Conversation

@ChipKerchner
Copy link
Copy Markdown
Contributor

@ChipKerchner ChipKerchner commented Jan 17, 2020 •
edited
Loading

Reduce LLC loads, stores and multiplies by 2x on MulTransposed - 8% faster

@ChipKerchner
Copy link
Copy Markdown
Contributor Author

ChipKerchner commented Jan 17, 2020

Last failure doesn't seem to be related to my check-in

alalek
alalek reviewed Jan 17, 2020
View reviewed changes
modules/core/src/matmul.simd.hpp Outdated
double s0 = 0, s1 = 0, s2 = 0, s3 = 0;
const sT *tsrc = src + j;
#if CV_SIMD_64F
if (is_same<sT, double>::value && is_same<dT, double>::value)
Copy link
Copy Markdown
Member

@alalek alalek Jan 17, 2020

Choose a reason for hiding this comment

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

Please try type traits from OpenCV:

DataType<sT>::depth == CV_64F && DataType<dT>::depth == CV_64F

Copy link
Copy Markdown
Contributor Author

@ChipKerchner ChipKerchner Jan 17, 2020

Choose a reason for hiding this comment

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

Done

@alalek
Copy link
Copy Markdown
Member

alalek commented Jan 20, 2020

/cc @terfendail BTW, dispatching is blocked by CV_MULTRANSPOSED_BASELINE_ONLY for this function.

@terfendail
Copy link
Copy Markdown
Contributor

terfendail commented Jan 23, 2020

The change looks fine for me.
However I can't reproduce the performance gain. There is no dedicated performance test for the function, so I've used accuracy test evaluation time. For SSE3 baseline I've got performance gain of about 3 per cent, for SSE4_2 baseline I've got performance degradation of 2-3 per cent. Both results look like random fluctuation rather than stable performance change.

@ChipKerchner
Copy link
Copy Markdown
Contributor Author

ChipKerchner commented Jan 23, 2020

The performance gains were measured on a Power9 VSX system. It is possible that the gains only show up on a non-Intel platform since there are different stalls related to memory access for this platform.

@alalek alalek merged commit 4d2da2d into opencv:3.4 Jan 24, 2020
@ChipKerchner ChipKerchner deleted the vectorizeMultTranspose branch January 27, 2020 15:53
@alalek alalek mentioned this pull request Jan 28, 2020
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@alalek alalek alalek left review comments

@terfendail terfendail Awaiting requested review from terfendail

Assignees

@terfendail terfendail

Labels

category: core optimization

Projects

None yet

Milestone

3.4.10

Development

Successfully merging this pull request may close these issues.

3 participants

@ChipKerchner @alalek @terfendail