Skip to content

Unroll multiply and add instructions in dotProd_32f#15136

Merged
alalek merged 2 commits intoopencv:3.4from
ChipKerchner:dotProd_unroll
Jul 25, 2019
Merged

Unroll multiply and add instructions in dotProd_32f#15136
alalek merged 2 commits intoopencv:3.4from
ChipKerchner:dotProd_unroll

Conversation

@ChipKerchner
Copy link
Copy Markdown
Contributor

@ChipKerchner ChipKerchner commented Jul 24, 2019

Unroll multiply and add instructions (absorb latencies) into separate accumulators in dotProd_32f - 35% faster.

@ChipKerchner ChipKerchner changed the base branch from master to 3.4 July 24, 2019 18:58
@alalek
Copy link
Copy Markdown
Member

alalek commented Jul 24, 2019

@terfendail Could you collect performance changes on IA?

tomoaki0705
tomoaki0705 reviewed Jul 24, 2019
View reviewed changes
modules/core/src/matmul.simd.hpp Outdated
vx_load(src2 + j + (cWidth * 3)), v_sum3);
}

r += v_reduce_sum(v_sum1) + v_reduce_sum(v_sum2) + v_reduce_sum(v_sum3);
Copy link
Copy Markdown
Contributor

@tomoaki0705 tomoaki0705 Jul 24, 2019

Choose a reason for hiding this comment

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

v_reduce_sum is not really fast especially if the HW doesn't have a horizontal sum up instruction.
It's better to reduce the number of call to v_reduce_sum

r += v_reduce_sum(v_sum1 + v_sum2 + v_sum3)

Copy link
Copy Markdown
Member

@alalek alalek Jul 25, 2019

Choose a reason for hiding this comment

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

I believe we can eliminate v_reduce_sum() call from here completely:

v_sum += v_sum1 + v_sum2 + v_sum3;

@alalek alalek merged commit 0db4fb1 into opencv:3.4 Jul 25, 2019
@alalek alalek mentioned this pull request Jul 25, 2019
Merged
@terfendail
Copy link
Copy Markdown
Contributor

terfendail commented Aug 1, 2019

[58/1943] Performance for SSE2 baseline
Performance test Reference time PR time Speedup
dot::MatType_Length::(8UC1, 32) 0.000 0.000 1.00
dot::MatType_Length::(8UC1, 64) 0.000 0.000 1.00
dot::MatType_Length::(8UC1, 128) 0.001 0.001 1.00
dot::MatType_Length::(8UC1, 256) 0.004 0.004 1.01
dot::MatType_Length::(8UC1, 512) 0.018 0.018 0.97
dot::MatType_Length::(8UC1, 1024) 0.082 0.083 0.98
dot::MatType_Length::(32SC1, 32) 0.001 0.001 1.03
dot::MatType_Length::(32SC1, 64) 0.002 0.002 1.00
dot::MatType_Length::(32SC1, 128) 0.008 0.008 1.00
dot::MatType_Length::(32SC1, 256) 0.033 0.033 1.00
dot::MatType_Length::(32SC1, 512) 0.129 0.129 1.00
dot::MatType_Length::(32SC1, 1024) 0.535 0.536 1.00
dot::MatType_Length::(32FC1, 32) 0.000 0.000 2.21
dot::MatType_Length::(32FC1, 64) 0.001 0.001 1.92
dot::MatType_Length::(32FC1, 128) 0.004 0.002 1.80
dot::MatType_Length::(32FC1, 256) 0.016 0.009 1.78
dot::MatType_Length::(32FC1, 512) 0.076 0.077 0.98
dot::MatType_Length::(32FC1, 1024) 0.322 0.332 0.97
[34/1943] Performance for SSE3 baseline
Performance test Reference time PR time Speedup
dot::MatType_Length::(8UC1, 32) 0.000 0.000 0.93
dot::MatType_Length::(8UC1, 64) 0.000 0.000 1.13
dot::MatType_Length::(8UC1, 128) 0.001 0.001 1.03
dot::MatType_Length::(8UC1, 256) 0.004 0.004 1.02
dot::MatType_Length::(8UC1, 512) 0.018 0.017 1.03
dot::MatType_Length::(8UC1, 1024) 0.083 0.083 1.00
dot::MatType_Length::(32SC1, 32) 0.001 0.001 0.99
dot::MatType_Length::(32SC1, 64) 0.002 0.002 1.02
dot::MatType_Length::(32SC1, 128) 0.008 0.008 1.03
dot::MatType_Length::(32SC1, 256) 0.033 0.032 1.03
dot::MatType_Length::(32SC1, 512) 0.129 0.129 1.00
dot::MatType_Length::(32SC1, 1024) 0.536 0.511 1.05
dot::MatType_Length::(32FC1, 32) 0.000 0.000 1.78
dot::MatType_Length::(32FC1, 64) 0.001 0.001 1.92
dot::MatType_Length::(32FC1, 128) 0.004 0.002 1.85
dot::MatType_Length::(32FC1, 256) 0.016 0.009 1.84
dot::MatType_Length::(32FC1, 512) 0.078 0.077 1.01
dot::MatType_Length::(32FC1, 1024) 0.329 0.330 1.00
[10/1943] Performance for SSE4_2 baseline
Performance test Reference time PR time Speedup
dot::MatType_Length::(8UC1, 32) 0.000 0.000 1.02
dot::MatType_Length::(8UC1, 64) 0.000 0.000 1.00
dot::MatType_Length::(8UC1, 128) 0.001 0.001 1.00
dot::MatType_Length::(8UC1, 256) 0.004 0.004 1.00
dot::MatType_Length::(8UC1, 512) 0.017 0.017 1.00
dot::MatType_Length::(8UC1, 1024) 0.090 0.081 1.11
dot::MatType_Length::(32SC1, 32) 0.001 0.001 1.00
dot::MatType_Length::(32SC1, 64) 0.002 0.002 1.00
dot::MatType_Length::(32SC1, 128) 0.008 0.008 1.00
dot::MatType_Length::(32SC1, 256) 0.033 0.033 1.00
dot::MatType_Length::(32SC1, 512) 0.129 0.130 1.00
dot::MatType_Length::(32SC1, 1024) 0.541 0.535 1.01
dot::MatType_Length::(32FC1, 32) 0.000 0.000 2.21
dot::MatType_Length::(32FC1, 64) 0.001 0.001 1.95
dot::MatType_Length::(32FC1, 128) 0.004 0.002 1.79
dot::MatType_Length::(32FC1, 256) 0.016 0.009 1.79
dot::MatType_Length::(32FC1, 512) 0.075 0.073 1.03
dot::MatType_Length::(32FC1, 1024) 0.340 0.331 1.03
Performance for AVX2 baseline
Performance test Reference time PR time Speedup
dot::MatType_Length::(8UC1, 32) 0.000 0.000 1.00
dot::MatType_Length::(8UC1, 64) 0.000 0.000 1.00
dot::MatType_Length::(8UC1, 128) 0.001 0.001 1.00
dot::MatType_Length::(8UC1, 256) 0.003 0.003 1.00
dot::MatType_Length::(8UC1, 512) 0.013 0.013 1.00
dot::MatType_Length::(8UC1, 1024) 0.078 0.078 1.01
dot::MatType_Length::(32SC1, 32) 0.000 0.000 1.02
dot::MatType_Length::(32SC1, 64) 0.002 0.002 1.02
dot::MatType_Length::(32SC1, 128) 0.007 0.007 1.01
dot::MatType_Length::(32SC1, 256) 0.027 0.027 1.00
dot::MatType_Length::(32SC1, 512) 0.109 0.108 1.01
dot::MatType_Length::(32SC1, 1024) 0.448 0.453 0.99
dot::MatType_Length::(32FC1, 32) 0.000 0.000 1.36
dot::MatType_Length::(32FC1, 64) 0.001 0.001 1.18
dot::MatType_Length::(32FC1, 128) 0.002 0.002 1.18
dot::MatType_Length::(32FC1, 256) 0.009 0.008 1.22
dot::MatType_Length::(32FC1, 512) 0.078 0.079 0.99
dot::MatType_Length::(32FC1, 1024) 0.326 0.323 1.01

dvd42 pushed a commit to dvd42/opencv that referenced this pull request Aug 6, 2019
@ChipKerchner @dvd42
Merge pull request opencv#15136 from ChipKerchner:dotProd_unroll
0276ca0 
* Unroll multiply and add instructions in dotProd_32f - 35% faster.

* Eliminate unnecessary v_reduce_sum instructions.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@alalek alalek alalek approved these changes

@terfendail terfendail Awaiting requested review from terfendail

+1 more reviewer

@tomoaki0705 tomoaki0705 tomoaki0705 left review comments

Reviewers whose approvals may not affect merge requirements

Assignees

@alalek alalek

Labels

category: core optimization

Projects

None yet

Milestone

3.4.7

Development

Successfully merging this pull request may close these issues.

4 participants

@ChipKerchner @alalek @terfendail @tomoaki0705