Optimize calculation of dot product for double vectors with AVX

[stweil]

This improves the performance with best models and should also
make training faster.

Signed-off-by: Stefan Weil [email protected]

[stweil]

This replaces pull request #954. See timing results for OCR there.

Feedback on more timing results on other hardware with AVX (also with training) would be interesting.

[stweil]

The current test results show a large performance increase on server hardware, but on typical other platforms (tested on a virtual machine and on a MacBook Pro) the increase is small (about 2 %).

[amitdo]

You might want to try using FMA.

https://en.wikipedia.org/wiki/FMA_instruction_set
https://en.wikipedia.org/wiki/Multiply%E2%80%93accumulate_operation

[stweil]

All my machines with AVX also support FMA, so I can try that. Thank you for the hint.

[zdenop]

thanks.

[amitdo]

@stweil, a reminder about FMA...

[stweil]

Thank you for the reminder. The first test results with debug code look promissing:

$ time -p tesseract test/testing/phototest.tif - -l tessdata_best/script/Latin -c dotproduct=fma
Page 1
This is a lot of 12 point text to test the
ocr code and see if it works on all types
of file format.

The quick brown dog jumped over the
lazy fox. The quick brown dog jumped
over the lazy fox. The quick brown dog
jumped over the lazy fox. The quick
brown dog jumped over the lazy fox.

real 10.25
user 9.87
sys 0.26
$ time -p tesseract test/testing/phototest.tif - -l tessdata_best/script/Latin -c dotproduct=fma
Page 1
This is a lot of 12 point text to test the
ocr code and see if it works on all types
of file format.

The quick brown dog jumped over the
lazy fox. The quick brown dog jumped
over the lazy fox. The quick brown dog
jumped over the lazy fox. The quick
brown dog jumped over the lazy fox.

real 10.35
user 9.94
sys 0.27
$ time -p tesseract test/testing/phototest.tif - -l tessdata_best/script/Latin -c dotproduct=avx
Page 1
This is a lot of 12 point text to test the
ocr code and see if it works on all types
of file format.

The quick brown dog jumped over the
lazy fox. The quick brown dog jumped
over the lazy fox. The quick brown dog
jumped over the lazy fox. The quick
brown dog jumped over the lazy fox.

real 12.39
user 11.96
sys 0.27
$ time -p tesseract test/testing/phototest.tif - -l tessdata_best/script/Latin -c dotproduct=avx
Page 1
This is a lot of 12 point text to test the
ocr code and see if it works on all types
of file format.

The quick brown dog jumped over the
lazy fox. The quick brown dog jumped
over the lazy fox. The quick brown dog
jumped over the lazy fox. The quick
brown dog jumped over the lazy fox.

real 12.73
user 12.17
sys 0.30

With production code, both AVX and FMA take the same time for this simple image (real 2.86).

[stweil]

FMA is supported in the latest code and gives about the same performance as AVX.

[amitdo]

Thanks!