Optimize opencv dft by vectorizing radix2 and radix3.#18158
Merged
opencv-pushbot merged 1 commit intoopencv:3.4from Oct 30, 2020
Merged
Optimize opencv dft by vectorizing radix2 and radix3.#18158opencv-pushbot merged 1 commit intoopencv:3.4from
opencv-pushbot merged 1 commit intoopencv:3.4from
Conversation
acfa28a to
c420c72
Compare
Contributor
Author
|
I've removed the constexpr since this is supposed to go in 3.4. From the coding style:
|
This is useful for non power-of-two sizes when WITH_IPP is not an option. This shows consistent improvement over openCV benchmarks, and we measure even larger improvements on our internal workloads. For example, for 320x480, `32FC*`, we can see a ~5% improvement}, as `320=2^6*5` and `480=2^5*3*5`, so the improved radix3 version is used. `64FC*` is flat as expected, as we do not specialize the functors for `double` in this change. ``` dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, 0, false) 1.239 1.153 1.07 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, 0, true) 0.991 0.926 1.07 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_COMPLEX_OUTPUT, false) 1.367 1.281 1.07 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_COMPLEX_OUTPUT, true) 1.114 1.049 1.06 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_INVERSE, false) 1.313 1.254 1.05 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_INVERSE, true) 1.027 0.977 1.05 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 1.296 1.217 1.06 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 1.039 0.963 1.08 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_ROWS, false) 0.542 0.524 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_ROWS, true) 0.293 0.277 1.06 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_SCALE, false) 1.265 1.175 1.08 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC1, DFT_SCALE, true) 1.004 0.942 1.07 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, 0, false) 1.292 1.280 1.01 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, 0, true) 1.038 1.030 1.01 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_COMPLEX_OUTPUT, false) 1.484 1.488 1.00 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_COMPLEX_OUTPUT, true) 1.222 1.224 1.00 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_INVERSE, false) 1.380 1.355 1.02 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_INVERSE, true) 1.117 1.133 0.99 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 1.372 1.383 0.99 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 1.117 1.127 0.99 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_ROWS, false) 0.546 0.539 1.01 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_ROWS, true) 0.293 0.299 0.98 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_SCALE, false) 1.351 1.339 1.01 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 64FC1, DFT_SCALE, true) 1.099 1.092 1.01 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, 0, false) 2.235 2.123 1.05 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, 0, true) 1.843 1.727 1.07 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_COMPLEX_OUTPUT, false) 2.189 2.109 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_COMPLEX_OUTPUT, true) 1.827 1.754 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_INVERSE, false) 2.392 2.309 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_INVERSE, true) 1.951 1.865 1.05 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 2.391 2.293 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 1.954 1.882 1.04 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_ROWS, false) 0.811 0.815 0.99 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_ROWS, true) 0.426 0.437 0.98 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_SCALE, false) 2.268 2.152 1.05 dft::Size_MatType_FlagsType_NzeroRows::(320x480, 32FC2, DFT_SCALE, true) 1.893 1.788 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, 0, false) 4.546 4.395 1.03 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, 0, true) 3.616 3.426 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_COMPLEX_OUTPUT, false) 4.843 4.668 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_COMPLEX_OUTPUT, true) 3.825 3.748 1.02 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_INVERSE, false) 4.720 4.525 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_INVERSE, true) 3.743 3.601 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 4.755 4.527 1.05 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 3.744 3.586 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_ROWS, false) 1.992 2.012 0.99 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_ROWS, true) 1.048 1.048 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_SCALE, false) 4.625 4.451 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC1, DFT_SCALE, true) 3.643 3.491 1.04 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, 0, false) 4.499 4.488 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, 0, true) 3.559 3.555 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_COMPLEX_OUTPUT, false) 5.155 5.165 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_COMPLEX_OUTPUT, true) 4.103 4.101 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_INVERSE, false) 5.484 5.474 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_INVERSE, true) 4.617 4.518 1.02 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 5.547 5.509 1.01 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 4.553 4.554 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_ROWS, false) 2.067 2.018 1.02 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_ROWS, true) 1.104 1.079 1.02 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_SCALE, false) 4.665 4.619 1.01 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 64FC1, DFT_SCALE, true) 3.698 3.681 1.00 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, 0, false) 8.774 8.275 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, 0, true) 6.975 6.527 1.07 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_COMPLEX_OUTPUT, false) 8.720 8.270 1.05 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_COMPLEX_OUTPUT, true) 6.928 6.532 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_INVERSE, false) 9.272 8.862 1.05 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_INVERSE, true) 7.323 6.946 1.05 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 9.262 8.768 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 7.298 6.871 1.06 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_ROWS, false) 3.766 3.639 1.03 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_ROWS, true) 1.932 1.889 1.02 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_SCALE, false) 8.865 8.417 1.05 dft::Size_MatType_FlagsType_NzeroRows::(800x600, 32FC2, DFT_SCALE, true) 7.067 6.643 1.06 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, 0, false) 10.014 10.141 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, 0, true) 7.600 7.632 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_COMPLEX_OUTPUT, false) 11.059 11.283 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_COMPLEX_OUTPUT, true) 8.475 8.552 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_INVERSE, false) 12.678 12.789 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_INVERSE, true) 10.445 10.359 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 12.626 12.925 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 10.538 10.553 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_ROWS, false) 5.041 5.084 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_ROWS, true) 2.595 2.607 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_SCALE, false) 10.231 10.330 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC1, DFT_SCALE, true) 7.786 7.815 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, 0, false) 13.597 13.302 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, 0, true) 10.377 10.207 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_COMPLEX_OUTPUT, false) 15.940 15.545 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_COMPLEX_OUTPUT, true) 12.299 12.230 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_INVERSE, false) 15.270 15.181 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_INVERSE, true) 12.757 12.339 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 15.512 15.157 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 12.505 12.635 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_ROWS, false) 6.359 6.255 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_ROWS, true) 3.314 3.248 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_SCALE, false) 13.937 13.733 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 64FC1, DFT_SCALE, true) 10.782 10.495 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, 0, false) 18.985 18.926 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, 0, true) 14.256 14.509 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_COMPLEX_OUTPUT, false) 18.696 19.021 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_COMPLEX_OUTPUT, true) 14.290 14.429 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_INVERSE, false) 20.135 20.296 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_INVERSE, true) 15.390 15.512 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 20.121 20.354 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 15.341 15.605 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_ROWS, false) 8.932 9.084 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_ROWS, true) 4.539 4.649 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_SCALE, false) 19.137 19.303 0.99 dft::Size_MatType_FlagsType_NzeroRows::(1280x1024, 32FC2, DFT_SCALE, true) 14.565 14.808 0.98 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, 0, false) 22.553 21.171 1.07 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, 0, true) 17.850 16.390 1.09 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_COMPLEX_OUTPUT, false) 24.062 22.634 1.06 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_COMPLEX_OUTPUT, true) 19.342 17.932 1.08 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_INVERSE, false) 28.609 27.326 1.05 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_INVERSE, true) 24.591 23.289 1.06 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 28.667 27.467 1.04 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 24.671 23.309 1.06 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_ROWS, false) 9.458 9.077 1.04 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_ROWS, true) 4.709 4.566 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_SCALE, false) 22.791 21.583 1.06 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC1, DFT_SCALE, true) 18.029 16.691 1.08 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, 0, false) 25.238 24.427 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, 0, true) 19.636 19.270 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_COMPLEX_OUTPUT, false) 28.342 27.957 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_COMPLEX_OUTPUT, true) 22.413 22.477 1.00 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_INVERSE, false) 26.465 26.085 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_INVERSE, true) 21.972 21.704 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 26.497 26.127 1.01 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 22.010 21.523 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_ROWS, false) 11.188 10.774 1.04 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_ROWS, true) 6.094 5.916 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_SCALE, false) 25.728 24.934 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 64FC1, DFT_SCALE, true) 20.077 19.653 1.02 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, 0, false) 43.834 40.726 1.08 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, 0, true) 35.198 32.218 1.09 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_COMPLEX_OUTPUT, false) 43.743 40.897 1.07 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_COMPLEX_OUTPUT, true) 35.240 32.226 1.09 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_INVERSE, false) 46.022 42.612 1.08 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_INVERSE, true) 36.779 33.961 1.08 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 46.396 42.723 1.09 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 37.025 33.874 1.09 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_ROWS, false) 17.334 16.832 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_ROWS, true) 9.212 8.970 1.03 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_SCALE, false) 44.190 41.211 1.07 dft::Size_MatType_FlagsType_NzeroRows::(1920x1080, 32FC2, DFT_SCALE, true) 35.900 32.888 1.09 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, 0, false) 40.948 38.256 1.07 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, 0, true) 33.825 30.759 1.10 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_COMPLEX_OUTPUT, false) 53.210 53.584 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_COMPLEX_OUTPUT, true) 46.356 46.712 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_INVERSE, false) 47.471 47.213 1.01 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_INVERSE, true) 40.491 41.363 0.98 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 46.724 47.049 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 40.834 41.381 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_ROWS, false) 14.508 14.490 1.00 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_ROWS, true) 7.832 7.828 1.00 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_SCALE, false) 41.491 38.341 1.08 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC1, DFT_SCALE, true) 34.587 31.208 1.11 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, 0, false) 65.155 63.173 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, 0, true) 56.091 54.752 1.02 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_COMPLEX_OUTPUT, false) 71.549 70.626 1.01 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_COMPLEX_OUTPUT, true) 62.319 61.437 1.01 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_INVERSE, false) 61.480 59.540 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_INVERSE, true) 54.047 52.650 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 61.752 61.366 1.01 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 54.400 53.665 1.01 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_ROWS, false) 20.219 19.704 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_ROWS, true) 11.145 10.868 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_SCALE, false) 66.220 64.525 1.03 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 64FC1, DFT_SCALE, true) 57.389 56.114 1.02 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, 0, false) 86.761 88.128 0.98 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, 0, true) 75.528 76.725 0.98 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_COMPLEX_OUTPUT, false) 86.750 88.223 0.98 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_COMPLEX_OUTPUT, true) 75.830 76.809 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_INVERSE, false) 91.728 92.161 1.00 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_INVERSE, true) 78.797 79.876 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, false) 92.163 92.177 1.00 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_INVERSE|DFT_COMPLEX_OUTPUT, true) 78.957 79.863 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_ROWS, false) 24.781 25.576 0.97 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_ROWS, true) 13.226 13.695 0.97 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_SCALE, false) 87.990 89.324 0.99 dft::Size_MatType_FlagsType_NzeroRows::(2048x2048, 32FC2, DFT_SCALE, true) 76.732 77.869 0.99 ```
c420c72 to
da555a2
Compare
alalek
reviewed
Aug 21, 2020
| // vector register. | ||
| inline __m128 complexMul(const Complex<float>* const a, const Complex<float>* const b) { | ||
| const __m128 z = _mm_setzero_ps(); | ||
| const __m128 neg_elem0 = _mm_set_ps(0.0f,0.0f,0.0f,-0.0f); |
Member
There was a problem hiding this comment.
New code with native SSE/AVX intrinsics is not accepted anymore without a strong reason.
Please try to use approach with "Universal SIMD Intrinsics" instead (which covers multiple platforms like ARM NEON, Power VSX, etc).
Some words with motivation: https://github.com/opencv/opencv/wiki/OE-27.-Wide-Universal-Intrinsics
BTW, Modern compilers are smart enough with optimization or auto-vectorization of floating-point code.
Member
There was a problem hiding this comment.
Tested with disabled added native SSE code. Regression is about 7-9%.
Also there is DFT_VecR4 implementation with native SSE code which is not added by this patch.
So just removal of native intrinsics is not a good option.
Unfortunately there is no resources to port on OpenCV SIMD UI too, so lets merge this "as is".
Contributor
|
@legrosbuffle, thank you for the contribution! Can you please convert your code to OpenCV's universal intrinsics? |
Contributor
Author
|
Hi Alexander, Vadim, Sorry I missed then previous message. I'll have a loot at it ASAP. |
alalek
approved these changes
Oct 30, 2020
Merged
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is useful for non power-of-two sizes when WITH_IPP is not an option.
This shows consistent improvement on openCV benchmarks, and we measure
even larger improvements on our internal workloads.
For example, for 320x480,
32FC*, we can see a ~5% improvement, as320=2^6*5and480=2^5*3*5, so the improved radix3 version is used.64FC*is flat as expected, as we do not specialize the functors fordoublein this change.
Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.