mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-28 21:20:18 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #15136 from ChipKerchner:dotProd_unroll

Browse Source 
* Unroll multiply and add instructions in dotProd_32f - 35% faster.

* Eliminate unnecessary v_reduce_sum instructions.
This commit is contained in:
Chip Kerchner 2019-07-25 14:21:32 -04:00 committed by Alexander Alekhin
parent ac425f67e4
commit 0db4fb1835
1 changed files with 22 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
modules/core/src/matmul.simd.hpp
View File

@ -2511,6 +2511,27 @@ double dotProd_32f(const float* src1, const float* src2, int len)
int j = 0;
int cWidth = v_float32::nlanes;
#if CV_ENABLE_UNROLLED
v_float32 v_sum1 = vx_setzero_f32();
v_float32 v_sum2 = vx_setzero_f32();
v_float32 v_sum3 = vx_setzero_f32();
for (; j <= blockSize - (cWidth * 4); j += (cWidth * 4))
{
v_sum = v_muladd(vx_load(src1 + j),
vx_load(src2 + j), v_sum);
v_sum1 = v_muladd(vx_load(src1 + j + cWidth),
vx_load(src2 + j + cWidth), v_sum1);
v_sum2 = v_muladd(vx_load(src1 + j + (cWidth * 2)),
vx_load(src2 + j + (cWidth * 2)), v_sum2);
v_sum3 = v_muladd(vx_load(src1 + j + (cWidth * 3)),
vx_load(src2 + j + (cWidth * 3)), v_sum3);
}
v_sum += v_sum1 + v_sum2 + v_sum3;
#endif
for (; j <= blockSize - cWidth; j += cWidth)
v_sum = v_muladd(vx_load(src1 + j), vx_load(src2 + j), v_sum);
@ -2532,4 +2553,4 @@ double dotProd_64f(const double* src1, const double* src2, int len)
#endif
CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace
} // namespace