mirror of
https://github.com/opencv/opencv.git
synced 2024-12-15 18:09:11 +08:00
0de26fd78e
Zlib-ng is zlib replacement with optimizations for "next generation" systems. Its optimization may benifits image library decode and encode speed such as libpng. In our tests, if using zlib-ng and libpng combination on a x86_64 machine with AVX2, the time of `imdecode` amd `imencode` will drop 20% approximately. This patch enables zlib-ng's optimization if `CV_DISABLE_OPTIMIZATION` is OFF. Since Zlib-ng can dispatch intrinsics on the fly, port work is much easier. Related discussion: https://github.com/opencv/opencv/issues/22573
33 lines
1.3 KiB
C
33 lines
1.3 KiB
C
/* adler32_avx2_p.h -- adler32 avx2 utility functions
|
|
* Copyright (C) 2022 Adam Stylinski
|
|
* For conditions of distribution and use, see copyright notice in zlib.h
|
|
*/
|
|
|
|
#ifndef ADLER32_AVX2_P_H_
|
|
#define ADLER32_AVX2_P_H_
|
|
|
|
#if defined(X86_AVX2) || defined(X86_AVX512VNNI)
|
|
|
|
/* 32 bit horizontal sum, adapted from Agner Fog's vector library. */
|
|
static inline uint32_t hsum256(__m256i x) {
|
|
__m128i sum1 = _mm_add_epi32(_mm256_extracti128_si256(x, 1),
|
|
_mm256_castsi256_si128(x));
|
|
__m128i sum2 = _mm_add_epi32(sum1, _mm_unpackhi_epi64(sum1, sum1));
|
|
__m128i sum3 = _mm_add_epi32(sum2, _mm_shuffle_epi32(sum2, 1));
|
|
return (uint32_t)_mm_cvtsi128_si32(sum3);
|
|
}
|
|
|
|
static inline uint32_t partial_hsum256(__m256i x) {
|
|
/* We need a permutation vector to extract every other integer. The
|
|
* rest are going to be zeros */
|
|
const __m256i perm_vec = _mm256_setr_epi32(0, 2, 4, 6, 1, 1, 1, 1);
|
|
__m256i non_zero = _mm256_permutevar8x32_epi32(x, perm_vec);
|
|
__m128i non_zero_sse = _mm256_castsi256_si128(non_zero);
|
|
__m128i sum2 = _mm_add_epi32(non_zero_sse,_mm_unpackhi_epi64(non_zero_sse, non_zero_sse));
|
|
__m128i sum3 = _mm_add_epi32(sum2, _mm_shuffle_epi32(sum2, 1));
|
|
return (uint32_t)_mm_cvtsi128_si32(sum3);
|
|
}
|
|
#endif
|
|
|
|
#endif
|