mirror of
https://github.com/opencv/opencv.git
synced 2024-12-27 11:28:14 +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
40 lines
1.1 KiB
C
40 lines
1.1 KiB
C
/*
|
|
* AVX2 optimized hash slide, based on Intel's slide_sse implementation
|
|
*
|
|
* Copyright (C) 2017 Intel Corporation
|
|
* Authors:
|
|
* Arjan van de Ven <arjan@linux.intel.com>
|
|
* Jim Kukunas <james.t.kukunas@linux.intel.com>
|
|
* Mika T. Lindqvist <postmaster@raasu.org>
|
|
*
|
|
* For conditions of distribution and use, see copyright notice in zlib.h
|
|
*/
|
|
#include "../../zbuild.h"
|
|
#include "../../deflate.h"
|
|
|
|
#include <immintrin.h>
|
|
|
|
static inline void slide_hash_chain(Pos *table, uint32_t entries, const __m256i wsize) {
|
|
table += entries;
|
|
table -= 16;
|
|
|
|
do {
|
|
__m256i value, result;
|
|
|
|
value = _mm256_loadu_si256((__m256i *)table);
|
|
result = _mm256_subs_epu16(value, wsize);
|
|
_mm256_storeu_si256((__m256i *)table, result);
|
|
|
|
table -= 16;
|
|
entries -= 16;
|
|
} while (entries > 0);
|
|
}
|
|
|
|
Z_INTERNAL void slide_hash_avx2(deflate_state *s) {
|
|
uint16_t wsize = (uint16_t)s->w_size;
|
|
const __m256i ymm_wsize = _mm256_set1_epi16((short)wsize);
|
|
|
|
slide_hash_chain(s->head, HASH_SIZE, ymm_wsize);
|
|
slide_hash_chain(s->prev, wsize, ymm_wsize);
|
|
}
|