opencv/3rdparty/zlib-ng/arch/x86/chunkset_avx2.c
Letu Ren 0de26fd78e Add zlib-ng as an alternative zlib implementation
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
2024-01-14 14:58:47 +08:00

134 lines
5.1 KiB
C

/* chunkset_avx2.c -- AVX2 inline functions to copy small data chunks.
* For conditions of distribution and use, see copyright notice in zlib.h
*/
#include "zbuild.h"
#ifdef X86_AVX2
#include <immintrin.h>
#include "../generic/chunk_permute_table.h"
typedef __m256i chunk_t;
#define CHUNK_SIZE 32
#define HAVE_CHUNKMEMSET_2
#define HAVE_CHUNKMEMSET_4
#define HAVE_CHUNKMEMSET_8
#define HAVE_CHUNK_MAG
/* Populate don't cares so that this is a direct lookup (with some indirection into the permute table), because dist can
* never be 0 - 2, we'll start with an offset, subtracting 3 from the input */
static const lut_rem_pair perm_idx_lut[29] = {
{ 0, 2}, /* 3 */
{ 0, 0}, /* don't care */
{ 1 * 32, 2}, /* 5 */
{ 2 * 32, 2}, /* 6 */
{ 3 * 32, 4}, /* 7 */
{ 0 * 32, 0}, /* don't care */
{ 4 * 32, 5}, /* 9 */
{ 5 * 32, 22}, /* 10 */
{ 6 * 32, 21}, /* 11 */
{ 7 * 32, 20}, /* 12 */
{ 8 * 32, 6}, /* 13 */
{ 9 * 32, 4}, /* 14 */
{10 * 32, 2}, /* 15 */
{ 0 * 32, 0}, /* don't care */
{11 * 32, 15}, /* 17 */
{11 * 32 + 16, 14}, /* 18 */
{11 * 32 + 16 * 2, 13}, /* 19 */
{11 * 32 + 16 * 3, 12}, /* 20 */
{11 * 32 + 16 * 4, 11}, /* 21 */
{11 * 32 + 16 * 5, 10}, /* 22 */
{11 * 32 + 16 * 6, 9}, /* 23 */
{11 * 32 + 16 * 7, 8}, /* 24 */
{11 * 32 + 16 * 8, 7}, /* 25 */
{11 * 32 + 16 * 9, 6}, /* 26 */
{11 * 32 + 16 * 10, 5}, /* 27 */
{11 * 32 + 16 * 11, 4}, /* 28 */
{11 * 32 + 16 * 12, 3}, /* 29 */
{11 * 32 + 16 * 13, 2}, /* 30 */
{11 * 32 + 16 * 14, 1} /* 31 */
};
static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
int16_t tmp;
memcpy(&tmp, from, sizeof(tmp));
*chunk = _mm256_set1_epi16(tmp);
}
static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
int32_t tmp;
memcpy(&tmp, from, sizeof(tmp));
*chunk = _mm256_set1_epi32(tmp);
}
static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
int64_t tmp;
memcpy(&tmp, from, sizeof(tmp));
*chunk = _mm256_set1_epi64x(tmp);
}
static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
*chunk = _mm256_loadu_si256((__m256i *)s);
}
static inline void storechunk(uint8_t *out, chunk_t *chunk) {
_mm256_storeu_si256((__m256i *)out, *chunk);
}
static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, uint32_t *chunk_rem, uint32_t dist) {
lut_rem_pair lut_rem = perm_idx_lut[dist - 3];
__m256i ret_vec;
/* While technically we only need to read 4 or 8 bytes into this vector register for a lot of cases, GCC is
* compiling this to a shared load for all branches, preferring the simpler code. Given that the buf value isn't in
* GPRs to begin with the 256 bit load is _probably_ just as inexpensive */
*chunk_rem = lut_rem.remval;
/* See note in chunkset_ssse3.c for why this is ok */
__msan_unpoison(buf + dist, 32 - dist);
if (dist < 16) {
/* This simpler case still requires us to shuffle in 128 bit lanes, so we must apply a static offset after
* broadcasting the first vector register to both halves. This is _marginally_ faster than doing two separate
* shuffles and combining the halves later */
const __m256i permute_xform =
_mm256_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16);
__m256i perm_vec = _mm256_load_si256((__m256i*)(permute_table+lut_rem.idx));
__m128i ret_vec0 = _mm_loadu_si128((__m128i*)buf);
perm_vec = _mm256_add_epi8(perm_vec, permute_xform);
ret_vec = _mm256_inserti128_si256(_mm256_castsi128_si256(ret_vec0), ret_vec0, 1);
ret_vec = _mm256_shuffle_epi8(ret_vec, perm_vec);
} else if (dist == 16) {
__m128i ret_vec0 = _mm_loadu_si128((__m128i*)buf);
return _mm256_inserti128_si256(_mm256_castsi128_si256(ret_vec0), ret_vec0, 1);
} else {
__m128i ret_vec0 = _mm_loadu_si128((__m128i*)buf);
__m128i ret_vec1 = _mm_loadu_si128((__m128i*)(buf + 16));
/* Take advantage of the fact that only the latter half of the 256 bit vector will actually differ */
__m128i perm_vec1 = _mm_load_si128((__m128i*)(permute_table + lut_rem.idx));
__m128i xlane_permutes = _mm_cmpgt_epi8(_mm_set1_epi8(16), perm_vec1);
__m128i xlane_res = _mm_shuffle_epi8(ret_vec0, perm_vec1);
/* Since we can't wrap twice, we can simply keep the later half exactly how it is instead of having to _also_
* shuffle those values */
__m128i latter_half = _mm_blendv_epi8(ret_vec1, xlane_res, xlane_permutes);
ret_vec = _mm256_inserti128_si256(_mm256_castsi128_si256(ret_vec0), latter_half, 1);
}
return ret_vec;
}
#define CHUNKSIZE chunksize_avx2
#define CHUNKCOPY chunkcopy_avx2
#define CHUNKUNROLL chunkunroll_avx2
#define CHUNKMEMSET chunkmemset_avx2
#define CHUNKMEMSET_SAFE chunkmemset_safe_avx2
#include "chunkset_tpl.h"
#define INFLATE_FAST inflate_fast_avx2
#include "inffast_tpl.h"
#endif