aaru-dps/Aaru.Checksums.Native
1
0
Fork 0
mirror of https://github.com/aaru-dps/Aaru.Checksums.Native.git synced 2025-12-16 11:14:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix Adler and Fletcher calculations using SIMD when dataset is smaller than block size.

Browse Source
This commit is contained in:
Natalia Portillo
2023-09-24 19:33:25 +01:00
parent 89382334ec
commit 0d9d1d92eb
9 changed files with 750 additions and 699 deletions
Download Patch File Download Diff File Expand all files Collapse all files

107
fletcher32_ssse3.c
View File

@@ -59,68 +59,72 @@ fletcher32_ssse3(uint16_t *sum1, uint16_t *sum2, const uint8_t *data, long len)
* Process the data in blocks.
*/
const unsigned BLOCK_SIZE = 1 << 5;
long blocks = len / BLOCK_SIZE;
len -= blocks * BLOCK_SIZE;
while(blocks)
if(len >= BLOCK_SIZE)
{
unsigned n = NMAX / BLOCK_SIZE; /* The NMAX constraint. */
if(n > blocks) n = (unsigned)blocks;
blocks -= n;
const __m128i tap1 = _mm_setr_epi8(32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17);
const __m128i tap2 = _mm_setr_epi8(16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
const __m128i zero = _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
const __m128i ones = _mm_set_epi16(1, 1, 1, 1, 1, 1, 1, 1);
/*
* Process n blocks of data. At most NMAX data bytes can be
* processed before s2 must be reduced modulo BASE.
*/
__m128i v_ps = _mm_set_epi32(0, 0, 0, s1 * n);
__m128i v_s2 = _mm_set_epi32(0, 0, 0, s2);
__m128i v_s1 = _mm_set_epi32(0, 0, 0, 0);
do
long blocks = len / BLOCK_SIZE;
len -= blocks * BLOCK_SIZE;
while(blocks)
{
unsigned n = NMAX / BLOCK_SIZE; /* The NMAX constraint. */
if(n > blocks) n = (unsigned)blocks;
blocks -= n;
const __m128i tap1 = _mm_setr_epi8(32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17);
const __m128i tap2 = _mm_setr_epi8(16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
const __m128i zero = _mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
const __m128i ones = _mm_set_epi16(1, 1, 1, 1, 1, 1, 1, 1);
/*
* Load 32 input bytes.
* Process n blocks of data. At most NMAX data bytes can be
* processed before s2 must be reduced modulo BASE.
*/
const __m128i bytes1 = _mm_loadu_si128((__m128i *)(data));
const __m128i bytes2 = _mm_loadu_si128((__m128i *)(data + 16));
__m128i v_ps = _mm_set_epi32(0, 0, 0, s1 * n);
__m128i v_s2 = _mm_set_epi32(0, 0, 0, s2);
__m128i v_s1 = _mm_set_epi32(0, 0, 0, 0);
do
{
/*
* Load 32 input bytes.
*/
const __m128i bytes1 = _mm_loadu_si128((__m128i *)(data));
const __m128i bytes2 = _mm_loadu_si128((__m128i *)(data + 16));
/*
* Add previous block byte sum to v_ps.
*/
v_ps = _mm_add_epi32(v_ps, v_s1);
/*
* Horizontally add the bytes for s1, multiply-adds the
* bytes by [ 32, 31, 30, ... ] for s2.
*/
v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes1, zero));
const __m128i mad1 = _mm_maddubs_epi16(bytes1, tap1);
v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad1, ones));
v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes2, zero));
const __m128i mad2 = _mm_maddubs_epi16(bytes2, tap2);
v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad2, ones));
data += BLOCK_SIZE;
}
while(--n);
v_s2 = _mm_add_epi32(v_s2, _mm_slli_epi32(v_ps, 5));
/*
* Add previous block byte sum to v_ps.
* Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
*/
v_ps = _mm_add_epi32(v_ps, v_s1);
/*
* Horizontally add the bytes for s1, multiply-adds the
* bytes by [ 32, 31, 30, ... ] for s2.
*/
v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes1, zero));
const __m128i mad1 = _mm_maddubs_epi16(bytes1, tap1);
v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad1, ones));
v_s1 = _mm_add_epi32(v_s1, _mm_sad_epu8(bytes2, zero));
const __m128i mad2 = _mm_maddubs_epi16(bytes2, tap2);
v_s2 = _mm_add_epi32(v_s2, _mm_madd_epi16(mad2, ones));
data += BLOCK_SIZE;
}
while(--n);
v_s2 = _mm_add_epi32(v_s2, _mm_slli_epi32(v_ps, 5));
/*
* Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
*/
#define S23O1 _MM_SHUFFLE(2, 3, 0, 1) /* A B C D -> B A D C */
#define S1O32 _MM_SHUFFLE(1, 0, 3, 2) /* A B C D -> C D A B */
v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S23O1));
v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S1O32));
s1 += _mm_cvtsi128_si32(v_s1);
v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S23O1));
v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S1O32));
s2 = _mm_cvtsi128_si32(v_s2);
v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S23O1));
v_s1 = _mm_add_epi32(v_s1, _mm_shuffle_epi32(v_s1, S1O32));
s1 += _mm_cvtsi128_si32(v_s1);
v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S23O1));
v_s2 = _mm_add_epi32(v_s2, _mm_shuffle_epi32(v_s2, S1O32));
s2 = _mm_cvtsi128_si32(v_s2);
#undef S23O1
#undef S1O32
/*
* Reduce.
*/
s1 %= FLETCHER32_MODULE;
s2 %= FLETCHER32_MODULE;
/*
* Reduce.
*/
s1 %= FLETCHER32_MODULE;
s2 %= FLETCHER32_MODULE;
}
}
/*
* Handle leftover data.
*/
@@ -151,6 +155,7 @@ fletcher32_ssse3(uint16_t *sum1, uint16_t *sum2, const uint8_t *data, long len)
if(s1 >= FLETCHER32_MODULE) s1 -= FLETCHER32_MODULE;
s2 %= FLETCHER32_MODULE;
}
/*
* Return the recombined sums.
*/