aaru-dps/Aaru.Checksums.Native
1
0
Fork 0
mirror of https://github.com/aaru-dps/Aaru.Checksums.Native.git synced 2025-12-16 19:24:29 +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

199
adler32_neon.c
View File

@@ -55,118 +55,124 @@ TARGET_WITH_NEON void adler32_neon(uint16_t *sum1, uint16_t *sum2, const uint8_t
*/
uint32_t s1 = *sum1;
uint32_t s2 = *sum2;
/*
* Serially compute s1 & s2, until the data is 16-byte aligned.
*/
if((uintptr_t)data & 15)
{
while((uintptr_t)data & 15)
{
s2 += (s1 += *data++);
--len;
}
if(s1 >= ADLER_MODULE) s1 -= ADLER_MODULE;
s2 %= ADLER_MODULE;
}
/*
* Process the data in blocks.
*/
const unsigned BLOCK_SIZE = 1 << 5;
uint32_t 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;
/*
* Process n blocks of data. At most NMAX data bytes can be
* processed before s2 must be reduced modulo ADLER_MODULE.
* Serially compute s1 & s2, until the data is 16-byte aligned.
*/
#ifdef _MSC_VER
uint32x4_t v_s2 = {.n128_u32 = {0, 0, 0, s1 * n}};
uint32x4_t v_s1 = {.n128_u32 = {0, 0, 0, 0}};
#else
uint32x4_t v_s2 = (uint32x4_t){0, 0, 0, s1 * n};
uint32x4_t v_s1 = (uint32x4_t){0, 0, 0, 0};
#endif
uint16x8_t v_column_sum_1 = vdupq_n_u16(0);
uint16x8_t v_column_sum_2 = vdupq_n_u16(0);
uint16x8_t v_column_sum_3 = vdupq_n_u16(0);
uint16x8_t v_column_sum_4 = vdupq_n_u16(0);
do
if((uintptr_t)data & 15)
{
/*
* Load 32 input bytes.
*/
const uint8x16_t bytes1 = vld1q_u8((uint8_t *)(data));
const uint8x16_t bytes2 = vld1q_u8((uint8_t *)(data + 16));
/*
* Add previous block byte sum to v_s2.
*/
v_s2 = vaddq_u32(v_s2, v_s1);
/*
* Horizontally add the bytes for s1.
*/
v_s1 = vpadalq_u16(v_s1, vpadalq_u8(vpaddlq_u8(bytes1), bytes2));
/*
* Vertically add the bytes for s2.
*/
v_column_sum_1 = vaddw_u8(v_column_sum_1, vget_low_u8(bytes1));
v_column_sum_2 = vaddw_u8(v_column_sum_2, vget_high_u8(bytes1));
v_column_sum_3 = vaddw_u8(v_column_sum_3, vget_low_u8(bytes2));
v_column_sum_4 = vaddw_u8(v_column_sum_4, vget_high_u8(bytes2));
data += BLOCK_SIZE;
while((uintptr_t)data & 15)
{
s2 += (s1 += *data++);
--len;
}
if(s1 >= ADLER_MODULE) s1 -= ADLER_MODULE;
s2 %= ADLER_MODULE;
}
while(--n);
v_s2 = vshlq_n_u32(v_s2, 5);
/*
* Multiply-add bytes by [ 32, 31, 30, ... ] for s2.
*/
uint32_t 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;
/*
* Process n blocks of data. At most NMAX data bytes can be
* processed before s2 must be reduced modulo ADLER_MODULE.
*/
#ifdef _MSC_VER
uint32x4_t v_s2 = {.n128_u32 = {0, 0, 0, s1 * n}};
uint32x4_t v_s1 = {.n128_u32 = {0, 0, 0, 0}};
#else
uint32x4_t v_s2 = (uint32x4_t){0, 0, 0, s1 * n};
uint32x4_t v_s1 = (uint32x4_t){0, 0, 0, 0};
#endif
uint16x8_t v_column_sum_1 = vdupq_n_u16(0);
uint16x8_t v_column_sum_2 = vdupq_n_u16(0);
uint16x8_t v_column_sum_3 = vdupq_n_u16(0);
uint16x8_t v_column_sum_4 = vdupq_n_u16(0);
do
{
/*
* Load 32 input bytes.
*/
const uint8x16_t bytes1 = vld1q_u8((uint8_t *)(data));
const uint8x16_t bytes2 = vld1q_u8((uint8_t *)(data + 16));
/*
* Add previous block byte sum to v_s2.
*/
v_s2 = vaddq_u32(v_s2, v_s1);
/*
* Horizontally add the bytes for s1.
*/
v_s1 = vpadalq_u16(v_s1, vpadalq_u8(vpaddlq_u8(bytes1), bytes2));
/*
* Vertically add the bytes for s2.
*/
v_column_sum_1 = vaddw_u8(v_column_sum_1, vget_low_u8(bytes1));
v_column_sum_2 = vaddw_u8(v_column_sum_2, vget_high_u8(bytes1));
v_column_sum_3 = vaddw_u8(v_column_sum_3, vget_low_u8(bytes2));
v_column_sum_4 = vaddw_u8(v_column_sum_4, vget_high_u8(bytes2));
data += BLOCK_SIZE;
}
while(--n);
v_s2 = vshlq_n_u32(v_s2, 5);
/*
* Multiply-add bytes by [ 32, 31, 30, ... ] for s2.
*/
#ifdef _MSC_VER
#ifdef _M_ARM64
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), neon_ld1m_16((uint16_t[]){32, 31, 30, 29}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), neon_ld1m_16((uint16_t[]){28, 27, 26, 25}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), neon_ld1m_16((uint16_t[]){24, 23, 22, 21}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), neon_ld1m_16((uint16_t[]){20, 19, 18, 17}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), neon_ld1m_16((uint16_t[]){16, 15, 14, 13}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), neon_ld1m_16((uint16_t[]){12, 11, 10, 9}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), neon_ld1m_16((uint16_t[]){8, 7, 6, 5}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), neon_ld1m_16((uint16_t[]){4, 3, 2, 1}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), neon_ld1m_16((uint16_t[]){32, 31, 30, 29}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), neon_ld1m_16((uint16_t[]){28, 27, 26, 25}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), neon_ld1m_16((uint16_t[]){24, 23, 22, 21}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), neon_ld1m_16((uint16_t[]){20, 19, 18, 17}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), neon_ld1m_16((uint16_t[]){16, 15, 14, 13}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), neon_ld1m_16((uint16_t[]){12, 11, 10, 9}));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), neon_ld1m_16((uint16_t[]){8, 7, 6, 5}));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), neon_ld1m_16((uint16_t[]){4, 3, 2, 1}));
#else
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), vld1_u16(((uint16_t[]){32, 31, 30, 29})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), vld1_u16(((uint16_t[]){28, 27, 26, 25})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), vld1_u16(((uint16_t[]){24, 23, 22, 21})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), vld1_u16(((uint16_t[]){20, 19, 18, 17})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), vld1_u16(((uint16_t[]){16, 15, 14, 13})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), vld1_u16(((uint16_t[]){12, 11, 10, 9})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), vld1_u16(((uint16_t[]){8, 7, 6, 5})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), vld1_u16(((uint16_t[]){4, 3, 2, 1})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), vld1_u16(((uint16_t[]){32, 31, 30, 29})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), vld1_u16(((uint16_t[]){28, 27, 26, 25})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), vld1_u16(((uint16_t[]){24, 23, 22, 21})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), vld1_u16(((uint16_t[]){20, 19, 18, 17})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), vld1_u16(((uint16_t[]){16, 15, 14, 13})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), vld1_u16(((uint16_t[]){12, 11, 10, 9})));
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), vld1_u16(((uint16_t[]){8, 7, 6, 5})));
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), vld1_u16(((uint16_t[]){4, 3, 2, 1})));
#endif
#else
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), (uint16x4_t){32, 31, 30, 29});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), (uint16x4_t){28, 27, 26, 25});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), (uint16x4_t){24, 23, 22, 21});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), (uint16x4_t){20, 19, 18, 17});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), (uint16x4_t){16, 15, 14, 13});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), (uint16x4_t){12, 11, 10, 9});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), (uint16x4_t){8, 7, 6, 5});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), (uint16x4_t){4, 3, 2, 1});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_1), (uint16x4_t){32, 31, 30, 29});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_1), (uint16x4_t){28, 27, 26, 25});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_2), (uint16x4_t){24, 23, 22, 21});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_2), (uint16x4_t){20, 19, 18, 17});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_3), (uint16x4_t){16, 15, 14, 13});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_3), (uint16x4_t){12, 11, 10, 9});
v_s2 = vmlal_u16(v_s2, vget_low_u16(v_column_sum_4), (uint16x4_t){8, 7, 6, 5});
v_s2 = vmlal_u16(v_s2, vget_high_u16(v_column_sum_4), (uint16x4_t){4, 3, 2, 1});
#endif
/*
* Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
*/
uint32x2_t t_s1 = vpadd_u32(vget_low_u32(v_s1), vget_high_u32(v_s1));
uint32x2_t t_s2 = vpadd_u32(vget_low_u32(v_s2), vget_high_u32(v_s2));
uint32x2_t s1s2 = vpadd_u32(t_s1, t_s2);
s1 += vget_lane_u32(s1s2, 0);
s2 += vget_lane_u32(s1s2, 1);
/*
* Reduce.
*/
s1 %= ADLER_MODULE;
s2 %= ADLER_MODULE;
/*
* Sum epi32 ints v_s1(s2) and accumulate in s1(s2).
*/
uint32x2_t t_s1 = vpadd_u32(vget_low_u32(v_s1), vget_high_u32(v_s1));
uint32x2_t t_s2 = vpadd_u32(vget_low_u32(v_s2), vget_high_u32(v_s2));
uint32x2_t s1s2 = vpadd_u32(t_s1, t_s2);
s1 += vget_lane_u32(s1s2, 0);
s2 += vget_lane_u32(s1s2, 1);
/*
* Reduce.
*/
s1 %= ADLER_MODULE;
s2 %= ADLER_MODULE;
}
}
/*
* Handle leftover data.
*/
@@ -197,6 +203,7 @@ TARGET_WITH_NEON void adler32_neon(uint16_t *sum1, uint16_t *sum2, const uint8_t
if(s1 >= ADLER_MODULE) s1 -= ADLER_MODULE;
s2 %= ADLER_MODULE;
}
/*
* Return the recombined sums.
*/