Mirror/aicodix___code
1
0
Fork 0
mirror of https://github.com/aicodix/code.git synced 2026-04-27 22:35:44 +00:00
Code Issues Projects Releases Packages Wiki Activity

added vshuf()

Browse source
This commit is contained in:
Ahmet Inan 2021-06-21 08:18:15 +02:00
commit 778d51ac0f
4 changed files with 160 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

32
avx2.hh
View file

@ -1134,3 +1134,35 @@ inline SIMD<int64_t, 4> vclamp(SIMD<int64_t, 4> x, int64_t a, int64_t b)
return tmp;
}
template <>
inline SIMD<uint8_t, 32> vshuf(SIMD<uint8_t, 32> a, SIMD<uint8_t, 32> b)
{
SIMD<uint8_t, 32> tmp;
__m256i c = _mm256_sub_epi8(b.m, _mm256_set1_epi8(16));
__m256i d = _mm256_or_si256(b.m, _mm256_cmpgt_epi8(b.m, _mm256_set1_epi8(15)));
__m256i e = _mm256_shuffle_epi8(_mm256_permute2x128_si256(a.m, a.m, 0), d);
__m256i f = _mm256_shuffle_epi8(_mm256_permute2x128_si256(a.m, a.m, 17), c);
tmp.m = _mm256_or_si256(e, f);
return tmp;
}
template <>
inline SIMD<int8_t, 32> vshuf(SIMD<int8_t, 32> a, SIMD<uint8_t, 32> b)
{
SIMD<int8_t, 32> tmp;
__m256i c = _mm256_sub_epi8(b.m, _mm256_set1_epi8(16));
__m256i d = _mm256_or_si256(b.m, _mm256_cmpgt_epi8(b.m, _mm256_set1_epi8(15)));
__m256i e = _mm256_shuffle_epi8(_mm256_permute2x128_si256(a.m, a.m, 0), d);
__m256i f = _mm256_shuffle_epi8(_mm256_permute2x128_si256(a.m, a.m, 17), c);
tmp.m = _mm256_or_si256(e, f);
return tmp;
}
template <>
inline SIMD<float, 8> vshuf(SIMD<float, 8> a, SIMD<uint32_t, 8> b)
{
SIMD<float, 8> tmp;
tmp.m = _mm256_permutevar8x32_ps(a.m, b.m);
return tmp;
}

22
neon.hh
View file

@ -938,3 +938,25 @@ inline SIMD<int32_t, 4> vclamp(SIMD<int32_t, 4> x, int32_t a, int32_t b)
return tmp;
}
template <>
inline SIMD<uint8_t, 16> vshuf(SIMD<uint8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<uint8_t, 16> tmp;
uint8x8x2_t c { vget_low_u8(a.m), vget_high_u8(a.m) };
uint8x8_t d = vtbl2_u8(c, vget_low_u8(b.m));
uint8x8_t e = vtbl2_u8(c, vget_high_u8(b.m));
tmp.m = vcombine_u8(d, e);
return tmp;
}
template <>
inline SIMD<int8_t, 16> vshuf(SIMD<int8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<int8_t, 16> tmp;
int8x8x2_t c { vget_low_s8(a.m), vget_high_s8(a.m) };
int8x8_t d = vtbl2_s8(c, vget_low_s8((int8x16_t)b.m));
int8x8_t e = vtbl2_s8(c, vget_high_s8((int8x16_t)b.m));
tmp.m = vcombine_s8(d, e);
return tmp;
}

90
simd.hh
View file

@ -1387,6 +1387,96 @@ static inline SIMD<int64_t, WIDTH> vsign(SIMD<int64_t, WIDTH> a, SIMD<int64_t, W
return tmp;
}
template <int WIDTH>
static inline SIMD<uint8_t, WIDTH> vshuf(SIMD<uint8_t, WIDTH> a, SIMD<uint8_t, WIDTH> b)
{
SIMD<uint8_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<uint16_t, WIDTH> vshuf(SIMD<uint16_t, WIDTH> a, SIMD<uint16_t, WIDTH> b)
{
SIMD<uint16_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<uint32_t, WIDTH> vshuf(SIMD<uint32_t, WIDTH> a, SIMD<uint32_t, WIDTH> b)
{
SIMD<uint32_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<uint64_t, WIDTH> vshuf(SIMD<uint64_t, WIDTH> a, SIMD<uint64_t, WIDTH> b)
{
SIMD<uint64_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<int8_t, WIDTH> vshuf(SIMD<int8_t, WIDTH> a, SIMD<uint8_t, WIDTH> b)
{
SIMD<int8_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<int16_t, WIDTH> vshuf(SIMD<int16_t, WIDTH> a, SIMD<uint16_t, WIDTH> b)
{
SIMD<int16_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<int32_t, WIDTH> vshuf(SIMD<int32_t, WIDTH> a, SIMD<uint32_t, WIDTH> b)
{
SIMD<int32_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<int64_t, WIDTH> vshuf(SIMD<int64_t, WIDTH> a, SIMD<uint64_t, WIDTH> b)
{
SIMD<int64_t, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0;
return tmp;
}
template <int WIDTH>
static inline SIMD<float, WIDTH> vshuf(SIMD<float, WIDTH> a, SIMD<uint32_t, WIDTH> b)
{
SIMD<float, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0.f;
return tmp;
}
template <int WIDTH>
static inline SIMD<double, WIDTH> vshuf(SIMD<double, WIDTH> a, SIMD<uint64_t, WIDTH> b)
{
SIMD<double, WIDTH> tmp;
for (int i = 0; i < WIDTH; ++i)
tmp.v[i] = b.v[i] < WIDTH ? a.v[b.v[i]] : 0.;
return tmp;
}
#if 1
#ifdef __AVX2__
#include "avx2.hh"

16
sse4_1.hh
View file

@ -1127,3 +1127,19 @@ inline SIMD<int32_t, 4> vclamp(SIMD<int32_t, 4> x, int32_t a, int32_t b)
return tmp;
}
template <>
inline SIMD<uint8_t, 16> vshuf(SIMD<uint8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<uint8_t, 16> tmp;
tmp.m = _mm_shuffle_epi8(a.m, b.m);
return tmp;
}
template <>
inline SIMD<int8_t, 16> vshuf(SIMD<int8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<int8_t, 16> tmp;
tmp.m = _mm_shuffle_epi8(a.m, b.m);
return tmp;
}