Mirror/aicodix___code
1
0
Fork 0
mirror of https://github.com/aicodix/code.git synced 2026-04-27 14:30:36 +00:00
Code Issues Projects Releases Packages Wiki Activity

use faster vqtbl1q_u8 on ARM64

Browse source
This commit is contained in:
Ahmet Inan 2023-05-05 08:37:13 +02:00
commit 2b7645ced3
2 changed files with 13 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

8
neon.hh
View file

@ -953,10 +953,14 @@ template <>
inline SIMD<uint8_t, 16> vshuf(SIMD<uint8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<uint8_t, 16> tmp;
#ifdef __aarch64__
tmp.m = vqtbl1q_u8(a.m, b.m);
#else
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);
#endif
return tmp;
}
@ -964,10 +968,14 @@ template <>
inline SIMD<int8_t, 16> vshuf(SIMD<int8_t, 16> a, SIMD<uint8_t, 16> b)
{
SIMD<int8_t, 16> tmp;
#ifdef __aarch64__
tmp.m = vqtbl1q_s8(a.m, b.m);
#else
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);
#endif
return tmp;
}

6
rotate.hh
View file

@ -51,11 +51,15 @@ public:
{
if (s < 0)
s += WIDTH;
TYPE ret;
#ifdef __aarch64__
ret.m = vqtbl1q_s8(a.m, vunsigned(rot[s]).m);
#else
int8x8x2_t b { vget_low_s8(a.m), vget_high_s8(a.m) };
int8x8_t c = vtbl2_s8(b, vget_low_s8(rot[s].m));
int8x8_t d = vtbl2_s8(b, vget_high_s8(rot[s].m));
TYPE ret;
ret.m = vcombine_s8(c, d);
#endif
return ret;
}
};