Mirrors/bx
1
0
Fork 0
mirror of https://github.com/bkaradzic/bx.git synced 2026-02-17 20:52:37 +01:00
Code Issues Packages Projects Releases Wiki Activity

Neon: Improved rcp, rsqrt accuracy. Issue #298.

Browse Source
This commit is contained in:
Бранимир Караџић
2023-04-13 21:18:36 -07:00
parent 4e67e34c4b
commit 7bbda806d9
2 changed files with 33 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
include/bx/inline/simd128_neon.inl
View File

@@ -536,7 +536,13 @@ BX_SIMD128_IMPLEMENT_TEST(yzw, yzww);
template<>
BX_SIMD_FORCE_INLINE simd128_neon_t simd_rcp(simd128_neon_t _a)
{
return simd_rcp_ni(_a);
const simd128_neon_t tmp0 = simd_rcp_est(_a);
const simd128_neon_t tmp1 = vrecpsq_f32(_a, tmp0);
const simd128_neon_t tmp2 = simd_mul(tmp0, tmp1);
const simd128_neon_t tmp3 = vrecpsq_f32(_a, tmp2);
const simd128_neon_t result = simd_mul(tmp2, tmp3);
return result;
}
template<>
@@ -620,13 +626,21 @@ BX_SIMD128_IMPLEMENT_TEST(yzw, yzww);
template<>
BX_SIMD_FORCE_INLINE simd128_neon_t simd_rsqrt(simd128_neon_t _a)
{
return simd_rsqrt_ni(_a);
const simd128_neon_t tmp0 = simd_rsqrt_est(_a);
const simd128_neon_t tmp1 = simd_mul(_a, tmp0);
const simd128_neon_t tmp2 = vrsqrtsq_f32(tmp1, tmp0);
const simd128_neon_t tmp3 = simd_mul(tmp0, tmp2);
const simd128_neon_t tmp4 = simd_mul(_a, tmp3);
const simd128_neon_t tmp5 = vrsqrtsq_f32(tmp4, tmp3);
const simd128_neon_t result = simd_mul(tmp3, tmp5);
return result;
}
template<>
BX_SIMD_FORCE_INLINE simd128_neon_t simd_rsqrt_nr(simd128_neon_t _a)
{
return simd_rsqrt_nr_ni(_a);
return simd_rsqrt(_a);
}
template<>
@@ -636,15 +650,18 @@ BX_SIMD128_IMPLEMENT_TEST(yzw, yzww);
}
template<>
BX_SIMD_FORCE_INLINE simd128_neon_t simd_sqrt_nr(simd128_neon_t _a)
BX_SIMD_FORCE_INLINE simd128_neon_t simd_sqrt(simd128_neon_t _a)
{
return simd_sqrt_nr_ni(_a);
const simd128_neon_t rsqrt = simd_rsqrt(_a);
const simd128_neon_t result = simd_rcp(rsqrt);
return result;
}
template<>
BX_SIMD_FORCE_INLINE simd128_neon_t simd_sqrt(simd128_neon_t _a)
BX_SIMD_FORCE_INLINE simd128_neon_t simd_sqrt_nr(simd128_neon_t _a)
{
return simd_sqrt_nr_ni(_a);
return simd_sqrt(_a);
}
template<>