diff options
| author | DeeJayLSP <djlsplays@gmail.com> | 2022-11-24 11:45:59 -0300 |
|---|---|---|
| committer | Rémi Verschelde <rverschelde@gmail.com> | 2022-11-25 13:09:04 +0100 |
| commit | 5e4158eb4869427ac13a0fe57e9b688ea4c3b0f1 (patch) | |
| tree | 1a659e532fcba81af33d1f874c65311f093e4535 /thirdparty/embree/common/simd/vfloat16_avx512.h | |
| parent | f16c5b564b569497d04deb965a4fd63b3ea2ab2f (diff) | |
Update embree to 3.13.5
Diffstat (limited to 'thirdparty/embree/common/simd/vfloat16_avx512.h')
| -rw-r--r-- | thirdparty/embree/common/simd/vfloat16_avx512.h | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/thirdparty/embree/common/simd/vfloat16_avx512.h b/thirdparty/embree/common/simd/vfloat16_avx512.h index 9f1e2459c4..75c471cc0c 100644 --- a/thirdparty/embree/common/simd/vfloat16_avx512.h +++ b/thirdparty/embree/common/simd/vfloat16_avx512.h @@ -177,9 +177,10 @@ namespace embree __forceinline vfloat16 abs (const vfloat16& a) { return _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(a),_mm512_set1_epi32(0x7FFFFFFF))); } __forceinline vfloat16 signmsk(const vfloat16& a) { return _mm512_castsi512_ps(_mm512_and_epi32(_mm512_castps_si512(a),_mm512_set1_epi32(0x80000000))); } - __forceinline vfloat16 rcp(const vfloat16& a) { + __forceinline vfloat16 rcp(const vfloat16& a) + { const vfloat16 r = _mm512_rcp14_ps(a); - return _mm512_mul_ps(r, _mm512_fnmadd_ps(r, a, vfloat16(2.0f))); + return _mm512_fmadd_ps(r, _mm512_fnmadd_ps(a, r, vfloat16(1.0)), r); // computes r + r * (1 - a*r) } __forceinline vfloat16 sqr (const vfloat16& a) { return _mm512_mul_ps(a,a); } |