diff options
| author | Sergey Sharybin <sergey.vfx@gmail.com> | 2018-09-18 12:49:08 +0300 |
|---|---|---|
| committer | Sergey Sharybin <sergey.vfx@gmail.com> | 2018-09-18 12:49:08 +0300 |
| commit | 5aa5066c3ba8f11ad5d435e4f571b3aabd31cb4e (patch) | |
| tree | f4ffa9aea6baf05cbbbaf11ae67a13848bdc62b1 /intern | |
| parent | eadf0f3d705c7f3cd2ba9e14f5568c421d88cb2a (diff) | |
| parent | d04eb330e873488d8b7cbe49da5600d6f64ea6db (diff) | |
Merge branch 'master' into blender2.8
Diffstat (limited to 'intern')
| -rw-r--r-- | intern/cycles/util/util_avxb.h | 18 | ||||
| -rw-r--r-- | intern/cycles/util/util_avxf.h | 6 |
2 files changed, 20 insertions, 4 deletions
diff --git a/intern/cycles/util/util_avxb.h b/intern/cycles/util/util_avxb.h index b6d77857c6f..4add69bcb91 100644 --- a/intern/cycles/util/util_avxb.h +++ b/intern/cycles/util/util_avxb.h @@ -49,7 +49,7 @@ struct avxb //////////////////////////////////////////////////////////////////////////////// __forceinline avxb( FalseTy ) : m256(_mm256_setzero_ps()) {} - __forceinline avxb( TrueTy ) : m256(_mm256_castsi256_ps(_mm256_cmpeq_epi32(_mm256_setzero_si256(), _mm256_setzero_si256()))) {} + __forceinline avxb( TrueTy ) : m256(_mm256_castsi256_ps(_mm256_set1_epi32(-1))) {} //////////////////////////////////////////////////////////////////////////////// /// Array Access @@ -86,7 +86,21 @@ __forceinline const avxb operator ^=( avxb& a, const avxb& b ) { return a = a ^ //////////////////////////////////////////////////////////////////////////////// __forceinline const avxb operator !=( const avxb& a, const avxb& b ) { return _mm256_xor_ps(a, b); } -__forceinline const avxb operator ==( const avxb& a, const avxb& b ) { return _mm256_castsi256_ps(_mm256_cmpeq_epi32(a, b)); } +__forceinline const avxb operator ==( const avxb& a, const avxb& b ) +{ +#ifdef __KERNEL_AVX2__ + return _mm256_castsi256_ps(_mm256_cmpeq_epi32(a, b)); +#else + __m128i a_lo = _mm_castps_si128(_mm256_extractf128_ps(a, 0)); + __m128i a_hi = _mm_castps_si128(_mm256_extractf128_ps(a, 1)); + __m128i b_lo = _mm_castps_si128(_mm256_extractf128_ps(b, 0)); + __m128i b_hi = _mm_castps_si128(_mm256_extractf128_ps(b, 1)); + __m128i c_lo = _mm_cmpeq_epi32(a_lo, b_lo); + __m128i c_hi = _mm_cmpeq_epi32(a_hi, b_hi); + __m256i result = _mm256_insertf128_si256(_mm256_castsi128_si256(c_lo), c_hi, 1); + return _mm256_castsi256_ps(result); +#endif +} __forceinline const avxb select( const avxb& m, const avxb& t, const avxb& f ) { #if defined(__KERNEL_SSE41__) diff --git a/intern/cycles/util/util_avxf.h b/intern/cycles/util/util_avxf.h index 5596702ca20..0c15ba5cbbd 100644 --- a/intern/cycles/util/util_avxf.h +++ b/intern/cycles/util/util_avxf.h @@ -214,17 +214,19 @@ __forceinline const avxf nmadd(const avxf& a, const avxf& b, const avxf& c) { #endif } __forceinline const avxf msub(const avxf& a, const avxf& b, const avxf& c) { +#ifdef __KERNEL_AVX2__ return _mm256_fmsub_ps(a, b, c); +#else + return (a*b) - c; +#endif } //////////////////////////////////////////////////////////////////////////////// /// Comparison Operators //////////////////////////////////////////////////////////////////////////////// -#ifdef __KERNEL_AVX2__ __forceinline const avxb operator <=(const avxf& a, const avxf& b) { return _mm256_cmp_ps(a.m256, b.m256, _CMP_LE_OS); } -#endif #endif |