diff options
| author | Tim Rowley <timothy.o.rowley@intel.com> | 2017-11-13 18:39:38 -0600 |
|---|---|---|
| committer | Andres Gomez <agomez@igalia.com> | 2017-11-21 18:16:46 +0200 |
| commit | 8edbc8f1091595d677621b407c648512437f1dd7 (patch) | |
| tree | a4c2d658a73b3517c05213593ece6f4398c8c31f | |
| parent | 0f4dfee254e1f054c851a729deb4217cc4e14003 (diff) | |
| download | mesa-8edbc8f1091595d677621b407c648512437f1dd7.tar.gz | |
swr/rast: Faster emulated simd16 permute
Speed up simd16 frontend (default) on avx/avx2 platforms;
fixes performance regression caused by switch to simdlib.
Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit d8489517a572c7e5c5405ebf510db9d20b1e2591)
| -rw-r--r-- | src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl | 34 |
1 files changed, 11 insertions, 23 deletions
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl index a45429f4b6b..fd248b3792b 100644 --- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl +++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl @@ -518,36 +518,24 @@ SIMD_IWRAPPER_2(packus_epi32); // See documentation for _mm256_packus_epi32 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer a, Integer swiz) // return a[swiz[i]] for each 32-bit lane i (int32) { - Integer result; - - // Ugly slow implementation - uint32_t const *pA = reinterpret_cast<uint32_t const*>(&a); - uint32_t const *pSwiz = reinterpret_cast<uint32_t const*>(&swiz); - uint32_t *pResult = reinterpret_cast<uint32_t *>(&result); - - for (uint32_t i = 0; i < SIMD_WIDTH; ++i) - { - pResult[i] = pA[0xF & pSwiz[i]]; - } - - return result; + return castps_si(permute_ps(castsi_ps(a), swiz)); } static SIMDINLINE Float SIMDCALL permute_ps(Float a, Integer swiz) // return a[swiz[i]] for each 32-bit lane i (float) { - Float result; + const auto mask = SIMD256T::set1_epi32(7); - // Ugly slow implementation - float const *pA = reinterpret_cast<float const*>(&a); - uint32_t const *pSwiz = reinterpret_cast<uint32_t const*>(&swiz); - float *pResult = reinterpret_cast<float *>(&result); + auto lolo = SIMD256T::permute_ps(a.v8[0], SIMD256T::and_si(swiz.v8[0], mask)); + auto lohi = SIMD256T::permute_ps(a.v8[1], SIMD256T::and_si(swiz.v8[0], mask)); - for (uint32_t i = 0; i < SIMD_WIDTH; ++i) - { - pResult[i] = pA[0xF & pSwiz[i]]; - } + auto hilo = SIMD256T::permute_ps(a.v8[0], SIMD256T::and_si(swiz.v8[1], mask)); + auto hihi = SIMD256T::permute_ps(a.v8[1], SIMD256T::and_si(swiz.v8[1], mask)); - return result; + return Float + { + SIMD256T::blendv_ps(lolo, lohi, SIMD256T::castsi_ps(SIMD256T::cmpgt_epi32(swiz.v8[0], mask))), + SIMD256T::blendv_ps(hilo, hihi, SIMD256T::castsi_ps(SIMD256T::cmpgt_epi32(swiz.v8[1], mask))), + }; } // All of the 512-bit permute2f128_XX intrinsics do the following: |