[Mesa-dev] [PATCH] swr/rast: Use gather instruction for i32gather_ps on simd16/avx512

[Cherniak, Bruce]

Reviewed-by: Bruce Cherniak <bruce.cherniak at intel.com> 

> On Nov 13, 2017, at 8:03 PM, Tim Rowley <timothy.o.rowley at intel.com> wrote:
> 
> Speed up avx512 platforms; fixes performance regression caused
> by swithc to simdlib.
> 
> Cc: mesa-stable at lists.freedesktop.org
> ---
> .../drivers/swr/rasterizer/common/simdlib_512_avx512.inl     | 12 +-----------
> 1 file changed, 1 insertion(+), 11 deletions(-)
> 
> diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
> index 95e4c31909..c13b9f616a 100644
> --- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
> +++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
> @@ -484,17 +484,7 @@ SIMD_WRAPPER_2(unpacklo_ps);
> template<ScaleFactor ScaleT>
> static SIMDINLINE Float SIMDCALL i32gather_ps(float const* p, Integer idx) // return *(float*)(((int8*)p) + (idx * ScaleT))
> {
> -    uint32_t *pOffsets = (uint32_t*)&idx;
> -    Float vResult;
> -    float* pResult = (float*)&vResult;
> -    for (uint32_t i = 0; i < SIMD_WIDTH; ++i)
> -    {
> -        uint32_t offset = pOffsets[i];
> -        offset = offset * static_cast<uint32_t>(ScaleT);
> -        pResult[i] = *(float const*)(((uint8_t const*)p + offset));
> -    }
> -
> -    return vResult;
> +    return _mm512_i32gather_ps(idx, p, static_cast<int>(ScaleT));
> }
> 
> static SIMDINLINE Float SIMDCALL load1_ps(float const *p)  // return *p    (broadcast 1 value to all elements)
> -- 
> 2.14.1
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev