swr/rast: Use gather instruction for i32gather_ps on simd16/avx512

Speed up avx512 platforms; fixes performance regression caused
by swithc to simdlib.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>
Cc: mesa-stable@lists.freedesktop.org
(cherry picked from commit 439904847e)
This commit is contained in:
Tim Rowley 2017-11-13 15:11:21 -06:00 committed by Emil Velikov
parent f3caa303cf
commit c798200543

View file

@ -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)