diff options
| author | Chris Robinson <[email protected]> | 2014-12-15 15:27:31 -0800 |
|---|---|---|
| committer | Chris Robinson <[email protected]> | 2014-12-15 15:27:31 -0800 |
| commit | a0e7ad493b5f6290e9e664f6ff8e130d10804bec (patch) | |
| tree | 845e9826087890d2ba37acddbc5d9c14d0e14c11 /Alc/mixer_sse2.c | |
| parent | 177dd956866aeed61e549f2163ac8beadd596ba1 (diff) | |
Load samples with _mm_loadu_ps in the cubic SSE resamplers
Diffstat (limited to 'Alc/mixer_sse2.c')
| -rw-r--r-- | Alc/mixer_sse2.c | 9 |
1 files changed, 5 insertions, 4 deletions
diff --git a/Alc/mixer_sse2.c b/Alc/mixer_sse2.c index 68a3a903..1c71b458 100644 --- a/Alc/mixer_sse2.c +++ b/Alc/mixer_sse2.c @@ -96,16 +96,17 @@ const ALfloat *Resample_cubic32_SSE2(const ALfloat *src, ALuint frac, ALuint inc --src; for(i = 0;numsamples-i > 3;i += 4) { - const __m128 val0 = _mm_setr_ps(src[pos_.i[0] ], src[pos_.i[1] ], src[pos_.i[2] ], src[pos_.i[3] ]); - const __m128 val1 = _mm_setr_ps(src[pos_.i[0]+1], src[pos_.i[1]+1], src[pos_.i[2]+1], src[pos_.i[3]+1]); - const __m128 val2 = _mm_setr_ps(src[pos_.i[0]+2], src[pos_.i[1]+2], src[pos_.i[2]+2], src[pos_.i[3]+2]); - const __m128 val3 = _mm_setr_ps(src[pos_.i[0]+3], src[pos_.i[1]+3], src[pos_.i[2]+3], src[pos_.i[3]+3]); + __m128 val0 = _mm_loadu_ps(&src[pos_.i[0]]); + __m128 val1 = _mm_loadu_ps(&src[pos_.i[1]]); + __m128 val2 = _mm_loadu_ps(&src[pos_.i[2]]); + __m128 val3 = _mm_loadu_ps(&src[pos_.i[3]]); __m128 k0 = _mm_load_ps(CubicLUT[frac_.i[0]]); __m128 k1 = _mm_load_ps(CubicLUT[frac_.i[1]]); __m128 k2 = _mm_load_ps(CubicLUT[frac_.i[2]]); __m128 k3 = _mm_load_ps(CubicLUT[frac_.i[3]]); __m128 out; + _MM_TRANSPOSE4_PS(val0, val1, val2, val3); _MM_TRANSPOSE4_PS(k0, k1, k2, k3); /* k0*val0 + k1*val1 + k2*val2 + k3*val3 */ |