Dima353/openal-soft
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix matrix multiply used by the SSE cubic resampler

Browse Source 
Also remove the 4-sample loop. It's not terribly effective.
This commit is contained in:
Chris Robinson
2012-09-18 18:08:57 -07:00
parent f0732426f2
commit d1c93525ed
1 changed files with 22 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Alc/mixer_sse.c
+22 -63
View File
@@ -60,84 +60,43 @@ void Resample_lerp32_SSE(const ALfloat *data, ALuint frac,
}
void Resample_cubic32_SSE(const ALfloat *data, ALuint frac,
ALuint increment, ALuint NumChannels, ALfloat *RESTRICT OutBuffer,
ALuint increment, ALuint channels, ALfloat *RESTRICT OutBuffer,
ALuint BufferSize)
{
/* Cubic interpolation mainly consists of a matrix4 * vector4 operation,
* followed by scalars being applied to the resulting elements before all
* four are added together for the final sample. */
static const __m128 matrix[4] = {
{ -0.5, 1.0f, -0.5f, 0.0f },
{ 1.5, -2.5f, 0.0f, 1.0f },
{ -1.5, 2.0f, 0.5f, 0.0f },
{ 0.5, -0.5f, 0.0f, 0.0f },
{ -0.5f, 1.0f, -0.5f, 0.0f },
{ 1.5f, -2.5f, 0.0f, 1.0f },
{ -1.5f, 2.0f, 0.5f, 0.0f },
{ 0.5f, -0.5f, 0.0f, 0.0f },
};
ALIGN(16) float value[4];
ALuint pos = 0;
ALuint i, j;
ALuint i;
for(i = 0;i < BufferSize+1-3;i+=4)
for(i = 0;i < BufferSize+1;i++)
{
__m128 result, final[4];
for(j = 0;j < 4;j++)
{
__m128 val4, s;
ALfloat mu;
val4 = _mm_set_ps(data[(pos-1)*NumChannels],
data[(pos )*NumChannels],
data[(pos+1)*NumChannels],
data[(pos+2)*NumChannels]);
mu = frac * (1.0f/FRACTIONONE);
s = _mm_set_ps(1.0f, mu, mu*mu, mu*mu*mu);
/* result = matrix * val4 */
result = _mm_mul_ps(val4, matrix[0]) ;
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[1]));
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[2]));
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[3]));
/* final[j] = result * { mu^0, mu^1, mu^2, mu^3 } */
final[j] = _mm_mul_ps(result, s);
frac += increment;
pos += frac>>FRACTIONBITS;
frac &= FRACTIONMASK;
}
/* Transpose the final "matrix" so adding the rows will give the four
* samples. TODO: Is this faster than doing..
* _mm_store_ps(value, result);
* OutBuffer[i] = value[0] + value[1] + value[2] + value[3];
* ..for each sample?
*/
_MM_TRANSPOSE4_PS(final[0], final[1], final[2], final[3]);
result = _mm_add_ps(_mm_add_ps(final[0], final[1]),
_mm_add_ps(final[2], final[3]));
_mm_store_ps(&OutBuffer[i], result);
}
for(;i < BufferSize+1;i++)
{
__m128 val4, s, result;
__m128 res1, res2;
ALfloat mu;
val4 = _mm_set_ps(data[(pos-1)*NumChannels],
data[(pos )*NumChannels],
data[(pos+1)*NumChannels],
data[(pos+2)*NumChannels]);
/* matrix * { samples } */
res1 = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(data[(pos-1)*channels]), matrix[0]),
_mm_mul_ps(_mm_set1_ps(data[(pos )*channels]), matrix[1]));
res2 = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(data[(pos+1)*channels]), matrix[2]),
_mm_mul_ps(_mm_set1_ps(data[(pos+2)*channels]), matrix[3]));
res1 = _mm_add_ps(res1, res2);
/* res1 * { mu^3, mu^2, mu^1, mu^0 } */
mu = frac * (1.0f/FRACTIONONE);
s = _mm_set_ps(1.0f, mu, mu*mu, mu*mu*mu);
/* result = matrix * val4 */
result = _mm_mul_ps(val4, matrix[0]) ;
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[1]));
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[2]));
result = _mm_add_ps(result, _mm_mul_ps(val4, matrix[3]));
/* value = result * { mu^0, mu^1, mu^2, mu^3 } */
_mm_store_ps(value, _mm_mul_ps(result, s));
value[0] = mu*mu*mu;
value[1] = mu*mu;
value[2] = mu;
value[3] = 1.0f;
res1 = _mm_mul_ps(res1, _mm_load_ps(value));
_mm_store_ps(value, res1);
OutBuffer[i] = value[0] + value[1] + value[2] + value[3];
frac += increment;