mirrors/pipewire
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pipewire/pipewire.git synced 2025-11-02 09:01:50 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

audiomixer: optimize avx mixer some more

Browse source 
Add avx mixer to test and benchmark
Rework and unroll the avx mixer some more.
The SSE one is 10 times faster than the C one, The AVX is 20 times
faster. The SSE2 function is 5 times faster than the C one.
This commit is contained in:
Wim Taymans 2022-07-10 23:13:24 +02:00
parent 23984f8790
commit 8fe83e5304
5 changed files with 77 additions and 58 deletions
Download patch file Download diff file Expand all files Collapse all files

81
spa/plugins/audiomixer/mix-ops-avx.c
View file

@ -86,50 +86,59 @@ static inline void mix_4(float * dst,
static inline void mix_2(float * dst, const float * SPA_RESTRICT src, uint32_t n_samples)
{
uint32_t n, unrolled;
if (SPA_IS_ALIGNED(src, 32) &&
SPA_IS_ALIGNED(dst, 32))
unrolled = n_samples & ~15;
else
unrolled = 0;
for (n = 0; n < unrolled; n += 16) {
__m256 in1[2], in2[2];
in1[0] = _mm256_load_ps(&dst[n + 0]);
in1[1] = _mm256_load_ps(&dst[n + 8]);
in2[0] = _mm256_load_ps(&src[n + 0]);
in2[1] = _mm256_load_ps(&src[n + 8]);
in1[0] = _mm256_add_ps(in1[0], in2[0]);
in1[1] = _mm256_add_ps(in1[1], in2[1]);
_mm256_store_ps(&dst[n + 0], in1[0]);
_mm256_store_ps(&dst[n + 8], in1[1]);
}
for (; n < n_samples; n++) {
__m128 in1[1], in2[1];
in1[0] = _mm_load_ss(&dst[n]),
in2[0] = _mm_load_ss(&src[n]),
in1[0] = _mm_add_ss(in1[0], in2[0]);
_mm_store_ss(&dst[n], in1[0]);
}
}
void
mix_f32_avx(struct mix_ops *ops, void * SPA_RESTRICT dst, const void * SPA_RESTRICT src[],
uint32_t n_src, uint32_t n_samples)
{
uint32_t i;
n_samples *= ops->n_channels;
if (n_src == 0)
memset(dst, 0, n_samples * ops->n_channels * sizeof(float));
else if (dst != src[0])
spa_memcpy(dst, src[0], n_samples * ops->n_channels * sizeof(float));
else if (n_src == 1) {
if (dst != src[0])
spa_memcpy(dst, src[0], n_samples * sizeof(float));
} else {
uint32_t i, n, unrolled;
const float **s = (const float **)src;
float *d = dst;
for (i = 1; i + 2 < n_src; i += 3)
mix_4(dst, src[i], src[i + 1], src[i + 2], n_samples);
for (; i < n_src; i++)
mix_2(dst, src[i], n_samples * ops->n_channels);
if (SPA_LIKELY(SPA_IS_ALIGNED(dst, 32))) {
unrolled = n_samples & ~31;
for (i = 0; i < n_src; i++) {
if (SPA_UNLIKELY(!SPA_IS_ALIGNED(src[i], 32))) {
unrolled = 0;
break;
}
}
} else
unrolled = 0;
for (n = 0; n < unrolled; n += 32) {
__m256 in[4];
in[0] = _mm256_load_ps(&s[0][n + 0]);
in[1] = _mm256_load_ps(&s[0][n + 8]);
in[2] = _mm256_load_ps(&s[0][n + 16]);
in[3] = _mm256_load_ps(&s[0][n + 24]);
for (i = 1; i < n_src; i++) {
in[0] = _mm256_add_ps(in[0], _mm256_load_ps(&s[i][n + 0]));
in[1] = _mm256_add_ps(in[1], _mm256_load_ps(&s[i][n + 8]));
in[2] = _mm256_add_ps(in[2], _mm256_load_ps(&s[i][n + 16]));
in[3] = _mm256_add_ps(in[3], _mm256_load_ps(&s[i][n + 24]));
}
_mm256_store_ps(&d[n + 0], in[0]);
_mm256_store_ps(&d[n + 8], in[1]);
_mm256_store_ps(&d[n + 16], in[2]);
_mm256_store_ps(&d[n + 24], in[3]);
}
for (; n < n_samples; n++) {
__m128 in[1];
in[0] = _mm_load_ss(&s[0][n]);
for (i = 1; i < n_src; i++)
in[0] = _mm_add_ss(in[0], _mm_load_ss(&s[i][n]));
_mm_store_ss(&d[n], in[0]);
}
}
}