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: improve sse and sse2 mixer

Browse source 
Add mixer benchmark.
Improve sse and sse2 mixer function by removin some read/write to the
temporary buffer at the expense of more scattered reads.
This commit is contained in:
Wim Taymans 2022-07-10 21:13:12 +02:00
parent 240d212822
commit 7a0445cb28
5 changed files with 339 additions and 95 deletions
Download patch file Download diff file Expand all files Collapse all files

93
spa/plugins/audiomixer/mix-ops-sse.c
View file

@ -32,58 +32,57 @@
#include <xmmintrin.h>
static inline void mix_2(float * dst, const float * SPA_RESTRICT src, uint32_t n_samples)
{
uint32_t n, unrolled;
__m128 in1[4], in2[4];
if (SPA_LIKELY(SPA_IS_ALIGNED(src, 16) &&
SPA_IS_ALIGNED(dst, 16)))
unrolled = n_samples & ~15;
else
unrolled = 0;
for (n = 0; n < unrolled; n += 16) {
in1[0] = _mm_load_ps(&dst[n+ 0]);
in1[1] = _mm_load_ps(&dst[n+ 4]);
in1[2] = _mm_load_ps(&dst[n+ 8]);
in1[3] = _mm_load_ps(&dst[n+12]);
in2[0] = _mm_load_ps(&src[n+ 0]);
in2[1] = _mm_load_ps(&src[n+ 4]);
in2[2] = _mm_load_ps(&src[n+ 8]);
in2[3] = _mm_load_ps(&src[n+12]);
in1[0] = _mm_add_ps(in1[0], in2[0]);
in1[1] = _mm_add_ps(in1[1], in2[1]);
in1[2] = _mm_add_ps(in1[2], in2[2]);
in1[3] = _mm_add_ps(in1[3], in2[3]);
_mm_store_ps(&dst[n+ 0], in1[0]);
_mm_store_ps(&dst[n+ 4], in1[1]);
_mm_store_ps(&dst[n+ 8], in1[2]);
_mm_store_ps(&dst[n+12], in1[3]);
}
for (; n < n_samples; n++) {
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_sse(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));
if (n_src == 0) {
memset(dst, 0, n_samples * sizeof(float));
} else if (n_src == 1) {
if (dst != src[0])
spa_memcpy(dst, src[0], n_samples * sizeof(float));
} else {
uint32_t n, i, unrolled;
__m128 in[4];
const float **s = (const float **)src;
float *d = dst;
bool aligned = true;
for (i = 1; i < n_src; i++) {
mix_2(dst, src[i], n_samples * ops->n_channels);
if (SPA_UNLIKELY(!SPA_IS_ALIGNED(dst, 16)))
aligned = false;
else {
for (i = 0; i < n_src && aligned; i++) {
if (SPA_UNLIKELY(!SPA_IS_ALIGNED(src[i], 16)))
aligned = false;
}
}
unrolled = aligned ? n_samples & ~15 : 0;
for (n = 0; n < unrolled; n += 16) {
in[0] = _mm_load_ps(&s[0][n+ 0]);
in[1] = _mm_load_ps(&s[0][n+ 4]);
in[2] = _mm_load_ps(&s[0][n+ 8]);
in[3] = _mm_load_ps(&s[0][n+12]);
for (i = 1; i < n_src; i++) {
in[0] = _mm_add_ps(in[0], _mm_load_ps(&s[i][n+ 0]));
in[1] = _mm_add_ps(in[1], _mm_load_ps(&s[i][n+ 4]));
in[2] = _mm_add_ps(in[2], _mm_load_ps(&s[i][n+ 8]));
in[3] = _mm_add_ps(in[3], _mm_load_ps(&s[i][n+12]));
}
_mm_store_ps(&d[n+ 0], in[0]);
_mm_store_ps(&d[n+ 4], in[1]);
_mm_store_ps(&d[n+ 8], in[2]);
_mm_store_ps(&d[n+12], in[3]);
}
for (; n < n_samples; n++) {
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]);
}
}
}