mirrors/pipewire
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pipewire/pipewire.git synced 2026-04-24 06:46:37 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

resample: use independent accumulators for SSE and SSSE3

Browse source
This commit is contained in:
Wim Taymans 2026-04-23 09:18:08 +02:00
parent 37f9f7773c
commit 75e432a49a
2 changed files with 22 additions and 41 deletions
Download patch file Download diff file Expand all files Collapse all files

38
spa/plugins/audioconvert/resample-native-sse.c
View file

@ -9,43 +9,23 @@
static inline void inner_product_sse(float *d, const float * SPA_RESTRICT s,
const float * SPA_RESTRICT taps, uint32_t n_taps)
{
__m128 sum = _mm_setzero_ps();
uint32_t i = 0;
#if 0
uint32_t unrolled = n_taps & ~15;
__m128 sum[2] = { _mm_setzero_ps(), _mm_setzero_ps() };
uint32_t i;
for (i = 0; i < unrolled; i += 16) {
sum = _mm_add_ps(sum,
for (i = 0; i < n_taps; i += 8) {
sum[0] = _mm_add_ps(sum[0],
_mm_mul_ps(
_mm_loadu_ps(s + i + 0),
_mm_load_ps(taps + i + 0)));
sum = _mm_add_ps(sum,
_mm_mul_ps(
_mm_loadu_ps(s + i + 4),
_mm_load_ps(taps + i + 4)));
sum = _mm_add_ps(sum,
_mm_mul_ps(
_mm_loadu_ps(s + i + 8),
_mm_load_ps(taps + i + 8)));
sum = _mm_add_ps(sum,
_mm_mul_ps(
_mm_loadu_ps(s + i + 12),
_mm_load_ps(taps + i + 12)));
}
#endif
for (; i < n_taps; i += 8) {
sum = _mm_add_ps(sum,
_mm_mul_ps(
_mm_loadu_ps(s + i + 0),
_mm_load_ps(taps + i + 0)));
sum = _mm_add_ps(sum,
sum[1] = _mm_add_ps(sum[1],
_mm_mul_ps(
_mm_loadu_ps(s + i + 4),
_mm_load_ps(taps + i + 4)));
}
sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum));
sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55));
_mm_store_ss(d, sum);
sum[0] = _mm_add_ps(sum[0], sum[1]);
sum[0] = _mm_add_ps(sum[0], _mm_movehl_ps(sum[0], sum[0]));
sum[0] = _mm_add_ss(sum[0], _mm_shuffle_ps(sum[0], sum[0], 0x55));
_mm_store_ss(d, sum[0]);
}
static inline void inner_product_ip_sse(float *d, const float * SPA_RESTRICT s,