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
pipewire/spa/plugins/audioconvert/resample-native-ssse3.c

155 lines
4.9 KiB
C
Raw Normal View History

treewide: use SPDX tags to specify copyright information SPDX tags make the licensing information easy to understand and clear, and they are machine parseable. See https://spdx.dev for more information.
2023-02-08 18:12:00 +01:00
/* Spa */
/* SPDX-FileCopyrightText: Copyright © 2019 Wim Taymans */
/* SPDX-License-Identifier: MIT */
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
audioconvert: handle more optimizations Compile an optimized library for the given CPU with the right flags, then link it with the main library.
2019-03-27 17:58:48 +01:00
#include "resample-native-impl.h"
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
#include <tmmintrin.h>
audioconvert: refactor peaks resampler Use common code in macro and generate arch specific version. Compile with -Ofast to optimize some fmaxf calls.
2022-09-07 16:00:31 +02:00
static inline void inner_product_ssse3(float *d, const float * SPA_RESTRICT s,
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
const float * SPA_RESTRICT taps, uint32_t n_taps)
{
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
__m128 sum[2] = { _mm_setzero_ps(), _mm_setzero_ps() };
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
__m128 t0, t1;
uint32_t i;
switch (SPA_PTR_ALIGNMENT(s, 16)) {
case 0:
for (i = 0; i < n_taps; i += 8) {
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[0] = _mm_add_ps(sum[0],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(
_mm_load_ps(s + i + 0),
_mm_load_ps(taps + i + 0)));
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[1] = _mm_add_ps(sum[1],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(
_mm_load_ps(s + i + 4),
_mm_load_ps(taps + i + 4)));
}
break;
case 4:
t0 = _mm_load_ps(s - 1);
for (i = 0; i < n_taps; i += 8) {
t1 = _mm_load_ps(s + i + 3);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 4);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[0] = _mm_add_ps(sum[0],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 0)));
t0 = t1;
t1 = _mm_load_ps(s + i + 7);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 4);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[1] = _mm_add_ps(sum[1],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 4)));
t0 = t1;
}
break;
case 8:
t0 = _mm_load_ps(s - 2);
for (i = 0; i < n_taps; i += 8) {
t1 = _mm_load_ps(s + i + 2);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 8);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[0] = _mm_add_ps(sum[0],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 0)));
t0 = t1;
t1 = _mm_load_ps(s + i + 6);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 8);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[1] = _mm_add_ps(sum[1],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 4)));
t0 = t1;
}
break;
case 12:
t0 = _mm_load_ps(s - 3);
for (i = 0; i < n_taps; i += 8) {
t1 = _mm_load_ps(s + i + 1);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 12);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[0] = _mm_add_ps(sum[0],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 0)));
t0 = t1;
t1 = _mm_load_ps(s + i + 5);
t0 = (__m128)_mm_alignr_epi8((__m128i)t1, (__m128i)t0, 12);
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[1] = _mm_add_ps(sum[1],
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
_mm_mul_ps(t0, _mm_load_ps(taps + i + 4)));
t0 = t1;
}
break;
}
resample: use independent accumulators for SSE and SSSE3
2026-04-23 09:18:08 +02:00
sum[0] = _mm_add_ps(sum[0], sum[1]);
sum[0] = _mm_add_ps(sum[0], _mm_movehdup_ps(sum[0]));
sum[0] = _mm_add_ss(sum[0], _mm_movehl_ps(sum[0], sum[0]));
_mm_store_ss(d, sum[0]);
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
}
audioconvert: refactor peaks resampler Use common code in macro and generate arch specific version. Compile with -Ofast to optimize some fmaxf calls.
2022-09-07 16:00:31 +02:00
static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s,
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
uint32_t n_taps)
{
resample: implement inter for ssse3
2026-04-22 18:28:25 +02:00
__m128 sum[2] = { _mm_setzero_ps(), _mm_setzero_ps() };
__m128 r0, r1, r;
audioconvert: handle more optimizations Compile an optimized library for the given CPU with the right flags, then link it with the main library.
2019-03-27 17:58:48 +01:00
uint32_t i;
resample: implement inter for ssse3
2026-04-22 18:28:25 +02:00
switch (SPA_PTR_ALIGNMENT(s, 16)) {
case 0:
for (i = 0; i < n_taps; i += 8) {
r = _mm_load_ps(s + i + 0);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 0)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 0)));
r = _mm_load_ps(s + i + 4);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 4)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
}
break;
case 4:
r0 = _mm_load_ps(s - 1);
for (i = 0; i < n_taps; i += 8) {
r1 = _mm_load_ps(s + i + 3);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 4);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 0)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 0)));
r0 = r1;
r1 = _mm_load_ps(s + i + 7);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 4);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 4)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1;
}
break;
case 8:
r0 = _mm_load_ps(s - 2);
for (i = 0; i < n_taps; i += 8) {
r1 = _mm_load_ps(s + i + 2);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 8);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 0)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 0)));
r0 = r1;
r1 = _mm_load_ps(s + i + 6);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 8);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 4)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1;
}
break;
case 12:
r0 = _mm_load_ps(s - 3);
for (i = 0; i < n_taps; i += 8) {
r1 = _mm_load_ps(s + i + 1);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 12);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 0)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 0)));
r0 = r1;
r1 = _mm_load_ps(s + i + 5);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 12);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + i + 4)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1;
}
break;
audioconvert: handle more optimizations Compile an optimized library for the given CPU with the right flags, then link it with the main library.
2019-03-27 17:58:48 +01:00
}
resample: implement inter for ssse3
2026-04-22 18:28:25 +02:00
sum[1] = _mm_mul_ps(_mm_sub_ps(sum[1], sum[0]), _mm_load1_ps(&x));
sum[0] = _mm_add_ps(sum[0], sum[1]);
sum[0] = _mm_add_ps(sum[0], _mm_movehdup_ps(sum[0]));
sum[0] = _mm_add_ss(sum[0], _mm_movehl_ps(sum[0], sum[0]));
_mm_store_ss(d, sum[0]);
audioconvert: add simple and faster resampler Add a sinc based resampler that, unlike speex, avoids memcpy and works directly on the source data. It also allows for ssse3 optimizations and aligned loads. It will later switch to table interpolation when doing variable rate.
2019-03-26 12:58:26 +01:00
}
MAKE_RESAMPLER_FULL(ssse3);
MAKE_RESAMPLER_INTER(ssse3);
Reference in a new issue Copy permalink