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

filter-chain: add optimized sum function

Browse source
This commit is contained in:
Wim Taymans 2022-12-16 10:05:24 +01:00
parent cd38d7b53b
commit 0f2f113bdc
6 changed files with 177 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

10
src/modules/meson.build
View file

@ -71,6 +71,16 @@ if have_sse
simd_cargs += ['-DHAVE_SSE']
simd_dependencies += filter_chain_sse
endif
if have_avx
filter_chain_avx = static_library('filter_chain_avx',
['module-filter-chain/dsp-ops-avx.c' ],
c_args : [avx_args, fma_args,'-O3', '-DHAVE_AVX'],
dependencies : [ spa_dep ],
install : false
)
simd_cargs += ['-DHAVE_AVX']
simd_dependencies += filter_chain_avx
endif
if have_neon
filter_chain_neon = static_library('filter_chain_neon',
['module-filter-chain/pffft.c' ],

85
src/modules/module-filter-chain/dsp-ops-avx.c Normal file
View file

@ -0,0 +1,85 @@
/* Spa
*
* Copyright © 2022 Wim Taymans
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
* DEALINGS IN THE SOFTWARE.
*/
#include <string.h>
#include <stdio.h>
#include <math.h>
#include <spa/utils/defs.h>
#include "dsp-ops.h"
#include <immintrin.h>
void dsp_sum_avx(struct dsp_ops *ops, float *r, const float *a, const float *b, uint32_t n_samples)
{
uint32_t n, unrolled;
__m256 in[4];
unrolled = n_samples & ~31;
if (SPA_LIKELY(SPA_IS_ALIGNED(r, 32)) &&
SPA_LIKELY(SPA_IS_ALIGNED(a, 32)) &&
SPA_LIKELY(SPA_IS_ALIGNED(b, 32))) {
for (n = 0; n < unrolled; n += 32) {
in[0] = _mm256_load_ps(&a[n+ 0]);
in[1] = _mm256_load_ps(&a[n+ 8]);
in[2] = _mm256_load_ps(&a[n+16]);
in[3] = _mm256_load_ps(&a[n+24]);
in[0] = _mm256_add_ps(in[0], _mm256_load_ps(&b[n+ 0]));
in[1] = _mm256_add_ps(in[1], _mm256_load_ps(&b[n+ 8]));
in[2] = _mm256_add_ps(in[2], _mm256_load_ps(&b[n+16]));
in[3] = _mm256_add_ps(in[3], _mm256_load_ps(&b[n+24]));
_mm256_store_ps(&r[n+ 0], in[0]);
_mm256_store_ps(&r[n+ 8], in[1]);
_mm256_store_ps(&r[n+16], in[2]);
_mm256_store_ps(&r[n+24], in[3]);
}
} else {
for (n = 0; n < unrolled; n += 16) {
in[0] = _mm256_loadu_ps(&a[n+ 0]);
in[1] = _mm256_loadu_ps(&a[n+ 8]);
in[2] = _mm256_loadu_ps(&a[n+16]);
in[3] = _mm256_loadu_ps(&a[n+24]);
in[0] = _mm256_add_ps(in[0], _mm256_loadu_ps(&b[n+ 0]));
in[1] = _mm256_add_ps(in[1], _mm256_loadu_ps(&b[n+ 8]));
in[2] = _mm256_add_ps(in[2], _mm256_loadu_ps(&b[n+16]));
in[3] = _mm256_add_ps(in[3], _mm256_loadu_ps(&b[n+24]));
_mm256_storeu_ps(&r[n+ 0], in[0]);
_mm256_storeu_ps(&r[n+ 8], in[1]);
_mm256_storeu_ps(&r[n+16], in[2]);
_mm256_storeu_ps(&r[n+24], in[3]);
}
}
for (; n < n_samples; n++) {
__m128 in[1];
in[0] = _mm_load_ss(&a[n]);
in[0] = _mm_add_ss(in[0], _mm_load_ss(&b[n]));
_mm_store_ss(&r[n], in[0]);
}
}

8
src/modules/module-filter-chain/dsp-ops-c.c
View file

@ -133,3 +133,11 @@ void dsp_biquad_run_c(struct dsp_ops *ops, struct biquad *bq,
#undef F
}
void dsp_sum_c(struct dsp_ops *ops, float * dst,
const float * SPA_RESTRICT a, const float * SPA_RESTRICT b, uint32_t n_samples)
{
uint32_t i;
for (i = 0; i < n_samples; i++)
dst[i] = a[i] + b[i];
}

51
src/modules/module-filter-chain/dsp-ops-sse.c
View file

@ -89,3 +89,54 @@ void dsp_mix_gain_sse(struct dsp_ops *ops,
}
}
}
void dsp_sum_sse(struct dsp_ops *ops, float *r, const float *a, const float *b, uint32_t n_samples)
{
uint32_t n, unrolled;
__m128 in[4];
unrolled = n_samples & ~15;
if (SPA_LIKELY(SPA_IS_ALIGNED(r, 16)) &&
SPA_LIKELY(SPA_IS_ALIGNED(a, 16)) &&
SPA_LIKELY(SPA_IS_ALIGNED(b, 16))) {
for (n = 0; n < unrolled; n += 16) {
in[0] = _mm_load_ps(&a[n+ 0]);
in[1] = _mm_load_ps(&a[n+ 4]);
in[2] = _mm_load_ps(&a[n+ 8]);
in[3] = _mm_load_ps(&a[n+12]);
in[0] = _mm_add_ps(in[0], _mm_load_ps(&b[n+ 0]));
in[1] = _mm_add_ps(in[1], _mm_load_ps(&b[n+ 4]));
in[2] = _mm_add_ps(in[2], _mm_load_ps(&b[n+ 8]));
in[3] = _mm_add_ps(in[3], _mm_load_ps(&b[n+12]));
_mm_store_ps(&r[n+ 0], in[0]);
_mm_store_ps(&r[n+ 4], in[1]);
_mm_store_ps(&r[n+ 8], in[2]);
_mm_store_ps(&r[n+12], in[3]);
}
} else {
for (n = 0; n < unrolled; n += 16) {
in[0] = _mm_loadu_ps(&a[n+ 0]);
in[1] = _mm_loadu_ps(&a[n+ 4]);
in[2] = _mm_loadu_ps(&a[n+ 8]);
in[3] = _mm_loadu_ps(&a[n+12]);
in[0] = _mm_add_ps(in[0], _mm_loadu_ps(&b[n+ 0]));
in[1] = _mm_add_ps(in[1], _mm_loadu_ps(&b[n+ 4]));
in[2] = _mm_add_ps(in[2], _mm_loadu_ps(&b[n+ 8]));
in[3] = _mm_add_ps(in[3], _mm_loadu_ps(&b[n+12]));
_mm_storeu_ps(&r[n+ 0], in[0]);
_mm_storeu_ps(&r[n+ 4], in[1]);
_mm_storeu_ps(&r[n+ 8], in[2]);
_mm_storeu_ps(&r[n+12], in[3]);
}
}
for (; n < n_samples; n++) {
in[0] = _mm_load_ss(&a[n]);
in[0] = _mm_add_ss(in[0], _mm_load_ss(&b[n]));
_mm_store_ss(&r[n], in[0]);
}
}

11
src/modules/module-filter-chain/dsp-ops.c
View file

@ -40,12 +40,22 @@ struct dsp_info {
static struct dsp_info dsp_table[] =
{
#if defined (HAVE_AVX)
{ SPA_CPU_FLAG_AVX,
.funcs.clear = dsp_clear_c,
.funcs.copy = dsp_copy_c,
.funcs.mix_gain = dsp_mix_gain_sse,
.funcs.biquad_run = dsp_biquad_run_c,
.funcs.sum = dsp_sum_avx,
},
#endif
#if defined (HAVE_SSE)
{ SPA_CPU_FLAG_SSE,
.funcs.clear = dsp_clear_c,
.funcs.copy = dsp_copy_c,
.funcs.mix_gain = dsp_mix_gain_sse,
.funcs.biquad_run = dsp_biquad_run_c,
.funcs.sum = dsp_sum_sse,
},
#endif
{ 0,
@ -53,6 +63,7 @@ static struct dsp_info dsp_table[] =
.funcs.copy = dsp_copy_c,
.funcs.mix_gain = dsp_mix_gain_c,
.funcs.biquad_run = dsp_biquad_run_c,
.funcs.sum = dsp_sum_c,
},
};

12
src/modules/module-filter-chain/dsp-ops.h
View file

@ -42,6 +42,9 @@ struct dsp_ops_funcs {
float gain[], uint32_t n_src, uint32_t n_samples);
void (*biquad_run) (struct dsp_ops *ops, struct biquad *bq,
float *out, const float *in, uint32_t n_samples);
void (*sum) (struct dsp_ops *ops,
float * dst, const float * SPA_RESTRICT a,
const float * SPA_RESTRICT b, uint32_t n_samples);
};
struct dsp_ops {
@ -62,6 +65,7 @@ int dsp_ops_init(struct dsp_ops *ops);
#define dsp_ops_copy(ops,...) (ops)->funcs.copy(ops, __VA_ARGS__)
#define dsp_ops_mix_gain(ops,...) (ops)->funcs.mix_gain(ops, __VA_ARGS__)
#define dsp_ops_biquad_run(ops,...) (ops)->funcs.biquad_run(ops, __VA_ARGS__)
#define dsp_ops_sum(ops,...) (ops)->funcs.sum(ops, __VA_ARGS__)
#define MAKE_CLEAR_FUNC(arch) \
void dsp_clear_##arch(struct dsp_ops *ops, void * SPA_RESTRICT dst, uint32_t n_samples)
@ -74,14 +78,22 @@ void dsp_mix_gain_##arch(struct dsp_ops *ops, void * SPA_RESTRICT dst, \
#define MAKE_BIQUAD_RUN_FUNC(arch) \
void dsp_biquad_run_##arch (struct dsp_ops *ops, struct biquad *bq, \
float *out, const float *in, uint32_t n_samples)
#define MAKE_SUM_FUNC(arch) \
void dsp_sum_##arch (struct dsp_ops *ops, float * SPA_RESTRICT dst, \
const float * SPA_RESTRICT a, const float * SPA_RESTRICT b, uint32_t n_samples);
MAKE_CLEAR_FUNC(c);
MAKE_COPY_FUNC(c);
MAKE_MIX_GAIN_FUNC(c);
MAKE_BIQUAD_RUN_FUNC(c);
MAKE_SUM_FUNC(c);
#if defined (HAVE_SSE)
MAKE_MIX_GAIN_FUNC(sse);
MAKE_SUM_FUNC(sse);
#endif
#if defined (HAVE_AVX)
MAKE_SUM_FUNC(avx);
#endif
#endif /* DSP_OPS_H */