mirror of
https://gitlab.freedesktop.org/pipewire/pipewire.git
synced 2025-11-03 09:01:54 -05:00
audioconvert: ensure alignment
Allocate a little bit more data to ensure alignment and overread of the dither data. Ensure sse2 can load aligned data in all cases.
This commit is contained in:
parent
b41d52cfd1
commit
00998ffd7e
4 changed files with 19 additions and 18 deletions
|
|
@ -49,7 +49,7 @@ void dither_f32_c(struct dither *dt, void * SPA_RESTRICT dst[],
|
||||||
uint32_t i, n, m, chunk;
|
uint32_t i, n, m, chunk;
|
||||||
const float **s = (const float**)src;
|
const float **s = (const float**)src;
|
||||||
float **d = (float**)dst;
|
float **d = (float**)dst;
|
||||||
float *t = dt->dither;
|
float *dither = dt->dither;
|
||||||
|
|
||||||
chunk = SPA_MIN(n_samples, dt->dither_size);
|
chunk = SPA_MIN(n_samples, dt->dither_size);
|
||||||
update_dither_c(dt, chunk);
|
update_dither_c(dt, chunk);
|
||||||
|
|
@ -62,7 +62,7 @@ void dither_f32_c(struct dither *dt, void * SPA_RESTRICT dst[],
|
||||||
const float *si = &s[i][n];
|
const float *si = &s[i][n];
|
||||||
|
|
||||||
for (m = 0; m < chunk; m++)
|
for (m = 0; m < chunk; m++)
|
||||||
di[m] = si[m] + t[m];
|
di[m] = si[m] + dither[m];
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
|
||||||
|
|
@ -29,25 +29,25 @@
|
||||||
static inline void update_dither_sse2(struct dither *dt, uint32_t n_samples)
|
static inline void update_dither_sse2(struct dither *dt, uint32_t n_samples)
|
||||||
{
|
{
|
||||||
uint32_t n;
|
uint32_t n;
|
||||||
const uint32_t *r = dt->random;
|
const uint32_t *r = SPA_PTR_ALIGN(dt->random, 16, uint32_t);
|
||||||
|
float *dither = SPA_PTR_ALIGN(dt->dither, 16, float);
|
||||||
__m128 scale = _mm_set1_ps(dt->scale), out[1];
|
__m128 scale = _mm_set1_ps(dt->scale), out[1];
|
||||||
__m128i in[1], t[1];
|
__m128i in[1], t[1];
|
||||||
|
|
||||||
for (n = 0; n < n_samples; n += 4) {
|
for (n = 0; n < n_samples; n += 4) {
|
||||||
|
|
||||||
/* 32 bit xorshift PRNG, see https://en.wikipedia.org/wiki/Xorshift */
|
/* 32 bit xorshift PRNG, see https://en.wikipedia.org/wiki/Xorshift */
|
||||||
in[0] = _mm_loadu_si128((__m128i*)r);
|
in[0] = _mm_load_si128((__m128i*)r);
|
||||||
t[0] = _mm_slli_epi32(in[0], 13);
|
t[0] = _mm_slli_epi32(in[0], 13);
|
||||||
in[0] = _mm_xor_si128(in[0], t[0]);
|
in[0] = _mm_xor_si128(in[0], t[0]);
|
||||||
t[0] = _mm_srli_epi32(in[0], 17);
|
t[0] = _mm_srli_epi32(in[0], 17);
|
||||||
in[0] = _mm_xor_si128(in[0], t[0]);
|
in[0] = _mm_xor_si128(in[0], t[0]);
|
||||||
t[0] = _mm_slli_epi32(in[0], 5);
|
t[0] = _mm_slli_epi32(in[0], 5);
|
||||||
in[0] = _mm_xor_si128(in[0], t[0]);
|
in[0] = _mm_xor_si128(in[0], t[0]);
|
||||||
_mm_storeu_si128((__m128i*)r, in[0]);
|
_mm_store_si128((__m128i*)r, in[0]);
|
||||||
|
|
||||||
out[0] = _mm_cvtepi32_ps(in[0]);
|
out[0] = _mm_cvtepi32_ps(in[0]);
|
||||||
out[0] = _mm_mul_ps(out[0], scale);
|
out[0] = _mm_mul_ps(out[0], scale);
|
||||||
_mm_storeu_ps(&dt->dither[n], out[0]);
|
_mm_store_ps(&dither[n], out[0]);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
@ -57,7 +57,7 @@ void dither_f32_sse2(struct dither *dt, void * SPA_RESTRICT dst[],
|
||||||
uint32_t i, n, m, chunk, unrolled;
|
uint32_t i, n, m, chunk, unrolled;
|
||||||
const float **s = (const float**)src;
|
const float **s = (const float**)src;
|
||||||
float **d = (float**)dst;
|
float **d = (float**)dst;
|
||||||
float *t = dt->dither;
|
float *dither = SPA_PTR_ALIGN(dt->dither, 16, float);
|
||||||
__m128 in[4];
|
__m128 in[4];
|
||||||
|
|
||||||
chunk = SPA_MIN(n_samples, dt->dither_size);
|
chunk = SPA_MIN(n_samples, dt->dither_size);
|
||||||
|
|
@ -81,17 +81,17 @@ void dither_f32_sse2(struct dither *dt, void * SPA_RESTRICT dst[],
|
||||||
in[1] = _mm_load_ps(&si[m + 4]);
|
in[1] = _mm_load_ps(&si[m + 4]);
|
||||||
in[2] = _mm_load_ps(&si[m + 8]);
|
in[2] = _mm_load_ps(&si[m + 8]);
|
||||||
in[3] = _mm_load_ps(&si[m + 12]);
|
in[3] = _mm_load_ps(&si[m + 12]);
|
||||||
in[0] = _mm_add_ps(in[0], _mm_load_ps(&t[m ]));
|
in[0] = _mm_add_ps(in[0], _mm_load_ps(&dither[m ]));
|
||||||
in[1] = _mm_add_ps(in[1], _mm_load_ps(&t[m + 4]));
|
in[1] = _mm_add_ps(in[1], _mm_load_ps(&dither[m + 4]));
|
||||||
in[2] = _mm_add_ps(in[2], _mm_load_ps(&t[m + 8]));
|
in[2] = _mm_add_ps(in[2], _mm_load_ps(&dither[m + 8]));
|
||||||
in[3] = _mm_add_ps(in[3], _mm_load_ps(&t[m + 12]));
|
in[3] = _mm_add_ps(in[3], _mm_load_ps(&dither[m + 12]));
|
||||||
_mm_store_ps(&di[m ], in[0]);
|
_mm_store_ps(&di[m ], in[0]);
|
||||||
_mm_store_ps(&di[m + 4], in[1]);
|
_mm_store_ps(&di[m + 4], in[1]);
|
||||||
_mm_store_ps(&di[m + 8], in[2]);
|
_mm_store_ps(&di[m + 8], in[2]);
|
||||||
_mm_store_ps(&di[m + 12], in[3]);
|
_mm_store_ps(&di[m + 12], in[3]);
|
||||||
}
|
}
|
||||||
for (; m < chunk; m++)
|
for (; m < chunk; m++)
|
||||||
di[m] = si[m] + t[m];
|
di[m] = si[m] + dither[m];
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
|
||||||
|
|
@ -85,14 +85,14 @@ int dither_init(struct dither *d)
|
||||||
d->scale = 1.0f / (1ULL << (31 + d->intensity));
|
d->scale = 1.0f / (1ULL << (31 + d->intensity));
|
||||||
|
|
||||||
d->dither_size = DITHER_SIZE;
|
d->dither_size = DITHER_SIZE;
|
||||||
d->dither = calloc(d->dither_size + 8, sizeof(float));
|
d->dither = calloc(d->dither_size + DITHER_OPS_MAX_OVERREAD +
|
||||||
|
DITHER_OPS_MAX_ALIGN / sizeof(float), sizeof(float));
|
||||||
if (d->dither == NULL)
|
if (d->dither == NULL)
|
||||||
return -errno;
|
return -errno;
|
||||||
|
|
||||||
for (i = 0; i < SPA_N_ELEMENTS(d->random); i++)
|
for (i = 0; i < SPA_N_ELEMENTS(d->random); i++)
|
||||||
d->random[i] = random();
|
d->random[i] = random();
|
||||||
|
|
||||||
|
|
||||||
d->free = impl_dither_free;
|
d->free = impl_dither_free;
|
||||||
d->process = info->process;
|
d->process = info->process;
|
||||||
return 0;
|
return 0;
|
||||||
|
|
|
||||||
|
|
@ -29,6 +29,9 @@
|
||||||
#include <spa/utils/string.h>
|
#include <spa/utils/string.h>
|
||||||
#include <spa/param/audio/raw.h>
|
#include <spa/param/audio/raw.h>
|
||||||
|
|
||||||
|
#define DITHER_OPS_MAX_ALIGN 16
|
||||||
|
#define DITHER_OPS_MAX_OVERREAD 16
|
||||||
|
|
||||||
struct dither {
|
struct dither {
|
||||||
uint32_t intensity;
|
uint32_t intensity;
|
||||||
#define DITHER_METHOD_NONE 0
|
#define DITHER_METHOD_NONE 0
|
||||||
|
|
@ -45,7 +48,7 @@ struct dither {
|
||||||
const void * SPA_RESTRICT src[], uint32_t n_samples);
|
const void * SPA_RESTRICT src[], uint32_t n_samples);
|
||||||
void (*free) (struct dither *d);
|
void (*free) (struct dither *d);
|
||||||
|
|
||||||
uint32_t random[16];
|
uint32_t random[16 + DITHER_OPS_MAX_ALIGN/sizeof(uint32_t)];
|
||||||
float *dither;
|
float *dither;
|
||||||
uint32_t dither_size;
|
uint32_t dither_size;
|
||||||
float scale;
|
float scale;
|
||||||
|
|
@ -83,8 +86,6 @@ void dither_##name##_##arch(struct dither *d, \
|
||||||
const void * SPA_RESTRICT src[], \
|
const void * SPA_RESTRICT src[], \
|
||||||
uint32_t n_samples);
|
uint32_t n_samples);
|
||||||
|
|
||||||
#define DITHER_OPS_MAX_ALIGN 16
|
|
||||||
|
|
||||||
DEFINE_FUNCTION(f32, c);
|
DEFINE_FUNCTION(f32, c);
|
||||||
#if defined(HAVE_SSE2)
|
#if defined(HAVE_SSE2)
|
||||||
DEFINE_FUNCTION(f32, sse2);
|
DEFINE_FUNCTION(f32, sse2);
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue