mirror of
https://gitlab.freedesktop.org/pipewire/pipewire.git
synced 2025-11-01 22:58:50 -04:00
resample-native: small tweaks
This commit is contained in:
parent
d260cb19be
commit
d47353c0e6
2 changed files with 23 additions and 33 deletions
|
|
@ -30,19 +30,20 @@
|
|||
static void inner_product_avx(float *d, const float * SPA_RESTRICT s,
|
||||
const float * SPA_RESTRICT taps, uint32_t n_taps)
|
||||
{
|
||||
__m256 sy[2] = { _mm256_setzero_ps(), _mm256_setzero_ps() };
|
||||
__m256 sy[2] = { _mm256_setzero_ps(), _mm256_setzero_ps() }, ty;
|
||||
__m128 sx[2];
|
||||
uint32_t i = 0;
|
||||
uint32_t n_taps4 = n_taps & ~0xf;
|
||||
|
||||
for (; i < n_taps4; i += 16) {
|
||||
sy[0] = _mm256_fmadd_ps(_mm256_loadu_ps(s + i + 0), _mm256_load_ps(taps + i + 0), sy[0]);
|
||||
sy[1] = _mm256_fmadd_ps(_mm256_loadu_ps(s + i + 8), _mm256_load_ps(taps + i + 8), sy[1]);
|
||||
ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 0));
|
||||
sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(taps + i + 0), sy[0]);
|
||||
ty = (__m256)_mm256_lddqu_si256((__m256i*)(s + i + 8));
|
||||
sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(taps + i + 8), sy[1]);
|
||||
}
|
||||
sy[0] = _mm256_add_ps(sy[0], sy[1]);
|
||||
|
||||
sx[0] = _mm256_extractf128_ps(sy[0], 0);
|
||||
sy[0] = _mm256_add_ps(sy[1], sy[0]);
|
||||
sx[1] = _mm256_extractf128_ps(sy[0], 1);
|
||||
sx[0] = _mm256_extractf128_ps(sy[0], 0);
|
||||
for (; i < n_taps; i += 8) {
|
||||
sx[0] = _mm_fmadd_ps(_mm_loadu_ps(s + i + 0), _mm_load_ps(taps + i + 0), sx[0]);
|
||||
sx[1] = _mm_fmadd_ps(_mm_loadu_ps(s + i + 4), _mm_load_ps(taps + i + 4), sx[1]);
|
||||
|
|
@ -69,9 +70,8 @@ static void inner_product_ip_avx(float *d, const float * SPA_RESTRICT s,
|
|||
sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 8), sy[0]);
|
||||
sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 8), sy[1]);
|
||||
}
|
||||
sy[0] = _mm256_add_ps(sy[0], sy[1]);
|
||||
sx[0] = _mm256_extractf128_ps(sy[0], 0);
|
||||
sx[1] = _mm256_extractf128_ps(sy[0], 1);
|
||||
sx[0] = _mm256_extractf128_ps(_mm256_hadd_ps(sy[0], sy[0]), 0);
|
||||
sx[1] = _mm256_extractf128_ps(_mm256_hadd_ps(sy[1], sy[1]), 0);
|
||||
for (; i < n_taps; i += 8) {
|
||||
tx = _mm_loadu_ps(s + i + 0);
|
||||
sx[0] = _mm_fmadd_ps(tx, _mm_load_ps(t0 + i + 0), sx[0]);
|
||||
|
|
|
|||
|
|
@ -51,23 +51,13 @@ struct native_data {
|
|||
float *hist_mem;
|
||||
};
|
||||
|
||||
#define DEFINE_RESAMPLER_COPY(arch) \
|
||||
void do_resample_copy_##arch(struct resample *r, \
|
||||
const void * SPA_RESTRICT src[], uint32_t *in_len, \
|
||||
void * SPA_RESTRICT dst[], uint32_t offs, uint32_t *out_len)
|
||||
|
||||
#define DEFINE_RESAMPLER_FULL(arch) \
|
||||
void do_resample_full_##arch(struct resample *r, \
|
||||
const void * SPA_RESTRICT src[], uint32_t *in_len, \
|
||||
void * SPA_RESTRICT dst[], uint32_t offs, uint32_t *out_len)
|
||||
|
||||
#define DEFINE_RESAMPLER_INTER(arch) \
|
||||
void do_resample_inter_##arch(struct resample *r, \
|
||||
#define DEFINE_RESAMPLER(type,arch) \
|
||||
void do_resample_##type##_##arch(struct resample *r, \
|
||||
const void * SPA_RESTRICT src[], uint32_t *in_len, \
|
||||
void * SPA_RESTRICT dst[], uint32_t offs, uint32_t *out_len)
|
||||
|
||||
#define MAKE_RESAMPLER_COPY(arch) \
|
||||
DEFINE_RESAMPLER_COPY(arch) \
|
||||
DEFINE_RESAMPLER(copy,arch) \
|
||||
{ \
|
||||
struct native_data *data = r->data; \
|
||||
uint32_t index, n_taps = data->n_taps; \
|
||||
|
|
@ -94,7 +84,7 @@ DEFINE_RESAMPLER_COPY(arch) \
|
|||
}
|
||||
|
||||
#define MAKE_RESAMPLER_FULL(arch) \
|
||||
DEFINE_RESAMPLER_FULL(arch) \
|
||||
DEFINE_RESAMPLER(full,arch) \
|
||||
{ \
|
||||
struct native_data *data = r->data; \
|
||||
uint32_t n_taps = data->n_taps, stride = data->filter_stride_os; \
|
||||
|
|
@ -133,7 +123,7 @@ DEFINE_RESAMPLER_FULL(arch) \
|
|||
}
|
||||
|
||||
#define MAKE_RESAMPLER_INTER(arch) \
|
||||
DEFINE_RESAMPLER_INTER(arch) \
|
||||
DEFINE_RESAMPLER(inter,arch) \
|
||||
{ \
|
||||
struct native_data *data = r->data; \
|
||||
uint32_t index, phase, stride = data->filter_stride; \
|
||||
|
|
@ -180,19 +170,19 @@ DEFINE_RESAMPLER_INTER(arch) \
|
|||
}
|
||||
|
||||
|
||||
DEFINE_RESAMPLER_COPY(c);
|
||||
DEFINE_RESAMPLER_FULL(c);
|
||||
DEFINE_RESAMPLER_INTER(c);
|
||||
DEFINE_RESAMPLER(copy,c);
|
||||
DEFINE_RESAMPLER(full,c);
|
||||
DEFINE_RESAMPLER(inter,c);
|
||||
|
||||
#if defined (HAVE_SSE)
|
||||
DEFINE_RESAMPLER_FULL(sse);
|
||||
DEFINE_RESAMPLER_INTER(sse);
|
||||
DEFINE_RESAMPLER(full,sse);
|
||||
DEFINE_RESAMPLER(inter,sse);
|
||||
#endif
|
||||
#if defined (HAVE_SSSE3)
|
||||
DEFINE_RESAMPLER_FULL(ssse3);
|
||||
DEFINE_RESAMPLER_INTER(ssse3);
|
||||
DEFINE_RESAMPLER(full,ssse3);
|
||||
DEFINE_RESAMPLER(inter,ssse3);
|
||||
#endif
|
||||
#if defined (HAVE_AVX) && defined(HAVE_FMA)
|
||||
DEFINE_RESAMPLER_FULL(avx);
|
||||
DEFINE_RESAMPLER_INTER(avx);
|
||||
DEFINE_RESAMPLER(full,avx);
|
||||
DEFINE_RESAMPLER(inter,avx);
|
||||
#endif
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue