mirrors/pipewire
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pipewire/pipewire.git synced 2026-04-25 06:46:40 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

resample: use some extra accumulators to improve pipelining

Browse source
This commit is contained in:
Wim Taymans 2026-04-23 09:32:31 +02:00
parent 75e432a49a
commit 596047aaef
3 changed files with 25 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

11
spa/plugins/audioconvert/resample-native-avx2.c
View file

@ -40,8 +40,9 @@ static inline void inner_product_ip_avx2(float *d, const float * SPA_RESTRICT s,
const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x, const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
uint32_t n_taps) uint32_t n_taps)
{ {
__m256 sy[2] = { _mm256_setzero_ps(), _mm256_setzero_ps() }, ty; __m256 sy[4] = { _mm256_setzero_ps(), _mm256_setzero_ps(),
__m128 sx[2], tx; _mm256_setzero_ps(), _mm256_setzero_ps() }, ty;
__m128 sx[4], tx;
uint32_t i, n_taps4 = n_taps & ~0xf; uint32_t i, n_taps4 = n_taps & ~0xf;
for (i = 0; i < n_taps4; i += 16) { for (i = 0; i < n_taps4; i += 16) {
@ -49,9 +50,11 @@ static inline void inner_product_ip_avx2(float *d, const float * SPA_RESTRICT s,
sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 0), sy[0]); sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 0), sy[0]);
sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 0), sy[1]); sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 0), sy[1]);
ty = _mm256_loadu_ps(s + i + 8); ty = _mm256_loadu_ps(s + i + 8);
sy[0] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 8), sy[0]); sy[2] = _mm256_fmadd_ps(ty, _mm256_load_ps(t0 + i + 8), sy[2]);
sy[1] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 8), sy[1]); sy[3] = _mm256_fmadd_ps(ty, _mm256_load_ps(t1 + i + 8), sy[3]);
} }
sy[0] = _mm256_add_ps(sy[0], sy[2]);
sy[1] = _mm256_add_ps(sy[1], sy[3]);
sx[0] = _mm_add_ps(_mm256_extractf128_ps(sy[0], 0), _mm256_extractf128_ps(sy[0], 1)); sx[0] = _mm_add_ps(_mm256_extractf128_ps(sy[0], 0), _mm256_extractf128_ps(sy[0], 1));
sx[1] = _mm_add_ps(_mm256_extractf128_ps(sy[1], 0), _mm256_extractf128_ps(sy[1], 1)); sx[1] = _mm_add_ps(_mm256_extractf128_ps(sy[1], 0), _mm256_extractf128_ps(sy[1], 1));

9
spa/plugins/audioconvert/resample-native-sse.c
View file

@ -32,7 +32,8 @@ static inline void inner_product_ip_sse(float *d, const float * SPA_RESTRICT s,
const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x, const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
uint32_t n_taps) uint32_t n_taps)
{ {
__m128 sum[2] = { _mm_setzero_ps (), _mm_setzero_ps () }, t; __m128 sum[4] = { _mm_setzero_ps(), _mm_setzero_ps(),
_mm_setzero_ps(), _mm_setzero_ps() }, t;
uint32_t i; uint32_t i;
for (i = 0; i < n_taps; i += 8) { for (i = 0; i < n_taps; i += 8) {
@ -40,9 +41,11 @@ static inline void inner_product_ip_sse(float *d, const float * SPA_RESTRICT s,
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(t, _mm_load_ps(t0 + i + 0))); sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(t, _mm_load_ps(t0 + i + 0)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(t, _mm_load_ps(t1 + i + 0))); sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(t, _mm_load_ps(t1 + i + 0)));
t = _mm_loadu_ps(s + i + 4); t = _mm_loadu_ps(s + i + 4);
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(t, _mm_load_ps(t0 + i + 4))); sum[2] = _mm_add_ps(sum[2], _mm_mul_ps(t, _mm_load_ps(t0 + i + 4)));
sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(t, _mm_load_ps(t1 + i + 4))); sum[3] = _mm_add_ps(sum[3], _mm_mul_ps(t, _mm_load_ps(t1 + i + 4)));
} }
sum[0] = _mm_add_ps(sum[0], sum[2]);
sum[1] = _mm_add_ps(sum[1], sum[3]);
sum[1] = _mm_mul_ps(_mm_sub_ps(sum[1], sum[0]), _mm_load1_ps(&x)); 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], sum[1]);
sum[0] = _mm_add_ps(sum[0], _mm_movehl_ps(sum[0], sum[0])); sum[0] = _mm_add_ps(sum[0], _mm_movehl_ps(sum[0], sum[0]));

21
spa/plugins/audioconvert/resample-native-ssse3.c
View file

@ -82,7 +82,8 @@ static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s
const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x, const float * SPA_RESTRICT t0, const float * SPA_RESTRICT t1, float x,
uint32_t n_taps) uint32_t n_taps)
{ {
__m128 sum[2] = { _mm_setzero_ps(), _mm_setzero_ps() }; __m128 sum[4] = { _mm_setzero_ps(), _mm_setzero_ps(),
_mm_setzero_ps(), _mm_setzero_ps() };
__m128 r0, r1, r; __m128 r0, r1, r;
uint32_t i; uint32_t i;
@ -93,8 +94,8 @@ static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s
sum[0] = _mm_add_ps(sum[0], _mm_mul_ps(r, _mm_load_ps(t0 + 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))); sum[1] = _mm_add_ps(sum[1], _mm_mul_ps(r, _mm_load_ps(t1 + i + 0)));
r = _mm_load_ps(s + i + 4); 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[2] = _mm_add_ps(sum[2], _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))); sum[3] = _mm_add_ps(sum[3], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
} }
break; break;
case 4: case 4:
@ -107,8 +108,8 @@ static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s
r0 = r1; r0 = r1;
r1 = _mm_load_ps(s + i + 7); r1 = _mm_load_ps(s + i + 7);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 4); 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[2] = _mm_add_ps(sum[2], _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))); sum[3] = _mm_add_ps(sum[3], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1; r0 = r1;
} }
break; break;
@ -122,8 +123,8 @@ static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s
r0 = r1; r0 = r1;
r1 = _mm_load_ps(s + i + 6); r1 = _mm_load_ps(s + i + 6);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 8); 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[2] = _mm_add_ps(sum[2], _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))); sum[3] = _mm_add_ps(sum[3], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1; r0 = r1;
} }
break; break;
@ -137,12 +138,14 @@ static inline void inner_product_ip_ssse3(float *d, const float * SPA_RESTRICT s
r0 = r1; r0 = r1;
r1 = _mm_load_ps(s + i + 5); r1 = _mm_load_ps(s + i + 5);
r = (__m128)_mm_alignr_epi8((__m128i)r1, (__m128i)r0, 12); 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[2] = _mm_add_ps(sum[2], _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))); sum[3] = _mm_add_ps(sum[3], _mm_mul_ps(r, _mm_load_ps(t1 + i + 4)));
r0 = r1; r0 = r1;
} }
break; break;
} }
sum[0] = _mm_add_ps(sum[0], sum[2]);
sum[1] = _mm_add_ps(sum[1], sum[3]);
sum[1] = _mm_mul_ps(_mm_sub_ps(sum[1], sum[0]), _mm_load1_ps(&x)); 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], sum[1]);
sum[0] = _mm_add_ps(sum[0], _mm_movehdup_ps(sum[0])); sum[0] = _mm_add_ps(sum[0], _mm_movehdup_ps(sum[0]));