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

audioconvert: optimise f32d to/from s16s conversions

Browse source 
They are mostly used for sending and receiving RTP so it might be worth
to make them a bit faster.
This commit is contained in:
Wim Taymans 2025-01-24 10:35:52 +01:00
parent edcecd8975
commit 0c8f803d59
4 changed files with 427 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

179
spa/plugins/audioconvert/fmt-ops-avx2.c
View file

@ -23,6 +23,13 @@
#define _MM_CLAMP_SS(r,min,max) \
_mm_min_ss(_mm_max_ss(r, min), max)
#define _MM256_BSWAP_EPI16(x) \
({ \
_mm256_or_si256( \
_mm256_slli_epi16(x, 8), \
_mm256_srli_epi16(x, 8)); \
})
static void
conv_s16_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src,
uint32_t n_channels, uint32_t n_samples)
@ -74,6 +81,59 @@ conv_s16_to_f32d_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const voi
conv_s16_to_f32d_1s_avx2(conv, &dst[i], &s[i], n_channels, n_samples);
}
static void
conv_s16s_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src,
uint32_t n_channels, uint32_t n_samples)
{
const uint16_t *s = src;
float *d0 = dst[0];
uint32_t n, unrolled;
__m256i in = _mm256_setzero_si256();
__m256 out, factor = _mm256_set1_ps(1.0f / S16_SCALE);
if (SPA_LIKELY(SPA_IS_ALIGNED(d0, 32)))
unrolled = n_samples & ~7;
else
unrolled = 0;
for(n = 0; n < unrolled; n += 8) {
in = _mm256_insert_epi16(in, s[0*n_channels], 1);
in = _mm256_insert_epi16(in, s[1*n_channels], 3);
in = _mm256_insert_epi16(in, s[2*n_channels], 5);
in = _mm256_insert_epi16(in, s[3*n_channels], 7);
in = _mm256_insert_epi16(in, s[4*n_channels], 9);
in = _mm256_insert_epi16(in, s[5*n_channels], 11);
in = _mm256_insert_epi16(in, s[6*n_channels], 13);
in = _mm256_insert_epi16(in, s[7*n_channels], 15);
in = _MM256_BSWAP_EPI16(in);
in = _mm256_srai_epi32(in, 16);
out = _mm256_cvtepi32_ps(in);
out = _mm256_mul_ps(out, factor);
_mm256_store_ps(&d0[n], out);
s += 8*n_channels;
}
for(; n < n_samples; n++) {
__m128 out, factor = _mm_set1_ps(1.0f / S16_SCALE);
out = _mm_cvtsi32_ss(factor, (int16_t)bswap_16(s[0]));
out = _mm_mul_ss(out, factor);
_mm_store_ss(&d0[n], out);
s += n_channels;
}
}
void
conv_s16s_to_f32d_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src[],
uint32_t n_samples)
{
const uint16_t *s = src[0];
uint32_t i = 0, n_channels = conv->n_channels;
for(; i < n_channels; i++)
conv_s16s_to_f32d_1s_avx2(conv, &dst[i], &s[i], n_channels, n_samples);
}
void
conv_s16_to_f32d_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src[],
uint32_t n_samples)
@ -132,6 +192,66 @@ conv_s16_to_f32d_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v
}
}
void
conv_s16s_to_f32d_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src[],
uint32_t n_samples)
{
const uint16_t *s = src[0];
float *d0 = dst[0], *d1 = dst[1];
uint32_t n, unrolled;
__m256i in[2], t[4];
__m256 out[4], factor = _mm256_set1_ps(1.0f / S16_SCALE);
if (SPA_IS_ALIGNED(s, 32) &&
SPA_IS_ALIGNED(d0, 32) &&
SPA_IS_ALIGNED(d1, 32))
unrolled = n_samples & ~15;
else
unrolled = 0;
for(n = 0; n < unrolled; n += 16) {
in[0] = _mm256_load_si256((__m256i*)(s + 0));
in[1] = _mm256_load_si256((__m256i*)(s + 16));
in[0] = _MM256_BSWAP_EPI16(in[0]);
in[1] = _MM256_BSWAP_EPI16(in[1]);
t[0] = _mm256_slli_epi32(in[0], 16);
t[0] = _mm256_srai_epi32(t[0], 16);
out[0] = _mm256_cvtepi32_ps(t[0]);
out[0] = _mm256_mul_ps(out[0], factor);
t[1] = _mm256_srai_epi32(in[0], 16);
out[1] = _mm256_cvtepi32_ps(t[1]);
out[1] = _mm256_mul_ps(out[1], factor);
t[2] = _mm256_slli_epi32(in[1], 16);
t[2] = _mm256_srai_epi32(t[2], 16);
out[2] = _mm256_cvtepi32_ps(t[2]);
out[2] = _mm256_mul_ps(out[2], factor);
t[3] = _mm256_srai_epi32(in[1], 16);
out[3] = _mm256_cvtepi32_ps(t[3]);
out[3] = _mm256_mul_ps(out[3], factor);
_mm256_store_ps(&d0[n + 0], out[0]);
_mm256_store_ps(&d1[n + 0], out[1]);
_mm256_store_ps(&d0[n + 8], out[2]);
_mm256_store_ps(&d1[n + 8], out[3]);
s += 32;
}
for(; n < n_samples; n++) {
__m128 out[4], factor = _mm_set1_ps(1.0f / S16_SCALE);
out[0] = _mm_cvtsi32_ss(factor, (int16_t)bswap_16(s[0]));
out[0] = _mm_mul_ss(out[0], factor);
out[1] = _mm_cvtsi32_ss(factor, (int16_t)bswap_16(s[1]));
out[1] = _mm_mul_ss(out[1], factor);
_mm_store_ss(&d0[n], out[0]);
_mm_store_ss(&d1[n], out[1]);
s += 2;
}
}
static void
conv_s24_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src,
uint32_t n_channels, uint32_t n_samples)
@ -1006,3 +1126,62 @@ conv_f32d_to_s16_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const v
d += 2;
}
}
void
conv_f32d_to_s16s_2_avx2(struct convert *conv, void * SPA_RESTRICT dst[], const void * SPA_RESTRICT src[],
uint32_t n_samples)
{
const float *s0 = src[0], *s1 = src[1];
uint16_t *d = dst[0];
uint32_t n, unrolled;
__m256 in[4];
__m256i out[4], t[4];
__m256 int_scale = _mm256_set1_ps(S16_SCALE);
if (SPA_IS_ALIGNED(s0, 32) &&
SPA_IS_ALIGNED(s1, 32))
unrolled = n_samples & ~15;
else
unrolled = 0;
for(n = 0; n < unrolled; n += 16) {
in[0] = _mm256_mul_ps(_mm256_load_ps(&s0[n+0]), int_scale);
in[1] = _mm256_mul_ps(_mm256_load_ps(&s1[n+0]), int_scale);
in[2] = _mm256_mul_ps(_mm256_load_ps(&s0[n+8]), int_scale);
in[3] = _mm256_mul_ps(_mm256_load_ps(&s1[n+8]), int_scale);
out[0] = _mm256_cvtps_epi32(in[0]); /* a0 a1 a2 a3 a4 a5 a6 a7 */
out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */
out[2] = _mm256_cvtps_epi32(in[2]); /* a0 a1 a2 a3 a4 a5 a6 a7 */
out[3] = _mm256_cvtps_epi32(in[3]); /* b0 b1 b2 b3 b4 b5 b6 b7 */
t[0] = _mm256_unpacklo_epi32(out[0], out[1]); /* a0 b0 a1 b1 a4 b4 a5 b5 */
t[1] = _mm256_unpackhi_epi32(out[0], out[1]); /* a2 b2 a3 b3 a6 b6 a7 b7 */
t[2] = _mm256_unpacklo_epi32(out[2], out[3]); /* a0 b0 a1 b1 a4 b4 a5 b5 */
t[3] = _mm256_unpackhi_epi32(out[2], out[3]); /* a2 b2 a3 b3 a6 b6 a7 b7 */
out[0] = _mm256_packs_epi32(t[0], t[1]); /* a0 b0 a1 b1 a2 b2 a3 b3 a4 b4 a5 b5 a6 b6 a7 b7 */
out[1] = _mm256_packs_epi32(t[2], t[3]); /* a0 b0 a1 b1 a2 b2 a3 b3 a4 b4 a5 b5 a6 b6 a7 b7 */
out[0] = _MM256_BSWAP_EPI16(out[0]);
out[1] = _MM256_BSWAP_EPI16(out[1]);
_mm256_store_si256((__m256i*)(d+0), out[0]);
_mm256_store_si256((__m256i*)(d+16), out[1]);
d += 32;
}
for(; n < n_samples; n++) {
__m128 in[4];
__m128 int_scale = _mm_set1_ps(S16_SCALE);
__m128 int_max = _mm_set1_ps(S16_MAX);
__m128 int_min = _mm_set1_ps(S16_MIN);
in[0] = _mm_mul_ss(_mm_load_ss(&s0[n]), int_scale);
in[1] = _mm_mul_ss(_mm_load_ss(&s1[n]), int_scale);
in[0] = _MM_CLAMP_SS(in[0], int_min, int_max);
in[1] = _MM_CLAMP_SS(in[1], int_min, int_max);
d[0] = bswap_16((uint16_t)_mm_cvtss_si32(in[0]));
d[1] = bswap_16((uint16_t)_mm_cvtss_si32(in[1]));
d += 2;
}
}