mirrors/pipewire
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pipewire/pipewire.git synced 2025-11-01 22:58:50 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

audioconvert: somewhat avoid precision loss in F32 to S32 conversion

Browse source 
At the very least, we should go through s25_32 intermediate
instead of s24_32, to avoid needlessly loosing 1 LSB precision bit.

FIXME: the noise codepath is not covered with tests.
This commit is contained in:
Roman Lebedev 2024-06-14 04:48:03 +03:00
parent 2a035ac49e
commit 175d533b56
No known key found for this signature in database
GPG key ID: 083C3EBB4A1689E0
4 changed files with 66 additions and 71 deletions
Download patch file Download diff file Expand all files Collapse all files

58
spa/plugins/audioconvert/fmt-ops-avx2.c
View file

@ -490,9 +490,9 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
uint32_t n, unrolled;
__m128 in[1];
__m128i out[4];
__m128 scale = _mm_set1_ps(S24_SCALE);
__m128 int_max = _mm_set1_ps(S24_MAX);
__m128 int_min = _mm_set1_ps(S24_MIN);
__m128 scale = _mm_set1_ps(S25_SCALE);
__m128 int_max = _mm_set1_ps(S25_MAX);
__m128 int_min = _mm_set1_ps(S25_MIN);
if (SPA_IS_ALIGNED(s0, 16))
unrolled = n_samples & ~3;
@ -503,7 +503,7 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
in[0] = _mm_mul_ps(_mm_load_ps(&s0[n]), scale);
in[0] = _MM_CLAMP_PS(in[0], int_min, int_max);
out[0] = _mm_cvtps_epi32(in[0]);
out[0] = _mm_slli_epi32(out[0], 8);
out[0] = _mm_slli_epi32(out[0], 7);
out[1] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(0, 3, 2, 1));
out[2] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(1, 0, 3, 2));
out[3] = _mm_shuffle_epi32(out[0], _MM_SHUFFLE(2, 1, 0, 3));
@ -518,7 +518,7 @@ conv_f32d_to_s32_1s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
in[0] = _mm_load_ss(&s0[n]);
in[0] = _mm_mul_ss(in[0], scale);
in[0] = _MM_CLAMP_SS(in[0], int_min, int_max);
*d = _mm_cvtss_si32(in[0]) << 8;
*d = _mm_cvtss_si32(in[0]) << 7;
d += n_channels;
}
}
@ -538,12 +538,12 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
uint32_t n, unrolled;
__m256 in[2];
__m256i out[2], t[2];
__m256 scale = _mm256_set1_ps(S24_SCALE);
__m256 int_min = _mm256_set1_ps(S24_MIN);
__m256 int_max = _mm256_set1_ps(S24_MAX);
__m256 scale = _mm256_set1_ps(S25_SCALE);
__m256 int_min = _mm256_set1_ps(S25_MIN);
__m256 int_max = _mm256_set1_ps(S25_MAX);
if (SPA_IS_ALIGNED(s0, 32) &&
SPA_IS_ALIGNED(s1, 32))
SPA_IS_ALIGNED(s1, 32))
unrolled = n_samples & ~7;
else
unrolled = 0;
@ -557,8 +557,8 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
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[0] = _mm256_slli_epi32(out[0], 8);
out[1] = _mm256_slli_epi32(out[1], 8);
out[0] = _mm256_slli_epi32(out[0], 7);
out[1] = _mm256_slli_epi32(out[1], 7);
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 */
@ -587,9 +587,9 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
for(; n < n_samples; n++) {
__m128 in[2];
__m128i out[2];
__m128 scale = _mm_set1_ps(S24_SCALE);
__m128 int_min = _mm_set1_ps(S24_MIN);
__m128 int_max = _mm_set1_ps(S24_MAX);
__m128 scale = _mm_set1_ps(S25_SCALE);
__m128 int_min = _mm_set1_ps(S25_MIN);
__m128 int_max = _mm_set1_ps(S25_MAX);
in[0] = _mm_load_ss(&s0[n]);
in[1] = _mm_load_ss(&s1[n]);
@ -599,7 +599,7 @@ conv_f32d_to_s32_2s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
in[0] = _mm_mul_ps(in[0], scale);
in[0] = _MM_CLAMP_PS(in[0], int_min, int_max);
out[0] = _mm_cvtps_epi32(in[0]);
out[0] = _mm_slli_epi32(out[0], 8);
out[0] = _mm_slli_epi32(out[0], 7);
_mm_storel_epi64((__m128i*)d, out[0]);
d += n_channels;
}
@ -614,14 +614,14 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
uint32_t n, unrolled;
__m256 in[4];
__m256i out[4], t[4];
__m256 scale = _mm256_set1_ps(S24_SCALE);
__m256 int_min = _mm256_set1_ps(S24_MIN);
__m256 int_max = _mm256_set1_ps(S24_MAX);
__m256 scale = _mm256_set1_ps(S25_SCALE);
__m256 int_min = _mm256_set1_ps(S25_MIN);
__m256 int_max = _mm256_set1_ps(S25_MAX);
if (SPA_IS_ALIGNED(s0, 32) &&
SPA_IS_ALIGNED(s1, 32) &&
SPA_IS_ALIGNED(s2, 32) &&
SPA_IS_ALIGNED(s3, 32))
SPA_IS_ALIGNED(s1, 32) &&
SPA_IS_ALIGNED(s2, 32) &&
SPA_IS_ALIGNED(s3, 32))
unrolled = n_samples & ~7;
else
unrolled = 0;
@ -641,10 +641,10 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
out[1] = _mm256_cvtps_epi32(in[1]); /* b0 b1 b2 b3 b4 b5 b6 b7 */
out[2] = _mm256_cvtps_epi32(in[2]); /* c0 c1 c2 c3 c4 c5 c6 c7 */
out[3] = _mm256_cvtps_epi32(in[3]); /* d0 d1 d2 d3 d4 d5 d6 d7 */
out[0] = _mm256_slli_epi32(out[0], 8);
out[1] = _mm256_slli_epi32(out[1], 8);
out[2] = _mm256_slli_epi32(out[2], 8);
out[3] = _mm256_slli_epi32(out[3], 8);
out[0] = _mm256_slli_epi32(out[0], 7);
out[1] = _mm256_slli_epi32(out[1], 7);
out[2] = _mm256_slli_epi32(out[2], 7);
out[3] = _mm256_slli_epi32(out[3], 7);
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 */
@ -669,9 +669,9 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
for(; n < n_samples; n++) {
__m128 in[4];
__m128i out[4];
__m128 scale = _mm_set1_ps(S24_SCALE);
__m128 int_min = _mm_set1_ps(S24_MIN);
__m128 int_max = _mm_set1_ps(S24_MAX);
__m128 scale = _mm_set1_ps(S25_SCALE);
__m128 int_min = _mm_set1_ps(S25_MIN);
__m128 int_max = _mm_set1_ps(S25_MAX);
in[0] = _mm_load_ss(&s0[n]);
in[1] = _mm_load_ss(&s1[n]);
@ -685,7 +685,7 @@ conv_f32d_to_s32_4s_avx2(void *data, void * SPA_RESTRICT dst, const void * SPA_R
in[0] = _mm_mul_ps(in[0], scale);
in[0] = _MM_CLAMP_PS(in[0], int_min, int_max);
out[0] = _mm_cvtps_epi32(in[0]);
out[0] = _mm_slli_epi32(out[0], 8);
out[0] = _mm_slli_epi32(out[0], 7);
_mm_storeu_si128((__m128i*)d, out[0]);
d += n_channels;
}