mirror of
https://gitlab.freedesktop.org/pipewire/pipewire
synced 2024-10-01 13:44:40 +00:00
audioconvert: avoid even more precision loss in S32 to F32 conversion
There's really no point in doing that s25_32 intermediate step, to be honest i don't have a clue why the original implementation did that \_(ツ)_/¯. Both `S25_SCALE` and `S32_SCALE` are powers of two, and thus are both exactly representable as floats, and reprocial of power-of-two is also exactly representable, so it's not like that rescaling results in precision loss. This additionally avoids right-shift, and thus is even faster. As `test_lossless_s32_lossless_subset` shows, if the integer is in the form of s25+shift, the maximal absolute error became even lower, but not zero, because F32->S32 still goes through S25 intermediate. I think we could theoretically do better, but then the clamping becomes pretty finicky, so i don't feel like touching that here.
This commit is contained in:
parent
c517865864
commit
f4c89b1b40
|
@ -316,7 +316,7 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
float *d0 = dst[0], *d1 = dst[1], *d2 = dst[2], *d3 = dst[3];
|
||||
uint32_t n, unrolled;
|
||||
__m256i in[4];
|
||||
__m256 out[4], factor = _mm256_set1_ps(1.0f / S25_SCALE);
|
||||
__m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE);
|
||||
__m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels,
|
||||
4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels);
|
||||
|
||||
|
@ -334,11 +334,6 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
in[2] = _mm256_i32gather_epi32((int*)&s[2], mask1, 4);
|
||||
in[3] = _mm256_i32gather_epi32((int*)&s[3], mask1, 4);
|
||||
|
||||
in[0] = _mm256_srai_epi32(in[0], 7);
|
||||
in[1] = _mm256_srai_epi32(in[1], 7);
|
||||
in[2] = _mm256_srai_epi32(in[2], 7);
|
||||
in[3] = _mm256_srai_epi32(in[3], 7);
|
||||
|
||||
out[0] = _mm256_cvtepi32_ps(in[0]);
|
||||
out[1] = _mm256_cvtepi32_ps(in[1]);
|
||||
out[2] = _mm256_cvtepi32_ps(in[2]);
|
||||
|
@ -357,11 +352,11 @@ conv_s32_to_f32d_4s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
s += 8*n_channels;
|
||||
}
|
||||
for(; n < n_samples; n++) {
|
||||
__m128 out[4], factor = _mm_set1_ps(1.0f / S25_SCALE);
|
||||
out[0] = _mm_cvtsi32_ss(factor, s[0] >> 7);
|
||||
out[1] = _mm_cvtsi32_ss(factor, s[1] >> 7);
|
||||
out[2] = _mm_cvtsi32_ss(factor, s[2] >> 7);
|
||||
out[3] = _mm_cvtsi32_ss(factor, s[3] >> 7);
|
||||
__m128 out[4], factor = _mm_set1_ps(1.0f / S32_SCALE);
|
||||
out[0] = _mm_cvtsi32_ss(factor, s[0]);
|
||||
out[1] = _mm_cvtsi32_ss(factor, s[1]);
|
||||
out[2] = _mm_cvtsi32_ss(factor, s[2]);
|
||||
out[3] = _mm_cvtsi32_ss(factor, s[3]);
|
||||
out[0] = _mm_mul_ss(out[0], factor);
|
||||
out[1] = _mm_mul_ss(out[1], factor);
|
||||
out[2] = _mm_mul_ss(out[2], factor);
|
||||
|
@ -382,7 +377,7 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
float *d0 = dst[0], *d1 = dst[1];
|
||||
uint32_t n, unrolled;
|
||||
__m256i in[4];
|
||||
__m256 out[4], factor = _mm256_set1_ps(1.0f / S25_SCALE);
|
||||
__m256 out[4], factor = _mm256_set1_ps(1.0f / S32_SCALE);
|
||||
__m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels,
|
||||
4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels);
|
||||
|
||||
|
@ -396,9 +391,6 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
in[0] = _mm256_i32gather_epi32((int*)&s[0], mask1, 4);
|
||||
in[1] = _mm256_i32gather_epi32((int*)&s[1], mask1, 4);
|
||||
|
||||
in[0] = _mm256_srai_epi32(in[0], 7);
|
||||
in[1] = _mm256_srai_epi32(in[1], 7);
|
||||
|
||||
out[0] = _mm256_cvtepi32_ps(in[0]);
|
||||
out[1] = _mm256_cvtepi32_ps(in[1]);
|
||||
|
||||
|
@ -411,9 +403,9 @@ conv_s32_to_f32d_2s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
s += 8*n_channels;
|
||||
}
|
||||
for(; n < n_samples; n++) {
|
||||
__m128 out[2], factor = _mm_set1_ps(1.0f / S25_SCALE);
|
||||
out[0] = _mm_cvtsi32_ss(factor, s[0] >> 7);
|
||||
out[1] = _mm_cvtsi32_ss(factor, s[1] >> 7);
|
||||
__m128 out[2], factor = _mm_set1_ps(1.0f / S32_SCALE);
|
||||
out[0] = _mm_cvtsi32_ss(factor, s[0]);
|
||||
out[1] = _mm_cvtsi32_ss(factor, s[1]);
|
||||
out[0] = _mm_mul_ss(out[0], factor);
|
||||
out[1] = _mm_mul_ss(out[1], factor);
|
||||
_mm_store_ss(&d0[n], out[0]);
|
||||
|
@ -430,7 +422,7 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
float *d0 = dst[0];
|
||||
uint32_t n, unrolled;
|
||||
__m256i in[2];
|
||||
__m256 out[2], factor = _mm256_set1_ps(1.0f / S25_SCALE);
|
||||
__m256 out[2], factor = _mm256_set1_ps(1.0f / S32_SCALE);
|
||||
__m256i mask1 = _mm256_setr_epi32(0*n_channels, 1*n_channels, 2*n_channels, 3*n_channels,
|
||||
4*n_channels, 5*n_channels, 6*n_channels, 7*n_channels);
|
||||
|
||||
|
@ -443,9 +435,6 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
in[0] = _mm256_i32gather_epi32(&s[0*n_channels], mask1, 4);
|
||||
in[1] = _mm256_i32gather_epi32(&s[8*n_channels], mask1, 4);
|
||||
|
||||
in[0] = _mm256_srai_epi32(in[0], 7);
|
||||
in[1] = _mm256_srai_epi32(in[1], 7);
|
||||
|
||||
out[0] = _mm256_cvtepi32_ps(in[0]);
|
||||
out[1] = _mm256_cvtepi32_ps(in[1]);
|
||||
|
||||
|
@ -458,8 +447,8 @@ conv_s32_to_f32d_1s_avx2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
s += 16*n_channels;
|
||||
}
|
||||
for(; n < n_samples; n++) {
|
||||
__m128 out, factor = _mm_set1_ps(1.0f / S25_SCALE);
|
||||
out = _mm_cvtsi32_ss(factor, s[0] >> 7);
|
||||
__m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE);
|
||||
out = _mm_cvtsi32_ss(factor, s[0]);
|
||||
out = _mm_mul_ss(out, factor);
|
||||
_mm_store_ss(&d0[n], out);
|
||||
s += n_channels;
|
||||
|
|
|
@ -335,7 +335,7 @@ conv_s32_to_f32d_1s_sse2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
float *d0 = dst[0];
|
||||
uint32_t n, unrolled;
|
||||
__m128i in;
|
||||
__m128 out, factor = _mm_set1_ps(1.0f / S25_SCALE);
|
||||
__m128 out, factor = _mm_set1_ps(1.0f / S32_SCALE);
|
||||
|
||||
if (SPA_IS_ALIGNED(d0, 16))
|
||||
unrolled = n_samples & ~3;
|
||||
|
@ -347,14 +347,13 @@ conv_s32_to_f32d_1s_sse2(void *data, void * SPA_RESTRICT dst[], const void * SPA
|
|||
s[1*n_channels],
|
||||
s[2*n_channels],
|
||||
s[3*n_channels]);
|
||||
in = _mm_srai_epi32(in, 7);
|
||||
out = _mm_cvtepi32_ps(in);
|
||||
out = _mm_mul_ps(out, factor);
|
||||
_mm_store_ps(&d0[n], out);
|
||||
s += 4*n_channels;
|
||||
}
|
||||
for(; n < n_samples; n++) {
|
||||
out = _mm_cvtsi32_ss(factor, s[0]>>7);
|
||||
out = _mm_cvtsi32_ss(factor, s[0]);
|
||||
out = _mm_mul_ss(out, factor);
|
||||
_mm_store_ss(&d0[n], out);
|
||||
s += n_channels;
|
||||
|
|
|
@ -119,9 +119,10 @@
|
|||
|
||||
#define S32_TO_S24_32(v) (((int32_t)(v)) >> 8)
|
||||
|
||||
#define S32_MIN (S24_MIN * 256)
|
||||
#define S32_MAX (S24_MAX * 256)
|
||||
#define S32_TO_F32(v) ITOF(int32_t, S32_TO_S25_32(v), S25_SCALE, 0.0f)
|
||||
#define S32_MIN -2147483648
|
||||
#define S32_MAX 2147483647
|
||||
#define S32_SCALE 2147483648.0f
|
||||
#define S32_TO_F32(v) ITOF(int32_t, v, S32_SCALE, 0.0f)
|
||||
#define S32S_TO_F32(v) S32_TO_F32(bswap_32(v))
|
||||
#define F32_TO_S32_D(v,d) S25_32_TO_S32(F32_TO_S25_32_D(v,d))
|
||||
#define F32_TO_S32(v) F32_TO_S32_D(v, 0.0f)
|
||||
|
|
|
@ -336,7 +336,7 @@ static void test_s32_f32(void)
|
|||
0xFFFFFF00, 0x0200, 0xFFFFFE00
|
||||
};
|
||||
|
||||
static const float out[] = { 0.e+00f, 9.99999940395355224609375e-01f, -1.e+00f,
|
||||
static const float out[] = { 0.e+00f, 1.e+00f, -1.e+00f,
|
||||
9.9999988079071044921875e-01f, -9.9999988079071044921875e-01f, 5.e-01f,
|
||||
-5.e-01f, 5.9604644775390625e-08f, -5.9604644775390625e-08f,
|
||||
1.1920928955078125e-07f, -1.1920928955078125e-07f,
|
||||
|
@ -672,15 +672,15 @@ static void test_lossless_s32(void)
|
|||
{
|
||||
int64_t i;
|
||||
|
||||
int32_t max_abs_err = -1;
|
||||
const int32_t expected_max_abs_err = 127;
|
||||
fprintf(stderr, "test %s:\n", __func__);
|
||||
for (i = S32_MIN; i < S32_MAX; i += (expected_max_abs_err >> 1)) {
|
||||
for (i = S32_MIN; i < S32_MAX; i += 63) {
|
||||
float v = S32_TO_F32(i);
|
||||
int32_t t = F32_TO_S32(v);
|
||||
max_abs_err = SPA_MAX(max_abs_err, SPA_ABS(i - t));
|
||||
spa_assert_se(SPA_ABS(i - t) <= 126);
|
||||
// NOTE: 126 is the maximal absolute error given step=1,
|
||||
// for wider steps it may (errneously) be lower,
|
||||
// because we may not check some integer that would bump it.
|
||||
}
|
||||
spa_assert_se(max_abs_err == expected_max_abs_err);
|
||||
}
|
||||
|
||||
static void test_lossless_s32_lossless_subset(void)
|
||||
|
@ -700,7 +700,7 @@ static void test_lossless_s32_lossless_subset(void)
|
|||
}
|
||||
}
|
||||
spa_assert_se(!all_lossless);
|
||||
spa_assert_se(max_abs_err == 127);
|
||||
spa_assert_se(max_abs_err == 64);
|
||||
}
|
||||
|
||||
static void test_lossless_u32(void)
|
||||
|
|
Loading…
Reference in a new issue