fmt: improve s24 to float conversion

This commit is contained in:
Wim Taymans 2018-11-08 09:56:58 +01:00
parent 9825932b88
commit 3267667d94
2 changed files with 70 additions and 6 deletions

View file

@ -105,6 +105,49 @@ conv_s16_to_f32d_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const v
conv_s16_to_f32d_1_sse(data, n_dst, &dst[i], &s[i], n_bytes);
}
static void
conv_s24_to_f32d_1_sse(void *data, int n_dst, void *dst[n_dst], const void *src, int n_bytes)
{
const uint8_t *s = src;
float **d = (float **) dst;
float *d0 = d[0];
int n, n_samples, unrolled;
__m128i in;
__m128 out, factor = _mm_set1_ps(1.0f / S24_SCALE);
n_samples = n_bytes / (3 * n_dst);
unrolled = n_samples / 4;
n_samples = n_samples & 3;
for(n = 0; unrolled--; n += 4) {
in = _mm_set_epi32(READ24(&s[9*n_dst]),
READ24(&s[6*n_dst]),
READ24(&s[3*n_dst]),
READ24(s));
out = _mm_cvtepi32_ps(in);
out = _mm_mul_ps(out, factor);
_mm_storeu_ps(&d0[n], out);
s += 12 * n_dst;
}
for(; n_samples--; n++) {
out = _mm_cvtsi32_ss(out, READ24(s));
out = _mm_mul_ss(out, factor);
_mm_store_ss(&d0[n], out);
s += 3 * n_dst;
}
}
static void
conv_s24_to_f32d_sse(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_bytes)
{
const int8_t *s = src[0];
int i = 0;
for(; i < n_dst; i++)
conv_s24_to_f32d_1_sse(data, n_dst, &dst[i], &s[3*i], n_bytes);
}
static void
conv_f32d_to_s32_1_sse(void *data, void *dst, int n_src, const void *src[n_src], int n_bytes)
{

View file

@ -46,6 +46,29 @@
#define S32_MAX 2147483647
#define S32_SCALE 2147483647
union s24_data {
struct {
#if __BYTE_ORDER == __LITTLE_ENDIAN
uint16_t lo;
int16_t hi;
#else
int16_t hi;
uint16_t lo;
#endif
} s24;
int32_t value;
};
static inline int32_t read_s24(const void *src)
{
union s24_data d;
const uint8_t *s = src;
d.s24.hi = ((int8_t*)s)[2];
d.s24.lo = *((uint16_t*)s);
return d.value;
}
#define READ24(s) read_s24(s)
#if defined (__SSE__)
#include "fmt-ops-sse.c"
#endif
@ -192,13 +215,8 @@ conv_s32d_to_f32(void *data, int n_dst, void *dst[n_dst], int n_src, const void
}
}
#if __BYTE_ORDER == __LITTLE_ENDIAN
#define READ24(s) (((uint32_t)s[2] << 16) | ((uint32_t)s[1] << 8) | ((uint32_t)s[0]))
#else
#define READ24(s) (((uint32_t)s[0] << 16) | ((uint32_t)s[1] << 8) | ((uint32_t)s[2]))
#endif
#define S24_TO_F32(v) ((((int32_t)v)<<8) * (1.0f / S32_SCALE))
#define S24_TO_F32(v) (((int32_t)(v)) * (1.0f / S24_SCALE))
static void
conv_s24_to_f32(void *data, int n_dst, void *dst[n_dst], int n_src, const void *src[n_src], int n_bytes)
@ -711,6 +729,9 @@ static const struct conv_info {
{ SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32, 0, conv_s24_to_f32 },
{ SPA_AUDIO_FORMAT_S24P, SPA_AUDIO_FORMAT_F32P, 0, conv_s24_to_f32 },
#if defined (__SSE2__)
{ SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32P, FEATURE_SSE, conv_s24_to_f32d_sse },
#endif
{ SPA_AUDIO_FORMAT_S24, SPA_AUDIO_FORMAT_F32P, 0, conv_s24_to_f32d },
{ SPA_AUDIO_FORMAT_S24P, SPA_AUDIO_FORMAT_F32, 0, conv_s24d_to_f32 },