eessppeelllloo/veejay
1
0
Fork 0
mirror of https://github.com/game-stop/veejay.git synced 2026-01-06 15:05:28 +01:00
Code Issues Packages Projects Releases Wiki Activity

fix slow (interpolation) ALT + [number], optimize for SSE/AVX and NEON

Browse Source
This commit is contained in:
niels
2025-12-08 02:07:44 +01:00
parent 8b05cb7603
commit 1bbe8dd413
1 changed files with 134 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

172
veejay-current/veejay-core/libyuv/yuvconv.c
View File

@@ -39,6 +39,15 @@
#include <assert.h>
#endif
#if defined(HAVE_ASM_SSE2) || defined(HAVE_ASM_SSE4_1) || defined(HAVE_ASM_AVX) || defined(HAVE_ASM_AVX2)
#include <emmintrin.h> // SSE2
#include <tmmintrin.h> // SSSE3 (for _mm_srai_epi16)
#include <smmintrin.h> // SSE4.1
#include <immintrin.h> // AVX
#define Q_SHIFT 15
#define Q_MAX 32767
#endif
#define Y4M_CHROMA_420JPEG 0 /* 4:2:0, H/V centered, for JPEG/MPEG-1 */
#define Y4M_CHROMA_420MPEG2 1 /* 4:2:0, H cosited, for MPEG-2 */
#define Y4M_CHROMA_420PALDV 2 /* 4:2:0, alternating Cb/Cr, for PAL-DV */
@@ -1013,42 +1022,35 @@ void yuv422_to_yuyv(uint8_t *yuv422[3], uint8_t *pixels, int w, int h)
#ifdef HAVE_ARM
void yuv422_to_yuyv(uint8_t *yuv422[3], uint8_t *pixels, int w, int h) {
int x, y;
uint8_t *Y = yuv422[0];
uint8_t *U = yuv422[1];
uint8_t *V = yuv422[2];
uint8_t *Y_plane = yuv422[0];
uint8_t *U_plane = yuv422[1];
uint8_t *V_plane = yuv422[2];
for (y = 0; y < h; y++) {
Y = yuv422[0] + y * w;
U = yuv422[1] + (y >> 1) * w;
V = yuv422[2] + (y >> 1) * w;
uint8_t *Y = Y_plane + y * w;
uint8_t *U = U_plane + (y >> 1) * w;
uint8_t *V = V_plane + (y >> 1) * w;
uint8_t *dst = pixels + y * w * 2;
for (x = 0; x < w; x += 16) {
uint8x8_t y0 = vld1_u8(Y);
uint8x8_t u0 = vld1_u8(U);
uint8x8_t y1 = vld1_u8(Y + 8);
uint8x8_t v0 = vld1_u8(V);
for (x = 0; x < w; x += 16) {
uint8x16_t y_data = vld1q_u8(Y);
uint8x8_t u_data = vld1_u8(U);
uint8x8_t v_data = vld1_u8(V);
uint8x8x2_t yuyv0;
yuyv0.val[0] = y0;
yuyv0.val[1] = u0;
yuyv0.val[0] = vext_u8(y0, v0, 1);
uint8x8x2_t uv_pair = vzip_u8(u_data, v_data);
uint8x8x2_t yuyv1;
yuyv1.val[0] = y1;
yuyv1.val[1] = vext_u8(u0, y1, 7);
yuyv1.val[0] = v0;
vst2_u8(pixels, yuyv0);
vst2_u8(pixels + 16, yuyv1);
pixels += 32;
uint8x16x2_t yuyv_pair = vzipq_u8(y_data, vcombine_u8(uv_pair.val[0], uv_pair.val[1]));
vst1q_u8(dst, yuyv_pair.val[0]);
vst1q_u8(dst + 16, yuyv_pair.val[1]);
dst += 32;
Y += 16;
U += 8;
V += 8;
}
}
}
#endif
@@ -1571,21 +1573,36 @@ void yuv420to422planar(uint8_t *src[3], uint8_t *dst[3], int w, int h) {
const int wid = w >> 1;
const int hei = h >> 1;
uint8_t *u_dst = dst[1];
uint8_t *v_dst = dst[2];
uint8_t *u_src = src[1];
uint8_t *v_src = src[2];
uint8_t *u_dst = dst[1];
uint8_t *v_dst = dst[2];
for (int y = 0; y < hei; y++) {
uint8x8_t u_values = vdup_n_u8(u_src[y * wid]);
uint8x8_t v_values = vdup_n_u8(v_src[y * wid]);
uint8_t *a = u_src + y * wid;
uint8_t *b = v_src + y * wid;
uint8_t *u_even = u_dst + (y << 1) * wid;
uint8_t *u_odd = u_dst + ((y << 1) + 1) * wid;
uint8_t *v_even = v_dst + (y << 1) * wid;
uint8_t *v_odd = v_dst + ((y << 1) + 1) * wid;
for (int x = 0; x < wid; x += 8) {
vst1_u8(u_dst, u_values);
vst1_u8(v_dst, v_values);
uint8x8_t u_values = vld1_u8(a);
uint8x8_t v_values = vld1_u8(b);
u_dst += 8;
v_dst += 8;
vst1_u8(u_even, u_values);
vst1_u8(v_even, v_values);
vst1_u8(u_odd, u_values);
vst1_u8(v_odd, v_values);
a += 8;
b += 8;
u_even += 8;
u_odd += 8;
v_even += 8;
v_odd += 8;
}
}
}
@@ -1615,7 +1632,7 @@ void yuv420to422planar( uint8_t *src[3], uint8_t *dst[3], int w, int h )
uint8_t *v2 = dst[2];
for( y = 0; y < hei; y ++ ) {
u = dst[1] + ( (y << 1 ) * wid );
u2 = dst[1] + (( (y+1)<<1) * wid);
u2 = dst[1] + ( ((y << 1) + 1) * wid );
a = src[1] + ( y * wid );
for( x = 0; x < work; x ++ ) {
copy8( u,u2, a );
@@ -1626,7 +1643,7 @@ void yuv420to422planar( uint8_t *src[3], uint8_t *dst[3], int w, int h )
}
for( y = 0; y < hei; y ++ ) {
v = dst[2] + ( (y << 1 ) * wid );
v2 = dst[2] + (( (y+1)<<1) * wid );
v2 = dst[2] + ( (( y << 1 ) + 1 ) * wid );
b = src[2] + ( y * wid );
for( x = 0; x < work; x ++ ) {
copy8( v,v2, b );
@@ -1828,12 +1845,91 @@ void yuv444_yvu444_1plane(
}
#ifdef HAVE_ARM_NEON
void yuv_interpolate_frames( uint8_t *dst, uint8_t *a, uint8_t *b, const int len, const float frac )
{
int i;
#pragma omp simd
for( i = 0; i < len; i ++ ) {
dst[i] = a[i] + ( frac * ( b[i] - a[i] ) );
const int step = 16;
const int16_t fixed_frac = (int16_t)(frac * 32767.0f);
int16x8_t v_frac = vdupq_n_s16(fixed_frac);
for( i = 0; i < len - step; i += step )
{
uint8x16_t va_u8 = vld1q_u8(a + i);
uint8x16_t vb_u8 = vld1q_u8(b + i);
int16x8_t va_low = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(va_u8)));
int16x8_t va_high = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(va_u8)));
int16x8_t vb_low = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vb_u8)));
int16x8_t vb_high = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(vb_u8)));
int16x8_t diff_low = vsubq_s16(vb_low, va_low);
int16x8_t diff_high = vsubq_s16(vb_high, va_high);
int16x8_t term_low = vshrq_n_s16(vmulq_s16(diff_low, v_frac), 15);
int16x8_t term_high = vshrq_n_s16(vmulq_s16(diff_high, v_frac), 15);
int16x8_t result_low = vaddq_s16(va_low, term_low);
int16x8_t result_high = vaddq_s16(va_high, term_high);
uint8x16_t vdst_u8 = vcombine_u8(vmovnq_u16(vreinterpretq_u16_s16(result_low)),
vmovnq_u16(vreinterpretq_u16_s16(result_high)));
vst1q_u8(dst + i, vdst_u8);
}
for( ; i < len; i++ ) {
dst[i] = a[i] + ( frac * ( b[i] - a[i] ) );
}
}
#elif defined(HAVE_ASM_SSE2) || defined(HAVE_ASM_SSE4_1) || defined(HAVE_ASM_AVX) || defined(HAVE_ASM_AVX2)
void yuv_interpolate_frames( uint8_t *dst, uint8_t *a, uint8_t *b, const int len, const float frac )
{
int i;
const int step = 16;
const int16_t fixed_frac = (int16_t)(frac * (float)Q_MAX);
__m128i v_frac = _mm_set1_epi16(fixed_frac);
__m128i v_zero = _mm_setzero_si128();
for( i = 0; i < len - step; i += step )
{
__m128i va_u8 = _mm_loadu_si128((__m128i const*)(a + i));
__m128i vb_u8 = _mm_loadu_si128((__m128i const*)(b + i));
__m128i va_low = _mm_unpacklo_epi8(va_u8, v_zero);
__m128i va_high = _mm_unpackhi_epi8(va_u8, v_zero);
__m128i vb_low = _mm_unpacklo_epi8(vb_u8, v_zero);
__m128i vb_high = _mm_unpackhi_epi8(vb_u8, v_zero);
__m128i diff_low = _mm_sub_epi16(vb_low, va_low);
__m128i diff_high = _mm_sub_epi16(vb_high, va_high);
__m128i term_mul_low = _mm_mullo_epi16(diff_low, v_frac);
__m128i term_mul_high = _mm_mullo_epi16(diff_high, v_frac);
__m128i term_low = _mm_srai_epi16(term_mul_low, Q_SHIFT);
__m128i term_high = _mm_srai_epi16(term_mul_high, Q_SHIFT);
__m128i result_low = _mm_add_epi16(va_low, term_low);
__m128i result_high = _mm_add_epi16(va_high, term_high);
__m128i vdst_u8 = _mm_packus_epi16(result_low, result_high);
_mm_storeu_si128((__m128i *)(dst + i), vdst_u8);
}
for( ; i < len; i++ ) {
dst[i] = a[i] + ( frac * ( b[i] - a[i] ) );
}
}
#else
void yuv_interpolate_frames( uint8_t *dst, uint8_t *a, uint8_t *b, const int len, const float frac )
{
int i;
for( i = 0; i < len; i ++) {
dst[i] = a[i] + ( frac * ( b[i] - a[i] ) );
}
}
#endif