hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 19:24:06 +00:00
Code Issues Actions Projects Releases Wiki Activity

Fixed (some) bugs in filter_inter_halfpel_chroma and inter_recon

Browse source 
Optimizations for filter to only filter horizontal or vertical pixels
This commit is contained in:
Marko Viitanen 2013-09-24 15:43:20 +03:00
parent 469644dd5b
commit 13e058abce
3 changed files with 58 additions and 49 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

29
src/filter.c
View file

@ -371,7 +371,7 @@ void filter_deblock(encoder_control* encoder)
* \param dst_stride stride of destination image * \param dst_stride stride of destination image
* *
*/ */
void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, int height, int16_t *dst, int16_t dst_stride) void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, int height, int16_t *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag)
{ {
/* ____________ /* ____________
* | B0,0|ae0,0| * | B0,0|ae0,0|
@ -385,6 +385,7 @@ void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, in
int32_t x, y; int32_t x, y;
int32_t shift1 = g_bitdepth-8; int32_t shift1 = g_bitdepth-8;
int32_t shift2 = 6; int32_t shift2 = 6;
int32_t shift3 = 14-g_bitdepth;
// Loop source pixels and generate four filtered half-pel pixels on each round // Loop source pixels and generate four filtered half-pel pixels on each round
for (y = 0; y < height; y++) { for (y = 0; y < height; y++) {
@ -397,23 +398,31 @@ void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, in
int src_pos = src_pos_y+x; int src_pos = src_pos_y+x;
// Temporary variables.. // Temporary variables..
int ae_temp1,ae_temp2,ae_temp3; int32_t ae_temp1,ae_temp2,ae_temp3;
dst[dst_pos] = src[src_pos]; //B0,0 // Original pixel (not really needed)
dst[dst_pos + 1] = (-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1; //ae0,0 //dst[dst_pos] = src[src_pos]; //B0,0
dst[dst_pos + 1*dst_stride] = ( -4*src[src_pos - src_stride] + 36*src[src_pos] + 36*src[src_pos + src_stride] - 4*src[src_pos + 2*src_stride] ) >> shift1; //ea0,0
// We need this only when hor_flag and for ee0,0
if (hor_flag) {
dst[dst_pos + 1] = ((-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1) >> shift3; //ae0,0
} else if(ver_flag) { // This one only needed if ver_flag and !hor_flag
dst[dst_pos + 1*dst_stride] = ((-4*src[src_pos - src_stride] + 36*src[src_pos] + 36*src[src_pos + src_stride] - 4*src[src_pos + 2*src_stride] ) >> shift1) >> shift3; //ea0,0
}
// When both flags, we _only_ use this pixel (but still need ae0,0 for it)
if (hor_flag && ver_flag) {
// Calculate temporary values.. // Calculate temporary values..
//TODO: optimization, store these values //TODO: optimization, store these values
src_pos -= src_stride; //0,-1 src_pos -= src_stride; //0,-1
ae_temp1 = (-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1; //ae0,-1 ae_temp1 = ((-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1) >> shift3; //ae0,-1
src_pos += src_stride; //0,1 src_pos += src_stride; //0,1
ae_temp2 = (-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1; //ae0,1 ae_temp2 = ((-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1) >> shift3; //ae0,1
src_pos += src_stride; //0,2 src_pos += src_stride; //0,2
ae_temp2 = (-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1; //ae0,2 ae_temp2 = ((-4*src[src_pos - 1] + 36*src[src_pos] + 36*src[src_pos + 1] - 4*src[src_pos + 2]) >> shift1) >> shift3; //ae0,2
dst[dst_pos + 1*dst_stride + 1] = ( -4*ae_temp1 + 36 * dst[dst_pos+1] + 36*ae_temp2 - 4*ae_temp2) >> shift2; //ee0,0 dst[dst_pos + 1*dst_stride + 1] = (( -4*ae_temp1 + 36*dst[dst_pos + 1] + 36*ae_temp2 - 4*ae_temp2) >> shift2); //ee0,0
}
} }
} }
} }

4
src/filter.h
View file

@ -37,8 +37,8 @@ void filter_deblock_chroma(uint8_t *src, int32_t offset, int32_t tc,
int8_t part_p_nofilter, int8_t part_q_nofilter); int8_t part_p_nofilter, int8_t part_q_nofilter);
// INTERPOLATION // INTERPOLATION
void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, void filter_inter_halfpel_chroma(int16_t *src, int16_t src_stride, int width, int height,
int height, int16_t *dst, int16_t dst_stride); int16_t *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag);
// SAO // SAO

52
src/inter.c
View file

@ -68,41 +68,40 @@ void inter_recon(picture* ref,int32_t xpos, int32_t ypos,int32_t width, int16_t
int32_t ref_width_c = ref->width>>1; //!< Reference picture width in chroma pixels int32_t ref_width_c = ref->width>>1; //!< Reference picture width in chroma pixels
// negative overflow flag // negative overflow flag
int8_t overflow_neg_x = (xpos + mv[0] < 0)?1:0; int8_t overflow_neg_x = (xpos + (mv[0]>>2) < 0)?1:0;
int8_t overflow_neg_y = (ypos + mv[1] < 0)?1:0; int8_t overflow_neg_y = (ypos + (mv[1]>>2) < 0)?1:0;
// positive overflow flag // positive overflow flag
int8_t overflow_pos_x = (xpos + mv[0] + width > ref->width )?1:0; int8_t overflow_pos_x = (xpos + (mv[0]>>2) + width >= ref->width )?1:0;
int8_t overflow_pos_y = (ypos + mv[1] + width > ref->height)?1:0; int8_t overflow_pos_y = (ypos + (mv[1]>>2) + width >= ref->height)?1:0;
// Chroma half-pel // Chroma half-pel
#define HALFPEL_CHROMA_WIDTH ((LCU_WIDTH>>1) + 8) #define HALFPEL_CHROMA_WIDTH ((LCU_WIDTH>>1) + 8)
int8_t chroma_halfpel = (mv[0]&4 || mv[1]&4); //!< third lsb is set -> chroma is half-pel int8_t chroma_halfpel = ((mv[0]>>2)&1) | ((mv[1]>>2)&1); //!< third lsb is set -> chroma is half-pel
int16_t halfpel_src_u[HALFPEL_CHROMA_WIDTH * HALFPEL_CHROMA_WIDTH]; int16_t halfpel_src_u[HALFPEL_CHROMA_WIDTH * HALFPEL_CHROMA_WIDTH];
int16_t halfpel_src_v[HALFPEL_CHROMA_WIDTH * HALFPEL_CHROMA_WIDTH]; int16_t halfpel_src_v[HALFPEL_CHROMA_WIDTH * HALFPEL_CHROMA_WIDTH];
int16_t *halfpel_src_off_u = &halfpel_src_u[HALFPEL_CHROMA_WIDTH*4 + 4]; //!< halfpel_src_u with offset (-4,-4) int16_t *halfpel_src_off_u = &halfpel_src_u[HALFPEL_CHROMA_WIDTH*4 + 4]; //!< halfpel_src_u with offset (-4,-4)
int16_t *halfpel_src_off_v = &halfpel_src_v[HALFPEL_CHROMA_WIDTH*4 + 4]; //!< halfpel_src_v with offset (-4,-4) int16_t *halfpel_src_off_v = &halfpel_src_v[HALFPEL_CHROMA_WIDTH*4 + 4]; //!< halfpel_src_v with offset (-4,-4)
int16_t halfpel_u[LCU_WIDTH * LCU_WIDTH]; int16_t halfpel_u[LCU_WIDTH * LCU_WIDTH]; //!< interpolated 2X x 2Y block (u)
int16_t halfpel_v[LCU_WIDTH * LCU_WIDTH]; int16_t halfpel_v[LCU_WIDTH * LCU_WIDTH]; //!< interpolated 2X x 2Y block (v)
// TODO: Fractional pixel support // TODO: Fractional pixel support
mv[0] = mv[0]>>2; mv[0] = mv[0]>>2;
mv[1] = mv[1]>>2; mv[1] = mv[1]>>2;
// Chroma half-pel // Chroma half-pel
// get half-pel interpolated block and push it to output
if(chroma_halfpel) { if(chroma_halfpel) {
int halfpel_y,halfpel_x; int halfpel_y,halfpel_x;
int abs_mv_x = abs(mv[0])&1; int abs_mv_x = mv[0]&1;
int abs_mv_y = abs(mv[1])&1; int abs_mv_y = mv[1]&1;
for (halfpel_y = 0, y = (ypos>>1) - 4; y < ((ypos + width)>>1) + 4; halfpel_y++, y++) { for (halfpel_y = 0, y = (ypos>>1) - 4; y < ((ypos + width)>>1) + 4; halfpel_y++, y++) {
// calculate y-pixel offset // calculate y-pixel offset
coord_y = (y + (mv[1]>>1)); coord_y = (y + (mv[1]>>1));
overflow_neg_y = (y + (mv[1]>>1) < 0)?1:0;
overflow_pos_y = (coord_y >= ref->height>>1)?1:0;
// On y-overflow set coord_y accordingly // On y-overflow set coord_y accordingly
overflow_neg_y = (coord_y < 0)?1:0;
overflow_pos_y = (coord_y >= ref->height>>1)?1:0;
if (overflow_neg_y) coord_y = 0; if (overflow_neg_y) coord_y = 0;
else if (overflow_pos_y) coord_y = ((ref->height>>1) - 1); else if (overflow_pos_y) coord_y = ((ref->height>>1) - 1);
@ -111,26 +110,27 @@ void inter_recon(picture* ref,int32_t xpos, int32_t ypos,int32_t width, int16_t
for (halfpel_x = 0, x = (xpos>>1) - 4; x < ((xpos + width)>>1) + 4; halfpel_x++, x++) { for (halfpel_x = 0, x = (xpos>>1) - 4; x < ((xpos + width)>>1) + 4; halfpel_x++, x++) {
coord_x = x + (mv[0]>>1); coord_x = x + (mv[0]>>1);
// On x-overflow set coord_x accordingly
overflow_neg_x = (coord_x < 0) ? 1 : 0; overflow_neg_x = (coord_x < 0) ? 1 : 0;
overflow_pos_x = (coord_x >= ref->width>>1) ? 1 : 0; overflow_pos_x = (coord_x >= ref->width>>1) ? 1 : 0;
// On x-overflow set coord_x accordingly
if (overflow_neg_x) coord_x = 0; if (overflow_neg_x) coord_x = 0;
else if (overflow_pos_x) coord_x = (ref->width>>1) - 1; else if (overflow_pos_x) coord_x = (ref->width>>1) - 1;
// Store source block data (with extended borders)
halfpel_src_u[halfpel_y*HALFPEL_CHROMA_WIDTH + halfpel_x] = ref->u_recdata[coord_y + coord_x]; halfpel_src_u[halfpel_y*HALFPEL_CHROMA_WIDTH + halfpel_x] = ref->u_recdata[coord_y + coord_x];
halfpel_src_u[halfpel_y*HALFPEL_CHROMA_WIDTH + halfpel_x] = ref->v_recdata[coord_y + coord_x]; halfpel_src_v[halfpel_y*HALFPEL_CHROMA_WIDTH + halfpel_x] = ref->v_recdata[coord_y + coord_x];
} }
} }
filter_inter_halfpel_chroma(halfpel_src_off_u, HALFPEL_CHROMA_WIDTH, width>>1, width>>1, halfpel_u, LCU_WIDTH);
filter_inter_halfpel_chroma(halfpel_src_off_v, HALFPEL_CHROMA_WIDTH, width>>1, width>>1, halfpel_v, LCU_WIDTH);
for (y = ypos>>1; y < (ypos + width)>>1; y++) { // Filter the block to half-pel resolution
halfpel_y = (y-(ypos>>1))<<1; filter_inter_halfpel_chroma(halfpel_src_off_u, HALFPEL_CHROMA_WIDTH, width>>1, width>>1, halfpel_u, LCU_WIDTH, abs_mv_x, abs_mv_y);
coord_y = (y + (mv[1]>>1)) * ref_width_c; // pre-calculate filter_inter_halfpel_chroma(halfpel_src_off_v, HALFPEL_CHROMA_WIDTH, width>>1, width>>1, halfpel_v, LCU_WIDTH, abs_mv_x, abs_mv_y);
for (x = xpos>>1; x < (xpos + width)>>1; x++) {
halfpel_x = (x-(xpos>>1))<<1; // Assign filtered pixels to output
dst->u_recdata[y*dst_width_c + x] = (uint8_t)halfpel_u[(halfpel_y + abs_mv_y)*LCU_WIDTH + halfpel_x + abs_mv_x]; for (halfpel_y = abs_mv_y, y = ypos>>1; y < (ypos + width)>>1; halfpel_y += 2, y++) {
dst->v_recdata[y*dst_width_c + x] = (uint8_t)halfpel_v[(halfpel_y + abs_mv_y)*LCU_WIDTH + halfpel_x + abs_mv_x]; for (halfpel_x = abs_mv_x, x = xpos>>1; x < (xpos + width)>>1; halfpel_x += 2, x++) {
dst->u_recdata[y*dst_width_c + x] = (uint8_t)halfpel_u[halfpel_y*LCU_WIDTH + halfpel_x];
dst->v_recdata[y*dst_width_c + x] = (uint8_t)halfpel_v[halfpel_y*LCU_WIDTH + halfpel_x];
} }
} }
} }
@ -167,7 +167,7 @@ void inter_recon(picture* ref,int32_t xpos, int32_t ypos,int32_t width, int16_t
} }
} }
if(chroma_halfpel) { if(!chroma_halfpel) {
// Copy Chroma with boundary checking // Copy Chroma with boundary checking
// TODO: chroma fractional pixel interpolation // TODO: chroma fractional pixel interpolation
for (y = ypos>>1; y < (ypos + width)>>1; y++) { for (y = ypos>>1; y < (ypos + width)>>1; y++) {
@ -211,7 +211,7 @@ void inter_recon(picture* ref,int32_t xpos, int32_t ypos,int32_t width, int16_t
} }
} }
if(chroma_halfpel) { if(!chroma_halfpel) {
// Copy Chroma // Copy Chroma
// TODO: chroma fractional pixel interpolation // TODO: chroma fractional pixel interpolation
for (y = ypos>>1; y < (ypos + width)>>1; y++) { for (y = ypos>>1; y < (ypos + width)>>1; y++) {