/*
- * Motion estimation
- * Copyright (c) 2000,2001 Gerard Lantau.
- *
+ * Motion estimation
+ * Copyright (c) 2000,2001 Fabrice Bellard
+ * Copyright (c) 2002-2004 Michael Niedermayer
*
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
+ * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
*
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
*
- * You should have received a copy of the GNU General Public License
- * along with this program; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
*
- * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
+
+/**
+ * @file libavcodec/motion_est.c
+ * Motion estimation.
+ */
+
#include <stdlib.h>
#include <stdio.h>
+#include <limits.h>
#include "avcodec.h"
#include "dsputil.h"
+#include "mathops.h"
#include "mpegvideo.h"
-#define ABS(a) ((a)>0 ? (a) : -(a))
-#define MAX(a,b) ((a) > (b) ? (a) : (b))
-#define INTER_BIAS 257
+#undef NDEBUG
+#include <assert.h>
-static int halfpel_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int dmin,
- int xmin, int ymin, int xmax, int ymax,
- int pred_x, int pred_y, uint8_t *ref_picture);
+#define SQ(a) ((a)*(a))
-static int pix_sum(UINT8 * pix, int line_size)
-{
- int s, i, j;
+#define P_LEFT P[1]
+#define P_TOP P[2]
+#define P_TOPRIGHT P[3]
+#define P_MEDIAN P[4]
+#define P_MV1 P[9]
- s = 0;
- for (i = 0; i < 16; i++) {
- for (j = 0; j < 16; j += 8) {
- s += pix[0];
- s += pix[1];
- s += pix[2];
- s += pix[3];
- s += pix[4];
- s += pix[5];
- s += pix[6];
- s += pix[7];
- pix += 8;
- }
- pix += line_size - 16;
- }
- return s;
-}
+static inline int sad_hpel_motion_search(MpegEncContext * s,
+ int *mx_ptr, int *my_ptr, int dmin,
+ int src_index, int ref_index,
+ int size, int h);
-static int pix_dev(UINT8 * pix, int line_size, int mean)
+static inline int update_map_generation(MotionEstContext *c)
{
- int s, i, j;
-
- s = 0;
- for (i = 0; i < 16; i++) {
- for (j = 0; j < 16; j += 8) {
- s += ABS(pix[0]-mean);
- s += ABS(pix[1]-mean);
- s += ABS(pix[2]-mean);
- s += ABS(pix[3]-mean);
- s += ABS(pix[4]-mean);
- s += ABS(pix[5]-mean);
- s += ABS(pix[6]-mean);
- s += ABS(pix[7]-mean);
- pix += 8;
- }
- pix += line_size - 16;
+ c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
+ if(c->map_generation==0){
+ c->map_generation= 1<<(ME_MAP_MV_BITS*2);
+ memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
}
- return s;
+ return c->map_generation;
}
-static int pix_norm1(UINT8 * pix, int line_size)
-{
- int s, i, j;
- UINT32 *sq = squareTbl + 256;
+/* shape adaptive search stuff */
+typedef struct Minima{
+ int height;
+ int x, y;
+ int checked;
+}Minima;
- s = 0;
- for (i = 0; i < 16; i++) {
- for (j = 0; j < 16; j += 8) {
- s += sq[pix[0]];
- s += sq[pix[1]];
- s += sq[pix[2]];
- s += sq[pix[3]];
- s += sq[pix[4]];
- s += sq[pix[5]];
- s += sq[pix[6]];
- s += sq[pix[7]];
- pix += 8;
- }
- pix += line_size - 16;
- }
- return s;
-}
+static int minima_cmp(const void *a, const void *b){
+ const Minima *da = (const Minima *) a;
+ const Minima *db = (const Minima *) b;
-static int pix_norm(UINT8 * pix1, UINT8 * pix2, int line_size)
-{
- int s, i, j;
- UINT32 *sq = squareTbl + 256;
+ return da->height - db->height;
+}
- s = 0;
- for (i = 0; i < 16; i++) {
- for (j = 0; j < 16; j += 8) {
- s += sq[pix1[0] - pix2[0]];
- s += sq[pix1[1] - pix2[1]];
- s += sq[pix1[2] - pix2[2]];
- s += sq[pix1[3] - pix2[3]];
- s += sq[pix1[4] - pix2[4]];
- s += sq[pix1[5] - pix2[5]];
- s += sq[pix1[6] - pix2[6]];
- s += sq[pix1[7] - pix2[7]];
- pix1 += 8;
- pix2 += 8;
- }
- pix1 += line_size - 16;
- pix2 += line_size - 16;
+#define FLAG_QPEL 1 //must be 1
+#define FLAG_CHROMA 2
+#define FLAG_DIRECT 4
+
+static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
+ const int offset[3]= {
+ y*c-> stride + x,
+ ((y*c->uvstride + x)>>1),
+ ((y*c->uvstride + x)>>1),
+ };
+ int i;
+ for(i=0; i<3; i++){
+ c->src[0][i]= src [i] + offset[i];
+ c->ref[0][i]= ref [i] + offset[i];
+ }
+ if(ref_index){
+ for(i=0; i<3; i++){
+ c->ref[ref_index][i]= ref2[i] + offset[i];
+ }
}
- return s;
}
-static void no_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr)
-{
- *mx_ptr = 16 * s->mb_x;
- *my_ptr = 16 * s->mb_y;
+static int get_flags(MotionEstContext *c, int direct, int chroma){
+ return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
+ + (direct ? FLAG_DIRECT : 0)
+ + (chroma ? FLAG_CHROMA : 0);
}
-static int full_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y;
- int mx, my, dmin, d;
- UINT8 *pix;
-
- xx = 16 * s->mb_x;
- yy = 16 * s->mb_y;
- x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
- if (x1 < xmin)
- x1 = xmin;
- x2 = xx + range - 1;
- if (x2 > xmax)
- x2 = xmax;
- y1 = yy - range + 1;
- if (y1 < ymin)
- y1 = ymin;
- y2 = yy + range - 1;
- if (y2 > ymax)
- y2 = ymax;
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
- dmin = 0x7fffffff;
- mx = 0;
- my = 0;
- for (y = y1; y <= y2; y++) {
- for (x = x1; x <= x2; x++) {
- d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
- s->linesize);
- if (d < dmin ||
- (d == dmin &&
- (abs(x - xx) + abs(y - yy)) <
- (abs(mx - xx) + abs(my - yy)))) {
- dmin = d;
- mx = x;
- my = y;
- }
- }
+/*! \brief compares a block (either a full macroblock or a partition thereof)
+ against a proposed motion-compensated prediction of that block
+ */
+static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
+ const int size, const int h, int ref_index, int src_index,
+ me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
+ MotionEstContext * const c= &s->me;
+ const int stride= c->stride;
+ const int uvstride= c->uvstride;
+ const int qpel= flags&FLAG_QPEL;
+ const int chroma= flags&FLAG_CHROMA;
+ const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
+ const int hx= subx + (x<<(1+qpel));
+ const int hy= suby + (y<<(1+qpel));
+ uint8_t * const * const ref= c->ref[ref_index];
+ uint8_t * const * const src= c->src[src_index];
+ int d;
+ //FIXME check chroma 4mv, (no crashes ...)
+ if(flags&FLAG_DIRECT){
+ assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
+ if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
+ const int time_pp= s->pp_time;
+ const int time_pb= s->pb_time;
+ const int mask= 2*qpel+1;
+ if(s->mv_type==MV_TYPE_8X8){
+ int i;
+ for(i=0; i<4; i++){
+ int fx = c->direct_basis_mv[i][0] + hx;
+ int fy = c->direct_basis_mv[i][1] + hy;
+ int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
+ int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
+ int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
+ int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
+
+ uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
+ if(qpel){
+ c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
+ c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
+ }else{
+ c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
+ c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
+ }
+ }
+ }else{
+ int fx = c->direct_basis_mv[0][0] + hx;
+ int fy = c->direct_basis_mv[0][1] + hy;
+ int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
+ int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
+ int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
+ int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
+
+ if(qpel){
+ c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
+ c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
+ c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
+ c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
+ c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
+ c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
+ c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
+ c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
+ }else{
+ assert((fx>>1) + 16*s->mb_x >= -16);
+ assert((fy>>1) + 16*s->mb_y >= -16);
+ assert((fx>>1) + 16*s->mb_x <= s->width);
+ assert((fy>>1) + 16*s->mb_y <= s->height);
+ assert((bx>>1) + 16*s->mb_x >= -16);
+ assert((by>>1) + 16*s->mb_y >= -16);
+ assert((bx>>1) + 16*s->mb_x <= s->width);
+ assert((by>>1) + 16*s->mb_y <= s->height);
+
+ c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
+ c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
+ }
+ }
+ d = cmp_func(s, c->temp, src[0], stride, 16);
+ }else
+ d= 256*256*256*32;
+ }else{
+ int uvdxy; /* no, it might not be used uninitialized */
+ if(dxy){
+ if(qpel){
+ c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
+ if(chroma){
+ int cx= hx/2;
+ int cy= hy/2;
+ cx= (cx>>1)|(cx&1);
+ cy= (cy>>1)|(cy&1);
+ uvdxy= (cx&1) + 2*(cy&1);
+ //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
+ }
+ }else{
+ c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
+ if(chroma)
+ uvdxy= dxy | (x&1) | (2*(y&1));
+ }
+ d = cmp_func(s, c->temp, src[0], stride, h);
+ }else{
+ d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
+ if(chroma)
+ uvdxy= (x&1) + 2*(y&1);
+ }
+ if(chroma){
+ uint8_t * const uvtemp= c->temp + 16*stride;
+ c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
+ c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
+ d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
+ d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
+ }
}
-
- *mx_ptr = mx;
- *my_ptr = my;
-
#if 0
- if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
- *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
- fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
+ if(full_pel){
+ const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
+ score_map[index]= d;
}
-#endif
- return dmin;
-}
-
-static int log_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y;
- int mx, my, dmin, d;
- UINT8 *pix;
-
- xx = s->mb_x << 4;
- yy = s->mb_y << 4;
-
- /* Left limit */
- x1 = xx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- /* Right limit */
- x2 = xx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- /* Upper limit */
- y1 = yy - range;
- if (y1 < ymin)
- y1 = ymin;
-
- /* Lower limit */
- y2 = yy + range;
- if (y2 > ymax)
- y2 = ymax;
-
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
- dmin = 0x7fffffff;
- mx = 0;
- my = 0;
-
- do {
- for (y = y1; y <= y2; y += range) {
- for (x = x1; x <= x2; x += range) {
- d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
- if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dmin = d;
- mx = x;
- my = y;
- }
- }
- }
-
- range = range >> 1;
-
- x1 = mx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- x2 = mx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- y1 = my - range;
- if (y1 < ymin)
- y1 = ymin;
-
- y2 = my + range;
- if (y2 > ymax)
- y2 = ymax;
-
- } while (range >= 1);
-
-#ifdef DEBUG
- fprintf(stderr, "log - MX: %d\tMY: %d\n", mx, my);
+ d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
#endif
- *mx_ptr = mx;
- *my_ptr = my;
- return dmin;
+ return d;
}
-static int phods_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
- int mx, my, dminx, dminy;
- UINT8 *pix;
-
- xx = s->mb_x << 4;
- yy = s->mb_y << 4;
-
- /* Left limit */
- x1 = xx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- /* Right limit */
- x2 = xx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- /* Upper limit */
- y1 = yy - range;
- if (y1 < ymin)
- y1 = ymin;
-
- /* Lower limit */
- y2 = yy + range;
- if (y2 > ymax)
- y2 = ymax;
-
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
- mx = 0;
- my = 0;
-
- x = xx;
- y = yy;
- do {
- dminx = 0x7fffffff;
- dminy = 0x7fffffff;
-
- lastx = x;
- for (x = x1; x <= x2; x += range) {
- d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
- if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dminx = d;
- mx = x;
- }
- }
-
- x = lastx;
- for (y = y1; y <= y2; y += range) {
- d = pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
- if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dminy = d;
- my = y;
- }
- }
-
- range = range >> 1;
-
- x = mx;
- y = my;
- x1 = mx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- x2 = mx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- y1 = my - range;
- if (y1 < ymin)
- y1 = ymin;
-
- y2 = my + range;
- if (y2 > ymax)
- y2 = ymax;
-
- } while (range >= 1);
-
-#ifdef DEBUG
- fprintf(stderr, "phods - MX: %d\tMY: %d\n", mx, my);
-#endif
+#include "motion_est_template.c"
- /* half pixel search */
- *mx_ptr = mx;
- *my_ptr = my;
- return dminy;
+static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
+ return 0;
}
-
-#define Z_THRESHOLD 256
-
-#define CHECK_MV(x,y)\
-{\
- d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
- d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
- if(d<dmin){\
- best[0]=x;\
- best[1]=y;\
- dmin=d;\
- }\
+static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
}
-#define CHECK_MV_DIR(x,y,new_dir)\
-{\
- d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
- d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
- if(d<dmin){\
- best[0]=x;\
- best[1]=y;\
- dmin=d;\
- next_dir= new_dir;\
- }\
-}
+int ff_init_me(MpegEncContext *s){
+ MotionEstContext * const c= &s->me;
+ int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
+ int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
-#define CHECK_MV4(x,y)\
-{\
- d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
- d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
- if(d<dmin){\
- best[0]=x;\
- best[1]=y;\
- dmin=d;\
- }\
-}
+ if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
+ av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
+ return -1;
+ }
+ //special case of snow is needed because snow uses its own iterative ME code
+ if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
+ av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
+ return -1;
+ }
-#define CHECK_MV4_DIR(x,y,new_dir)\
-{\
- d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
- d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
- if(d<dmin){\
- best[0]=x;\
- best[1]=y;\
- dmin=d;\
- next_dir= new_dir;\
- }\
-}
+ c->avctx= s->avctx;
+ if(cache_size < 2*dia_size && !c->stride){
+ av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
+ }
-#define check(x,y,S,v)\
-if( (x)<(xmin<<(S)) ) printf("%d %d %d %d xmin" #v, (x), (y), s->mb_x, s->mb_y);\
-if( (x)>(xmax<<(S)) ) printf("%d %d %d %d xmax" #v, (x), (y), s->mb_x, s->mb_y);\
-if( (y)<(ymin<<(S)) ) printf("%d %d %d %d ymin" #v, (x), (y), s->mb_x, s->mb_y);\
-if( (y)>(ymax<<(S)) ) printf("%d %d %d %d ymax" #v, (x), (y), s->mb_x, s->mb_y);\
+ ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
+ ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
+ ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
+ ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
+
+ c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
+ c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
+ c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
+
+/*FIXME s->no_rounding b_type*/
+ if(s->flags&CODEC_FLAG_QPEL){
+ c->sub_motion_search= qpel_motion_search;
+ c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
+ if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
+ else c->qpel_put= s->dsp.put_qpel_pixels_tab;
+ }else{
+ if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
+ c->sub_motion_search= hpel_motion_search;
+ else if( c->avctx->me_sub_cmp == FF_CMP_SAD
+ && c->avctx-> me_cmp == FF_CMP_SAD
+ && c->avctx-> mb_cmp == FF_CMP_SAD)
+ c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
+ else
+ c->sub_motion_search= hpel_motion_search;
+ }
+ c->hpel_avg= s->dsp.avg_pixels_tab;
+ if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
+ else c->hpel_put= s->dsp.put_pixels_tab;
+ if(s->linesize){
+ c->stride = s->linesize;
+ c->uvstride= s->uvlinesize;
+ }else{
+ c->stride = 16*s->mb_width + 32;
+ c->uvstride= 8*s->mb_width + 16;
+ }
-static inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
- UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
- int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
- int xmin, int ymin, int xmax, int ymax, int shift)
-{
- int next_dir=-1;
-
- for(;;){
- int d;
- const int dir= next_dir;
- const int x= best[0];
- const int y= best[1];
- next_dir=-1;
-
-//printf("%d", dir);
- if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
- if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
- if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
- if(dir!=1 && y<ymax) CHECK_MV_DIR(x , y+1, 3)
-
- if(next_dir==-1){
- return dmin;
+ /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
+ * not have yet, and even if we had, the motion estimation code
+ * does not expect it. */
+ if(s->codec_id != CODEC_ID_SNOW){
+ if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
+ s->dsp.me_cmp[2]= zero_cmp;
+ }
+ if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
+ s->dsp.me_sub_cmp[2]= zero_cmp;
}
+ c->hpel_put[2][0]= c->hpel_put[2][1]=
+ c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
}
-/* for(;;){
- int d;
- const int x= best[0];
- const int y= best[1];
- const int last_min=dmin;
- if(x>xmin) CHECK_MV(x-1, y )
- if(y>xmin) CHECK_MV(x , y-1)
- if(x<xmax) CHECK_MV(x+1, y )
- if(y<xmax) CHECK_MV(x , y+1)
- if(x>xmin && y>ymin) CHECK_MV(x-1, y-1)
- if(x>xmin && y<ymax) CHECK_MV(x-1, y+1)
- if(x<xmax && y>ymin) CHECK_MV(x+1, y-1)
- if(x<xmax && y<ymax) CHECK_MV(x+1, y+1)
- if(x-1>xmin) CHECK_MV(x-2, y )
- if(y-1>xmin) CHECK_MV(x , y-2)
- if(x+1<xmax) CHECK_MV(x+2, y )
- if(y+1<xmax) CHECK_MV(x , y+2)
- if(x-1>xmin && y-1>ymin) CHECK_MV(x-2, y-2)
- if(x-1>xmin && y+1<ymax) CHECK_MV(x-2, y+2)
- if(x+1<xmax && y-1>ymin) CHECK_MV(x+2, y-2)
- if(x+1<xmax && y+1<ymax) CHECK_MV(x+2, y+2)
- if(dmin==last_min) return dmin;
+ if(s->codec_id == CODEC_ID_H261){
+ c->sub_motion_search= no_sub_motion_search;
}
- */
+
+ return 0;
}
-static inline int small_diamond_search4MV(MpegEncContext * s, int *best, int dmin,
- UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
- int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
- int xmin, int ymin, int xmax, int ymax, int shift)
+#if 0
+static int pix_dev(uint8_t * pix, int line_size, int mean)
{
- int next_dir=-1;
-
- for(;;){
- int d;
- const int dir= next_dir;
- const int x= best[0];
- const int y= best[1];
- next_dir=-1;
-
-//printf("%d", dir);
- if(dir!=2 && x>xmin) CHECK_MV4_DIR(x-1, y , 0)
- if(dir!=3 && y>ymin) CHECK_MV4_DIR(x , y-1, 1)
- if(dir!=0 && x<xmax) CHECK_MV4_DIR(x+1, y , 2)
- if(dir!=1 && y<ymax) CHECK_MV4_DIR(x , y+1, 3)
-
- if(next_dir==-1){
- return dmin;
+ int s, i, j;
+
+ s = 0;
+ for (i = 0; i < 16; i++) {
+ for (j = 0; j < 16; j += 8) {
+ s += FFABS(pix[0]-mean);
+ s += FFABS(pix[1]-mean);
+ s += FFABS(pix[2]-mean);
+ s += FFABS(pix[3]-mean);
+ s += FFABS(pix[4]-mean);
+ s += FFABS(pix[5]-mean);
+ s += FFABS(pix[6]-mean);
+ s += FFABS(pix[7]-mean);
+ pix += 8;
}
+ pix += line_size - 16;
}
+ return s;
}
+#endif
-static inline int snake_search(MpegEncContext * s, int *best, int dmin,
- UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
- int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
- int xmin, int ymin, int xmax, int ymax, int shift)
+static inline void no_motion_search(MpegEncContext * s,
+ int *mx_ptr, int *my_ptr)
{
- int dir=0;
- int c=1;
- static int x_dir[8]= {1,1,0,-1,-1,-1, 0, 1};
- static int y_dir[8]= {0,1,1, 1, 0,-1,-1,-1};
- int fails=0;
- int last_d[2]={dmin, dmin};
-
-/*static int good=0;
-static int bad=0;
-static int point=0;
-
-point++;
-if(256*256*256*64%point==0)
+ *mx_ptr = 16 * s->mb_x;
+ *my_ptr = 16 * s->mb_y;
+}
+
+#define Z_THRESHOLD 256
+
+#define CHECK_SAD_HALF_MV(suffix, x, y) \
+{\
+ d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
+ d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
+ COPY3_IF_LT(dminh, d, dx, x, dy, y)\
+}
+
+static inline int sad_hpel_motion_search(MpegEncContext * s,
+ int *mx_ptr, int *my_ptr, int dmin,
+ int src_index, int ref_index,
+ int size, int h)
{
- printf("%d %d %d\n", good, bad, point);
-}*/
-
- for(;;){
- int x= best[0];
- int y= best[1];
- int d;
- x+=x_dir[dir];
- y+=y_dir[dir];
- if(x>=xmin && x<=xmax && y>=ymin && y<=ymax){
- d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);
- d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;
- }else{
- d = dmin + 10000; //FIXME smarter boundary handling
- }
- if(d<dmin){
- best[0]=x;
- best[1]=y;
- dmin=d;
-
- if(last_d[1] - last_d[0] > last_d[0] - d) c= -c;
- dir+=c;
-
- fails=0;
-//good++;
- last_d[1]=last_d[0];
- last_d[0]=d;
+ MotionEstContext * const c= &s->me;
+ const int penalty_factor= c->sub_penalty_factor;
+ int mx, my, dminh;
+ uint8_t *pix, *ptr;
+ int stride= c->stride;
+ const int flags= c->sub_flags;
+ LOAD_COMMON
+
+ assert(flags == 0);
+
+ if(c->skip){
+// printf("S");
+ *mx_ptr = 0;
+ *my_ptr = 0;
+ return dmin;
+ }
+// printf("N");
+
+ pix = c->src[src_index][0];
+
+ mx = *mx_ptr;
+ my = *my_ptr;
+ ptr = c->ref[ref_index][0] + (my * stride) + mx;
+
+ dminh = dmin;
+
+ if (mx > xmin && mx < xmax &&
+ my > ymin && my < ymax) {
+ int dx=0, dy=0;
+ int d, pen_x, pen_y;
+ const int index= (my<<ME_MAP_SHIFT) + mx;
+ const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
+ const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
+ const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
+ const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
+ mx<<=1;
+ my<<=1;
+
+
+ pen_x= pred_x + mx;
+ pen_y= pred_y + my;
+
+ ptr-= stride;
+ if(t<=b){
+ CHECK_SAD_HALF_MV(y2 , 0, -1)
+ if(l<=r){
+ CHECK_SAD_HALF_MV(xy2, -1, -1)
+ if(t+r<=b+l){
+ CHECK_SAD_HALF_MV(xy2, +1, -1)
+ ptr+= stride;
+ }else{
+ ptr+= stride;
+ CHECK_SAD_HALF_MV(xy2, -1, +1)
+ }
+ CHECK_SAD_HALF_MV(x2 , -1, 0)
+ }else{
+ CHECK_SAD_HALF_MV(xy2, +1, -1)
+ if(t+l<=b+r){
+ CHECK_SAD_HALF_MV(xy2, -1, -1)
+ ptr+= stride;
+ }else{
+ ptr+= stride;
+ CHECK_SAD_HALF_MV(xy2, +1, +1)
+ }
+ CHECK_SAD_HALF_MV(x2 , +1, 0)
+ }
}else{
-//bad++;
- if(fails){
- if(fails>=3) return dmin;
+ if(l<=r){
+ if(t+l<=b+r){
+ CHECK_SAD_HALF_MV(xy2, -1, -1)
+ ptr+= stride;
+ }else{
+ ptr+= stride;
+ CHECK_SAD_HALF_MV(xy2, +1, +1)
+ }
+ CHECK_SAD_HALF_MV(x2 , -1, 0)
+ CHECK_SAD_HALF_MV(xy2, -1, +1)
}else{
- c= -c;
+ if(t+r<=b+l){
+ CHECK_SAD_HALF_MV(xy2, +1, -1)
+ ptr+= stride;
+ }else{
+ ptr+= stride;
+ CHECK_SAD_HALF_MV(xy2, -1, +1)
+ }
+ CHECK_SAD_HALF_MV(x2 , +1, 0)
+ CHECK_SAD_HALF_MV(xy2, +1, +1)
}
- dir+=c*2;
- fails++;
+ CHECK_SAD_HALF_MV(y2 , 0, +1)
}
- dir&=7;
+ mx+=dx;
+ my+=dy;
+
+ }else{
+ mx<<=1;
+ my<<=1;
}
+
+ *mx_ptr = mx;
+ *my_ptr = my;
+ return dminh;
}
-static int epzs_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr,
- int P[5][2], int pred_x, int pred_y,
- int xmin, int ymin, int xmax, int ymax, uint8_t * ref_picture)
+static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
{
- int best[2]={0, 0};
- int d, dmin;
- UINT8 *new_pic, *old_pic;
- const int pic_stride= s->linesize;
- const int pic_xy= (s->mb_y*pic_stride + s->mb_x)*16;
- UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
- int quant= s->qscale; // qscale of the prev frame
- const int shift= 1+s->quarter_sample;
+ const int xy= s->mb_x + s->mb_y*s->mb_stride;
- new_pic = s->new_picture[0] + pic_xy;
- old_pic = ref_picture + pic_xy;
-
- dmin = pix_abs16x16(new_pic, old_pic, pic_stride);
- if(dmin<Z_THRESHOLD){
- *mx_ptr= 0;
- *my_ptr= 0;
-//printf("Z");
- return dmin;
+ s->p_mv_table[xy][0] = mx;
+ s->p_mv_table[xy][1] = my;
+
+ /* has already been set to the 4 MV if 4MV is done */
+ if(mv4){
+ int mot_xy= s->block_index[0];
+
+ s->current_picture.motion_val[0][mot_xy ][0]= mx;
+ s->current_picture.motion_val[0][mot_xy ][1]= my;
+ s->current_picture.motion_val[0][mot_xy+1][0]= mx;
+ s->current_picture.motion_val[0][mot_xy+1][1]= my;
+
+ mot_xy += s->b8_stride;
+ s->current_picture.motion_val[0][mot_xy ][0]= mx;
+ s->current_picture.motion_val[0][mot_xy ][1]= my;
+ s->current_picture.motion_val[0][mot_xy+1][0]= mx;
+ s->current_picture.motion_val[0][mot_xy+1][1]= my;
}
+}
- /* first line */
- if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
- CHECK_MV(P[1][0]>>shift, P[1][1]>>shift)
- }else{
- CHECK_MV(P[4][0]>>shift, P[4][1]>>shift)
- if(dmin<Z_THRESHOLD){
- *mx_ptr= P[4][0]>>shift;
- *my_ptr= P[4][1]>>shift;
-//printf("M\n");
- return dmin;
- }
- CHECK_MV(P[1][0]>>shift, P[1][1]>>shift)
- CHECK_MV(P[2][0]>>shift, P[2][1]>>shift)
- CHECK_MV(P[3][0]>>shift, P[3][1]>>shift)
+/**
+ * get fullpel ME search limits.
+ */
+static inline void get_limits(MpegEncContext *s, int x, int y)
+{
+ MotionEstContext * const c= &s->me;
+ int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
+/*
+ if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
+ else c->range= 16;
+*/
+ if (s->unrestricted_mv) {
+ c->xmin = - x - 16;
+ c->ymin = - y - 16;
+ c->xmax = - x + s->mb_width *16;
+ c->ymax = - y + s->mb_height*16;
+ } else if (s->out_format == FMT_H261){
+ // Search range of H261 is different from other codec standards
+ c->xmin = (x > 15) ? - 15 : 0;
+ c->ymin = (y > 15) ? - 15 : 0;
+ c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
+ c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
+ } else {
+ c->xmin = - x;
+ c->ymin = - y;
+ c->xmax = - x + s->mb_width *16 - 16;
+ c->ymax = - y + s->mb_height*16 - 16;
+ }
+ if(range){
+ c->xmin = FFMAX(c->xmin,-range);
+ c->xmax = FFMIN(c->xmax, range);
+ c->ymin = FFMAX(c->ymin,-range);
+ c->ymax = FFMIN(c->ymax, range);
}
- CHECK_MV(P[0][0]>>shift, P[0][1]>>shift)
+}
-//check(best[0],best[1],0, b0)
- if(s->me_method==ME_EPZS)
- dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride,
- pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);
- else
- dmin= snake_search(s, best, dmin, new_pic, old_pic, pic_stride,
- pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);
-//check(best[0],best[1],0, b1)
- *mx_ptr= best[0];
- *my_ptr= best[1];
+static inline void init_mv4_ref(MotionEstContext *c){
+ const int stride= c->stride;
-// printf("%d %d %d \n", best[0], best[1], dmin);
- return dmin;
+ c->ref[1][0] = c->ref[0][0] + 8;
+ c->ref[2][0] = c->ref[0][0] + 8*stride;
+ c->ref[3][0] = c->ref[2][0] + 8;
+ c->src[1][0] = c->src[0][0] + 8;
+ c->src[2][0] = c->src[0][0] + 8*stride;
+ c->src[3][0] = c->src[2][0] + 8;
}
-static int epzs_motion_search4(MpegEncContext * s, int block,
- int *mx_ptr, int *my_ptr,
- int P[6][2], int pred_x, int pred_y,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
+static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
{
- int best[2]={0, 0};
- int d, dmin;
- UINT8 *new_pic, *old_pic;
- const int pic_stride= s->linesize;
- const int pic_xy= ((s->mb_y*2 + (block>>1))*pic_stride + s->mb_x*2 + (block&1))*8;
- UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
- int quant= s->qscale; // qscale of the prev frame
- const int shift= 1+s->quarter_sample;
+ MotionEstContext * const c= &s->me;
+ const int size= 1;
+ const int h=8;
+ int block;
+ int P[10][2];
+ int dmin_sum=0, mx4_sum=0, my4_sum=0;
+ int same=1;
+ const int stride= c->stride;
+ uint8_t *mv_penalty= c->current_mv_penalty;
+
+ init_mv4_ref(c);
+
+ for(block=0; block<4; block++){
+ int mx4, my4;
+ int pred_x4, pred_y4;
+ int dmin4;
+ static const int off[4]= {2, 1, 1, -1};
+ const int mot_stride = s->b8_stride;
+ const int mot_xy = s->block_index[block];
+
+ P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
+ P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
+
+ if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
+
+ /* special case for first line */
+ if (s->first_slice_line && block<2) {
+ c->pred_x= pred_x4= P_LEFT[0];
+ c->pred_y= pred_y4= P_LEFT[1];
+ } else {
+ P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
+ P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
+ P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
+ P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
+ if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
+ if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
+ if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
+ if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
+
+ c->pred_x= pred_x4 = P_MEDIAN[0];
+ c->pred_y= pred_y4 = P_MEDIAN[1];
+ }
+ P_MV1[0]= mx;
+ P_MV1[1]= my;
- new_pic = s->new_picture[0] + pic_xy;
- old_pic = ref_picture + pic_xy;
-
- dmin = pix_abs8x8(new_pic, old_pic, pic_stride);
+ dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
- /* first line */
- if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {
- CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)
- }else{
- CHECK_MV4(P[4][0]>>shift, P[4][1]>>shift)
- if(dmin<Z_THRESHOLD){
- *mx_ptr= P[4][0]>>shift;
- *my_ptr= P[4][1]>>shift;
-//printf("M\n");
- return dmin;
- }
- CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)
- CHECK_MV4(P[2][0]>>shift, P[2][1]>>shift)
- CHECK_MV4(P[3][0]>>shift, P[3][1]>>shift)
- }
- CHECK_MV4(P[0][0]>>shift, P[0][1]>>shift)
- CHECK_MV4(P[5][0]>>shift, P[5][1]>>shift)
+ dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
+
+ if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
+ int dxy;
+ const int offset= ((block&1) + (block>>1)*stride)*8;
+ uint8_t *dest_y = c->scratchpad + offset;
+ if(s->quarter_sample){
+ uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
+ dxy = ((my4 & 3) << 2) | (mx4 & 3);
+
+ if(s->no_rounding)
+ s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
+ else
+ s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
+ }else{
+ uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
+ dxy = ((my4 & 1) << 1) | (mx4 & 1);
-//check(best[0],best[1],0, b0)
- dmin= small_diamond_search4MV(s, best, dmin, new_pic, old_pic, pic_stride,
- pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);
-//check(best[0],best[1],0, b1)
- *mx_ptr= best[0];
- *my_ptr= best[1];
+ if(s->no_rounding)
+ s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
+ else
+ s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
+ }
+ dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
+ }else
+ dmin_sum+= dmin4;
-// printf("%d %d %d \n", best[0], best[1], dmin);
- return dmin;
-}
+ if(s->quarter_sample){
+ mx4_sum+= mx4/2;
+ my4_sum+= my4/2;
+ }else{
+ mx4_sum+= mx4;
+ my4_sum+= my4;
+ }
-#define CHECK_HALF_MV(suffix, x, y) \
- d= pix_abs16x16_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
- d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*quant;\
- if(d<dminh){\
- dminh= d;\
- mx= mx1 + x;\
- my= my1 + y;\
+ s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
+ s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
+
+ if(mx4 != mx || my4 != my) same=0;
}
-#define CHECK_HALF_MV4(suffix, x, y) \
- d= pix_abs8x8_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
- d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*quant;\
- if(d<dminh){\
- dminh= d;\
- mx= mx1 + x;\
- my= my1 + y;\
+ if(same)
+ return INT_MAX;
+
+ if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
+ dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
}
-
-/* The idea would be to make half pel ME after Inter/Intra decision to
- save time. */
-static inline int halfpel_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int dmin,
- int xmin, int ymin, int xmax, int ymax,
- int pred_x, int pred_y, uint8_t *ref_picture)
-{
- UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
- const int quant= s->qscale;
- int pen_x, pen_y;
- int mx, my, mx1, my1, d, xx, yy, dminh;
- UINT8 *pix, *ptr;
- mx = *mx_ptr;
- my = *my_ptr;
- ptr = ref_picture + (my * s->linesize) + mx;
+ if(c->avctx->mb_cmp&FF_CMP_CHROMA){
+ int dxy;
+ int mx, my;
+ int offset;
- xx = 16 * s->mb_x;
- yy = 16 * s->mb_y;
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
-
- dminh = dmin;
+ mx= ff_h263_round_chroma(mx4_sum);
+ my= ff_h263_round_chroma(my4_sum);
+ dxy = ((my & 1) << 1) | (mx & 1);
- if (mx > xmin && mx < xmax &&
- my > ymin && my < ymax) {
+ offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
- mx= mx1= 2*(mx - xx);
- my= my1= 2*(my - yy);
- if(dmin < Z_THRESHOLD && mx==0 && my==0){
- *mx_ptr = 0;
- *my_ptr = 0;
- return dmin;
+ if(s->no_rounding){
+ s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
+ s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
+ }else{
+ s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
+ s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
}
-
- pen_x= pred_x + mx;
- pen_y= pred_y + my;
- ptr-= s->linesize;
- CHECK_HALF_MV(xy2, -1, -1)
- CHECK_HALF_MV(y2 , 0, -1)
- CHECK_HALF_MV(xy2, +1, -1)
-
- ptr+= s->linesize;
- CHECK_HALF_MV(x2 , -1, 0)
- CHECK_HALF_MV(x2 , +1, 0)
- CHECK_HALF_MV(xy2, -1, +1)
- CHECK_HALF_MV(y2 , 0, +1)
- CHECK_HALF_MV(xy2, +1, +1)
+ dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
+ dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
+ }
- }else{
- mx= 2*(mx - xx);
- my= 2*(my - yy);
+ c->pred_x= mx;
+ c->pred_y= my;
+
+ switch(c->avctx->mb_cmp&0xFF){
+ /*case FF_CMP_SSE:
+ return dmin_sum+ 32*s->qscale*s->qscale;*/
+ case FF_CMP_RD:
+ return dmin_sum;
+ default:
+ return dmin_sum+ 11*c->mb_penalty_factor;
}
+}
- *mx_ptr = mx;
- *my_ptr = my;
- return dminh;
+static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
+ MotionEstContext * const c= &s->me;
+
+ c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
+ c->src[1][0] = c->src[0][0] + s->linesize;
+ if(c->flags & FLAG_CHROMA){
+ c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
+ c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
+ c->src[1][1] = c->src[0][1] + s->uvlinesize;
+ c->src[1][2] = c->src[0][2] + s->uvlinesize;
+ }
}
-static inline void halfpel_motion_search4(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int dmin,
- int xmin, int ymin, int xmax, int ymax,
- int pred_x, int pred_y, int block_x, int block_y,
- uint8_t *ref_picture)
+static int interlaced_search(MpegEncContext *s, int ref_index,
+ int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
{
- UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
- const int quant= s->qscale;
- int pen_x, pen_y;
- int mx, my, mx1, my1, d, xx, yy, dminh;
- UINT8 *pix, *ptr;
-
- xx = 8 * block_x;
- yy = 8 * block_y;
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
-
- mx = *mx_ptr;
- my = *my_ptr;
- ptr = ref_picture + ((yy+my) * s->linesize) + xx + mx;
+ MotionEstContext * const c= &s->me;
+ const int size=0;
+ const int h=8;
+ int block;
+ int P[10][2];
+ uint8_t * const mv_penalty= c->current_mv_penalty;
+ int same=1;
+ const int stride= 2*s->linesize;
+ int dmin_sum= 0;
+ const int mot_stride= s->mb_stride;
+ const int xy= s->mb_x + s->mb_y*mot_stride;
+
+ c->ymin>>=1;
+ c->ymax>>=1;
+ c->stride<<=1;
+ c->uvstride<<=1;
+ init_interlaced_ref(s, ref_index);
+
+ for(block=0; block<2; block++){
+ int field_select;
+ int best_dmin= INT_MAX;
+ int best_field= -1;
+
+ for(field_select=0; field_select<2; field_select++){
+ int dmin, mx_i, my_i;
+ int16_t (*mv_table)[2]= mv_tables[block][field_select];
+
+ if(user_field_select){
+ assert(field_select==0 || field_select==1);
+ assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
+ if(field_select_tables[block][xy] != field_select)
+ continue;
+ }
- dminh = dmin;
+ P_LEFT[0] = mv_table[xy - 1][0];
+ P_LEFT[1] = mv_table[xy - 1][1];
+ if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
+
+ c->pred_x= P_LEFT[0];
+ c->pred_y= P_LEFT[1];
+
+ if(!s->first_slice_line){
+ P_TOP[0] = mv_table[xy - mot_stride][0];
+ P_TOP[1] = mv_table[xy - mot_stride][1];
+ P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
+ P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
+ if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
+ if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
+ if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
+ if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
+ }
+ P_MV1[0]= mx; //FIXME not correct if block != field_select
+ P_MV1[1]= my / 2;
- if (mx > xmin && mx < xmax &&
- my > ymin && my < ymax) {
+ dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
- mx= mx1= 2*mx;
- my= my1= 2*my;
- if(dmin < Z_THRESHOLD && mx==0 && my==0){
- *mx_ptr = 0;
- *my_ptr = 0;
- return;
+ dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
+
+ mv_table[xy][0]= mx_i;
+ mv_table[xy][1]= my_i;
+
+ if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
+ int dxy;
+
+ //FIXME chroma ME
+ uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
+ dxy = ((my_i & 1) << 1) | (mx_i & 1);
+
+ if(s->no_rounding){
+ s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
+ }else{
+ s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
+ }
+ dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
+ dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
+ }else
+ dmin+= c->mb_penalty_factor; //field_select bits
+
+ dmin += field_select != block; //slightly prefer same field
+
+ if(dmin < best_dmin){
+ best_dmin= dmin;
+ best_field= field_select;
+ }
}
-
- pen_x= pred_x + mx;
- pen_y= pred_y + my;
+ {
+ int16_t (*mv_table)[2]= mv_tables[block][best_field];
- ptr-= s->linesize;
- CHECK_HALF_MV4(xy2, -1, -1)
- CHECK_HALF_MV4(y2 , 0, -1)
- CHECK_HALF_MV4(xy2, +1, -1)
-
- ptr+= s->linesize;
- CHECK_HALF_MV4(x2 , -1, 0)
- CHECK_HALF_MV4(x2 , +1, 0)
- CHECK_HALF_MV4(xy2, -1, +1)
- CHECK_HALF_MV4(y2 , 0, +1)
- CHECK_HALF_MV4(xy2, +1, +1)
+ if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
+ if(mv_table[xy][1]&1) same=0;
+ if(mv_table[xy][1]*2 != my) same=0;
+ if(best_field != block) same=0;
+ }
- }else{
- mx*=2;
- my*=2;
+ field_select_tables[block][xy]= best_field;
+ dmin_sum += best_dmin;
}
- *mx_ptr = mx;
- *my_ptr = my;
+ c->ymin<<=1;
+ c->ymax<<=1;
+ c->stride>>=1;
+ c->uvstride>>=1;
+
+ if(same)
+ return INT_MAX;
+
+ switch(c->avctx->mb_cmp&0xFF){
+ /*case FF_CMP_SSE:
+ return dmin_sum+ 32*s->qscale*s->qscale;*/
+ case FF_CMP_RD:
+ return dmin_sum;
+ default:
+ return dmin_sum+ 11*c->mb_penalty_factor;
+ }
}
-static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my)
-{
- const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
-
- s->p_mv_table[xy][0] = mx;
- s->p_mv_table[xy][1] = my;
+static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
+ int ymax= s->me.ymax>>interlaced;
+ int ymin= s->me.ymin>>interlaced;
- /* has allready been set to the 4 MV if 4MV is done */
- if(!(s->flags&CODEC_FLAG_4MV)){
- int mot_xy= s->block_index[0];
+ if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
+ if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
+ if(mv[1] < ymin) mv[1] = ymin;
+ if(mv[1] > ymax) mv[1] = ymax;
+}
- s->motion_val[mot_xy ][0]= mx;
- s->motion_val[mot_xy ][1]= my;
- s->motion_val[mot_xy+1][0]= mx;
- s->motion_val[mot_xy+1][1]= my;
+static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
+ MotionEstContext * const c= &s->me;
+ Picture *p= s->current_picture_ptr;
+ int mb_xy= mb_x + mb_y*s->mb_stride;
+ int xy= 2*mb_x + 2*mb_y*s->b8_stride;
+ int mb_type= s->current_picture.mb_type[mb_xy];
+ int flags= c->flags;
+ int shift= (flags&FLAG_QPEL) + 1;
+ int mask= (1<<shift)-1;
+ int x, y, i;
+ int d=0;
+ me_cmp_func cmpf= s->dsp.sse[0];
+ me_cmp_func chroma_cmpf= s->dsp.sse[1];
+
+ if(p_type && USES_LIST(mb_type, 1)){
+ av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
+ return INT_MAX/2;
+ }
+ assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
- mot_xy += s->block_wrap[0];
- s->motion_val[mot_xy ][0]= mx;
- s->motion_val[mot_xy ][1]= my;
- s->motion_val[mot_xy+1][0]= mx;
- s->motion_val[mot_xy+1][1]= my;
+ for(i=0; i<4; i++){
+ int xy= s->block_index[i];
+ clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
+ clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
}
-}
-static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
-{
- *range = 8 * (1 << (f_code - 1));
- /* XXX: temporary kludge to avoid overflow for msmpeg4 */
- if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
- *range *= 2;
+ if(IS_INTERLACED(mb_type)){
+ int xy2= xy + s->b8_stride;
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
+ c->stride<<=1;
+ c->uvstride<<=1;
- if (s->unrestricted_mv) {
- *xmin = -16;
- *ymin = -16;
- if (s->h263_plus)
- *range *= 2;
- if(s->avctx==NULL || s->avctx->codec->id!=CODEC_ID_MPEG4){
- *xmax = s->mb_width*16;
- *ymax = s->mb_height*16;
- }else {
- /* XXX: dunno if this is correct but ffmpeg4 decoder wont like it otherwise
- (cuz the drawn edge isnt large enough))*/
- *xmax = s->width;
- *ymax = s->height;
+ if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
+ av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
+ return INT_MAX/2;
}
- } else {
- *xmin = 0;
- *ymin = 0;
- *xmax = s->mb_width*16 - 16;
- *ymax = s->mb_height*16 - 16;
+
+ if(USES_LIST(mb_type, 0)){
+ int field_select0= p->ref_index[0][xy ];
+ int field_select1= p->ref_index[0][xy2];
+ assert(field_select0==0 ||field_select0==1);
+ assert(field_select1==0 ||field_select1==1);
+ init_interlaced_ref(s, 0);
+
+ if(p_type){
+ s->p_field_select_table[0][mb_xy]= field_select0;
+ s->p_field_select_table[1][mb_xy]= field_select1;
+ *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
+ *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
+ }else{
+ s->b_field_select_table[0][0][mb_xy]= field_select0;
+ s->b_field_select_table[0][1][mb_xy]= field_select1;
+ *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
+ *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
+ s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
+ }
+
+ x= p->motion_val[0][xy ][0];
+ y= p->motion_val[0][xy ][1];
+ d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
+ x= p->motion_val[0][xy2][0];
+ y= p->motion_val[0][xy2][1];
+ d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
+ }
+ if(USES_LIST(mb_type, 1)){
+ int field_select0= p->ref_index[1][xy ];
+ int field_select1= p->ref_index[1][xy2];
+ assert(field_select0==0 ||field_select0==1);
+ assert(field_select1==0 ||field_select1==1);
+ init_interlaced_ref(s, 2);
+
+ s->b_field_select_table[1][0][mb_xy]= field_select0;
+ s->b_field_select_table[1][1][mb_xy]= field_select1;
+ *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
+ *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
+ if(USES_LIST(mb_type, 0)){
+ s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
+ }else{
+ s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
+ }
+
+ x= p->motion_val[1][xy ][0];
+ y= p->motion_val[1][xy ][1];
+ d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
+ x= p->motion_val[1][xy2][0];
+ y= p->motion_val[1][xy2][1];
+ d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
+ //FIXME bidir scores
+ }
+ c->stride>>=1;
+ c->uvstride>>=1;
+ }else if(IS_8X8(mb_type)){
+ if(!(s->flags & CODEC_FLAG_4MV)){
+ av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
+ return INT_MAX/2;
+ }
+ cmpf= s->dsp.sse[1];
+ chroma_cmpf= s->dsp.sse[1];
+ init_mv4_ref(c);
+ for(i=0; i<4; i++){
+ xy= s->block_index[i];
+ x= p->motion_val[0][xy][0];
+ y= p->motion_val[0][xy][1];
+ d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
+ }
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
+ }else{
+ if(USES_LIST(mb_type, 0)){
+ if(p_type){
+ *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
+ }else if(USES_LIST(mb_type, 1)){
+ *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+ *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
+ }else{
+ *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
+ }
+ x= p->motion_val[0][xy][0];
+ y= p->motion_val[0][xy][1];
+ d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
+ }else if(USES_LIST(mb_type, 1)){
+ *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
+
+ x= p->motion_val[1][xy][0];
+ y= p->motion_val[1][xy][1];
+ d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
+ }else
+ s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
}
+ return d;
}
void ff_estimate_p_frame_motion(MpegEncContext * s,
int mb_x, int mb_y)
{
- UINT8 *pix, *ppix;
- int sum, varc, vard, mx, my, range, dmin, xx, yy;
- int xmin, ymin, xmax, ymax;
- int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
- int pred_x=0, pred_y=0;
- int P[6][2];
+ MotionEstContext * const c= &s->me;
+ uint8_t *pix, *ppix;
+ int sum, mx, my, dmin;
+ int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
+ int vard; ///< sum of squared differences with the estimated motion vector
+ int P[10][2];
const int shift= 1+s->quarter_sample;
int mb_type=0;
- uint8_t *ref_picture= s->last_picture[0];
+ Picture * const pic= &s->current_picture;
+
+ init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
+
+ assert(s->quarter_sample==0 || s->quarter_sample==1);
+ assert(s->linesize == c->stride);
+ assert(s->uvlinesize == c->uvstride);
+
+ c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
+ c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
+ c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
+ c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
- get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
+ get_limits(s, 16*mb_x, 16*mb_y);
+ c->skip=0;
+
+ /* intra / predictive decision */
+ pix = c->src[0][0];
+ sum = s->dsp.pix_sum(pix, s->linesize);
+ varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
+
+ pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
+ pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
+ c->mb_var_sum_temp += (varc+128)>>8;
+
+ if(c->avctx->me_threshold){
+ vard= check_input_motion(s, mb_x, mb_y, 1);
+
+ if((vard+128)>>8 < c->avctx->me_threshold){
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
+ c->mc_mb_var_sum_temp += (vard+128)>>8;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
+ return;
+ }
+ if((vard+128)>>8 < c->avctx->mb_threshold)
+ mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
+ }
switch(s->me_method) {
case ME_ZERO:
default:
- no_motion_search(s, &mx, &my);
+ no_motion_search(s, &mx, &my);
+ mx-= mb_x*16;
+ my-= mb_y*16;
dmin = 0;
break;
- case ME_FULL:
- dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
- break;
- case ME_LOG:
- dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
- break;
- case ME_PHODS:
- dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
- break;
case ME_X1:
case ME_EPZS:
{
- const int mot_stride = s->block_wrap[0];
+ const int mot_stride = s->b8_stride;
const int mot_xy = s->block_index[0];
- rel_xmin= xmin - mb_x*16;
- rel_xmax= xmax - mb_x*16;
- rel_ymin= ymin - mb_y*16;
- rel_ymax= ymax - mb_y*16;
-
- P[0][0] = s->motion_val[mot_xy ][0];
- P[0][1] = s->motion_val[mot_xy ][1];
- P[1][0] = s->motion_val[mot_xy - 1][0];
- P[1][1] = s->motion_val[mot_xy - 1][1];
- if(P[1][0] > (rel_xmax<<shift)) P[1][0]= (rel_xmax<<shift);
-
- /* special case for first line */
- if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
- P[4][0] = P[1][0];
- P[4][1] = P[1][1];
- } else {
- P[2][0] = s->motion_val[mot_xy - mot_stride ][0];
- P[2][1] = s->motion_val[mot_xy - mot_stride ][1];
- P[3][0] = s->motion_val[mot_xy - mot_stride + 2 ][0];
- P[3][1] = s->motion_val[mot_xy - mot_stride + 2 ][1];
- if(P[2][1] > (rel_ymax<<shift)) P[2][1]= (rel_ymax<<shift);
- if(P[3][0] < (rel_xmin<<shift)) P[3][0]= (rel_xmin<<shift);
- if(P[3][1] > (rel_ymax<<shift)) P[3][1]= (rel_ymax<<shift);
-
- P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
- P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
- }
- if(s->out_format == FMT_H263){
- pred_x = P[4][0];
- pred_y = P[4][1];
- }else { /* mpeg1 at least */
- pred_x= P[1][0];
- pred_y= P[1][1];
+ P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
+ P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
+
+ if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
+
+ if(!s->first_slice_line) {
+ P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
+ P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
+ P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
+ P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
+ if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
+ if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
+ if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
+
+ if(s->out_format == FMT_H263){
+ c->pred_x = P_MEDIAN[0];
+ c->pred_y = P_MEDIAN[1];
+ }else { /* mpeg1 at least */
+ c->pred_x= P_LEFT[0];
+ c->pred_y= P_LEFT[1];
+ }
+ }else{
+ c->pred_x= P_LEFT[0];
+ c->pred_y= P_LEFT[1];
}
+
}
- dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);
-
- mx+= mb_x*16;
- my+= mb_y*16;
+ dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
+
break;
}
-
- if(s->flags&CODEC_FLAG_4MV){
- int block;
-
- mb_type|= MB_TYPE_INTER4V;
-
- for(block=0; block<4; block++){
- int mx4, my4;
- int pred_x4, pred_y4;
- int dmin4;
- static const int off[4]= {2, 1, 1, -1};
- const int mot_stride = s->block_wrap[0];
- const int mot_xy = s->block_index[block];
- const int block_x= mb_x*2 + (block&1);
- const int block_y= mb_y*2 + (block>>1);
-
- const int rel_xmin4= xmin - block_x*8;
- const int rel_xmax4= xmax - block_x*8 + 8;
- const int rel_ymin4= ymin - block_y*8;
- const int rel_ymax4= ymax - block_y*8 + 8;
-
- P[0][0] = s->motion_val[mot_xy ][0];
- P[0][1] = s->motion_val[mot_xy ][1];
- P[1][0] = s->motion_val[mot_xy - 1][0];
- P[1][1] = s->motion_val[mot_xy - 1][1];
- if(P[1][0] > (rel_xmax4<<shift)) P[1][0]= (rel_xmax4<<shift);
- /* special case for first line */
- if ((mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {
- P[4][0] = P[1][0];
- P[4][1] = P[1][1];
- } else {
- P[2][0] = s->motion_val[mot_xy - mot_stride ][0];
- P[2][1] = s->motion_val[mot_xy - mot_stride ][1];
- P[3][0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
- P[3][1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
- if(P[2][1] > (rel_ymax4<<shift)) P[2][1]= (rel_ymax4<<shift);
- if(P[3][0] < (rel_xmin4<<shift)) P[3][0]= (rel_xmin4<<shift);
- if(P[3][0] > (rel_xmax4<<shift)) P[3][0]= (rel_xmax4<<shift);
- if(P[3][1] > (rel_ymax4<<shift)) P[3][1]= (rel_ymax4<<shift);
-
- P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
- P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
- }
- if(s->out_format == FMT_H263){
- pred_x4 = P[4][0];
- pred_y4 = P[4][1];
- }else { /* mpeg1 at least */
- pred_x4= P[1][0];
- pred_y4= P[1][1];
- }
- P[5][0]= mx - mb_x*16;
- P[5][1]= my - mb_y*16;
+ /* At this point (mx,my) are full-pell and the relative displacement */
+ ppix = c->ref[0][0] + (my * s->linesize) + mx;
- dmin4 = epzs_motion_search4(s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, ref_picture);
+ vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
- halfpel_motion_search4(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
- pred_x4, pred_y4, block_x, block_y, ref_picture);
-
- s->motion_val[ s->block_index[block] ][0]= mx4;
- s->motion_val[ s->block_index[block] ][1]= my4;
- }
- }
+ pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
+// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
+ c->mc_mb_var_sum_temp += (vard+128)>>8;
- /* intra / predictive decision */
- xx = mb_x * 16;
- yy = mb_y * 16;
-
- pix = s->new_picture[0] + (yy * s->linesize) + xx;
- /* At this point (mx,my) are full-pell and the absolute displacement */
- ppix = ref_picture + (my * s->linesize) + mx;
-
- sum = pix_sum(pix, s->linesize);
-#if 0
- varc = pix_dev(pix, s->linesize, (sum+128)>>8) + INTER_BIAS;
- vard = pix_abs16x16(pix, ppix, s->linesize);
-#else
- sum= (sum+8)>>4;
- varc = ((pix_norm1(pix, s->linesize) - sum*sum + 128 + 500)>>8);
- vard = (pix_norm(pix, ppix, s->linesize)+128)>>8;
-#endif
-
- s->mb_var[s->mb_width * mb_y + mb_x] = varc;
- s->avg_mb_var+= varc;
- s->mc_mb_var += vard;
-
-
#if 0
printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
- varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
+ varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
- if(s->flags&CODEC_FLAG_HQ){
- if (vard*2 + 200 > varc)
- mb_type|= MB_TYPE_INTRA;
- if (varc*2 + 200 > vard){
- mb_type|= MB_TYPE_INTER;
- halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);
+ if(mb_type){
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
+
+ if(mb_type == CANDIDATE_MB_TYPE_INTER){
+ c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
+ set_p_mv_tables(s, mx, my, 1);
}else{
- mx = mx*2 - mb_x*32;
- my = my*2 - mb_y*32;
+ mx <<=shift;
+ my <<=shift;
+ }
+ if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
+ h263_mv4_search(s, mx, my, shift);
+
+ set_p_mv_tables(s, mx, my, 0);
+ }
+ if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
+ interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
+ }
+ }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
+
+ if (vard*2 + 200*256 > varc)
+ mb_type|= CANDIDATE_MB_TYPE_INTRA;
+ if (varc*2 + 200*256 > vard || s->qscale > 24){
+// if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
+ mb_type|= CANDIDATE_MB_TYPE_INTER;
+ c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
+ if(s->flags&CODEC_FLAG_MV0)
+ if(mx || my)
+ mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
+ }else{
+ mx <<=shift;
+ my <<=shift;
+ }
+ if((s->flags&CODEC_FLAG_4MV)
+ && !c->skip && varc>50<<8 && vard>10<<8){
+ if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
+ mb_type|=CANDIDATE_MB_TYPE_INTER4V;
+
+ set_p_mv_tables(s, mx, my, 0);
+ }else
+ set_p_mv_tables(s, mx, my, 1);
+ if((s->flags&CODEC_FLAG_INTERLACED_ME)
+ && !c->skip){ //FIXME varc/d checks
+ if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
+ mb_type |= CANDIDATE_MB_TYPE_INTER_I;
}
}else{
- if (vard <= 64 || vard < varc) {
- mb_type|= MB_TYPE_INTER;
- if (s->me_method != ME_ZERO) {
- halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);
- } else {
- mx -= 16 * mb_x;
- my -= 16 * mb_y;
+ int intra_score, i;
+ mb_type= CANDIDATE_MB_TYPE_INTER;
+
+ dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
+ if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
+ dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
+
+ if((s->flags&CODEC_FLAG_4MV)
+ && !c->skip && varc>50<<8 && vard>10<<8){
+ int dmin4= h263_mv4_search(s, mx, my, shift);
+ if(dmin4 < dmin){
+ mb_type= CANDIDATE_MB_TYPE_INTER4V;
+ dmin=dmin4;
}
-#if 0
- if (vard < 10) {
- skip++;
- fprintf(stderr,"\nEarly skip: %d vard: %2d varc: %5d dmin: %d",
- skip, vard, varc, dmin);
+ }
+ if((s->flags&CODEC_FLAG_INTERLACED_ME)
+ && !c->skip){ //FIXME varc/d checks
+ int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
+ if(dmin_i < dmin){
+ mb_type = CANDIDATE_MB_TYPE_INTER_I;
+ dmin= dmin_i;
}
-#endif
+ }
+
+// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
+ set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
+
+ /* get intra luma score */
+ if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
+ intra_score= varc - 500;
}else{
- mb_type|= MB_TYPE_INTRA;
- mx = 0;//mx*2 - 32 * mb_x;
- my = 0;//my*2 - 32 * mb_y;
+ int mean= (sum+128)>>8;
+ mean*= 0x01010101;
+
+ for(i=0; i<16; i++){
+ *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
+ *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
+ *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
+ *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
+ }
+
+ intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
}
+#if 0 //FIXME
+ /* get chroma score */
+ if(c->avctx->mb_cmp&FF_CMP_CHROMA){
+ for(i=1; i<3; i++){
+ uint8_t *dest_c;
+ int mean;
+
+ if(s->out_format == FMT_H263){
+ mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
+ }else{
+ mean= (s->last_dc[i] + 4)>>3;
+ }
+ dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
+
+ mean*= 0x01010101;
+ for(i=0; i<8; i++){
+ *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
+ *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
+ }
+
+ intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
+ }
+ }
+#endif
+ intra_score += c->mb_penalty_factor*16;
+
+ if(intra_score < dmin){
+ mb_type= CANDIDATE_MB_TYPE_INTRA;
+ s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
+ }else
+ s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
+
+ {
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
+ }
+ }
+
+ s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
+}
+
+int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
+ int mb_x, int mb_y)
+{
+ MotionEstContext * const c= &s->me;
+ int mx, my, dmin;
+ int P[10][2];
+ const int shift= 1+s->quarter_sample;
+ const int xy= mb_x + mb_y*s->mb_stride;
+ init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
+
+ assert(s->quarter_sample==0 || s->quarter_sample==1);
+
+ c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
+ c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
+
+ get_limits(s, 16*mb_x, 16*mb_y);
+ c->skip=0;
+
+ P_LEFT[0] = s->p_mv_table[xy + 1][0];
+ P_LEFT[1] = s->p_mv_table[xy + 1][1];
+
+ if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
+
+ /* special case for first line */
+ if (s->first_slice_line) {
+ c->pred_x= P_LEFT[0];
+ c->pred_y= P_LEFT[1];
+ P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
+ P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
+ } else {
+ P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
+ P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
+ P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
+ P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
+ if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
+ if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
+ if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
+
+ c->pred_x = P_MEDIAN[0];
+ c->pred_y = P_MEDIAN[1];
}
- s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
- set_p_mv_tables(s, mx, my);
+ dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
+
+ s->p_mv_table[xy][0] = mx<<shift;
+ s->p_mv_table[xy][1] = my<<shift;
+
+ return dmin;
}
-int ff_estimate_motion_b(MpegEncContext * s,
- int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *ref_picture, int f_code)
+static int ff_estimate_motion_b(MpegEncContext * s,
+ int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
{
- int mx, my, range, dmin;
- int xmin, ymin, xmax, ymax;
- int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
- int pred_x=0, pred_y=0;
- int P[6][2];
+ MotionEstContext * const c= &s->me;
+ int mx, my, dmin;
+ int P[10][2];
const int shift= 1+s->quarter_sample;
- const int mot_stride = s->mb_width + 2;
- const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
-
- get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
+ const int mot_stride = s->mb_stride;
+ const int mot_xy = mb_y*mot_stride + mb_x;
+ uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
+ int mv_scale;
+
+ c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
+ c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
+ c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
+ c->current_mv_penalty= mv_penalty;
+
+ get_limits(s, 16*mb_x, 16*mb_y);
switch(s->me_method) {
case ME_ZERO:
default:
- no_motion_search(s, &mx, &my);
+ no_motion_search(s, &mx, &my);
dmin = 0;
- break;
- case ME_FULL:
- dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
- break;
- case ME_LOG:
- dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
- break;
- case ME_PHODS:
- dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
+ mx-= mb_x*16;
+ my-= mb_y*16;
break;
case ME_X1:
case ME_EPZS:
{
+ P_LEFT[0] = mv_table[mot_xy - 1][0];
+ P_LEFT[1] = mv_table[mot_xy - 1][1];
- rel_xmin= xmin - mb_x*16;
- rel_xmax= xmax - mb_x*16;
- rel_ymin= ymin - mb_y*16;
- rel_ymax= ymax - mb_y*16;
-
- P[0][0] = mv_table[mot_xy ][0];
- P[0][1] = mv_table[mot_xy ][1];
- P[1][0] = mv_table[mot_xy - 1][0];
- P[1][1] = mv_table[mot_xy - 1][1];
- if(P[1][0] > (rel_xmax<<shift)) P[1][0]= (rel_xmax<<shift);
+ if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
/* special case for first line */
- if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
- P[4][0] = P[1][0];
- P[4][1] = P[1][1];
- } else {
- P[2][0] = mv_table[mot_xy - mot_stride ][0];
- P[2][1] = mv_table[mot_xy - mot_stride ][1];
- P[3][0] = mv_table[mot_xy - mot_stride + 1 ][0];
- P[3][1] = mv_table[mot_xy - mot_stride + 1 ][1];
- if(P[2][1] > (rel_ymax<<shift)) P[2][1]= (rel_ymax<<shift);
- if(P[3][0] < (rel_xmin<<shift)) P[3][0]= (rel_xmin<<shift);
- if(P[3][1] > (rel_ymax<<shift)) P[3][1]= (rel_ymax<<shift);
-
- P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
- P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
+ if (!s->first_slice_line) {
+ P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
+ P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
+ P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
+ P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
+ if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
+ if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
+ if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
}
- pred_x= P[1][0];
- pred_y= P[1][1];
+ c->pred_x= P_LEFT[0];
+ c->pred_y= P_LEFT[1];
}
- dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);
-
- mx+= mb_x*16;
- my+= mb_y*16;
+
+ if(mv_table == s->b_forw_mv_table){
+ mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
+ }else{
+ mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
+ }
+
+ dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
+
break;
}
-
- /* intra / predictive decision */
-// xx = mb_x * 16;
-// yy = mb_y * 16;
-// pix = s->new_picture[0] + (yy * s->linesize) + xx;
- /* At this point (mx,my) are full-pell and the absolute displacement */
-// ppix = ref_picture + (my * s->linesize) + mx;
-
- dmin= halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);
+ dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
+ if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
+ dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
+
+//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
// s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
mv_table[mot_xy][0]= mx;
mv_table[mot_xy][1]= my;
+
return dmin;
}
-
static inline int check_bidir_mv(MpegEncContext * s,
- int mb_x, int mb_y,
int motion_fx, int motion_fy,
int motion_bx, int motion_by,
int pred_fx, int pred_fy,
- int pred_bx, int pred_by)
+ int pred_bx, int pred_by,
+ int size, int h)
{
//FIXME optimize?
- UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
- uint8_t *dest_y = s->me_scratchpad;
+ //FIXME better f_code prediction (max mv & distance)
+ //FIXME pointers
+ MotionEstContext * const c= &s->me;
+ uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
+ uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
+ int stride= c->stride;
+ uint8_t *dest_y = c->scratchpad;
uint8_t *ptr;
int dxy;
int src_x, src_y;
int fbmin;
+ uint8_t **src_data= c->src[0];
+ uint8_t **ref_data= c->ref[0];
+ uint8_t **ref2_data= c->ref[2];
+
+ if(s->quarter_sample){
+ dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
+ src_x = motion_fx >> 2;
+ src_y = motion_fy >> 2;
+
+ ptr = ref_data[0] + (src_y * stride) + src_x;
+ s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
+
+ dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
+ src_x = motion_bx >> 2;
+ src_y = motion_by >> 2;
+
+ ptr = ref2_data[0] + (src_y * stride) + src_x;
+ s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
+ }else{
+ dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
+ src_x = motion_fx >> 1;
+ src_y = motion_fy >> 1;
+
+ ptr = ref_data[0] + (src_y * stride) + src_x;
+ s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
+
+ dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
+ src_x = motion_bx >> 1;
+ src_y = motion_by >> 1;
+
+ ptr = ref2_data[0] + (src_y * stride) + src_x;
+ s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
+ }
+
+ fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
+ +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
+ + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
+
+ if(c->avctx->mb_cmp&FF_CMP_CHROMA){
+ }
+ //FIXME CHROMA !!!
- fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->qscale;
-
- dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
- src_x = mb_x * 16 + (motion_fx >> 1);
- src_y = mb_y * 16 + (motion_fy >> 1);
-
- ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
- put_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
- put_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
-
- fbmin += (mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->qscale;
-
- dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
- src_x = mb_x * 16 + (motion_bx >> 1);
- src_y = mb_y * 16 + (motion_by >> 1);
-
- ptr = s->next_picture[0] + (src_y * s->linesize) + src_x;
- avg_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
- avg_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
-
- fbmin += pix_abs16x16(s->new_picture[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
return fbmin;
}
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
-static inline int bidir_refine(MpegEncContext * s,
- int mb_x, int mb_y)
+static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
{
- const int mot_stride = s->mb_width + 2;
- const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ MotionEstContext * const c= &s->me;
+ const int mot_stride = s->mb_stride;
+ const int xy = mb_y *mot_stride + mb_x;
int fbmin;
int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
-
- //FIXME do refinement and add flag
-
- fbmin= check_bidir_mv(s, mb_x, mb_y,
- motion_fx, motion_fy,
+ const int flags= c->sub_flags;
+ const int qpel= flags&FLAG_QPEL;
+ const int shift= 1+qpel;
+ const int xmin= c->xmin<<shift;
+ const int ymin= c->ymin<<shift;
+ const int xmax= c->xmax<<shift;
+ const int ymax= c->ymax<<shift;
+ uint8_t map[8][8][8][8];
+
+ memset(map,0,sizeof(map));
+#define BIDIR_MAP(fx,fy,bx,by) \
+ map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
+ BIDIR_MAP(0,0,0,0) = 1;
+
+ fbmin= check_bidir_mv(s, motion_fx, motion_fy,
motion_bx, motion_by,
pred_fx, pred_fy,
- pred_bx, pred_by);
+ pred_bx, pred_by,
+ 0, 16);
+
+ if(s->avctx->bidir_refine){
+ int score, end;
+#define CHECK_BIDIR(fx,fy,bx,by)\
+ if( !BIDIR_MAP(fx,fy,bx,by)\
+ &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
+ &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
+ BIDIR_MAP(fx,fy,bx,by) = 1;\
+ score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
+ if(score < fbmin){\
+ fbmin= score;\
+ motion_fx+=fx;\
+ motion_fy+=fy;\
+ motion_bx+=bx;\
+ motion_by+=by;\
+ end=0;\
+ }\
+ }
+#define CHECK_BIDIR2(a,b,c,d)\
+CHECK_BIDIR(a,b,c,d)\
+CHECK_BIDIR(-(a),-(b),-(c),-(d))
+
+#define CHECK_BIDIRR(a,b,c,d)\
+CHECK_BIDIR2(a,b,c,d)\
+CHECK_BIDIR2(b,c,d,a)\
+CHECK_BIDIR2(c,d,a,b)\
+CHECK_BIDIR2(d,a,b,c)
+
+ do{
+ end=1;
+
+ CHECK_BIDIRR( 0, 0, 0, 1)
+ if(s->avctx->bidir_refine > 1){
+ CHECK_BIDIRR( 0, 0, 1, 1)
+ CHECK_BIDIR2( 0, 1, 0, 1)
+ CHECK_BIDIR2( 1, 0, 1, 0)
+ CHECK_BIDIRR( 0, 0,-1, 1)
+ CHECK_BIDIR2( 0,-1, 0, 1)
+ CHECK_BIDIR2(-1, 0, 1, 0)
+ if(s->avctx->bidir_refine > 2){
+ CHECK_BIDIRR( 0, 1, 1, 1)
+ CHECK_BIDIRR( 0,-1, 1, 1)
+ CHECK_BIDIRR( 0, 1,-1, 1)
+ CHECK_BIDIRR( 0, 1, 1,-1)
+ if(s->avctx->bidir_refine > 3){
+ CHECK_BIDIR2( 1, 1, 1, 1)
+ CHECK_BIDIRR( 1, 1, 1,-1)
+ CHECK_BIDIR2( 1, 1,-1,-1)
+ CHECK_BIDIR2( 1,-1,-1, 1)
+ CHECK_BIDIR2( 1,-1, 1,-1)
+ }
+ }
+ }
+ }while(!end);
+ }
+
+ s->b_bidir_forw_mv_table[xy][0]= motion_fx;
+ s->b_bidir_forw_mv_table[xy][1]= motion_fy;
+ s->b_bidir_back_mv_table[xy][0]= motion_bx;
+ s->b_bidir_back_mv_table[xy][1]= motion_by;
- return fbmin;
+ return fbmin;
}
-static inline int direct_search(MpegEncContext * s,
- int mb_x, int mb_y)
+static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
{
- int P[6][2];
- const int mot_stride = s->mb_width + 2;
- const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
- int dmin, dmin2;
- int motion_fx, motion_fy, motion_bx, motion_by, motion_bx0, motion_by0;
- int motion_dx, motion_dy;
- const int motion_px= s->p_mv_table[mot_xy][0];
- const int motion_py= s->p_mv_table[mot_xy][1];
+ MotionEstContext * const c= &s->me;
+ int P[10][2];
+ const int mot_stride = s->mb_stride;
+ const int mot_xy = mb_y*mot_stride + mb_x;
+ const int shift= 1+s->quarter_sample;
+ int dmin, i;
const int time_pp= s->pp_time;
- const int time_bp= s->bp_time;
- const int time_pb= time_pp - time_bp;
- int bx, by;
- int mx, my, mx2, my2;
- uint8_t *ref_picture= s->me_scratchpad - (mb_x + 1 + (mb_y + 1)*s->linesize)*16;
+ const int time_pb= s->pb_time;
+ int mx, my, xmin, xmax, ymin, ymax;
int16_t (*mv_table)[2]= s->b_direct_mv_table;
- uint16_t *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
-
- /* thanks to iso-mpeg the rounding is different for the zero vector, so we need to handle that ... */
- motion_fx= (motion_px*time_pb)/time_pp;
- motion_fy= (motion_py*time_pb)/time_pp;
- motion_bx0= (-motion_px*time_bp)/time_pp;
- motion_by0= (-motion_py*time_bp)/time_pp;
- motion_dx= motion_dy=0;
- dmin2= check_bidir_mv(s, mb_x, mb_y,
- motion_fx, motion_fy,
- motion_bx0, motion_by0,
- motion_fx, motion_fy,
- motion_bx0, motion_by0) - s->qscale;
-
- motion_bx= motion_fx - motion_px;
- motion_by= motion_fy - motion_py;
- for(by=-1; by<2; by++){
- for(bx=-1; bx<2; bx++){
- uint8_t *dest_y = s->me_scratchpad + (by+1)*s->linesize*16 + (bx+1)*16;
- uint8_t *ptr;
- int dxy;
- int src_x, src_y;
- const int width= s->width;
- const int height= s->height;
-
- dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
- src_x = (mb_x + bx) * 16 + (motion_fx >> 1);
- src_y = (mb_y + by) * 16 + (motion_fy >> 1);
- src_x = clip(src_x, -16, width);
- if (src_x == width) dxy &= ~1;
- src_y = clip(src_y, -16, height);
- if (src_y == height) dxy &= ~2;
-
- ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
- put_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
- put_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
-
- dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
- src_x = (mb_x + bx) * 16 + (motion_bx >> 1);
- src_y = (mb_y + by) * 16 + (motion_by >> 1);
- src_x = clip(src_x, -16, width);
- if (src_x == width) dxy &= ~1;
- src_y = clip(src_y, -16, height);
- if (src_y == height) dxy &= ~2;
-
- avg_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
- avg_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
- }
- }
- P[0][0] = mv_table[mot_xy ][0];
- P[0][1] = mv_table[mot_xy ][1];
- P[1][0] = mv_table[mot_xy - 1][0];
- P[1][1] = mv_table[mot_xy - 1][1];
+ c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
+ ymin= xmin=(-32)>>shift;
+ ymax= xmax= 31>>shift;
- /* special case for first line */
- if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
- P[4][0] = P[1][0];
- P[4][1] = P[1][1];
- } else {
- P[2][0] = mv_table[mot_xy - mot_stride ][0];
- P[2][1] = mv_table[mot_xy - mot_stride ][1];
- P[3][0] = mv_table[mot_xy - mot_stride + 1 ][0];
- P[3][1] = mv_table[mot_xy - mot_stride + 1 ][1];
-
- P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
- P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
+ if(IS_8X8(s->next_picture.mb_type[mot_xy])){
+ s->mv_type= MV_TYPE_8X8;
+ }else{
+ s->mv_type= MV_TYPE_16X16;
}
- dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, -16, -16, 15, 15, ref_picture);
- if(mx==0 && my==0) dmin=99999999; // not representable, due to rounding stuff
- if(dmin2<dmin){
- dmin= dmin2;
- mx=0;
- my=0;
+
+ for(i=0; i<4; i++){
+ int index= s->block_index[i];
+ int min, max;
+
+ c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
+ c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
+ c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
+ c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
+// c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
+// c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
+
+ max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
+ min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
+ max+= 16*mb_x + 1; // +-1 is for the simpler rounding
+ min+= 16*mb_x - 1;
+ xmax= FFMIN(xmax, s->width - max);
+ xmin= FFMAX(xmin, - 16 - min);
+
+ max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
+ min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
+ max+= 16*mb_y + 1; // +-1 is for the simpler rounding
+ min+= 16*mb_y - 1;
+ ymax= FFMIN(ymax, s->height - max);
+ ymin= FFMAX(ymin, - 16 - min);
+
+ if(s->mv_type == MV_TYPE_16X16) break;
}
-#if 1
- mx2= mx= mx*2;
- my2= my= my*2;
- for(by=-1; by<2; by++){
- if(my2+by < -32) continue;
- for(bx=-1; bx<2; bx++){
- if(bx==0 && by==0) continue;
- if(mx2+bx < -32) continue;
- dmin2= check_bidir_mv(s, mb_x, mb_y,
- mx2+bx+motion_fx, my2+by+motion_fy,
- mx2+bx+motion_bx, my2+by+motion_by,
- mx2+bx+motion_fx, my2+by+motion_fy,
- motion_bx, motion_by) - s->qscale;
-
- if(dmin2<dmin){
- dmin=dmin2;
- mx= mx2 + bx;
- my= my2 + by;
- }
- }
+
+ assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
+
+ if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
+ s->b_direct_mv_table[mot_xy][0]= 0;
+ s->b_direct_mv_table[mot_xy][1]= 0;
+
+ return 256*256*256*64;
}
-#else
- mx*=2; my*=2;
-#endif
- if(mx==0 && my==0){
- motion_bx= motion_bx0;
- motion_by= motion_by0;
+
+ c->xmin= xmin;
+ c->ymin= ymin;
+ c->xmax= xmax;
+ c->ymax= ymax;
+ c->flags |= FLAG_DIRECT;
+ c->sub_flags |= FLAG_DIRECT;
+ c->pred_x=0;
+ c->pred_y=0;
+
+ P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
+ P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
+
+ /* special case for first line */
+ if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
+ P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
+ P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
+ P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
+ P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
+
+ P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
+ P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
}
- s->b_direct_mv_table[mot_xy][0]= mx;
- s->b_direct_mv_table[mot_xy][1]= my;
- s->b_direct_forw_mv_table[mot_xy][0]= motion_fx + mx;
- s->b_direct_forw_mv_table[mot_xy][1]= motion_fy + my;
- s->b_direct_back_mv_table[mot_xy][0]= motion_bx + mx;
- s->b_direct_back_mv_table[mot_xy][1]= motion_by + my;
+ dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
+ if(c->sub_flags&FLAG_QPEL)
+ dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
+ else
+ dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
+
+ if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
+ dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
+
+ get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
+
+ mv_table[mot_xy][0]= mx;
+ mv_table[mot_xy][1]= my;
+ c->flags &= ~FLAG_DIRECT;
+ c->sub_flags &= ~FLAG_DIRECT;
+
return dmin;
}
void ff_estimate_b_frame_motion(MpegEncContext * s,
int mb_x, int mb_y)
{
- const int mot_stride = s->mb_width + 2;
- const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
- const int quant= s->qscale;
- int fmin, bmin, dmin, fbmin;
+ MotionEstContext * const c= &s->me;
+ const int penalty_factor= c->mb_penalty_factor;
+ int fmin, bmin, dmin, fbmin, bimin, fimin;
int type=0;
- int motion_fx, motion_fy, motion_bx, motion_by;
-
- dmin= direct_search(s, mb_x, mb_y);
+ const int xy = mb_y*s->mb_stride + mb_x;
+ init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
- fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);
- bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code) - quant;
-//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
+ get_limits(s, 16*mb_x, 16*mb_y);
- fbmin= bidir_refine(s, mb_x, mb_y);
+ c->skip=0;
- if(s->flags&CODEC_FLAG_HQ){
- type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT;
- }else{
- int score= dmin;
- type=MB_TYPE_DIRECT;
-
- if(fmin<score){
- score=fmin;
- type= MB_TYPE_FORWARD;
+ if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
+ int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
+
+ score= ((unsigned)(score*score + 128*256))>>16;
+ c->mc_mb_var_sum_temp += score;
+ s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
+ s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
+
+ return;
+ }
+
+ if(c->avctx->me_threshold){
+ int vard= check_input_motion(s, mb_x, mb_y, 0);
+
+ if((vard+128)>>8 < c->avctx->me_threshold){
+// pix = c->src[0][0];
+// sum = s->dsp.pix_sum(pix, s->linesize);
+// varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
+
+// pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
+ s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
+/* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
+ c->mb_var_sum_temp += (varc+128)>>8;*/
+ c->mc_mb_var_sum_temp += (vard+128)>>8;
+/* if (vard <= 64<<8 || vard < varc) {
+ c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
+ }else{
+ c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
+ }*/
+ return;
+ }
+ if((vard+128)>>8 < c->avctx->mb_threshold){
+ type= s->mb_type[mb_y*s->mb_stride + mb_x];
+ if(type == CANDIDATE_MB_TYPE_DIRECT){
+ direct_search(s, mb_x, mb_y);
+ }
+ if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
+ c->skip=0;
+ ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
+ }
+ if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
+ c->skip=0;
+ ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
+ }
+ if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
+ c->skip=0;
+ c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
+ interlaced_search(s, 0,
+ s->b_field_mv_table[0], s->b_field_select_table[0],
+ s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
+ }
+ if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
+ c->skip=0;
+ c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
+ interlaced_search(s, 2,
+ s->b_field_mv_table[1], s->b_field_select_table[1],
+ s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
+ }
+ return;
+ }
+ }
+
+ if (s->codec_id == CODEC_ID_MPEG4)
+ dmin= direct_search(s, mb_x, mb_y);
+ else
+ dmin= INT_MAX;
+//FIXME penalty stuff for non mpeg4
+ c->skip=0;
+ fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
+
+ c->skip=0;
+ bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
+//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
+
+ c->skip=0;
+ fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
+//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
+
+ if(s->flags & CODEC_FLAG_INTERLACED_ME){
+//FIXME mb type penalty
+ c->skip=0;
+ c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
+ fimin= interlaced_search(s, 0,
+ s->b_field_mv_table[0], s->b_field_select_table[0],
+ s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
+ c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
+ bimin= interlaced_search(s, 2,
+ s->b_field_mv_table[1], s->b_field_select_table[1],
+ s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
+ }else
+ fimin= bimin= INT_MAX;
+
+ {
+ int score= fmin;
+ type = CANDIDATE_MB_TYPE_FORWARD;
+
+ if (dmin <= score){
+ score = dmin;
+ type = CANDIDATE_MB_TYPE_DIRECT;
}
if(bmin<score){
score=bmin;
- type= MB_TYPE_BACKWARD;
+ type= CANDIDATE_MB_TYPE_BACKWARD;
}
if(fbmin<score){
score=fbmin;
- type= MB_TYPE_BIDIR;
+ type= CANDIDATE_MB_TYPE_BIDIR;
+ }
+ if(fimin<score){
+ score=fimin;
+ type= CANDIDATE_MB_TYPE_FORWARD_I;
}
- s->mc_mb_var += score;
+ if(bimin<score){
+ score=bimin;
+ type= CANDIDATE_MB_TYPE_BACKWARD_I;
+ }
+
+ score= ((unsigned)(score*score + 128*256))>>16;
+ c->mc_mb_var_sum_temp += score;
+ s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
+ }
+
+ if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
+ type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
+ if(fimin < INT_MAX)
+ type |= CANDIDATE_MB_TYPE_FORWARD_I;
+ if(bimin < INT_MAX)
+ type |= CANDIDATE_MB_TYPE_BACKWARD_I;
+ if(fimin < INT_MAX && bimin < INT_MAX){
+ type |= CANDIDATE_MB_TYPE_BIDIR_I;
+ }
+ //FIXME something smarter
+ if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
+ if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
+ type |= CANDIDATE_MB_TYPE_DIRECT0;
+#if 0
+ if(s->out_format == FMT_MPEG1)
+ type |= CANDIDATE_MB_TYPE_INTRA;
+#endif
}
- s->mb_type[mb_y*s->mb_width + mb_x]= type;
+ s->mb_type[mb_y*s->mb_stride + mb_x]= type;
}
/* find best f_code for ME which do unlimited searches */
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
{
- int f_code;
-
if(s->me_method>=ME_EPZS){
- int mv_num[8];
- int i, y;
- int loose=0;
- UINT8 * fcode_tab= s->fcode_tab;
+ int score[8];
+ int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
+ uint8_t * fcode_tab= s->fcode_tab;
+ int best_fcode=-1;
+ int best_score=-10000000;
- for(i=0; i<8; i++) mv_num[i]=0;
+ if(s->msmpeg4_version)
+ range= FFMIN(range, 16);
+ else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
+ range= FFMIN(range, 256);
+
+ for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
for(y=0; y<s->mb_height; y++){
int x;
- int xy= (y+1)* (s->mb_width+2) + 1;
- i= y*s->mb_width;
+ int xy= y*s->mb_stride;
for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i] & type){
- mv_num[ fcode_tab[mv_table[xy][0] + MAX_MV] ]++;
- mv_num[ fcode_tab[mv_table[xy][1] + MAX_MV] ]++;
-//printf("%d %d %d\n", s->mv_table[0][i], fcode_tab[s->mv_table[0][i] + MAX_MV], i);
+ if(s->mb_type[xy] & type){
+ int mx= mv_table[xy][0];
+ int my= mv_table[xy][1];
+ int fcode= FFMAX(fcode_tab[mx + MAX_MV],
+ fcode_tab[my + MAX_MV]);
+ int j;
+
+ if(mx >= range || mx < -range ||
+ my >= range || my < -range)
+ continue;
+
+ for(j=0; j<fcode && j<8; j++){
+ if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
+ score[j]-= 170;
+ }
}
- i++;
xy++;
}
}
- for(i=MAX_FCODE; i>1; i--){
- int threshold;
- loose+= mv_num[i];
-
- if(s->pict_type==B_TYPE) threshold= 0;
- else threshold= s->mb_num/20; //FIXME
- if(loose > threshold) break;
+ for(i=1; i<8; i++){
+ if(score[i] > best_score){
+ best_score= score[i];
+ best_fcode= i;
+ }
+// printf("%d %d\n", i, score[i]);
}
+
// printf("fcode: %d type: %d\n", i, s->pict_type);
- return i;
+ return best_fcode;
/* for(i=0; i<=MAX_FCODE; i++){
printf("%d ", mv_num[i]);
}
void ff_fix_long_p_mvs(MpegEncContext * s)
{
+ MotionEstContext * const c= &s->me;
const int f_code= s->f_code;
- int y;
- UINT8 * fcode_tab= s->fcode_tab;
+ int y, range;
+ assert(s->pict_type==FF_P_TYPE);
- /* clip / convert to intra 16x16 type MVs */
- for(y=0; y<s->mb_height; y++){
- int x;
- int xy= (y+1)* (s->mb_width+2)+1;
- int i= y*s->mb_width;
- for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i]&MB_TYPE_INTER){
- if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
- || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
- || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
- || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
- s->mb_type[i] &= ~MB_TYPE_INTER;
- s->mb_type[i] |= MB_TYPE_INTRA;
- s->p_mv_table[xy][0] = 0;
- s->p_mv_table[xy][1] = 0;
- }
- }
- xy++;
- i++;
- }
- }
+ range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
+ assert(range <= 16 || !s->msmpeg4_version);
+ assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
+
+ if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
+
+//printf("%d no:%d %d//\n", clip, noclip, f_code);
if(s->flags&CODEC_FLAG_4MV){
- const int wrap= 2+ s->mb_width*2;
+ const int wrap= s->b8_stride;
/* clip / convert to intra 8x8 type MVs */
for(y=0; y<s->mb_height; y++){
- int xy= (y*2 + 1)*wrap + 1;
- int i= y*s->mb_width;
+ int xy= y*2*wrap;
+ int i= y*s->mb_stride;
int x;
for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i]&MB_TYPE_INTER4V){
+ if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
int block;
for(block=0; block<4; block++){
int off= (block& 1) + (block>>1)*wrap;
- int mx= s->motion_val[ xy + off ][0];
- int my= s->motion_val[ xy + off ][1];
-
- if( fcode_tab[mx + MAX_MV] > f_code
- || fcode_tab[mx + MAX_MV] == 0
- || fcode_tab[my + MAX_MV] > f_code
- || fcode_tab[my + MAX_MV] == 0 ){
- s->mb_type[i] &= ~MB_TYPE_INTER4V;
- s->mb_type[i] |= MB_TYPE_INTRA;
+ int mx= s->current_picture.motion_val[0][ xy + off ][0];
+ int my= s->current_picture.motion_val[0][ xy + off ][1];
+
+ if( mx >=range || mx <-range
+ || my >=range || my <-range){
+ s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
+ s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
+ s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
}
}
- xy+=2;
- i++;
}
+ xy+=2;
+ i++;
}
}
}
}
-void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
+/**
+ *
+ * @param truncate 1 for truncation, 0 for using intra
+ */
+void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
+ int16_t (*mv_table)[2], int f_code, int type, int truncate)
{
- int y;
- UINT8 * fcode_tab= s->fcode_tab;
+ MotionEstContext * const c= &s->me;
+ int y, h_range, v_range;
+
+ // RAL: 8 in MPEG-1, 16 in MPEG-4
+ int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
+
+ if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
+
+ h_range= range;
+ v_range= field_select_table ? range>>1 : range;
/* clip / convert to intra 16x16 type MVs */
for(y=0; y<s->mb_height; y++){
int x;
- int xy= (y+1)* (s->mb_width+2)+1;
- int i= y*s->mb_width;
+ int xy= y*s->mb_stride;
for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i]&type){
- if( fcode_tab[mv_table[xy][0] + MAX_MV] > f_code
- || fcode_tab[mv_table[xy][0] + MAX_MV] == 0
- || fcode_tab[mv_table[xy][1] + MAX_MV] > f_code
- || fcode_tab[mv_table[xy][1] + MAX_MV] == 0 ){
- if(s->mb_type[i]&(~type)) s->mb_type[i] &= ~type;
- else{
- mv_table[xy][0] = 0;
- mv_table[xy][1] = 0;
- //this is certainly bad FIXME
+ if (s->mb_type[xy] & type){ // RAL: "type" test added...
+ if(field_select_table==NULL || field_select_table[xy] == field_select){
+ if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
+ || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
+
+ if(truncate){
+ if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
+ else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
+ if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
+ else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
+ }else{
+ s->mb_type[xy] &= ~type;
+ s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
+ mv_table[xy][0]=
+ mv_table[xy][1]= 0;
+ }
}
}
}
xy++;
- i++;
}
}
}