michael@0: /* michael@0: * Copyright (c) 2010 The WebM project authors. All Rights Reserved. michael@0: * michael@0: * Use of this source code is governed by a BSD-style license michael@0: * that can be found in the LICENSE file in the root of the source michael@0: * tree. An additional intellectual property rights grant can be found michael@0: * in the file PATENTS. All contributing project authors may michael@0: * be found in the AUTHORS file in the root of the source tree. michael@0: */ michael@0: michael@0: #include michael@0: #include michael@0: #include michael@0: michael@0: #include "./vpx_config.h" michael@0: michael@0: #include "vpx_mem/vpx_mem.h" michael@0: michael@0: #include "vp9/common/vp9_findnearmv.h" michael@0: #include "vp9/common/vp9_common.h" michael@0: michael@0: #include "vp9/encoder/vp9_onyx_int.h" michael@0: #include "vp9/encoder/vp9_mcomp.h" michael@0: michael@0: // #define NEW_DIAMOND_SEARCH michael@0: michael@0: void vp9_clamp_mv_min_max(MACROBLOCK *x, MV *mv) { michael@0: const int col_min = (mv->col >> 3) - MAX_FULL_PEL_VAL + (mv->col & 7 ? 1 : 0); michael@0: const int row_min = (mv->row >> 3) - MAX_FULL_PEL_VAL + (mv->row & 7 ? 1 : 0); michael@0: const int col_max = (mv->col >> 3) + MAX_FULL_PEL_VAL; michael@0: const int row_max = (mv->row >> 3) + MAX_FULL_PEL_VAL; michael@0: michael@0: // Get intersection of UMV window and valid MV window to reduce # of checks michael@0: // in diamond search. michael@0: if (x->mv_col_min < col_min) michael@0: x->mv_col_min = col_min; michael@0: if (x->mv_col_max > col_max) michael@0: x->mv_col_max = col_max; michael@0: if (x->mv_row_min < row_min) michael@0: x->mv_row_min = row_min; michael@0: if (x->mv_row_max > row_max) michael@0: x->mv_row_max = row_max; michael@0: } michael@0: michael@0: int vp9_init_search_range(VP9_COMP *cpi, int size) { michael@0: int sr = 0; michael@0: michael@0: // Minimum search size no matter what the passed in value. michael@0: size = MAX(16, size); michael@0: michael@0: while ((size << sr) < MAX_FULL_PEL_VAL) michael@0: sr++; michael@0: michael@0: if (sr) michael@0: sr--; michael@0: michael@0: sr += cpi->sf.reduce_first_step_size; michael@0: sr = MIN(sr, (cpi->sf.max_step_search_steps - 2)); michael@0: return sr; michael@0: } michael@0: michael@0: static INLINE int mv_cost(const MV *mv, michael@0: const int *joint_cost, int *comp_cost[2]) { michael@0: return joint_cost[vp9_get_mv_joint(mv)] + michael@0: comp_cost[0][mv->row] + comp_cost[1][mv->col]; michael@0: } michael@0: michael@0: int vp9_mv_bit_cost(const MV *mv, const MV *ref, michael@0: const int *mvjcost, int *mvcost[2], int weight) { michael@0: const MV diff = { mv->row - ref->row, michael@0: mv->col - ref->col }; michael@0: return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7); michael@0: } michael@0: michael@0: static int mv_err_cost(const MV *mv, const MV *ref, michael@0: const int *mvjcost, int *mvcost[2], michael@0: int error_per_bit) { michael@0: if (mvcost) { michael@0: const MV diff = { mv->row - ref->row, michael@0: mv->col - ref->col }; michael@0: return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * michael@0: error_per_bit, 13); michael@0: } michael@0: return 0; michael@0: } michael@0: michael@0: static int mvsad_err_cost(const MV *mv, const MV *ref, michael@0: const int *mvjsadcost, int *mvsadcost[2], michael@0: int error_per_bit) { michael@0: if (mvsadcost) { michael@0: const MV diff = { mv->row - ref->row, michael@0: mv->col - ref->col }; michael@0: return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjsadcost, mvsadcost) * michael@0: error_per_bit, 8); michael@0: } michael@0: return 0; michael@0: } michael@0: michael@0: void vp9_init_dsmotion_compensation(MACROBLOCK *x, int stride) { michael@0: int len; michael@0: int search_site_count = 0; michael@0: michael@0: // Generate offsets for 4 search sites per step. michael@0: x->ss[search_site_count].mv.col = 0; michael@0: x->ss[search_site_count].mv.row = 0; michael@0: x->ss[search_site_count].offset = 0; michael@0: search_site_count++; michael@0: michael@0: for (len = MAX_FIRST_STEP; len > 0; len /= 2) { michael@0: // Compute offsets for search sites. michael@0: x->ss[search_site_count].mv.col = 0; michael@0: x->ss[search_site_count].mv.row = -len; michael@0: x->ss[search_site_count].offset = -len * stride; michael@0: search_site_count++; michael@0: michael@0: // Compute offsets for search sites. michael@0: x->ss[search_site_count].mv.col = 0; michael@0: x->ss[search_site_count].mv.row = len; michael@0: x->ss[search_site_count].offset = len * stride; michael@0: search_site_count++; michael@0: michael@0: // Compute offsets for search sites. michael@0: x->ss[search_site_count].mv.col = -len; michael@0: x->ss[search_site_count].mv.row = 0; michael@0: x->ss[search_site_count].offset = -len; michael@0: search_site_count++; michael@0: michael@0: // Compute offsets for search sites. michael@0: x->ss[search_site_count].mv.col = len; michael@0: x->ss[search_site_count].mv.row = 0; michael@0: x->ss[search_site_count].offset = len; michael@0: search_site_count++; michael@0: } michael@0: michael@0: x->ss_count = search_site_count; michael@0: x->searches_per_step = 4; michael@0: } michael@0: michael@0: void vp9_init3smotion_compensation(MACROBLOCK *x, int stride) { michael@0: int len, ss_count = 1; michael@0: michael@0: x->ss[0].mv.col = x->ss[0].mv.row = 0; michael@0: x->ss[0].offset = 0; michael@0: michael@0: for (len = MAX_FIRST_STEP; len > 0; len /= 2) { michael@0: // Generate offsets for 8 search sites per step. michael@0: const MV ss_mvs[8] = { michael@0: {-len, 0 }, {len, 0 }, { 0, -len}, {0, len}, michael@0: {-len, -len}, {-len, len}, {len, -len}, {len, len} michael@0: }; michael@0: int i; michael@0: for (i = 0; i < 8; ++i) { michael@0: search_site *const ss = &x->ss[ss_count++]; michael@0: ss->mv = ss_mvs[i]; michael@0: ss->offset = ss->mv.row * stride + ss->mv.col; michael@0: } michael@0: } michael@0: michael@0: x->ss_count = ss_count; michael@0: x->searches_per_step = 8; michael@0: } michael@0: michael@0: /* michael@0: * To avoid the penalty for crossing cache-line read, preload the reference michael@0: * area in a small buffer, which is aligned to make sure there won't be crossing michael@0: * cache-line read while reading from this buffer. This reduced the cpu michael@0: * cycles spent on reading ref data in sub-pixel filter functions. michael@0: * TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x michael@0: * 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we michael@0: * could reduce the area. michael@0: */ michael@0: michael@0: /* estimated cost of a motion vector (r,c) */ michael@0: #define MVC(r, c) \ michael@0: (mvcost ? \ michael@0: ((mvjcost[((r) != rr) * 2 + ((c) != rc)] + \ michael@0: mvcost[0][((r) - rr)] + mvcost[1][((c) - rc)]) * \ michael@0: error_per_bit + 4096) >> 13 : 0) michael@0: michael@0: michael@0: #define SP(x) (((x) & 7) << 1) // convert motion vector component to offset michael@0: // for svf calc michael@0: michael@0: #define IFMVCV(r, c, s, e) \ michael@0: if (c >= minc && c <= maxc && r >= minr && r <= maxr) \ michael@0: s \ michael@0: else \ michael@0: e; michael@0: michael@0: /* pointer to predictor base of a motionvector */ michael@0: #define PRE(r, c) (y + (((r) >> 3) * y_stride + ((c) >> 3) -(offset))) michael@0: michael@0: /* returns subpixel variance error function */ michael@0: #define DIST(r, c) \ michael@0: vfp->svf(PRE(r, c), y_stride, SP(c), SP(r), z, src_stride, &sse) michael@0: michael@0: /* checks if (r, c) has better score than previous best */ michael@0: #define CHECK_BETTER(v, r, c) \ michael@0: IFMVCV(r, c, { \ michael@0: thismse = (DIST(r, c)); \ michael@0: if ((v = MVC(r, c) + thismse) < besterr) { \ michael@0: besterr = v; \ michael@0: br = r; \ michael@0: bc = c; \ michael@0: *distortion = thismse; \ michael@0: *sse1 = sse; \ michael@0: } \ michael@0: }, \ michael@0: v = INT_MAX;) michael@0: michael@0: #define FIRST_LEVEL_CHECKS \ michael@0: { \ michael@0: unsigned int left, right, up, down, diag; \ michael@0: CHECK_BETTER(left, tr, tc - hstep); \ michael@0: CHECK_BETTER(right, tr, tc + hstep); \ michael@0: CHECK_BETTER(up, tr - hstep, tc); \ michael@0: CHECK_BETTER(down, tr + hstep, tc); \ michael@0: whichdir = (left < right ? 0 : 1) + \ michael@0: (up < down ? 0 : 2); \ michael@0: switch (whichdir) { \ michael@0: case 0: \ michael@0: CHECK_BETTER(diag, tr - hstep, tc - hstep); \ michael@0: break; \ michael@0: case 1: \ michael@0: CHECK_BETTER(diag, tr - hstep, tc + hstep); \ michael@0: break; \ michael@0: case 2: \ michael@0: CHECK_BETTER(diag, tr + hstep, tc - hstep); \ michael@0: break; \ michael@0: case 3: \ michael@0: CHECK_BETTER(diag, tr + hstep, tc + hstep); \ michael@0: break; \ michael@0: } \ michael@0: } michael@0: michael@0: #define SECOND_LEVEL_CHECKS \ michael@0: { \ michael@0: int kr, kc; \ michael@0: unsigned int second; \ michael@0: if (tr != br && tc != bc) { \ michael@0: kr = br - tr; \ michael@0: kc = bc - tc; \ michael@0: CHECK_BETTER(second, tr + kr, tc + 2 * kc); \ michael@0: CHECK_BETTER(second, tr + 2 * kr, tc + kc); \ michael@0: } else if (tr == br && tc != bc) { \ michael@0: kc = bc - tc; \ michael@0: CHECK_BETTER(second, tr + hstep, tc + 2 * kc); \ michael@0: CHECK_BETTER(second, tr - hstep, tc + 2 * kc); \ michael@0: switch (whichdir) { \ michael@0: case 0: \ michael@0: case 1: \ michael@0: CHECK_BETTER(second, tr + hstep, tc + kc); \ michael@0: break; \ michael@0: case 2: \ michael@0: case 3: \ michael@0: CHECK_BETTER(second, tr - hstep, tc + kc); \ michael@0: break; \ michael@0: } \ michael@0: } else if (tr != br && tc == bc) { \ michael@0: kr = br - tr; \ michael@0: CHECK_BETTER(second, tr + 2 * kr, tc + hstep); \ michael@0: CHECK_BETTER(second, tr + 2 * kr, tc - hstep); \ michael@0: switch (whichdir) { \ michael@0: case 0: \ michael@0: case 2: \ michael@0: CHECK_BETTER(second, tr + kr, tc + hstep); \ michael@0: break; \ michael@0: case 1: \ michael@0: case 3: \ michael@0: CHECK_BETTER(second, tr + kr, tc - hstep); \ michael@0: break; \ michael@0: } \ michael@0: } \ michael@0: } michael@0: michael@0: int vp9_find_best_sub_pixel_iterative(MACROBLOCK *x, michael@0: MV *bestmv, const MV *ref_mv, michael@0: int allow_hp, michael@0: int error_per_bit, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int forced_stop, michael@0: int iters_per_step, michael@0: int *mvjcost, int *mvcost[2], michael@0: int *distortion, michael@0: unsigned int *sse1) { michael@0: uint8_t *z = x->plane[0].src.buf; michael@0: int src_stride = x->plane[0].src.stride; michael@0: MACROBLOCKD *xd = &x->e_mbd; michael@0: michael@0: unsigned int besterr = INT_MAX; michael@0: unsigned int sse; michael@0: unsigned int whichdir; michael@0: unsigned int halfiters = iters_per_step; michael@0: unsigned int quarteriters = iters_per_step; michael@0: unsigned int eighthiters = iters_per_step; michael@0: int thismse; michael@0: michael@0: const int y_stride = xd->plane[0].pre[0].stride; michael@0: const int offset = bestmv->row * y_stride + bestmv->col; michael@0: uint8_t *y = xd->plane[0].pre[0].buf + offset; michael@0: michael@0: int rr = ref_mv->row; michael@0: int rc = ref_mv->col; michael@0: int br = bestmv->row * 8; michael@0: int bc = bestmv->col * 8; michael@0: int hstep = 4; michael@0: const int minc = MAX(x->mv_col_min * 8, ref_mv->col - MV_MAX); michael@0: const int maxc = MIN(x->mv_col_max * 8, ref_mv->col + MV_MAX); michael@0: const int minr = MAX(x->mv_row_min * 8, ref_mv->row - MV_MAX); michael@0: const int maxr = MIN(x->mv_row_max * 8, ref_mv->row + MV_MAX); michael@0: michael@0: int tr = br; michael@0: int tc = bc; michael@0: michael@0: // central mv michael@0: bestmv->row <<= 3; michael@0: bestmv->col <<= 3; michael@0: michael@0: // calculate central point error michael@0: besterr = vfp->vf(y, y_stride, z, src_stride, sse1); michael@0: *distortion = besterr; michael@0: besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit); michael@0: michael@0: // TODO(jbb): Each subsequent iteration checks at least one point in michael@0: // common with the last iteration could be 2 if diagonal is selected. michael@0: while (halfiters--) { michael@0: // 1/2 pel michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: michael@0: // TODO(yaowu): Each subsequent iteration checks at least one point in common michael@0: // with the last iteration could be 2 if diagonal is selected. michael@0: michael@0: // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only michael@0: if (forced_stop != 2) { michael@0: hstep >>= 1; michael@0: while (quarteriters--) { michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: } michael@0: michael@0: if (allow_hp && vp9_use_mv_hp(ref_mv) && forced_stop == 0) { michael@0: hstep >>= 1; michael@0: while (eighthiters--) { michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: } michael@0: michael@0: bestmv->row = br; michael@0: bestmv->col = bc; michael@0: michael@0: if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) || michael@0: (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3))) michael@0: return INT_MAX; michael@0: michael@0: return besterr; michael@0: } michael@0: michael@0: int vp9_find_best_sub_pixel_tree(MACROBLOCK *x, michael@0: MV *bestmv, const MV *ref_mv, michael@0: int allow_hp, michael@0: int error_per_bit, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int forced_stop, michael@0: int iters_per_step, michael@0: int *mvjcost, int *mvcost[2], michael@0: int *distortion, michael@0: unsigned int *sse1) { michael@0: uint8_t *z = x->plane[0].src.buf; michael@0: const int src_stride = x->plane[0].src.stride; michael@0: MACROBLOCKD *xd = &x->e_mbd; michael@0: unsigned int besterr = INT_MAX; michael@0: unsigned int sse; michael@0: unsigned int whichdir; michael@0: int thismse; michael@0: unsigned int halfiters = iters_per_step; michael@0: unsigned int quarteriters = iters_per_step; michael@0: unsigned int eighthiters = iters_per_step; michael@0: michael@0: const int y_stride = xd->plane[0].pre[0].stride; michael@0: const int offset = bestmv->row * y_stride + bestmv->col; michael@0: uint8_t *y = xd->plane[0].pre[0].buf + offset; michael@0: michael@0: int rr = ref_mv->row; michael@0: int rc = ref_mv->col; michael@0: int br = bestmv->row * 8; michael@0: int bc = bestmv->col * 8; michael@0: int hstep = 4; michael@0: const int minc = MAX(x->mv_col_min * 8, ref_mv->col - MV_MAX); michael@0: const int maxc = MIN(x->mv_col_max * 8, ref_mv->col + MV_MAX); michael@0: const int minr = MAX(x->mv_row_min * 8, ref_mv->row - MV_MAX); michael@0: const int maxr = MIN(x->mv_row_max * 8, ref_mv->row + MV_MAX); michael@0: michael@0: int tr = br; michael@0: int tc = bc; michael@0: michael@0: // central mv michael@0: bestmv->row *= 8; michael@0: bestmv->col *= 8; michael@0: michael@0: // calculate central point error michael@0: besterr = vfp->vf(y, y_stride, z, src_stride, sse1); michael@0: *distortion = besterr; michael@0: besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit); michael@0: michael@0: // 1/2 pel michael@0: FIRST_LEVEL_CHECKS; michael@0: if (halfiters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: michael@0: // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only michael@0: if (forced_stop != 2) { michael@0: hstep >>= 1; michael@0: FIRST_LEVEL_CHECKS; michael@0: if (quarteriters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: michael@0: if (allow_hp && vp9_use_mv_hp(ref_mv) && forced_stop == 0) { michael@0: hstep >>= 1; michael@0: FIRST_LEVEL_CHECKS; michael@0: if (eighthiters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: michael@0: bestmv->row = br; michael@0: bestmv->col = bc; michael@0: michael@0: if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) || michael@0: (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3))) michael@0: return INT_MAX; michael@0: michael@0: return besterr; michael@0: } michael@0: michael@0: #undef DIST michael@0: /* returns subpixel variance error function */ michael@0: #define DIST(r, c) \ michael@0: vfp->svaf(PRE(r, c), y_stride, SP(c), SP(r), \ michael@0: z, src_stride, &sse, second_pred) michael@0: michael@0: int vp9_find_best_sub_pixel_comp_iterative(MACROBLOCK *x, michael@0: MV *bestmv, const MV *ref_mv, michael@0: int allow_hp, michael@0: int error_per_bit, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int forced_stop, michael@0: int iters_per_step, michael@0: int *mvjcost, int *mvcost[2], michael@0: int *distortion, michael@0: unsigned int *sse1, michael@0: const uint8_t *second_pred, michael@0: int w, int h) { michael@0: uint8_t *const z = x->plane[0].src.buf; michael@0: const int src_stride = x->plane[0].src.stride; michael@0: MACROBLOCKD *const xd = &x->e_mbd; michael@0: michael@0: unsigned int besterr = INT_MAX; michael@0: unsigned int sse; michael@0: unsigned int whichdir; michael@0: unsigned int halfiters = iters_per_step; michael@0: unsigned int quarteriters = iters_per_step; michael@0: unsigned int eighthiters = iters_per_step; michael@0: int thismse; michael@0: michael@0: DECLARE_ALIGNED_ARRAY(16, uint8_t, comp_pred, 64 * 64); michael@0: const int y_stride = xd->plane[0].pre[0].stride; michael@0: const int offset = bestmv->row * y_stride + bestmv->col; michael@0: uint8_t *const y = xd->plane[0].pre[0].buf + offset; michael@0: michael@0: int rr = ref_mv->row; michael@0: int rc = ref_mv->col; michael@0: int br = bestmv->row * 8; michael@0: int bc = bestmv->col * 8; michael@0: int hstep = 4; michael@0: const int minc = MAX(x->mv_col_min * 8, ref_mv->col - MV_MAX); michael@0: const int maxc = MIN(x->mv_col_max * 8, ref_mv->col + MV_MAX); michael@0: const int minr = MAX(x->mv_row_min * 8, ref_mv->row - MV_MAX); michael@0: const int maxr = MIN(x->mv_row_max * 8, ref_mv->row + MV_MAX); michael@0: michael@0: int tr = br; michael@0: int tc = bc; michael@0: michael@0: // central mv michael@0: bestmv->row *= 8; michael@0: bestmv->col *= 8; michael@0: michael@0: // calculate central point error michael@0: // TODO(yunqingwang): central pointer error was already calculated in full- michael@0: // pixel search, and can be passed in this function. michael@0: comp_avg_pred(comp_pred, second_pred, w, h, y, y_stride); michael@0: besterr = vfp->vf(comp_pred, w, z, src_stride, sse1); michael@0: *distortion = besterr; michael@0: besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit); michael@0: michael@0: // Each subsequent iteration checks at least one point in michael@0: // common with the last iteration could be 2 ( if diag selected) michael@0: while (halfiters--) { michael@0: // 1/2 pel michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: michael@0: // Each subsequent iteration checks at least one point in common with michael@0: // the last iteration could be 2 ( if diag selected) 1/4 pel michael@0: michael@0: // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only michael@0: if (forced_stop != 2) { michael@0: hstep >>= 1; michael@0: while (quarteriters--) { michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: } michael@0: michael@0: if (allow_hp && vp9_use_mv_hp(ref_mv) && forced_stop == 0) { michael@0: hstep >>= 1; michael@0: while (eighthiters--) { michael@0: FIRST_LEVEL_CHECKS; michael@0: // no reason to check the same one again. michael@0: if (tr == br && tc == bc) michael@0: break; michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: } michael@0: bestmv->row = br; michael@0: bestmv->col = bc; michael@0: michael@0: if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) || michael@0: (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3))) michael@0: return INT_MAX; michael@0: michael@0: return besterr; michael@0: } michael@0: michael@0: int vp9_find_best_sub_pixel_comp_tree(MACROBLOCK *x, michael@0: MV *bestmv, const MV *ref_mv, michael@0: int allow_hp, michael@0: int error_per_bit, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int forced_stop, michael@0: int iters_per_step, michael@0: int *mvjcost, int *mvcost[2], michael@0: int *distortion, michael@0: unsigned int *sse1, michael@0: const uint8_t *second_pred, michael@0: int w, int h) { michael@0: uint8_t *z = x->plane[0].src.buf; michael@0: const int src_stride = x->plane[0].src.stride; michael@0: MACROBLOCKD *xd = &x->e_mbd; michael@0: unsigned int besterr = INT_MAX; michael@0: unsigned int sse; michael@0: unsigned int whichdir; michael@0: int thismse; michael@0: unsigned int halfiters = iters_per_step; michael@0: unsigned int quarteriters = iters_per_step; michael@0: unsigned int eighthiters = iters_per_step; michael@0: michael@0: DECLARE_ALIGNED_ARRAY(16, uint8_t, comp_pred, 64 * 64); michael@0: const int y_stride = xd->plane[0].pre[0].stride; michael@0: const int offset = bestmv->row * y_stride + bestmv->col; michael@0: uint8_t *y = xd->plane[0].pre[0].buf + offset; michael@0: michael@0: int rr = ref_mv->row; michael@0: int rc = ref_mv->col; michael@0: int br = bestmv->row * 8; michael@0: int bc = bestmv->col * 8; michael@0: int hstep = 4; michael@0: const int minc = MAX(x->mv_col_min * 8, ref_mv->col - MV_MAX); michael@0: const int maxc = MIN(x->mv_col_max * 8, ref_mv->col + MV_MAX); michael@0: const int minr = MAX(x->mv_row_min * 8, ref_mv->row - MV_MAX); michael@0: const int maxr = MIN(x->mv_row_max * 8, ref_mv->row + MV_MAX); michael@0: michael@0: int tr = br; michael@0: int tc = bc; michael@0: michael@0: // central mv michael@0: bestmv->row *= 8; michael@0: bestmv->col *= 8; michael@0: michael@0: // calculate central point error michael@0: // TODO(yunqingwang): central pointer error was already calculated in full- michael@0: // pixel search, and can be passed in this function. michael@0: comp_avg_pred(comp_pred, second_pred, w, h, y, y_stride); michael@0: besterr = vfp->vf(comp_pred, w, z, src_stride, sse1); michael@0: *distortion = besterr; michael@0: besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit); michael@0: michael@0: // Each subsequent iteration checks at least one point in michael@0: // common with the last iteration could be 2 ( if diag selected) michael@0: // 1/2 pel michael@0: FIRST_LEVEL_CHECKS; michael@0: if (halfiters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: michael@0: // Each subsequent iteration checks at least one point in common with michael@0: // the last iteration could be 2 ( if diag selected) 1/4 pel michael@0: michael@0: // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only michael@0: if (forced_stop != 2) { michael@0: hstep >>= 1; michael@0: FIRST_LEVEL_CHECKS; michael@0: if (quarteriters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: michael@0: if (allow_hp && vp9_use_mv_hp(ref_mv) && forced_stop == 0) { michael@0: hstep >>= 1; michael@0: FIRST_LEVEL_CHECKS; michael@0: if (eighthiters > 1) { michael@0: SECOND_LEVEL_CHECKS; michael@0: } michael@0: tr = br; michael@0: tc = bc; michael@0: } michael@0: bestmv->row = br; michael@0: bestmv->col = bc; michael@0: michael@0: if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) || michael@0: (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3))) michael@0: return INT_MAX; michael@0: michael@0: return besterr; michael@0: } michael@0: michael@0: #undef MVC michael@0: #undef PRE michael@0: #undef DIST michael@0: #undef IFMVCV michael@0: #undef CHECK_BETTER michael@0: #undef SP michael@0: michael@0: #define CHECK_BOUNDS(range) \ michael@0: {\ michael@0: all_in = 1;\ michael@0: all_in &= ((br-range) >= x->mv_row_min);\ michael@0: all_in &= ((br+range) <= x->mv_row_max);\ michael@0: all_in &= ((bc-range) >= x->mv_col_min);\ michael@0: all_in &= ((bc+range) <= x->mv_col_max);\ michael@0: } michael@0: michael@0: #define CHECK_POINT \ michael@0: {\ michael@0: if (this_mv.col < x->mv_col_min) continue;\ michael@0: if (this_mv.col > x->mv_col_max) continue;\ michael@0: if (this_mv.row < x->mv_row_min) continue;\ michael@0: if (this_mv.row > x->mv_row_max) continue;\ michael@0: } michael@0: michael@0: #define CHECK_BETTER \ michael@0: {\ michael@0: if (thissad < bestsad)\ michael@0: {\ michael@0: if (use_mvcost) \ michael@0: thissad += mvsad_err_cost(&this_mv, &fcenter_mv.as_mv, \ michael@0: mvjsadcost, mvsadcost, \ michael@0: sad_per_bit);\ michael@0: if (thissad < bestsad)\ michael@0: {\ michael@0: bestsad = thissad;\ michael@0: best_site = i;\ michael@0: }\ michael@0: }\ michael@0: } michael@0: michael@0: #define get_next_chkpts(list, i, n) \ michael@0: list[0] = ((i) == 0 ? (n) - 1 : (i) - 1); \ michael@0: list[1] = (i); \ michael@0: list[2] = ((i) == (n) - 1 ? 0 : (i) + 1); michael@0: michael@0: #define MAX_PATTERN_SCALES 11 michael@0: #define MAX_PATTERN_CANDIDATES 8 // max number of canddiates per scale michael@0: #define PATTERN_CANDIDATES_REF 3 // number of refinement candidates michael@0: michael@0: // Generic pattern search function that searches over multiple scales. michael@0: // Each scale can have a different number of candidates and shape of michael@0: // candidates as indicated in the num_candidates and candidates arrays michael@0: // passed into this function michael@0: static int vp9_pattern_search(MACROBLOCK *x, michael@0: MV *ref_mv, michael@0: int search_param, michael@0: int sad_per_bit, michael@0: int do_init_search, michael@0: int do_refine, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int use_mvcost, michael@0: const MV *center_mv, MV *best_mv, michael@0: const int num_candidates[MAX_PATTERN_SCALES], michael@0: const MV candidates[MAX_PATTERN_SCALES] michael@0: [MAX_PATTERN_CANDIDATES]) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = { michael@0: 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, michael@0: }; michael@0: int i, j, s, t; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: int br, bc; michael@0: MV this_mv; michael@0: int bestsad = INT_MAX; michael@0: int thissad; michael@0: uint8_t *base_offset; michael@0: uint8_t *this_offset; michael@0: int k = -1; michael@0: int all_in; michael@0: int best_site = -1; michael@0: int_mv fcenter_mv; michael@0: int best_init_s = search_param_to_steps[search_param]; michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->col >> 3; michael@0: michael@0: // adjust ref_mv to make sure it is within MV range michael@0: clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); michael@0: br = ref_mv->row; michael@0: bc = ref_mv->col; michael@0: michael@0: // Work out the start point for the search michael@0: base_offset = (uint8_t *)(xd->plane[0].pre[0].buf); michael@0: this_offset = base_offset + (br * in_what_stride) + bc; michael@0: this_mv.row = br; michael@0: this_mv.col = bc; michael@0: bestsad = vfp->sdf(what, what_stride, this_offset, in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&this_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // Search all possible scales upto the search param around the center point michael@0: // pick the scale of the point that is best as the starting scale of michael@0: // further steps around it. michael@0: if (do_init_search) { michael@0: s = best_init_s; michael@0: best_init_s = -1; michael@0: for (t = 0; t <= s; ++t) { michael@0: best_site = -1; michael@0: CHECK_BOUNDS((1 << t)) michael@0: if (all_in) { michael@0: for (i = 0; i < num_candidates[t]; i++) { michael@0: this_mv.row = br + candidates[t][i].row; michael@0: this_mv.col = bc + candidates[t][i].col; michael@0: this_offset = base_offset + (this_mv.row * in_what_stride) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } else { michael@0: for (i = 0; i < num_candidates[t]; i++) { michael@0: this_mv.row = br + candidates[t][i].row; michael@0: this_mv.col = bc + candidates[t][i].col; michael@0: CHECK_POINT michael@0: this_offset = base_offset + (this_mv.row * in_what_stride) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } michael@0: if (best_site == -1) { michael@0: continue; michael@0: } else { michael@0: best_init_s = t; michael@0: k = best_site; michael@0: } michael@0: } michael@0: if (best_init_s != -1) { michael@0: br += candidates[best_init_s][k].row; michael@0: bc += candidates[best_init_s][k].col; michael@0: } michael@0: } michael@0: michael@0: // If the center point is still the best, just skip this and move to michael@0: // the refinement step. michael@0: if (best_init_s != -1) { michael@0: s = best_init_s; michael@0: best_site = -1; michael@0: do { michael@0: // No need to search all 6 points the 1st time if initial search was used michael@0: if (!do_init_search || s != best_init_s) { michael@0: CHECK_BOUNDS((1 << s)) michael@0: if (all_in) { michael@0: for (i = 0; i < num_candidates[s]; i++) { michael@0: this_mv.row = br + candidates[s][i].row; michael@0: this_mv.col = bc + candidates[s][i].col; michael@0: this_offset = base_offset + (this_mv.row * in_what_stride) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } else { michael@0: for (i = 0; i < num_candidates[s]; i++) { michael@0: this_mv.row = br + candidates[s][i].row; michael@0: this_mv.col = bc + candidates[s][i].col; michael@0: CHECK_POINT michael@0: this_offset = base_offset + (this_mv.row * in_what_stride) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } michael@0: michael@0: if (best_site == -1) { michael@0: continue; michael@0: } else { michael@0: br += candidates[s][best_site].row; michael@0: bc += candidates[s][best_site].col; michael@0: k = best_site; michael@0: } michael@0: } michael@0: michael@0: do { michael@0: int next_chkpts_indices[PATTERN_CANDIDATES_REF]; michael@0: best_site = -1; michael@0: CHECK_BOUNDS((1 << s)) michael@0: michael@0: get_next_chkpts(next_chkpts_indices, k, num_candidates[s]); michael@0: if (all_in) { michael@0: for (i = 0; i < PATTERN_CANDIDATES_REF; i++) { michael@0: this_mv.row = br + candidates[s][next_chkpts_indices[i]].row; michael@0: this_mv.col = bc + candidates[s][next_chkpts_indices[i]].col; michael@0: this_offset = base_offset + (this_mv.row * (in_what_stride)) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } else { michael@0: for (i = 0; i < PATTERN_CANDIDATES_REF; i++) { michael@0: this_mv.row = br + candidates[s][next_chkpts_indices[i]].row; michael@0: this_mv.col = bc + candidates[s][next_chkpts_indices[i]].col; michael@0: CHECK_POINT michael@0: this_offset = base_offset + (this_mv.row * (in_what_stride)) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } michael@0: michael@0: if (best_site != -1) { michael@0: k = next_chkpts_indices[best_site]; michael@0: br += candidates[s][k].row; michael@0: bc += candidates[s][k].col; michael@0: } michael@0: } while (best_site != -1); michael@0: } while (s--); michael@0: } michael@0: michael@0: // Check 4 1-away neighbors if do_refine is true. michael@0: // For most well-designed schemes do_refine will not be necessary. michael@0: if (do_refine) { michael@0: static const MV neighbors[4] = { michael@0: {0, -1}, { -1, 0}, {1, 0}, {0, 1}, michael@0: }; michael@0: for (j = 0; j < 16; j++) { michael@0: best_site = -1; michael@0: CHECK_BOUNDS(1) michael@0: if (all_in) { michael@0: for (i = 0; i < 4; i++) { michael@0: this_mv.row = br + neighbors[i].row; michael@0: this_mv.col = bc + neighbors[i].col; michael@0: this_offset = base_offset + (this_mv.row * (in_what_stride)) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } else { michael@0: for (i = 0; i < 4; i++) { michael@0: this_mv.row = br + neighbors[i].row; michael@0: this_mv.col = bc + neighbors[i].col; michael@0: CHECK_POINT michael@0: this_offset = base_offset + (this_mv.row * (in_what_stride)) + michael@0: this_mv.col; michael@0: thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, michael@0: bestsad); michael@0: CHECK_BETTER michael@0: } michael@0: } michael@0: michael@0: if (best_site == -1) { michael@0: break; michael@0: } else { michael@0: br += neighbors[best_site].row; michael@0: bc += neighbors[best_site].col; michael@0: } michael@0: } michael@0: } michael@0: michael@0: best_mv->row = br; michael@0: best_mv->col = bc; michael@0: michael@0: this_offset = base_offset + (best_mv->row * in_what_stride) + michael@0: best_mv->col; michael@0: this_mv.row = best_mv->row * 8; michael@0: this_mv.col = best_mv->col * 8; michael@0: if (bestsad == INT_MAX) michael@0: return INT_MAX; michael@0: michael@0: return vfp->vf(what, what_stride, this_offset, in_what_stride, michael@0: (unsigned int *)&bestsad) + michael@0: use_mvcost ? mv_err_cost(&this_mv, center_mv, michael@0: x->nmvjointcost, x->mvcost, x->errorperbit) michael@0: : 0; michael@0: } michael@0: michael@0: michael@0: int vp9_hex_search(MACROBLOCK *x, michael@0: MV *ref_mv, michael@0: int search_param, michael@0: int sad_per_bit, michael@0: int do_init_search, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int use_mvcost, michael@0: const MV *center_mv, MV *best_mv) { michael@0: // First scale has 8-closest points, the rest have 6 points in hex shape michael@0: // at increasing scales michael@0: static const int hex_num_candidates[MAX_PATTERN_SCALES] = { michael@0: 8, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6 michael@0: }; michael@0: // Note that the largest candidate step at each scale is 2^scale michael@0: static const MV hex_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = { michael@0: {{-1, -1}, {0, -1}, {1, -1}, {1, 0}, {1, 1}, { 0, 1}, { -1, 1}, {-1, 0}}, michael@0: {{-1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0}}, michael@0: {{-2, -4}, {2, -4}, {4, 0}, {2, 4}, { -2, 4}, { -4, 0}}, michael@0: {{-4, -8}, {4, -8}, {8, 0}, {4, 8}, { -4, 8}, { -8, 0}}, michael@0: {{-8, -16}, {8, -16}, {16, 0}, {8, 16}, { -8, 16}, { -16, 0}}, michael@0: {{-16, -32}, {16, -32}, {32, 0}, {16, 32}, { -16, 32}, { -32, 0}}, michael@0: {{-32, -64}, {32, -64}, {64, 0}, {32, 64}, { -32, 64}, { -64, 0}}, michael@0: {{-64, -128}, {64, -128}, {128, 0}, {64, 128}, { -64, 128}, { -128, 0}}, michael@0: {{-128, -256}, {128, -256}, {256, 0}, {128, 256}, { -128, 256}, { -256, 0}}, michael@0: {{-256, -512}, {256, -512}, {512, 0}, {256, 512}, { -256, 512}, { -512, 0}}, michael@0: {{-512, -1024}, {512, -1024}, {1024, 0}, {512, 1024}, { -512, 1024}, michael@0: { -1024, 0}}, michael@0: }; michael@0: return michael@0: vp9_pattern_search(x, ref_mv, search_param, sad_per_bit, michael@0: do_init_search, 0, vfp, use_mvcost, michael@0: center_mv, best_mv, michael@0: hex_num_candidates, hex_candidates); michael@0: } michael@0: michael@0: int vp9_bigdia_search(MACROBLOCK *x, michael@0: MV *ref_mv, michael@0: int search_param, michael@0: int sad_per_bit, michael@0: int do_init_search, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int use_mvcost, michael@0: const MV *center_mv, michael@0: MV *best_mv) { michael@0: // First scale has 4-closest points, the rest have 8 points in diamond michael@0: // shape at increasing scales michael@0: static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = { michael@0: 4, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, michael@0: }; michael@0: // Note that the largest candidate step at each scale is 2^scale michael@0: static const MV bigdia_candidates[MAX_PATTERN_SCALES] michael@0: [MAX_PATTERN_CANDIDATES] = { michael@0: {{0, -1}, {1, 0}, { 0, 1}, {-1, 0}}, michael@0: {{-1, -1}, {0, -2}, {1, -1}, {2, 0}, {1, 1}, {0, 2}, {-1, 1}, {-2, 0}}, michael@0: {{-2, -2}, {0, -4}, {2, -2}, {4, 0}, {2, 2}, {0, 4}, {-2, 2}, {-4, 0}}, michael@0: {{-4, -4}, {0, -8}, {4, -4}, {8, 0}, {4, 4}, {0, 8}, {-4, 4}, {-8, 0}}, michael@0: {{-8, -8}, {0, -16}, {8, -8}, {16, 0}, {8, 8}, {0, 16}, {-8, 8}, {-16, 0}}, michael@0: {{-16, -16}, {0, -32}, {16, -16}, {32, 0}, {16, 16}, {0, 32}, michael@0: {-16, 16}, {-32, 0}}, michael@0: {{-32, -32}, {0, -64}, {32, -32}, {64, 0}, {32, 32}, {0, 64}, michael@0: {-32, 32}, {-64, 0}}, michael@0: {{-64, -64}, {0, -128}, {64, -64}, {128, 0}, {64, 64}, {0, 128}, michael@0: {-64, 64}, {-128, 0}}, michael@0: {{-128, -128}, {0, -256}, {128, -128}, {256, 0}, {128, 128}, {0, 256}, michael@0: {-128, 128}, {-256, 0}}, michael@0: {{-256, -256}, {0, -512}, {256, -256}, {512, 0}, {256, 256}, {0, 512}, michael@0: {-256, 256}, {-512, 0}}, michael@0: {{-512, -512}, {0, -1024}, {512, -512}, {1024, 0}, {512, 512}, {0, 1024}, michael@0: {-512, 512}, {-1024, 0}}, michael@0: }; michael@0: return vp9_pattern_search(x, ref_mv, search_param, sad_per_bit, michael@0: do_init_search, 0, vfp, use_mvcost, michael@0: center_mv, best_mv, michael@0: bigdia_num_candidates, bigdia_candidates); michael@0: } michael@0: michael@0: int vp9_square_search(MACROBLOCK *x, michael@0: MV *ref_mv, michael@0: int search_param, michael@0: int sad_per_bit, michael@0: int do_init_search, michael@0: const vp9_variance_fn_ptr_t *vfp, michael@0: int use_mvcost, michael@0: const MV *center_mv, michael@0: MV *best_mv) { michael@0: // All scales have 8 closest points in square shape michael@0: static const int square_num_candidates[MAX_PATTERN_SCALES] = { michael@0: 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, michael@0: }; michael@0: // Note that the largest candidate step at each scale is 2^scale michael@0: static const MV square_candidates[MAX_PATTERN_SCALES] michael@0: [MAX_PATTERN_CANDIDATES] = { michael@0: {{-1, -1}, {0, -1}, {1, -1}, {1, 0}, {1, 1}, {0, 1}, {-1, 1}, {-1, 0}}, michael@0: {{-2, -2}, {0, -2}, {2, -2}, {2, 0}, {2, 2}, {0, 2}, {-2, 2}, {-2, 0}}, michael@0: {{-4, -4}, {0, -4}, {4, -4}, {4, 0}, {4, 4}, {0, 4}, {-4, 4}, {-4, 0}}, michael@0: {{-8, -8}, {0, -8}, {8, -8}, {8, 0}, {8, 8}, {0, 8}, {-8, 8}, {-8, 0}}, michael@0: {{-16, -16}, {0, -16}, {16, -16}, {16, 0}, {16, 16}, {0, 16}, michael@0: {-16, 16}, {-16, 0}}, michael@0: {{-32, -32}, {0, -32}, {32, -32}, {32, 0}, {32, 32}, {0, 32}, michael@0: {-32, 32}, {-32, 0}}, michael@0: {{-64, -64}, {0, -64}, {64, -64}, {64, 0}, {64, 64}, {0, 64}, michael@0: {-64, 64}, {-64, 0}}, michael@0: {{-128, -128}, {0, -128}, {128, -128}, {128, 0}, {128, 128}, {0, 128}, michael@0: {-128, 128}, {-128, 0}}, michael@0: {{-256, -256}, {0, -256}, {256, -256}, {256, 0}, {256, 256}, {0, 256}, michael@0: {-256, 256}, {-256, 0}}, michael@0: {{-512, -512}, {0, -512}, {512, -512}, {512, 0}, {512, 512}, {0, 512}, michael@0: {-512, 512}, {-512, 0}}, michael@0: {{-1024, -1024}, {0, -1024}, {1024, -1024}, {1024, 0}, {1024, 1024}, michael@0: {0, 1024}, {-1024, 1024}, {-1024, 0}}, michael@0: }; michael@0: return vp9_pattern_search(x, ref_mv, search_param, sad_per_bit, michael@0: do_init_search, 0, vfp, use_mvcost, michael@0: center_mv, best_mv, michael@0: square_num_candidates, square_candidates); michael@0: }; michael@0: michael@0: #undef CHECK_BOUNDS michael@0: #undef CHECK_POINT michael@0: #undef CHECK_BETTER michael@0: michael@0: int vp9_diamond_search_sad_c(MACROBLOCK *x, michael@0: int_mv *ref_mv, int_mv *best_mv, michael@0: int search_param, int sad_per_bit, int *num00, michael@0: vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost, michael@0: int *mvcost[2], int_mv *center_mv) { michael@0: int i, j, step; michael@0: michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: uint8_t *in_what; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *best_address; michael@0: michael@0: int tot_steps; michael@0: int_mv this_mv; michael@0: michael@0: int bestsad = INT_MAX; michael@0: int best_site = 0; michael@0: int last_site = 0; michael@0: michael@0: int ref_row, ref_col; michael@0: int this_row_offset, this_col_offset; michael@0: search_site *ss; michael@0: michael@0: uint8_t *check_here; michael@0: int thissad; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: clamp_mv(&ref_mv->as_mv, michael@0: x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); michael@0: ref_row = ref_mv->as_mv.row; michael@0: ref_col = ref_mv->as_mv.col; michael@0: *num00 = 0; michael@0: best_mv->as_mv.row = ref_row; michael@0: best_mv->as_mv.col = ref_col; michael@0: michael@0: // Work out the start point for the search michael@0: in_what = (uint8_t *)(xd->plane[0].pre[0].buf + michael@0: (ref_row * (xd->plane[0].pre[0].stride)) + ref_col); michael@0: best_address = in_what; michael@0: michael@0: // Check the starting position michael@0: bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&best_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // search_param determines the length of the initial step and hence the number michael@0: // of iterations michael@0: // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 = michael@0: // (MAX_FIRST_STEP/4) pel... etc. michael@0: ss = &x->ss[search_param * x->searches_per_step]; michael@0: tot_steps = (x->ss_count / x->searches_per_step) - search_param; michael@0: michael@0: i = 1; michael@0: michael@0: for (step = 0; step < tot_steps; step++) { michael@0: for (j = 0; j < x->searches_per_step; j++) { michael@0: // Trap illegal vectors michael@0: this_row_offset = best_mv->as_mv.row + ss[i].mv.row; michael@0: this_col_offset = best_mv->as_mv.col + ss[i].mv.col; michael@0: michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = ss[i].offset + best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_site = i; michael@0: } michael@0: } michael@0: } michael@0: michael@0: i++; michael@0: } michael@0: michael@0: if (best_site != last_site) { michael@0: best_mv->as_mv.row += ss[best_site].mv.row; michael@0: best_mv->as_mv.col += ss[best_site].mv.col; michael@0: best_address += ss[best_site].offset; michael@0: last_site = best_site; michael@0: #if defined(NEW_DIAMOND_SEARCH) michael@0: while (1) { michael@0: this_row_offset = best_mv->as_mv.row + ss[best_site].mv.row; michael@0: this_col_offset = best_mv->as_mv.col + ss[best_site].mv.col; michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = ss[best_site].offset + best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row += ss[best_site].mv.row; michael@0: best_mv->as_mv.col += ss[best_site].mv.col; michael@0: best_address += ss[best_site].offset; michael@0: continue; michael@0: } michael@0: } michael@0: } michael@0: break; michael@0: }; michael@0: #endif michael@0: } else if (best_address == in_what) { michael@0: (*num00)++; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = best_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = best_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad == INT_MAX) michael@0: return INT_MAX; michael@0: michael@0: return fn_ptr->vf(what, what_stride, best_address, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: } michael@0: michael@0: int vp9_diamond_search_sadx4(MACROBLOCK *x, michael@0: int_mv *ref_mv, int_mv *best_mv, int search_param, michael@0: int sad_per_bit, int *num00, michael@0: vp9_variance_fn_ptr_t *fn_ptr, michael@0: int *mvjcost, int *mvcost[2], int_mv *center_mv) { michael@0: int i, j, step; michael@0: michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: uint8_t *in_what; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *best_address; michael@0: michael@0: int tot_steps; michael@0: int_mv this_mv; michael@0: michael@0: unsigned int bestsad = INT_MAX; michael@0: int best_site = 0; michael@0: int last_site = 0; michael@0: michael@0: int ref_row; michael@0: int ref_col; michael@0: int this_row_offset; michael@0: int this_col_offset; michael@0: search_site *ss; michael@0: michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: clamp_mv(&ref_mv->as_mv, michael@0: x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max); michael@0: ref_row = ref_mv->as_mv.row; michael@0: ref_col = ref_mv->as_mv.col; michael@0: *num00 = 0; michael@0: best_mv->as_mv.row = ref_row; michael@0: best_mv->as_mv.col = ref_col; michael@0: michael@0: // Work out the start point for the search michael@0: in_what = (uint8_t *)(xd->plane[0].pre[0].buf + michael@0: (ref_row * (xd->plane[0].pre[0].stride)) + ref_col); michael@0: best_address = in_what; michael@0: michael@0: // Check the starting position michael@0: bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&best_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // search_param determines the length of the initial step and hence the number michael@0: // of iterations. michael@0: // 0 = initial step (MAX_FIRST_STEP) pel michael@0: // 1 = (MAX_FIRST_STEP/2) pel, michael@0: // 2 = (MAX_FIRST_STEP/4) pel... michael@0: ss = &x->ss[search_param * x->searches_per_step]; michael@0: tot_steps = (x->ss_count / x->searches_per_step) - search_param; michael@0: michael@0: i = 1; michael@0: michael@0: for (step = 0; step < tot_steps; step++) { michael@0: int all_in = 1, t; michael@0: michael@0: // All_in is true if every one of the points we are checking are within michael@0: // the bounds of the image. michael@0: all_in &= ((best_mv->as_mv.row + ss[i].mv.row) > x->mv_row_min); michael@0: all_in &= ((best_mv->as_mv.row + ss[i + 1].mv.row) < x->mv_row_max); michael@0: all_in &= ((best_mv->as_mv.col + ss[i + 2].mv.col) > x->mv_col_min); michael@0: all_in &= ((best_mv->as_mv.col + ss[i + 3].mv.col) < x->mv_col_max); michael@0: michael@0: // If all the pixels are within the bounds we don't check whether the michael@0: // search point is valid in this loop, otherwise we check each point michael@0: // for validity.. michael@0: if (all_in) { michael@0: unsigned int sad_array[4]; michael@0: michael@0: for (j = 0; j < x->searches_per_step; j += 4) { michael@0: unsigned char const *block_offset[4]; michael@0: michael@0: for (t = 0; t < 4; t++) michael@0: block_offset[t] = ss[i + t].offset + best_address; michael@0: michael@0: fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, michael@0: sad_array); michael@0: michael@0: for (t = 0; t < 4; t++, i++) { michael@0: if (sad_array[t] < bestsad) { michael@0: this_mv.as_mv.row = best_mv->as_mv.row + ss[i].mv.row; michael@0: this_mv.as_mv.col = best_mv->as_mv.col + ss[i].mv.col; michael@0: sad_array[t] += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (sad_array[t] < bestsad) { michael@0: bestsad = sad_array[t]; michael@0: best_site = i; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: } else { michael@0: for (j = 0; j < x->searches_per_step; j++) { michael@0: // Trap illegal vectors michael@0: this_row_offset = best_mv->as_mv.row + ss[i].mv.row; michael@0: this_col_offset = best_mv->as_mv.col + ss[i].mv.col; michael@0: michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = ss[i].offset + best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_site = i; michael@0: } michael@0: } michael@0: } michael@0: i++; michael@0: } michael@0: } michael@0: if (best_site != last_site) { michael@0: best_mv->as_mv.row += ss[best_site].mv.row; michael@0: best_mv->as_mv.col += ss[best_site].mv.col; michael@0: best_address += ss[best_site].offset; michael@0: last_site = best_site; michael@0: #if defined(NEW_DIAMOND_SEARCH) michael@0: while (1) { michael@0: this_row_offset = best_mv->as_mv.row + ss[best_site].mv.row; michael@0: this_col_offset = best_mv->as_mv.col + ss[best_site].mv.col; michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = ss[best_site].offset + best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row += ss[best_site].mv.row; michael@0: best_mv->as_mv.col += ss[best_site].mv.col; michael@0: best_address += ss[best_site].offset; michael@0: continue; michael@0: } michael@0: } michael@0: } michael@0: break; michael@0: }; michael@0: #endif michael@0: } else if (best_address == in_what) { michael@0: (*num00)++; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = best_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = best_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad == INT_MAX) michael@0: return INT_MAX; michael@0: michael@0: return fn_ptr->vf(what, what_stride, best_address, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: } michael@0: michael@0: /* do_refine: If last step (1-away) of n-step search doesn't pick the center michael@0: point as the best match, we will do a final 1-away diamond michael@0: refining search */ michael@0: michael@0: int vp9_full_pixel_diamond(VP9_COMP *cpi, MACROBLOCK *x, michael@0: int_mv *mvp_full, int step_param, michael@0: int sadpb, int further_steps, michael@0: int do_refine, vp9_variance_fn_ptr_t *fn_ptr, michael@0: int_mv *ref_mv, int_mv *dst_mv) { michael@0: int_mv temp_mv; michael@0: int thissme, n, num00; michael@0: int bestsme = cpi->diamond_search_sad(x, mvp_full, &temp_mv, michael@0: step_param, sadpb, &num00, michael@0: fn_ptr, x->nmvjointcost, michael@0: x->mvcost, ref_mv); michael@0: dst_mv->as_int = temp_mv.as_int; michael@0: michael@0: n = num00; michael@0: num00 = 0; michael@0: michael@0: /* If there won't be more n-step search, check to see if refining search is michael@0: * needed. */ michael@0: if (n > further_steps) michael@0: do_refine = 0; michael@0: michael@0: while (n < further_steps) { michael@0: n++; michael@0: michael@0: if (num00) { michael@0: num00--; michael@0: } else { michael@0: thissme = cpi->diamond_search_sad(x, mvp_full, &temp_mv, michael@0: step_param + n, sadpb, &num00, michael@0: fn_ptr, x->nmvjointcost, x->mvcost, michael@0: ref_mv); michael@0: michael@0: /* check to see if refining search is needed. */ michael@0: if (num00 > (further_steps - n)) michael@0: do_refine = 0; michael@0: michael@0: if (thissme < bestsme) { michael@0: bestsme = thissme; michael@0: dst_mv->as_int = temp_mv.as_int; michael@0: } michael@0: } michael@0: } michael@0: michael@0: /* final 1-away diamond refining search */ michael@0: if (do_refine == 1) { michael@0: int search_range = 8; michael@0: int_mv best_mv; michael@0: best_mv.as_int = dst_mv->as_int; michael@0: thissme = cpi->refining_search_sad(x, &best_mv, sadpb, search_range, michael@0: fn_ptr, x->nmvjointcost, x->mvcost, michael@0: ref_mv); michael@0: michael@0: if (thissme < bestsme) { michael@0: bestsme = thissme; michael@0: dst_mv->as_int = best_mv.as_int; michael@0: } michael@0: } michael@0: return bestsme; michael@0: } michael@0: michael@0: int vp9_full_search_sad_c(MACROBLOCK *x, int_mv *ref_mv, michael@0: int sad_per_bit, int distance, michael@0: vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost, michael@0: int *mvcost[2], michael@0: int_mv *center_mv, int n) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: uint8_t *in_what; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: int mv_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *bestaddress; michael@0: int_mv *best_mv = &x->e_mbd.mi_8x8[0]->bmi[n].as_mv[0]; michael@0: int_mv this_mv; michael@0: int bestsad = INT_MAX; michael@0: int r, c; michael@0: michael@0: uint8_t *check_here; michael@0: int thissad; michael@0: michael@0: int ref_row = ref_mv->as_mv.row; michael@0: int ref_col = ref_mv->as_mv.col; michael@0: michael@0: int row_min = ref_row - distance; michael@0: int row_max = ref_row + distance; michael@0: int col_min = ref_col - distance; michael@0: int col_max = ref_col + distance; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: // Work out the mid point for the search michael@0: in_what = xd->plane[0].pre[0].buf; michael@0: bestaddress = in_what + (ref_row * xd->plane[0].pre[0].stride) + ref_col; michael@0: michael@0: best_mv->as_mv.row = ref_row; michael@0: best_mv->as_mv.col = ref_col; michael@0: michael@0: // Baseline value at the centre michael@0: bestsad = fn_ptr->sdf(what, what_stride, bestaddress, michael@0: in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&best_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // Apply further limits to prevent us looking using vectors that stretch michael@0: // beyond the UMV border michael@0: col_min = MAX(col_min, x->mv_col_min); michael@0: col_max = MIN(col_max, x->mv_col_max); michael@0: row_min = MAX(row_min, x->mv_row_min); michael@0: row_max = MIN(row_max, x->mv_row_max); michael@0: michael@0: for (r = row_min; r < row_max; r++) { michael@0: this_mv.as_mv.row = r; michael@0: check_here = r * mv_stride + in_what + col_min; michael@0: michael@0: for (c = col_min; c < col_max; c++) { michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: michael@0: check_here++; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = best_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = best_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) michael@0: return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: else michael@0: return INT_MAX; michael@0: } michael@0: michael@0: int vp9_full_search_sadx3(MACROBLOCK *x, int_mv *ref_mv, michael@0: int sad_per_bit, int distance, michael@0: vp9_variance_fn_ptr_t *fn_ptr, int *mvjcost, michael@0: int *mvcost[2], int_mv *center_mv, int n) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: uint8_t *in_what; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: int mv_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *bestaddress; michael@0: int_mv *best_mv = &x->e_mbd.mi_8x8[0]->bmi[n].as_mv[0]; michael@0: int_mv this_mv; michael@0: unsigned int bestsad = INT_MAX; michael@0: int r, c; michael@0: michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: michael@0: int ref_row = ref_mv->as_mv.row; michael@0: int ref_col = ref_mv->as_mv.col; michael@0: michael@0: int row_min = ref_row - distance; michael@0: int row_max = ref_row + distance; michael@0: int col_min = ref_col - distance; michael@0: int col_max = ref_col + distance; michael@0: michael@0: unsigned int sad_array[3]; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: // Work out the mid point for the search michael@0: in_what = xd->plane[0].pre[0].buf; michael@0: bestaddress = in_what + (ref_row * xd->plane[0].pre[0].stride) + ref_col; michael@0: michael@0: best_mv->as_mv.row = ref_row; michael@0: best_mv->as_mv.col = ref_col; michael@0: michael@0: // Baseline value at the centre michael@0: bestsad = fn_ptr->sdf(what, what_stride, michael@0: bestaddress, in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&best_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // Apply further limits to prevent us looking using vectors that stretch michael@0: // beyond the UMV border michael@0: col_min = MAX(col_min, x->mv_col_min); michael@0: col_max = MIN(col_max, x->mv_col_max); michael@0: row_min = MAX(row_min, x->mv_row_min); michael@0: row_max = MIN(row_max, x->mv_row_max); michael@0: michael@0: for (r = row_min; r < row_max; r++) { michael@0: this_mv.as_mv.row = r; michael@0: check_here = r * mv_stride + in_what + col_min; michael@0: c = col_min; michael@0: michael@0: while ((c + 2) < col_max) { michael@0: int i; michael@0: michael@0: fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array); michael@0: michael@0: for (i = 0; i < 3; i++) { michael@0: thissad = sad_array[i]; michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: } michael@0: michael@0: check_here++; michael@0: c++; michael@0: } michael@0: } michael@0: michael@0: while (c < col_max) { michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: } michael@0: michael@0: check_here++; michael@0: c++; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = best_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = best_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) michael@0: return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: else michael@0: return INT_MAX; michael@0: } michael@0: michael@0: int vp9_full_search_sadx8(MACROBLOCK *x, int_mv *ref_mv, michael@0: int sad_per_bit, int distance, michael@0: vp9_variance_fn_ptr_t *fn_ptr, michael@0: int *mvjcost, int *mvcost[2], michael@0: int_mv *center_mv, int n) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: int what_stride = x->plane[0].src.stride; michael@0: uint8_t *in_what; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: int mv_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *bestaddress; michael@0: int_mv *best_mv = &x->e_mbd.mi_8x8[0]->bmi[n].as_mv[0]; michael@0: int_mv this_mv; michael@0: unsigned int bestsad = INT_MAX; michael@0: int r, c; michael@0: michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: michael@0: int ref_row = ref_mv->as_mv.row; michael@0: int ref_col = ref_mv->as_mv.col; michael@0: michael@0: int row_min = ref_row - distance; michael@0: int row_max = ref_row + distance; michael@0: int col_min = ref_col - distance; michael@0: int col_max = ref_col + distance; michael@0: michael@0: DECLARE_ALIGNED_ARRAY(16, uint32_t, sad_array8, 8); michael@0: unsigned int sad_array[3]; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: // Work out the mid point for the search michael@0: in_what = xd->plane[0].pre[0].buf; michael@0: bestaddress = in_what + (ref_row * xd->plane[0].pre[0].stride) + ref_col; michael@0: michael@0: best_mv->as_mv.row = ref_row; michael@0: best_mv->as_mv.col = ref_col; michael@0: michael@0: // Baseline value at the centre michael@0: bestsad = fn_ptr->sdf(what, what_stride, michael@0: bestaddress, in_what_stride, 0x7fffffff) michael@0: + mvsad_err_cost(&best_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: // Apply further limits to prevent us looking using vectors that stretch michael@0: // beyond the UMV border michael@0: col_min = MAX(col_min, x->mv_col_min); michael@0: col_max = MIN(col_max, x->mv_col_max); michael@0: row_min = MAX(row_min, x->mv_row_min); michael@0: row_max = MIN(row_max, x->mv_row_max); michael@0: michael@0: for (r = row_min; r < row_max; r++) { michael@0: this_mv.as_mv.row = r; michael@0: check_here = r * mv_stride + in_what + col_min; michael@0: c = col_min; michael@0: michael@0: while ((c + 7) < col_max) { michael@0: int i; michael@0: michael@0: fn_ptr->sdx8f(what, what_stride, check_here, in_what_stride, sad_array8); michael@0: michael@0: for (i = 0; i < 8; i++) { michael@0: thissad = (unsigned int)sad_array8[i]; michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: } michael@0: michael@0: check_here++; michael@0: c++; michael@0: } michael@0: } michael@0: michael@0: while ((c + 2) < col_max && fn_ptr->sdx3f != NULL) { michael@0: int i; michael@0: michael@0: fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array); michael@0: michael@0: for (i = 0; i < 3; i++) { michael@0: thissad = sad_array[i]; michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: } michael@0: michael@0: check_here++; michael@0: c++; michael@0: } michael@0: } michael@0: michael@0: while (c < col_max) { michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.col = c; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, sad_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_mv->as_mv.row = r; michael@0: best_mv->as_mv.col = c; michael@0: bestaddress = check_here; michael@0: } michael@0: } michael@0: michael@0: check_here++; michael@0: c++; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = best_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = best_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) michael@0: return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: else michael@0: return INT_MAX; michael@0: } michael@0: int vp9_refining_search_sad_c(MACROBLOCK *x, michael@0: int_mv *ref_mv, int error_per_bit, michael@0: int search_range, vp9_variance_fn_ptr_t *fn_ptr, michael@0: int *mvjcost, int *mvcost[2], int_mv *center_mv) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}}; michael@0: int i, j; michael@0: int this_row_offset, this_col_offset; michael@0: michael@0: int what_stride = x->plane[0].src.stride; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: uint8_t *best_address = xd->plane[0].pre[0].buf + michael@0: (ref_mv->as_mv.row * xd->plane[0].pre[0].stride) + michael@0: ref_mv->as_mv.col; michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: int_mv this_mv; michael@0: unsigned int bestsad = INT_MAX; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: bestsad = fn_ptr->sdf(what, what_stride, best_address, michael@0: in_what_stride, 0x7fffffff) + michael@0: mvsad_err_cost(&ref_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: for (i = 0; i < search_range; i++) { michael@0: int best_site = -1; michael@0: michael@0: for (j = 0; j < 4; j++) { michael@0: this_row_offset = ref_mv->as_mv.row + neighbors[j].row; michael@0: this_col_offset = ref_mv->as_mv.col + neighbors[j].col; michael@0: michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = (neighbors[j].row) * in_what_stride + neighbors[j].col + michael@0: best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_site = j; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: michael@0: if (best_site == -1) { michael@0: break; michael@0: } else { michael@0: ref_mv->as_mv.row += neighbors[best_site].row; michael@0: ref_mv->as_mv.col += neighbors[best_site].col; michael@0: best_address += (neighbors[best_site].row) * in_what_stride + michael@0: neighbors[best_site].col; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = ref_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = ref_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) michael@0: return fn_ptr->vf(what, what_stride, best_address, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: else michael@0: return INT_MAX; michael@0: } michael@0: michael@0: int vp9_refining_search_sadx4(MACROBLOCK *x, michael@0: int_mv *ref_mv, int error_per_bit, michael@0: int search_range, vp9_variance_fn_ptr_t *fn_ptr, michael@0: int *mvjcost, int *mvcost[2], int_mv *center_mv) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}}; michael@0: int i, j; michael@0: int this_row_offset, this_col_offset; michael@0: michael@0: int what_stride = x->plane[0].src.stride; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: uint8_t *best_address = xd->plane[0].pre[0].buf + michael@0: (ref_mv->as_mv.row * xd->plane[0].pre[0].stride) + michael@0: ref_mv->as_mv.col; michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: int_mv this_mv; michael@0: unsigned int bestsad = INT_MAX; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: bestsad = fn_ptr->sdf(what, what_stride, best_address, michael@0: in_what_stride, 0x7fffffff) + michael@0: mvsad_err_cost(&ref_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: for (i = 0; i < search_range; i++) { michael@0: int best_site = -1; michael@0: int all_in = ((ref_mv->as_mv.row - 1) > x->mv_row_min) & michael@0: ((ref_mv->as_mv.row + 1) < x->mv_row_max) & michael@0: ((ref_mv->as_mv.col - 1) > x->mv_col_min) & michael@0: ((ref_mv->as_mv.col + 1) < x->mv_col_max); michael@0: michael@0: if (all_in) { michael@0: unsigned int sad_array[4]; michael@0: unsigned char const *block_offset[4]; michael@0: block_offset[0] = best_address - in_what_stride; michael@0: block_offset[1] = best_address - 1; michael@0: block_offset[2] = best_address + 1; michael@0: block_offset[3] = best_address + in_what_stride; michael@0: michael@0: fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, michael@0: sad_array); michael@0: michael@0: for (j = 0; j < 4; j++) { michael@0: if (sad_array[j] < bestsad) { michael@0: this_mv.as_mv.row = ref_mv->as_mv.row + neighbors[j].row; michael@0: this_mv.as_mv.col = ref_mv->as_mv.col + neighbors[j].col; michael@0: sad_array[j] += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: if (sad_array[j] < bestsad) { michael@0: bestsad = sad_array[j]; michael@0: best_site = j; michael@0: } michael@0: } michael@0: } michael@0: } else { michael@0: for (j = 0; j < 4; j++) { michael@0: this_row_offset = ref_mv->as_mv.row + neighbors[j].row; michael@0: this_col_offset = ref_mv->as_mv.col + neighbors[j].col; michael@0: michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = (neighbors[j].row) * in_what_stride + neighbors[j].col + michael@0: best_address; michael@0: thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, michael@0: bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_site = j; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: } michael@0: michael@0: if (best_site == -1) { michael@0: break; michael@0: } else { michael@0: ref_mv->as_mv.row += neighbors[best_site].row; michael@0: ref_mv->as_mv.col += neighbors[best_site].col; michael@0: best_address += (neighbors[best_site].row) * in_what_stride + michael@0: neighbors[best_site].col; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = ref_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = ref_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) michael@0: return fn_ptr->vf(what, what_stride, best_address, in_what_stride, michael@0: (unsigned int *)(&thissad)) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: else michael@0: return INT_MAX; michael@0: } michael@0: michael@0: /* This function is called when we do joint motion search in comp_inter_inter michael@0: * mode. michael@0: */ michael@0: int vp9_refining_search_8p_c(MACROBLOCK *x, michael@0: int_mv *ref_mv, int error_per_bit, michael@0: int search_range, vp9_variance_fn_ptr_t *fn_ptr, michael@0: int *mvjcost, int *mvcost[2], int_mv *center_mv, michael@0: const uint8_t *second_pred, int w, int h) { michael@0: const MACROBLOCKD* const xd = &x->e_mbd; michael@0: MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}, michael@0: {-1, -1}, {1, -1}, {-1, 1}, {1, 1}}; michael@0: int i, j; michael@0: int this_row_offset, this_col_offset; michael@0: michael@0: int what_stride = x->plane[0].src.stride; michael@0: int in_what_stride = xd->plane[0].pre[0].stride; michael@0: uint8_t *what = x->plane[0].src.buf; michael@0: uint8_t *best_address = xd->plane[0].pre[0].buf + michael@0: (ref_mv->as_mv.row * xd->plane[0].pre[0].stride) + michael@0: ref_mv->as_mv.col; michael@0: uint8_t *check_here; michael@0: unsigned int thissad; michael@0: int_mv this_mv; michael@0: unsigned int bestsad = INT_MAX; michael@0: int_mv fcenter_mv; michael@0: michael@0: int *mvjsadcost = x->nmvjointsadcost; michael@0: int *mvsadcost[2] = {x->nmvsadcost[0], x->nmvsadcost[1]}; michael@0: michael@0: fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3; michael@0: fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3; michael@0: michael@0: /* Get compound pred by averaging two pred blocks. */ michael@0: bestsad = fn_ptr->sdaf(what, what_stride, best_address, in_what_stride, michael@0: second_pred, 0x7fffffff) + michael@0: mvsad_err_cost(&ref_mv->as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: michael@0: for (i = 0; i < search_range; i++) { michael@0: int best_site = -1; michael@0: michael@0: for (j = 0; j < 8; j++) { michael@0: this_row_offset = ref_mv->as_mv.row + neighbors[j].row; michael@0: this_col_offset = ref_mv->as_mv.col + neighbors[j].col; michael@0: michael@0: if ((this_col_offset > x->mv_col_min) && michael@0: (this_col_offset < x->mv_col_max) && michael@0: (this_row_offset > x->mv_row_min) && michael@0: (this_row_offset < x->mv_row_max)) { michael@0: check_here = (neighbors[j].row) * in_what_stride + neighbors[j].col + michael@0: best_address; michael@0: michael@0: /* Get compound block and use it to calculate SAD. */ michael@0: thissad = fn_ptr->sdaf(what, what_stride, check_here, in_what_stride, michael@0: second_pred, bestsad); michael@0: michael@0: if (thissad < bestsad) { michael@0: this_mv.as_mv.row = this_row_offset; michael@0: this_mv.as_mv.col = this_col_offset; michael@0: thissad += mvsad_err_cost(&this_mv.as_mv, &fcenter_mv.as_mv, michael@0: mvjsadcost, mvsadcost, error_per_bit); michael@0: if (thissad < bestsad) { michael@0: bestsad = thissad; michael@0: best_site = j; michael@0: } michael@0: } michael@0: } michael@0: } michael@0: michael@0: if (best_site == -1) { michael@0: break; michael@0: } else { michael@0: ref_mv->as_mv.row += neighbors[best_site].row; michael@0: ref_mv->as_mv.col += neighbors[best_site].col; michael@0: best_address += (neighbors[best_site].row) * in_what_stride + michael@0: neighbors[best_site].col; michael@0: } michael@0: } michael@0: michael@0: this_mv.as_mv.row = ref_mv->as_mv.row * 8; michael@0: this_mv.as_mv.col = ref_mv->as_mv.col * 8; michael@0: michael@0: if (bestsad < INT_MAX) { michael@0: // FIXME(rbultje, yunqing): add full-pixel averaging variance functions michael@0: // so we don't have to use the subpixel with xoff=0,yoff=0 here. michael@0: return fn_ptr->svaf(best_address, in_what_stride, 0, 0, what, what_stride, michael@0: (unsigned int *)(&thissad), second_pred) + michael@0: mv_err_cost(&this_mv.as_mv, ¢er_mv->as_mv, michael@0: mvjcost, mvcost, x->errorperbit); michael@0: } else { michael@0: return INT_MAX; michael@0: } michael@0: }