1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/vp9_mbgraph.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,432 @@
1.4 +/*
1.5 + * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include <limits.h>
1.15 +
1.16 +#include "vpx_mem/vpx_mem.h"
1.17 +#include "vp9/encoder/vp9_encodeintra.h"
1.18 +#include "vp9/encoder/vp9_rdopt.h"
1.19 +#include "vp9/encoder/vp9_segmentation.h"
1.20 +#include "vp9/encoder/vp9_mcomp.h"
1.21 +#include "vp9/common/vp9_blockd.h"
1.22 +#include "vp9/common/vp9_reconinter.h"
1.23 +#include "vp9/common/vp9_reconintra.h"
1.24 +#include "vp9/common/vp9_systemdependent.h"
1.25 +
1.26 +
1.27 +
1.28 +static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
1.29 + int_mv *ref_mv,
1.30 + int_mv *dst_mv,
1.31 + int mb_row,
1.32 + int mb_col) {
1.33 + MACROBLOCK *const x = &cpi->mb;
1.34 + MACROBLOCKD *const xd = &x->e_mbd;
1.35 + vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
1.36 + unsigned int best_err;
1.37 +
1.38 + const int tmp_col_min = x->mv_col_min;
1.39 + const int tmp_col_max = x->mv_col_max;
1.40 + const int tmp_row_min = x->mv_row_min;
1.41 + const int tmp_row_max = x->mv_row_max;
1.42 + int_mv ref_full;
1.43 +
1.44 + // Further step/diamond searches as necessary
1.45 + int step_param = cpi->sf.reduce_first_step_size +
1.46 + (cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2);
1.47 + step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
1.48 +
1.49 + vp9_clamp_mv_min_max(x, &ref_mv->as_mv);
1.50 +
1.51 + ref_full.as_mv.col = ref_mv->as_mv.col >> 3;
1.52 + ref_full.as_mv.row = ref_mv->as_mv.row >> 3;
1.53 +
1.54 + /*cpi->sf.search_method == HEX*/
1.55 + best_err = vp9_hex_search(x, &ref_full.as_mv, step_param, x->errorperbit,
1.56 + 0, &v_fn_ptr,
1.57 + 0, &ref_mv->as_mv, &dst_mv->as_mv);
1.58 +
1.59 + // Try sub-pixel MC
1.60 + // if (bestsme > error_thresh && bestsme < INT_MAX)
1.61 + {
1.62 + int distortion;
1.63 + unsigned int sse;
1.64 + best_err = cpi->find_fractional_mv_step(
1.65 + x,
1.66 + &dst_mv->as_mv, &ref_mv->as_mv,
1.67 + cpi->common.allow_high_precision_mv,
1.68 + x->errorperbit, &v_fn_ptr,
1.69 + 0, cpi->sf.subpel_iters_per_step, NULL, NULL,
1.70 + & distortion, &sse);
1.71 + }
1.72 +
1.73 + vp9_set_mbmode_and_mvs(x, NEWMV, dst_mv);
1.74 + vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
1.75 + best_err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
1.76 + xd->plane[0].dst.buf, xd->plane[0].dst.stride,
1.77 + INT_MAX);
1.78 +
1.79 + /* restore UMV window */
1.80 + x->mv_col_min = tmp_col_min;
1.81 + x->mv_col_max = tmp_col_max;
1.82 + x->mv_row_min = tmp_row_min;
1.83 + x->mv_row_max = tmp_row_max;
1.84 +
1.85 + return best_err;
1.86 +}
1.87 +
1.88 +static int do_16x16_motion_search(VP9_COMP *cpi, int_mv *ref_mv, int_mv *dst_mv,
1.89 + int mb_row, int mb_col) {
1.90 + MACROBLOCK *const x = &cpi->mb;
1.91 + MACROBLOCKD *const xd = &x->e_mbd;
1.92 + unsigned int err, tmp_err;
1.93 + int_mv tmp_mv;
1.94 +
1.95 + // Try zero MV first
1.96 + // FIXME should really use something like near/nearest MV and/or MV prediction
1.97 + err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
1.98 + xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
1.99 + INT_MAX);
1.100 + dst_mv->as_int = 0;
1.101 +
1.102 + // Test last reference frame using the previous best mv as the
1.103 + // starting point (best reference) for the search
1.104 + tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
1.105 + if (tmp_err < err) {
1.106 + err = tmp_err;
1.107 + dst_mv->as_int = tmp_mv.as_int;
1.108 + }
1.109 +
1.110 + // If the current best reference mv is not centered on 0,0 then do a 0,0
1.111 + // based search as well.
1.112 + if (ref_mv->as_int) {
1.113 + unsigned int tmp_err;
1.114 + int_mv zero_ref_mv, tmp_mv;
1.115 +
1.116 + zero_ref_mv.as_int = 0;
1.117 + tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
1.118 + mb_row, mb_col);
1.119 + if (tmp_err < err) {
1.120 + dst_mv->as_int = tmp_mv.as_int;
1.121 + err = tmp_err;
1.122 + }
1.123 + }
1.124 +
1.125 + return err;
1.126 +}
1.127 +
1.128 +static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
1.129 + MACROBLOCK *const x = &cpi->mb;
1.130 + MACROBLOCKD *const xd = &x->e_mbd;
1.131 + unsigned int err;
1.132 +
1.133 + // Try zero MV first
1.134 + // FIXME should really use something like near/nearest MV and/or MV prediction
1.135 + err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
1.136 + xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
1.137 + INT_MAX);
1.138 +
1.139 + dst_mv->as_int = 0;
1.140 +
1.141 + return err;
1.142 +}
1.143 +static int find_best_16x16_intra(VP9_COMP *cpi,
1.144 + int mb_y_offset,
1.145 + MB_PREDICTION_MODE *pbest_mode) {
1.146 + MACROBLOCK *const x = &cpi->mb;
1.147 + MACROBLOCKD *const xd = &x->e_mbd;
1.148 + MB_PREDICTION_MODE best_mode = -1, mode;
1.149 + unsigned int best_err = INT_MAX;
1.150 +
1.151 + // calculate SATD for each intra prediction mode;
1.152 + // we're intentionally not doing 4x4, we just want a rough estimate
1.153 + for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1.154 + unsigned int err;
1.155 +
1.156 + xd->mi_8x8[0]->mbmi.mode = mode;
1.157 + vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
1.158 + x->plane[0].src.buf, x->plane[0].src.stride,
1.159 + xd->plane[0].dst.buf, xd->plane[0].dst.stride);
1.160 + err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
1.161 + xd->plane[0].dst.buf, xd->plane[0].dst.stride, best_err);
1.162 +
1.163 + // find best
1.164 + if (err < best_err) {
1.165 + best_err = err;
1.166 + best_mode = mode;
1.167 + }
1.168 + }
1.169 +
1.170 + if (pbest_mode)
1.171 + *pbest_mode = best_mode;
1.172 +
1.173 + return best_err;
1.174 +}
1.175 +
1.176 +static void update_mbgraph_mb_stats
1.177 +(
1.178 + VP9_COMP *cpi,
1.179 + MBGRAPH_MB_STATS *stats,
1.180 + YV12_BUFFER_CONFIG *buf,
1.181 + int mb_y_offset,
1.182 + YV12_BUFFER_CONFIG *golden_ref,
1.183 + int_mv *prev_golden_ref_mv,
1.184 + int gld_y_offset,
1.185 + YV12_BUFFER_CONFIG *alt_ref,
1.186 + int_mv *prev_alt_ref_mv,
1.187 + int arf_y_offset,
1.188 + int mb_row,
1.189 + int mb_col
1.190 +) {
1.191 + MACROBLOCK *const x = &cpi->mb;
1.192 + MACROBLOCKD *const xd = &x->e_mbd;
1.193 + int intra_error;
1.194 + VP9_COMMON *cm = &cpi->common;
1.195 +
1.196 + // FIXME in practice we're completely ignoring chroma here
1.197 + x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
1.198 + x->plane[0].src.stride = buf->y_stride;
1.199 +
1.200 + xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
1.201 + xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
1.202 +
1.203 + // do intra 16x16 prediction
1.204 + intra_error = find_best_16x16_intra(cpi, mb_y_offset,
1.205 + &stats->ref[INTRA_FRAME].m.mode);
1.206 + if (intra_error <= 0)
1.207 + intra_error = 1;
1.208 + stats->ref[INTRA_FRAME].err = intra_error;
1.209 +
1.210 + // Golden frame MV search, if it exists and is different than last frame
1.211 + if (golden_ref) {
1.212 + int g_motion_error;
1.213 + xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
1.214 + xd->plane[0].pre[0].stride = golden_ref->y_stride;
1.215 + g_motion_error = do_16x16_motion_search(cpi,
1.216 + prev_golden_ref_mv,
1.217 + &stats->ref[GOLDEN_FRAME].m.mv,
1.218 + mb_row, mb_col);
1.219 + stats->ref[GOLDEN_FRAME].err = g_motion_error;
1.220 + } else {
1.221 + stats->ref[GOLDEN_FRAME].err = INT_MAX;
1.222 + stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
1.223 + }
1.224 +
1.225 + // Do an Alt-ref frame MV search, if it exists and is different than
1.226 + // last/golden frame.
1.227 + if (alt_ref) {
1.228 + int a_motion_error;
1.229 + xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
1.230 + xd->plane[0].pre[0].stride = alt_ref->y_stride;
1.231 + a_motion_error = do_16x16_zerozero_search(cpi,
1.232 + &stats->ref[ALTREF_FRAME].m.mv);
1.233 +
1.234 + stats->ref[ALTREF_FRAME].err = a_motion_error;
1.235 + } else {
1.236 + stats->ref[ALTREF_FRAME].err = INT_MAX;
1.237 + stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
1.238 + }
1.239 +}
1.240 +
1.241 +static void update_mbgraph_frame_stats(VP9_COMP *cpi,
1.242 + MBGRAPH_FRAME_STATS *stats,
1.243 + YV12_BUFFER_CONFIG *buf,
1.244 + YV12_BUFFER_CONFIG *golden_ref,
1.245 + YV12_BUFFER_CONFIG *alt_ref) {
1.246 + MACROBLOCK *const x = &cpi->mb;
1.247 + MACROBLOCKD *const xd = &x->e_mbd;
1.248 + VP9_COMMON *const cm = &cpi->common;
1.249 +
1.250 + int mb_col, mb_row, offset = 0;
1.251 + int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
1.252 + int_mv arf_top_mv, gld_top_mv;
1.253 + MODE_INFO mi_local = { { 0 } };
1.254 +
1.255 + // Set up limit values for motion vectors to prevent them extending outside
1.256 + // the UMV borders.
1.257 + arf_top_mv.as_int = 0;
1.258 + gld_top_mv.as_int = 0;
1.259 + x->mv_row_min = -BORDER_MV_PIXELS_B16;
1.260 + x->mv_row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
1.261 + xd->up_available = 0;
1.262 + xd->plane[0].dst.stride = buf->y_stride;
1.263 + xd->plane[0].pre[0].stride = buf->y_stride;
1.264 + xd->plane[1].dst.stride = buf->uv_stride;
1.265 + xd->mi_8x8[0] = &mi_local;
1.266 + mi_local.mbmi.sb_type = BLOCK_16X16;
1.267 + mi_local.mbmi.ref_frame[0] = LAST_FRAME;
1.268 + mi_local.mbmi.ref_frame[1] = NONE;
1.269 +
1.270 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
1.271 + int_mv arf_left_mv, gld_left_mv;
1.272 + int mb_y_in_offset = mb_y_offset;
1.273 + int arf_y_in_offset = arf_y_offset;
1.274 + int gld_y_in_offset = gld_y_offset;
1.275 +
1.276 + // Set up limit values for motion vectors to prevent them extending outside
1.277 + // the UMV borders.
1.278 + arf_left_mv.as_int = arf_top_mv.as_int;
1.279 + gld_left_mv.as_int = gld_top_mv.as_int;
1.280 + x->mv_col_min = -BORDER_MV_PIXELS_B16;
1.281 + x->mv_col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
1.282 + xd->left_available = 0;
1.283 +
1.284 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
1.285 + MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
1.286 +
1.287 + update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
1.288 + golden_ref, &gld_left_mv, gld_y_in_offset,
1.289 + alt_ref, &arf_left_mv, arf_y_in_offset,
1.290 + mb_row, mb_col);
1.291 + arf_left_mv.as_int = mb_stats->ref[ALTREF_FRAME].m.mv.as_int;
1.292 + gld_left_mv.as_int = mb_stats->ref[GOLDEN_FRAME].m.mv.as_int;
1.293 + if (mb_col == 0) {
1.294 + arf_top_mv.as_int = arf_left_mv.as_int;
1.295 + gld_top_mv.as_int = gld_left_mv.as_int;
1.296 + }
1.297 + xd->left_available = 1;
1.298 + mb_y_in_offset += 16;
1.299 + gld_y_in_offset += 16;
1.300 + arf_y_in_offset += 16;
1.301 + x->mv_col_min -= 16;
1.302 + x->mv_col_max -= 16;
1.303 + }
1.304 + xd->up_available = 1;
1.305 + mb_y_offset += buf->y_stride * 16;
1.306 + gld_y_offset += golden_ref->y_stride * 16;
1.307 + if (alt_ref)
1.308 + arf_y_offset += alt_ref->y_stride * 16;
1.309 + x->mv_row_min -= 16;
1.310 + x->mv_row_max -= 16;
1.311 + offset += cm->mb_cols;
1.312 + }
1.313 +}
1.314 +
1.315 +// void separate_arf_mbs_byzz
1.316 +static void separate_arf_mbs(VP9_COMP *cpi) {
1.317 + VP9_COMMON *const cm = &cpi->common;
1.318 + int mb_col, mb_row, offset, i;
1.319 + int mi_row, mi_col;
1.320 + int ncnt[4] = { 0 };
1.321 + int n_frames = cpi->mbgraph_n_frames;
1.322 +
1.323 + int *arf_not_zz;
1.324 +
1.325 + CHECK_MEM_ERROR(cm, arf_not_zz,
1.326 + vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
1.327 + 1));
1.328 +
1.329 + // We are not interested in results beyond the alt ref itself.
1.330 + if (n_frames > cpi->frames_till_gf_update_due)
1.331 + n_frames = cpi->frames_till_gf_update_due;
1.332 +
1.333 + // defer cost to reference frames
1.334 + for (i = n_frames - 1; i >= 0; i--) {
1.335 + MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
1.336 +
1.337 + for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
1.338 + offset += cm->mb_cols, mb_row++) {
1.339 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
1.340 + MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
1.341 +
1.342 + int altref_err = mb_stats->ref[ALTREF_FRAME].err;
1.343 + int intra_err = mb_stats->ref[INTRA_FRAME ].err;
1.344 + int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
1.345 +
1.346 + // Test for altref vs intra and gf and that its mv was 0,0.
1.347 + if (altref_err > 1000 ||
1.348 + altref_err > intra_err ||
1.349 + altref_err > golden_err) {
1.350 + arf_not_zz[offset + mb_col]++;
1.351 + }
1.352 + }
1.353 + }
1.354 + }
1.355 +
1.356 + // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
1.357 + // of bound access in segmentation_map
1.358 + for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
1.359 + for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
1.360 + // If any of the blocks in the sequence failed then the MB
1.361 + // goes in segment 0
1.362 + if (arf_not_zz[mi_row/2*cm->mb_cols + mi_col/2]) {
1.363 + ncnt[0]++;
1.364 + cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
1.365 + } else {
1.366 + cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
1.367 + ncnt[1]++;
1.368 + }
1.369 + }
1.370 + }
1.371 +
1.372 + // Only bother with segmentation if over 10% of the MBs in static segment
1.373 + // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
1.374 + if (1) {
1.375 + // Note % of blocks that are marked as static
1.376 + if (cm->MBs)
1.377 + cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
1.378 +
1.379 + // This error case should not be reachable as this function should
1.380 + // never be called with the common data structure uninitialized.
1.381 + else
1.382 + cpi->static_mb_pct = 0;
1.383 +
1.384 + cpi->seg0_cnt = ncnt[0];
1.385 + vp9_enable_segmentation((VP9_PTR)cpi);
1.386 + } else {
1.387 + cpi->static_mb_pct = 0;
1.388 + vp9_disable_segmentation((VP9_PTR)cpi);
1.389 + }
1.390 +
1.391 + // Free localy allocated storage
1.392 + vpx_free(arf_not_zz);
1.393 +}
1.394 +
1.395 +void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
1.396 + VP9_COMMON *const cm = &cpi->common;
1.397 + int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
1.398 + YV12_BUFFER_CONFIG *golden_ref =
1.399 + &cm->yv12_fb[cm->ref_frame_map[cpi->gld_fb_idx]];
1.400 +
1.401 + // we need to look ahead beyond where the ARF transitions into
1.402 + // being a GF - so exit if we don't look ahead beyond that
1.403 + if (n_frames <= cpi->frames_till_gf_update_due)
1.404 + return;
1.405 + if (n_frames > (int)cpi->frames_till_alt_ref_frame)
1.406 + n_frames = cpi->frames_till_alt_ref_frame;
1.407 + if (n_frames > MAX_LAG_BUFFERS)
1.408 + n_frames = MAX_LAG_BUFFERS;
1.409 +
1.410 + cpi->mbgraph_n_frames = n_frames;
1.411 + for (i = 0; i < n_frames; i++) {
1.412 + MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
1.413 + vpx_memset(frame_stats->mb_stats, 0,
1.414 + cm->mb_rows * cm->mb_cols *
1.415 + sizeof(*cpi->mbgraph_stats[i].mb_stats));
1.416 + }
1.417 +
1.418 + // do motion search to find contribution of each reference to data
1.419 + // later on in this GF group
1.420 + // FIXME really, the GF/last MC search should be done forward, and
1.421 + // the ARF MC search backwards, to get optimal results for MV caching
1.422 + for (i = 0; i < n_frames; i++) {
1.423 + MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
1.424 + struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
1.425 +
1.426 + assert(q_cur != NULL);
1.427 +
1.428 + update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
1.429 + golden_ref, cpi->Source);
1.430 + }
1.431 +
1.432 + vp9_clear_system_state(); // __asm emms;
1.433 +
1.434 + separate_arf_mbs(cpi);
1.435 +}
