1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/common/mfqe.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,385 @@
1.4 +/*
1.5 + * Copyright (c) 2012 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 +
1.15 +/* MFQE: Multiframe Quality Enhancement
1.16 + * In rate limited situations keyframes may cause significant visual artifacts
1.17 + * commonly referred to as "popping." This file implements a postproccesing
1.18 + * algorithm which blends data from the preceeding frame when there is no
1.19 + * motion and the q from the previous frame is lower which indicates that it is
1.20 + * higher quality.
1.21 + */
1.22 +
1.23 +#include "postproc.h"
1.24 +#include "variance.h"
1.25 +#include "vpx_mem/vpx_mem.h"
1.26 +#include "vp8_rtcd.h"
1.27 +#include "vpx_scale/yv12config.h"
1.28 +
1.29 +#include <limits.h>
1.30 +#include <stdlib.h>
1.31 +
1.32 +static void filter_by_weight(unsigned char *src, int src_stride,
1.33 + unsigned char *dst, int dst_stride,
1.34 + int block_size, int src_weight)
1.35 +{
1.36 + int dst_weight = (1 << MFQE_PRECISION) - src_weight;
1.37 + int rounding_bit = 1 << (MFQE_PRECISION - 1);
1.38 + int r, c;
1.39 +
1.40 + for (r = 0; r < block_size; r++)
1.41 + {
1.42 + for (c = 0; c < block_size; c++)
1.43 + {
1.44 + dst[c] = (src[c] * src_weight +
1.45 + dst[c] * dst_weight +
1.46 + rounding_bit) >> MFQE_PRECISION;
1.47 + }
1.48 + src += src_stride;
1.49 + dst += dst_stride;
1.50 + }
1.51 +}
1.52 +
1.53 +void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride,
1.54 + unsigned char *dst, int dst_stride,
1.55 + int src_weight)
1.56 +{
1.57 + filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight);
1.58 +}
1.59 +
1.60 +void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride,
1.61 + unsigned char *dst, int dst_stride,
1.62 + int src_weight)
1.63 +{
1.64 + filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight);
1.65 +}
1.66 +
1.67 +void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride,
1.68 + unsigned char *dst, int dst_stride,
1.69 + int src_weight)
1.70 +{
1.71 + filter_by_weight(src, src_stride, dst, dst_stride, 4, src_weight);
1.72 +}
1.73 +
1.74 +static void apply_ifactor(unsigned char *y_src,
1.75 + int y_src_stride,
1.76 + unsigned char *y_dst,
1.77 + int y_dst_stride,
1.78 + unsigned char *u_src,
1.79 + unsigned char *v_src,
1.80 + int uv_src_stride,
1.81 + unsigned char *u_dst,
1.82 + unsigned char *v_dst,
1.83 + int uv_dst_stride,
1.84 + int block_size,
1.85 + int src_weight)
1.86 +{
1.87 + if (block_size == 16)
1.88 + {
1.89 + vp8_filter_by_weight16x16(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
1.90 + vp8_filter_by_weight8x8(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
1.91 + vp8_filter_by_weight8x8(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
1.92 + }
1.93 + else /* if (block_size == 8) */
1.94 + {
1.95 + vp8_filter_by_weight8x8(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
1.96 + vp8_filter_by_weight4x4(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
1.97 + vp8_filter_by_weight4x4(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
1.98 + }
1.99 +}
1.100 +
1.101 +static unsigned int int_sqrt(unsigned int x)
1.102 +{
1.103 + unsigned int y = x;
1.104 + unsigned int guess;
1.105 + int p = 1;
1.106 + while (y>>=1) p++;
1.107 + p>>=1;
1.108 +
1.109 + guess=0;
1.110 + while (p>=0)
1.111 + {
1.112 + guess |= (1<<p);
1.113 + if (x<guess*guess)
1.114 + guess -= (1<<p);
1.115 + p--;
1.116 + }
1.117 + /* choose between guess or guess+1 */
1.118 + return guess+(guess*guess+guess+1<=x);
1.119 +}
1.120 +
1.121 +#define USE_SSD
1.122 +static void multiframe_quality_enhance_block
1.123 +(
1.124 + int blksize, /* Currently only values supported are 16, 8 */
1.125 + int qcurr,
1.126 + int qprev,
1.127 + unsigned char *y,
1.128 + unsigned char *u,
1.129 + unsigned char *v,
1.130 + int y_stride,
1.131 + int uv_stride,
1.132 + unsigned char *yd,
1.133 + unsigned char *ud,
1.134 + unsigned char *vd,
1.135 + int yd_stride,
1.136 + int uvd_stride
1.137 +)
1.138 +{
1.139 + static const unsigned char VP8_ZEROS[16]=
1.140 + {
1.141 + 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
1.142 + };
1.143 + int uvblksize = blksize >> 1;
1.144 + int qdiff = qcurr - qprev;
1.145 +
1.146 + int i;
1.147 + unsigned char *up;
1.148 + unsigned char *udp;
1.149 + unsigned char *vp;
1.150 + unsigned char *vdp;
1.151 +
1.152 + unsigned int act, actd, sad, usad, vsad, sse, thr, thrsq, actrisk;
1.153 +
1.154 + if (blksize == 16)
1.155 + {
1.156 + actd = (vp8_variance16x16(yd, yd_stride, VP8_ZEROS, 0, &sse)+128)>>8;
1.157 + act = (vp8_variance16x16(y, y_stride, VP8_ZEROS, 0, &sse)+128)>>8;
1.158 +#ifdef USE_SSD
1.159 + sad = (vp8_variance16x16(y, y_stride, yd, yd_stride, &sse));
1.160 + sad = (sse + 128)>>8;
1.161 + usad = (vp8_variance8x8(u, uv_stride, ud, uvd_stride, &sse));
1.162 + usad = (sse + 32)>>6;
1.163 + vsad = (vp8_variance8x8(v, uv_stride, vd, uvd_stride, &sse));
1.164 + vsad = (sse + 32)>>6;
1.165 +#else
1.166 + sad = (vp8_sad16x16(y, y_stride, yd, yd_stride, UINT_MAX) + 128) >> 8;
1.167 + usad = (vp8_sad8x8(u, uv_stride, ud, uvd_stride, UINT_MAX) + 32) >> 6;
1.168 + vsad = (vp8_sad8x8(v, uv_stride, vd, uvd_stride, UINT_MAX)+ 32) >> 6;
1.169 +#endif
1.170 + }
1.171 + else /* if (blksize == 8) */
1.172 + {
1.173 + actd = (vp8_variance8x8(yd, yd_stride, VP8_ZEROS, 0, &sse)+32)>>6;
1.174 + act = (vp8_variance8x8(y, y_stride, VP8_ZEROS, 0, &sse)+32)>>6;
1.175 +#ifdef USE_SSD
1.176 + sad = (vp8_variance8x8(y, y_stride, yd, yd_stride, &sse));
1.177 + sad = (sse + 32)>>6;
1.178 + usad = (vp8_variance4x4(u, uv_stride, ud, uvd_stride, &sse));
1.179 + usad = (sse + 8)>>4;
1.180 + vsad = (vp8_variance4x4(v, uv_stride, vd, uvd_stride, &sse));
1.181 + vsad = (sse + 8)>>4;
1.182 +#else
1.183 + sad = (vp8_sad8x8(y, y_stride, yd, yd_stride, UINT_MAX) + 32) >> 6;
1.184 + usad = (vp8_sad4x4(u, uv_stride, ud, uvd_stride, UINT_MAX) + 8) >> 4;
1.185 + vsad = (vp8_sad4x4(v, uv_stride, vd, uvd_stride, UINT_MAX) + 8) >> 4;
1.186 +#endif
1.187 + }
1.188 +
1.189 + actrisk = (actd > act * 5);
1.190 +
1.191 + /* thr = qdiff/16 + log2(act) + log4(qprev) */
1.192 + thr = (qdiff >> 4);
1.193 + while (actd >>= 1) thr++;
1.194 + while (qprev >>= 2) thr++;
1.195 +
1.196 +#ifdef USE_SSD
1.197 + thrsq = thr * thr;
1.198 + if (sad < thrsq &&
1.199 + /* additional checks for color mismatch and excessive addition of
1.200 + * high-frequencies */
1.201 + 4 * usad < thrsq && 4 * vsad < thrsq && !actrisk)
1.202 +#else
1.203 + if (sad < thr &&
1.204 + /* additional checks for color mismatch and excessive addition of
1.205 + * high-frequencies */
1.206 + 2 * usad < thr && 2 * vsad < thr && !actrisk)
1.207 +#endif
1.208 + {
1.209 + int ifactor;
1.210 +#ifdef USE_SSD
1.211 + /* TODO: optimize this later to not need sqr root */
1.212 + sad = int_sqrt(sad);
1.213 +#endif
1.214 + ifactor = (sad << MFQE_PRECISION) / thr;
1.215 + ifactor >>= (qdiff >> 5);
1.216 +
1.217 + if (ifactor)
1.218 + {
1.219 + apply_ifactor(y, y_stride, yd, yd_stride,
1.220 + u, v, uv_stride,
1.221 + ud, vd, uvd_stride,
1.222 + blksize, ifactor);
1.223 + }
1.224 + }
1.225 + else /* else implicitly copy from previous frame */
1.226 + {
1.227 + if (blksize == 16)
1.228 + {
1.229 + vp8_copy_mem16x16(y, y_stride, yd, yd_stride);
1.230 + vp8_copy_mem8x8(u, uv_stride, ud, uvd_stride);
1.231 + vp8_copy_mem8x8(v, uv_stride, vd, uvd_stride);
1.232 + }
1.233 + else /* if (blksize == 8) */
1.234 + {
1.235 + vp8_copy_mem8x8(y, y_stride, yd, yd_stride);
1.236 + for (up = u, udp = ud, i = 0; i < uvblksize; ++i, up += uv_stride, udp += uvd_stride)
1.237 + vpx_memcpy(udp, up, uvblksize);
1.238 + for (vp = v, vdp = vd, i = 0; i < uvblksize; ++i, vp += uv_stride, vdp += uvd_stride)
1.239 + vpx_memcpy(vdp, vp, uvblksize);
1.240 + }
1.241 + }
1.242 +}
1.243 +
1.244 +static int qualify_inter_mb(const MODE_INFO *mode_info_context, int *map)
1.245 +{
1.246 + if (mode_info_context->mbmi.mb_skip_coeff)
1.247 + map[0] = map[1] = map[2] = map[3] = 1;
1.248 + else if (mode_info_context->mbmi.mode==SPLITMV)
1.249 + {
1.250 + static int ndx[4][4] =
1.251 + {
1.252 + {0, 1, 4, 5},
1.253 + {2, 3, 6, 7},
1.254 + {8, 9, 12, 13},
1.255 + {10, 11, 14, 15}
1.256 + };
1.257 + int i, j;
1.258 + for (i=0; i<4; ++i)
1.259 + {
1.260 + map[i] = 1;
1.261 + for (j=0; j<4 && map[j]; ++j)
1.262 + map[i] &= (mode_info_context->bmi[ndx[i][j]].mv.as_mv.row <= 2 &&
1.263 + mode_info_context->bmi[ndx[i][j]].mv.as_mv.col <= 2);
1.264 + }
1.265 + }
1.266 + else
1.267 + {
1.268 + map[0] = map[1] = map[2] = map[3] =
1.269 + (mode_info_context->mbmi.mode > B_PRED &&
1.270 + abs(mode_info_context->mbmi.mv.as_mv.row) <= 2 &&
1.271 + abs(mode_info_context->mbmi.mv.as_mv.col) <= 2);
1.272 + }
1.273 + return (map[0]+map[1]+map[2]+map[3]);
1.274 +}
1.275 +
1.276 +void vp8_multiframe_quality_enhance
1.277 +(
1.278 + VP8_COMMON *cm
1.279 +)
1.280 +{
1.281 + YV12_BUFFER_CONFIG *show = cm->frame_to_show;
1.282 + YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer;
1.283 +
1.284 + FRAME_TYPE frame_type = cm->frame_type;
1.285 + /* Point at base of Mb MODE_INFO list has motion vectors etc */
1.286 + const MODE_INFO *mode_info_context = cm->show_frame_mi;
1.287 + int mb_row;
1.288 + int mb_col;
1.289 + int totmap, map[4];
1.290 + int qcurr = cm->base_qindex;
1.291 + int qprev = cm->postproc_state.last_base_qindex;
1.292 +
1.293 + unsigned char *y_ptr, *u_ptr, *v_ptr;
1.294 + unsigned char *yd_ptr, *ud_ptr, *vd_ptr;
1.295 +
1.296 + /* Set up the buffer pointers */
1.297 + y_ptr = show->y_buffer;
1.298 + u_ptr = show->u_buffer;
1.299 + v_ptr = show->v_buffer;
1.300 + yd_ptr = dest->y_buffer;
1.301 + ud_ptr = dest->u_buffer;
1.302 + vd_ptr = dest->v_buffer;
1.303 +
1.304 + /* postprocess each macro block */
1.305 + for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
1.306 + {
1.307 + for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
1.308 + {
1.309 + /* if motion is high there will likely be no benefit */
1.310 + if (frame_type == INTER_FRAME) totmap = qualify_inter_mb(mode_info_context, map);
1.311 + else totmap = (frame_type == KEY_FRAME ? 4 : 0);
1.312 + if (totmap)
1.313 + {
1.314 + if (totmap < 4)
1.315 + {
1.316 + int i, j;
1.317 + for (i=0; i<2; ++i)
1.318 + for (j=0; j<2; ++j)
1.319 + {
1.320 + if (map[i*2+j])
1.321 + {
1.322 + multiframe_quality_enhance_block(8, qcurr, qprev,
1.323 + y_ptr + 8*(i*show->y_stride+j),
1.324 + u_ptr + 4*(i*show->uv_stride+j),
1.325 + v_ptr + 4*(i*show->uv_stride+j),
1.326 + show->y_stride,
1.327 + show->uv_stride,
1.328 + yd_ptr + 8*(i*dest->y_stride+j),
1.329 + ud_ptr + 4*(i*dest->uv_stride+j),
1.330 + vd_ptr + 4*(i*dest->uv_stride+j),
1.331 + dest->y_stride,
1.332 + dest->uv_stride);
1.333 + }
1.334 + else
1.335 + {
1.336 + /* copy a 8x8 block */
1.337 + int k;
1.338 + unsigned char *up = u_ptr + 4*(i*show->uv_stride+j);
1.339 + unsigned char *udp = ud_ptr + 4*(i*dest->uv_stride+j);
1.340 + unsigned char *vp = v_ptr + 4*(i*show->uv_stride+j);
1.341 + unsigned char *vdp = vd_ptr + 4*(i*dest->uv_stride+j);
1.342 + vp8_copy_mem8x8(y_ptr + 8*(i*show->y_stride+j), show->y_stride,
1.343 + yd_ptr + 8*(i*dest->y_stride+j), dest->y_stride);
1.344 + for (k = 0; k < 4; ++k, up += show->uv_stride, udp += dest->uv_stride,
1.345 + vp += show->uv_stride, vdp += dest->uv_stride)
1.346 + {
1.347 + vpx_memcpy(udp, up, 4);
1.348 + vpx_memcpy(vdp, vp, 4);
1.349 + }
1.350 + }
1.351 + }
1.352 + }
1.353 + else /* totmap = 4 */
1.354 + {
1.355 + multiframe_quality_enhance_block(16, qcurr, qprev, y_ptr,
1.356 + u_ptr, v_ptr,
1.357 + show->y_stride,
1.358 + show->uv_stride,
1.359 + yd_ptr, ud_ptr, vd_ptr,
1.360 + dest->y_stride,
1.361 + dest->uv_stride);
1.362 + }
1.363 + }
1.364 + else
1.365 + {
1.366 + vp8_copy_mem16x16(y_ptr, show->y_stride, yd_ptr, dest->y_stride);
1.367 + vp8_copy_mem8x8(u_ptr, show->uv_stride, ud_ptr, dest->uv_stride);
1.368 + vp8_copy_mem8x8(v_ptr, show->uv_stride, vd_ptr, dest->uv_stride);
1.369 + }
1.370 + y_ptr += 16;
1.371 + u_ptr += 8;
1.372 + v_ptr += 8;
1.373 + yd_ptr += 16;
1.374 + ud_ptr += 8;
1.375 + vd_ptr += 8;
1.376 + mode_info_context++; /* step to next MB */
1.377 + }
1.378 +
1.379 + y_ptr += show->y_stride * 16 - 16 * cm->mb_cols;
1.380 + u_ptr += show->uv_stride * 8 - 8 * cm->mb_cols;
1.381 + v_ptr += show->uv_stride * 8 - 8 * cm->mb_cols;
1.382 + yd_ptr += dest->y_stride * 16 - 16 * cm->mb_cols;
1.383 + ud_ptr += dest->uv_stride * 8 - 8 * cm->mb_cols;
1.384 + vd_ptr += dest->uv_stride * 8 - 8 * cm->mb_cols;
1.385 +
1.386 + mode_info_context++; /* Skip border mb */
1.387 + }
1.388 +}
