1 /* 2 * Copyright (c) 2012 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 /* MFQE: Multiframe Quality Enhancement 12 * In rate limited situations keyframes may cause significant visual artifacts 13 * commonly referred to as "popping." This file implements a postproccesing 14 * algorithm which blends data from the preceeding frame when there is no 15 * motion and the q from the previous frame is lower which indicates that it is 16 * higher quality. 17 */ 18 19 #include "./vp8_rtcd.h" 20 #include "./vpx_dsp_rtcd.h" 21 #include "vp8/common/postproc.h" 22 #include "vpx_dsp/variance.h" 23 #include "vpx_mem/vpx_mem.h" 24 #include "vpx_scale/yv12config.h" 25 26 #include <limits.h> 27 #include <stdlib.h> 28 29 static void filter_by_weight(unsigned char *src, int src_stride, 30 unsigned char *dst, int dst_stride, int block_size, 31 int src_weight) { 32 int dst_weight = (1 << MFQE_PRECISION) - src_weight; 33 int rounding_bit = 1 << (MFQE_PRECISION - 1); 34 int r, c; 35 36 for (r = 0; r < block_size; ++r) { 37 for (c = 0; c < block_size; ++c) { 38 dst[c] = (src[c] * src_weight + dst[c] * dst_weight + rounding_bit) >> 39 MFQE_PRECISION; 40 } 41 src += src_stride; 42 dst += dst_stride; 43 } 44 } 45 46 void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, 47 unsigned char *dst, int dst_stride, 48 int src_weight) { 49 filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight); 50 } 51 52 void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, 53 unsigned char *dst, int dst_stride, 54 int src_weight) { 55 filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight); 56 } 57 58 void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, 59 unsigned char *dst, int dst_stride, 60 int src_weight) { 61 filter_by_weight(src, src_stride, dst, dst_stride, 4, src_weight); 62 } 63 64 static void apply_ifactor(unsigned char *y_src, int y_src_stride, 65 unsigned char *y_dst, int y_dst_stride, 66 unsigned char *u_src, unsigned char *v_src, 67 int uv_src_stride, unsigned char *u_dst, 68 unsigned char *v_dst, int uv_dst_stride, 69 int block_size, int src_weight) { 70 if (block_size == 16) { 71 vp8_filter_by_weight16x16(y_src, y_src_stride, y_dst, y_dst_stride, 72 src_weight); 73 vp8_filter_by_weight8x8(u_src, uv_src_stride, u_dst, uv_dst_stride, 74 src_weight); 75 vp8_filter_by_weight8x8(v_src, uv_src_stride, v_dst, uv_dst_stride, 76 src_weight); 77 } else { 78 vp8_filter_by_weight8x8(y_src, y_src_stride, y_dst, y_dst_stride, 79 src_weight); 80 vp8_filter_by_weight4x4(u_src, uv_src_stride, u_dst, uv_dst_stride, 81 src_weight); 82 vp8_filter_by_weight4x4(v_src, uv_src_stride, v_dst, uv_dst_stride, 83 src_weight); 84 } 85 } 86 87 static unsigned int int_sqrt(unsigned int x) { 88 unsigned int y = x; 89 unsigned int guess; 90 int p = 1; 91 while (y >>= 1) p++; 92 p >>= 1; 93 94 guess = 0; 95 while (p >= 0) { 96 guess |= (1 << p); 97 if (x < guess * guess) guess -= (1 << p); 98 p--; 99 } 100 /* choose between guess or guess+1 */ 101 return guess + (guess * guess + guess + 1 <= x); 102 } 103 104 #define USE_SSD 105 static void multiframe_quality_enhance_block( 106 int blksize, /* Currently only values supported are 16, 8 */ 107 int qcurr, int qprev, unsigned char *y, unsigned char *u, unsigned char *v, 108 int y_stride, int uv_stride, unsigned char *yd, unsigned char *ud, 109 unsigned char *vd, int yd_stride, int uvd_stride) { 110 static const unsigned char VP8_ZEROS[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 111 0, 0, 0, 0, 0, 0, 0, 0 }; 112 int uvblksize = blksize >> 1; 113 int qdiff = qcurr - qprev; 114 115 int i; 116 unsigned char *up; 117 unsigned char *udp; 118 unsigned char *vp; 119 unsigned char *vdp; 120 121 unsigned int act, actd, sad, usad, vsad, sse, thr, thrsq, actrisk; 122 123 if (blksize == 16) { 124 actd = (vpx_variance16x16(yd, yd_stride, VP8_ZEROS, 0, &sse) + 128) >> 8; 125 act = (vpx_variance16x16(y, y_stride, VP8_ZEROS, 0, &sse) + 128) >> 8; 126 #ifdef USE_SSD 127 vpx_variance16x16(y, y_stride, yd, yd_stride, &sse); 128 sad = (sse + 128) >> 8; 129 vpx_variance8x8(u, uv_stride, ud, uvd_stride, &sse); 130 usad = (sse + 32) >> 6; 131 vpx_variance8x8(v, uv_stride, vd, uvd_stride, &sse); 132 vsad = (sse + 32) >> 6; 133 #else 134 sad = (vpx_sad16x16(y, y_stride, yd, yd_stride) + 128) >> 8; 135 usad = (vpx_sad8x8(u, uv_stride, ud, uvd_stride) + 32) >> 6; 136 vsad = (vpx_sad8x8(v, uv_stride, vd, uvd_stride) + 32) >> 6; 137 #endif 138 } else { 139 actd = (vpx_variance8x8(yd, yd_stride, VP8_ZEROS, 0, &sse) + 32) >> 6; 140 act = (vpx_variance8x8(y, y_stride, VP8_ZEROS, 0, &sse) + 32) >> 6; 141 #ifdef USE_SSD 142 vpx_variance8x8(y, y_stride, yd, yd_stride, &sse); 143 sad = (sse + 32) >> 6; 144 vpx_variance4x4(u, uv_stride, ud, uvd_stride, &sse); 145 usad = (sse + 8) >> 4; 146 vpx_variance4x4(v, uv_stride, vd, uvd_stride, &sse); 147 vsad = (sse + 8) >> 4; 148 #else 149 sad = (vpx_sad8x8(y, y_stride, yd, yd_stride) + 32) >> 6; 150 usad = (vpx_sad4x4(u, uv_stride, ud, uvd_stride) + 8) >> 4; 151 vsad = (vpx_sad4x4(v, uv_stride, vd, uvd_stride) + 8) >> 4; 152 #endif 153 } 154 155 actrisk = (actd > act * 5); 156 157 /* thr = qdiff/16 + log2(act) + log4(qprev) */ 158 thr = (qdiff >> 4); 159 while (actd >>= 1) thr++; 160 while (qprev >>= 2) thr++; 161 162 #ifdef USE_SSD 163 thrsq = thr * thr; 164 if (sad < thrsq && 165 /* additional checks for color mismatch and excessive addition of 166 * high-frequencies */ 167 4 * usad < thrsq && 4 * vsad < thrsq && !actrisk) 168 #else 169 if (sad < thr && 170 /* additional checks for color mismatch and excessive addition of 171 * high-frequencies */ 172 2 * usad < thr && 2 * vsad < thr && !actrisk) 173 #endif 174 { 175 int ifactor; 176 #ifdef USE_SSD 177 /* TODO: optimize this later to not need sqr root */ 178 sad = int_sqrt(sad); 179 #endif 180 ifactor = (sad << MFQE_PRECISION) / thr; 181 ifactor >>= (qdiff >> 5); 182 183 if (ifactor) { 184 apply_ifactor(y, y_stride, yd, yd_stride, u, v, uv_stride, ud, vd, 185 uvd_stride, blksize, ifactor); 186 } 187 } else { /* else implicitly copy from previous frame */ 188 if (blksize == 16) { 189 vp8_copy_mem16x16(y, y_stride, yd, yd_stride); 190 vp8_copy_mem8x8(u, uv_stride, ud, uvd_stride); 191 vp8_copy_mem8x8(v, uv_stride, vd, uvd_stride); 192 } else { 193 vp8_copy_mem8x8(y, y_stride, yd, yd_stride); 194 for (up = u, udp = ud, i = 0; i < uvblksize; 195 ++i, up += uv_stride, udp += uvd_stride) { 196 memcpy(udp, up, uvblksize); 197 } 198 for (vp = v, vdp = vd, i = 0; i < uvblksize; 199 ++i, vp += uv_stride, vdp += uvd_stride) { 200 memcpy(vdp, vp, uvblksize); 201 } 202 } 203 } 204 } 205 206 static int qualify_inter_mb(const MODE_INFO *mode_info_context, int *map) { 207 if (mode_info_context->mbmi.mb_skip_coeff) { 208 map[0] = map[1] = map[2] = map[3] = 1; 209 } else if (mode_info_context->mbmi.mode == SPLITMV) { 210 static int ndx[4][4] = { 211 { 0, 1, 4, 5 }, { 2, 3, 6, 7 }, { 8, 9, 12, 13 }, { 10, 11, 14, 15 } 212 }; 213 int i, j; 214 for (i = 0; i < 4; ++i) { 215 map[i] = 1; 216 for (j = 0; j < 4 && map[j]; ++j) { 217 map[i] &= (mode_info_context->bmi[ndx[i][j]].mv.as_mv.row <= 2 && 218 mode_info_context->bmi[ndx[i][j]].mv.as_mv.col <= 2); 219 } 220 } 221 } else { 222 map[0] = map[1] = map[2] = map[3] = 223 (mode_info_context->mbmi.mode > B_PRED && 224 abs(mode_info_context->mbmi.mv.as_mv.row) <= 2 && 225 abs(mode_info_context->mbmi.mv.as_mv.col) <= 2); 226 } 227 return (map[0] + map[1] + map[2] + map[3]); 228 } 229 230 void vp8_multiframe_quality_enhance(VP8_COMMON *cm) { 231 YV12_BUFFER_CONFIG *show = cm->frame_to_show; 232 YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer; 233 234 FRAME_TYPE frame_type = cm->frame_type; 235 /* Point at base of Mb MODE_INFO list has motion vectors etc */ 236 const MODE_INFO *mode_info_context = cm->show_frame_mi; 237 int mb_row; 238 int mb_col; 239 int totmap, map[4]; 240 int qcurr = cm->base_qindex; 241 int qprev = cm->postproc_state.last_base_qindex; 242 243 unsigned char *y_ptr, *u_ptr, *v_ptr; 244 unsigned char *yd_ptr, *ud_ptr, *vd_ptr; 245 246 /* Set up the buffer pointers */ 247 y_ptr = show->y_buffer; 248 u_ptr = show->u_buffer; 249 v_ptr = show->v_buffer; 250 yd_ptr = dest->y_buffer; 251 ud_ptr = dest->u_buffer; 252 vd_ptr = dest->v_buffer; 253 254 /* postprocess each macro block */ 255 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) { 256 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) { 257 /* if motion is high there will likely be no benefit */ 258 if (frame_type == INTER_FRAME) { 259 totmap = qualify_inter_mb(mode_info_context, map); 260 } else { 261 totmap = (frame_type == KEY_FRAME ? 4 : 0); 262 } 263 if (totmap) { 264 if (totmap < 4) { 265 int i, j; 266 for (i = 0; i < 2; ++i) { 267 for (j = 0; j < 2; ++j) { 268 if (map[i * 2 + j]) { 269 multiframe_quality_enhance_block( 270 8, qcurr, qprev, y_ptr + 8 * (i * show->y_stride + j), 271 u_ptr + 4 * (i * show->uv_stride + j), 272 v_ptr + 4 * (i * show->uv_stride + j), show->y_stride, 273 show->uv_stride, yd_ptr + 8 * (i * dest->y_stride + j), 274 ud_ptr + 4 * (i * dest->uv_stride + j), 275 vd_ptr + 4 * (i * dest->uv_stride + j), dest->y_stride, 276 dest->uv_stride); 277 } else { 278 /* copy a 8x8 block */ 279 int k; 280 unsigned char *up = u_ptr + 4 * (i * show->uv_stride + j); 281 unsigned char *udp = ud_ptr + 4 * (i * dest->uv_stride + j); 282 unsigned char *vp = v_ptr + 4 * (i * show->uv_stride + j); 283 unsigned char *vdp = vd_ptr + 4 * (i * dest->uv_stride + j); 284 vp8_copy_mem8x8( 285 y_ptr + 8 * (i * show->y_stride + j), show->y_stride, 286 yd_ptr + 8 * (i * dest->y_stride + j), dest->y_stride); 287 for (k = 0; k < 4; ++k, up += show->uv_stride, 288 udp += dest->uv_stride, vp += show->uv_stride, 289 vdp += dest->uv_stride) { 290 memcpy(udp, up, 4); 291 memcpy(vdp, vp, 4); 292 } 293 } 294 } 295 } 296 } else { /* totmap = 4 */ 297 multiframe_quality_enhance_block( 298 16, qcurr, qprev, y_ptr, u_ptr, v_ptr, show->y_stride, 299 show->uv_stride, yd_ptr, ud_ptr, vd_ptr, dest->y_stride, 300 dest->uv_stride); 301 } 302 } else { 303 vp8_copy_mem16x16(y_ptr, show->y_stride, yd_ptr, dest->y_stride); 304 vp8_copy_mem8x8(u_ptr, show->uv_stride, ud_ptr, dest->uv_stride); 305 vp8_copy_mem8x8(v_ptr, show->uv_stride, vd_ptr, dest->uv_stride); 306 } 307 y_ptr += 16; 308 u_ptr += 8; 309 v_ptr += 8; 310 yd_ptr += 16; 311 ud_ptr += 8; 312 vd_ptr += 8; 313 mode_info_context++; /* step to next MB */ 314 } 315 316 y_ptr += show->y_stride * 16 - 16 * cm->mb_cols; 317 u_ptr += show->uv_stride * 8 - 8 * cm->mb_cols; 318 v_ptr += show->uv_stride * 8 - 8 * cm->mb_cols; 319 yd_ptr += dest->y_stride * 16 - 16 * cm->mb_cols; 320 ud_ptr += dest->uv_stride * 8 - 8 * cm->mb_cols; 321 vd_ptr += dest->uv_stride * 8 - 8 * cm->mb_cols; 322 323 mode_info_context++; /* Skip border mb */ 324 } 325 } 326