vp8/encoder/picklpf.c

   1 /*
   2  *  Copyright (c) 2010 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
  12 #include "onyxc_int.h"
  13 #include "onyx_int.h"
  14 #include "quantize.h"
  15 #include "vpx_mem/vpx_mem.h"
  16 #include "vpx_scale/yv12extend.h"
  17 #include "vpx_scale/vpxscale.h"
  18 #include "alloccommon.h"
  19 #if ARCH_ARM
  20 #include "vpx_ports/arm.h"
  21 #endif
  22
  23 extern void vp8_loop_filter_frame(VP8_COMMON *cm,    MACROBLOCKD *mbd,  int filt_val);
  24 extern void vp8_loop_filter_frame_yonly(VP8_COMMON *cm,    MACROBLOCKD *mbd,  int filt_val, int sharpness_lvl);
  25 extern int vp8_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest, const vp8_variance_rtcd_vtable_t *rtcd);
  26 #if HAVE_ARMV7
  27 extern void vp8_yv12_copy_frame_yonly_no_extend_frame_borders_neon(YV12_BUFFER_CONFIG *src_ybc, YV12_BUFFER_CONFIG *dst_ybc);
  28 #endif
  29
  30 #if CONFIG_RUNTIME_CPU_DETECT
  31 #define IF_RTCD(x) (x)
  32 #else
  33 #define IF_RTCD(x) NULL
  34 #endif
  35
  36 extern void
  37 (*vp8_yv12_copy_partial_frame_ptr)(YV12_BUFFER_CONFIG *src_ybc,
  38                                    YV12_BUFFER_CONFIG *dst_ybc,
  39                                    int Fraction);
  40 void
  41 vp8_yv12_copy_partial_frame(YV12_BUFFER_CONFIG *src_ybc, YV12_BUFFER_CONFIG *dst_ybc, int Fraction)
  42 {
  43     unsigned char *src_y, *dst_y;
  44     int yheight;
  45     int ystride;
  46     int border;
  47     int yoffset;
  48     int linestocopy;
  49
  50     border   = src_ybc->border;
  51     yheight  = src_ybc->y_height;
  52     ystride  = src_ybc->y_stride;
  53
  54     linestocopy = (yheight >> (Fraction + 4));
  55
  56     if (linestocopy < 1)
  57         linestocopy = 1;
  58
  59     linestocopy <<= 4;
  60
  61     yoffset  = ystride * ((yheight >> 5) * 16 - 8);
  62     src_y = src_ybc->y_buffer + yoffset;
  63     dst_y = dst_ybc->y_buffer + yoffset;
  64
  65     vpx_memcpy(dst_y, src_y, ystride *(linestocopy + 16));
  66 }
  67
  68 static int vp8_calc_partial_ssl_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest, int Fraction, const vp8_variance_rtcd_vtable_t *rtcd)
  69 {
  70     int i, j;
  71     int Total = 0;
  72     int srcoffset, dstoffset;
  73     unsigned char *src = source->y_buffer;
  74     unsigned char *dst = dest->y_buffer;
  75
  76     int linestocopy = (source->y_height >> (Fraction + 4));
  77     (void)rtcd;
  78
  79     if (linestocopy < 1)
  80         linestocopy = 1;
  81
  82     linestocopy <<= 4;
  83
  84
  85     srcoffset = source->y_stride   * (dest->y_height >> 5) * 16;
  86     dstoffset = dest->y_stride     * (dest->y_height >> 5) * 16;
  87
  88     src += srcoffset;
  89     dst += dstoffset;
  90
  91     // Loop through the Y plane raw and reconstruction data summing (square differences)
  92     for (i = 0; i < linestocopy; i += 16)
  93     {
  94         for (j = 0; j < source->y_width; j += 16)
  95         {
  96             unsigned int sse;
  97             Total += VARIANCE_INVOKE(rtcd, mse16x16)(src + j, source->y_stride, dst + j, dest->y_stride, &sse);
  98         }
  99
 100         src += 16 * source->y_stride;
 101         dst += 16 * dest->y_stride;
 102     }
 103
 104     return Total;
 105 }
 106
 107 extern void vp8_loop_filter_partial_frame
 108 (
 109     VP8_COMMON *cm,
 110     MACROBLOCKD *mbd,
 111     int default_filt_lvl,
 112     int sharpness_lvl,
 113     int Fraction
 114 );
 115
 116 // Enforce a minimum filter level based upon baseline Q
 117 static int get_min_filter_level(VP8_COMP *cpi, int base_qindex)
 118 {
 119     int min_filter_level;
 120
 121     if (cpi->source_alt_ref_active && cpi->common.refresh_golden_frame && !cpi->common.refresh_alt_ref_frame)
 122         min_filter_level = 0;
 123     else
 124     {
 125         if (base_qindex <= 6)
 126             min_filter_level = 0;
 127         else if (base_qindex <= 16)
 128             min_filter_level = 1;
 129         else
 130             min_filter_level = (base_qindex / 8);
 131     }
 132
 133     return min_filter_level;
 134 }
 135
 136 // Enforce a maximum filter level based upon baseline Q
 137 static int get_max_filter_level(VP8_COMP *cpi, int base_qindex)
 138 {
 139     // PGW August 2006: Highest filter values almost always a bad idea
 140
 141     // jbb chg: 20100118 - not so any more with this overquant stuff allow high values
 142     // with lots of intra coming in.
 143     int max_filter_level = MAX_LOOP_FILTER ;//* 3 / 4;
 144
 145     if (cpi->section_intra_rating > 8)
 146         max_filter_level = MAX_LOOP_FILTER * 3 / 4;
 147
 148     (void) cpi;
 149     (void) base_qindex;
 150
 151     return max_filter_level;
 152 }
 153
 154 void vp8cx_pick_filter_level_fast(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
 155 {
 156     VP8_COMMON *cm = &cpi->common;
 157
 158     int best_err = 0;
 159     int filt_err = 0;
 160     int min_filter_level = 0;
 161     int max_filter_level = MAX_LOOP_FILTER * 3 / 4;   // PGW August 2006: Highest filter values almost always a bad idea
 162     int filt_val;
 163     int best_filt_val = cm->filter_level;
 164
 165     //  Make a copy of the unfiltered / processed recon buffer
 166     //vp8_yv12_copy_frame_ptr( cm->frame_to_show, &cpi->last_frame_uf  );
 167     vp8_yv12_copy_partial_frame_ptr(cm->frame_to_show, &cpi->last_frame_uf, 3);
 168
 169     if (cm->frame_type == KEY_FRAME)
 170         cm->sharpness_level = 0;
 171     else
 172         cm->sharpness_level = cpi->oxcf.Sharpness;
 173
 174     // Enforce a minimum filter level based upon Q
 175     min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
 176     max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
 177
 178     // Start the search at the previous frame filter level unless it is now out of range.
 179     if (cm->filter_level < min_filter_level)
 180         cm->filter_level = min_filter_level;
 181     else if (cm->filter_level > max_filter_level)
 182         cm->filter_level = max_filter_level;
 183
 184     filt_val = cm->filter_level;
 185     best_filt_val = filt_val;
 186
 187     // Set up alternate filter values
 188
 189     // Get the err using the previous frame's filter value.
 190     vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val, 0  , 3);
 191     cm->last_frame_type = cm->frame_type;
 192     cm->last_filter_type = cm->filter_type;
 193     cm->last_sharpness_level = cm->sharpness_level;
 194
 195     best_err = vp8_calc_partial_ssl_err(sd, cm->frame_to_show, 3, IF_RTCD(&cpi->rtcd.variance));
 196
 197     //  Re-instate the unfiltered frame
 198     vp8_yv12_copy_partial_frame_ptr(&cpi->last_frame_uf, cm->frame_to_show, 3);
 199
 200     filt_val -= (1 + ((filt_val > 10) ? 1 : 0));
 201
 202     // Search lower filter levels
 203     while (filt_val >= min_filter_level)
 204     {
 205         // Apply the loop filter
 206         vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val, 0, 3);
 207         cm->last_frame_type = cm->frame_type;
 208         cm->last_filter_type = cm->filter_type;
 209         cm->last_sharpness_level = cm->sharpness_level;
 210
 211         // Get the err for filtered frame
 212         filt_err = vp8_calc_partial_ssl_err(sd, cm->frame_to_show, 3, IF_RTCD(&cpi->rtcd.variance));
 213
 214
 215         //  Re-instate the unfiltered frame
 216         vp8_yv12_copy_partial_frame_ptr(&cpi->last_frame_uf, cm->frame_to_show, 3);
 217
 218
 219         // Update the best case record or exit loop.
 220         if (filt_err < best_err)
 221         {
 222             best_err = filt_err;
 223             best_filt_val = filt_val;
 224         }
 225         else
 226             break;
 227
 228         // Adjust filter level
 229         filt_val -= (1 + ((filt_val > 10) ? 1 : 0));
 230     }
 231
 232     // Search up (note that we have already done filt_val = cm->filter_level)
 233     filt_val = cm->filter_level + (1 + ((filt_val > 10) ? 1 : 0));
 234
 235     if (best_filt_val == cm->filter_level)
 236     {
 237         // Resist raising filter level for very small gains
 238         best_err -= (best_err >> 10);
 239
 240         while (filt_val < max_filter_level)
 241         {
 242             // Apply the loop filter
 243             vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val, 0, 3);
 244             cm->last_frame_type = cm->frame_type;
 245             cm->last_filter_type = cm->filter_type;
 246             cm->last_sharpness_level = cm->sharpness_level;
 247
 248             // Get the err for filtered frame
 249             filt_err = vp8_calc_partial_ssl_err(sd, cm->frame_to_show, 3, IF_RTCD(&cpi->rtcd.variance));
 250
 251             //  Re-instate the unfiltered frame
 252             vp8_yv12_copy_partial_frame_ptr(&cpi->last_frame_uf, cm->frame_to_show, 3);
 253
 254             // Update the best case record or exit loop.
 255             if (filt_err < best_err)
 256             {
 257                 // Do not raise filter level if improvement is < 1 part in 4096
 258                 best_err = filt_err - (filt_err >> 10);
 259
 260                 best_filt_val = filt_val;
 261             }
 262             else
 263                 break;
 264
 265             // Adjust filter level
 266             filt_val += (1 + ((filt_val > 10) ? 1 : 0));
 267         }
 268     }
 269
 270     cm->filter_level = best_filt_val;
 271
 272     if (cm->filter_level < min_filter_level)
 273         cm->filter_level = min_filter_level;
 274
 275     if (cm->filter_level > max_filter_level)
 276         cm->filter_level = max_filter_level;
 277 }
 278
 279 // Stub function for now Alt LF not used
 280 void vp8cx_set_alt_lf_level(VP8_COMP *cpi, int filt_val)
 281 {
 282     MACROBLOCKD *mbd = &cpi->mb.e_mbd;
 283     (void) filt_val;
 284
 285     mbd->segment_feature_data[MB_LVL_ALT_LF][0] = cpi->segment_feature_data[MB_LVL_ALT_LF][0];
 286     mbd->segment_feature_data[MB_LVL_ALT_LF][1] = cpi->segment_feature_data[MB_LVL_ALT_LF][1];
 287     mbd->segment_feature_data[MB_LVL_ALT_LF][2] = cpi->segment_feature_data[MB_LVL_ALT_LF][2];
 288     mbd->segment_feature_data[MB_LVL_ALT_LF][3] = cpi->segment_feature_data[MB_LVL_ALT_LF][3];
 289 }
 290
 291 void vp8cx_pick_filter_level(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
 292 {
 293     VP8_COMMON *cm = &cpi->common;
 294
 295     int best_err = 0;
 296     int filt_err = 0;
 297     int min_filter_level;
 298     int max_filter_level;
 299
 300     int filter_step;
 301     int filt_high = 0;
 302     int filt_mid = cm->filter_level;      // Start search at previous frame filter level
 303     int filt_low = 0;
 304     int filt_best;
 305     int filt_direction = 0;
 306
 307     int Bias = 0;                       // Bias against raising loop filter and in favour of lowering it
 308
 309     //  Make a copy of the unfiltered / processed recon buffer
 310 #if HAVE_ARMV7
 311 #if CONFIG_RUNTIME_CPU_DETECT
 312     if (cm->rtcd.flags & HAS_NEON)
 313 #endif
 314     {
 315         vp8_yv12_copy_frame_yonly_no_extend_frame_borders_neon(cm->frame_to_show, &cpi->last_frame_uf);
 316     }
 317 #if CONFIG_RUNTIME_CPU_DETECT
 318     else
 319 #endif
 320 #endif
 321 #if !HAVE_ARMV7 || CONFIG_RUNTIME_CPU_DETECT
 322     {
 323         vp8_yv12_copy_frame_ptr(cm->frame_to_show, &cpi->last_frame_uf);
 324     }
 325 #endif
 326
 327     if (cm->frame_type == KEY_FRAME)
 328         cm->sharpness_level = 0;
 329     else
 330         cm->sharpness_level = cpi->oxcf.Sharpness;
 331
 332     // Enforce a minimum filter level based upon Q
 333     min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
 334     max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
 335
 336     // Start the search at the previous frame filter level unless it is now out of range.
 337     filt_mid = cm->filter_level;
 338
 339     if (filt_mid < min_filter_level)
 340         filt_mid = min_filter_level;
 341     else if (filt_mid > max_filter_level)
 342         filt_mid = max_filter_level;
 343
 344     // Define the initial step size
 345     filter_step = (filt_mid < 16) ? 4 : filt_mid / 4;
 346
 347     // Get baseline error score
 348     vp8cx_set_alt_lf_level(cpi, filt_mid);
 349     vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_mid, 0);
 350     cm->last_frame_type = cm->frame_type;
 351     cm->last_filter_type = cm->filter_type;
 352     cm->last_sharpness_level = cm->sharpness_level;
 353
 354     best_err = vp8_calc_ss_err(sd, cm->frame_to_show, IF_RTCD(&cpi->rtcd.variance));
 355     filt_best = filt_mid;
 356
 357     //  Re-instate the unfiltered frame
 358 #if HAVE_ARMV7
 359 #if CONFIG_RUNTIME_CPU_DETECT
 360     if (cm->rtcd.flags & HAS_NEON)
 361 #endif
 362     {
 363         vp8_yv12_copy_frame_yonly_no_extend_frame_borders_neon(&cpi->last_frame_uf, cm->frame_to_show);
 364     }
 365 #if CONFIG_RUNTIME_CPU_DETECT
 366     else
 367 #endif
 368 #endif
 369 #if !HAVE_ARMV7 || CONFIG_RUNTIME_CPU_DETECT
 370     {
 371         vp8_yv12_copy_frame_yonly_ptr(&cpi->last_frame_uf, cm->frame_to_show);
 372     }
 373 #endif
 374
 375     while (filter_step > 0)
 376     {
 377         Bias = (best_err >> (15 - (filt_mid / 8))) * filter_step; //PGW change 12/12/06 for small images
 378
 379         // jbb chg: 20100118 - in sections with lots of new material coming in don't bias as much to a low filter value
 380         if (cpi->section_intra_rating < 20)
 381             Bias = Bias * cpi->section_intra_rating / 20;
 382
 383         filt_high = ((filt_mid + filter_step) > max_filter_level) ? max_filter_level : (filt_mid + filter_step);
 384         filt_low = ((filt_mid - filter_step) < min_filter_level) ? min_filter_level : (filt_mid - filter_step);
 385
 386         if ((filt_direction <= 0) && (filt_low != filt_mid))
 387         {
 388             // Get Low filter error score
 389             vp8cx_set_alt_lf_level(cpi, filt_low);
 390             vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_low, 0);
 391             cm->last_frame_type = cm->frame_type;
 392             cm->last_filter_type = cm->filter_type;
 393             cm->last_sharpness_level = cm->sharpness_level;
 394
 395             filt_err = vp8_calc_ss_err(sd, cm->frame_to_show, IF_RTCD(&cpi->rtcd.variance));
 396
 397             //  Re-instate the unfiltered frame
 398 #if HAVE_ARMV7
 399 #if CONFIG_RUNTIME_CPU_DETECT
 400             if (cm->rtcd.flags & HAS_NEON)
 401 #endif
 402             {
 403                 vp8_yv12_copy_frame_yonly_no_extend_frame_borders_neon(&cpi->last_frame_uf, cm->frame_to_show);
 404             }
 405 #if CONFIG_RUNTIME_CPU_DETECT
 406             else
 407 #endif
 408 #endif
 409 #if !HAVE_ARMV7 || CONFIG_RUNTIME_CPU_DETECT
 410             {
 411                 vp8_yv12_copy_frame_yonly_ptr(&cpi->last_frame_uf, cm->frame_to_show);
 412             }
 413 #endif
 414
 415             // If value is close to the best so far then bias towards a lower loop filter value.
 416             if ((filt_err - Bias) < best_err)
 417             {
 418                 // Was it actually better than the previous best?
 419                 if (filt_err < best_err)
 420                     best_err = filt_err;
 421
 422                 filt_best = filt_low;
 423             }
 424         }
 425
 426         // Now look at filt_high
 427         if ((filt_direction >= 0) && (filt_high != filt_mid))
 428         {
 429             vp8cx_set_alt_lf_level(cpi, filt_high);
 430             vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_high, 0);
 431             cm->last_frame_type = cm->frame_type;
 432             cm->last_filter_type = cm->filter_type;
 433             cm->last_sharpness_level = cm->sharpness_level;
 434
 435             filt_err = vp8_calc_ss_err(sd, cm->frame_to_show, IF_RTCD(&cpi->rtcd.variance));
 436
 437             //  Re-instate the unfiltered frame
 438 #if HAVE_ARMV7
 439 #if CONFIG_RUNTIME_CPU_DETECT
 440             if (cm->rtcd.flags & HAS_NEON)
 441 #endif
 442             {
 443                 vp8_yv12_copy_frame_yonly_no_extend_frame_borders_neon(&cpi->last_frame_uf, cm->frame_to_show);
 444             }
 445 #if CONFIG_RUNTIME_CPU_DETECT
 446             else
 447 #endif
 448 #endif
 449 #if !HAVE_ARMV7 || CONFIG_RUNTIME_CPU_DETECT
 450             {
 451                 vp8_yv12_copy_frame_yonly_ptr(&cpi->last_frame_uf, cm->frame_to_show);
 452             }
 453 #endif
 454
 455             // Was it better than the previous best?
 456             if (filt_err < (best_err - Bias))
 457             {
 458                 best_err = filt_err;
 459                 filt_best = filt_high;
 460             }
 461         }
 462
 463         // Half the step distance if the best filter value was the same as last time
 464         if (filt_best == filt_mid)
 465         {
 466             filter_step = filter_step / 2;
 467             filt_direction = 0;
 468         }
 469         else
 470         {
 471             filt_direction = (filt_best < filt_mid) ? -1 : 1;
 472             filt_mid = filt_best;
 473         }
 474     }
 475
 476     cm->filter_level = filt_best;
 477     cpi->last_auto_filt_val = filt_best;
 478     cpi->last_auto_filt_q  = cm->base_qindex;
 479
 480     cpi->frames_since_auto_filter = 0;
 481 }