cinelerra/maskengine.C

   1 #include "bcsignals.h"
   2 #include "condition.h"
   3 #include "clip.h"
   4 #include "maskauto.h"
   5 #include "maskautos.h"
   6 #include "maskengine.h"
   7 #include "mutex.h"
   8 #include "vframe.h"
   9
  10 #include <math.h>
  11 #include <stdint.h>
  12 #include <string.h>
  13 #include <limits.h>
  14
  15 #include "feather.h"
  16
  17
  18 int64_t get_difference(struct timeval *start_time)
  19 {
  20         struct timeval new_time;
  21
  22         gettimeofday(&new_time, 0);
  23
  24         new_time.tv_usec -= start_time->tv_usec;
  25         new_time.tv_sec -= start_time->tv_sec;
  26         if(new_time.tv_usec < 0)
  27         {
  28                 new_time.tv_usec += 1000000;
  29                 new_time.tv_sec--;
  30         }
  31
  32         return (int64_t)new_time.tv_sec * 1000000 +
  33                 (int64_t)new_time.tv_usec;
  34
  35 }
  36
  37
  38
  39 MaskPackage::MaskPackage()
  40 {
  41 }
  42
  43 MaskPackage::~MaskPackage()
  44 {
  45 }
  46
  47
  48
  49
  50
  51 MaskUnit::MaskUnit(MaskEngine *engine)
  52  : LoadClient(engine)
  53 {
  54         this->engine = engine;
  55         row_spans_h = 0;
  56         row_spans = 0;
  57 }
  58
  59
  60 MaskUnit::~MaskUnit()
  61 {
  62         if (row_spans)
  63         {
  64                 for (int i = 0; i < row_spans_h; i++)
  65                         free(row_spans[i]);
  66                 delete [] row_spans;
  67         }
  68 }
  69
  70 #ifndef SQR
  71 #define SQR(x) ((x) * (x))
  72 #endif
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89 inline void MaskUnit::draw_line_clamped(
  90         int draw_x1,
  91         int draw_y1,
  92         int draw_x2,
  93         int draw_y2,
  94         int w,
  95         int h,
  96         int hoffset)
  97 {
  98 //printf("MaskUnit::draw_line_clamped 1 %d %d %d %d\n", x1, y1, x2, y2);
  99         if (draw_y1 == draw_y2) return;
 100
 101         if(draw_y2 < draw_y1)
 102         { /* change the order */
 103                 int tmp;
 104                 tmp = draw_x1;
 105                 draw_x1 = draw_x2;
 106                 draw_x2 = tmp;
 107                 tmp = draw_y1;
 108                 draw_y1 = draw_y2;
 109                 draw_y2 = tmp;
 110         }
 111
 112         float slope = ((float)draw_x2 - draw_x1) / ((float)draw_y2 - draw_y1);
 113         w--;
 114         for(int y_i = draw_y1; y_i < draw_y2; y_i++)
 115         {
 116                 if (y_i >= h)
 117                         return; // since y gets larger, there is no point in continuing
 118                 else if(y_i >= 0)
 119                 {
 120                         int x = (int)(slope * (y_i - draw_y1) + draw_x1);
 121                         int x_i = CLIP(x, 0, w);
 122
 123                         /* now insert into span in order */
 124                         short *span = row_spans[y_i + hoffset];
 125                         if (span[0] >= span[1]) { /* do the reallocation */
 126                                 span[1] *= 2;
 127                                 span = row_spans[y_i + hoffset] = (short *) realloc (span, span[1] * sizeof(short)); /* be careful! row_spans has to be updated! */
 128                         };
 129
 130                         short index = 2;
 131                         while (index < span[0]  && span[index] < x_i)
 132                                 index++;
 133                         for (int j = span[0]; j > index; j--) {       // move forward
 134                                 span[j] = span[j-1];
 135                         }
 136                         span[index] = x_i;
 137                         span[0] ++;
 138                 }
 139         }
 140 }
 141
 142 template<class T>
 143 void MaskUnit::blur_strip(float *val_p,
 144         float *val_m,
 145         float *dst,
 146         float *src,
 147         int size,
 148         T max)
 149 {
 150         float *sp_p = src;
 151         float *sp_m = src + size - 1;
 152         float *vp = val_p;
 153         float *vm = val_m + size - 1;
 154         float initial_p = sp_p[0];
 155         float initial_m = sp_m[0];
 156
 157 //printf("MaskUnit::blur_strip %d\n", size);
 158         for(int k = 0; k < size; k++)
 159         {
 160                 int terms = (k < 4) ? k : 4;
 161                 int l;
 162                 for(l = 0; l <= terms; l++)
 163                 {
 164                         *vp += n_p[l] * sp_p[-l] - d_p[l] * vp[-l];
 165                         *vm += n_m[l] * sp_m[l] - d_m[l] * vm[l];
 166                 }
 167
 168                 for( ; l <= 4; l++)
 169                 {
 170                         *vp += (n_p[l] - bd_p[l]) * initial_p;
 171                         *vm += (n_m[l] - bd_m[l]) * initial_m;
 172                 }
 173                 sp_p++;
 174                 sp_m--;
 175                 vp++;
 176                 vm--;
 177         }
 178
 179         for(int i = 0; i < size; i++)
 180         {
 181                 float sum = val_p[i] + val_m[i];
 182                 CLAMP(sum, 0, max);
 183                 dst[i] = sum;
 184         }
 185 }
 186
 187
 188
 189 int MaskUnit::do_feather_2(VFrame *output,
 190         VFrame *input,
 191         float feather,
 192         int start_out,
 193         int end_out)
 194 {
 195
 196         int fint = (int)feather;
 197         DO_FEATHER_N(unsigned char, uint32_t, 0xffff, fint);
 198
 199 }
 200
 201
 202 void MaskUnit::do_feather(VFrame *output,
 203         VFrame *input,
 204         float feather,
 205         int start_out,
 206         int end_out)
 207 {
 208 //printf("MaskUnit::do_feather %f\n", feather);
 209 // Get constants
 210         double constants[8];
 211         double div;
 212         double std_dev = sqrt(-(double)(feather * feather) / (2 * log(1.0 / 255.0)));
 213         div = sqrt(2 * M_PI) * std_dev;
 214         constants[0] = -1.783 / std_dev;
 215         constants[1] = -1.723 / std_dev;
 216         constants[2] = 0.6318 / std_dev;
 217         constants[3] = 1.997  / std_dev;
 218         constants[4] = 1.6803 / div;
 219         constants[5] = 3.735 / div;
 220         constants[6] = -0.6803 / div;
 221         constants[7] = -0.2598 / div;
 222
 223         n_p[0] = constants[4] + constants[6];
 224         n_p[1] = exp(constants[1]) *
 225                                 (constants[7] * sin(constants[3]) -
 226                                 (constants[6] + 2 * constants[4]) * cos(constants[3])) +
 227                                 exp(constants[0]) *
 228                                 (constants[5] * sin(constants[2]) -
 229                                 (2 * constants[6] + constants[4]) * cos(constants[2]));
 230
 231         n_p[2] = 2 * exp(constants[0] + constants[1]) *
 232                                 ((constants[4] + constants[6]) * cos(constants[3]) *
 233                                 cos(constants[2]) - constants[5] *
 234                                 cos(constants[3]) * sin(constants[2]) -
 235                                 constants[7] * cos(constants[2]) * sin(constants[3])) +
 236                                 constants[6] * exp(2 * constants[0]) +
 237                                 constants[4] * exp(2 * constants[1]);
 238
 239         n_p[3] = exp(constants[1] + 2 * constants[0]) *
 240                                 (constants[7] * sin(constants[3]) -
 241                                 constants[6] * cos(constants[3])) +
 242                                 exp(constants[0] + 2 * constants[1]) *
 243                                 (constants[5] * sin(constants[2]) - constants[4] *
 244                                 cos(constants[2]));
 245         n_p[4] = 0.0;
 246
 247         d_p[0] = 0.0;
 248         d_p[1] = -2 * exp(constants[1]) * cos(constants[3]) -
 249                                 2 * exp(constants[0]) * cos(constants[2]);
 250
 251         d_p[2] = 4 * cos(constants[3]) * cos(constants[2]) *
 252                                 exp(constants[0] + constants[1]) +
 253                                 exp(2 * constants[1]) + exp (2 * constants[0]);
 254
 255         d_p[3] = -2 * cos(constants[2]) * exp(constants[0] + 2 * constants[1]) -
 256                                 2 * cos(constants[3]) * exp(constants[1] + 2 * constants[0]);
 257
 258         d_p[4] = exp(2 * constants[0] + 2 * constants[1]);
 259
 260         for(int i = 0; i < 5; i++) d_m[i] = d_p[i];
 261
 262         n_m[0] = 0.0;
 263         for(int i = 1; i <= 4; i++)
 264                 n_m[i] = n_p[i] - d_p[i] * n_p[0];
 265
 266         double sum_n_p, sum_n_m, sum_d;
 267         double a, b;
 268
 269         sum_n_p = 0.0;
 270         sum_n_m = 0.0;
 271         sum_d = 0.0;
 272         for(int i = 0; i < 5; i++)
 273         {
 274                 sum_n_p += n_p[i];
 275                 sum_n_m += n_m[i];
 276                 sum_d += d_p[i];
 277         }
 278
 279         a = sum_n_p / (1 + sum_d);
 280         b = sum_n_m / (1 + sum_d);
 281
 282         for(int i = 0; i < 5; i++)
 283         {
 284                 bd_p[i] = d_p[i] * a;
 285                 bd_m[i] = d_m[i] * b;
 286         }
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309 #define DO_FEATHER(type, max) \
 310 { \
 311         int frame_w = input->get_w(); \
 312         int frame_h = input->get_h(); \
 313         int size = MAX(frame_w, frame_h); \
 314         float *src = new float[size]; \
 315         float *dst = new float[size]; \
 316         float *val_p = new float[size]; \
 317         float *val_m = new float[size]; \
 318         int start_in = start_out - (int)feather; \
 319         int end_in = end_out + (int)feather; \
 320         if(start_in < 0) start_in = 0; \
 321         if(end_in > frame_h) end_in = frame_h; \
 322         int strip_size = end_in - start_in; \
 323         type **in_rows = (type**)input->get_rows(); \
 324         type **out_rows = (type**)output->get_rows(); \
 325         int j; \
 326  \
 327 /* printf("DO_FEATHER 1\n"); */ \
 328         for(j = 0; j < frame_w; j++) \
 329         { \
 330 /* printf("DO_FEATHER 1.1 %d\n", j); */ \
 331                 bzero(val_p, sizeof(float) * (end_in - start_in)); \
 332                 bzero(val_m, sizeof(float) * (end_in - start_in)); \
 333                 for(int l = 0, k = start_in; k < end_in; l++, k++) \
 334                 { \
 335                         src[l] = (float)in_rows[k][j]; \
 336                 } \
 337  \
 338                 blur_strip(val_p, val_m, dst, src, strip_size, max); \
 339  \
 340                 for(int l = start_out - start_in, k = start_out; k < end_out; l++, k++) \
 341                 { \
 342                         out_rows[k][j] = (type)dst[l]; \
 343                 } \
 344         } \
 345  \
 346         for(j = start_out; j < end_out; j++) \
 347         { \
 348 /* printf("DO_FEATHER 2 %d\n", j); */ \
 349                 bzero(val_p, sizeof(float) * frame_w); \
 350                 bzero(val_m, sizeof(float) * frame_w); \
 351                 for(int k = 0; k < frame_w; k++) \
 352                 { \
 353                         src[k] = (float)out_rows[j][k]; \
 354                 } \
 355  \
 356                 blur_strip(val_p, val_m, dst, src, frame_w, max); \
 357  \
 358                 for(int k = 0; k < frame_w; k++) \
 359                 { \
 360                         out_rows[j][k] = (type)dst[k]; \
 361                 } \
 362         } \
 363  \
 364 /* printf("DO_FEATHER 3\n"); */ \
 365  \
 366         delete [] src; \
 367         delete [] dst; \
 368         delete [] val_p; \
 369         delete [] val_m; \
 370 /* printf("DO_FEATHER 4\n"); */ \
 371 }
 372
 373
 374
 375
 376
 377
 378
 379
 380 //printf("do_feather %d\n", frame->get_color_model());
 381         switch(input->get_color_model())
 382         {
 383                 case BC_A8:
 384                         DO_FEATHER(unsigned char, 0xff);
 385                         break;
 386
 387                 case BC_A16:
 388                         DO_FEATHER(uint16_t, 0xffff);
 389                         break;
 390
 391                 case BC_A_FLOAT:
 392                         DO_FEATHER(float, 1.0f);
 393                         break;
 394         }
 395
 396
 397
 398
 399 }
 400
 401 void MaskUnit::process_package(LoadPackage *package)
 402 {
 403         MaskPackage *ptr = (MaskPackage*)package;
 404
 405         int start_row = SHRT_MIN;         // part for which mask exists
 406         int end_row;
 407         if(engine->recalculate)
 408         {
 409                 VFrame *mask;
 410                 if(engine->feather > 0)
 411                         mask = engine->temp_mask;
 412                 else
 413                         mask = engine->mask;
 414
 415 SET_TRACE
 416 // Generated oversampling frame
 417                 int mask_w = mask->get_w();
 418                 int mask_h = mask->get_h();
 419                 int mask_color_model = mask->get_color_model();
 420                 int oversampled_package_w = mask_w * OVERSAMPLE;
 421                 int oversampled_package_h = (ptr->row2 - ptr->row1) * OVERSAMPLE;
 422 //printf("MaskUnit::process_package 1\n");
 423
 424 SET_TRACE
 425
 426                 int local_first_nonempty_rowspan = SHRT_MIN;
 427                 int local_last_nonempty_rowspan = SHRT_MIN;
 428
 429                 if (!row_spans || row_spans_h != mask_h * OVERSAMPLE) {
 430                         int i;
 431                         if (row_spans) {   /* size change */
 432                                 for (i = 0; i < row_spans_h; i++)
 433                                         free(row_spans[i]);
 434                                 delete [] row_spans;
 435                         }
 436                         row_spans_h = mask_h * OVERSAMPLE;
 437                         row_spans = new short *[mask_h * OVERSAMPLE];
 438                         for (i= 0; i<mask_h * OVERSAMPLE; i++) {
 439                                 /* we use malloc so we can use realloc */
 440                                 row_spans[i] = (short *)malloc(sizeof(short) * NUM_SPANS);
 441                                 /* [0] is initialized later */
 442                                 row_spans[i][1] = NUM_SPANS;
 443                         }
 444                 }
 445
 446 SET_TRACE
 447 //printf("MaskUnit::process_package 1 %d\n", engine->point_sets.total);
 448
 449 SET_TRACE
 450
 451 // Draw bezier curves onto span buffer
 452 //struct timeval start_time;
 453 //gettimeofday(&start_time, 0);
 454
 455                 for(int k = 0; k < engine->point_sets.total; k++)
 456                 {
 457                         int old_x, old_y;
 458                         old_x = SHRT_MIN; // sentinel
 459                         ArrayList<MaskPoint*> *points = engine->point_sets.values[k];
 460
 461                         if(points->total < 2) continue;
 462 //printf("MaskUnit::process_package 2 %d %d\n", k, points->total);
 463                         for (int i = ptr->row1 * OVERSAMPLE; i < ptr->row2 * OVERSAMPLE; i++)
 464                                 row_spans[i][0] = 2; /* initialize to zero */
 465                         (ptr->row1*OVERSAMPLE, ptr->row2*OVERSAMPLE); // init just my rows
 466                         for(int i = 0; i < points->total; i++)
 467                         {
 468                                 MaskPoint *point1 = points->values[i];
 469                                 MaskPoint *point2 = (i >= points->total - 1) ?
 470                                         points->values[0] :
 471                                         points->values[i + 1];
 472
 473                                 float x0 = point1->x;
 474                                 float y0 = point1->y;
 475                                 float x1 = point1->x + point1->control_x2;
 476                                 float y1 = point1->y + point1->control_y2;
 477                                 float x2 = point2->x + point2->control_x1;
 478                                 float y2 = point2->y + point2->control_y1;
 479                                 float x3 = point2->x;
 480                                 float y3 = point2->y;
 481
 482                                 // possible optimization here... since these coordinates are bounding box for curve
 483                                 // we can continue with next curve if they are out of our range
 484
 485                                 // forward differencing bezier curves implementation taken from GPL code at
 486                                 // http://cvs.sourceforge.net/viewcvs.py/guliverkli/guliverkli/src/subtitles/Rasterizer.cpp?rev=1.3
 487
 488
 489
 490                                 float cx3, cx2, cx1, cx0, cy3, cy2, cy1, cy0;
 491
 492
 493                                 // [-1 +3 -3 +1]
 494                                 // [+3 -6 +3  0]
 495                                 // [-3 +3  0  0]
 496                                 // [+1  0  0  0]
 497
 498                                 cx3 = (-  x0 + 3*x1 - 3*x2 + x3) * OVERSAMPLE;
 499                                 cx2 = ( 3*x0 - 6*x1 + 3*x2) * OVERSAMPLE;
 500                                 cx1 = (-3*x0 + 3*x1) * OVERSAMPLE;
 501                                 cx0 = (   x0) * OVERSAMPLE;
 502
 503                                 cy3 = (-  y0 + 3*y1 - 3*y2 + y3) * OVERSAMPLE;
 504                                 cy2 = ( 3*y0 - 6*y1 + 3*y2) * OVERSAMPLE;
 505                                 cy1 = (-3*y0 + 3*y1) * OVERSAMPLE;
 506                                 cy0 = (   y0 - ptr->row1) * OVERSAMPLE;
 507
 508                                 float maxaccel1 = fabs(2*cy2) + fabs(6*cy3);
 509                                 float maxaccel2 = fabs(2*cx2) + fabs(6*cx3);
 510
 511                                 float maxaccel = maxaccel1 > maxaccel2 ? maxaccel1 : maxaccel2;
 512                                 float h = 1.0;
 513
 514                                 if(maxaccel > 8.0 * OVERSAMPLE) h = sqrt((8.0 * OVERSAMPLE) / maxaccel);
 515
 516                                 for(float t = 0.0; t < 1.0; t += h)
 517                                 {
 518                                         int x = (int) (cx0 + t*(cx1 + t*(cx2 + t*cx3)));
 519                                         int y = (int) (cy0 + t*(cy1 + t*(cy2 + t*cy3)));
 520
 521                                         if (old_x != SHRT_MIN)
 522                                                 draw_line_clamped(old_x, old_y, x, y, oversampled_package_w, oversampled_package_h, ptr->row1 * OVERSAMPLE);
 523                                         old_x = x;
 524                                         old_y = y;
 525                                 }
 526
 527                                 int x = (int)(x3 * OVERSAMPLE);
 528                                 int y = (int)((y3 - ptr->row1) * OVERSAMPLE);
 529                                 draw_line_clamped(old_x, old_y, x, y, oversampled_package_w, oversampled_package_h, ptr->row1 * OVERSAMPLE);
 530                                 old_x = (int)x;
 531                                 old_y = (int)y;
 532
 533                         }
 534 //printf("MaskUnit::process_package 1\n");
 535
 536                         // Now we have ordered spans ready!
 537                         //printf("Segment : %i , row1: %i\n", oversampled_package_h, ptr->row1);
 538                         uint16_t value;
 539                         if (mask_color_model == BC_A8)
 540                                 value = (int)((float)engine->value / 100 * 0xff);
 541                         else
 542                                 value = (int)((float)engine->value / 100 * 0xffff);     // also for BC_A_FLOAT
 543
 544                         /* Scaneline sampling, inspired by Graphics gems I, page 81 */
 545                         for (int i = ptr->row1; i < ptr->row2; i++)
 546                         {
 547                                 short min_x = SHRT_MAX;
 548                                 short max_x = SHRT_MIN;
 549                                 int j;                          /* universal counter for 0..OVERSAMPLE-1 */
 550                                 short *span;                    /* current span - set inside loops with j */
 551                                 short span_p[OVERSAMPLE];       /* pointers to current positions in spans */
 552                                 #define P (span_p[j])           /* current span pointer */
 553                                 #define MAXP (span[0])          /* current span length */
 554                                 int num_empty_spans = 0;
 555                                 /* get the initial span pointers ready */
 556                                 for (j = 0; j < OVERSAMPLE; j++)
 557                                 {
 558                                         span = row_spans[j + i * OVERSAMPLE];
 559                                         P = 2;              /* starting pointers to spans */
 560                                                 /* hypotetical hypotetical fix goes here: take care that there is maximum one empty span for every subpixel */
 561                                         if (MAXP != 2) {                                        /* if span is not empty */
 562                                                 if (span[2] < min_x) min_x = span[2];           /* take start of the first span */
 563                                                 if (span[MAXP-1] > max_x) max_x = span[MAXP-1]; /* and end of last */
 564                                         } else
 565                                         {       /* span is empty */
 566                                                 num_empty_spans ++;
 567                                         }
 568                                 }
 569                                 if (num_empty_spans == OVERSAMPLE)
 570                                         continue; /* no work for us here */
 571                                 else
 572                                 {       /* if we have engaged first nonempty rowspan... remember it to speed up mask applying */
 573                                         if (local_first_nonempty_rowspan < 0 || i < local_first_nonempty_rowspan)
 574                                                 local_first_nonempty_rowspan = i;
 575                                         if (i > local_last_nonempty_rowspan) local_last_nonempty_rowspan = i;
 576                                 }
 577                                 /* we have some pixels to fill, do coverage calculation for span */
 578
 579                                 void *output_row = (unsigned char*)mask->get_rows()[i];
 580                                 min_x = min_x / OVERSAMPLE;
 581                                 max_x = (max_x + OVERSAMPLE - 1) / OVERSAMPLE;
 582
 583                                 /* printf("row %i, pixel range: %i %i, spans0: %i\n", i, min_x, max_x, row_spans[i*OVERSAMPLE][0]-2); */
 584
 585                                 /* this is not a full loop, since we jump trough h if possible */
 586                                 for (int h = min_x; h <= max_x; h++)
 587                                 {
 588                                         short pixelleft = h * OVERSAMPLE;  /* leftmost subpixel of pixel*/
 589                                         short pixelright = pixelleft + OVERSAMPLE - 1; /* rightmost subpixel of pixel */
 590                                         uint32_t coverage = 0;
 591                                         int num_left = 0;               /* number of spans that have start left of the next pixel */
 592                                         short right_end = SHRT_MAX;     /* leftmost end of any span - right end of a full scanline */
 593                                         short right_start = SHRT_MAX;   /* leftmost start of any span - left end of empty scanline */
 594
 595                                         for (j=0; j< OVERSAMPLE; j++)
 596                                         {
 597                                                 char chg = 1;
 598                                                 span = row_spans[j + i * OVERSAMPLE];
 599                                                 while (P < MAXP && chg)
 600                                                 {
 601                                                 //      printf("Sp: %i %i\n", span[P], span[P+1]);
 602                                                         if (span[P] == span[P+1])           /* ignore empty spans */
 603                                                         {
 604                                                                 P +=2;
 605                                                                 continue;
 606                                                         }
 607                                                         if (span[P] <= pixelright)          /* if span start is before the end of pixel */
 608                                                                 coverage += MIN(span[P+1], pixelright)  /* 'clip' the span to pixel */
 609                                                                           - MAX(span[P], pixelleft) + 1;
 610                                                         if (span[P+1] <= pixelright)
 611                                                                 P += 2;
 612                                                         else
 613                                                                 chg = 0;
 614                                                 }
 615                                                 if (P == MAXP)
 616                                                         num_left = -OVERSAMPLE; /* just take care that num_left cannot equal OVERSAMPLE or zero again */
 617                                                 else
 618                                                 {
 619                                                         if (span[P] <= pixelright)  /* if span starts before subpixel in the pixel on the right */
 620                                                         {    /* useful for determining filled space till next non-fully-filled pixel */
 621                                                                 num_left ++;
 622                                                                 if (span[P+1] < right_end) right_end = span[P+1];
 623                                                         } else
 624                                                         {    /* useful for determining empty space till next non-empty pixel */
 625                                                                 if (span[P] < right_start) right_start = span[P];
 626                                                         }
 627                                                 }
 628                                         }
 629                                         // calculate coverage
 630                                         coverage *= value;
 631                                         coverage /= OVERSAMPLE * OVERSAMPLE;
 632
 633                                         // when we have multiple masks the highest coverage wins
 634                                         switch (mask_color_model)
 635                                         {
 636                                         case BC_A8:
 637                                                 if (((unsigned char *) output_row)[h] < coverage)
 638                                                         ((unsigned char*)output_row)[h] = coverage;
 639                                                 break;
 640                                         case BC_A16:
 641                                                 if (((uint16_t *) output_row)[h] < coverage)
 642                                                         ((uint16_t *) output_row)[h] = coverage;
 643                                                 break;
 644                                         case BC_A_FLOAT:
 645                                                 if (((float *) output_row)[h] < coverage/float(0xffff))
 646                                                         ((float *) output_row)[h] = coverage/float(0xffff);
 647                                                 break;
 648                                         }
 649                                         /* possible optimization: do joining of multiple masks by span logics, not by bitmap logics*/
 650
 651                                         if (num_left == OVERSAMPLE)
 652                                         {
 653                                                 /* all current spans start more left than next pixel */
 654                                                 /* this means we can probably (if lucky) draw a longer horizontal line */
 655                                                 right_end = (right_end / OVERSAMPLE) - 1; /* last fully covered pixel */
 656                                                 if (right_end > h)
 657                                                 {
 658                                                         if (mask_color_model == BC_A8)
 659                                                                 memset((char *)output_row + h + 1, value, right_end - h);
 660                                                         else {
 661                                                                 /* we are fucked, since there is no 16bit memset */
 662                                                                 if (mask_color_model == BC_A16) {
 663                                                                         for (int z = h +1; z <= right_end; z++)
 664                                                                                 ((uint16_t *) output_row)[z] =  value;
 665                                                                 } else {
 666                                                                         for (int z = h +1; z <= right_end; z++)
 667                                                                                 ((float *) output_row)[z] =  value/float(0xffff);
 668                                                                 }
 669                                                         }
 670                                                         h = right_end;
 671                                                 }
 672                                         } else
 673                                         if (num_left == 0)
 674                                         {
 675                                                 /* all current spans start right of next pixel */
 676                                                 /* this means we can probably (if lucky) skip some pixels */
 677                                                 right_start = (right_start / OVERSAMPLE) - 1; /* last fully empty pixel */
 678                                                 if (right_start > h)
 679                                                 {
 680                                                         h = right_start;
 681                                                 }
 682                                         }
 683                                 }
 684                         }
 685                 }
 686                 engine->protect_data.lock();
 687                 if (local_first_nonempty_rowspan < engine->first_nonempty_rowspan)
 688                         engine->first_nonempty_rowspan = local_first_nonempty_rowspan;
 689                 if (local_last_nonempty_rowspan > engine->last_nonempty_rowspan)
 690                         engine->last_nonempty_rowspan = local_last_nonempty_rowspan;
 691                 engine->protect_data.unlock();
 692
 693
 694 //              int64_t dif= get_difference(&start_time);
 695 //              printf("diff: %lli\n", dif);
 696         }       /* END OF RECALCULATION! */
 697
 698 SET_TRACE
 699
 700         /* possible optimization: this could be useful for do_feather also */
 701
 702         // Feather polygon
 703         if(engine->recalculate && engine->feather > 0)
 704         {
 705                 /* first take care that all packages are already drawn onto mask */
 706                 pthread_mutex_lock(&engine->stage1_finished_mutex);
 707                 engine->stage1_finished_count ++;
 708                 if (engine->stage1_finished_count == engine->get_total_packages())
 709                 {
 710                         // let others pass
 711                         pthread_cond_broadcast(&engine->stage1_finished_cond);
 712                 }
 713                 else
 714                 {
 715                         // wait until all are finished
 716                         while (engine->stage1_finished_count < engine->get_total_packages())
 717                                 pthread_cond_wait(&engine->stage1_finished_cond, &engine->stage1_finished_mutex);
 718                 }
 719                 pthread_mutex_unlock(&engine->stage1_finished_mutex);
 720
 721                 /* now do the feather */
 722 //printf("MaskUnit::process_package 3 %f\n", engine->feather);
 723
 724         struct timeval start_time;
 725         gettimeofday(&start_time, 0);
 726
 727         /*
 728         {
 729         // EXPERIMENTAL CODE to find out how values between old and new do_feather map
 730         // create a testcase and find out the closest match between do_feather_2 at 3 and do_feather
 731         //                      2       3       4       5       6       7       8       10      13      15
 732         // do_feather_2         3       5       7       9       11      13      15      19      25      29
 733         // do_feather_1         2.683   3.401   4.139   4.768   5.315   5.819   6.271   7.093   8.170   8.844
 734         // diff                         0.718   0.738   0.629   0.547   0.504   0.452
 735         // {(2,2.683),(3,3.401),(4,4.139),(5,4.768),(6,5.315),(7,5.819),(8,6.271),(10,7.093),(13,8.170),(15,8.844)}
 736         // use http://mss.math.vanderbilt.edu/cgi-bin/MSSAgent/~pscrooke/MSS/fitpoly.def
 737         // for calculating the coefficients
 738
 739                 VFrame *df2 = new VFrame (*engine->mask);
 740                 VFrame *one_sample = new VFrame(*engine->mask);
 741                 do_feather_2(df2,
 742                         engine->temp_mask,
 743                         25,
 744                         ptr->row1,
 745                         ptr->row2);
 746                 float ftmp;
 747                 for (ftmp = 8.15; ftmp <8.18; ftmp += 0.001)
 748                 {
 749                         do_feather(one_sample,
 750                         engine->temp_mask,
 751                         ftmp,
 752                         ptr->row1,
 753                         ptr->row2);
 754                         double squarediff = 0;
 755                         for (int i=0; i< engine->mask->get_h(); i++)
 756                                 for (int j = 0; j< engine->mask->get_w(); j++)
 757                                 {
 758                                         double v1= ((unsigned char *)one_sample->get_rows()[i])[j];
 759                                         double v2= ((unsigned char *)df2->get_rows()[i])[j];
 760                                         squarediff += (v1-v2)*(v1-v2);
 761                                 }
 762                         squarediff = sqrt(squarediff);
 763                         printf("for value 3: ftmp: %2.3f, squarediff: %f\n", ftmp, squarediff);
 764                 }
 765         }
 766         */
 767
 768                 int done = 0;
 769                 done = do_feather_2(engine->mask,        // try if we have super fast implementation ready
 770                                 engine->temp_mask,
 771                                 engine->feather * 2 - 1,
 772                                 ptr->row1,
 773                                 ptr->row2);
 774                 if (done) {
 775                         engine->realfeather = engine->feather;
 776                 }
 777                 if (!done)
 778                 {
 779                 //      printf("not done\n");
 780                         float feather = engine->feather;
 781                         engine->realfeather = 0.878441 + 0.988534*feather - 0.0490204 *feather*feather  + 0.0012359 *feather*feather*feather;
 782                         do_feather(engine->mask,
 783                                 engine->temp_mask,
 784                                 engine->realfeather,
 785                                 ptr->row1,
 786                                 ptr->row2);
 787                 }
 788                 int64_t dif= get_difference(&start_time);
 789                 printf("diff: %lli\n", dif);
 790         } else
 791         if (engine->feather <= 0) {
 792                 engine->realfeather = 0;
 793         }
 794         start_row = MAX (ptr->row1, engine->first_nonempty_rowspan - (int)ceil(engine->realfeather));
 795         end_row = MIN (ptr->row2, engine->last_nonempty_rowspan + 1 + (int)ceil(engine->realfeather));
 796
 797
 798
 799 // Apply mask
 800
 801
 802 /* use the info about first and last column that are coloured from rowspan!  */
 803 /* possible optimisation: also remember total spans */
 804 /* possible optimisation: lookup for  X * (max - *mask_row) / max, where max is known mask_row and X are variabiles */
 805 #define APPLY_MASK_SUBTRACT_ALPHA(type, max, components, do_yuv) \
 806 { \
 807         type chroma_offset = (max + 1) / 2; \
 808         for(int i = start_row; i < end_row; i++) \
 809         { \
 810         type *output_row = (type*)engine->output->get_rows()[i]; \
 811         type *mask_row = (type*)engine->mask->get_rows()[i]; \
 812         \
 813  \
 814         for(int j  = 0; j < mask_w; j++) \
 815         { \
 816                 if(components == 4) \
 817                 { \
 818                         output_row[3] = output_row[3] * (max - *mask_row) / max; \
 819                 } \
 820                 else \
 821                 { \
 822                         output_row[0] = output_row[0] * (max - *mask_row) / max; \
 823  \
 824                         output_row[1] = output_row[1] * (max - *mask_row) / max; \
 825                         output_row[2] = output_row[2] * (max - *mask_row) / max; \
 826  \
 827                         if(do_yuv) \
 828                         { \
 829                                 output_row[1] += chroma_offset * *mask_row / max; \
 830                                 output_row[2] += chroma_offset * *mask_row / max; \
 831                         } \
 832                 } \
 833                 output_row += components; \
 834                 mask_row += 1;           \
 835         } \
 836         } \
 837 }
 838
 839 #define APPLY_MASK_MULTIPLY_ALPHA(type, max, components, do_yuv) \
 840 { \
 841         type chroma_offset = (max + 1) / 2; \
 842                 for(int i = ptr->row1; i < ptr->row2; i++) \
 843                 { \
 844         type *output_row = (type*)engine->output->get_rows()[i]; \
 845         type *mask_row = (type*)engine->mask->get_rows()[i]; \
 846  \
 847         if (components == 4) output_row += 3; \
 848         for(int j  = mask_w; j != 0;  j--) \
 849         { \
 850                 if(components == 4) \
 851                 { \
 852                         *output_row = *output_row * *mask_row / max; \
 853                 } \
 854                 else \
 855                 { \
 856                         output_row[0] = output_row[3] * *mask_row / max; \
 857  \
 858                         output_row[1] = output_row[1] * *mask_row / max; \
 859                         output_row[2] = output_row[2] * *mask_row / max; \
 860  \
 861                         if(do_yuv) \
 862                         { \
 863                                 output_row[1] += chroma_offset * (max - *mask_row) / max; \
 864                                 output_row[2] += chroma_offset * (max - *mask_row) / max; \
 865                         } \
 866                 } \
 867                 output_row += components; \
 868                 mask_row += 1;           \
 869         } \
 870         } \
 871 }
 872
 873
 874 //struct timeval start_time;
 875 //gettimeofday(&start_time, 0);
 876
 877 //printf("MaskUnit::process_package 1 %d\n", engine->mode);
 878         int mask_w = engine->mask->get_w();
 879         switch(engine->mode)
 880         {
 881                 case MASK_MULTIPLY_ALPHA:
 882                         switch(engine->output->get_color_model())
 883                         {
 884                                 case BC_RGB888:
 885                                         APPLY_MASK_MULTIPLY_ALPHA(unsigned char, 0xff, 3, 0);
 886                                         break;
 887                                 case BC_YUV888:
 888                                         APPLY_MASK_MULTIPLY_ALPHA(unsigned char, 0xff, 3, 1);
 889                                         break;
 890                                 case BC_YUVA8888:
 891                                 case BC_RGBA8888:
 892                                         APPLY_MASK_MULTIPLY_ALPHA(unsigned char, 0xff, 4, 0);
 893                                         break;
 894                                 case BC_RGB161616:
 895                                         APPLY_MASK_MULTIPLY_ALPHA(uint16_t, 0xffff, 3, 0);
 896                                         break;
 897                                 case BC_YUV161616:
 898                                         APPLY_MASK_MULTIPLY_ALPHA(uint16_t, 0xffff, 3, 1);
 899                                         break;
 900                                 case BC_YUVA16161616:
 901                                 case BC_RGBA16161616:
 902                                         APPLY_MASK_MULTIPLY_ALPHA(uint16_t, 0xffff, 4, 0);
 903                                         break;
 904                                 case BC_RGB_FLOAT:
 905                                         APPLY_MASK_MULTIPLY_ALPHA(float, 1.0f, 3, 0);
 906                                         break;
 907                                 case BC_RGBA_FLOAT:
 908                                         APPLY_MASK_MULTIPLY_ALPHA(float, 1.0f, 4, 0);
 909                                         break;
 910                         }
 911                         break;
 912
 913                 case MASK_SUBTRACT_ALPHA:
 914                         switch(engine->output->get_color_model())
 915                         {
 916                                 case BC_RGB888:
 917                                         APPLY_MASK_SUBTRACT_ALPHA(unsigned char, 0xff, 3, 0);
 918                                         break;
 919                                 case BC_YUV888:
 920                                         APPLY_MASK_SUBTRACT_ALPHA(unsigned char, 0xff, 3, 1);
 921                                         break;
 922                                 case BC_YUVA8888:
 923                                 case BC_RGBA8888:
 924                                         APPLY_MASK_SUBTRACT_ALPHA(unsigned char, 0xff, 4, 0);
 925                                         break;
 926                                 case BC_RGB161616:
 927                                         APPLY_MASK_SUBTRACT_ALPHA(uint16_t, 0xffff, 3, 0);
 928                                         break;
 929                                 case BC_YUV161616:
 930                                         APPLY_MASK_SUBTRACT_ALPHA(uint16_t, 0xffff, 3, 1);
 931                                         break;
 932                                 case BC_YUVA16161616:
 933                                 case BC_RGBA16161616:
 934                                         APPLY_MASK_SUBTRACT_ALPHA(uint16_t, 0xffff, 4, 0);
 935                                         break;
 936                                 case BC_RGB_FLOAT:
 937                                         APPLY_MASK_SUBTRACT_ALPHA(float, 1.0f, 3, 0);
 938                                         break;
 939                                 case BC_RGBA_FLOAT:
 940                                         APPLY_MASK_SUBTRACT_ALPHA(float, 1.0f, 4, 0);
 941                                         break;
 942                         }
 943                         break;
 944         }
 945 //      int64_t dif= get_difference(&start_time);
 946 //      printf("diff: %lli\n", dif);
 947 //printf("diff2: %lli\n", get_difference(&start_time));
 948 //printf("MaskUnit::process_package 4 %d\n", get_package_number());
 949 }
 950
 951
 952
 953
 954
 955 MaskEngine::MaskEngine(int cpus)
 956  : LoadServer(cpus, cpus )      /* these two HAVE to be the same, since packages communicate  */
 957 // : LoadServer(1, 2)
 958 {
 959         mask = 0;
 960         pthread_mutex_init(&stage1_finished_mutex, NULL);
 961         pthread_cond_init(&stage1_finished_cond, NULL);
 962 }
 963
 964 MaskEngine::~MaskEngine()
 965 {
 966         pthread_cond_destroy(&stage1_finished_cond);
 967         pthread_mutex_destroy(&stage1_finished_mutex);
 968         if(mask)
 969         {
 970                 delete mask;
 971                 delete temp_mask;
 972         }
 973
 974         for(int i = 0; i < point_sets.total; i++)
 975         {
 976                 ArrayList<MaskPoint*> *points = point_sets.values[i];
 977                 points->remove_all_objects();
 978         }
 979         point_sets.remove_all_objects();
 980 }
 981
 982 int MaskEngine::points_equivalent(ArrayList<MaskPoint*> *new_points,
 983         ArrayList<MaskPoint*> *points)
 984 {
 985 //printf("MaskEngine::points_equivalent %d %d\n", new_points->total, points->total);
 986         if(new_points->total != points->total) return 0;
 987
 988         for(int i = 0; i < new_points->total; i++)
 989         {
 990                 if(!(*new_points->values[i] == *points->values[i])) return 0;
 991         }
 992
 993         return 1;
 994 }
 995
 996 void MaskEngine::do_mask(VFrame *output,
 997         int64_t start_position,
 998         double frame_rate,
 999         double project_frame_rate,
1000         MaskAutos *keyframe_set,
1001         int direction,
1002         int before_plugins)
1003 {
1004         int64_t start_position_project = (int64_t)(start_position *
1005                 project_frame_rate /
1006                 frame_rate);
1007         Auto *current = 0;
1008         MaskAuto *default_auto = (MaskAuto*)keyframe_set->default_auto;
1009         MaskAuto *keyframe = (MaskAuto*)keyframe_set->get_prev_auto(start_position_project,
1010                 direction,
1011                 current);
1012
1013         if (keyframe->apply_before_plugins != before_plugins)
1014                 return;
1015
1016
1017         int total_points = 0;
1018         for(int i = 0; i < keyframe->masks.total; i++)
1019         {
1020                 SubMask *mask = keyframe->get_submask(i);
1021                 int submask_points = mask->points.total;
1022                 if(submask_points > 1) total_points += submask_points;
1023         }
1024
1025 //printf("MaskEngine::do_mask 1 %d %d\n", total_points, keyframe->value);
1026 // Ignore certain masks
1027         if(total_points < 2 ||
1028                 (keyframe->value == 0 && default_auto->mode == MASK_SUBTRACT_ALPHA))
1029         {
1030                 return;
1031         }
1032
1033 // Fake certain masks
1034         if(keyframe->value == 0 && default_auto->mode == MASK_MULTIPLY_ALPHA)
1035         {
1036                 output->clear_frame();
1037                 return;
1038         }
1039
1040 //printf("MaskEngine::do_mask 1\n");
1041
1042         int new_color_model = 0;
1043         recalculate = 0;
1044
1045         switch(output->get_color_model())
1046         {
1047                 case BC_RGB_FLOAT:
1048                 case BC_RGBA_FLOAT:
1049                         new_color_model = BC_A_FLOAT;
1050                         break;
1051
1052                 case BC_RGB888:
1053                 case BC_RGBA8888:
1054                 case BC_YUV888:
1055                 case BC_YUVA8888:
1056                         new_color_model = BC_A8;
1057                         break;
1058
1059                 case BC_RGB161616:
1060                 case BC_RGBA16161616:
1061                 case BC_YUV161616:
1062                 case BC_YUVA16161616:
1063                         new_color_model = BC_A16;
1064                         break;
1065         }
1066
1067 // Determine if recalculation is needed
1068 SET_TRACE
1069
1070         if(mask &&
1071                 (mask->get_w() != output->get_w() ||
1072                 mask->get_h() != output->get_h() ||
1073                 mask->get_color_model() != new_color_model))
1074         {
1075                 delete mask;
1076                 delete temp_mask;
1077                 mask = 0;
1078                 recalculate = 1;
1079         }
1080
1081         if(!recalculate)
1082         {
1083                 if(point_sets.total != keyframe_set->total_submasks(start_position_project,
1084                         direction))
1085                         recalculate = 1;
1086         }
1087
1088         if(!recalculate)
1089         {
1090                 for(int i = 0;
1091                         i < keyframe_set->total_submasks(start_position_project,
1092                                 direction) && !recalculate;
1093                         i++)
1094                 {
1095                         ArrayList<MaskPoint*> *new_points = new ArrayList<MaskPoint*>;
1096                         keyframe_set->get_points(new_points,
1097                                 i,
1098                                 start_position_project,
1099                                 direction);
1100                         if(!points_equivalent(new_points, point_sets.values[i])) recalculate = 1;
1101                         new_points->remove_all_objects();
1102                         delete new_points;
1103                 }
1104         }
1105
1106         if(recalculate ||
1107                 !EQUIV(keyframe->feather, feather) ||
1108                 !EQUIV(keyframe->value, value))
1109         {
1110                 recalculate = 1;
1111                 if(!mask)
1112                 {
1113                         mask = new VFrame(0,
1114                                         output->get_w(),
1115                                         output->get_h(),
1116                                         new_color_model);
1117                         temp_mask = new VFrame(0,
1118                                         output->get_w(),
1119                                         output->get_h(),
1120                                         new_color_model);
1121                 }
1122                 if(keyframe->feather > 0)
1123                         temp_mask->clear_frame();
1124                 else
1125                         mask->clear_frame();
1126
1127                 for(int i = 0; i < point_sets.total; i++)
1128                 {
1129                         ArrayList<MaskPoint*> *points = point_sets.values[i];
1130                         points->remove_all_objects();
1131                 }
1132                 point_sets.remove_all_objects();
1133
1134                 for(int i = 0;
1135                         i < keyframe_set->total_submasks(start_position_project,
1136                                 direction);
1137                         i++)
1138                 {
1139                         ArrayList<MaskPoint*> *new_points = new ArrayList<MaskPoint*>;
1140                         keyframe_set->get_points(new_points,
1141                                 i,
1142                                 start_position_project,
1143                                 direction);
1144                         point_sets.append(new_points);
1145                 }
1146         }
1147
1148
1149
1150         this->output = output;
1151         this->mode = default_auto->mode;
1152         this->feather = keyframe->feather;
1153         this->value = keyframe->value;
1154
1155
1156 // Run units
1157 SET_TRACE
1158         process_packages();
1159 SET_TRACE
1160
1161
1162 }
1163
1164 void MaskEngine::init_packages()
1165 {
1166 SET_TRACE
1167 //printf("MaskEngine::init_packages 1\n");
1168         int division = (int)((float)output->get_h() / (get_total_packages()) + 0.5);
1169         if(division < 1) division = 1;
1170
1171         stage1_finished_count = 0;
1172         if (recalculate) {
1173                 last_nonempty_rowspan = SHRT_MIN;
1174                 first_nonempty_rowspan = SHRT_MAX;
1175         }
1176 SET_TRACE
1177 // Always a multiple of 2 packages exist
1178         for(int i = 0; i < get_total_packages(); i++)
1179         {
1180                 MaskPackage *pkg = (MaskPackage*)get_package(i);
1181                 pkg->row1 = division * i;
1182                 pkg->row2 = MIN (division * i + division, output->get_h());
1183
1184                 if(i == get_total_packages() - 1)  // last package
1185                 {
1186                         pkg->row2 = pkg->row2 = output->get_h();
1187                 }
1188
1189         }
1190 SET_TRACE
1191 //printf("MaskEngine::init_packages 2\n");
1192 }
1193
1194 LoadClient* MaskEngine::new_client()
1195 {
1196         return new MaskUnit(this);
1197 }
1198
1199 LoadPackage* MaskEngine::new_package()
1200 {
1201         return new MaskPackage;
1202 }
1203