cinelerra/maskengine.C

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