git.sesse.net Git - qscale/blob - libqscale.c

   1 /*
   2  * qscale: Quick, high-quality JPEG-to-JPEG scaler.
   3  * Copyright (C) 2008 Steinar H. Gunderson <sgunderson@bigfoot.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation, version 2 of the License.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA
  17  */
  18
  19 #include <math.h>
  20 #include <stdio.h>
  21 #include <stdlib.h>
  22 #include <string.h>
  23 #include <malloc.h>
  24
  25 #include "libqscale.h"
  26
  27 /* The number of pixels to process at a time when scaling vertically. */
  28 #define CACHE_LINE_FACTOR 16
  29
  30 /* Whether to use SSE for horizontal scaling or not (requires SSE3). */
  31 #define USE_HORIZONTAL_SSE 1
  32
  33 /* Whether to use SSE for vertical scaling or not (requires only SSE1). */
  34 #define USE_VERTICAL_SSE 1
  35
  36 #if USE_VERTICAL_SSE
  37 #undef CACHE_LINE_FACTOR
  38 #define CACHE_LINE_FACTOR 16
  39 #endif
  40
  41 #ifndef M_PI
  42 #define M_PI 3.14159265358979323846264
  43 #endif
  44
  45 qscale_img *qscale_load_jpeg(const char *filename)
  46 {
  47         FILE *file = fopen(filename, "rb");
  48         qscale_img *img;
  49         if (file == NULL) {
  50                 return NULL;
  51         }
  52
  53         img = qscale_load_jpeg_from_stdio(file);
  54
  55         fclose(file);
  56         return img;
  57 }
  58
  59 qscale_img *qscale_load_jpeg_from_stdio(FILE *file)
  60 {
  61         qscale_img *img = (qscale_img *)malloc(sizeof(qscale_img));
  62         if (img == NULL) {
  63                 return NULL;
  64         }
  65
  66         img->data_y = img->data_cb = img->data_cr = NULL;
  67
  68         /* FIXME: Better error handling here (ie., return NULL). */
  69         struct jpeg_decompress_struct dinfo;
  70         struct jpeg_error_mgr jerr;
  71         dinfo.err = jpeg_std_error(&jerr);
  72         jpeg_create_decompress(&dinfo);
  73         jpeg_stdio_src(&dinfo, file);
  74         jpeg_read_header(&dinfo, TRUE);
  75         dinfo.raw_data_out = TRUE;
  76         jpeg_start_decompress(&dinfo);
  77
  78         /* We do not handle anything but YCbCr images (yet?). */
  79         if (dinfo.num_components != 3) {
  80                 qscale_destroy(img);
  81                 return NULL;
  82         }
  83
  84         img->width = dinfo.image_width;
  85         img->height = dinfo.image_height;
  86
  87         img->w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
  88         img->h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
  89
  90         img->w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
  91         img->h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
  92
  93         img->w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
  94         img->h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
  95
  96         img->samp_h0 = dinfo.comp_info[0].h_samp_factor;
  97         img->samp_v0 = dinfo.comp_info[0].v_samp_factor;
  98
  99         img->samp_h1 = dinfo.comp_info[1].h_samp_factor;
 100         img->samp_v1 = dinfo.comp_info[1].v_samp_factor;
 101
 102         img->samp_h2 = dinfo.comp_info[2].h_samp_factor;
 103         img->samp_v2 = dinfo.comp_info[2].v_samp_factor;
 104
 105         img->data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 106         img->data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 107         img->data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 108
 109         if (img->data_y == NULL || img->data_cb == NULL || img->data_cr == NULL) {
 110                 qscale_destroy(img);
 111                 return NULL;
 112         }
 113
 114         int total_lines = 0, blocks = 0;
 115         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 116                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 117
 118                 JSAMPROW y_row_ptrs[max_lines];
 119                 JSAMPROW cb_row_ptrs[max_lines];
 120                 JSAMPROW cr_row_ptrs[max_lines];
 121                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 122
 123                 int i;
 124                 for (i = 0; i < max_lines; ++i) {
 125                         y_row_ptrs[i]  = img->data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
 126                         cb_row_ptrs[i] = img->data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
 127                         cr_row_ptrs[i] = img->data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
 128                 }
 129
 130                 total_lines += max_lines;
 131                 ++blocks;
 132
 133                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 134                         break;
 135         }
 136
 137         jpeg_destroy_decompress(&dinfo);
 138         return img;
 139 }
 140
 141 void qscale_destroy(qscale_img *img)
 142 {
 143         free(img->data_y);
 144         free(img->data_cb);
 145         free(img->data_cr);
 146         free(img);
 147 }
 148
 149
 150 static double sinc(double x)
 151 {
 152         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
 153
 154         if (abs(x) < cutoff) {
 155                 /* For small |x|, use Taylor series instead */
 156                 const double x2 = x * x;
 157                 const double x4 = x2 * x2;
 158
 159                 return 1.0 - x2 / 6.0 + x4 / 120.0;
 160         } else {
 161                 return sin(x) / x;
 162         }
 163 }
 164
 165 static double lanczos_tap(double x)
 166 {
 167         if (x < -3.0 || x > 3.0)
 168                 return 0.0;
 169         if (x < 0.0)
 170                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
 171         else
 172                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
 173 }
 174
 175 static double mitchell_tap(double x)
 176 {
 177         const double b = 1.0 / 3.0;
 178         const double c = 1.0 / 3.0;
 179         const double p0 = (  6.0 -  2.0*b         ) / 6.0;
 180         const double p2 = (-18.0 + 12.0*b +  6.0*c) / 6.0;
 181         const double p3 = ( 12.0 -  9.0*b -  6.0*c) / 6.0;
 182         const double q0 = (         8.0*b + 24.0*c) / 6.0;
 183         const double q1 = (      - 12.0*b - 48.0*c) / 6.0;
 184         const double q2 = (         6.0*b + 30.0*c) / 6.0;
 185         const double q3 = (      -      b -  6.0*c) / 6.0;
 186
 187         if (x < -2.0) {
 188                 return 0.0;
 189         } else if (x < -1.0) {
 190                 return q0 - x * (q1 - x * (q2 - x * q3));
 191         } else if (x < 0.0) {
 192                 return p0 + x * x * (p2 - x * p3);
 193         } else if (x < 1.0) {
 194                 return p0 + x * x * (p2 + x * p3);
 195         } else if (x < 2.0) {
 196                 return q0 + x * (q1 + x * (q2 + x * q3));
 197         } else {
 198                 return 0.0;
 199         }
 200 }
 201
 202 struct pix_desc {
 203         unsigned start, end;
 204         unsigned startcoeff;
 205 };
 206
 207 static void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride, enum qscale_scaling_filter scaling_filter)
 208 {
 209         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
 210         int size_coeffs = 8;
 211         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 212         int num_coeffs = 0;
 213         int x, y;
 214         double sf = (double)w / (double)nw;
 215         double support;
 216
 217         if (scaling_filter == LANCZOS) {
 218                 support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
 219         } else {  /* Mitchell */
 220                 support = (w > nw) ? (2.0 * sf) : (2.0 / sf);
 221         }
 222
 223         /* calculate the filter */
 224         for (x = 0; x < nw; ++x) {
 225                 int start = ceil(x * sf - support);
 226                 int end = floor(x * sf + support);
 227                 int sx;
 228                 double sum = 0.0;
 229
 230                 if (start < 0) {
 231                         start = 0;
 232                 }
 233                 if (end > w - 1) {
 234                         end = w - 1;
 235                 }
 236
 237 #if USE_HORIZONTAL_SSE
 238                 /* round up so we get a multiple of four for the SSE code */
 239                 int num = (end - start + 1);
 240                 if (num % 4 != 0) {
 241                         /* prefer aligning it if possible */
 242                         if (start % 4 != 0 && start % 4 <= num % 4) {
 243                                 num += start % 4;
 244                                 start -= start % 4;
 245                         }
 246                         if (num % 4 != 0) {
 247                                 end += 4 - (num % 4);
 248                         }
 249                 }
 250 #endif
 251
 252                 pd[x].start = start;
 253                 pd[x].end = end;
 254                 pd[x].startcoeff = num_coeffs;
 255
 256                 for (sx = start; sx <= end; ++sx) {
 257                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
 258                         double f;
 259                         if (scaling_filter == LANCZOS) {
 260                                 f = lanczos_tap(nd);
 261                         } else {  /* Mitchell */
 262                                 f = mitchell_tap(nd);
 263                         }
 264                         if (num_coeffs == size_coeffs) {
 265                                 size_coeffs <<= 1;
 266                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 267                         }
 268
 269                         coeffs[num_coeffs++] = f;
 270                         sum += f;
 271                 }
 272
 273                 for (sx = start; sx <= end; ++sx) {
 274                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 275                 }
 276         }
 277
 278         for (y = 0; y < h; ++y) {
 279                 float *sptr = pix + y*sstride;
 280                 unsigned char *dptr = npix + y*dstride;
 281                 unsigned char ch;
 282                 for (x = 0; x < nw; ++x) {
 283 #if USE_HORIZONTAL_SSE
 284                         int result;
 285                         float acc;
 286                         long tmp;
 287                         static const float low = 0.0, high = 255.0;
 288                         __asm__ (
 289                                 "pxor %1, %1               \n"
 290                                 "xor %2, %2                \n"
 291                                 "0:                        \n"
 292                                 "movups (%4,%2),%%xmm1     \n"
 293                                 "movups (%3,%2),%%xmm2     \n"
 294                                 "mulps %%xmm2,%%xmm1       \n"
 295                                 "addps %%xmm1,%1           \n"
 296                                 "add $16,%2                \n"
 297                                 "dec %5                    \n"
 298                                 "jnz 0b                    \n"
 299                                 "haddps %1,%1              \n"
 300                                 "haddps %1,%1              \n"
 301                                 "maxss %6,%1               \n"
 302                                 "minss %7,%1               \n"
 303                                 "cvtss2si %1,%0            \n"
 304                                 : "=r" (result),
 305                                   "=&x" (acc),
 306                                   "=&r" (tmp)
 307                                 : "r" (&coeffs[pd[x].startcoeff]),
 308                                   "r" (&sptr[pd[x].start]),
 309                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 310                                   "m" (low),
 311                                   "m" (high)
 312                                 : "memory", "xmm1", "xmm2"
 313                         );
 314
 315                         *dptr++ = (unsigned char)result;
 316 #else
 317                         float acc = 0.0;
 318                         float *cf = &coeffs[pd[x].startcoeff];
 319                         unsigned sx;
 320
 321                         for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
 322                                 acc += sptr[sx] * *cf++;
 323                         }
 324
 325                         if (acc < 0.0)
 326                                 ch = 0;
 327                         else if (acc > 255.0)
 328                                 ch = 255;
 329                         else
 330                                 ch = (unsigned char)acc;
 331                         *dptr++ = ch;
 332 #endif
 333                 }
 334                 ch = dptr[-1];
 335                 for ( ; x < dstride; ++x) {
 336                         *dptr++ = ch;
 337                 }
 338         }
 339 }
 340
 341 static void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride, enum qscale_scaling_filter scaling_filter)
 342 {
 343         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 344         int size_coeffs = 8;
 345         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 346         int num_coeffs = 0;
 347         int x, y, sy;
 348         double sf = (double)h / (double)nh;
 349         double support;
 350
 351         if (scaling_filter == LANCZOS) {
 352                 support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 353         } else {  /* Mitchell */
 354                 support = (h > nh) ? (2.0 * sf) : (2.0 / sf);
 355         }
 356
 357         /* calculate the filter */
 358         for (y = 0; y < nh; ++y) {
 359                 int start = ceil(y * sf - support);
 360                 int end = floor(y * sf + support);
 361                 double sum = 0.0;
 362
 363                 if (start < 0) {
 364                         start = 0;
 365                 }
 366                 if (end > h - 1) {
 367                         end = h - 1;
 368                 }
 369
 370                 pd[y].start = start;
 371                 pd[y].end = end;
 372                 pd[y].startcoeff = num_coeffs;
 373
 374                 for (sy = start; sy <= end; ++sy) {
 375                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 376                         double f;
 377                         if (scaling_filter == LANCZOS) {
 378                                 f = lanczos_tap(nd);
 379                         } else {  /* Mitchell */
 380                                 f = mitchell_tap(nd);
 381                         }
 382                         if (num_coeffs == size_coeffs) {
 383                                 size_coeffs <<= 1;
 384                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 385                         }
 386
 387                         coeffs[num_coeffs++] = f;
 388                         sum += f;
 389                 }
 390
 391                 for (sy = start; sy <= end; ++sy) {
 392                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 393                 }
 394         }
 395
 396 #if CACHE_LINE_FACTOR > 1
 397         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 398                 unsigned char *sptr = pix + x;
 399                 float *dptr = npix + x;
 400                 for (y = 0; y < nh; ++y) {
 401 #if USE_VERTICAL_SSE
 402                         /*
 403                          * xmm0 - xmm3: acc[0..15]
 404                          * xmm4: current filter coefficient
 405                          * xmm5, xmm6, xmm7: scratchpad
 406                          */
 407                         __asm__ (
 408                                 /* clear */
 409                                 "pxor %%xmm0, %%xmm0          \n"
 410                                 "pxor %%xmm1, %%xmm1          \n"
 411                                 "pxor %%xmm2, %%xmm2          \n"
 412                                 "pxor %%xmm3, %%xmm3          \n"
 413
 414                                 /* main loop */
 415                                 "0:                           \n"
 416
 417                                 /* a zero is useful during unpacking */
 418                                 "pxor %%xmm4, %%xmm4          \n"
 419
 420                                 /* fetch all 16 source bytes */
 421                                 "movups (%0), %%xmm5          \n"
 422                                 "prefetcht0 (%0,%3,4)         \n"
 423
 424                                 /* unpack into words (xmm5, xmm7) */
 425                                 "movaps %%xmm5, %%xmm7        \n"
 426                                 "punpcklbw %%xmm4, %%xmm5     \n"
 427                                 "punpckhbw %%xmm4, %%xmm7     \n"
 428
 429                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 430                                 "movaps %%xmm5, %%xmm6        \n"
 431                                 "punpcklwd %%xmm4, %%xmm5     \n"
 432                                 "punpckhwd %%xmm4, %%xmm6     \n"
 433
 434                                 /* convert xmm5, xmm6 to floats */
 435                                 "cvtdq2ps %%xmm5, %%xmm5      \n"
 436                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 437
 438                                 /* fetch the coefficient */
 439                                 "movss (%2), %%xmm4           \n"
 440                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 441
 442                                 /* do the muls for xmm5 and xmm6 */
 443                                 "mulps %%xmm4, %%xmm5         \n"
 444                                 "mulps %%xmm4, %%xmm6         \n"
 445                                 "addps %%xmm5, %%xmm0         \n"
 446                                 "addps %%xmm6, %%xmm1         \n"
 447
 448                                 /* get the zero back again */
 449                                 "pxor %%xmm4, %%xmm4          \n"
 450
 451                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 452                                 "movaps %%xmm7, %%xmm6        \n"
 453                                 "punpcklwd %%xmm4, %%xmm7     \n"
 454                                 "punpckhwd %%xmm4, %%xmm6     \n"
 455
 456                                 /* convert xmm7, xmm6 to floats */
 457                                 "cvtdq2ps %%xmm7, %%xmm7      \n"
 458                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 459
 460                                 /* fetch the coefficient */
 461                                 "movss (%2), %%xmm4           \n"
 462                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 463
 464                                 /* do the second set of muls */
 465                                 "mulps %%xmm4, %%xmm7         \n"
 466                                 "mulps %%xmm4, %%xmm6         \n"
 467                                 "addps %%xmm7, %%xmm2         \n"
 468                                 "addps %%xmm6, %%xmm3         \n"
 469
 470                                 /* move along, and loop */
 471                                 "add $4, %2                   \n"
 472                                 "add %3, %0                   \n"
 473                                 "dec %1                       \n"
 474                                 "jnz 0b                       \n"
 475
 476                                 /* store the values */
 477                                 "movaps %%xmm0, (%4)          \n"
 478                                 "movaps %%xmm1, 16(%4)        \n"
 479                                 "movaps %%xmm2, 32(%4)        \n"
 480                                 "movaps %%xmm3, 48(%4)        \n"
 481                                 : :
 482                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 483                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 484                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 485                                 "r" ((long)w),                       /* 3: stride */
 486                                 "r" (dptr)                           /* 4: dstptr base */
 487                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 488                         );
 489 #else
 490                         int i;
 491                         float acc[CACHE_LINE_FACTOR];
 492                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 493                                 acc[i] = 0.0;
 494                         float *cf = &coeffs[pd[y].startcoeff];
 495                         unsigned sy;
 496
 497                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 498                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 499                                         acc[i] += sptr[sy * w + i] * *cf;
 500                                 }
 501                                 ++cf;
 502                         }
 503
 504                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 505                                 dptr[i] = acc[i];
 506                         }
 507 #endif
 508                         dptr += dstride;
 509                 }
 510         }
 511         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 512 #else
 513         for (x = 0; x < w; ++x) {
 514 #endif
 515                 unsigned char *sptr = pix + x;
 516                 float *dptr = npix + x;
 517                 for (y = 0; y < nh; ++y) {
 518                         float acc = 0.0;
 519                         float *cf = &coeffs[pd[y].startcoeff];
 520                         unsigned sy;
 521
 522                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 523                                 acc += sptr[sy * w] * *cf++;
 524                         }
 525
 526                         *dptr = acc;
 527                         dptr += dstride;
 528                 }
 529         }
 530 }
 531
 532 qscale_img *qscale_clone(const qscale_img *img)
 533 {
 534         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 535         if (dst == NULL) {
 536                 return NULL;
 537         }
 538
 539         *dst = *img;
 540
 541         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 542         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 543         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 544
 545         /* FIXME: handle out-of-memory gracefully */
 546         {
 547                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 548                 memcpy(dst->data_y, img->data_y, dst->h0 * dstride0);
 549         }
 550         {
 551                 dst->data_cb = (unsigned char *)malloc(dst->h1 * dstride1);
 552                 memcpy(dst->data_cb, img->data_cb, dst->h1 * dstride1);
 553         }
 554         {
 555                 dst->data_cr = (unsigned char *)malloc(dst->h2 * dstride2);
 556                 memcpy(dst->data_cr, img->data_cr, dst->h2 * dstride2);
 557         }
 558
 559         return dst;
 560 }
 561
 562 qscale_img *qscale_scale(qscale_img *src, unsigned width, unsigned height, unsigned samp_h0, unsigned samp_v0, unsigned samp_h1, unsigned samp_v1, unsigned samp_h2, unsigned samp_v2, enum qscale_scaling_filter scaling_filter)
 563 {
 564         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 565         if (dst == NULL) {
 566                 return NULL;
 567         }
 568
 569         dst->width = width;
 570         dst->height = height;
 571
 572         unsigned max_samp_h, max_samp_v;
 573         max_samp_h = samp_h0;
 574         if (samp_h1 > max_samp_h)
 575                 max_samp_h = samp_h1;
 576         if (samp_h2 > max_samp_h)
 577                 max_samp_h = samp_h2;
 578
 579         max_samp_v = samp_v0;
 580         if (samp_v1 > max_samp_v)
 581                 max_samp_v = samp_v1;
 582         if (samp_v2 > max_samp_v)
 583                 max_samp_v = samp_v2;
 584
 585         dst->w0 = width * samp_h0 / max_samp_h;
 586         dst->h0 = height * samp_v0 / max_samp_v;
 587
 588         dst->w1 = width * samp_h1 / max_samp_h;
 589         dst->h1 = height * samp_v1 / max_samp_v;
 590
 591         dst->w2 = width * samp_h2 / max_samp_h;
 592         dst->h2 = height * samp_v2 / max_samp_v;
 593
 594         dst->samp_h0 = samp_h0;
 595         dst->samp_v0 = samp_v0;
 596
 597         dst->samp_h1 = samp_h1;
 598         dst->samp_v1 = samp_v1;
 599
 600         dst->samp_h2 = samp_h2;
 601         dst->samp_v2 = samp_v2;
 602
 603         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 604         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 605         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 606
 607         unsigned sstride0 = (src->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 608         unsigned sstride1 = (src->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 609         unsigned sstride2 = (src->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 610
 611         /* FIXME: handle out-of-memory gracefully */
 612         {
 613                 float *npix = (float*)memalign(16, sstride0 * dst->h0 * sizeof(float));
 614                 vscale(src->data_y, npix, sstride0, src->h0, dst->h0, sstride0, scaling_filter);
 615                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 616                 hscale(npix, dst->data_y, src->w0, dst->h0, dst->w0, sstride0, dstride0, scaling_filter);
 617                 free(npix);
 618         }
 619         {
 620                 float *npix = (float*)memalign(16, sstride1 * dst->h1 * sizeof(float));
 621                 vscale(src->data_cr, npix, sstride1, src->h1, dst->h1, sstride1, scaling_filter);
 622                 dst->data_cr = (unsigned char *)malloc(dst->h1 * dstride1);
 623                 hscale(npix, dst->data_cr, src->w1, dst->h1, dst->w1, sstride1, dstride1, scaling_filter);
 624                 free(npix);
 625         }
 626         {
 627                 float *npix = (float*)memalign(16, sstride2 * dst->h2 * sizeof(float));
 628                 vscale(src->data_cb, npix, sstride2, src->h2, dst->h2, sstride2, scaling_filter);
 629                 dst->data_cb = (unsigned char *)malloc(dst->h2 * dstride2);
 630                 hscale(npix, dst->data_cb, src->w2, dst->h2, dst->w2, sstride2, dstride2, scaling_filter);
 631                 free(npix);
 632         }
 633
 634         return dst;
 635 }
 636
 637 int qscale_save_jpeg(const qscale_img *img, const char *filename, unsigned jpeg_quality, enum qscale_jpeg_mode jpeg_mode)
 638 {
 639         FILE *file = fopen(filename, "wb");
 640         if (file == NULL) {
 641                 return -1;
 642         }
 643
 644         int err = qscale_save_jpeg_to_stdio(img, file, jpeg_quality, jpeg_mode);
 645
 646         fclose(file);
 647         return err;
 648 }
 649
 650 int qscale_save_jpeg_to_stdio(const qscale_img *img, FILE *file, unsigned jpeg_quality, enum qscale_jpeg_mode jpeg_mode)
 651 {
 652         struct jpeg_compress_struct cinfo;
 653         struct jpeg_error_mgr jerr;
 654         cinfo.err = jpeg_std_error(&jerr);
 655         jpeg_create_compress(&cinfo);
 656         jpeg_stdio_dest(&cinfo, file);
 657         cinfo.input_components = 3;
 658         jpeg_set_defaults(&cinfo);
 659         jpeg_set_quality(&cinfo, jpeg_quality, FALSE);
 660
 661         if (jpeg_mode == PROGRESSIVE) {
 662                 jpeg_simple_progression(&cinfo);
 663         }
 664
 665         cinfo.image_width = img->width;
 666         cinfo.image_height = img->height;
 667         cinfo.raw_data_in = TRUE;
 668         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 669         cinfo.comp_info[0].h_samp_factor = img->samp_h0;
 670         cinfo.comp_info[0].v_samp_factor = img->samp_v0;
 671         cinfo.comp_info[1].h_samp_factor = img->samp_h1;
 672         cinfo.comp_info[1].v_samp_factor = img->samp_v1;
 673         cinfo.comp_info[2].h_samp_factor = img->samp_h2;
 674         cinfo.comp_info[2].v_samp_factor = img->samp_v2;
 675         jpeg_start_compress(&cinfo, TRUE);
 676
 677         unsigned dstride0 = (img->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 678         unsigned dstride1 = (img->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 679         unsigned dstride2 = (img->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 680
 681         int total_lines = 0;
 682         int blocks = 0;
 683         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 684                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 685
 686                 JSAMPROW y_row_ptrs[max_lines];
 687                 JSAMPROW cb_row_ptrs[max_lines];
 688                 JSAMPROW cr_row_ptrs[max_lines];
 689                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 690                 int i;
 691
 692                 for (i = 0; i < max_lines; ++i) {
 693                         /* simple edge extension */
 694                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 695                         if (yline > img->h0 - 1)
 696                                 yline = img->h0 - 1;
 697
 698                         int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 699                         if (cbline > img->h1 - 1)
 700                                 cbline = img->h1 - 1;
 701
 702                         int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 703                         if (crline > img->h2 - 1)
 704                                 crline = img->h2 - 1;
 705
 706                         y_row_ptrs[i]  = img->data_y  + yline * dstride0;
 707                         cb_row_ptrs[i] = img->data_cb + cbline * dstride1;
 708                         cr_row_ptrs[i] = img->data_cr + crline * dstride2;
 709                 }
 710
 711                 total_lines += max_lines;
 712                 ++blocks;
 713
 714                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 715         }
 716         jpeg_finish_compress(&cinfo);
 717         jpeg_destroy_compress(&cinfo);
 718
 719         return 0;
 720 }