git.sesse.net Git - qscale/blob - qscale.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 <stdio.h>
  20 #include <malloc.h>
  21 #include <math.h>
  22 #include <string.h>
  23 #include <stdlib.h>
  24 #include "jpeglib.h"
  25
  26 /* The number of pixels to process at a time when scaling vertically. */
  27 #define CACHE_LINE_FACTOR 16
  28
  29 /* Whether to use SSE for horizontal scaling or not (requires SSE3). */
  30 #define USE_HORIZONTAL_SSE 1
  31
  32 /* Whether to use SSE for vertical scaling or not (requires only SSE1). */
  33 #define USE_VERTICAL_SSE 1
  34
  35 #if USE_VERTICAL_SSE
  36 #undef CACHE_LINE_FACTOR
  37 #define CACHE_LINE_FACTOR 16
  38 #endif
  39
  40 #ifndef M_PI
  41 #define M_PI 3.14159265358979323846264
  42 #endif
  43
  44 double sinc(double x)
  45 {
  46         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
  47
  48         if (abs(x) < cutoff) {
  49                 /* For small |x|, use Taylor series instead */
  50                 const double x2 = x * x;
  51                 const double x4 = x2 * x2;
  52
  53                 return 1.0 - x2 / 6.0 + x4 / 120.0;
  54         } else {
  55                 return sin(x) / x;
  56         }
  57 }
  58
  59 double lanczos_tap(double x)
  60 {
  61         if (x < -3.0 || x > 3.0)
  62                 return 0.0;
  63         if (x < 0.0)
  64                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
  65         else
  66                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
  67 }
  68
  69
  70 struct pix_desc {
  71         unsigned start, end;
  72         unsigned startcoeff;
  73 };
  74
  75 void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
  76 {
  77         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
  78         int size_coeffs = 8;
  79         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
  80         int num_coeffs = 0;
  81         int x, y;
  82         double sf = (double)w / (double)nw;
  83         double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
  84
  85         /* calculate the filter */
  86         for (x = 0; x < nw; ++x) {
  87                 int start = ceil(x * sf - support);
  88                 int end = floor(x * sf + support);
  89                 int sx;
  90                 double sum = 0.0;
  91
  92                 if (start < 0) {
  93                         start = 0;
  94                 }
  95                 if (end > w - 1) {
  96                         end = w - 1;
  97                 }
  98
  99 #if USE_HORIZONTAL_SSE
 100                 /* round up so we get a multiple of four for the SSE code */
 101                 int num = (end - start + 1);
 102                 if (num % 4 != 0) {
 103                         /* prefer aligning it if possible */
 104                         if (start % 4 != 0 && start % 4 <= num % 4) {
 105                                 num += start % 4;
 106                                 start -= start % 4;
 107                         }
 108                         if (num % 4 != 0) {
 109                                 end += 4 - (num % 4);
 110                         }
 111                 }
 112 #endif
 113
 114                 pd[x].start = start;
 115                 pd[x].end = end;
 116                 pd[x].startcoeff = num_coeffs;
 117
 118                 for (sx = start; sx <= end; ++sx) {
 119                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
 120                         double f = lanczos_tap(nd);
 121                         if (num_coeffs == size_coeffs) {
 122                                 size_coeffs <<= 1;
 123                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 124                         }
 125
 126                         coeffs[num_coeffs++] = f;
 127                         sum += f;
 128                 }
 129
 130                 for (sx = start; sx <= end; ++sx) {
 131                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 132                 }
 133         }
 134
 135         for (y = 0; y < h; ++y) {
 136                 float *sptr = pix + y*sstride;
 137                 unsigned char *dptr = npix + y*dstride;
 138                 unsigned char ch;
 139                 for (x = 0; x < nw; ++x) {
 140 #if USE_HORIZONTAL_SSE
 141                         int result;
 142                         float acc;
 143                         long tmp;
 144                         static const float low = 0.0, high = 255.0;
 145                         __asm__ (
 146                                 "pxor %1, %1               \n"
 147                                 "xor %2, %2                \n"
 148                                 "0:                        \n"
 149                                 "movups (%4,%2),%%xmm1     \n"
 150                                 "movups (%3,%2),%%xmm2     \n"
 151                                 "mulps %%xmm2,%%xmm1       \n"
 152                                 "addps %%xmm1,%1           \n"
 153                                 "add $16,%2                \n"
 154                                 "dec %5                    \n"
 155                                 "jnz 0b                    \n"
 156                                 "haddps %1,%1              \n"
 157                                 "haddps %1,%1              \n"
 158                                 "maxss %6,%1               \n"
 159                                 "minss %7,%1               \n"
 160                                 "cvtss2si %1,%0            \n"
 161                                 : "=r" (result),
 162                                   "=&x" (acc),
 163                                   "=&r" (tmp)
 164                                 : "r" (&coeffs[pd[x].startcoeff]),
 165                                   "r" (&sptr[pd[x].start]),
 166                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 167                                   "m" (low),
 168                                   "m" (high)
 169                                 : "memory", "xmm1", "xmm2"
 170                         );
 171
 172                         *dptr++ = (unsigned char)result;
 173 #else
 174                         float acc = 0.0;
 175                         float *cf = &coeffs[pd[x].startcoeff];
 176                         unsigned sx;
 177
 178                         for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
 179                                 acc += sptr[sx] * *cf++;
 180                         }
 181
 182                         if (acc < 0.0)
 183                                 ch = 0;
 184                         else if (acc > 255.0)
 185                                 ch = 255;
 186                         else
 187                                 ch = (unsigned char)acc;
 188                         *dptr++ = ch;
 189 #endif
 190                 }
 191                 ch = dptr[-1];
 192                 for ( ; x < dstride; ++x) {
 193                         *dptr++ = ch;
 194                 }
 195         }
 196 }
 197
 198 void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 199 {
 200         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 201         int size_coeffs = 8;
 202         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 203         int num_coeffs = 0;
 204         int x, y, sy;
 205         double sf = (double)h / (double)nh;
 206         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 207
 208         /* calculate the filter */
 209         for (y = 0; y < nh; ++y) {
 210                 int start = ceil(y * sf - support);
 211                 int end = floor(y * sf + support);
 212                 double sum = 0.0;
 213
 214                 if (start < 0) {
 215                         start = 0;
 216                 }
 217                 if (end > h - 1) {
 218                         end = h - 1;
 219                 }
 220
 221                 pd[y].start = start;
 222                 pd[y].end = end;
 223                 pd[y].startcoeff = num_coeffs;
 224
 225                 for (sy = start; sy <= end; ++sy) {
 226                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 227                         double f = lanczos_tap(nd);
 228                         if (num_coeffs == size_coeffs) {
 229                                 size_coeffs <<= 1;
 230                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 231                         }
 232
 233                         coeffs[num_coeffs++] = f;
 234                         sum += f;
 235                 }
 236
 237                 for (sy = start; sy <= end; ++sy) {
 238                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 239                 }
 240         }
 241
 242 #if CACHE_LINE_FACTOR > 1
 243         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 244                 unsigned char *sptr = pix + x;
 245                 float *dptr = npix + x;
 246                 for (y = 0; y < nh; ++y) {
 247 #if USE_VERTICAL_SSE
 248                         /*
 249                          * xmm0 - xmm3: acc[0..15]
 250                          * xmm4: current filter coefficient
 251                          * xmm5, xmm6, xmm7: scratchpad
 252                          */
 253                         __asm__ (
 254                                 /* clear */
 255                                 "pxor %%xmm0, %%xmm0          \n"
 256                                 "pxor %%xmm1, %%xmm1          \n"
 257                                 "pxor %%xmm2, %%xmm2          \n"
 258                                 "pxor %%xmm3, %%xmm3          \n"
 259
 260                                 /* main loop */
 261                                 "0:                           \n"
 262
 263                                 /* a zero is useful during unpacking */
 264                                 "pxor %%xmm4, %%xmm4          \n"
 265
 266                                 /* fetch all 16 source bytes */
 267                                 "movups (%0), %%xmm5          \n"
 268                                 "prefetcht0 (%0,%3,4)         \n"
 269
 270                                 /* unpack into words (xmm5, xmm7) */
 271                                 "movaps %%xmm5, %%xmm7        \n"
 272                                 "punpcklbw %%xmm4, %%xmm5     \n"
 273                                 "punpckhbw %%xmm4, %%xmm7     \n"
 274
 275                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 276                                 "movaps %%xmm5, %%xmm6        \n"
 277                                 "punpcklwd %%xmm4, %%xmm5     \n"
 278                                 "punpckhwd %%xmm4, %%xmm6     \n"
 279
 280                                 /* convert xmm5, xmm6 to floats */
 281                                 "cvtdq2ps %%xmm5, %%xmm5      \n"
 282                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 283
 284                                 /* fetch the coefficient */
 285                                 "movss (%2), %%xmm4           \n"
 286                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 287
 288                                 /* do the muls for xmm5 and xmm6 */
 289                                 "mulps %%xmm4, %%xmm5         \n"
 290                                 "mulps %%xmm4, %%xmm6         \n"
 291                                 "addps %%xmm5, %%xmm0         \n"
 292                                 "addps %%xmm6, %%xmm1         \n"
 293
 294                                 /* get the zero back again */
 295                                 "pxor %%xmm4, %%xmm4          \n"
 296
 297                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 298                                 "movaps %%xmm7, %%xmm6        \n"
 299                                 "punpcklwd %%xmm4, %%xmm7     \n"
 300                                 "punpckhwd %%xmm4, %%xmm6     \n"
 301
 302                                 /* convert xmm7, xmm6 to floats */
 303                                 "cvtdq2ps %%xmm7, %%xmm7      \n"
 304                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 305
 306                                 /* fetch the coefficient */
 307                                 "movss (%2), %%xmm4           \n"
 308                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 309
 310                                 /* do the second set of muls */
 311                                 "mulps %%xmm4, %%xmm7         \n"
 312                                 "mulps %%xmm4, %%xmm6         \n"
 313                                 "addps %%xmm7, %%xmm2         \n"
 314                                 "addps %%xmm6, %%xmm3         \n"
 315
 316                                 /* move along, and loop */
 317                                 "add $4, %2                   \n"
 318                                 "add %3, %0                   \n"
 319                                 "dec %1                       \n"
 320                                 "jnz 0b                       \n"
 321
 322                                 /* store the values */
 323                                 "movaps %%xmm0, (%4)          \n"
 324                                 "movaps %%xmm1, 16(%4)        \n"
 325                                 "movaps %%xmm2, 32(%4)        \n"
 326                                 "movaps %%xmm3, 48(%4)        \n"
 327                                 : :
 328                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 329                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 330                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 331                                 "r" ((long)w),                       /* 3: stride */
 332                                 "r" (dptr)                           /* 4: dstptr base */
 333                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 334                         );
 335 #else
 336                         int i;
 337                         float acc[CACHE_LINE_FACTOR];
 338                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 339                                 acc[i] = 0.0;
 340                         float *cf = &coeffs[pd[y].startcoeff];
 341                         unsigned sy;
 342
 343                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 344                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 345                                         acc[i] += sptr[sy * w + i] * *cf;
 346                                 }
 347                                 ++cf;
 348                         }
 349
 350                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 351                                 dptr[i] = acc[i];
 352                         }
 353 #endif
 354                         dptr += dstride;
 355                 }
 356         }
 357         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 358 #else
 359         for (x = 0; x < w; ++x) {
 360 #endif
 361                 unsigned char *sptr = pix + x;
 362                 float *dptr = npix + x;
 363                 for (y = 0; y < nh; ++y) {
 364                         float acc = 0.0;
 365                         float *cf = &coeffs[pd[y].startcoeff];
 366                         unsigned sy;
 367
 368                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 369                                 acc += sptr[sy * w] * *cf++;
 370                         }
 371
 372                         *dptr = acc;
 373                         dptr += dstride;
 374                 }
 375         }
 376 }
 377
 378 int main(int argc, char **argv)
 379 {
 380         /* user-settable parameters */
 381         unsigned nominal_w = atoi(argv[1]);
 382         unsigned nominal_h = atoi(argv[2]);
 383         unsigned samp_h0 = 2, samp_v0 = 2;
 384         unsigned samp_h1 = 1, samp_v1 = 1;
 385         unsigned samp_h2 = 1, samp_v2 = 1;
 386         unsigned jpeg_quality = 85;
 387         /* end */
 388
 389         unsigned max_samp_h, max_samp_v;
 390         max_samp_h = samp_h0;
 391         if (samp_h1 > max_samp_h)
 392                 max_samp_h = samp_h1;
 393         if (samp_h2 > max_samp_h)
 394                 max_samp_h = samp_h2;
 395
 396         max_samp_v = samp_v0;
 397         if (samp_v1 > max_samp_v)
 398                 max_samp_v = samp_v1;
 399         if (samp_v2 > max_samp_v)
 400                 max_samp_v = samp_v2;
 401
 402         unsigned nw0 = nominal_w * samp_h0 / max_samp_h, nh0 = nominal_h * samp_v0 / max_samp_v;
 403         unsigned nw1 = nominal_w * samp_h1 / max_samp_h, nh1 = nominal_h * samp_v1 / max_samp_v;
 404         unsigned nw2 = nominal_w * samp_h2 / max_samp_h, nh2 = nominal_h * samp_v2 / max_samp_v;
 405
 406         unsigned stride0 = (nw0 + DCTSIZE-1) & ~(DCTSIZE-1);
 407         unsigned stride1 = (nw1 + DCTSIZE-1) & ~(DCTSIZE-1);
 408         unsigned stride2 = (nw2 + DCTSIZE-1) & ~(DCTSIZE-1);
 409
 410         struct jpeg_decompress_struct dinfo;
 411         struct jpeg_error_mgr jerr;
 412         dinfo.err = jpeg_std_error(&jerr);
 413         jpeg_create_decompress(&dinfo);
 414         jpeg_stdio_src(&dinfo, stdin);
 415         jpeg_read_header(&dinfo, TRUE);
 416         dinfo.raw_data_out = TRUE;
 417         jpeg_start_decompress(&dinfo);
 418
 419         unsigned w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
 420         unsigned h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
 421
 422         unsigned w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
 423         unsigned h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
 424
 425         unsigned w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
 426         unsigned h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
 427
 428         fprintf(stderr, "Scaling using Lanczos filter:\n");
 429         fprintf(stderr, "  Y component: %ux%u -> %ux%u\n", dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nw0, nh0);
 430         fprintf(stderr, "  Cb component: %ux%u -> %ux%u\n", dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nw1, nh1);
 431         fprintf(stderr, "  Cr component: %ux%u -> %ux%u\n", dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nw2, nh2);
 432
 433         JSAMPLE *data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 434         JSAMPLE *data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 435         JSAMPLE *data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 436         JSAMPLE *data_ny, *data_ncb, *data_ncr;
 437
 438         int total_lines = 0, blocks = 0;
 439         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 440                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 441
 442                 JSAMPROW y_row_ptrs[max_lines];
 443                 JSAMPROW cb_row_ptrs[max_lines];
 444                 JSAMPROW cr_row_ptrs[max_lines];
 445                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 446                 int i;
 447
 448                 for (i = 0; i < max_lines; ++i) {
 449                         y_row_ptrs[i]  = data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
 450                         cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
 451                         cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
 452                 }
 453
 454                 total_lines += max_lines;
 455                 ++blocks;
 456
 457                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 458                         break;
 459         }
 460
 461         {
 462                 float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));
 463                 vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, h0, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
 464                 data_ny = (unsigned char *)malloc(nh0 * stride0);
 465                 hscale(npix, data_ny, w0, nh0, nw0, dinfo.comp_info[0].width_in_blocks * DCTSIZE, stride0);
 466                 free(npix);
 467         }
 468         {
 469                 float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));
 470                 vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, h1, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
 471                 data_ncr = (unsigned char *)malloc(nh1 * stride1);
 472                 hscale(npix, data_ncr, w1, nh1, nw1, dinfo.comp_info[1].width_in_blocks * DCTSIZE, stride1);
 473                 free(npix);
 474         }
 475         {
 476                 float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));
 477                 vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, h2, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
 478                 data_ncb = (unsigned char *)malloc(nh2 * stride2);
 479                 hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2);
 480                 free(npix);
 481         }
 482         jpeg_destroy_decompress(&dinfo);
 483
 484         struct jpeg_compress_struct cinfo;
 485         cinfo.err = jpeg_std_error(&jerr);
 486         jpeg_create_compress(&cinfo);
 487         jpeg_stdio_dest(&cinfo, stdout);
 488         cinfo.input_components = 3;
 489         jpeg_set_defaults(&cinfo);
 490         jpeg_set_quality(&cinfo, jpeg_quality, FALSE);
 491         cinfo.image_width = nominal_w;
 492         cinfo.image_height = nominal_h;
 493         cinfo.raw_data_in = TRUE;
 494         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 495         cinfo.comp_info[0].h_samp_factor = samp_h0;
 496         cinfo.comp_info[0].v_samp_factor = samp_v0;
 497         cinfo.comp_info[1].h_samp_factor = samp_h1;
 498         cinfo.comp_info[1].v_samp_factor = samp_v1;
 499         cinfo.comp_info[2].h_samp_factor = samp_h2;
 500         cinfo.comp_info[2].v_samp_factor = samp_v2;
 501         jpeg_start_compress(&cinfo, TRUE);
 502
 503         total_lines = 0;
 504         blocks = 0;
 505         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 506                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 507
 508                 JSAMPROW y_row_ptrs[max_lines];
 509                 JSAMPROW cb_row_ptrs[max_lines];
 510                 JSAMPROW cr_row_ptrs[max_lines];
 511                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 512                 int i;
 513
 514                 for (i = 0; i < max_lines; ++i) {
 515                         /* simple edge extension */
 516                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 517                         if (yline > nh0 - 1)
 518                                 yline = nh0 - 1;
 519
 520                         int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 521                         if (cbline > nh1 - 1)
 522                                 cbline = nh1 - 1;
 523
 524                         int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 525                         if (crline > nh2 - 1)
 526                                 crline = nh2 - 1;
 527
 528                         y_row_ptrs[i]  = data_ny  + yline * stride0;
 529                         cb_row_ptrs[i] = data_ncb + cbline * stride1;
 530                         cr_row_ptrs[i] = data_ncr + crline * stride2;
 531                 }
 532
 533                 total_lines += max_lines;
 534                 ++blocks;
 535
 536                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 537         }
 538         jpeg_finish_compress(&cinfo);
 539         jpeg_destroy_compress(&cinfo);
 540
 541         return 0;
 542 }
 543