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