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 #if USE_VERTICAL_SSE || USE_HORIZONTAL_SSE
  46 typedef float v4sf __attribute__((vector_size(16)));
  47 typedef int v4si __attribute__((vector_size(16)));
  48 typedef short v8hi __attribute__((vector_size(16)));
  49 typedef char v16qi __attribute__((vector_size(16)));
  50 #endif
  51
  52 qscale_img *qscale_load_jpeg(const char *filename)
  53 {
  54         FILE *file = fopen(filename, "rb");
  55         qscale_img *img;
  56         if (file == NULL) {
  57                 return NULL;
  58         }
  59
  60         img = qscale_load_jpeg_from_stdio(file);
  61
  62         fclose(file);
  63         return img;
  64 }
  65
  66 qscale_img *qscale_load_jpeg_from_stdio(FILE *file)
  67 {
  68         qscale_img *img = (qscale_img *)malloc(sizeof(qscale_img));
  69         if (img == NULL) {
  70                 return NULL;
  71         }
  72
  73         img->data_y = img->data_cb = img->data_cr = NULL;
  74
  75         /* FIXME: Better error handling here (ie., return NULL). */
  76         struct jpeg_decompress_struct dinfo;
  77         struct jpeg_error_mgr jerr;
  78         dinfo.err = jpeg_std_error(&jerr);
  79         jpeg_create_decompress(&dinfo);
  80         jpeg_stdio_src(&dinfo, file);
  81         jpeg_read_header(&dinfo, TRUE);
  82         dinfo.raw_data_out = TRUE;
  83         jpeg_start_decompress(&dinfo);
  84
  85         if (dinfo.num_components != 1 && dinfo.num_components != 3) {
  86                 qscale_destroy(img);
  87                 return NULL;
  88         }
  89         img->num_components = dinfo.num_components;
  90
  91         img->width = dinfo.image_width;
  92         img->height = dinfo.image_height;
  93
  94         img->w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
  95         img->h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
  96
  97         if (img->num_components == 3) {
  98                 img->w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
  99                 img->h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
 100
 101                 img->w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
 102                 img->h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
 103         }
 104
 105         img->samp_h0 = dinfo.comp_info[0].h_samp_factor;
 106         img->samp_v0 = dinfo.comp_info[0].v_samp_factor;
 107
 108         if (img->num_components == 3) {
 109                 img->samp_h1 = dinfo.comp_info[1].h_samp_factor;
 110                 img->samp_v1 = dinfo.comp_info[1].v_samp_factor;
 111
 112                 img->samp_h2 = dinfo.comp_info[2].h_samp_factor;
 113                 img->samp_v2 = dinfo.comp_info[2].v_samp_factor;
 114         }
 115
 116         img->data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 117         if (img->data_y == NULL) {
 118                 qscale_destroy(img);
 119                 return NULL;
 120         }
 121
 122         if (img->num_components == 3) {
 123                 img->data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 124                 img->data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 125                 if (img->data_cb == NULL || img->data_cr == NULL) {
 126                         qscale_destroy(img);
 127                         return NULL;
 128                 }
 129         }
 130
 131         int total_lines = 0, blocks = 0;
 132         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 133                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 134
 135                 JSAMPROW y_row_ptrs[max_lines];
 136                 JSAMPROW cb_row_ptrs[max_lines];
 137                 JSAMPROW cr_row_ptrs[max_lines];
 138                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 139
 140                 int i;
 141                 for (i = 0; i < max_lines; ++i) {
 142                         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;
 143                         if (img->num_components == 3) {
 144                                 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;
 145                                 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;
 146                         }
 147                 }
 148
 149                 total_lines += max_lines;
 150                 ++blocks;
 151
 152                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 153                         break;
 154         }
 155
 156         jpeg_destroy_decompress(&dinfo);
 157         return img;
 158 }
 159
 160 void qscale_destroy(qscale_img *img)
 161 {
 162         free(img->data_y);
 163         free(img->data_cb);
 164         free(img->data_cr);
 165         free(img);
 166 }
 167
 168
 169 static double sinc(double x)
 170 {
 171         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
 172
 173         if (abs(x) < cutoff) {
 174                 /* For small |x|, use Taylor series instead */
 175                 const double x2 = x * x;
 176                 const double x4 = x2 * x2;
 177
 178                 return 1.0 - x2 / 6.0 + x4 / 120.0;
 179         } else {
 180                 return sin(x) / x;
 181         }
 182 }
 183
 184 static double lanczos_tap(double x)
 185 {
 186         if (x < -3.0 || x > 3.0)
 187                 return 0.0;
 188         if (x < 0.0)
 189                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
 190         else
 191                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
 192 }
 193
 194 static double mitchell_tap(double x)
 195 {
 196         const double b = 1.0 / 3.0;
 197         const double c = 1.0 / 3.0;
 198         const double p0 = (  6.0 -  2.0*b         ) / 6.0;
 199         const double p2 = (-18.0 + 12.0*b +  6.0*c) / 6.0;
 200         const double p3 = ( 12.0 -  9.0*b -  6.0*c) / 6.0;
 201         const double q0 = (         8.0*b + 24.0*c) / 6.0;
 202         const double q1 = (      - 12.0*b - 48.0*c) / 6.0;
 203         const double q2 = (         6.0*b + 30.0*c) / 6.0;
 204         const double q3 = (      -      b -  6.0*c) / 6.0;
 205
 206         if (x < -2.0) {
 207                 return 0.0;
 208         } else if (x < -1.0) {
 209                 return q0 - x * (q1 - x * (q2 - x * q3));
 210         } else if (x < 0.0) {
 211                 return p0 + x * x * (p2 - x * p3);
 212         } else if (x < 1.0) {
 213                 return p0 + x * x * (p2 + x * p3);
 214         } else if (x < 2.0) {
 215                 return q0 + x * (q1 + x * (q2 + x * q3));
 216         } else {
 217                 return 0.0;
 218         }
 219 }
 220
 221 struct pix_desc {
 222         unsigned start, end;
 223         unsigned startcoeff;
 224 };
 225
 226 static void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride, enum qscale_scaling_filter scaling_filter)
 227 {
 228         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
 229         int size_coeffs = 8;
 230         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 231         int num_coeffs = 0;
 232         int x, y;
 233         double sf = (double)w / (double)nw;
 234         double support;
 235
 236         if (scaling_filter == LANCZOS) {
 237                 support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
 238         } else {  /* Mitchell */
 239                 support = (w > nw) ? (2.0 * sf) : (2.0 / sf);
 240         }
 241
 242         /* calculate the filter */
 243         for (x = 0; x < nw; ++x) {
 244                 int start = ceil(x * sf - support);
 245                 int end = floor(x * sf + support);
 246                 int sx;
 247                 double sum = 0.0;
 248
 249                 if (start < 0) {
 250                         start = 0;
 251                 }
 252                 if (end > w - 1) {
 253                         end = w - 1;
 254                 }
 255
 256 #if USE_HORIZONTAL_SSE
 257                 /* round up so we get a multiple of four for the SSE code */
 258                 int num = (end - start + 1);
 259                 if (num % 4 != 0) {
 260                         /* prefer aligning it if possible */
 261                         if (start % 4 != 0 && start % 4 <= num % 4) {
 262                                 num += start % 4;
 263                                 start -= start % 4;
 264                         }
 265                         if (num % 4 != 0) {
 266                                 end += 4 - (num % 4);
 267                         }
 268                 }
 269 #endif
 270
 271                 pd[x].start = start;
 272                 pd[x].end = end;
 273                 pd[x].startcoeff = num_coeffs;
 274
 275                 for (sx = start; sx <= end; ++sx) {
 276                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
 277                         double f;
 278                         if (scaling_filter == LANCZOS) {
 279                                 f = lanczos_tap(nd);
 280                         } else {  /* Mitchell */
 281                                 f = mitchell_tap(nd);
 282                         }
 283                         if (num_coeffs == size_coeffs) {
 284                                 size_coeffs <<= 1;
 285                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 286                         }
 287
 288                         coeffs[num_coeffs++] = f;
 289                         sum += f;
 290                 }
 291
 292                 for (sx = start; sx <= end; ++sx) {
 293                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 294                 }
 295         }
 296
 297         for (y = 0; y < h; ++y) {
 298                 float *sptr = pix + y*sstride;
 299                 unsigned char *dptr = npix + y*dstride;
 300                 unsigned char ch;
 301                 for (x = 0; x < nw; ++x) {
 302 #if USE_HORIZONTAL_SSE
 303                         v4sf acc = { 0.0f, 0.0f, 0.0f, 0.0f };
 304                         static const v4sf low = { 0.0f, 0.0f, 0.0f, 0.0f };
 305                         static const v4sf high = { 255.0f, 255.0f, 255.0f, 255.0f };
 306                         int result;
 307                         int i;
 308
 309                         const float *sptr_xmm = &sptr[pd[x].start];
 310                         const float *coeffptr = &coeffs[pd[x].startcoeff];
 311                         const int filter_len = (pd[x].end - pd[x].start + 1) / 4;
 312
 313                         for (i = 0; i < filter_len; ++i) {
 314                                 v4sf pixels = __builtin_ia32_loadups(&sptr_xmm[i * 4]);
 315                                 v4sf coeffs = __builtin_ia32_loadups(&coeffptr[i * 4]);
 316                                 acc = __builtin_ia32_addps(acc, __builtin_ia32_mulps(pixels, coeffs));
 317                         }
 318                         acc = __builtin_ia32_haddps(acc, acc);
 319                         acc = __builtin_ia32_haddps(acc, acc);
 320                         acc = __builtin_ia32_maxss(acc, low);
 321                         acc = __builtin_ia32_minss(acc, high);
 322                         result = __builtin_ia32_cvtss2si(acc);
 323
 324                         *dptr++ = (unsigned char)result;
 325 #else
 326                         float acc = 0.0;
 327                         float *cf = &coeffs[pd[x].startcoeff];
 328                         unsigned sx;
 329
 330                         for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
 331                                 acc += sptr[sx] * *cf++;
 332                         }
 333
 334                         if (acc < 0.0)
 335                                 ch = 0;
 336                         else if (acc > 255.0)
 337                                 ch = 255;
 338                         else
 339                                 ch = (unsigned char)acc;
 340                         *dptr++ = ch;
 341 #endif
 342                 }
 343                 ch = dptr[-1];
 344                 for ( ; x < dstride; ++x) {
 345                         *dptr++ = ch;
 346                 }
 347         }
 348 }
 349
 350 static void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride, enum qscale_scaling_filter scaling_filter)
 351 {
 352         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 353         int size_coeffs = 8;
 354         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 355         int num_coeffs = 0;
 356         int x, y, sy;
 357         double sf = (double)h / (double)nh;
 358         double support;
 359
 360         if (scaling_filter == LANCZOS) {
 361                 support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 362         } else {  /* Mitchell */
 363                 support = (h > nh) ? (2.0 * sf) : (2.0 / sf);
 364         }
 365
 366         /* calculate the filter */
 367         for (y = 0; y < nh; ++y) {
 368                 int start = ceil(y * sf - support);
 369                 int end = floor(y * sf + support);
 370                 double sum = 0.0;
 371
 372                 if (start < 0) {
 373                         start = 0;
 374                 }
 375                 if (end > h - 1) {
 376                         end = h - 1;
 377                 }
 378
 379                 pd[y].start = start;
 380                 pd[y].end = end;
 381                 pd[y].startcoeff = num_coeffs;
 382
 383                 for (sy = start; sy <= end; ++sy) {
 384                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 385                         double f;
 386                         if (scaling_filter == LANCZOS) {
 387                                 f = lanczos_tap(nd);
 388                         } else {  /* Mitchell */
 389                                 f = mitchell_tap(nd);
 390                         }
 391                         if (num_coeffs == size_coeffs) {
 392                                 size_coeffs <<= 1;
 393                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 394                         }
 395
 396                         coeffs[num_coeffs++] = f;
 397                         sum += f;
 398                 }
 399
 400                 for (sy = start; sy <= end; ++sy) {
 401                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 402                 }
 403         }
 404
 405 #if CACHE_LINE_FACTOR > 1
 406         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 407                 unsigned char *sptr = pix + x;
 408                 float *dptr = npix + x;
 409                 for (y = 0; y < nh; ++y) {
 410 #if USE_VERTICAL_SSE
 411                         /* A zero is useful during unpacking. */
 412                         static const v4sf zero = { 0.0f, 0.0f, 0.0f, 0.0f };
 413                         const unsigned char *sptr_xmm = &sptr[pd[y].start * w];
 414                         const float *coeffptr = &coeffs[pd[y].startcoeff];
 415                         const int filter_len = pd[y].end - pd[y].start + 1;
 416                         int i;
 417
 418                         v4sf acc0 = { 0.0f, 0.0f, 0.0f, 0.0f };
 419                         v4sf acc1 = { 0.0f, 0.0f, 0.0f, 0.0f };
 420                         v4sf acc2 = { 0.0f, 0.0f, 0.0f, 0.0f };
 421                         v4sf acc3 = { 0.0f, 0.0f, 0.0f, 0.0f };
 422
 423                         for (i = 0; i < filter_len; ++i, ++coeffptr, sptr_xmm += w) {
 424                                 __builtin_prefetch(sptr_xmm + w, 0);
 425                                 v16qi src = (v16qi)__builtin_ia32_loadups((float*)sptr_xmm);
 426
 427                                 // unpack into words
 428                                 v8hi src_lo = (v8hi)__builtin_ia32_punpcklbw128(src, (v16qi)zero);
 429                                 v8hi src_hi = (v8hi)__builtin_ia32_punpckhbw128(src, (v16qi)zero);
 430
 431                                 // unpack into dwords, convert to floats
 432                                 v4si src0_i = (v4si)__builtin_ia32_punpcklwd128(src_lo, (v8hi)zero);
 433                                 v4si src1_i = (v4si)__builtin_ia32_punpckhwd128(src_lo, (v8hi)zero);
 434                                 v4si src2_i = (v4si)__builtin_ia32_punpcklwd128(src_hi, (v8hi)zero);
 435                                 v4si src3_i = (v4si)__builtin_ia32_punpckhwd128(src_hi, (v8hi)zero);
 436
 437                                 v4sf src0 = __builtin_ia32_cvtdq2ps(src0_i);
 438                                 v4sf src1 = __builtin_ia32_cvtdq2ps(src1_i);
 439                                 v4sf src2 = __builtin_ia32_cvtdq2ps(src2_i);
 440                                 v4sf src3 = __builtin_ia32_cvtdq2ps(src3_i);
 441
 442                                 // fetch the coefficient, and replicate it
 443                                 v4sf coeff = { *coeffptr, *coeffptr, *coeffptr, *coeffptr };
 444
 445                                 // do the actual muladds
 446                                 acc0 = __builtin_ia32_addps(acc0, __builtin_ia32_mulps(src0, coeff));
 447                                 acc1 = __builtin_ia32_addps(acc1, __builtin_ia32_mulps(src1, coeff));
 448                                 acc2 = __builtin_ia32_addps(acc2, __builtin_ia32_mulps(src2, coeff));
 449                                 acc3 = __builtin_ia32_addps(acc3, __builtin_ia32_mulps(src3, coeff));
 450                         }
 451
 452                         *(v4sf *)(&dptr[0]) = acc0;
 453                         *(v4sf *)(&dptr[4]) = acc1;
 454                         *(v4sf *)(&dptr[8]) = acc2;
 455                         *(v4sf *)(&dptr[12]) = acc3;
 456 #else
 457                         int i;
 458                         float acc[CACHE_LINE_FACTOR];
 459                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 460                                 acc[i] = 0.0;
 461                         float *cf = &coeffs[pd[y].startcoeff];
 462                         unsigned sy;
 463
 464                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 465                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 466                                         acc[i] += sptr[sy * w + i] * *cf;
 467                                 }
 468                                 ++cf;
 469                         }
 470
 471                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 472                                 dptr[i] = acc[i];
 473                         }
 474 #endif
 475                         dptr += dstride;
 476                 }
 477         }
 478         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 479 #else
 480         for (x = 0; x < w; ++x) {
 481 #endif
 482                 unsigned char *sptr = pix + x;
 483                 float *dptr = npix + x;
 484                 for (y = 0; y < nh; ++y) {
 485                         float acc = 0.0;
 486                         float *cf = &coeffs[pd[y].startcoeff];
 487                         unsigned sy;
 488
 489                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 490                                 acc += sptr[sy * w] * *cf++;
 491                         }
 492
 493                         *dptr = acc;
 494                         dptr += dstride;
 495                 }
 496         }
 497 }
 498
 499 qscale_img *qscale_clone(const qscale_img *img)
 500 {
 501         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 502         if (dst == NULL) {
 503                 return NULL;
 504         }
 505
 506         *dst = *img;
 507
 508         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 509         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 510         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 511
 512         /* FIXME: handle out-of-memory gracefully */
 513         {
 514                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 515                 memcpy(dst->data_y, img->data_y, dst->h0 * dstride0);
 516         }
 517         {
 518                 dst->data_cb = (unsigned char *)malloc(dst->h1 * dstride1);
 519                 memcpy(dst->data_cb, img->data_cb, dst->h1 * dstride1);
 520         }
 521         {
 522                 dst->data_cr = (unsigned char *)malloc(dst->h2 * dstride2);
 523                 memcpy(dst->data_cr, img->data_cr, dst->h2 * dstride2);
 524         }
 525
 526         return dst;
 527 }
 528
 529 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)
 530 {
 531         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 532         if (dst == NULL) {
 533                 return NULL;
 534         }
 535
 536         dst->width = width;
 537         dst->height = height;
 538         dst->num_components = src->num_components;
 539
 540         unsigned max_samp_h, max_samp_v;
 541         max_samp_h = samp_h0;
 542         if (src->num_components == 3) {
 543                 if (samp_h1 > max_samp_h)
 544                         max_samp_h = samp_h1;
 545                 if (samp_h2 > max_samp_h)
 546                         max_samp_h = samp_h2;
 547         }
 548
 549         max_samp_v = samp_v0;
 550         if (src->num_components == 3) {
 551                 if (samp_v1 > max_samp_v)
 552                         max_samp_v = samp_v1;
 553                 if (samp_v2 > max_samp_v)
 554                         max_samp_v = samp_v2;
 555         }
 556
 557         dst->w0 = width * samp_h0 / max_samp_h;
 558         dst->h0 = height * samp_v0 / max_samp_v;
 559
 560         if (src->num_components == 3) {
 561                 dst->w1 = width * samp_h1 / max_samp_h;
 562                 dst->h1 = height * samp_v1 / max_samp_v;
 563
 564                 dst->w2 = width * samp_h2 / max_samp_h;
 565                 dst->h2 = height * samp_v2 / max_samp_v;
 566         }
 567
 568         dst->samp_h0 = samp_h0;
 569         dst->samp_v0 = samp_v0;
 570
 571         if (src->num_components == 3) {
 572                 dst->samp_h1 = samp_h1;
 573                 dst->samp_v1 = samp_v1;
 574
 575                 dst->samp_h2 = samp_h2;
 576                 dst->samp_v2 = samp_v2;
 577         }
 578
 579         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 580         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 581         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 582
 583         unsigned sstride0 = (src->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 584         unsigned sstride1 = (src->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 585         unsigned sstride2 = (src->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 586
 587         /* FIXME: handle out-of-memory gracefully */
 588         {
 589                 float *npix = (float*)memalign(16, sstride0 * dst->h0 * sizeof(float));
 590                 vscale(src->data_y, npix, sstride0, src->h0, dst->h0, sstride0, scaling_filter);
 591                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 592                 hscale(npix, dst->data_y, src->w0, dst->h0, dst->w0, sstride0, dstride0, scaling_filter);
 593                 free(npix);
 594         }
 595         if (src->num_components == 3) {
 596                 {
 597                         float *npix = (float*)memalign(16, sstride1 * dst->h1 * sizeof(float));
 598                         vscale(src->data_cr, npix, sstride1, src->h1, dst->h1, sstride1, scaling_filter);
 599                         dst->data_cr = (unsigned char *)malloc(dst->h1 * dstride1);
 600                         hscale(npix, dst->data_cr, src->w1, dst->h1, dst->w1, sstride1, dstride1, scaling_filter);
 601                         free(npix);
 602                 }
 603                 {
 604                         float *npix = (float*)memalign(16, sstride2 * dst->h2 * sizeof(float));
 605                         vscale(src->data_cb, npix, sstride2, src->h2, dst->h2, sstride2, scaling_filter);
 606                         dst->data_cb = (unsigned char *)malloc(dst->h2 * dstride2);
 607                         hscale(npix, dst->data_cb, src->w2, dst->h2, dst->w2, sstride2, dstride2, scaling_filter);
 608                         free(npix);
 609                 }
 610         }
 611
 612         return dst;
 613 }
 614
 615 int qscale_save_jpeg(const qscale_img *img, const char *filename, unsigned jpeg_quality, enum qscale_jpeg_mode jpeg_mode)
 616 {
 617         FILE *file = fopen(filename, "wb");
 618         if (file == NULL) {
 619                 return -1;
 620         }
 621
 622         int err = qscale_save_jpeg_to_stdio(img, file, jpeg_quality, jpeg_mode);
 623
 624         fclose(file);
 625         return err;
 626 }
 627
 628 int qscale_save_jpeg_to_stdio(const qscale_img *img, FILE *file, unsigned jpeg_quality, enum qscale_jpeg_mode jpeg_mode)
 629 {
 630         struct jpeg_compress_struct cinfo;
 631         struct jpeg_error_mgr jerr;
 632         cinfo.err = jpeg_std_error(&jerr);
 633         jpeg_create_compress(&cinfo);
 634         jpeg_stdio_dest(&cinfo, file);
 635         cinfo.input_components = img->num_components;
 636         jpeg_set_defaults(&cinfo);
 637         jpeg_set_quality(&cinfo, jpeg_quality, FALSE);
 638
 639         if (jpeg_mode == PROGRESSIVE) {
 640                 jpeg_simple_progression(&cinfo);
 641         }
 642
 643         cinfo.image_width = img->width;
 644         cinfo.image_height = img->height;
 645         cinfo.raw_data_in = TRUE;
 646         if (img->num_components == 3) {
 647                 jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 648         } else {
 649                 jpeg_set_colorspace(&cinfo, JCS_GRAYSCALE);
 650         }
 651         cinfo.comp_info[0].h_samp_factor = img->samp_h0;
 652         cinfo.comp_info[0].v_samp_factor = img->samp_v0;
 653         if (img->num_components == 3) {
 654                 cinfo.comp_info[1].h_samp_factor = img->samp_h1;
 655                 cinfo.comp_info[1].v_samp_factor = img->samp_v1;
 656                 cinfo.comp_info[2].h_samp_factor = img->samp_h2;
 657                 cinfo.comp_info[2].v_samp_factor = img->samp_v2;
 658         }
 659         jpeg_start_compress(&cinfo, TRUE);
 660
 661         unsigned dstride0 = (img->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 662         unsigned dstride1 = (img->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 663         unsigned dstride2 = (img->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 664
 665         int total_lines = 0;
 666         int blocks = 0;
 667         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 668                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 669
 670                 JSAMPROW y_row_ptrs[max_lines];
 671                 JSAMPROW cb_row_ptrs[max_lines];
 672                 JSAMPROW cr_row_ptrs[max_lines];
 673                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 674                 int i;
 675
 676                 for (i = 0; i < max_lines; ++i) {
 677                         /* simple edge extension */
 678                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 679                         if (yline > img->h0 - 1)
 680                                 yline = img->h0 - 1;
 681
 682                         y_row_ptrs[i]  = img->data_y  + yline * dstride0;
 683
 684                         if (img->num_components == 3) {
 685                                 int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 686                                 if (cbline > img->h1 - 1)
 687                                         cbline = img->h1 - 1;
 688
 689                                 int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 690                                 if (crline > img->h2 - 1)
 691                                         crline = img->h2 - 1;
 692
 693                                 cb_row_ptrs[i] = img->data_cb + cbline * dstride1;
 694                                 cr_row_ptrs[i] = img->data_cr + crline * dstride2;
 695                         }
 696                 }
 697
 698                 total_lines += max_lines;
 699                 ++blocks;
 700
 701                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 702         }
 703         jpeg_finish_compress(&cinfo);
 704         jpeg_destroy_compress(&cinfo);
 705
 706         return 0;
 707 }