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

   1 #include <math.h>
   2 #include <stdio.h>
   3 #include <stdlib.h>
   4 #include <malloc.h>
   5
   6 #include "libqscale.h"
   7
   8 /* The number of pixels to process at a time when scaling vertically. */
   9 #define CACHE_LINE_FACTOR 16
  10
  11 /* Whether to use SSE for horizontal scaling or not (requires SSE3). */
  12 #define USE_HORIZONTAL_SSE 1
  13
  14 /* Whether to use SSE for vertical scaling or not (requires only SSE1). */
  15 #define USE_VERTICAL_SSE 1
  16
  17 #if USE_VERTICAL_SSE
  18 #undef CACHE_LINE_FACTOR
  19 #define CACHE_LINE_FACTOR 16
  20 #endif
  21
  22 #ifndef M_PI
  23 #define M_PI 3.14159265358979323846264
  24 #endif
  25
  26 qscale_img *qscale_load_jpeg(const char *filename)
  27 {
  28         FILE *file = fopen(filename, "rb");
  29         qscale_img *img;
  30         if (file == NULL) {
  31                 return NULL;
  32         }
  33
  34         img = qscale_load_jpeg_from_stdio(file);
  35
  36         fclose(file);
  37         return img;
  38 }
  39
  40 qscale_img *qscale_load_jpeg_from_stdio(FILE *file)
  41 {
  42         qscale_img *img = (qscale_img *)malloc(sizeof(qscale_img));
  43         if (img == NULL) {
  44                 return NULL;
  45         }
  46
  47         img->data_y = img->data_cb = img->data_cr = NULL;
  48
  49         /* FIXME: Better error handling here (ie., return NULL). */
  50         struct jpeg_decompress_struct dinfo;
  51         struct jpeg_error_mgr jerr;
  52         dinfo.err = jpeg_std_error(&jerr);
  53         jpeg_create_decompress(&dinfo);
  54         jpeg_stdio_src(&dinfo, stdin);
  55         jpeg_read_header(&dinfo, TRUE);
  56         dinfo.raw_data_out = TRUE;
  57         jpeg_start_decompress(&dinfo);
  58
  59         /* We do not handle anything but YCbCr images (yet?). */
  60         if (dinfo.num_components != 3) {
  61                 qscale_destroy(img);
  62                 return NULL;
  63         }
  64
  65         img->width = dinfo.image_width;
  66         img->height = dinfo.image_height;
  67
  68         img->w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
  69         img->h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
  70
  71         img->w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
  72         img->h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
  73
  74         img->w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
  75         img->h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
  76
  77         img->data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
  78         img->data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
  79         img->data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
  80
  81         if (img->data_y == NULL || img->data_cb == NULL || img->data_cr == NULL) {
  82                 qscale_destroy(img);
  83                 return NULL;
  84         }
  85
  86         int total_lines = 0, blocks = 0;
  87         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
  88                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
  89
  90                 JSAMPROW y_row_ptrs[max_lines];
  91                 JSAMPROW cb_row_ptrs[max_lines];
  92                 JSAMPROW cr_row_ptrs[max_lines];
  93                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
  94
  95                 int i;
  96                 for (i = 0; i < max_lines; ++i) {
  97                         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;
  98                         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;
  99                         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;
 100                 }
 101
 102                 total_lines += max_lines;
 103                 ++blocks;
 104
 105                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 106                         break;
 107         }
 108
 109         jpeg_destroy_decompress(&dinfo);
 110         return img;
 111 }
 112
 113 void qscale_destroy(qscale_img *img)
 114 {
 115         free(img->data_y);
 116         free(img->data_cb);
 117         free(img->data_cr);
 118         free(img);
 119 }
 120
 121
 122 static double sinc(double x)
 123 {
 124         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
 125
 126         if (abs(x) < cutoff) {
 127                 /* For small |x|, use Taylor series instead */
 128                 const double x2 = x * x;
 129                 const double x4 = x2 * x2;
 130
 131                 return 1.0 - x2 / 6.0 + x4 / 120.0;
 132         } else {
 133                 return sin(x) / x;
 134         }
 135 }
 136
 137 static double lanczos_tap(double x)
 138 {
 139         if (x < -3.0 || x > 3.0)
 140                 return 0.0;
 141         if (x < 0.0)
 142                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
 143         else
 144                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
 145 }
 146
 147 struct pix_desc {
 148         unsigned start, end;
 149         unsigned startcoeff;
 150 };
 151
 152 static void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
 153 {
 154         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
 155         int size_coeffs = 8;
 156         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 157         int num_coeffs = 0;
 158         int x, y;
 159         double sf = (double)w / (double)nw;
 160         double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
 161
 162         /* calculate the filter */
 163         for (x = 0; x < nw; ++x) {
 164                 int start = ceil(x * sf - support);
 165                 int end = floor(x * sf + support);
 166                 int sx;
 167                 double sum = 0.0;
 168
 169                 if (start < 0) {
 170                         start = 0;
 171                 }
 172                 if (end > w - 1) {
 173                         end = w - 1;
 174                 }
 175
 176 #if USE_HORIZONTAL_SSE
 177                 /* round up so we get a multiple of four for the SSE code */
 178                 int num = (end - start + 1);
 179                 if (num % 4 != 0) {
 180                         /* prefer aligning it if possible */
 181                         if (start % 4 != 0 && start % 4 <= num % 4) {
 182                                 num += start % 4;
 183                                 start -= start % 4;
 184                         }
 185                         if (num % 4 != 0) {
 186                                 end += 4 - (num % 4);
 187                         }
 188                 }
 189 #endif
 190
 191                 pd[x].start = start;
 192                 pd[x].end = end;
 193                 pd[x].startcoeff = num_coeffs;
 194
 195                 for (sx = start; sx <= end; ++sx) {
 196                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
 197                         double f = lanczos_tap(nd);
 198                         if (num_coeffs == size_coeffs) {
 199                                 size_coeffs <<= 1;
 200                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 201                         }
 202
 203                         coeffs[num_coeffs++] = f;
 204                         sum += f;
 205                 }
 206
 207                 for (sx = start; sx <= end; ++sx) {
 208                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 209                 }
 210         }
 211
 212         for (y = 0; y < h; ++y) {
 213                 float *sptr = pix + y*sstride;
 214                 unsigned char *dptr = npix + y*dstride;
 215                 unsigned char ch;
 216                 for (x = 0; x < nw; ++x) {
 217 #if USE_HORIZONTAL_SSE
 218                         int result;
 219                         float acc;
 220                         long tmp;
 221                         static const float low = 0.0, high = 255.0;
 222                         __asm__ (
 223                                 "pxor %1, %1               \n"
 224                                 "xor %2, %2                \n"
 225                                 "0:                        \n"
 226                                 "movups (%4,%2),%%xmm1     \n"
 227                                 "movups (%3,%2),%%xmm2     \n"
 228                                 "mulps %%xmm2,%%xmm1       \n"
 229                                 "addps %%xmm1,%1           \n"
 230                                 "add $16,%2                \n"
 231                                 "dec %5                    \n"
 232                                 "jnz 0b                    \n"
 233                                 "haddps %1,%1              \n"
 234                                 "haddps %1,%1              \n"
 235                                 "maxss %6,%1               \n"
 236                                 "minss %7,%1               \n"
 237                                 "cvtss2si %1,%0            \n"
 238                                 : "=r" (result),
 239                                   "=&x" (acc),
 240                                   "=&r" (tmp)
 241                                 : "r" (&coeffs[pd[x].startcoeff]),
 242                                   "r" (&sptr[pd[x].start]),
 243                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 244                                   "m" (low),
 245                                   "m" (high)
 246                                 : "memory", "xmm1", "xmm2"
 247                         );
 248
 249                         *dptr++ = (unsigned char)result;
 250 #else
 251                         float acc = 0.0;
 252                         float *cf = &coeffs[pd[x].startcoeff];
 253                         unsigned sx;
 254
 255                         for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
 256                                 acc += sptr[sx] * *cf++;
 257                         }
 258
 259                         if (acc < 0.0)
 260                                 ch = 0;
 261                         else if (acc > 255.0)
 262                                 ch = 255;
 263                         else
 264                                 ch = (unsigned char)acc;
 265                         *dptr++ = ch;
 266 #endif
 267                 }
 268                 ch = dptr[-1];
 269                 for ( ; x < dstride; ++x) {
 270                         *dptr++ = ch;
 271                 }
 272         }
 273 }
 274
 275 static void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 276 {
 277         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 278         int size_coeffs = 8;
 279         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 280         int num_coeffs = 0;
 281         int x, y, sy;
 282         double sf = (double)h / (double)nh;
 283         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 284
 285         /* calculate the filter */
 286         for (y = 0; y < nh; ++y) {
 287                 int start = ceil(y * sf - support);
 288                 int end = floor(y * sf + support);
 289                 double sum = 0.0;
 290
 291                 if (start < 0) {
 292                         start = 0;
 293                 }
 294                 if (end > h - 1) {
 295                         end = h - 1;
 296                 }
 297
 298                 pd[y].start = start;
 299                 pd[y].end = end;
 300                 pd[y].startcoeff = num_coeffs;
 301
 302                 for (sy = start; sy <= end; ++sy) {
 303                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 304                         double f = lanczos_tap(nd);
 305                         if (num_coeffs == size_coeffs) {
 306                                 size_coeffs <<= 1;
 307                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 308                         }
 309
 310                         coeffs[num_coeffs++] = f;
 311                         sum += f;
 312                 }
 313
 314                 for (sy = start; sy <= end; ++sy) {
 315                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 316                 }
 317         }
 318
 319 #if CACHE_LINE_FACTOR > 1
 320         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 321                 unsigned char *sptr = pix + x;
 322                 float *dptr = npix + x;
 323                 for (y = 0; y < nh; ++y) {
 324 #if USE_VERTICAL_SSE
 325                         /*
 326                          * xmm0 - xmm3: acc[0..15]
 327                          * xmm4: current filter coefficient
 328                          * xmm5, xmm6, xmm7: scratchpad
 329                          */
 330                         __asm__ (
 331                                 /* clear */
 332                                 "pxor %%xmm0, %%xmm0          \n"
 333                                 "pxor %%xmm1, %%xmm1          \n"
 334                                 "pxor %%xmm2, %%xmm2          \n"
 335                                 "pxor %%xmm3, %%xmm3          \n"
 336
 337                                 /* main loop */
 338                                 "0:                           \n"
 339
 340                                 /* a zero is useful during unpacking */
 341                                 "pxor %%xmm4, %%xmm4          \n"
 342
 343                                 /* fetch all 16 source bytes */
 344                                 "movups (%0), %%xmm5          \n"
 345                                 "prefetcht0 (%0,%3,4)         \n"
 346
 347                                 /* unpack into words (xmm5, xmm7) */
 348                                 "movaps %%xmm5, %%xmm7        \n"
 349                                 "punpcklbw %%xmm4, %%xmm5     \n"
 350                                 "punpckhbw %%xmm4, %%xmm7     \n"
 351
 352                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 353                                 "movaps %%xmm5, %%xmm6        \n"
 354                                 "punpcklwd %%xmm4, %%xmm5     \n"
 355                                 "punpckhwd %%xmm4, %%xmm6     \n"
 356
 357                                 /* convert xmm5, xmm6 to floats */
 358                                 "cvtdq2ps %%xmm5, %%xmm5      \n"
 359                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 360
 361                                 /* fetch the coefficient */
 362                                 "movss (%2), %%xmm4           \n"
 363                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 364
 365                                 /* do the muls for xmm5 and xmm6 */
 366                                 "mulps %%xmm4, %%xmm5         \n"
 367                                 "mulps %%xmm4, %%xmm6         \n"
 368                                 "addps %%xmm5, %%xmm0         \n"
 369                                 "addps %%xmm6, %%xmm1         \n"
 370
 371                                 /* get the zero back again */
 372                                 "pxor %%xmm4, %%xmm4          \n"
 373
 374                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 375                                 "movaps %%xmm7, %%xmm6        \n"
 376                                 "punpcklwd %%xmm4, %%xmm7     \n"
 377                                 "punpckhwd %%xmm4, %%xmm6     \n"
 378
 379                                 /* convert xmm7, xmm6 to floats */
 380                                 "cvtdq2ps %%xmm7, %%xmm7      \n"
 381                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 382
 383                                 /* fetch the coefficient */
 384                                 "movss (%2), %%xmm4           \n"
 385                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 386
 387                                 /* do the second set of muls */
 388                                 "mulps %%xmm4, %%xmm7         \n"
 389                                 "mulps %%xmm4, %%xmm6         \n"
 390                                 "addps %%xmm7, %%xmm2         \n"
 391                                 "addps %%xmm6, %%xmm3         \n"
 392
 393                                 /* move along, and loop */
 394                                 "add $4, %2                   \n"
 395                                 "add %3, %0                   \n"
 396                                 "dec %1                       \n"
 397                                 "jnz 0b                       \n"
 398
 399                                 /* store the values */
 400                                 "movaps %%xmm0, (%4)          \n"
 401                                 "movaps %%xmm1, 16(%4)        \n"
 402                                 "movaps %%xmm2, 32(%4)        \n"
 403                                 "movaps %%xmm3, 48(%4)        \n"
 404                                 : :
 405                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 406                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 407                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 408                                 "r" ((long)w),                       /* 3: stride */
 409                                 "r" (dptr)                           /* 4: dstptr base */
 410                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 411                         );
 412 #else
 413                         int i;
 414                         float acc[CACHE_LINE_FACTOR];
 415                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 416                                 acc[i] = 0.0;
 417                         float *cf = &coeffs[pd[y].startcoeff];
 418                         unsigned sy;
 419
 420                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 421                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 422                                         acc[i] += sptr[sy * w + i] * *cf;
 423                                 }
 424                                 ++cf;
 425                         }
 426
 427                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 428                                 dptr[i] = acc[i];
 429                         }
 430 #endif
 431                         dptr += dstride;
 432                 }
 433         }
 434         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 435 #else
 436         for (x = 0; x < w; ++x) {
 437 #endif
 438                 unsigned char *sptr = pix + x;
 439                 float *dptr = npix + x;
 440                 for (y = 0; y < nh; ++y) {
 441                         float acc = 0.0;
 442                         float *cf = &coeffs[pd[y].startcoeff];
 443                         unsigned sy;
 444
 445                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 446                                 acc += sptr[sy * w] * *cf++;
 447                         }
 448
 449                         *dptr = acc;
 450                         dptr += dstride;
 451                 }
 452         }
 453 }
 454
 455 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)
 456 {
 457         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 458         if (dst == NULL) {
 459                 return NULL;
 460         }
 461
 462         dst->width = width;
 463         dst->height = height;
 464
 465         unsigned max_samp_h, max_samp_v;
 466         max_samp_h = samp_h0;
 467         if (samp_h1 > max_samp_h)
 468                 max_samp_h = samp_h1;
 469         if (samp_h2 > max_samp_h)
 470                 max_samp_h = samp_h2;
 471
 472         max_samp_v = samp_v0;
 473         if (samp_v1 > max_samp_v)
 474                 max_samp_v = samp_v1;
 475         if (samp_v2 > max_samp_v)
 476                 max_samp_v = samp_v2;
 477
 478         dst->w0 = width * samp_h0 / max_samp_h;
 479         dst->h0 = height * samp_v0 / max_samp_v;
 480
 481         dst->w1 = width * samp_h1 / max_samp_h;
 482         dst->h1 = height * samp_v1 / max_samp_v;
 483
 484         dst->w2 = width * samp_h2 / max_samp_h;
 485         dst->h2 = height * samp_v2 / max_samp_v;
 486
 487         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 488         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 489         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 490
 491         unsigned sstride0 = (src->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 492         unsigned sstride1 = (src->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 493         unsigned sstride2 = (src->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 494
 495         /* FIXME: handle out-of-memory gracefully */
 496         {
 497                 float *npix = (float*)memalign(16, sstride0 * dst->h0 * sizeof(float));
 498                 vscale(src->data_y, npix, sstride0, src->h0, dst->h0, sstride0);
 499                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 500                 hscale(npix, dst->data_y, src->w0, dst->h0, dst->w0, sstride0, dstride0);
 501                 free(npix);
 502         }
 503         {
 504                 float *npix = (float*)memalign(16, sstride1 * dst->h1 * sizeof(float));
 505                 vscale(src->data_cr, npix, sstride1, src->h1, dst->h1, sstride1);
 506                 dst->data_cr = (unsigned char *)malloc(dst->h1 * dstride1);
 507                 hscale(npix, dst->data_cr, src->w1, dst->h1, dst->w1, sstride1, dstride1);
 508                 free(npix);
 509         }
 510         {
 511                 float *npix = (float*)memalign(16, sstride2 * dst->h2 * sizeof(float));
 512                 vscale(src->data_cb, npix, sstride2, src->h2, dst->h2, sstride2);
 513                 dst->data_cb = (unsigned char *)malloc(dst->h2 * dstride2);
 514                 hscale(npix, dst->data_cb, src->w2, dst->h2, dst->w2, sstride2, dstride2);
 515                 free(npix);
 516         }
 517
 518         return dst;
 519 }