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->samp_h0 = dinfo.comp_info[0].h_samp_factor;
  78         img->samp_v0 = dinfo.comp_info[0].v_samp_factor;
  79
  80         img->samp_h1 = dinfo.comp_info[1].h_samp_factor;
  81         img->samp_v1 = dinfo.comp_info[1].v_samp_factor;
  82
  83         img->samp_h2 = dinfo.comp_info[2].h_samp_factor;
  84         img->samp_v2 = dinfo.comp_info[2].v_samp_factor;
  85
  86         img->data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
  87         img->data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
  88         img->data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
  89
  90         if (img->data_y == NULL || img->data_cb == NULL || img->data_cr == NULL) {
  91                 qscale_destroy(img);
  92                 return NULL;
  93         }
  94
  95         int total_lines = 0, blocks = 0;
  96         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
  97                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
  98
  99                 JSAMPROW y_row_ptrs[max_lines];
 100                 JSAMPROW cb_row_ptrs[max_lines];
 101                 JSAMPROW cr_row_ptrs[max_lines];
 102                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 103
 104                 int i;
 105                 for (i = 0; i < max_lines; ++i) {
 106                         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;
 107                         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;
 108                         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;
 109                 }
 110
 111                 total_lines += max_lines;
 112                 ++blocks;
 113
 114                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 115                         break;
 116         }
 117
 118         jpeg_destroy_decompress(&dinfo);
 119         return img;
 120 }
 121
 122 void qscale_destroy(qscale_img *img)
 123 {
 124         free(img->data_y);
 125         free(img->data_cb);
 126         free(img->data_cr);
 127         free(img);
 128 }
 129
 130
 131 static double sinc(double x)
 132 {
 133         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
 134
 135         if (abs(x) < cutoff) {
 136                 /* For small |x|, use Taylor series instead */
 137                 const double x2 = x * x;
 138                 const double x4 = x2 * x2;
 139
 140                 return 1.0 - x2 / 6.0 + x4 / 120.0;
 141         } else {
 142                 return sin(x) / x;
 143         }
 144 }
 145
 146 static double lanczos_tap(double x)
 147 {
 148         if (x < -3.0 || x > 3.0)
 149                 return 0.0;
 150         if (x < 0.0)
 151                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
 152         else
 153                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
 154 }
 155
 156 struct pix_desc {
 157         unsigned start, end;
 158         unsigned startcoeff;
 159 };
 160
 161 static void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
 162 {
 163         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
 164         int size_coeffs = 8;
 165         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 166         int num_coeffs = 0;
 167         int x, y;
 168         double sf = (double)w / (double)nw;
 169         double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
 170
 171         /* calculate the filter */
 172         for (x = 0; x < nw; ++x) {
 173                 int start = ceil(x * sf - support);
 174                 int end = floor(x * sf + support);
 175                 int sx;
 176                 double sum = 0.0;
 177
 178                 if (start < 0) {
 179                         start = 0;
 180                 }
 181                 if (end > w - 1) {
 182                         end = w - 1;
 183                 }
 184
 185 #if USE_HORIZONTAL_SSE
 186                 /* round up so we get a multiple of four for the SSE code */
 187                 int num = (end - start + 1);
 188                 if (num % 4 != 0) {
 189                         /* prefer aligning it if possible */
 190                         if (start % 4 != 0 && start % 4 <= num % 4) {
 191                                 num += start % 4;
 192                                 start -= start % 4;
 193                         }
 194                         if (num % 4 != 0) {
 195                                 end += 4 - (num % 4);
 196                         }
 197                 }
 198 #endif
 199
 200                 pd[x].start = start;
 201                 pd[x].end = end;
 202                 pd[x].startcoeff = num_coeffs;
 203
 204                 for (sx = start; sx <= end; ++sx) {
 205                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
 206                         double f = lanczos_tap(nd);
 207                         if (num_coeffs == size_coeffs) {
 208                                 size_coeffs <<= 1;
 209                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 210                         }
 211
 212                         coeffs[num_coeffs++] = f;
 213                         sum += f;
 214                 }
 215
 216                 for (sx = start; sx <= end; ++sx) {
 217                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 218                 }
 219         }
 220
 221         for (y = 0; y < h; ++y) {
 222                 float *sptr = pix + y*sstride;
 223                 unsigned char *dptr = npix + y*dstride;
 224                 unsigned char ch;
 225                 for (x = 0; x < nw; ++x) {
 226 #if USE_HORIZONTAL_SSE
 227                         int result;
 228                         float acc;
 229                         long tmp;
 230                         static const float low = 0.0, high = 255.0;
 231                         __asm__ (
 232                                 "pxor %1, %1               \n"
 233                                 "xor %2, %2                \n"
 234                                 "0:                        \n"
 235                                 "movups (%4,%2),%%xmm1     \n"
 236                                 "movups (%3,%2),%%xmm2     \n"
 237                                 "mulps %%xmm2,%%xmm1       \n"
 238                                 "addps %%xmm1,%1           \n"
 239                                 "add $16,%2                \n"
 240                                 "dec %5                    \n"
 241                                 "jnz 0b                    \n"
 242                                 "haddps %1,%1              \n"
 243                                 "haddps %1,%1              \n"
 244                                 "maxss %6,%1               \n"
 245                                 "minss %7,%1               \n"
 246                                 "cvtss2si %1,%0            \n"
 247                                 : "=r" (result),
 248                                   "=&x" (acc),
 249                                   "=&r" (tmp)
 250                                 : "r" (&coeffs[pd[x].startcoeff]),
 251                                   "r" (&sptr[pd[x].start]),
 252                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 253                                   "m" (low),
 254                                   "m" (high)
 255                                 : "memory", "xmm1", "xmm2"
 256                         );
 257
 258                         *dptr++ = (unsigned char)result;
 259 #else
 260                         float acc = 0.0;
 261                         float *cf = &coeffs[pd[x].startcoeff];
 262                         unsigned sx;
 263
 264                         for (sx = pd[x].start; sx <= pd[x].end; ++sx) {
 265                                 acc += sptr[sx] * *cf++;
 266                         }
 267
 268                         if (acc < 0.0)
 269                                 ch = 0;
 270                         else if (acc > 255.0)
 271                                 ch = 255;
 272                         else
 273                                 ch = (unsigned char)acc;
 274                         *dptr++ = ch;
 275 #endif
 276                 }
 277                 ch = dptr[-1];
 278                 for ( ; x < dstride; ++x) {
 279                         *dptr++ = ch;
 280                 }
 281         }
 282 }
 283
 284 static void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 285 {
 286         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 287         int size_coeffs = 8;
 288         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 289         int num_coeffs = 0;
 290         int x, y, sy;
 291         double sf = (double)h / (double)nh;
 292         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 293
 294         /* calculate the filter */
 295         for (y = 0; y < nh; ++y) {
 296                 int start = ceil(y * sf - support);
 297                 int end = floor(y * sf + support);
 298                 double sum = 0.0;
 299
 300                 if (start < 0) {
 301                         start = 0;
 302                 }
 303                 if (end > h - 1) {
 304                         end = h - 1;
 305                 }
 306
 307                 pd[y].start = start;
 308                 pd[y].end = end;
 309                 pd[y].startcoeff = num_coeffs;
 310
 311                 for (sy = start; sy <= end; ++sy) {
 312                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 313                         double f = lanczos_tap(nd);
 314                         if (num_coeffs == size_coeffs) {
 315                                 size_coeffs <<= 1;
 316                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 317                         }
 318
 319                         coeffs[num_coeffs++] = f;
 320                         sum += f;
 321                 }
 322
 323                 for (sy = start; sy <= end; ++sy) {
 324                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 325                 }
 326         }
 327
 328 #if CACHE_LINE_FACTOR > 1
 329         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 330                 unsigned char *sptr = pix + x;
 331                 float *dptr = npix + x;
 332                 for (y = 0; y < nh; ++y) {
 333 #if USE_VERTICAL_SSE
 334                         /*
 335                          * xmm0 - xmm3: acc[0..15]
 336                          * xmm4: current filter coefficient
 337                          * xmm5, xmm6, xmm7: scratchpad
 338                          */
 339                         __asm__ (
 340                                 /* clear */
 341                                 "pxor %%xmm0, %%xmm0          \n"
 342                                 "pxor %%xmm1, %%xmm1          \n"
 343                                 "pxor %%xmm2, %%xmm2          \n"
 344                                 "pxor %%xmm3, %%xmm3          \n"
 345
 346                                 /* main loop */
 347                                 "0:                           \n"
 348
 349                                 /* a zero is useful during unpacking */
 350                                 "pxor %%xmm4, %%xmm4          \n"
 351
 352                                 /* fetch all 16 source bytes */
 353                                 "movups (%0), %%xmm5          \n"
 354                                 "prefetcht0 (%0,%3,4)         \n"
 355
 356                                 /* unpack into words (xmm5, xmm7) */
 357                                 "movaps %%xmm5, %%xmm7        \n"
 358                                 "punpcklbw %%xmm4, %%xmm5     \n"
 359                                 "punpckhbw %%xmm4, %%xmm7     \n"
 360
 361                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 362                                 "movaps %%xmm5, %%xmm6        \n"
 363                                 "punpcklwd %%xmm4, %%xmm5     \n"
 364                                 "punpckhwd %%xmm4, %%xmm6     \n"
 365
 366                                 /* convert xmm5, xmm6 to floats */
 367                                 "cvtdq2ps %%xmm5, %%xmm5      \n"
 368                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 369
 370                                 /* fetch the coefficient */
 371                                 "movss (%2), %%xmm4           \n"
 372                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 373
 374                                 /* do the muls for xmm5 and xmm6 */
 375                                 "mulps %%xmm4, %%xmm5         \n"
 376                                 "mulps %%xmm4, %%xmm6         \n"
 377                                 "addps %%xmm5, %%xmm0         \n"
 378                                 "addps %%xmm6, %%xmm1         \n"
 379
 380                                 /* get the zero back again */
 381                                 "pxor %%xmm4, %%xmm4          \n"
 382
 383                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 384                                 "movaps %%xmm7, %%xmm6        \n"
 385                                 "punpcklwd %%xmm4, %%xmm7     \n"
 386                                 "punpckhwd %%xmm4, %%xmm6     \n"
 387
 388                                 /* convert xmm7, xmm6 to floats */
 389                                 "cvtdq2ps %%xmm7, %%xmm7      \n"
 390                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 391
 392                                 /* fetch the coefficient */
 393                                 "movss (%2), %%xmm4           \n"
 394                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 395
 396                                 /* do the second set of muls */
 397                                 "mulps %%xmm4, %%xmm7         \n"
 398                                 "mulps %%xmm4, %%xmm6         \n"
 399                                 "addps %%xmm7, %%xmm2         \n"
 400                                 "addps %%xmm6, %%xmm3         \n"
 401
 402                                 /* move along, and loop */
 403                                 "add $4, %2                   \n"
 404                                 "add %3, %0                   \n"
 405                                 "dec %1                       \n"
 406                                 "jnz 0b                       \n"
 407
 408                                 /* store the values */
 409                                 "movaps %%xmm0, (%4)          \n"
 410                                 "movaps %%xmm1, 16(%4)        \n"
 411                                 "movaps %%xmm2, 32(%4)        \n"
 412                                 "movaps %%xmm3, 48(%4)        \n"
 413                                 : :
 414                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 415                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 416                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 417                                 "r" ((long)w),                       /* 3: stride */
 418                                 "r" (dptr)                           /* 4: dstptr base */
 419                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 420                         );
 421 #else
 422                         int i;
 423                         float acc[CACHE_LINE_FACTOR];
 424                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 425                                 acc[i] = 0.0;
 426                         float *cf = &coeffs[pd[y].startcoeff];
 427                         unsigned sy;
 428
 429                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 430                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 431                                         acc[i] += sptr[sy * w + i] * *cf;
 432                                 }
 433                                 ++cf;
 434                         }
 435
 436                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 437                                 dptr[i] = acc[i];
 438                         }
 439 #endif
 440                         dptr += dstride;
 441                 }
 442         }
 443         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 444 #else
 445         for (x = 0; x < w; ++x) {
 446 #endif
 447                 unsigned char *sptr = pix + x;
 448                 float *dptr = npix + x;
 449                 for (y = 0; y < nh; ++y) {
 450                         float acc = 0.0;
 451                         float *cf = &coeffs[pd[y].startcoeff];
 452                         unsigned sy;
 453
 454                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 455                                 acc += sptr[sy * w] * *cf++;
 456                         }
 457
 458                         *dptr = acc;
 459                         dptr += dstride;
 460                 }
 461         }
 462 }
 463
 464 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)
 465 {
 466         qscale_img *dst = (qscale_img *)malloc(sizeof(qscale_img));
 467         if (dst == NULL) {
 468                 return NULL;
 469         }
 470
 471         dst->width = width;
 472         dst->height = height;
 473
 474         unsigned max_samp_h, max_samp_v;
 475         max_samp_h = samp_h0;
 476         if (samp_h1 > max_samp_h)
 477                 max_samp_h = samp_h1;
 478         if (samp_h2 > max_samp_h)
 479                 max_samp_h = samp_h2;
 480
 481         max_samp_v = samp_v0;
 482         if (samp_v1 > max_samp_v)
 483                 max_samp_v = samp_v1;
 484         if (samp_v2 > max_samp_v)
 485                 max_samp_v = samp_v2;
 486
 487         dst->w0 = width * samp_h0 / max_samp_h;
 488         dst->h0 = height * samp_v0 / max_samp_v;
 489
 490         dst->w1 = width * samp_h1 / max_samp_h;
 491         dst->h1 = height * samp_v1 / max_samp_v;
 492
 493         dst->w2 = width * samp_h2 / max_samp_h;
 494         dst->h2 = height * samp_v2 / max_samp_v;
 495
 496         dst->samp_h0 = samp_h0;
 497         dst->samp_v0 = samp_v0;
 498
 499         dst->samp_h1 = samp_h1;
 500         dst->samp_v1 = samp_v1;
 501
 502         dst->samp_h2 = samp_h2;
 503         dst->samp_v2 = samp_v2;
 504
 505         unsigned dstride0 = (dst->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 506         unsigned dstride1 = (dst->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 507         unsigned dstride2 = (dst->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 508
 509         unsigned sstride0 = (src->w0 + DCTSIZE-1) & ~(DCTSIZE-1);
 510         unsigned sstride1 = (src->w1 + DCTSIZE-1) & ~(DCTSIZE-1);
 511         unsigned sstride2 = (src->w2 + DCTSIZE-1) & ~(DCTSIZE-1);
 512
 513         /* FIXME: handle out-of-memory gracefully */
 514         {
 515                 float *npix = (float*)memalign(16, sstride0 * dst->h0 * sizeof(float));
 516                 vscale(src->data_y, npix, sstride0, src->h0, dst->h0, sstride0);
 517                 dst->data_y = (unsigned char *)malloc(dst->h0 * dstride0);
 518                 hscale(npix, dst->data_y, src->w0, dst->h0, dst->w0, sstride0, dstride0);
 519                 free(npix);
 520         }
 521         {
 522                 float *npix = (float*)memalign(16, sstride1 * dst->h1 * sizeof(float));
 523                 vscale(src->data_cr, npix, sstride1, src->h1, dst->h1, sstride1);
 524                 dst->data_cr = (unsigned char *)malloc(dst->h1 * dstride1);
 525                 hscale(npix, dst->data_cr, src->w1, dst->h1, dst->w1, sstride1, dstride1);
 526                 free(npix);
 527         }
 528         {
 529                 float *npix = (float*)memalign(16, sstride2 * dst->h2 * sizeof(float));
 530                 vscale(src->data_cb, npix, sstride2, src->h2, dst->h2, sstride2);
 531                 dst->data_cb = (unsigned char *)malloc(dst->h2 * dstride2);
 532                 hscale(npix, dst->data_cb, src->w2, dst->h2, dst->w2, sstride2, dstride2);
 533                 free(npix);
 534         }
 535
 536         return dst;
 537 }