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

   1 #include <stdio.h>
   2 #include <malloc.h>
   3 #include <math.h>
   4 #include <string.h>
   5 #include <stdlib.h>
   6 #include "jpeglib.h"
   7
   8 #define CACHE_LINE_FACTOR 16
   9
  10 #ifndef M_PI
  11 #define M_PI 3.14159265358979323846264
  12 #endif
  13
  14 double sinc(double x)
  15 {
  16         static const double cutoff = 1.220703668e-4;  /* sqrt(sqrt(eps)) */
  17
  18         if (abs(x) < cutoff) {
  19                 /* For small |x|, use Taylor series instead */
  20                 const double x2 = x * x;
  21                 const double x4 = x2 * x2;
  22
  23                 return 1.0 - x2 / 6.0 + x4 / 120.0;
  24         } else {
  25                 return sin(x) / x;
  26         }
  27 }
  28
  29 double lanczos_tap(double x)
  30 {
  31         if (x < -3.0 || x > 3.0)
  32                 return 0.0;
  33         if (x < 0.0)
  34                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
  35         else
  36                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
  37 }
  38
  39
  40 struct pix_desc {
  41         unsigned start, end;
  42         unsigned startcoeff;
  43 };
  44
  45 void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride)
  46 {
  47         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
  48         int size_coeffs = 8;
  49         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
  50         int num_coeffs = 0;
  51         int x, y;
  52         double sf = (double)w / (double)nw;
  53         double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
  54
  55         /* calculate the filter */
  56         for (x = 0; x < nw; ++x) {
  57                 int start = ceil(x * sf - support);
  58                 int end = floor(x * sf + support);
  59                 int sx;
  60                 double sum = 0.0;
  61
  62                 if (start < 0) {
  63                         start = 0;
  64                 }
  65                 if (end > w - 1) {
  66                         end = w - 1;
  67                 }
  68
  69                 /* round up so we get a multiple of four for the SSE code */
  70                 int num = (end - start + 1);
  71                 if (num % 4 != 0) {
  72                         /* prefer aligning it if possible */
  73                         if (start % 4 != 0 && start % 4 <= num % 4) {
  74                                 num += start % 4;
  75                                 start -= start % 4;
  76                         }
  77                         if (num % 4 != 0) {
  78                                 end += 4 - (num % 4);
  79                         }
  80                 }
  81
  82                 pd[x].start = start;
  83                 pd[x].end = end;
  84                 pd[x].startcoeff = num_coeffs;
  85
  86                 for (sx = start; sx <= end; ++sx) {
  87                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
  88                         double f = lanczos_tap(nd);
  89                         if (num_coeffs == size_coeffs) {
  90                                 size_coeffs <<= 1;
  91                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
  92                         }
  93
  94                         coeffs[num_coeffs++] = f;
  95                         sum += f;
  96                 }
  97
  98                 for (sx = start; sx <= end; ++sx) {
  99                         coeffs[pd[x].startcoeff + sx - start] /= sum;
 100                 }
 101         }
 102
 103         for (y = 0; y < h; ++y) {
 104                 float *sptr = pix + y*sstride;
 105                 unsigned char *dptr = npix + y*dstride;
 106                 unsigned char ch;
 107                 for (x = 0; x < nw; ++x) {
 108                         int result;
 109                         float acc;
 110                         long tmp;
 111                         static const float low = 0.0, high = 255.0;
 112                         __asm__ (
 113                                 "pxor %1, %1               \n"
 114                                 "xor %2, %2                \n"
 115                                 "0:                        \n"
 116                                 "movups (%4,%2),%%xmm1     \n"
 117                                 "movups (%3,%2),%%xmm2     \n"
 118                                 "mulps %%xmm2,%%xmm1       \n"
 119                                 "addps %%xmm1,%1           \n"
 120                                 "add $16,%2                \n"
 121                                 "dec %5                    \n"
 122                                 "jnz 0b                    \n"
 123                                 "haddps %1,%1              \n"
 124                                 "haddps %1,%1              \n"
 125                                 "maxss %6,%1               \n"
 126                                 "minss %7,%1               \n"
 127                                 "cvtss2si %1,%0            \n"
 128                                 : "=r" (result),
 129                                   "=&x" (acc),
 130                                   "=&r" (tmp)
 131                                 : "r" (&coeffs[pd[x].startcoeff]),
 132                                   "r" (&sptr[pd[x].start]),
 133                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 134                                   "m" (low),
 135                                   "m" (high)
 136                                 : "memory", "xmm1", "xmm2"
 137                         );
 138
 139 #if 0
 140                         if (acc < 0.0)
 141                                 ch = 0;
 142                         else if (acc > 255.0)
 143                                 ch = 255;
 144                         else
 145                                 ch = (unsigned char)acc;
 146                         *dptr++ = ch;
 147 #endif
 148                         *dptr++ = (unsigned char)result;
 149                 }
 150                 ch = dptr[-1];
 151                 for ( ; x < dstride; ++x) {
 152                         *dptr++ = ch;
 153                 }
 154         }
 155 }
 156
 157 void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 158 {
 159         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 160         int size_coeffs = 8;
 161         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 162         int num_coeffs = 0;
 163         int x, y, sy;
 164         double sf = (double)h / (double)nh;
 165         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 166
 167         /* calculate the filter */
 168         for (y = 0; y < nh; ++y) {
 169                 int start = ceil(y * sf - support);
 170                 int end = floor(y * sf + support);
 171                 double sum = 0.0;
 172
 173                 if (start < 0) {
 174                         start = 0;
 175                 }
 176                 if (end > h - 1) {
 177                         end = h - 1;
 178                 }
 179
 180                 pd[y].start = start;
 181                 pd[y].end = end;
 182                 pd[y].startcoeff = num_coeffs;
 183
 184                 for (sy = start; sy <= end; ++sy) {
 185                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 186                         double f = lanczos_tap(nd);
 187                         if (num_coeffs == size_coeffs) {
 188                                 size_coeffs <<= 1;
 189                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 190                         }
 191
 192                         coeffs[num_coeffs++] = f;
 193                         sum += f;
 194                 }
 195
 196                 for (sy = start; sy <= end; ++sy) {
 197                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 198                 }
 199         }
 200
 201 #if CACHE_LINE_FACTOR > 1
 202         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 203                 unsigned char *sptr = pix + x;
 204                 float *dptr = npix + x;
 205                 for (y = 0; y < nh; ++y) {
 206 #if 0
 207                         int i;
 208                         float acc[CACHE_LINE_FACTOR];
 209                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 210                                 acc[i] = 0.0;
 211                         float *cf = &coeffs[pd[y].startcoeff];
 212                         unsigned sy;
 213
 214                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 215                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 216                                         acc[i] += sptr[sy * w + i] * *cf;
 217                                 }
 218                                 ++cf;
 219                         }
 220
 221                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 222                                 dptr[i] = acc[i];
 223                         }
 224 #else
 225                         /*
 226                          * xmm0 - xmm3: acc[0..15]
 227                          * xmm4: current filter coefficient
 228                          * xmm5, xmm6, xmm7: scratchpad
 229                          */
 230                         __asm__ (
 231                                 /* clear */
 232                                 "pxor %%xmm0, %%xmm0          \n"
 233                                 "pxor %%xmm1, %%xmm1          \n"
 234                                 "pxor %%xmm2, %%xmm2          \n"
 235                                 "pxor %%xmm3, %%xmm3          \n"
 236
 237                                 /* main loop */
 238                                 "0:                           \n"
 239
 240                                 /* a zero is useful during unpacking */
 241                                 "pxor %%xmm4, %%xmm4          \n"
 242
 243                                 /* fetch all 16 source bytes */
 244                                 "movups (%0), %%xmm5          \n"
 245                                 "prefetcht0 (%0,%3,4)         \n"
 246
 247                                 /* unpack into words (xmm5, xmm7) */
 248                                 "movaps %%xmm5, %%xmm7        \n"
 249                                 "punpcklbw %%xmm4, %%xmm5     \n"
 250                                 "punpckhbw %%xmm4, %%xmm7     \n"
 251
 252                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 253                                 "movaps %%xmm5, %%xmm6        \n"
 254                                 "punpcklwd %%xmm4, %%xmm5     \n"
 255                                 "punpckhwd %%xmm4, %%xmm6     \n"
 256
 257                                 /* convert xmm5, xmm6 to floats */
 258                                 "cvtdq2ps %%xmm5, %%xmm5      \n"
 259                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 260
 261                                 /* fetch the coefficient */
 262                                 "movss (%2), %%xmm4           \n"
 263                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 264
 265                                 /* do the muls for xmm5 and xmm6 */
 266                                 "mulps %%xmm4, %%xmm5         \n"
 267                                 "mulps %%xmm4, %%xmm6         \n"
 268                                 "addps %%xmm5, %%xmm0         \n"
 269                                 "addps %%xmm6, %%xmm1         \n"
 270
 271                                 /* get the zero back again */
 272                                 "pxor %%xmm4, %%xmm4          \n"
 273
 274                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 275                                 "movaps %%xmm7, %%xmm6        \n"
 276                                 "punpcklwd %%xmm4, %%xmm7     \n"
 277                                 "punpckhwd %%xmm4, %%xmm6     \n"
 278
 279                                 /* convert xmm7, xmm6 to floats */
 280                                 "cvtdq2ps %%xmm7, %%xmm7      \n"
 281                                 "cvtdq2ps %%xmm6, %%xmm6      \n"
 282
 283                                 /* fetch the coefficient */
 284                                 "movss (%2), %%xmm4           \n"
 285                                 "shufps $0x0, %%xmm4, %%xmm4  \n"
 286
 287                                 /* do the second set of muls */
 288                                 "mulps %%xmm4, %%xmm7         \n"
 289                                 "mulps %%xmm4, %%xmm6         \n"
 290                                 "addps %%xmm7, %%xmm2         \n"
 291                                 "addps %%xmm6, %%xmm3         \n"
 292
 293                                 /* move along, and loop */
 294                                 "add $4, %2                   \n"
 295                                 "add %3, %0                   \n"
 296                                 "dec %1                       \n"
 297                                 "jnz 0b                       \n"
 298
 299                                 /* store the values */
 300                                 "movaps %%xmm0, (%4)          \n"
 301                                 "movaps %%xmm1, 16(%4)        \n"
 302                                 "movaps %%xmm2, 32(%4)        \n"
 303                                 "movaps %%xmm3, 48(%4)        \n"
 304                                 : :
 305                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 306                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 307                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 308                                 "r" ((long)w),                       /* 3: stride */
 309                                 "r" (dptr)                           /* 4: dstptr base */
 310                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 311                         );
 312 #endif
 313                         dptr += dstride;
 314                 }
 315         }
 316         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 317 #else
 318         for (x = 0; x < w; ++x) {
 319 #endif
 320                 unsigned char *sptr = pix + x;
 321                 float *dptr = npix + x;
 322                 for (y = 0; y < nh; ++y) {
 323                         float acc = 0.0;
 324                         float *cf = &coeffs[pd[y].startcoeff];
 325                         unsigned sy;
 326
 327                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 328                                 acc += sptr[sy * w] * *cf++;
 329                         }
 330
 331                         *dptr = acc;
 332                         dptr += dstride;
 333                 }
 334         }
 335 }
 336
 337 int main(int argc, char **argv)
 338 {
 339         /* user-settable parameters */
 340         unsigned nominal_w = atoi(argv[1]);
 341         unsigned nominal_h = atoi(argv[2]);
 342         unsigned samp_h0 = 2, samp_v0 = 2;
 343         unsigned samp_h1 = 1, samp_v1 = 1;
 344         unsigned samp_h2 = 1, samp_v2 = 1;
 345         unsigned jpeg_quality = 85;
 346         /* end */
 347
 348         unsigned max_samp_h, max_samp_v;
 349         max_samp_h = samp_h0;
 350         if (samp_h1 > max_samp_h)
 351                 max_samp_h = samp_h1;
 352         if (samp_h2 > max_samp_h)
 353                 max_samp_h = samp_h2;
 354
 355         max_samp_v = samp_v0;
 356         if (samp_v1 > max_samp_v)
 357                 max_samp_v = samp_v1;
 358         if (samp_v2 > max_samp_v)
 359                 max_samp_v = samp_v2;
 360
 361         unsigned nw0 = nominal_w * samp_h0 / max_samp_h, nh0 = nominal_h * samp_v0 / max_samp_v;
 362         unsigned nw1 = nominal_w * samp_h1 / max_samp_h, nh1 = nominal_h * samp_v1 / max_samp_v;
 363         unsigned nw2 = nominal_w * samp_h2 / max_samp_h, nh2 = nominal_h * samp_v2 / max_samp_v;
 364
 365         unsigned stride0 = (nw0 + DCTSIZE-1) & ~(DCTSIZE-1);
 366         unsigned stride1 = (nw1 + DCTSIZE-1) & ~(DCTSIZE-1);
 367         unsigned stride2 = (nw2 + DCTSIZE-1) & ~(DCTSIZE-1);
 368
 369         struct jpeg_decompress_struct dinfo;
 370         struct jpeg_error_mgr jerr;
 371         dinfo.err = jpeg_std_error(&jerr);
 372         jpeg_create_decompress(&dinfo);
 373         jpeg_stdio_src(&dinfo, stdin);
 374         jpeg_read_header(&dinfo, TRUE);
 375         dinfo.raw_data_out = TRUE;
 376         jpeg_start_decompress(&dinfo);
 377
 378         unsigned w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor;
 379         unsigned h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor;
 380
 381         unsigned w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor;
 382         unsigned h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor;
 383
 384         unsigned w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor;
 385         unsigned h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor;
 386
 387         fprintf(stderr, "Scaling using Lanczos filter:\n");
 388         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);
 389         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);
 390         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);
 391
 392         JSAMPLE *data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 393         JSAMPLE *data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 394         JSAMPLE *data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 395         JSAMPLE *data_ny, *data_ncb, *data_ncr;
 396
 397         int total_lines = 0, blocks = 0;
 398         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 399                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 400
 401                 JSAMPROW y_row_ptrs[max_lines];
 402                 JSAMPROW cb_row_ptrs[max_lines];
 403                 JSAMPROW cr_row_ptrs[max_lines];
 404                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 405                 int i;
 406
 407                 for (i = 0; i < max_lines; ++i) {
 408                         y_row_ptrs[i]  = data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
 409                         cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
 410                         cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
 411                 }
 412
 413                 total_lines += max_lines;
 414                 ++blocks;
 415
 416                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 417                         break;
 418         }
 419
 420         {
 421                 float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));
 422                 vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, h0, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
 423                 data_ny = (unsigned char *)malloc(nh0 * stride0);
 424                 hscale(npix, data_ny, w0, nh0, nw0, dinfo.comp_info[0].width_in_blocks * DCTSIZE, stride0);
 425                 free(npix);
 426         }
 427         {
 428                 float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));
 429                 vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, h1, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
 430                 data_ncr = (unsigned char *)malloc(nh1 * stride1);
 431                 hscale(npix, data_ncr, w1, nh1, nw1, dinfo.comp_info[1].width_in_blocks * DCTSIZE, stride1);
 432                 free(npix);
 433         }
 434         {
 435                 float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));
 436                 vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, h2, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
 437                 data_ncb = (unsigned char *)malloc(nh2 * stride2);
 438                 hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2);
 439                 free(npix);
 440         }
 441         jpeg_destroy_decompress(&dinfo);
 442
 443         struct jpeg_compress_struct cinfo;
 444         cinfo.err = jpeg_std_error(&jerr);
 445         jpeg_create_compress(&cinfo);
 446         jpeg_stdio_dest(&cinfo, stdout);
 447         cinfo.input_components = 3;
 448         jpeg_set_defaults(&cinfo);
 449         jpeg_set_quality(&cinfo, jpeg_quality, FALSE);
 450         cinfo.image_width = nominal_w;
 451         cinfo.image_height = nominal_h;
 452         cinfo.raw_data_in = TRUE;
 453         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 454         cinfo.comp_info[0].h_samp_factor = samp_h0;
 455         cinfo.comp_info[0].v_samp_factor = samp_v0;
 456         cinfo.comp_info[1].h_samp_factor = samp_h1;
 457         cinfo.comp_info[1].v_samp_factor = samp_v1;
 458         cinfo.comp_info[2].h_samp_factor = samp_h2;
 459         cinfo.comp_info[2].v_samp_factor = samp_v2;
 460         jpeg_start_compress(&cinfo, TRUE);
 461
 462         total_lines = 0;
 463         blocks = 0;
 464         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 465                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 466
 467                 JSAMPROW y_row_ptrs[max_lines];
 468                 JSAMPROW cb_row_ptrs[max_lines];
 469                 JSAMPROW cr_row_ptrs[max_lines];
 470                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 471                 int i;
 472
 473                 for (i = 0; i < max_lines; ++i) {
 474                         /* simple edge extension */
 475                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 476                         if (yline > nh0 - 1)
 477                                 yline = nh0 - 1;
 478
 479                         int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 480                         if (cbline > nh1 - 1)
 481                                 cbline = nh1 - 1;
 482
 483                         int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 484                         if (crline > nh2 - 1)
 485                                 crline = nh2 - 1;
 486
 487                         y_row_ptrs[i]  = data_ny  + yline * stride0;
 488                         cb_row_ptrs[i] = data_ncb + cbline * stride1;
 489                         cr_row_ptrs[i] = data_ncr + crline * stride2;
 490                 }
 491
 492                 total_lines += max_lines;
 493                 ++blocks;
 494
 495                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 496         }
 497         jpeg_finish_compress(&cinfo);
 498         jpeg_destroy_compress(&cinfo);
 499
 500         return 0;
 501 }
 502