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 double sinc(double x)
  11 {
  12         static const double cutoff = 1.220703668e-4;  // sqrt(sqrt(eps))
  13
  14         if (abs(x) < cutoff) {
  15                 // For small |x|, use Taylor series instead
  16                 const double x2 = x * x;
  17                 const double x4 = x2 * x2;
  18
  19                 return 1.0 - x2 / 6.0 + x4 / 120.0;
  20         } else {
  21                 return sin(x) / x;
  22         }
  23 }
  24
  25 double lanczos_tap(double x)
  26 {
  27         if (x < -3.0 || x > 3.0)
  28                 return 0.0;
  29         if (x < 0.0)
  30                 return sinc(-x*M_PI) * sinc(-x*M_PI / 3.0);
  31         else
  32                 return sinc(x*M_PI) * sinc(x*M_PI / 3.0);
  33 }
  34
  35
  36 struct pix_desc {
  37         unsigned start, end;
  38         unsigned startcoeff;
  39 };
  40
  41 void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned dstride)
  42 {
  43         struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc));
  44         int size_coeffs = 8;
  45         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
  46         int num_coeffs = 0;
  47         int x, y;
  48         double sf = (double)w / (double)nw;
  49         double support = (w > nw) ? (3.0 * sf) : (3.0 / sf);
  50
  51         /* calculate the filter */
  52         for (x = 0; x < nw; ++x) {
  53                 int start = ceil(x * sf - support);
  54                 int end = floor(x * sf + support);
  55                 int sx;
  56                 double sum = 0.0;
  57
  58                 if (start < 0) {
  59                         start = 0;
  60                 }
  61                 if (end > w - 1) {
  62                         end = w - 1;
  63                 }
  64
  65                 // round up so we get a multiple of four for the SSE code
  66                 int num = (end - start + 1);
  67                 if (num % 4 != 0) {
  68                         // prefer aligning it if possible
  69                         if (start % 4 != 0 && start % 4 <= num % 4) {
  70                                 num += start % 4;
  71                                 start -= start % 4;
  72                         }
  73                         if (num % 4 != 0) {
  74                                 end += 4 - (num % 4);
  75                         }
  76                 }
  77
  78                 pd[x].start = start;
  79                 pd[x].end = end;
  80                 pd[x].startcoeff = num_coeffs;
  81
  82                 for (sx = start; sx <= end; ++sx) {
  83                         double nd = (w > nw) ? (sx/sf - x) : (sx - x*sf);
  84                         double f = lanczos_tap(nd);
  85                         if (num_coeffs == size_coeffs) {
  86                                 size_coeffs <<= 1;
  87                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
  88                         }
  89
  90                         coeffs[num_coeffs++] = f;
  91                         sum += f;
  92                 }
  93
  94                 for (sx = start; sx <= end; ++sx) {
  95                         coeffs[pd[x].startcoeff + sx - start] /= sum;
  96                 }
  97         }
  98
  99         for (y = 0; y < h; ++y) {
 100                 float *sptr = pix + y*w;
 101                 unsigned char *dptr = npix + y*dstride;
 102                 unsigned char ch;
 103                 for (x = 0; x < nw; ++x) {
 104                         float acc;
 105                         static const float low = 0.0, high = 255.0;
 106                         asm (
 107                                 "pxor %0, %0               \n"
 108                                 "xor %%eax, %%eax          \n"
 109                                 ".lbl2:                    \n"
 110                                 "movups (%2,%%eax),%%xmm1  \n"
 111                                 "movups (%1,%%eax),%%xmm2  \n"
 112                                 "mulps %%xmm2,%%xmm1       \n"
 113                                 "addps %%xmm1,%0           \n"
 114                                 "addl $16,%%eax            \n"
 115                                 "dec %3                    \n"
 116                                 "jnz .lbl2                 \n"
 117                                 "haddps %0,%0              \n"
 118                                 "haddps %0,%0              \n"
 119                                 "maxss %4,%0               \n"
 120                                 "minss %5,%0               \n"
 121                                 : "=x" (acc)
 122                                 : "r" (&coeffs[pd[x].startcoeff]),
 123                                   "r" (&sptr[pd[x].start]),
 124                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 125                                   "m" (low),
 126                                   "m" (high)
 127                                 : "eax", "xmm1", "xmm2"
 128                         );
 129
 130 #if 0
 131                         if (acc < 0.0)
 132                                 ch = 0;
 133                         else if (acc > 255.0)
 134                                 ch = 255;
 135                         else
 136                                 ch = (unsigned char)acc;
 137                         *dptr++ = ch;
 138 #endif
 139                         *dptr++ = (unsigned char)acc;
 140                 }
 141                 ch = dptr[-1];
 142                 for ( ; x < dstride; ++x) {
 143                         *dptr++ = ch;
 144                 }
 145         }
 146 }
 147
 148 void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 149 {
 150         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 151         int size_coeffs = 8;
 152         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 153         int num_coeffs = 0;
 154         int x, y, sy;
 155         double sf = (double)h / (double)nh;
 156         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 157
 158         /* calculate the filter */
 159         for (y = 0; y < nh; ++y) {
 160                 int start = ceil(y * sf - support);
 161                 int end = floor(y * sf + support);
 162                 double sum = 0.0;
 163
 164                 if (start < 0) {
 165                         start = 0;
 166                 }
 167                 if (end > h - 1) {
 168                         end = h - 1;
 169                 }
 170
 171                 pd[y].start = start;
 172                 pd[y].end = end;
 173                 pd[y].startcoeff = num_coeffs;
 174
 175                 for (sy = start; sy <= end; ++sy) {
 176                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 177                         double f = lanczos_tap(nd);
 178                         if (num_coeffs == size_coeffs) {
 179                                 size_coeffs <<= 1;
 180                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 181                         }
 182
 183                         coeffs[num_coeffs++] = f;
 184                         sum += f;
 185                 }
 186
 187                 for (sy = start; sy <= end; ++sy) {
 188                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 189                 }
 190         }
 191
 192 #if CACHE_LINE_FACTOR > 1
 193         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 194                 unsigned char *sptr = pix + x;
 195                 float *dptr = npix + x;
 196                 for (y = 0; y < nh; ++y) {
 197 #if 0
 198                         int i;
 199                         float acc[CACHE_LINE_FACTOR];
 200                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 201                                 acc[i] = 0.0;
 202                         float *cf = &coeffs[pd[y].startcoeff];
 203                         unsigned sy;
 204
 205                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 206                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 207                                         acc[i] += sptr[sy * w + i] * *cf;
 208                                 }
 209                                 ++cf;
 210                         }
 211
 212                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 213                                 dptr[i] = acc[i];
 214                         }
 215 #else
 216                         /*
 217                          * xmm0 - xmm3: acc[0..15]
 218                          * xmm4: current filter coefficient
 219                          * xmm5, xmm6, xmm7: scratchpad
 220                          */
 221                         asm (
 222                                 /* clear */
 223                                 "pxor %%xmm0, %%xmm0    \n"
 224                                 "pxor %%xmm1, %%xmm1    \n"
 225                                 "pxor %%xmm2, %%xmm2    \n"
 226                                 "pxor %%xmm3, %%xmm3    \n"
 227
 228                                 /* main loop */
 229                                 ".lbl:                   \n"
 230
 231                                 /* a zero is useful during unpacking */
 232                                 "pxor %%xmm4, %%xmm4     \n"
 233
 234                                 /* fetch all 16 source bytes */
 235                                 "movups (%0), %%xmm5     \n"
 236                                 "prefetcht0 (%0,%3,4)    \n"
 237
 238                                 /* unpack into words (xmm5, xmm7) */
 239                                 "movaps %%xmm5, %%xmm7    \n"
 240                                 "punpcklbw %%xmm4, %%xmm5 \n"
 241                                 "punpckhbw %%xmm4, %%xmm7 \n"
 242
 243                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 244                                 "movaps %%xmm5, %%xmm6    \n"
 245                                 "punpcklwd %%xmm4, %%xmm5 \n"
 246                                 "punpckhwd %%xmm4, %%xmm6 \n"
 247
 248                                 /* convert xmm5, xmm6 to floats */
 249                                 "cvtdq2ps %%xmm5, %%xmm5 \n"
 250                                 "cvtdq2ps %%xmm6, %%xmm6 \n"
 251
 252                                 /* fetch the coefficient */
 253                                 "movss (%2), %%xmm4      \n"
 254                                 "shufps $0x0, %%xmm4, %%xmm4 \n"
 255
 256                                 /* do the muls for xmm5 and xmm6 */
 257                                 "mulps %%xmm4, %%xmm5    \n"
 258                                 "mulps %%xmm4, %%xmm6    \n"
 259                                 "addps %%xmm5, %%xmm0    \n"
 260                                 "addps %%xmm6, %%xmm1    \n"
 261
 262                                 /* get the zero back again */
 263                                 "pxor %%xmm4, %%xmm4     \n"
 264
 265                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 266                                 "movaps %%xmm7, %%xmm6    \n"
 267                                 "punpcklwd %%xmm4, %%xmm7 \n"
 268                                 "punpckhwd %%xmm4, %%xmm6 \n"
 269
 270                                 /* convert xmm7, xmm6 to floats */
 271                                 "cvtdq2ps %%xmm7, %%xmm7 \n"
 272                                 "cvtdq2ps %%xmm6, %%xmm6 \n"
 273
 274                                 /* fetch the coefficient */
 275                                 "movss (%2), %%xmm4      \n"
 276                                 "shufps $0x0, %%xmm4, %%xmm4 \n"
 277
 278                                 /* do the second set of muls */
 279                                 "mulps %%xmm4, %%xmm7    \n"
 280                                 "mulps %%xmm4, %%xmm6    \n"
 281                                 "addps %%xmm7, %%xmm2    \n"
 282                                 "addps %%xmm6, %%xmm3    \n"
 283
 284                                 /* move along, and loop */
 285                                 "add $4, %2              \n"
 286                                 "add %3, %0              \n"
 287                                 "dec %1                  \n"
 288                                 "jnz .lbl                \n"
 289
 290                                 /* store the values */
 291                                 "movaps %%xmm0, (%4)     \n"
 292                                 "movaps %%xmm1, 16(%4)   \n"
 293                                 "movaps %%xmm2, 32(%4)   \n"
 294                                 "movaps %%xmm3, 48(%4)   \n"
 295                                 : :
 296                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 297                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 298                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 299                                 "r" (w),                             /* 3: stride */
 300                                 "r" (dptr)                           /* 4: dstptr base */
 301                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 302                         );
 303 #endif
 304                         dptr += dstride;
 305                 }
 306         }
 307         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 308 #else
 309         for (x = 0; x < w; ++x) {
 310 #endif
 311                 unsigned char *sptr = pix + x;
 312                 float *dptr = npix + x;
 313                 for (y = 0; y < nh; ++y) {
 314                         float acc = 0.0;
 315                         float *cf = &coeffs[pd[y].startcoeff];
 316                         unsigned sy;
 317
 318                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 319                                 acc += sptr[sy * w] * *cf++;
 320                         }
 321
 322                         *dptr = acc;
 323                         dptr += dstride;
 324                 }
 325         }
 326 }
 327
 328 int main(int argc, char **argv)
 329 {
 330         unsigned nominal_w = atoi(argv[1]);
 331         unsigned nominal_h = atoi(argv[2]);
 332
 333         unsigned samp_h0 = 2, samp_v0 = 2;
 334         unsigned samp_h1 = 1, samp_v1 = 1;
 335         unsigned samp_h2 = 1, samp_v2 = 1;
 336         unsigned max_samp_h = 2, max_samp_v = 2;
 337
 338         unsigned nw0 = nominal_w * samp_h0 / max_samp_h, nh0 = nominal_h * samp_v0 / max_samp_v;
 339         unsigned nw1 = nominal_w * samp_h1 / max_samp_h, nh1 = nominal_h * samp_v1 / max_samp_v;
 340         unsigned nw2 = nominal_w * samp_h2 / max_samp_h, nh2 = nominal_h * samp_v2 / max_samp_v;
 341
 342         unsigned stride0 = (nw0 + DCTSIZE-1) & ~(DCTSIZE-1);
 343         unsigned stride1 = (nw1 + DCTSIZE-1) & ~(DCTSIZE-1);
 344         unsigned stride2 = (nw2 + DCTSIZE-1) & ~(DCTSIZE-1);
 345
 346         struct jpeg_decompress_struct dinfo;
 347         struct jpeg_error_mgr jerr;
 348         dinfo.err = jpeg_std_error(&jerr);
 349         jpeg_create_decompress(&dinfo);
 350         jpeg_stdio_src(&dinfo, stdin);
 351         jpeg_read_header(&dinfo, TRUE);
 352         dinfo.raw_data_out = TRUE;
 353         jpeg_start_decompress(&dinfo);
 354
 355         fprintf(stderr, "Scaling using Lanczos filter:\n");
 356         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);
 357         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);
 358         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);
 359
 360         JSAMPLE *data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 361         JSAMPLE *data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 362         JSAMPLE *data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 363         JSAMPLE *data_ny, *data_ncb, *data_ncr;
 364
 365         int total_lines = 0, blocks = 0;
 366         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 367                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 368
 369                 JSAMPROW y_row_ptrs[max_lines];
 370                 JSAMPROW cb_row_ptrs[max_lines];
 371                 JSAMPROW cr_row_ptrs[max_lines];
 372                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 373                 int i;
 374
 375                 for (i = 0; i < max_lines; ++i) {
 376                         y_row_ptrs[i]  = data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
 377                         cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
 378                         cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
 379                 }
 380
 381                 total_lines += max_lines;
 382                 ++blocks;
 383
 384                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 385                         break;
 386         }
 387
 388         {
 389                 float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));
 390                 vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
 391                 data_ny = (unsigned char *)malloc(nw0 * stride0);
 392                 hscale(npix, data_ny, dinfo.comp_info[0].width_in_blocks * DCTSIZE, nh0, nw0, stride0);
 393                 free(npix);
 394         }
 395         {
 396                 float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));
 397                 vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
 398                 data_ncr = (unsigned char *)malloc(nw1 * stride1);
 399                 hscale(npix, data_ncr, dinfo.comp_info[1].width_in_blocks * DCTSIZE, nh1, nw1, stride1);
 400                 free(npix);
 401         }
 402         {
 403                 float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));
 404                 vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
 405                 data_ncb = (unsigned char *)malloc(nw2 * stride2);
 406                 hscale(npix, data_ncb, dinfo.comp_info[2].width_in_blocks * DCTSIZE, nh2, nw2, stride2);
 407                 free(npix);
 408         }
 409         jpeg_destroy_decompress(&dinfo);
 410
 411         struct jpeg_compress_struct cinfo;
 412         cinfo.err = jpeg_std_error(&jerr);
 413         jpeg_create_compress(&cinfo);
 414         jpeg_stdio_dest(&cinfo, stdout);
 415         cinfo.input_components = 3;
 416         jpeg_set_defaults(&cinfo);
 417         jpeg_set_quality(&cinfo, 85, FALSE);
 418         cinfo.image_width = nominal_w;
 419         cinfo.image_height = nominal_h;
 420         cinfo.raw_data_in = TRUE;
 421         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 422         cinfo.comp_info[0].h_samp_factor = samp_h0;
 423         cinfo.comp_info[0].v_samp_factor = samp_v0;
 424         cinfo.comp_info[1].h_samp_factor = samp_h1;
 425         cinfo.comp_info[1].v_samp_factor = samp_v1;
 426         cinfo.comp_info[2].h_samp_factor = samp_h2;
 427         cinfo.comp_info[2].v_samp_factor = samp_v2;
 428         jpeg_start_compress(&cinfo, TRUE);
 429
 430         total_lines = 0;
 431         blocks = 0;
 432         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 433                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 434
 435                 JSAMPROW y_row_ptrs[max_lines];
 436                 JSAMPROW cb_row_ptrs[max_lines];
 437                 JSAMPROW cr_row_ptrs[max_lines];
 438                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 439                 int i;
 440
 441                 for (i = 0; i < max_lines; ++i) {
 442                         // simple edge extension
 443                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 444                         if (yline > nh0 - 1)
 445                                 yline = nh0 - 1;
 446
 447                         int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 448                         if (cbline > nh1 - 1)
 449                                 cbline = nh1 - 1;
 450
 451                         int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 452                         if (crline > nh2 - 1)
 453                                 crline = nh2 - 1;
 454
 455                         y_row_ptrs[i]  = data_ny  + yline * stride0;
 456                         cb_row_ptrs[i] = data_ncb + cbline * stride1;
 457                         cr_row_ptrs[i] = data_ncr + crline * stride2;
 458                 }
 459
 460                 total_lines += max_lines;
 461                 ++blocks;
 462
 463                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 464         }
 465         jpeg_finish_compress(&cinfo);
 466         jpeg_destroy_compress(&cinfo);
 467
 468         return 0;
 469 }
 470