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 sstride, 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*sstride;
 101                 unsigned char *dptr = npix + y*dstride;
 102                 unsigned char ch;
 103                 for (x = 0; x < nw; ++x) {
 104                         float acc;
 105                         int tmp;
 106                         static const float low = 0.0, high = 255.0;
 107                         asm (
 108                                 "pxor %0, %0               \n"
 109                                 "xor %1, %1                \n"
 110                                 ".lbl2:                    \n"
 111                                 "movups (%3,%1),%%xmm1     \n"
 112                                 "movups (%2,%1),%%xmm2     \n"
 113                                 "mulps %%xmm2,%%xmm1       \n"
 114                                 "addps %%xmm1,%0           \n"
 115                                 "addl $16,%1              \n"
 116                                 "dec %4                    \n"
 117                                 "jnz .lbl2                 \n"
 118                                 "haddps %0,%0              \n"
 119                                 "haddps %0,%0              \n"
 120                                 "maxss %5,%0               \n"
 121                                 "minss %6,%0               \n"
 122                                 : "=&x" (acc),
 123                                   "=&r" (tmp)
 124                                 : "r" (&coeffs[pd[x].startcoeff]),
 125                                   "r" (&sptr[pd[x].start]),
 126                                   "r" ((pd[x].end - pd[x].start + 1)/4),
 127                                   "m" (low),
 128                                   "m" (high)
 129                                 : "xmm1", "xmm2"
 130                         );
 131
 132 #if 0
 133                         if (acc < 0.0)
 134                                 ch = 0;
 135                         else if (acc > 255.0)
 136                                 ch = 255;
 137                         else
 138                                 ch = (unsigned char)acc;
 139                         *dptr++ = ch;
 140 #endif
 141                         *dptr++ = (unsigned char)acc;
 142                 }
 143                 ch = dptr[-1];
 144                 for ( ; x < dstride; ++x) {
 145                         *dptr++ = ch;
 146                 }
 147         }
 148 }
 149
 150 void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh, unsigned dstride)
 151 {
 152         struct pix_desc *pd = (struct pix_desc *)malloc(nh * sizeof(struct pix_desc));
 153         int size_coeffs = 8;
 154         float *coeffs = (float *)malloc(size_coeffs * sizeof(float));
 155         int num_coeffs = 0;
 156         int x, y, sy;
 157         double sf = (double)h / (double)nh;
 158         double support = (h > nh) ? (3.0 * sf) : (3.0 / sf);
 159
 160         /* calculate the filter */
 161         for (y = 0; y < nh; ++y) {
 162                 int start = ceil(y * sf - support);
 163                 int end = floor(y * sf + support);
 164                 double sum = 0.0;
 165
 166                 if (start < 0) {
 167                         start = 0;
 168                 }
 169                 if (end > h - 1) {
 170                         end = h - 1;
 171                 }
 172
 173                 pd[y].start = start;
 174                 pd[y].end = end;
 175                 pd[y].startcoeff = num_coeffs;
 176
 177                 for (sy = start; sy <= end; ++sy) {
 178                         double nd = (h > nh) ? (sy/sf - y) : (sy - y*sf);
 179                         double f = lanczos_tap(nd);
 180                         if (num_coeffs == size_coeffs) {
 181                                 size_coeffs <<= 1;
 182                                 coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float));
 183                         }
 184
 185                         coeffs[num_coeffs++] = f;
 186                         sum += f;
 187                 }
 188
 189                 for (sy = start; sy <= end; ++sy) {
 190                         coeffs[pd[y].startcoeff + sy - start] /= sum;
 191                 }
 192         }
 193
 194 #if CACHE_LINE_FACTOR > 1
 195         for (x = 0; x < (w/CACHE_LINE_FACTOR) * CACHE_LINE_FACTOR; x += CACHE_LINE_FACTOR) {
 196                 unsigned char *sptr = pix + x;
 197                 float *dptr = npix + x;
 198                 for (y = 0; y < nh; ++y) {
 199 #if 0
 200                         int i;
 201                         float acc[CACHE_LINE_FACTOR];
 202                         for (i = 0; i < CACHE_LINE_FACTOR; ++i)
 203                                 acc[i] = 0.0;
 204                         float *cf = &coeffs[pd[y].startcoeff];
 205                         unsigned sy;
 206
 207                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 208                                 for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 209                                         acc[i] += sptr[sy * w + i] * *cf;
 210                                 }
 211                                 ++cf;
 212                         }
 213
 214                         for (i = 0; i < CACHE_LINE_FACTOR; ++i) {
 215                                 dptr[i] = acc[i];
 216                         }
 217 #else
 218                         /*
 219                          * xmm0 - xmm3: acc[0..15]
 220                          * xmm4: current filter coefficient
 221                          * xmm5, xmm6, xmm7: scratchpad
 222                          */
 223                         asm (
 224                                 /* clear */
 225                                 "pxor %%xmm0, %%xmm0    \n"
 226                                 "pxor %%xmm1, %%xmm1    \n"
 227                                 "pxor %%xmm2, %%xmm2    \n"
 228                                 "pxor %%xmm3, %%xmm3    \n"
 229
 230                                 /* main loop */
 231                                 ".lbl:                   \n"
 232
 233                                 /* a zero is useful during unpacking */
 234                                 "pxor %%xmm4, %%xmm4     \n"
 235
 236                                 /* fetch all 16 source bytes */
 237                                 "movups (%0), %%xmm5     \n"
 238                                 "prefetcht0 (%0,%3,4)    \n"
 239
 240                                 /* unpack into words (xmm5, xmm7) */
 241                                 "movaps %%xmm5, %%xmm7    \n"
 242                                 "punpcklbw %%xmm4, %%xmm5 \n"
 243                                 "punpckhbw %%xmm4, %%xmm7 \n"
 244
 245                                 /* unpack xmm5 into dwords (xmm5, xmm6) */
 246                                 "movaps %%xmm5, %%xmm6    \n"
 247                                 "punpcklwd %%xmm4, %%xmm5 \n"
 248                                 "punpckhwd %%xmm4, %%xmm6 \n"
 249
 250                                 /* convert xmm5, xmm6 to floats */
 251                                 "cvtdq2ps %%xmm5, %%xmm5 \n"
 252                                 "cvtdq2ps %%xmm6, %%xmm6 \n"
 253
 254                                 /* fetch the coefficient */
 255                                 "movss (%2), %%xmm4      \n"
 256                                 "shufps $0x0, %%xmm4, %%xmm4 \n"
 257
 258                                 /* do the muls for xmm5 and xmm6 */
 259                                 "mulps %%xmm4, %%xmm5    \n"
 260                                 "mulps %%xmm4, %%xmm6    \n"
 261                                 "addps %%xmm5, %%xmm0    \n"
 262                                 "addps %%xmm6, %%xmm1    \n"
 263
 264                                 /* get the zero back again */
 265                                 "pxor %%xmm4, %%xmm4     \n"
 266
 267                                 /* unpack xmm7 into dwords (xmm7, xmm6) */
 268                                 "movaps %%xmm7, %%xmm6    \n"
 269                                 "punpcklwd %%xmm4, %%xmm7 \n"
 270                                 "punpckhwd %%xmm4, %%xmm6 \n"
 271
 272                                 /* convert xmm7, xmm6 to floats */
 273                                 "cvtdq2ps %%xmm7, %%xmm7 \n"
 274                                 "cvtdq2ps %%xmm6, %%xmm6 \n"
 275
 276                                 /* fetch the coefficient */
 277                                 "movss (%2), %%xmm4      \n"
 278                                 "shufps $0x0, %%xmm4, %%xmm4 \n"
 279
 280                                 /* do the second set of muls */
 281                                 "mulps %%xmm4, %%xmm7    \n"
 282                                 "mulps %%xmm4, %%xmm6    \n"
 283                                 "addps %%xmm7, %%xmm2    \n"
 284                                 "addps %%xmm6, %%xmm3    \n"
 285
 286                                 /* move along, and loop */
 287                                 "add $4, %2              \n"
 288                                 "add %3, %0              \n"
 289                                 "dec %1                  \n"
 290                                 "jnz .lbl                \n"
 291
 292                                 /* store the values */
 293                                 "movaps %%xmm0, (%4)     \n"
 294                                 "movaps %%xmm1, 16(%4)   \n"
 295                                 "movaps %%xmm2, 32(%4)   \n"
 296                                 "movaps %%xmm3, 48(%4)   \n"
 297                                 : :
 298                                 "r" (&sptr[pd[y].start * w]),        /* 0: srcptr base */
 299                                 "r" (pd[y].end - pd[y].start + 1),   /* 1: filter len */
 300                                 "r" (&coeffs[pd[y].startcoeff]),     /* 2: coeffs base */
 301                                 "r" (w),                             /* 3: stride */
 302                                 "r" (dptr)                           /* 4: dstptr base */
 303                                 : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
 304                         );
 305 #endif
 306                         dptr += dstride;
 307                 }
 308         }
 309         for (x = (x/CACHE_LINE_FACTOR)*CACHE_LINE_FACTOR; x < w; ++x) {
 310 #else
 311         for (x = 0; x < w; ++x) {
 312 #endif
 313                 unsigned char *sptr = pix + x;
 314                 float *dptr = npix + x;
 315                 for (y = 0; y < nh; ++y) {
 316                         float acc = 0.0;
 317                         float *cf = &coeffs[pd[y].startcoeff];
 318                         unsigned sy;
 319
 320                         for (sy = pd[y].start; sy <= pd[y].end; ++sy) {
 321                                 acc += sptr[sy * w] * *cf++;
 322                         }
 323
 324                         *dptr = acc;
 325                         dptr += dstride;
 326                 }
 327         }
 328 }
 329
 330 int main(int argc, char **argv)
 331 {
 332         unsigned nominal_w = atoi(argv[1]);
 333         unsigned nominal_h = atoi(argv[2]);
 334
 335         unsigned samp_h0 = 2, samp_v0 = 2;
 336         unsigned samp_h1 = 1, samp_v1 = 1;
 337         unsigned samp_h2 = 1, samp_v2 = 1;
 338         unsigned max_samp_h = 2, max_samp_v = 2;
 339
 340         unsigned nw0 = nominal_w * samp_h0 / max_samp_h, nh0 = nominal_h * samp_v0 / max_samp_v;
 341         unsigned nw1 = nominal_w * samp_h1 / max_samp_h, nh1 = nominal_h * samp_v1 / max_samp_v;
 342         unsigned nw2 = nominal_w * samp_h2 / max_samp_h, nh2 = nominal_h * samp_v2 / max_samp_v;
 343
 344         unsigned stride0 = (nw0 + DCTSIZE-1) & ~(DCTSIZE-1);
 345         unsigned stride1 = (nw1 + DCTSIZE-1) & ~(DCTSIZE-1);
 346         unsigned stride2 = (nw2 + DCTSIZE-1) & ~(DCTSIZE-1);
 347
 348         struct jpeg_decompress_struct dinfo;
 349         struct jpeg_error_mgr jerr;
 350         dinfo.err = jpeg_std_error(&jerr);
 351         jpeg_create_decompress(&dinfo);
 352         jpeg_stdio_src(&dinfo, stdin);
 353         jpeg_read_header(&dinfo, TRUE);
 354         dinfo.raw_data_out = TRUE;
 355         jpeg_start_decompress(&dinfo);
 356
 357         unsigned w0 = dinfo.image_width * samp_h0 / max_samp_h, h0 = dinfo.image_height * samp_v0 / max_samp_v;
 358         unsigned w1 = dinfo.image_width * samp_h1 / max_samp_h, h1 = dinfo.image_height * samp_v1 / max_samp_v;
 359         unsigned w2 = dinfo.image_width * samp_h2 / max_samp_h, h2 = dinfo.image_height * samp_v2 / max_samp_v;
 360
 361         fprintf(stderr, "Scaling using Lanczos filter:\n");
 362         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);
 363         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);
 364         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);
 365
 366         JSAMPLE *data_y  = (JSAMPLE*)memalign(16, dinfo.comp_info[0].height_in_blocks * dinfo.comp_info[0].width_in_blocks * DCTSIZE * DCTSIZE);
 367         JSAMPLE *data_cb = (JSAMPLE*)memalign(16, dinfo.comp_info[1].height_in_blocks * dinfo.comp_info[1].width_in_blocks * DCTSIZE * DCTSIZE);
 368         JSAMPLE *data_cr = (JSAMPLE*)memalign(16, dinfo.comp_info[2].height_in_blocks * dinfo.comp_info[2].width_in_blocks * DCTSIZE * DCTSIZE);
 369         JSAMPLE *data_ny, *data_ncb, *data_ncr;
 370
 371         int total_lines = 0, blocks = 0;
 372         while (total_lines < dinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 373                 unsigned max_lines = dinfo.max_v_samp_factor * DCTSIZE;
 374
 375                 JSAMPROW y_row_ptrs[max_lines];
 376                 JSAMPROW cb_row_ptrs[max_lines];
 377                 JSAMPROW cr_row_ptrs[max_lines];
 378                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 379                 int i;
 380
 381                 for (i = 0; i < max_lines; ++i) {
 382                         y_row_ptrs[i]  = data_y  + (i+blocks*DCTSIZE*dinfo.comp_info[0].v_samp_factor) * dinfo.comp_info[0].width_in_blocks * DCTSIZE;
 383                         cb_row_ptrs[i] = data_cb + (i+blocks*DCTSIZE*dinfo.comp_info[1].v_samp_factor) * dinfo.comp_info[1].width_in_blocks * DCTSIZE;
 384                         cr_row_ptrs[i] = data_cr + (i+blocks*DCTSIZE*dinfo.comp_info[2].v_samp_factor) * dinfo.comp_info[2].width_in_blocks * DCTSIZE;
 385                 }
 386
 387                 total_lines += max_lines;
 388                 ++blocks;
 389
 390                 if (jpeg_read_raw_data(&dinfo, ptrs, max_lines) == 0)
 391                         break;
 392         }
 393
 394         {
 395                 float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float));
 396                 vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, h0, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE);
 397                 data_ny = (unsigned char *)malloc(nw0 * stride0);
 398                 hscale(npix, data_ny, w0, nh0, nw0, dinfo.comp_info[0].width_in_blocks * DCTSIZE, stride0);
 399                 free(npix);
 400         }
 401         {
 402                 float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float));
 403                 vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, h1, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE);
 404                 data_ncr = (unsigned char *)malloc(nw1 * stride1);
 405                 hscale(npix, data_ncr, w1, nh1, nw1, dinfo.comp_info[1].width_in_blocks * DCTSIZE, stride1);
 406                 free(npix);
 407         }
 408         {
 409                 float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float));
 410                 vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, h2, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE);
 411                 data_ncb = (unsigned char *)malloc(nw2 * stride2);
 412                 hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2);
 413                 free(npix);
 414         }
 415         jpeg_destroy_decompress(&dinfo);
 416
 417         struct jpeg_compress_struct cinfo;
 418         cinfo.err = jpeg_std_error(&jerr);
 419         jpeg_create_compress(&cinfo);
 420         jpeg_stdio_dest(&cinfo, stdout);
 421         cinfo.input_components = 3;
 422         jpeg_set_defaults(&cinfo);
 423         jpeg_set_quality(&cinfo, 85, FALSE);
 424         cinfo.image_width = nominal_w;
 425         cinfo.image_height = nominal_h;
 426         cinfo.raw_data_in = TRUE;
 427         jpeg_set_colorspace(&cinfo, JCS_YCbCr);
 428         cinfo.comp_info[0].h_samp_factor = samp_h0;
 429         cinfo.comp_info[0].v_samp_factor = samp_v0;
 430         cinfo.comp_info[1].h_samp_factor = samp_h1;
 431         cinfo.comp_info[1].v_samp_factor = samp_v1;
 432         cinfo.comp_info[2].h_samp_factor = samp_h2;
 433         cinfo.comp_info[2].v_samp_factor = samp_v2;
 434         jpeg_start_compress(&cinfo, TRUE);
 435
 436         total_lines = 0;
 437         blocks = 0;
 438         while (total_lines < cinfo.comp_info[0].height_in_blocks * DCTSIZE) {
 439                 unsigned max_lines = cinfo.max_v_samp_factor * DCTSIZE;
 440
 441                 JSAMPROW y_row_ptrs[max_lines];
 442                 JSAMPROW cb_row_ptrs[max_lines];
 443                 JSAMPROW cr_row_ptrs[max_lines];
 444                 JSAMPROW* ptrs[] = { y_row_ptrs, cb_row_ptrs, cr_row_ptrs };
 445                 int i;
 446
 447                 for (i = 0; i < max_lines; ++i) {
 448                         // simple edge extension
 449                         int yline = i + blocks*DCTSIZE*cinfo.comp_info[0].v_samp_factor;
 450                         if (yline > nh0 - 1)
 451                                 yline = nh0 - 1;
 452
 453                         int cbline = i + blocks*DCTSIZE*cinfo.comp_info[1].v_samp_factor;
 454                         if (cbline > nh1 - 1)
 455                                 cbline = nh1 - 1;
 456
 457                         int crline = i + blocks*DCTSIZE*cinfo.comp_info[2].v_samp_factor;
 458                         if (crline > nh2 - 1)
 459                                 crline = nh2 - 1;
 460
 461                         y_row_ptrs[i]  = data_ny  + yline * stride0;
 462                         cb_row_ptrs[i] = data_ncb + cbline * stride1;
 463                         cr_row_ptrs[i] = data_ncr + crline * stride2;
 464                 }
 465
 466                 total_lines += max_lines;
 467                 ++blocks;
 468
 469                 jpeg_write_raw_data(&cinfo, ptrs, max_lines);
 470         }
 471         jpeg_finish_compress(&cinfo);
 472         jpeg_destroy_compress(&cinfo);
 473
 474         return 0;
 475 }
 476