istereo

annotate libs/libjpeg/jidctfst.c @ 28:c0ae8e668447

find changesets by author, revision, files, or words in the commit message
added vmath library
author John Tsiombikas <nuclear@mutantstargoat.com>
date Thu, 08 Sep 2011 08:30:42 +0300
parents
children
rev   line source
nuclear@26 1 /*
nuclear@26 2 * jidctfst.c
nuclear@26 3 *
nuclear@26 4 * Copyright (C) 1994-1998, Thomas G. Lane.
nuclear@26 5 * This file is part of the Independent JPEG Group's software.
nuclear@26 6 * For conditions of distribution and use, see the accompanying README file.
nuclear@26 7 *
nuclear@26 8 * This file contains a fast, not so accurate integer implementation of the
nuclear@26 9 * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
nuclear@26 10 * must also perform dequantization of the input coefficients.
nuclear@26 11 *
nuclear@26 12 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
nuclear@26 13 * on each row (or vice versa, but it's more convenient to emit a row at
nuclear@26 14 * a time). Direct algorithms are also available, but they are much more
nuclear@26 15 * complex and seem not to be any faster when reduced to code.
nuclear@26 16 *
nuclear@26 17 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
nuclear@26 18 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
nuclear@26 19 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
nuclear@26 20 * JPEG textbook (see REFERENCES section in file README). The following code
nuclear@26 21 * is based directly on figure 4-8 in P&M.
nuclear@26 22 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
nuclear@26 23 * possible to arrange the computation so that many of the multiplies are
nuclear@26 24 * simple scalings of the final outputs. These multiplies can then be
nuclear@26 25 * folded into the multiplications or divisions by the JPEG quantization
nuclear@26 26 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
nuclear@26 27 * to be done in the DCT itself.
nuclear@26 28 * The primary disadvantage of this method is that with fixed-point math,
nuclear@26 29 * accuracy is lost due to imprecise representation of the scaled
nuclear@26 30 * quantization values. The smaller the quantization table entry, the less
nuclear@26 31 * precise the scaled value, so this implementation does worse with high-
nuclear@26 32 * quality-setting files than with low-quality ones.
nuclear@26 33 */
nuclear@26 34
nuclear@26 35 #define JPEG_INTERNALS
nuclear@26 36 #include "jinclude.h"
nuclear@26 37 #include "jpeglib.h"
nuclear@26 38 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@26 39
nuclear@26 40 #ifdef DCT_IFAST_SUPPORTED
nuclear@26 41
nuclear@26 42
nuclear@26 43 /*
nuclear@26 44 * This module is specialized to the case DCTSIZE = 8.
nuclear@26 45 */
nuclear@26 46
nuclear@26 47 #if DCTSIZE != 8
nuclear@26 48 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@26 49 #endif
nuclear@26 50
nuclear@26 51
nuclear@26 52 /* Scaling decisions are generally the same as in the LL&M algorithm;
nuclear@26 53 * see jidctint.c for more details. However, we choose to descale
nuclear@26 54 * (right shift) multiplication products as soon as they are formed,
nuclear@26 55 * rather than carrying additional fractional bits into subsequent additions.
nuclear@26 56 * This compromises accuracy slightly, but it lets us save a few shifts.
nuclear@26 57 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
nuclear@26 58 * everywhere except in the multiplications proper; this saves a good deal
nuclear@26 59 * of work on 16-bit-int machines.
nuclear@26 60 *
nuclear@26 61 * The dequantized coefficients are not integers because the AA&N scaling
nuclear@26 62 * factors have been incorporated. We represent them scaled up by PASS1_BITS,
nuclear@26 63 * so that the first and second IDCT rounds have the same input scaling.
nuclear@26 64 * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
nuclear@26 65 * avoid a descaling shift; this compromises accuracy rather drastically
nuclear@26 66 * for small quantization table entries, but it saves a lot of shifts.
nuclear@26 67 * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
nuclear@26 68 * so we use a much larger scaling factor to preserve accuracy.
nuclear@26 69 *
nuclear@26 70 * A final compromise is to represent the multiplicative constants to only
nuclear@26 71 * 8 fractional bits, rather than 13. This saves some shifting work on some
nuclear@26 72 * machines, and may also reduce the cost of multiplication (since there
nuclear@26 73 * are fewer one-bits in the constants).
nuclear@26 74 */
nuclear@26 75
nuclear@26 76 #if BITS_IN_JSAMPLE == 8
nuclear@26 77 #define CONST_BITS 8
nuclear@26 78 #define PASS1_BITS 2
nuclear@26 79 #else
nuclear@26 80 #define CONST_BITS 8
nuclear@26 81 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */
nuclear@26 82 #endif
nuclear@26 83
nuclear@26 84 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
nuclear@26 85 * causing a lot of useless floating-point operations at run time.
nuclear@26 86 * To get around this we use the following pre-calculated constants.
nuclear@26 87 * If you change CONST_BITS you may want to add appropriate values.
nuclear@26 88 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
nuclear@26 89 */
nuclear@26 90
nuclear@26 91 #if CONST_BITS == 8
nuclear@26 92 #define FIX_1_082392200 ((INT32) 277) /* FIX(1.082392200) */
nuclear@26 93 #define FIX_1_414213562 ((INT32) 362) /* FIX(1.414213562) */
nuclear@26 94 #define FIX_1_847759065 ((INT32) 473) /* FIX(1.847759065) */
nuclear@26 95 #define FIX_2_613125930 ((INT32) 669) /* FIX(2.613125930) */
nuclear@26 96 #else
nuclear@26 97 #define FIX_1_082392200 FIX(1.082392200)
nuclear@26 98 #define FIX_1_414213562 FIX(1.414213562)
nuclear@26 99 #define FIX_1_847759065 FIX(1.847759065)
nuclear@26 100 #define FIX_2_613125930 FIX(2.613125930)
nuclear@26 101 #endif
nuclear@26 102
nuclear@26 103
nuclear@26 104 /* We can gain a little more speed, with a further compromise in accuracy,
nuclear@26 105 * by omitting the addition in a descaling shift. This yields an incorrectly
nuclear@26 106 * rounded result half the time...
nuclear@26 107 */
nuclear@26 108
nuclear@26 109 #ifndef USE_ACCURATE_ROUNDING
nuclear@26 110 #undef DESCALE
nuclear@26 111 #define DESCALE(x,n) RIGHT_SHIFT(x, n)
nuclear@26 112 #endif
nuclear@26 113
nuclear@26 114
nuclear@26 115 /* Multiply a DCTELEM variable by an INT32 constant, and immediately
nuclear@26 116 * descale to yield a DCTELEM result.
nuclear@26 117 */
nuclear@26 118
nuclear@26 119 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
nuclear@26 120
nuclear@26 121
nuclear@26 122 /* Dequantize a coefficient by multiplying it by the multiplier-table
nuclear@26 123 * entry; produce a DCTELEM result. For 8-bit data a 16x16->16
nuclear@26 124 * multiplication will do. For 12-bit data, the multiplier table is
nuclear@26 125 * declared INT32, so a 32-bit multiply will be used.
nuclear@26 126 */
nuclear@26 127
nuclear@26 128 #if BITS_IN_JSAMPLE == 8
nuclear@26 129 #define DEQUANTIZE(coef,quantval) (((IFAST_MULT_TYPE) (coef)) * (quantval))
nuclear@26 130 #else
nuclear@26 131 #define DEQUANTIZE(coef,quantval) \
nuclear@26 132 DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
nuclear@26 133 #endif
nuclear@26 134
nuclear@26 135
nuclear@26 136 /* Like DESCALE, but applies to a DCTELEM and produces an int.
nuclear@26 137 * We assume that int right shift is unsigned if INT32 right shift is.
nuclear@26 138 */
nuclear@26 139
nuclear@26 140 #ifdef RIGHT_SHIFT_IS_UNSIGNED
nuclear@26 141 #define ISHIFT_TEMPS DCTELEM ishift_temp;
nuclear@26 142 #if BITS_IN_JSAMPLE == 8
nuclear@26 143 #define DCTELEMBITS 16 /* DCTELEM may be 16 or 32 bits */
nuclear@26 144 #else
nuclear@26 145 #define DCTELEMBITS 32 /* DCTELEM must be 32 bits */
nuclear@26 146 #endif
nuclear@26 147 #define IRIGHT_SHIFT(x,shft) \
nuclear@26 148 ((ishift_temp = (x)) < 0 ? \
nuclear@26 149 (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
nuclear@26 150 (ishift_temp >> (shft)))
nuclear@26 151 #else
nuclear@26 152 #define ISHIFT_TEMPS
nuclear@26 153 #define IRIGHT_SHIFT(x,shft) ((x) >> (shft))
nuclear@26 154 #endif
nuclear@26 155
nuclear@26 156 #ifdef USE_ACCURATE_ROUNDING
nuclear@26 157 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
nuclear@26 158 #else
nuclear@26 159 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT(x, n))
nuclear@26 160 #endif
nuclear@26 161
nuclear@26 162
nuclear@26 163 /*
nuclear@26 164 * Perform dequantization and inverse DCT on one block of coefficients.
nuclear@26 165 */
nuclear@26 166
nuclear@26 167 GLOBAL(void)
nuclear@26 168 jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
nuclear@26 169 JCOEFPTR coef_block,
nuclear@26 170 JSAMPARRAY output_buf, JDIMENSION output_col)
nuclear@26 171 {
nuclear@26 172 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
nuclear@26 173 DCTELEM tmp10, tmp11, tmp12, tmp13;
nuclear@26 174 DCTELEM z5, z10, z11, z12, z13;
nuclear@26 175 JCOEFPTR inptr;
nuclear@26 176 IFAST_MULT_TYPE * quantptr;
nuclear@26 177 int * wsptr;
nuclear@26 178 JSAMPROW outptr;
nuclear@26 179 JSAMPLE *range_limit = IDCT_range_limit(cinfo);
nuclear@26 180 int ctr;
nuclear@26 181 int workspace[DCTSIZE2]; /* buffers data between passes */
nuclear@26 182 SHIFT_TEMPS /* for DESCALE */
nuclear@26 183 ISHIFT_TEMPS /* for IDESCALE */
nuclear@26 184
nuclear@26 185 /* Pass 1: process columns from input, store into work array. */
nuclear@26 186
nuclear@26 187 inptr = coef_block;
nuclear@26 188 quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
nuclear@26 189 wsptr = workspace;
nuclear@26 190 for (ctr = DCTSIZE; ctr > 0; ctr--) {
nuclear@26 191 /* Due to quantization, we will usually find that many of the input
nuclear@26 192 * coefficients are zero, especially the AC terms. We can exploit this
nuclear@26 193 * by short-circuiting the IDCT calculation for any column in which all
nuclear@26 194 * the AC terms are zero. In that case each output is equal to the
nuclear@26 195 * DC coefficient (with scale factor as needed).
nuclear@26 196 * With typical images and quantization tables, half or more of the
nuclear@26 197 * column DCT calculations can be simplified this way.
nuclear@26 198 */
nuclear@26 199
nuclear@26 200 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
nuclear@26 201 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
nuclear@26 202 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
nuclear@26 203 inptr[DCTSIZE*7] == 0) {
nuclear@26 204 /* AC terms all zero */
nuclear@26 205 int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
nuclear@26 206
nuclear@26 207 wsptr[DCTSIZE*0] = dcval;
nuclear@26 208 wsptr[DCTSIZE*1] = dcval;
nuclear@26 209 wsptr[DCTSIZE*2] = dcval;
nuclear@26 210 wsptr[DCTSIZE*3] = dcval;
nuclear@26 211 wsptr[DCTSIZE*4] = dcval;
nuclear@26 212 wsptr[DCTSIZE*5] = dcval;
nuclear@26 213 wsptr[DCTSIZE*6] = dcval;
nuclear@26 214 wsptr[DCTSIZE*7] = dcval;
nuclear@26 215
nuclear@26 216 inptr++; /* advance pointers to next column */
nuclear@26 217 quantptr++;
nuclear@26 218 wsptr++;
nuclear@26 219 continue;
nuclear@26 220 }
nuclear@26 221
nuclear@26 222 /* Even part */
nuclear@26 223
nuclear@26 224 tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
nuclear@26 225 tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
nuclear@26 226 tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
nuclear@26 227 tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
nuclear@26 228
nuclear@26 229 tmp10 = tmp0 + tmp2; /* phase 3 */
nuclear@26 230 tmp11 = tmp0 - tmp2;
nuclear@26 231
nuclear@26 232 tmp13 = tmp1 + tmp3; /* phases 5-3 */
nuclear@26 233 tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
nuclear@26 234
nuclear@26 235 tmp0 = tmp10 + tmp13; /* phase 2 */
nuclear@26 236 tmp3 = tmp10 - tmp13;
nuclear@26 237 tmp1 = tmp11 + tmp12;
nuclear@26 238 tmp2 = tmp11 - tmp12;
nuclear@26 239
nuclear@26 240 /* Odd part */
nuclear@26 241
nuclear@26 242 tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
nuclear@26 243 tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
nuclear@26 244 tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
nuclear@26 245 tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
nuclear@26 246
nuclear@26 247 z13 = tmp6 + tmp5; /* phase 6 */
nuclear@26 248 z10 = tmp6 - tmp5;
nuclear@26 249 z11 = tmp4 + tmp7;
nuclear@26 250 z12 = tmp4 - tmp7;
nuclear@26 251
nuclear@26 252 tmp7 = z11 + z13; /* phase 5 */
nuclear@26 253 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
nuclear@26 254
nuclear@26 255 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
nuclear@26 256 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
nuclear@26 257 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
nuclear@26 258
nuclear@26 259 tmp6 = tmp12 - tmp7; /* phase 2 */
nuclear@26 260 tmp5 = tmp11 - tmp6;
nuclear@26 261 tmp4 = tmp10 + tmp5;
nuclear@26 262
nuclear@26 263 wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
nuclear@26 264 wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
nuclear@26 265 wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
nuclear@26 266 wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
nuclear@26 267 wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
nuclear@26 268 wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
nuclear@26 269 wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
nuclear@26 270 wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
nuclear@26 271
nuclear@26 272 inptr++; /* advance pointers to next column */
nuclear@26 273 quantptr++;
nuclear@26 274 wsptr++;
nuclear@26 275 }
nuclear@26 276
nuclear@26 277 /* Pass 2: process rows from work array, store into output array. */
nuclear@26 278 /* Note that we must descale the results by a factor of 8 == 2**3, */
nuclear@26 279 /* and also undo the PASS1_BITS scaling. */
nuclear@26 280
nuclear@26 281 wsptr = workspace;
nuclear@26 282 for (ctr = 0; ctr < DCTSIZE; ctr++) {
nuclear@26 283 outptr = output_buf[ctr] + output_col;
nuclear@26 284 /* Rows of zeroes can be exploited in the same way as we did with columns.
nuclear@26 285 * However, the column calculation has created many nonzero AC terms, so
nuclear@26 286 * the simplification applies less often (typically 5% to 10% of the time).
nuclear@26 287 * On machines with very fast multiplication, it's possible that the
nuclear@26 288 * test takes more time than it's worth. In that case this section
nuclear@26 289 * may be commented out.
nuclear@26 290 */
nuclear@26 291
nuclear@26 292 #ifndef NO_ZERO_ROW_TEST
nuclear@26 293 if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
nuclear@26 294 wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
nuclear@26 295 /* AC terms all zero */
nuclear@26 296 JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
nuclear@26 297 & RANGE_MASK];
nuclear@26 298
nuclear@26 299 outptr[0] = dcval;
nuclear@26 300 outptr[1] = dcval;
nuclear@26 301 outptr[2] = dcval;
nuclear@26 302 outptr[3] = dcval;
nuclear@26 303 outptr[4] = dcval;
nuclear@26 304 outptr[5] = dcval;
nuclear@26 305 outptr[6] = dcval;
nuclear@26 306 outptr[7] = dcval;
nuclear@26 307
nuclear@26 308 wsptr += DCTSIZE; /* advance pointer to next row */
nuclear@26 309 continue;
nuclear@26 310 }
nuclear@26 311 #endif
nuclear@26 312
nuclear@26 313 /* Even part */
nuclear@26 314
nuclear@26 315 tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
nuclear@26 316 tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
nuclear@26 317
nuclear@26 318 tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
nuclear@26 319 tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
nuclear@26 320 - tmp13;
nuclear@26 321
nuclear@26 322 tmp0 = tmp10 + tmp13;
nuclear@26 323 tmp3 = tmp10 - tmp13;
nuclear@26 324 tmp1 = tmp11 + tmp12;
nuclear@26 325 tmp2 = tmp11 - tmp12;
nuclear@26 326
nuclear@26 327 /* Odd part */
nuclear@26 328
nuclear@26 329 z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
nuclear@26 330 z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
nuclear@26 331 z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
nuclear@26 332 z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
nuclear@26 333
nuclear@26 334 tmp7 = z11 + z13; /* phase 5 */
nuclear@26 335 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
nuclear@26 336
nuclear@26 337 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
nuclear@26 338 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
nuclear@26 339 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
nuclear@26 340
nuclear@26 341 tmp6 = tmp12 - tmp7; /* phase 2 */
nuclear@26 342 tmp5 = tmp11 - tmp6;
nuclear@26 343 tmp4 = tmp10 + tmp5;
nuclear@26 344
nuclear@26 345 /* Final output stage: scale down by a factor of 8 and range-limit */
nuclear@26 346
nuclear@26 347 outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
nuclear@26 348 & RANGE_MASK];
nuclear@26 349 outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
nuclear@26 350 & RANGE_MASK];
nuclear@26 351 outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
nuclear@26 352 & RANGE_MASK];
nuclear@26 353 outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
nuclear@26 354 & RANGE_MASK];
nuclear@26 355 outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
nuclear@26 356 & RANGE_MASK];
nuclear@26 357 outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
nuclear@26 358 & RANGE_MASK];
nuclear@26 359 outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
nuclear@26 360 & RANGE_MASK];
nuclear@26 361 outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
nuclear@26 362 & RANGE_MASK];
nuclear@26 363
nuclear@26 364 wsptr += DCTSIZE; /* advance pointer to next row */
nuclear@26 365 }
nuclear@26 366 }
nuclear@26 367
nuclear@26 368 #endif /* DCT_IFAST_SUPPORTED */