istereo

annotate libs/libjpeg/jidctint.c @ 26:862a3329a8f0

find changesets by author, revision, files, or words in the commit message
wohooo, added a shitload of code from zlib/libpng/libjpeg. When the good lord was raining shared libraries the iphone held a fucking umbrella...
author John Tsiombikas <nuclear@mutantstargoat.com>
date Thu, 08 Sep 2011 06:28:38 +0300
parents
children
rev   line source
nuclear@26 1 /*
nuclear@26 2 * jidctint.c
nuclear@26 3 *
nuclear@26 4 * Copyright (C) 1991-1998, 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 slow-but-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 an algorithm described in
nuclear@26 18 * C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
nuclear@26 19 * Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
nuclear@26 20 * Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
nuclear@26 21 * The primary algorithm described there uses 11 multiplies and 29 adds.
nuclear@26 22 * We use their alternate method with 12 multiplies and 32 adds.
nuclear@26 23 * The advantage of this method is that no data path contains more than one
nuclear@26 24 * multiplication; this allows a very simple and accurate implementation in
nuclear@26 25 * scaled fixed-point arithmetic, with a minimal number of shifts.
nuclear@26 26 */
nuclear@26 27
nuclear@26 28 #define JPEG_INTERNALS
nuclear@26 29 #include "jinclude.h"
nuclear@26 30 #include "jpeglib.h"
nuclear@26 31 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@26 32
nuclear@26 33 #ifdef DCT_ISLOW_SUPPORTED
nuclear@26 34
nuclear@26 35
nuclear@26 36 /*
nuclear@26 37 * This module is specialized to the case DCTSIZE = 8.
nuclear@26 38 */
nuclear@26 39
nuclear@26 40 #if DCTSIZE != 8
nuclear@26 41 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@26 42 #endif
nuclear@26 43
nuclear@26 44
nuclear@26 45 /*
nuclear@26 46 * The poop on this scaling stuff is as follows:
nuclear@26 47 *
nuclear@26 48 * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
nuclear@26 49 * larger than the true IDCT outputs. The final outputs are therefore
nuclear@26 50 * a factor of N larger than desired; since N=8 this can be cured by
nuclear@26 51 * a simple right shift at the end of the algorithm. The advantage of
nuclear@26 52 * this arrangement is that we save two multiplications per 1-D IDCT,
nuclear@26 53 * because the y0 and y4 inputs need not be divided by sqrt(N).
nuclear@26 54 *
nuclear@26 55 * We have to do addition and subtraction of the integer inputs, which
nuclear@26 56 * is no problem, and multiplication by fractional constants, which is
nuclear@26 57 * a problem to do in integer arithmetic. We multiply all the constants
nuclear@26 58 * by CONST_SCALE and convert them to integer constants (thus retaining
nuclear@26 59 * CONST_BITS bits of precision in the constants). After doing a
nuclear@26 60 * multiplication we have to divide the product by CONST_SCALE, with proper
nuclear@26 61 * rounding, to produce the correct output. This division can be done
nuclear@26 62 * cheaply as a right shift of CONST_BITS bits. We postpone shifting
nuclear@26 63 * as long as possible so that partial sums can be added together with
nuclear@26 64 * full fractional precision.
nuclear@26 65 *
nuclear@26 66 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
nuclear@26 67 * they are represented to better-than-integral precision. These outputs
nuclear@26 68 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
nuclear@26 69 * with the recommended scaling. (To scale up 12-bit sample data further, an
nuclear@26 70 * intermediate INT32 array would be needed.)
nuclear@26 71 *
nuclear@26 72 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
nuclear@26 73 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
nuclear@26 74 * shows that the values given below are the most effective.
nuclear@26 75 */
nuclear@26 76
nuclear@26 77 #if BITS_IN_JSAMPLE == 8
nuclear@26 78 #define CONST_BITS 13
nuclear@26 79 #define PASS1_BITS 2
nuclear@26 80 #else
nuclear@26 81 #define CONST_BITS 13
nuclear@26 82 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */
nuclear@26 83 #endif
nuclear@26 84
nuclear@26 85 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
nuclear@26 86 * causing a lot of useless floating-point operations at run time.
nuclear@26 87 * To get around this we use the following pre-calculated constants.
nuclear@26 88 * If you change CONST_BITS you may want to add appropriate values.
nuclear@26 89 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
nuclear@26 90 */
nuclear@26 91
nuclear@26 92 #if CONST_BITS == 13
nuclear@26 93 #define FIX_0_298631336 ((INT32) 2446) /* FIX(0.298631336) */
nuclear@26 94 #define FIX_0_390180644 ((INT32) 3196) /* FIX(0.390180644) */
nuclear@26 95 #define FIX_0_541196100 ((INT32) 4433) /* FIX(0.541196100) */
nuclear@26 96 #define FIX_0_765366865 ((INT32) 6270) /* FIX(0.765366865) */
nuclear@26 97 #define FIX_0_899976223 ((INT32) 7373) /* FIX(0.899976223) */
nuclear@26 98 #define FIX_1_175875602 ((INT32) 9633) /* FIX(1.175875602) */
nuclear@26 99 #define FIX_1_501321110 ((INT32) 12299) /* FIX(1.501321110) */
nuclear@26 100 #define FIX_1_847759065 ((INT32) 15137) /* FIX(1.847759065) */
nuclear@26 101 #define FIX_1_961570560 ((INT32) 16069) /* FIX(1.961570560) */
nuclear@26 102 #define FIX_2_053119869 ((INT32) 16819) /* FIX(2.053119869) */
nuclear@26 103 #define FIX_2_562915447 ((INT32) 20995) /* FIX(2.562915447) */
nuclear@26 104 #define FIX_3_072711026 ((INT32) 25172) /* FIX(3.072711026) */
nuclear@26 105 #else
nuclear@26 106 #define FIX_0_298631336 FIX(0.298631336)
nuclear@26 107 #define FIX_0_390180644 FIX(0.390180644)
nuclear@26 108 #define FIX_0_541196100 FIX(0.541196100)
nuclear@26 109 #define FIX_0_765366865 FIX(0.765366865)
nuclear@26 110 #define FIX_0_899976223 FIX(0.899976223)
nuclear@26 111 #define FIX_1_175875602 FIX(1.175875602)
nuclear@26 112 #define FIX_1_501321110 FIX(1.501321110)
nuclear@26 113 #define FIX_1_847759065 FIX(1.847759065)
nuclear@26 114 #define FIX_1_961570560 FIX(1.961570560)
nuclear@26 115 #define FIX_2_053119869 FIX(2.053119869)
nuclear@26 116 #define FIX_2_562915447 FIX(2.562915447)
nuclear@26 117 #define FIX_3_072711026 FIX(3.072711026)
nuclear@26 118 #endif
nuclear@26 119
nuclear@26 120
nuclear@26 121 /* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
nuclear@26 122 * For 8-bit samples with the recommended scaling, all the variable
nuclear@26 123 * and constant values involved are no more than 16 bits wide, so a
nuclear@26 124 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
nuclear@26 125 * For 12-bit samples, a full 32-bit multiplication will be needed.
nuclear@26 126 */
nuclear@26 127
nuclear@26 128 #if BITS_IN_JSAMPLE == 8
nuclear@26 129 #define MULTIPLY(var,const) MULTIPLY16C16(var,const)
nuclear@26 130 #else
nuclear@26 131 #define MULTIPLY(var,const) ((var) * (const))
nuclear@26 132 #endif
nuclear@26 133
nuclear@26 134
nuclear@26 135 /* Dequantize a coefficient by multiplying it by the multiplier-table
nuclear@26 136 * entry; produce an int result. In this module, both inputs and result
nuclear@26 137 * are 16 bits or less, so either int or short multiply will work.
nuclear@26 138 */
nuclear@26 139
nuclear@26 140 #define DEQUANTIZE(coef,quantval) (((ISLOW_MULT_TYPE) (coef)) * (quantval))
nuclear@26 141
nuclear@26 142
nuclear@26 143 /*
nuclear@26 144 * Perform dequantization and inverse DCT on one block of coefficients.
nuclear@26 145 */
nuclear@26 146
nuclear@26 147 GLOBAL(void)
nuclear@26 148 jpeg_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
nuclear@26 149 JCOEFPTR coef_block,
nuclear@26 150 JSAMPARRAY output_buf, JDIMENSION output_col)
nuclear@26 151 {
nuclear@26 152 INT32 tmp0, tmp1, tmp2, tmp3;
nuclear@26 153 INT32 tmp10, tmp11, tmp12, tmp13;
nuclear@26 154 INT32 z1, z2, z3, z4, z5;
nuclear@26 155 JCOEFPTR inptr;
nuclear@26 156 ISLOW_MULT_TYPE * quantptr;
nuclear@26 157 int * wsptr;
nuclear@26 158 JSAMPROW outptr;
nuclear@26 159 JSAMPLE *range_limit = IDCT_range_limit(cinfo);
nuclear@26 160 int ctr;
nuclear@26 161 int workspace[DCTSIZE2]; /* buffers data between passes */
nuclear@26 162 SHIFT_TEMPS
nuclear@26 163
nuclear@26 164 /* Pass 1: process columns from input, store into work array. */
nuclear@26 165 /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
nuclear@26 166 /* furthermore, we scale the results by 2**PASS1_BITS. */
nuclear@26 167
nuclear@26 168 inptr = coef_block;
nuclear@26 169 quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
nuclear@26 170 wsptr = workspace;
nuclear@26 171 for (ctr = DCTSIZE; ctr > 0; ctr--) {
nuclear@26 172 /* Due to quantization, we will usually find that many of the input
nuclear@26 173 * coefficients are zero, especially the AC terms. We can exploit this
nuclear@26 174 * by short-circuiting the IDCT calculation for any column in which all
nuclear@26 175 * the AC terms are zero. In that case each output is equal to the
nuclear@26 176 * DC coefficient (with scale factor as needed).
nuclear@26 177 * With typical images and quantization tables, half or more of the
nuclear@26 178 * column DCT calculations can be simplified this way.
nuclear@26 179 */
nuclear@26 180
nuclear@26 181 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
nuclear@26 182 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
nuclear@26 183 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
nuclear@26 184 inptr[DCTSIZE*7] == 0) {
nuclear@26 185 /* AC terms all zero */
nuclear@26 186 int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
nuclear@26 187
nuclear@26 188 wsptr[DCTSIZE*0] = dcval;
nuclear@26 189 wsptr[DCTSIZE*1] = dcval;
nuclear@26 190 wsptr[DCTSIZE*2] = dcval;
nuclear@26 191 wsptr[DCTSIZE*3] = dcval;
nuclear@26 192 wsptr[DCTSIZE*4] = dcval;
nuclear@26 193 wsptr[DCTSIZE*5] = dcval;
nuclear@26 194 wsptr[DCTSIZE*6] = dcval;
nuclear@26 195 wsptr[DCTSIZE*7] = dcval;
nuclear@26 196
nuclear@26 197 inptr++; /* advance pointers to next column */
nuclear@26 198 quantptr++;
nuclear@26 199 wsptr++;
nuclear@26 200 continue;
nuclear@26 201 }
nuclear@26 202
nuclear@26 203 /* Even part: reverse the even part of the forward DCT. */
nuclear@26 204 /* The rotator is sqrt(2)*c(-6). */
nuclear@26 205
nuclear@26 206 z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
nuclear@26 207 z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
nuclear@26 208
nuclear@26 209 z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
nuclear@26 210 tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
nuclear@26 211 tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
nuclear@26 212
nuclear@26 213 z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
nuclear@26 214 z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
nuclear@26 215
nuclear@26 216 tmp0 = (z2 + z3) << CONST_BITS;
nuclear@26 217 tmp1 = (z2 - z3) << CONST_BITS;
nuclear@26 218
nuclear@26 219 tmp10 = tmp0 + tmp3;
nuclear@26 220 tmp13 = tmp0 - tmp3;
nuclear@26 221 tmp11 = tmp1 + tmp2;
nuclear@26 222 tmp12 = tmp1 - tmp2;
nuclear@26 223
nuclear@26 224 /* Odd part per figure 8; the matrix is unitary and hence its
nuclear@26 225 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
nuclear@26 226 */
nuclear@26 227
nuclear@26 228 tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
nuclear@26 229 tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
nuclear@26 230 tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
nuclear@26 231 tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
nuclear@26 232
nuclear@26 233 z1 = tmp0 + tmp3;
nuclear@26 234 z2 = tmp1 + tmp2;
nuclear@26 235 z3 = tmp0 + tmp2;
nuclear@26 236 z4 = tmp1 + tmp3;
nuclear@26 237 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
nuclear@26 238
nuclear@26 239 tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
nuclear@26 240 tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
nuclear@26 241 tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
nuclear@26 242 tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
nuclear@26 243 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
nuclear@26 244 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
nuclear@26 245 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
nuclear@26 246 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
nuclear@26 247
nuclear@26 248 z3 += z5;
nuclear@26 249 z4 += z5;
nuclear@26 250
nuclear@26 251 tmp0 += z1 + z3;
nuclear@26 252 tmp1 += z2 + z4;
nuclear@26 253 tmp2 += z2 + z3;
nuclear@26 254 tmp3 += z1 + z4;
nuclear@26 255
nuclear@26 256 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
nuclear@26 257
nuclear@26 258 wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
nuclear@26 259 wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
nuclear@26 260 wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
nuclear@26 261 wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
nuclear@26 262 wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
nuclear@26 263 wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
nuclear@26 264 wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
nuclear@26 265 wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
nuclear@26 266
nuclear@26 267 inptr++; /* advance pointers to next column */
nuclear@26 268 quantptr++;
nuclear@26 269 wsptr++;
nuclear@26 270 }
nuclear@26 271
nuclear@26 272 /* Pass 2: process rows from work array, store into output array. */
nuclear@26 273 /* Note that we must descale the results by a factor of 8 == 2**3, */
nuclear@26 274 /* and also undo the PASS1_BITS scaling. */
nuclear@26 275
nuclear@26 276 wsptr = workspace;
nuclear@26 277 for (ctr = 0; ctr < DCTSIZE; ctr++) {
nuclear@26 278 outptr = output_buf[ctr] + output_col;
nuclear@26 279 /* Rows of zeroes can be exploited in the same way as we did with columns.
nuclear@26 280 * However, the column calculation has created many nonzero AC terms, so
nuclear@26 281 * the simplification applies less often (typically 5% to 10% of the time).
nuclear@26 282 * On machines with very fast multiplication, it's possible that the
nuclear@26 283 * test takes more time than it's worth. In that case this section
nuclear@26 284 * may be commented out.
nuclear@26 285 */
nuclear@26 286
nuclear@26 287 #ifndef NO_ZERO_ROW_TEST
nuclear@26 288 if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
nuclear@26 289 wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
nuclear@26 290 /* AC terms all zero */
nuclear@26 291 JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
nuclear@26 292 & RANGE_MASK];
nuclear@26 293
nuclear@26 294 outptr[0] = dcval;
nuclear@26 295 outptr[1] = dcval;
nuclear@26 296 outptr[2] = dcval;
nuclear@26 297 outptr[3] = dcval;
nuclear@26 298 outptr[4] = dcval;
nuclear@26 299 outptr[5] = dcval;
nuclear@26 300 outptr[6] = dcval;
nuclear@26 301 outptr[7] = dcval;
nuclear@26 302
nuclear@26 303 wsptr += DCTSIZE; /* advance pointer to next row */
nuclear@26 304 continue;
nuclear@26 305 }
nuclear@26 306 #endif
nuclear@26 307
nuclear@26 308 /* Even part: reverse the even part of the forward DCT. */
nuclear@26 309 /* The rotator is sqrt(2)*c(-6). */
nuclear@26 310
nuclear@26 311 z2 = (INT32) wsptr[2];
nuclear@26 312 z3 = (INT32) wsptr[6];
nuclear@26 313
nuclear@26 314 z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
nuclear@26 315 tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
nuclear@26 316 tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
nuclear@26 317
nuclear@26 318 tmp0 = ((INT32) wsptr[0] + (INT32) wsptr[4]) << CONST_BITS;
nuclear@26 319 tmp1 = ((INT32) wsptr[0] - (INT32) wsptr[4]) << CONST_BITS;
nuclear@26 320
nuclear@26 321 tmp10 = tmp0 + tmp3;
nuclear@26 322 tmp13 = tmp0 - tmp3;
nuclear@26 323 tmp11 = tmp1 + tmp2;
nuclear@26 324 tmp12 = tmp1 - tmp2;
nuclear@26 325
nuclear@26 326 /* Odd part per figure 8; the matrix is unitary and hence its
nuclear@26 327 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
nuclear@26 328 */
nuclear@26 329
nuclear@26 330 tmp0 = (INT32) wsptr[7];
nuclear@26 331 tmp1 = (INT32) wsptr[5];
nuclear@26 332 tmp2 = (INT32) wsptr[3];
nuclear@26 333 tmp3 = (INT32) wsptr[1];
nuclear@26 334
nuclear@26 335 z1 = tmp0 + tmp3;
nuclear@26 336 z2 = tmp1 + tmp2;
nuclear@26 337 z3 = tmp0 + tmp2;
nuclear@26 338 z4 = tmp1 + tmp3;
nuclear@26 339 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
nuclear@26 340
nuclear@26 341 tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
nuclear@26 342 tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
nuclear@26 343 tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
nuclear@26 344 tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
nuclear@26 345 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
nuclear@26 346 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
nuclear@26 347 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
nuclear@26 348 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
nuclear@26 349
nuclear@26 350 z3 += z5;
nuclear@26 351 z4 += z5;
nuclear@26 352
nuclear@26 353 tmp0 += z1 + z3;
nuclear@26 354 tmp1 += z2 + z4;
nuclear@26 355 tmp2 += z2 + z3;
nuclear@26 356 tmp3 += z1 + z4;
nuclear@26 357
nuclear@26 358 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
nuclear@26 359
nuclear@26 360 outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp3,
nuclear@26 361 CONST_BITS+PASS1_BITS+3)
nuclear@26 362 & RANGE_MASK];
nuclear@26 363 outptr[7] = range_limit[(int) DESCALE(tmp10 - tmp3,
nuclear@26 364 CONST_BITS+PASS1_BITS+3)
nuclear@26 365 & RANGE_MASK];
nuclear@26 366 outptr[1] = range_limit[(int) DESCALE(tmp11 + tmp2,
nuclear@26 367 CONST_BITS+PASS1_BITS+3)
nuclear@26 368 & RANGE_MASK];
nuclear@26 369 outptr[6] = range_limit[(int) DESCALE(tmp11 - tmp2,
nuclear@26 370 CONST_BITS+PASS1_BITS+3)
nuclear@26 371 & RANGE_MASK];
nuclear@26 372 outptr[2] = range_limit[(int) DESCALE(tmp12 + tmp1,
nuclear@26 373 CONST_BITS+PASS1_BITS+3)
nuclear@26 374 & RANGE_MASK];
nuclear@26 375 outptr[5] = range_limit[(int) DESCALE(tmp12 - tmp1,
nuclear@26 376 CONST_BITS+PASS1_BITS+3)
nuclear@26 377 & RANGE_MASK];
nuclear@26 378 outptr[3] = range_limit[(int) DESCALE(tmp13 + tmp0,
nuclear@26 379 CONST_BITS+PASS1_BITS+3)
nuclear@26 380 & RANGE_MASK];
nuclear@26 381 outptr[4] = range_limit[(int) DESCALE(tmp13 - tmp0,
nuclear@26 382 CONST_BITS+PASS1_BITS+3)
nuclear@26 383 & RANGE_MASK];
nuclear@26 384
nuclear@26 385 wsptr += DCTSIZE; /* advance pointer to next row */
nuclear@26 386 }
nuclear@26 387 }
nuclear@26 388
nuclear@26 389 #endif /* DCT_ISLOW_SUPPORTED */