1 #if !defined(_FX_JPEG_TURBO_) 2 /* 3 * jidctfst.c 4 * 5 * Copyright (C) 1994-1998, Thomas G. Lane. 6 * This file is part of the Independent JPEG Group's software. 7 * For conditions of distribution and use, see the accompanying README file. 8 * 9 * This file contains a fast, not so accurate integer implementation of the 10 * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine 11 * must also perform dequantization of the input coefficients. 12 * 13 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT 14 * on each row (or vice versa, but it's more convenient to emit a row at 15 * a time). Direct algorithms are also available, but they are much more 16 * complex and seem not to be any faster when reduced to code. 17 * 18 * This implementation is based on Arai, Agui, and Nakajima's algorithm for 19 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in 20 * Japanese, but the algorithm is described in the Pennebaker & Mitchell 21 * JPEG textbook (see REFERENCES section in file README). The following code 22 * is based directly on figure 4-8 in P&M. 23 * While an 8-point DCT cannot be done in less than 11 multiplies, it is 24 * possible to arrange the computation so that many of the multiplies are 25 * simple scalings of the final outputs. These multiplies can then be 26 * folded into the multiplications or divisions by the JPEG quantization 27 * table entries. The AA&N method leaves only 5 multiplies and 29 adds 28 * to be done in the DCT itself. 29 * The primary disadvantage of this method is that with fixed-point math, 30 * accuracy is lost due to imprecise representation of the scaled 31 * quantization values. The smaller the quantization table entry, the less 32 * precise the scaled value, so this implementation does worse with high- 33 * quality-setting files than with low-quality ones. 34 */ 35 36 #define JPEG_INTERNALS 37 #include "jinclude.h" 38 #include "jpeglib.h" 39 #include "jdct.h" /* Private declarations for DCT subsystem */ 40 41 #ifdef DCT_IFAST_SUPPORTED 42 43 44 /* 45 * This module is specialized to the case DCTSIZE = 8. 46 */ 47 48 #if DCTSIZE != 8 49 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ 50 #endif 51 52 53 /* Scaling decisions are generally the same as in the LL&M algorithm; 54 * see jidctint.c for more details. However, we choose to descale 55 * (right shift) multiplication products as soon as they are formed, 56 * rather than carrying additional fractional bits into subsequent additions. 57 * This compromises accuracy slightly, but it lets us save a few shifts. 58 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) 59 * everywhere except in the multiplications proper; this saves a good deal 60 * of work on 16-bit-int machines. 61 * 62 * The dequantized coefficients are not integers because the AA&N scaling 63 * factors have been incorporated. We represent them scaled up by PASS1_BITS, 64 * so that the first and second IDCT rounds have the same input scaling. 65 * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to 66 * avoid a descaling shift; this compromises accuracy rather drastically 67 * for small quantization table entries, but it saves a lot of shifts. 68 * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway, 69 * so we use a much larger scaling factor to preserve accuracy. 70 * 71 * A final compromise is to represent the multiplicative constants to only 72 * 8 fractional bits, rather than 13. This saves some shifting work on some 73 * machines, and may also reduce the cost of multiplication (since there 74 * are fewer one-bits in the constants). 75 */ 76 77 #if BITS_IN_JSAMPLE == 8 78 #define CONST_BITS 8 79 #define PASS1_BITS 2 80 #else 81 #define CONST_BITS 8 82 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */ 83 #endif 84 85 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus 86 * causing a lot of useless floating-point operations at run time. 87 * To get around this we use the following pre-calculated constants. 88 * If you change CONST_BITS you may want to add appropriate values. 89 * (With a reasonable C compiler, you can just rely on the FIX() macro...) 90 */ 91 92 #if CONST_BITS == 8 93 #define FIX_1_082392200 ((INT32) 277) /* FIX(1.082392200) */ 94 #define FIX_1_414213562 ((INT32) 362) /* FIX(1.414213562) */ 95 #define FIX_1_847759065 ((INT32) 473) /* FIX(1.847759065) */ 96 #define FIX_2_613125930 ((INT32) 669) /* FIX(2.613125930) */ 97 #else 98 #define FIX_1_082392200 FIX(1.082392200) 99 #define FIX_1_414213562 FIX(1.414213562) 100 #define FIX_1_847759065 FIX(1.847759065) 101 #define FIX_2_613125930 FIX(2.613125930) 102 #endif 103 104 105 /* We can gain a little more speed, with a further compromise in accuracy, 106 * by omitting the addition in a descaling shift. This yields an incorrectly 107 * rounded result half the time... 108 */ 109 110 #ifndef USE_ACCURATE_ROUNDING 111 #undef DESCALE 112 #define DESCALE(x,n) RIGHT_SHIFT(x, n) 113 #endif 114 115 116 /* Multiply a DCTELEM variable by an INT32 constant, and immediately 117 * descale to yield a DCTELEM result. 118 */ 119 120 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) 121 122 123 /* Dequantize a coefficient by multiplying it by the multiplier-table 124 * entry; produce a DCTELEM result. For 8-bit data a 16x16->16 125 * multiplication will do. For 12-bit data, the multiplier table is 126 * declared INT32, so a 32-bit multiply will be used. 127 */ 128 129 #if BITS_IN_JSAMPLE == 8 130 #define DEQUANTIZE(coef,quantval) (((IFAST_MULT_TYPE) (coef)) * (quantval)) 131 #else 132 #define DEQUANTIZE(coef,quantval) \ 133 DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS) 134 #endif 135 136 137 /* Like DESCALE, but applies to a DCTELEM and produces an int. 138 * We assume that int right shift is unsigned if INT32 right shift is. 139 */ 140 141 #ifdef RIGHT_SHIFT_IS_UNSIGNED 142 #define ISHIFT_TEMPS DCTELEM ishift_temp; 143 #if BITS_IN_JSAMPLE == 8 144 #define DCTELEMBITS 16 /* DCTELEM may be 16 or 32 bits */ 145 #else 146 #define DCTELEMBITS 32 /* DCTELEM must be 32 bits */ 147 #endif 148 #define IRIGHT_SHIFT(x,shft) \ 149 ((ishift_temp = (x)) < 0 ? \ 150 (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \ 151 (ishift_temp >> (shft))) 152 #else 153 #define ISHIFT_TEMPS 154 #define IRIGHT_SHIFT(x,shft) ((x) >> (shft)) 155 #endif 156 157 #ifdef USE_ACCURATE_ROUNDING 158 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n)) 159 #else 160 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT(x, n)) 161 #endif 162 163 164 /* 165 * Perform dequantization and inverse DCT on one block of coefficients. 166 */ 167 168 GLOBAL(void) 169 jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr, 170 JCOEFPTR coef_block, 171 JSAMPARRAY output_buf, JDIMENSION output_col) 172 { 173 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 174 DCTELEM tmp10, tmp11, tmp12, tmp13; 175 DCTELEM z5, z10, z11, z12, z13; 176 JCOEFPTR inptr; 177 IFAST_MULT_TYPE * quantptr; 178 int * wsptr; 179 JSAMPROW outptr; 180 JSAMPLE *range_limit = IDCT_range_limit(cinfo); 181 int ctr; 182 int workspace[DCTSIZE2]; /* buffers data between passes */ 183 SHIFT_TEMPS /* for DESCALE */ 184 ISHIFT_TEMPS /* for IDESCALE */ 185 186 /* Pass 1: process columns from input, store into work array. */ 187 188 inptr = coef_block; 189 quantptr = (IFAST_MULT_TYPE *) compptr->dct_table; 190 wsptr = workspace; 191 for (ctr = DCTSIZE; ctr > 0; ctr--) { 192 /* Due to quantization, we will usually find that many of the input 193 * coefficients are zero, especially the AC terms. We can exploit this 194 * by short-circuiting the IDCT calculation for any column in which all 195 * the AC terms are zero. In that case each output is equal to the 196 * DC coefficient (with scale factor as needed). 197 * With typical images and quantization tables, half or more of the 198 * column DCT calculations can be simplified this way. 199 */ 200 201 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 && 202 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 && 203 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 && 204 inptr[DCTSIZE*7] == 0) { 205 /* AC terms all zero */ 206 int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); 207 208 wsptr[DCTSIZE*0] = dcval; 209 wsptr[DCTSIZE*1] = dcval; 210 wsptr[DCTSIZE*2] = dcval; 211 wsptr[DCTSIZE*3] = dcval; 212 wsptr[DCTSIZE*4] = dcval; 213 wsptr[DCTSIZE*5] = dcval; 214 wsptr[DCTSIZE*6] = dcval; 215 wsptr[DCTSIZE*7] = dcval; 216 217 inptr++; /* advance pointers to next column */ 218 quantptr++; 219 wsptr++; 220 continue; 221 } 222 223 /* Even part */ 224 225 tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); 226 tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]); 227 tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]); 228 tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]); 229 230 tmp10 = tmp0 + tmp2; /* phase 3 */ 231 tmp11 = tmp0 - tmp2; 232 233 tmp13 = tmp1 + tmp3; /* phases 5-3 */ 234 tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */ 235 236 tmp0 = tmp10 + tmp13; /* phase 2 */ 237 tmp3 = tmp10 - tmp13; 238 tmp1 = tmp11 + tmp12; 239 tmp2 = tmp11 - tmp12; 240 241 /* Odd part */ 242 243 tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]); 244 tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]); 245 tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]); 246 tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]); 247 248 z13 = tmp6 + tmp5; /* phase 6 */ 249 z10 = tmp6 - tmp5; 250 z11 = tmp4 + tmp7; 251 z12 = tmp4 - tmp7; 252 253 tmp7 = z11 + z13; /* phase 5 */ 254 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */ 255 256 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */ 257 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */ 258 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */ 259 260 tmp6 = tmp12 - tmp7; /* phase 2 */ 261 tmp5 = tmp11 - tmp6; 262 tmp4 = tmp10 + tmp5; 263 264 wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7); 265 wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7); 266 wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6); 267 wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6); 268 wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5); 269 wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5); 270 wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4); 271 wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4); 272 273 inptr++; /* advance pointers to next column */ 274 quantptr++; 275 wsptr++; 276 } 277 278 /* Pass 2: process rows from work array, store into output array. */ 279 /* Note that we must descale the results by a factor of 8 == 2**3, */ 280 /* and also undo the PASS1_BITS scaling. */ 281 282 wsptr = workspace; 283 for (ctr = 0; ctr < DCTSIZE; ctr++) { 284 outptr = output_buf[ctr] + output_col; 285 /* Rows of zeroes can be exploited in the same way as we did with columns. 286 * However, the column calculation has created many nonzero AC terms, so 287 * the simplification applies less often (typically 5% to 10% of the time). 288 * On machines with very fast multiplication, it's possible that the 289 * test takes more time than it's worth. In that case this section 290 * may be commented out. 291 */ 292 293 #ifndef NO_ZERO_ROW_TEST 294 if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 && 295 wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) { 296 /* AC terms all zero */ 297 JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3) 298 & RANGE_MASK]; 299 300 outptr[0] = dcval; 301 outptr[1] = dcval; 302 outptr[2] = dcval; 303 outptr[3] = dcval; 304 outptr[4] = dcval; 305 outptr[5] = dcval; 306 outptr[6] = dcval; 307 outptr[7] = dcval; 308 309 wsptr += DCTSIZE; /* advance pointer to next row */ 310 continue; 311 } 312 #endif 313 314 /* Even part */ 315 316 tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]); 317 tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]); 318 319 tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]); 320 tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562) 321 - tmp13; 322 323 tmp0 = tmp10 + tmp13; 324 tmp3 = tmp10 - tmp13; 325 tmp1 = tmp11 + tmp12; 326 tmp2 = tmp11 - tmp12; 327 328 /* Odd part */ 329 330 z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3]; 331 z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3]; 332 z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7]; 333 z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7]; 334 335 tmp7 = z11 + z13; /* phase 5 */ 336 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */ 337 338 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */ 339 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */ 340 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */ 341 342 tmp6 = tmp12 - tmp7; /* phase 2 */ 343 tmp5 = tmp11 - tmp6; 344 tmp4 = tmp10 + tmp5; 345 346 /* Final output stage: scale down by a factor of 8 and range-limit */ 347 348 outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3) 349 & RANGE_MASK]; 350 outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3) 351 & RANGE_MASK]; 352 outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3) 353 & RANGE_MASK]; 354 outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3) 355 & RANGE_MASK]; 356 outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3) 357 & RANGE_MASK]; 358 outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3) 359 & RANGE_MASK]; 360 outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3) 361 & RANGE_MASK]; 362 outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3) 363 & RANGE_MASK]; 364 365 wsptr += DCTSIZE; /* advance pointer to next row */ 366 } 367 } 368 369 #endif /* DCT_IFAST_SUPPORTED */ 370 371 #endif //_FX_JPEG_TURBO_ 372