Libav 0.7.1
|
00001 /* 00002 * jfdctint.c 00003 * 00004 * This file is part of the Independent JPEG Group's software. 00005 * 00006 * The authors make NO WARRANTY or representation, either express or implied, 00007 * with respect to this software, its quality, accuracy, merchantability, or 00008 * fitness for a particular purpose. This software is provided "AS IS", and 00009 * you, its user, assume the entire risk as to its quality and accuracy. 00010 * 00011 * This software is copyright (C) 1991-1996, Thomas G. Lane. 00012 * All Rights Reserved except as specified below. 00013 * 00014 * Permission is hereby granted to use, copy, modify, and distribute this 00015 * software (or portions thereof) for any purpose, without fee, subject to 00016 * these conditions: 00017 * (1) If any part of the source code for this software is distributed, then 00018 * this README file must be included, with this copyright and no-warranty 00019 * notice unaltered; and any additions, deletions, or changes to the original 00020 * files must be clearly indicated in accompanying documentation. 00021 * (2) If only executable code is distributed, then the accompanying 00022 * documentation must state that "this software is based in part on the work 00023 * of the Independent JPEG Group". 00024 * (3) Permission for use of this software is granted only if the user accepts 00025 * full responsibility for any undesirable consequences; the authors accept 00026 * NO LIABILITY for damages of any kind. 00027 * 00028 * These conditions apply to any software derived from or based on the IJG 00029 * code, not just to the unmodified library. If you use our work, you ought 00030 * to acknowledge us. 00031 * 00032 * Permission is NOT granted for the use of any IJG author's name or company 00033 * name in advertising or publicity relating to this software or products 00034 * derived from it. This software may be referred to only as "the Independent 00035 * JPEG Group's software". 00036 * 00037 * We specifically permit and encourage the use of this software as the basis 00038 * of commercial products, provided that all warranty or liability claims are 00039 * assumed by the product vendor. 00040 * 00041 * This file contains a slow-but-accurate integer implementation of the 00042 * forward DCT (Discrete Cosine Transform). 00043 * 00044 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT 00045 * on each column. Direct algorithms are also available, but they are 00046 * much more complex and seem not to be any faster when reduced to code. 00047 * 00048 * This implementation is based on an algorithm described in 00049 * C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT 00050 * Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics, 00051 * Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991. 00052 * The primary algorithm described there uses 11 multiplies and 29 adds. 00053 * We use their alternate method with 12 multiplies and 32 adds. 00054 * The advantage of this method is that no data path contains more than one 00055 * multiplication; this allows a very simple and accurate implementation in 00056 * scaled fixed-point arithmetic, with a minimal number of shifts. 00057 */ 00058 00064 #include <stdlib.h> 00065 #include <stdio.h> 00066 #include "libavutil/common.h" 00067 #include "dsputil.h" 00068 00069 #define DCTSIZE 8 00070 #define BITS_IN_JSAMPLE 8 00071 #define GLOBAL(x) x 00072 #define RIGHT_SHIFT(x, n) ((x) >> (n)) 00073 #define MULTIPLY16C16(var,const) ((var)*(const)) 00074 00075 #if 1 //def USE_ACCURATE_ROUNDING 00076 #define DESCALE(x,n) RIGHT_SHIFT((x) + (1 << ((n) - 1)), n) 00077 #else 00078 #define DESCALE(x,n) RIGHT_SHIFT(x, n) 00079 #endif 00080 00081 00082 /* 00083 * This module is specialized to the case DCTSIZE = 8. 00084 */ 00085 00086 #if DCTSIZE != 8 00087 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ 00088 #endif 00089 00090 00091 /* 00092 * The poop on this scaling stuff is as follows: 00093 * 00094 * Each 1-D DCT step produces outputs which are a factor of sqrt(N) 00095 * larger than the true DCT outputs. The final outputs are therefore 00096 * a factor of N larger than desired; since N=8 this can be cured by 00097 * a simple right shift at the end of the algorithm. The advantage of 00098 * this arrangement is that we save two multiplications per 1-D DCT, 00099 * because the y0 and y4 outputs need not be divided by sqrt(N). 00100 * In the IJG code, this factor of 8 is removed by the quantization step 00101 * (in jcdctmgr.c), NOT in this module. 00102 * 00103 * We have to do addition and subtraction of the integer inputs, which 00104 * is no problem, and multiplication by fractional constants, which is 00105 * a problem to do in integer arithmetic. We multiply all the constants 00106 * by CONST_SCALE and convert them to integer constants (thus retaining 00107 * CONST_BITS bits of precision in the constants). After doing a 00108 * multiplication we have to divide the product by CONST_SCALE, with proper 00109 * rounding, to produce the correct output. This division can be done 00110 * cheaply as a right shift of CONST_BITS bits. We postpone shifting 00111 * as long as possible so that partial sums can be added together with 00112 * full fractional precision. 00113 * 00114 * The outputs of the first pass are scaled up by PASS1_BITS bits so that 00115 * they are represented to better-than-integral precision. These outputs 00116 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word 00117 * with the recommended scaling. (For 12-bit sample data, the intermediate 00118 * array is int32_t anyway.) 00119 * 00120 * To avoid overflow of the 32-bit intermediate results in pass 2, we must 00121 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis 00122 * shows that the values given below are the most effective. 00123 */ 00124 00125 #if BITS_IN_JSAMPLE == 8 00126 #define CONST_BITS 13 00127 #define PASS1_BITS 4 /* set this to 2 if 16x16 multiplies are faster */ 00128 #else 00129 #define CONST_BITS 13 00130 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */ 00131 #endif 00132 00133 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus 00134 * causing a lot of useless floating-point operations at run time. 00135 * To get around this we use the following pre-calculated constants. 00136 * If you change CONST_BITS you may want to add appropriate values. 00137 * (With a reasonable C compiler, you can just rely on the FIX() macro...) 00138 */ 00139 00140 #if CONST_BITS == 13 00141 #define FIX_0_298631336 ((int32_t) 2446) /* FIX(0.298631336) */ 00142 #define FIX_0_390180644 ((int32_t) 3196) /* FIX(0.390180644) */ 00143 #define FIX_0_541196100 ((int32_t) 4433) /* FIX(0.541196100) */ 00144 #define FIX_0_765366865 ((int32_t) 6270) /* FIX(0.765366865) */ 00145 #define FIX_0_899976223 ((int32_t) 7373) /* FIX(0.899976223) */ 00146 #define FIX_1_175875602 ((int32_t) 9633) /* FIX(1.175875602) */ 00147 #define FIX_1_501321110 ((int32_t) 12299) /* FIX(1.501321110) */ 00148 #define FIX_1_847759065 ((int32_t) 15137) /* FIX(1.847759065) */ 00149 #define FIX_1_961570560 ((int32_t) 16069) /* FIX(1.961570560) */ 00150 #define FIX_2_053119869 ((int32_t) 16819) /* FIX(2.053119869) */ 00151 #define FIX_2_562915447 ((int32_t) 20995) /* FIX(2.562915447) */ 00152 #define FIX_3_072711026 ((int32_t) 25172) /* FIX(3.072711026) */ 00153 #else 00154 #define FIX_0_298631336 FIX(0.298631336) 00155 #define FIX_0_390180644 FIX(0.390180644) 00156 #define FIX_0_541196100 FIX(0.541196100) 00157 #define FIX_0_765366865 FIX(0.765366865) 00158 #define FIX_0_899976223 FIX(0.899976223) 00159 #define FIX_1_175875602 FIX(1.175875602) 00160 #define FIX_1_501321110 FIX(1.501321110) 00161 #define FIX_1_847759065 FIX(1.847759065) 00162 #define FIX_1_961570560 FIX(1.961570560) 00163 #define FIX_2_053119869 FIX(2.053119869) 00164 #define FIX_2_562915447 FIX(2.562915447) 00165 #define FIX_3_072711026 FIX(3.072711026) 00166 #endif 00167 00168 00169 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result. 00170 * For 8-bit samples with the recommended scaling, all the variable 00171 * and constant values involved are no more than 16 bits wide, so a 00172 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply. 00173 * For 12-bit samples, a full 32-bit multiplication will be needed. 00174 */ 00175 00176 #if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2 00177 #define MULTIPLY(var,const) MULTIPLY16C16(var,const) 00178 #else 00179 #define MULTIPLY(var,const) ((var) * (const)) 00180 #endif 00181 00182 00183 static av_always_inline void row_fdct(DCTELEM * data){ 00184 int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 00185 int tmp10, tmp11, tmp12, tmp13; 00186 int z1, z2, z3, z4, z5; 00187 DCTELEM *dataptr; 00188 int ctr; 00189 00190 /* Pass 1: process rows. */ 00191 /* Note results are scaled up by sqrt(8) compared to a true DCT; */ 00192 /* furthermore, we scale the results by 2**PASS1_BITS. */ 00193 00194 dataptr = data; 00195 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { 00196 tmp0 = dataptr[0] + dataptr[7]; 00197 tmp7 = dataptr[0] - dataptr[7]; 00198 tmp1 = dataptr[1] + dataptr[6]; 00199 tmp6 = dataptr[1] - dataptr[6]; 00200 tmp2 = dataptr[2] + dataptr[5]; 00201 tmp5 = dataptr[2] - dataptr[5]; 00202 tmp3 = dataptr[3] + dataptr[4]; 00203 tmp4 = dataptr[3] - dataptr[4]; 00204 00205 /* Even part per LL&M figure 1 --- note that published figure is faulty; 00206 * rotator "sqrt(2)*c1" should be "sqrt(2)*c6". 00207 */ 00208 00209 tmp10 = tmp0 + tmp3; 00210 tmp13 = tmp0 - tmp3; 00211 tmp11 = tmp1 + tmp2; 00212 tmp12 = tmp1 - tmp2; 00213 00214 dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); 00215 dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); 00216 00217 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00218 dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00219 CONST_BITS-PASS1_BITS); 00220 dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00221 CONST_BITS-PASS1_BITS); 00222 00223 /* Odd part per figure 8 --- note paper omits factor of sqrt(2). 00224 * cK represents cos(K*pi/16). 00225 * i0..i3 in the paper are tmp4..tmp7 here. 00226 */ 00227 00228 z1 = tmp4 + tmp7; 00229 z2 = tmp5 + tmp6; 00230 z3 = tmp4 + tmp6; 00231 z4 = tmp5 + tmp7; 00232 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ 00233 00234 tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ 00235 tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ 00236 tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ 00237 tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ 00238 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ 00239 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ 00240 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ 00241 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ 00242 00243 z3 += z5; 00244 z4 += z5; 00245 00246 dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); 00247 dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); 00248 dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); 00249 dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); 00250 00251 dataptr += DCTSIZE; /* advance pointer to next row */ 00252 } 00253 } 00254 00255 /* 00256 * Perform the forward DCT on one block of samples. 00257 */ 00258 00259 GLOBAL(void) 00260 ff_jpeg_fdct_islow (DCTELEM * data) 00261 { 00262 int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 00263 int tmp10, tmp11, tmp12, tmp13; 00264 int z1, z2, z3, z4, z5; 00265 DCTELEM *dataptr; 00266 int ctr; 00267 00268 row_fdct(data); 00269 00270 /* Pass 2: process columns. 00271 * We remove the PASS1_BITS scaling, but leave the results scaled up 00272 * by an overall factor of 8. 00273 */ 00274 00275 dataptr = data; 00276 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { 00277 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; 00278 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; 00279 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; 00280 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; 00281 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; 00282 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; 00283 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; 00284 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; 00285 00286 /* Even part per LL&M figure 1 --- note that published figure is faulty; 00287 * rotator "sqrt(2)*c1" should be "sqrt(2)*c6". 00288 */ 00289 00290 tmp10 = tmp0 + tmp3; 00291 tmp13 = tmp0 - tmp3; 00292 tmp11 = tmp1 + tmp2; 00293 tmp12 = tmp1 - tmp2; 00294 00295 dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS); 00296 dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS); 00297 00298 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00299 dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00300 CONST_BITS+PASS1_BITS); 00301 dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00302 CONST_BITS+PASS1_BITS); 00303 00304 /* Odd part per figure 8 --- note paper omits factor of sqrt(2). 00305 * cK represents cos(K*pi/16). 00306 * i0..i3 in the paper are tmp4..tmp7 here. 00307 */ 00308 00309 z1 = tmp4 + tmp7; 00310 z2 = tmp5 + tmp6; 00311 z3 = tmp4 + tmp6; 00312 z4 = tmp5 + tmp7; 00313 z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ 00314 00315 tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ 00316 tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ 00317 tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ 00318 tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ 00319 z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ 00320 z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ 00321 z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ 00322 z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ 00323 00324 z3 += z5; 00325 z4 += z5; 00326 00327 dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, 00328 CONST_BITS+PASS1_BITS); 00329 dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, 00330 CONST_BITS+PASS1_BITS); 00331 dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, 00332 CONST_BITS+PASS1_BITS); 00333 dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, 00334 CONST_BITS+PASS1_BITS); 00335 00336 dataptr++; /* advance pointer to next column */ 00337 } 00338 } 00339 00340 /* 00341 * The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT 00342 * on the rows and then, instead of doing even and odd, part on the colums 00343 * you do even part two times. 00344 */ 00345 GLOBAL(void) 00346 ff_fdct248_islow (DCTELEM * data) 00347 { 00348 int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 00349 int tmp10, tmp11, tmp12, tmp13; 00350 int z1; 00351 DCTELEM *dataptr; 00352 int ctr; 00353 00354 row_fdct(data); 00355 00356 /* Pass 2: process columns. 00357 * We remove the PASS1_BITS scaling, but leave the results scaled up 00358 * by an overall factor of 8. 00359 */ 00360 00361 dataptr = data; 00362 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { 00363 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1]; 00364 tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3]; 00365 tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5]; 00366 tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7]; 00367 tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1]; 00368 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3]; 00369 tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5]; 00370 tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7]; 00371 00372 tmp10 = tmp0 + tmp3; 00373 tmp11 = tmp1 + tmp2; 00374 tmp12 = tmp1 - tmp2; 00375 tmp13 = tmp0 - tmp3; 00376 00377 dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS); 00378 dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS); 00379 00380 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00381 dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00382 CONST_BITS+PASS1_BITS); 00383 dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00384 CONST_BITS+PASS1_BITS); 00385 00386 tmp10 = tmp4 + tmp7; 00387 tmp11 = tmp5 + tmp6; 00388 tmp12 = tmp5 - tmp6; 00389 tmp13 = tmp4 - tmp7; 00390 00391 dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS); 00392 dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS); 00393 00394 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00395 dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00396 CONST_BITS+PASS1_BITS); 00397 dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00398 CONST_BITS+PASS1_BITS); 00399 00400 dataptr++; /* advance pointer to next column */ 00401 } 00402 }