FFmpeg: libavcodec/jfdctint.c Source File

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 SHIFT_TEMPS
00070 #define DCTSIZE 8
00071 #define BITS_IN_JSAMPLE 8
00072 #define GLOBAL(x) x
00073 #define RIGHT_SHIFT(x, n) ((x) >> (n))
00074 #define MULTIPLY16C16(var,const) ((var)*(const))
00075 
00076 #if 1 //def USE_ACCURATE_ROUNDING
00077 #define DESCALE(x,n)  RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
00078 #else
00079 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
00080 #endif
00081 
00082 
00083 /*
00084  * This module is specialized to the case DCTSIZE = 8.
00085  */
00086 
00087 #if DCTSIZE != 8
00088   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00089 #endif
00090 
00091 
00092 /*
00093  * The poop on this scaling stuff is as follows:
00094  *
00095  * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
00096  * larger than the true DCT outputs.  The final outputs are therefore
00097  * a factor of N larger than desired; since N=8 this can be cured by
00098  * a simple right shift at the end of the algorithm.  The advantage of
00099  * this arrangement is that we save two multiplications per 1-D DCT,
00100  * because the y0 and y4 outputs need not be divided by sqrt(N).
00101  * In the IJG code, this factor of 8 is removed by the quantization step
00102  * (in jcdctmgr.c), NOT in this module.
00103  *
00104  * We have to do addition and subtraction of the integer inputs, which
00105  * is no problem, and multiplication by fractional constants, which is
00106  * a problem to do in integer arithmetic.  We multiply all the constants
00107  * by CONST_SCALE and convert them to integer constants (thus retaining
00108  * CONST_BITS bits of precision in the constants).  After doing a
00109  * multiplication we have to divide the product by CONST_SCALE, with proper
00110  * rounding, to produce the correct output.  This division can be done
00111  * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
00112  * as long as possible so that partial sums can be added together with
00113  * full fractional precision.
00114  *
00115  * The outputs of the first pass are scaled up by PASS1_BITS bits so that
00116  * they are represented to better-than-integral precision.  These outputs
00117  * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
00118  * with the recommended scaling.  (For 12-bit sample data, the intermediate
00119  * array is int32_t anyway.)
00120  *
00121  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
00122  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
00123  * shows that the values given below are the most effective.
00124  */
00125 
00126 #if BITS_IN_JSAMPLE == 8
00127 #define CONST_BITS  13
00128 #define PASS1_BITS  4   /* set this to 2 if 16x16 multiplies are faster */
00129 #else
00130 #define CONST_BITS  13
00131 #define PASS1_BITS  1   /* lose a little precision to avoid overflow */
00132 #endif
00133 
00134 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00135  * causing a lot of useless floating-point operations at run time.
00136  * To get around this we use the following pre-calculated constants.
00137  * If you change CONST_BITS you may want to add appropriate values.
00138  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00139  */
00140 
00141 #if CONST_BITS == 13
00142 #define FIX_0_298631336  ((int32_t)  2446)      /* FIX(0.298631336) */
00143 #define FIX_0_390180644  ((int32_t)  3196)      /* FIX(0.390180644) */
00144 #define FIX_0_541196100  ((int32_t)  4433)      /* FIX(0.541196100) */
00145 #define FIX_0_765366865  ((int32_t)  6270)      /* FIX(0.765366865) */
00146 #define FIX_0_899976223  ((int32_t)  7373)      /* FIX(0.899976223) */
00147 #define FIX_1_175875602  ((int32_t)  9633)      /* FIX(1.175875602) */
00148 #define FIX_1_501321110  ((int32_t)  12299)     /* FIX(1.501321110) */
00149 #define FIX_1_847759065  ((int32_t)  15137)     /* FIX(1.847759065) */
00150 #define FIX_1_961570560  ((int32_t)  16069)     /* FIX(1.961570560) */
00151 #define FIX_2_053119869  ((int32_t)  16819)     /* FIX(2.053119869) */
00152 #define FIX_2_562915447  ((int32_t)  20995)     /* FIX(2.562915447) */
00153 #define FIX_3_072711026  ((int32_t)  25172)     /* FIX(3.072711026) */
00154 #else
00155 #define FIX_0_298631336  FIX(0.298631336)
00156 #define FIX_0_390180644  FIX(0.390180644)
00157 #define FIX_0_541196100  FIX(0.541196100)
00158 #define FIX_0_765366865  FIX(0.765366865)
00159 #define FIX_0_899976223  FIX(0.899976223)
00160 #define FIX_1_175875602  FIX(1.175875602)
00161 #define FIX_1_501321110  FIX(1.501321110)
00162 #define FIX_1_847759065  FIX(1.847759065)
00163 #define FIX_1_961570560  FIX(1.961570560)
00164 #define FIX_2_053119869  FIX(2.053119869)
00165 #define FIX_2_562915447  FIX(2.562915447)
00166 #define FIX_3_072711026  FIX(3.072711026)
00167 #endif
00168 
00169 
00170 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
00171  * For 8-bit samples with the recommended scaling, all the variable
00172  * and constant values involved are no more than 16 bits wide, so a
00173  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
00174  * For 12-bit samples, a full 32-bit multiplication will be needed.
00175  */
00176 
00177 #if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
00178 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
00179 #else
00180 #define MULTIPLY(var,const)  ((var) * (const))
00181 #endif
00182 
00183 
00184 static av_always_inline void row_fdct(DCTELEM * data){
00185   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00186   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00187   int_fast32_t z1, z2, z3, z4, z5;
00188   DCTELEM *dataptr;
00189   int ctr;
00190   SHIFT_TEMPS
00191 
00192   /* Pass 1: process rows. */
00193   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00194   /* furthermore, we scale the results by 2**PASS1_BITS. */
00195 
00196   dataptr = data;
00197   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00198     tmp0 = dataptr[0] + dataptr[7];
00199     tmp7 = dataptr[0] - dataptr[7];
00200     tmp1 = dataptr[1] + dataptr[6];
00201     tmp6 = dataptr[1] - dataptr[6];
00202     tmp2 = dataptr[2] + dataptr[5];
00203     tmp5 = dataptr[2] - dataptr[5];
00204     tmp3 = dataptr[3] + dataptr[4];
00205     tmp4 = dataptr[3] - dataptr[4];
00206 
00207     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00208      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00209      */
00210 
00211     tmp10 = tmp0 + tmp3;
00212     tmp13 = tmp0 - tmp3;
00213     tmp11 = tmp1 + tmp2;
00214     tmp12 = tmp1 - tmp2;
00215 
00216     dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
00217     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
00218 
00219     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00220     dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00221                                    CONST_BITS-PASS1_BITS);
00222     dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00223                                    CONST_BITS-PASS1_BITS);
00224 
00225     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00226      * cK represents cos(K*pi/16).
00227      * i0..i3 in the paper are tmp4..tmp7 here.
00228      */
00229 
00230     z1 = tmp4 + tmp7;
00231     z2 = tmp5 + tmp6;
00232     z3 = tmp4 + tmp6;
00233     z4 = tmp5 + tmp7;
00234     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00235 
00236     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00237     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00238     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00239     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00240     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00241     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00242     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00243     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00244 
00245     z3 += z5;
00246     z4 += z5;
00247 
00248     dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
00249     dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
00250     dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
00251     dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
00252 
00253     dataptr += DCTSIZE;         /* advance pointer to next row */
00254   }
00255 }
00256 
00257 /*
00258  * Perform the forward DCT on one block of samples.
00259  */
00260 
00261 GLOBAL(void)
00262 ff_jpeg_fdct_islow (DCTELEM * data)
00263 {
00264   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00265   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00266   int_fast32_t z1, z2, z3, z4, z5;
00267   DCTELEM *dataptr;
00268   int ctr;
00269   SHIFT_TEMPS
00270 
00271   row_fdct(data);
00272 
00273   /* Pass 2: process columns.
00274    * We remove the PASS1_BITS scaling, but leave the results scaled up
00275    * by an overall factor of 8.
00276    */
00277 
00278   dataptr = data;
00279   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00280     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00281     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00282     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00283     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00284     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00285     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00286     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00287     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00288 
00289     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00290      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00291      */
00292 
00293     tmp10 = tmp0 + tmp3;
00294     tmp13 = tmp0 - tmp3;
00295     tmp11 = tmp1 + tmp2;
00296     tmp12 = tmp1 - tmp2;
00297 
00298     dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00299     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00300 
00301     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00302     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00303                                            CONST_BITS+PASS1_BITS);
00304     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00305                                            CONST_BITS+PASS1_BITS);
00306 
00307     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00308      * cK represents cos(K*pi/16).
00309      * i0..i3 in the paper are tmp4..tmp7 here.
00310      */
00311 
00312     z1 = tmp4 + tmp7;
00313     z2 = tmp5 + tmp6;
00314     z3 = tmp4 + tmp6;
00315     z4 = tmp5 + tmp7;
00316     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00317 
00318     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00319     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00320     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00321     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00322     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00323     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00324     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00325     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00326 
00327     z3 += z5;
00328     z4 += z5;
00329 
00330     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
00331                                            CONST_BITS+PASS1_BITS);
00332     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
00333                                            CONST_BITS+PASS1_BITS);
00334     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
00335                                            CONST_BITS+PASS1_BITS);
00336     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
00337                                            CONST_BITS+PASS1_BITS);
00338 
00339     dataptr++;                  /* advance pointer to next column */
00340   }
00341 }
00342 
00343 /*
00344  * The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
00345  * on the rows and then, instead of doing even and odd, part on the colums
00346  * you do even part two times.
00347  */
00348 GLOBAL(void)
00349 ff_fdct248_islow (DCTELEM * data)
00350 {
00351   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00352   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00353   int_fast32_t z1;
00354   DCTELEM *dataptr;
00355   int ctr;
00356   SHIFT_TEMPS
00357 
00358   row_fdct(data);
00359 
00360   /* Pass 2: process columns.
00361    * We remove the PASS1_BITS scaling, but leave the results scaled up
00362    * by an overall factor of 8.
00363    */
00364 
00365   dataptr = data;
00366   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00367      tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
00368      tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00369      tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
00370      tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
00371      tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
00372      tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00373      tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
00374      tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
00375 
00376      tmp10 = tmp0 + tmp3;
00377      tmp11 = tmp1 + tmp2;
00378      tmp12 = tmp1 - tmp2;
00379      tmp13 = tmp0 - tmp3;
00380 
00381      dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00382      dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00383 
00384      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00385      dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00386                                             CONST_BITS+PASS1_BITS);
00387      dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00388                                             CONST_BITS+PASS1_BITS);
00389 
00390      tmp10 = tmp4 + tmp7;
00391      tmp11 = tmp5 + tmp6;
00392      tmp12 = tmp5 - tmp6;
00393      tmp13 = tmp4 - tmp7;
00394 
00395      dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00396      dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00397 
00398      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00399      dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00400                                             CONST_BITS+PASS1_BITS);
00401      dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00402                                             CONST_BITS+PASS1_BITS);
00403 
00404      dataptr++;                 /* advance pointer to next column */
00405   }
00406 }