FFmpeg: libavcodec/jfdctint_template.c Source File

00001 /*
00002  * This file is part of the Independent JPEG Group's software.
00003  *
00004  * The authors make NO WARRANTY or representation, either express or implied,
00005  * with respect to this software, its quality, accuracy, merchantability, or
00006  * fitness for a particular purpose.  This software is provided "AS IS", and
00007  * you, its user, assume the entire risk as to its quality and accuracy.
00008  *
00009  * This software is copyright (C) 1991-1996, Thomas G. Lane.
00010  * All Rights Reserved except as specified below.
00011  *
00012  * Permission is hereby granted to use, copy, modify, and distribute this
00013  * software (or portions thereof) for any purpose, without fee, subject to
00014  * these conditions:
00015  * (1) If any part of the source code for this software is distributed, then
00016  * this README file must be included, with this copyright and no-warranty
00017  * notice unaltered; and any additions, deletions, or changes to the original
00018  * files must be clearly indicated in accompanying documentation.
00019  * (2) If only executable code is distributed, then the accompanying
00020  * documentation must state that "this software is based in part on the work
00021  * of the Independent JPEG Group".
00022  * (3) Permission for use of this software is granted only if the user accepts
00023  * full responsibility for any undesirable consequences; the authors accept
00024  * NO LIABILITY for damages of any kind.
00025  *
00026  * These conditions apply to any software derived from or based on the IJG
00027  * code, not just to the unmodified library.  If you use our work, you ought
00028  * to acknowledge us.
00029  *
00030  * Permission is NOT granted for the use of any IJG author's name or company
00031  * name in advertising or publicity relating to this software or products
00032  * derived from it.  This software may be referred to only as "the Independent
00033  * JPEG Group's software".
00034  *
00035  * We specifically permit and encourage the use of this software as the basis
00036  * of commercial products, provided that all warranty or liability claims are
00037  * assumed by the product vendor.
00038  *
00039  * This file contains a slow-but-accurate integer implementation of the
00040  * forward DCT (Discrete Cosine Transform).
00041  *
00042  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
00043  * on each column.  Direct algorithms are also available, but they are
00044  * much more complex and seem not to be any faster when reduced to code.
00045  *
00046  * This implementation is based on an algorithm described in
00047  *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
00048  *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
00049  *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
00050  * The primary algorithm described there uses 11 multiplies and 29 adds.
00051  * We use their alternate method with 12 multiplies and 32 adds.
00052  * The advantage of this method is that no data path contains more than one
00053  * multiplication; this allows a very simple and accurate implementation in
00054  * scaled fixed-point arithmetic, with a minimal number of shifts.
00055  */
00056 
00062 #include "libavutil/common.h"
00063 #include "dsputil.h"
00064 
00065 #include "bit_depth_template.c"
00066 
00067 #define DCTSIZE 8
00068 #define BITS_IN_JSAMPLE BIT_DEPTH
00069 #define GLOBAL(x) x
00070 #define RIGHT_SHIFT(x, n) ((x) >> (n))
00071 #define MULTIPLY16C16(var,const) ((var)*(const))
00072 
00073 #if 1 //def USE_ACCURATE_ROUNDING
00074 #define DESCALE(x,n)  RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
00075 #else
00076 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
00077 #endif
00078 
00079 
00080 /*
00081  * This module is specialized to the case DCTSIZE = 8.
00082  */
00083 
00084 #if DCTSIZE != 8
00085 #error  "Sorry, this code only copes with 8x8 DCTs."
00086 #endif
00087 
00088 
00089 /*
00090  * The poop on this scaling stuff is as follows:
00091  *
00092  * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
00093  * larger than the true DCT outputs.  The final outputs are therefore
00094  * a factor of N larger than desired; since N=8 this can be cured by
00095  * a simple right shift at the end of the algorithm.  The advantage of
00096  * this arrangement is that we save two multiplications per 1-D DCT,
00097  * because the y0 and y4 outputs need not be divided by sqrt(N).
00098  * In the IJG code, this factor of 8 is removed by the quantization step
00099  * (in jcdctmgr.c), NOT in this module.
00100  *
00101  * We have to do addition and subtraction of the integer inputs, which
00102  * is no problem, and multiplication by fractional constants, which is
00103  * a problem to do in integer arithmetic.  We multiply all the constants
00104  * by CONST_SCALE and convert them to integer constants (thus retaining
00105  * CONST_BITS bits of precision in the constants).  After doing a
00106  * multiplication we have to divide the product by CONST_SCALE, with proper
00107  * rounding, to produce the correct output.  This division can be done
00108  * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
00109  * as long as possible so that partial sums can be added together with
00110  * full fractional precision.
00111  *
00112  * The outputs of the first pass are scaled up by PASS1_BITS bits so that
00113  * they are represented to better-than-integral precision.  These outputs
00114  * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
00115  * with the recommended scaling.  (For 12-bit sample data, the intermediate
00116  * array is int32_t anyway.)
00117  *
00118  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
00119  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
00120  * shows that the values given below are the most effective.
00121  */
00122 
00123 #undef CONST_BITS
00124 #undef PASS1_BITS
00125 #undef OUT_SHIFT
00126 
00127 #if BITS_IN_JSAMPLE == 8
00128 #define CONST_BITS  13
00129 #define PASS1_BITS  4   /* set this to 2 if 16x16 multiplies are faster */
00130 #define OUT_SHIFT   PASS1_BITS
00131 #else
00132 #define CONST_BITS  13
00133 #define PASS1_BITS  1   /* lose a little precision to avoid overflow */
00134 #define OUT_SHIFT   (PASS1_BITS + 1)
00135 #endif
00136 
00137 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00138  * causing a lot of useless floating-point operations at run time.
00139  * To get around this we use the following pre-calculated constants.
00140  * If you change CONST_BITS you may want to add appropriate values.
00141  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00142  */
00143 
00144 #if CONST_BITS == 13
00145 #define FIX_0_298631336  ((int32_t)  2446)      /* FIX(0.298631336) */
00146 #define FIX_0_390180644  ((int32_t)  3196)      /* FIX(0.390180644) */
00147 #define FIX_0_541196100  ((int32_t)  4433)      /* FIX(0.541196100) */
00148 #define FIX_0_765366865  ((int32_t)  6270)      /* FIX(0.765366865) */
00149 #define FIX_0_899976223  ((int32_t)  7373)      /* FIX(0.899976223) */
00150 #define FIX_1_175875602  ((int32_t)  9633)      /* FIX(1.175875602) */
00151 #define FIX_1_501321110  ((int32_t)  12299)     /* FIX(1.501321110) */
00152 #define FIX_1_847759065  ((int32_t)  15137)     /* FIX(1.847759065) */
00153 #define FIX_1_961570560  ((int32_t)  16069)     /* FIX(1.961570560) */
00154 #define FIX_2_053119869  ((int32_t)  16819)     /* FIX(2.053119869) */
00155 #define FIX_2_562915447  ((int32_t)  20995)     /* FIX(2.562915447) */
00156 #define FIX_3_072711026  ((int32_t)  25172)     /* FIX(3.072711026) */
00157 #else
00158 #define FIX_0_298631336  FIX(0.298631336)
00159 #define FIX_0_390180644  FIX(0.390180644)
00160 #define FIX_0_541196100  FIX(0.541196100)
00161 #define FIX_0_765366865  FIX(0.765366865)
00162 #define FIX_0_899976223  FIX(0.899976223)
00163 #define FIX_1_175875602  FIX(1.175875602)
00164 #define FIX_1_501321110  FIX(1.501321110)
00165 #define FIX_1_847759065  FIX(1.847759065)
00166 #define FIX_1_961570560  FIX(1.961570560)
00167 #define FIX_2_053119869  FIX(2.053119869)
00168 #define FIX_2_562915447  FIX(2.562915447)
00169 #define FIX_3_072711026  FIX(3.072711026)
00170 #endif
00171 
00172 
00173 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
00174  * For 8-bit samples with the recommended scaling, all the variable
00175  * and constant values involved are no more than 16 bits wide, so a
00176  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
00177  * For 12-bit samples, a full 32-bit multiplication will be needed.
00178  */
00179 
00180 #if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
00181 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
00182 #else
00183 #define MULTIPLY(var,const)  ((var) * (const))
00184 #endif
00185 
00186 
00187 static av_always_inline void FUNC(row_fdct)(DCTELEM *data)
00188 {
00189   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00190   int tmp10, tmp11, tmp12, tmp13;
00191   int z1, z2, z3, z4, z5;
00192   DCTELEM *dataptr;
00193   int ctr;
00194 
00195   /* Pass 1: process rows. */
00196   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00197   /* furthermore, we scale the results by 2**PASS1_BITS. */
00198 
00199   dataptr = data;
00200   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00201     tmp0 = dataptr[0] + dataptr[7];
00202     tmp7 = dataptr[0] - dataptr[7];
00203     tmp1 = dataptr[1] + dataptr[6];
00204     tmp6 = dataptr[1] - dataptr[6];
00205     tmp2 = dataptr[2] + dataptr[5];
00206     tmp5 = dataptr[2] - dataptr[5];
00207     tmp3 = dataptr[3] + dataptr[4];
00208     tmp4 = dataptr[3] - dataptr[4];
00209 
00210     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00211      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00212      */
00213 
00214     tmp10 = tmp0 + tmp3;
00215     tmp13 = tmp0 - tmp3;
00216     tmp11 = tmp1 + tmp2;
00217     tmp12 = tmp1 - tmp2;
00218 
00219     dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
00220     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
00221 
00222     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00223     dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00224                                    CONST_BITS-PASS1_BITS);
00225     dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00226                                    CONST_BITS-PASS1_BITS);
00227 
00228     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00229      * cK represents cos(K*pi/16).
00230      * i0..i3 in the paper are tmp4..tmp7 here.
00231      */
00232 
00233     z1 = tmp4 + tmp7;
00234     z2 = tmp5 + tmp6;
00235     z3 = tmp4 + tmp6;
00236     z4 = tmp5 + tmp7;
00237     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00238 
00239     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00240     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00241     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00242     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00243     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00244     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00245     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00246     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00247 
00248     z3 += z5;
00249     z4 += z5;
00250 
00251     dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
00252     dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
00253     dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
00254     dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
00255 
00256     dataptr += DCTSIZE;         /* advance pointer to next row */
00257   }
00258 }
00259 
00260 /*
00261  * Perform the forward DCT on one block of samples.
00262  */
00263 
00264 GLOBAL(void)
00265 FUNC(ff_jpeg_fdct_islow)(DCTELEM *data)
00266 {
00267   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00268   int tmp10, tmp11, tmp12, tmp13;
00269   int z1, z2, z3, z4, z5;
00270   DCTELEM *dataptr;
00271   int ctr;
00272 
00273   FUNC(row_fdct)(data);
00274 
00275   /* Pass 2: process columns.
00276    * We remove the PASS1_BITS scaling, but leave the results scaled up
00277    * by an overall factor of 8.
00278    */
00279 
00280   dataptr = data;
00281   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00282     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00283     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00284     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00285     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00286     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00287     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00288     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00289     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00290 
00291     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00292      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00293      */
00294 
00295     tmp10 = tmp0 + tmp3;
00296     tmp13 = tmp0 - tmp3;
00297     tmp11 = tmp1 + tmp2;
00298     tmp12 = tmp1 - tmp2;
00299 
00300     dataptr[DCTSIZE*0] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
00301     dataptr[DCTSIZE*4] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
00302 
00303     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00304     dataptr[DCTSIZE*2] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00305                                  CONST_BITS + OUT_SHIFT);
00306     dataptr[DCTSIZE*6] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00307                                  CONST_BITS + OUT_SHIFT);
00308 
00309     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00310      * cK represents cos(K*pi/16).
00311      * i0..i3 in the paper are tmp4..tmp7 here.
00312      */
00313 
00314     z1 = tmp4 + tmp7;
00315     z2 = tmp5 + tmp6;
00316     z3 = tmp4 + tmp6;
00317     z4 = tmp5 + tmp7;
00318     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00319 
00320     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00321     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00322     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00323     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00324     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00325     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00326     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00327     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00328 
00329     z3 += z5;
00330     z4 += z5;
00331 
00332     dataptr[DCTSIZE*7] = DESCALE(tmp4 + z1 + z3, CONST_BITS + OUT_SHIFT);
00333     dataptr[DCTSIZE*5] = DESCALE(tmp5 + z2 + z4, CONST_BITS + OUT_SHIFT);
00334     dataptr[DCTSIZE*3] = DESCALE(tmp6 + z2 + z3, CONST_BITS + OUT_SHIFT);
00335     dataptr[DCTSIZE*1] = DESCALE(tmp7 + z1 + z4, CONST_BITS + OUT_SHIFT);
00336 
00337     dataptr++;                  /* advance pointer to next column */
00338   }
00339 }
00340 
00341 /*
00342  * The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
00343  * on the rows and then, instead of doing even and odd, part on the columns
00344  * you do even part two times.
00345  */
00346 GLOBAL(void)
00347 FUNC(ff_fdct248_islow)(DCTELEM *data)
00348 {
00349   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00350   int tmp10, tmp11, tmp12, tmp13;
00351   int z1;
00352   DCTELEM *dataptr;
00353   int ctr;
00354 
00355   FUNC(row_fdct)(data);
00356 
00357   /* Pass 2: process columns.
00358    * We remove the PASS1_BITS scaling, but leave the results scaled up
00359    * by an overall factor of 8.
00360    */
00361 
00362   dataptr = data;
00363   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00364      tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
00365      tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00366      tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
00367      tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
00368      tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
00369      tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00370      tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
00371      tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
00372 
00373      tmp10 = tmp0 + tmp3;
00374      tmp11 = tmp1 + tmp2;
00375      tmp12 = tmp1 - tmp2;
00376      tmp13 = tmp0 - tmp3;
00377 
00378      dataptr[DCTSIZE*0] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
00379      dataptr[DCTSIZE*4] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
00380 
00381      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00382      dataptr[DCTSIZE*2] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00383                                   CONST_BITS+OUT_SHIFT);
00384      dataptr[DCTSIZE*6] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00385                                   CONST_BITS+OUT_SHIFT);
00386 
00387      tmp10 = tmp4 + tmp7;
00388      tmp11 = tmp5 + tmp6;
00389      tmp12 = tmp5 - tmp6;
00390      tmp13 = tmp4 - tmp7;
00391 
00392      dataptr[DCTSIZE*1] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
00393      dataptr[DCTSIZE*5] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
00394 
00395      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00396      dataptr[DCTSIZE*3] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00397                                   CONST_BITS + OUT_SHIFT);
00398      dataptr[DCTSIZE*7] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00399                                   CONST_BITS + OUT_SHIFT);
00400 
00401      dataptr++;                 /* advance pointer to next column */
00402   }
00403 }