• Main Page
  • Related Pages
  • Modules
  • Data Structures
  • Files
  • File List
  • Globals

libavcodec/jfdctfst.c

Go to the documentation of this 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) 1994-1996, 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 fast, not so 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 Arai, Agui, and Nakajima's algorithm for
00047  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
00048  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
00049  * JPEG textbook (see REFERENCES section in file README).  The following code
00050  * is based directly on figure 4-8 in P&M.
00051  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
00052  * possible to arrange the computation so that many of the multiplies are
00053  * simple scalings of the final outputs.  These multiplies can then be
00054  * folded into the multiplications or divisions by the JPEG quantization
00055  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
00056  * to be done in the DCT itself.
00057  * The primary disadvantage of this method is that with fixed-point math,
00058  * accuracy is lost due to imprecise representation of the scaled
00059  * quantization values.  The smaller the quantization table entry, the less
00060  * precise the scaled value, so this implementation does worse with high-
00061  * quality-setting files than with low-quality ones.
00062  */
00063 
00069 #include <stdlib.h>
00070 #include <stdio.h>
00071 #include "libavutil/common.h"
00072 #include "dsputil.h"
00073 
00074 #define DCTSIZE 8
00075 #define GLOBAL(x) x
00076 #define RIGHT_SHIFT(x, n) ((x) >> (n))
00077 
00078 /*
00079  * This module is specialized to the case DCTSIZE = 8.
00080  */
00081 
00082 #if DCTSIZE != 8
00083   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00084 #endif
00085 
00086 
00087 /* Scaling decisions are generally the same as in the LL&M algorithm;
00088  * see jfdctint.c for more details.  However, we choose to descale
00089  * (right shift) multiplication products as soon as they are formed,
00090  * rather than carrying additional fractional bits into subsequent additions.
00091  * This compromises accuracy slightly, but it lets us save a few shifts.
00092  * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
00093  * everywhere except in the multiplications proper; this saves a good deal
00094  * of work on 16-bit-int machines.
00095  *
00096  * Again to save a few shifts, the intermediate results between pass 1 and
00097  * pass 2 are not upscaled, but are represented only to integral precision.
00098  *
00099  * A final compromise is to represent the multiplicative constants to only
00100  * 8 fractional bits, rather than 13.  This saves some shifting work on some
00101  * machines, and may also reduce the cost of multiplication (since there
00102  * are fewer one-bits in the constants).
00103  */
00104 
00105 #define CONST_BITS  8
00106 
00107 
00108 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00109  * causing a lot of useless floating-point operations at run time.
00110  * To get around this we use the following pre-calculated constants.
00111  * If you change CONST_BITS you may want to add appropriate values.
00112  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00113  */
00114 
00115 #if CONST_BITS == 8
00116 #define FIX_0_382683433  ((int32_t)   98)       /* FIX(0.382683433) */
00117 #define FIX_0_541196100  ((int32_t)  139)       /* FIX(0.541196100) */
00118 #define FIX_0_707106781  ((int32_t)  181)       /* FIX(0.707106781) */
00119 #define FIX_1_306562965  ((int32_t)  334)       /* FIX(1.306562965) */
00120 #else
00121 #define FIX_0_382683433  FIX(0.382683433)
00122 #define FIX_0_541196100  FIX(0.541196100)
00123 #define FIX_0_707106781  FIX(0.707106781)
00124 #define FIX_1_306562965  FIX(1.306562965)
00125 #endif
00126 
00127 
00128 /* We can gain a little more speed, with a further compromise in accuracy,
00129  * by omitting the addition in a descaling shift.  This yields an incorrectly
00130  * rounded result half the time...
00131  */
00132 
00133 #ifndef USE_ACCURATE_ROUNDING
00134 #undef DESCALE
00135 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
00136 #endif
00137 
00138 
00139 /* Multiply a DCTELEM variable by an int32_t constant, and immediately
00140  * descale to yield a DCTELEM result.
00141  */
00142 
00143 #define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
00144 
00145 static av_always_inline void row_fdct(DCTELEM * data){
00146   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00147   int tmp10, tmp11, tmp12, tmp13;
00148   int z1, z2, z3, z4, z5, z11, z13;
00149   DCTELEM *dataptr;
00150   int ctr;
00151 
00152   /* Pass 1: process rows. */
00153 
00154   dataptr = data;
00155   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00156     tmp0 = dataptr[0] + dataptr[7];
00157     tmp7 = dataptr[0] - dataptr[7];
00158     tmp1 = dataptr[1] + dataptr[6];
00159     tmp6 = dataptr[1] - dataptr[6];
00160     tmp2 = dataptr[2] + dataptr[5];
00161     tmp5 = dataptr[2] - dataptr[5];
00162     tmp3 = dataptr[3] + dataptr[4];
00163     tmp4 = dataptr[3] - dataptr[4];
00164 
00165     /* Even part */
00166 
00167     tmp10 = tmp0 + tmp3;        /* phase 2 */
00168     tmp13 = tmp0 - tmp3;
00169     tmp11 = tmp1 + tmp2;
00170     tmp12 = tmp1 - tmp2;
00171 
00172     dataptr[0] = tmp10 + tmp11; /* phase 3 */
00173     dataptr[4] = tmp10 - tmp11;
00174 
00175     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
00176     dataptr[2] = tmp13 + z1;    /* phase 5 */
00177     dataptr[6] = tmp13 - z1;
00178 
00179     /* Odd part */
00180 
00181     tmp10 = tmp4 + tmp5;        /* phase 2 */
00182     tmp11 = tmp5 + tmp6;
00183     tmp12 = tmp6 + tmp7;
00184 
00185     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00186     z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
00187     z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5;    /* c2-c6 */
00188     z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5;    /* c2+c6 */
00189     z3 = MULTIPLY(tmp11, FIX_0_707106781);         /* c4 */
00190 
00191     z11 = tmp7 + z3;            /* phase 5 */
00192     z13 = tmp7 - z3;
00193 
00194     dataptr[5] = z13 + z2;      /* phase 6 */
00195     dataptr[3] = z13 - z2;
00196     dataptr[1] = z11 + z4;
00197     dataptr[7] = z11 - z4;
00198 
00199     dataptr += DCTSIZE;         /* advance pointer to next row */
00200   }
00201 }
00202 
00203 /*
00204  * Perform the forward DCT on one block of samples.
00205  */
00206 
00207 GLOBAL(void)
00208 fdct_ifast (DCTELEM * data)
00209 {
00210   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00211   int tmp10, tmp11, tmp12, tmp13;
00212   int z1, z2, z3, z4, z5, z11, z13;
00213   DCTELEM *dataptr;
00214   int ctr;
00215 
00216   row_fdct(data);
00217 
00218   /* Pass 2: process columns. */
00219 
00220   dataptr = data;
00221   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00222     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00223     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00224     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00225     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00226     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00227     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00228     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00229     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00230 
00231     /* Even part */
00232 
00233     tmp10 = tmp0 + tmp3;        /* phase 2 */
00234     tmp13 = tmp0 - tmp3;
00235     tmp11 = tmp1 + tmp2;
00236     tmp12 = tmp1 - tmp2;
00237 
00238     dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
00239     dataptr[DCTSIZE*4] = tmp10 - tmp11;
00240 
00241     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
00242     dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
00243     dataptr[DCTSIZE*6] = tmp13 - z1;
00244 
00245     /* Odd part */
00246 
00247     tmp10 = tmp4 + tmp5;        /* phase 2 */
00248     tmp11 = tmp5 + tmp6;
00249     tmp12 = tmp6 + tmp7;
00250 
00251     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00252     z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
00253     z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
00254     z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
00255     z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
00256 
00257     z11 = tmp7 + z3;            /* phase 5 */
00258     z13 = tmp7 - z3;
00259 
00260     dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
00261     dataptr[DCTSIZE*3] = z13 - z2;
00262     dataptr[DCTSIZE*1] = z11 + z4;
00263     dataptr[DCTSIZE*7] = z11 - z4;
00264 
00265     dataptr++;                  /* advance pointer to next column */
00266   }
00267 }
00268 
00269 /*
00270  * Perform the forward 2-4-8 DCT on one block of samples.
00271  */
00272 
00273 GLOBAL(void)
00274 fdct_ifast248 (DCTELEM * data)
00275 {
00276   int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00277   int tmp10, tmp11, tmp12, tmp13;
00278   int z1;
00279   DCTELEM *dataptr;
00280   int ctr;
00281 
00282   row_fdct(data);
00283 
00284   /* Pass 2: process columns. */
00285 
00286   dataptr = data;
00287   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00288     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
00289     tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00290     tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
00291     tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
00292     tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
00293     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00294     tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
00295     tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
00296 
00297     /* Even part */
00298 
00299     tmp10 = tmp0 + tmp3;
00300     tmp11 = tmp1 + tmp2;
00301     tmp12 = tmp1 - tmp2;
00302     tmp13 = tmp0 - tmp3;
00303 
00304     dataptr[DCTSIZE*0] = tmp10 + tmp11;
00305     dataptr[DCTSIZE*4] = tmp10 - tmp11;
00306 
00307     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
00308     dataptr[DCTSIZE*2] = tmp13 + z1;
00309     dataptr[DCTSIZE*6] = tmp13 - z1;
00310 
00311     tmp10 = tmp4 + tmp7;
00312     tmp11 = tmp5 + tmp6;
00313     tmp12 = tmp5 - tmp6;
00314     tmp13 = tmp4 - tmp7;
00315 
00316     dataptr[DCTSIZE*1] = tmp10 + tmp11;
00317     dataptr[DCTSIZE*5] = tmp10 - tmp11;
00318 
00319     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
00320     dataptr[DCTSIZE*3] = tmp13 + z1;
00321     dataptr[DCTSIZE*7] = tmp13 - z1;
00322 
00323     dataptr++;                        /* advance pointer to next column */
00324   }
00325 }
00326 
00327 
00328 #undef GLOBAL
00329 #undef CONST_BITS
00330 #undef DESCALE
00331 #undef FIX_0_541196100
00332 #undef FIX_1_306562965
Generated on Fri Feb 1 2013 14:34:37 for FFmpeg by doxygen 1.7.1