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

libavcodec/jfdctfst.c

Go to the documentation of this file.
00001 /*
00002  * jfdctfst.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) 1994-1996, 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 fast, not so 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 Arai, Agui, and Nakajima's algorithm for
00049  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
00050  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
00051  * JPEG textbook (see REFERENCES section in file README).  The following code
00052  * is based directly on figure 4-8 in P&M.
00053  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
00054  * possible to arrange the computation so that many of the multiplies are
00055  * simple scalings of the final outputs.  These multiplies can then be
00056  * folded into the multiplications or divisions by the JPEG quantization
00057  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
00058  * to be done in the DCT itself.
00059  * The primary disadvantage of this method is that with fixed-point math,
00060  * accuracy is lost due to imprecise representation of the scaled
00061  * quantization values.  The smaller the quantization table entry, the less
00062  * precise the scaled value, so this implementation does worse with high-
00063  * quality-setting files than with low-quality ones.
00064  */
00065 
00071 #include <stdlib.h>
00072 #include <stdio.h>
00073 #include "libavutil/common.h"
00074 #include "dsputil.h"
00075 
00076 #define DCTSIZE 8
00077 #define GLOBAL(x) x
00078 #define RIGHT_SHIFT(x, n) ((x) >> (n))
00079 
00080 /*
00081  * This module is specialized to the case DCTSIZE = 8.
00082  */
00083 
00084 #if DCTSIZE != 8
00085   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00086 #endif
00087 
00088 
00089 /* Scaling decisions are generally the same as in the LL&M algorithm;
00090  * see jfdctint.c for more details.  However, we choose to descale
00091  * (right shift) multiplication products as soon as they are formed,
00092  * rather than carrying additional fractional bits into subsequent additions.
00093  * This compromises accuracy slightly, but it lets us save a few shifts.
00094  * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
00095  * everywhere except in the multiplications proper; this saves a good deal
00096  * of work on 16-bit-int machines.
00097  *
00098  * Again to save a few shifts, the intermediate results between pass 1 and
00099  * pass 2 are not upscaled, but are represented only to integral precision.
00100  *
00101  * A final compromise is to represent the multiplicative constants to only
00102  * 8 fractional bits, rather than 13.  This saves some shifting work on some
00103  * machines, and may also reduce the cost of multiplication (since there
00104  * are fewer one-bits in the constants).
00105  */
00106 
00107 #define CONST_BITS  8
00108 
00109 
00110 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00111  * causing a lot of useless floating-point operations at run time.
00112  * To get around this we use the following pre-calculated constants.
00113  * If you change CONST_BITS you may want to add appropriate values.
00114  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00115  */
00116 
00117 #if CONST_BITS == 8
00118 #define FIX_0_382683433  ((int32_t)   98)       /* FIX(0.382683433) */
00119 #define FIX_0_541196100  ((int32_t)  139)       /* FIX(0.541196100) */
00120 #define FIX_0_707106781  ((int32_t)  181)       /* FIX(0.707106781) */
00121 #define FIX_1_306562965  ((int32_t)  334)       /* FIX(1.306562965) */
00122 #else
00123 #define FIX_0_382683433  FIX(0.382683433)
00124 #define FIX_0_541196100  FIX(0.541196100)
00125 #define FIX_0_707106781  FIX(0.707106781)
00126 #define FIX_1_306562965  FIX(1.306562965)
00127 #endif
00128 
00129 
00130 /* We can gain a little more speed, with a further compromise in accuracy,
00131  * by omitting the addition in a descaling shift.  This yields an incorrectly
00132  * rounded result half the time...
00133  */
00134 
00135 #ifndef USE_ACCURATE_ROUNDING
00136 #undef DESCALE
00137 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
00138 #endif
00139 
00140 
00141 /* Multiply a DCTELEM variable by an int32_t constant, and immediately
00142  * descale to yield a DCTELEM result.
00143  */
00144 
00145 #define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
00146 
00147 static av_always_inline void row_fdct(DCTELEM * data){
00148   int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00149   int_fast16_t tmp10, tmp11, tmp12, tmp13;
00150   int_fast16_t z1, z2, z3, z4, z5, z11, z13;
00151   DCTELEM *dataptr;
00152   int ctr;
00153 
00154   /* Pass 1: process rows. */
00155 
00156   dataptr = data;
00157   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00158     tmp0 = dataptr[0] + dataptr[7];
00159     tmp7 = dataptr[0] - dataptr[7];
00160     tmp1 = dataptr[1] + dataptr[6];
00161     tmp6 = dataptr[1] - dataptr[6];
00162     tmp2 = dataptr[2] + dataptr[5];
00163     tmp5 = dataptr[2] - dataptr[5];
00164     tmp3 = dataptr[3] + dataptr[4];
00165     tmp4 = dataptr[3] - dataptr[4];
00166 
00167     /* Even part */
00168 
00169     tmp10 = tmp0 + tmp3;        /* phase 2 */
00170     tmp13 = tmp0 - tmp3;
00171     tmp11 = tmp1 + tmp2;
00172     tmp12 = tmp1 - tmp2;
00173 
00174     dataptr[0] = tmp10 + tmp11; /* phase 3 */
00175     dataptr[4] = tmp10 - tmp11;
00176 
00177     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
00178     dataptr[2] = tmp13 + z1;    /* phase 5 */
00179     dataptr[6] = tmp13 - z1;
00180 
00181     /* Odd part */
00182 
00183     tmp10 = tmp4 + tmp5;        /* phase 2 */
00184     tmp11 = tmp5 + tmp6;
00185     tmp12 = tmp6 + tmp7;
00186 
00187     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00188     z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
00189     z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5;    /* c2-c6 */
00190     z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5;    /* c2+c6 */
00191     z3 = MULTIPLY(tmp11, FIX_0_707106781);         /* c4 */
00192 
00193     z11 = tmp7 + z3;            /* phase 5 */
00194     z13 = tmp7 - z3;
00195 
00196     dataptr[5] = z13 + z2;      /* phase 6 */
00197     dataptr[3] = z13 - z2;
00198     dataptr[1] = z11 + z4;
00199     dataptr[7] = z11 - z4;
00200 
00201     dataptr += DCTSIZE;         /* advance pointer to next row */
00202   }
00203 }
00204 
00205 /*
00206  * Perform the forward DCT on one block of samples.
00207  */
00208 
00209 GLOBAL(void)
00210 fdct_ifast (DCTELEM * data)
00211 {
00212   int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00213   int_fast16_t tmp10, tmp11, tmp12, tmp13;
00214   int_fast16_t z1, z2, z3, z4, z5, z11, z13;
00215   DCTELEM *dataptr;
00216   int ctr;
00217 
00218   row_fdct(data);
00219 
00220   /* Pass 2: process columns. */
00221 
00222   dataptr = data;
00223   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00224     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00225     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00226     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00227     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00228     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00229     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00230     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00231     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00232 
00233     /* Even part */
00234 
00235     tmp10 = tmp0 + tmp3;        /* phase 2 */
00236     tmp13 = tmp0 - tmp3;
00237     tmp11 = tmp1 + tmp2;
00238     tmp12 = tmp1 - tmp2;
00239 
00240     dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
00241     dataptr[DCTSIZE*4] = tmp10 - tmp11;
00242 
00243     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
00244     dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
00245     dataptr[DCTSIZE*6] = tmp13 - z1;
00246 
00247     /* Odd part */
00248 
00249     tmp10 = tmp4 + tmp5;        /* phase 2 */
00250     tmp11 = tmp5 + tmp6;
00251     tmp12 = tmp6 + tmp7;
00252 
00253     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00254     z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
00255     z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
00256     z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
00257     z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
00258 
00259     z11 = tmp7 + z3;            /* phase 5 */
00260     z13 = tmp7 - z3;
00261 
00262     dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
00263     dataptr[DCTSIZE*3] = z13 - z2;
00264     dataptr[DCTSIZE*1] = z11 + z4;
00265     dataptr[DCTSIZE*7] = z11 - z4;
00266 
00267     dataptr++;                  /* advance pointer to next column */
00268   }
00269 }
00270 
00271 /*
00272  * Perform the forward 2-4-8 DCT on one block of samples.
00273  */
00274 
00275 GLOBAL(void)
00276 fdct_ifast248 (DCTELEM * data)
00277 {
00278   int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00279   int_fast16_t tmp10, tmp11, tmp12, tmp13;
00280   int_fast16_t z1;
00281   DCTELEM *dataptr;
00282   int ctr;
00283 
00284   row_fdct(data);
00285 
00286   /* Pass 2: process columns. */
00287 
00288   dataptr = data;
00289   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00290     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
00291     tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00292     tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
00293     tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
00294     tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
00295     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00296     tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
00297     tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
00298 
00299     /* Even part */
00300 
00301     tmp10 = tmp0 + tmp3;
00302     tmp11 = tmp1 + tmp2;
00303     tmp12 = tmp1 - tmp2;
00304     tmp13 = tmp0 - tmp3;
00305 
00306     dataptr[DCTSIZE*0] = tmp10 + tmp11;
00307     dataptr[DCTSIZE*4] = tmp10 - tmp11;
00308 
00309     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
00310     dataptr[DCTSIZE*2] = tmp13 + z1;
00311     dataptr[DCTSIZE*6] = tmp13 - z1;
00312 
00313     tmp10 = tmp4 + tmp7;
00314     tmp11 = tmp5 + tmp6;
00315     tmp12 = tmp5 - tmp6;
00316     tmp13 = tmp4 - tmp7;
00317 
00318     dataptr[DCTSIZE*1] = tmp10 + tmp11;
00319     dataptr[DCTSIZE*5] = tmp10 - tmp11;
00320 
00321     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
00322     dataptr[DCTSIZE*3] = tmp13 + z1;
00323     dataptr[DCTSIZE*7] = tmp13 - z1;
00324 
00325     dataptr++;                        /* advance pointer to next column */
00326   }
00327 }
00328 
00329 
00330 #undef GLOBAL
00331 #undef CONST_BITS
00332 #undef DESCALE
00333 #undef FIX_0_541196100
00334 #undef FIX_1_306562965

Generated on Fri Sep 16 2011 17:17:38 for FFmpeg by  doxygen 1.7.1