Cinelerra: hvirtual/quicktime/ffmpeg/libavcodec/jfdctint.c Source File

00001 /*
00002  * jfdctint.c
00003  *
00004  * Copyright (C) 1991-1996, Thomas G. Lane.
00005  * This file is part of the Independent JPEG Group's software.
00006  * For conditions of distribution and use, see the accompanying README file.
00007  *
00008  * This file contains a slow-but-accurate integer implementation of the
00009  * forward DCT (Discrete Cosine Transform).
00010  *
00011  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
00012  * on each column.  Direct algorithms are also available, but they are
00013  * much more complex and seem not to be any faster when reduced to code.
00014  *
00015  * This implementation is based on an algorithm described in
00016  *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
00017  *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
00018  *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
00019  * The primary algorithm described there uses 11 multiplies and 29 adds.
00020  * We use their alternate method with 12 multiplies and 32 adds.
00021  * The advantage of this method is that no data path contains more than one
00022  * multiplication; this allows a very simple and accurate implementation in
00023  * scaled fixed-point arithmetic, with a minimal number of shifts.
00024  */
00025 
00031 #include <stdlib.h>
00032 #include <stdio.h>
00033 #include "common.h"
00034 #include "dsputil.h"
00035 
00036 #define SHIFT_TEMPS
00037 #define DCTSIZE 8
00038 #define BITS_IN_JSAMPLE 8
00039 #define GLOBAL(x) x
00040 #define RIGHT_SHIFT(x, n) ((x) >> (n))
00041 #define MULTIPLY16C16(var,const) ((var)*(const))
00042 
00043 #if 1 //def USE_ACCURATE_ROUNDING
00044 #define DESCALE(x,n)  RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
00045 #else
00046 #define DESCALE(x,n)  RIGHT_SHIFT(x, n)
00047 #endif
00048 
00049 
00050 /*
00051  * This module is specialized to the case DCTSIZE = 8.
00052  */
00053 
00054 #if DCTSIZE != 8
00055   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00056 #endif
00057 
00058 
00059 /*
00060  * The poop on this scaling stuff is as follows:
00061  *
00062  * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
00063  * larger than the true DCT outputs.  The final outputs are therefore
00064  * a factor of N larger than desired; since N=8 this can be cured by
00065  * a simple right shift at the end of the algorithm.  The advantage of
00066  * this arrangement is that we save two multiplications per 1-D DCT,
00067  * because the y0 and y4 outputs need not be divided by sqrt(N).
00068  * In the IJG code, this factor of 8 is removed by the quantization step
00069  * (in jcdctmgr.c), NOT in this module.
00070  *
00071  * We have to do addition and subtraction of the integer inputs, which
00072  * is no problem, and multiplication by fractional constants, which is
00073  * a problem to do in integer arithmetic.  We multiply all the constants
00074  * by CONST_SCALE and convert them to integer constants (thus retaining
00075  * CONST_BITS bits of precision in the constants).  After doing a
00076  * multiplication we have to divide the product by CONST_SCALE, with proper
00077  * rounding, to produce the correct output.  This division can be done
00078  * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
00079  * as long as possible so that partial sums can be added together with
00080  * full fractional precision.
00081  *
00082  * The outputs of the first pass are scaled up by PASS1_BITS bits so that
00083  * they are represented to better-than-integral precision.  These outputs
00084  * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
00085  * with the recommended scaling.  (For 12-bit sample data, the intermediate
00086  * array is int32_t anyway.)
00087  *
00088  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
00089  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
00090  * shows that the values given below are the most effective.
00091  */
00092 
00093 #if BITS_IN_JSAMPLE == 8
00094 #define CONST_BITS  13
00095 #define PASS1_BITS  4           /* set this to 2 if 16x16 multiplies are faster */
00096 #else
00097 #define CONST_BITS  13
00098 #define PASS1_BITS  1           /* lose a little precision to avoid overflow */
00099 #endif
00100 
00101 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00102  * causing a lot of useless floating-point operations at run time.
00103  * To get around this we use the following pre-calculated constants.
00104  * If you change CONST_BITS you may want to add appropriate values.
00105  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00106  */
00107 
00108 #if CONST_BITS == 13
00109 #define FIX_0_298631336  ((int32_t)  2446)      /* FIX(0.298631336) */
00110 #define FIX_0_390180644  ((int32_t)  3196)      /* FIX(0.390180644) */
00111 #define FIX_0_541196100  ((int32_t)  4433)      /* FIX(0.541196100) */
00112 #define FIX_0_765366865  ((int32_t)  6270)      /* FIX(0.765366865) */
00113 #define FIX_0_899976223  ((int32_t)  7373)      /* FIX(0.899976223) */
00114 #define FIX_1_175875602  ((int32_t)  9633)      /* FIX(1.175875602) */
00115 #define FIX_1_501321110  ((int32_t)  12299)     /* FIX(1.501321110) */
00116 #define FIX_1_847759065  ((int32_t)  15137)     /* FIX(1.847759065) */
00117 #define FIX_1_961570560  ((int32_t)  16069)     /* FIX(1.961570560) */
00118 #define FIX_2_053119869  ((int32_t)  16819)     /* FIX(2.053119869) */
00119 #define FIX_2_562915447  ((int32_t)  20995)     /* FIX(2.562915447) */
00120 #define FIX_3_072711026  ((int32_t)  25172)     /* FIX(3.072711026) */
00121 #else
00122 #define FIX_0_298631336  FIX(0.298631336)
00123 #define FIX_0_390180644  FIX(0.390180644)
00124 #define FIX_0_541196100  FIX(0.541196100)
00125 #define FIX_0_765366865  FIX(0.765366865)
00126 #define FIX_0_899976223  FIX(0.899976223)
00127 #define FIX_1_175875602  FIX(1.175875602)
00128 #define FIX_1_501321110  FIX(1.501321110)
00129 #define FIX_1_847759065  FIX(1.847759065)
00130 #define FIX_1_961570560  FIX(1.961570560)
00131 #define FIX_2_053119869  FIX(2.053119869)
00132 #define FIX_2_562915447  FIX(2.562915447)
00133 #define FIX_3_072711026  FIX(3.072711026)
00134 #endif
00135 
00136 
00137 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
00138  * For 8-bit samples with the recommended scaling, all the variable
00139  * and constant values involved are no more than 16 bits wide, so a
00140  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
00141  * For 12-bit samples, a full 32-bit multiplication will be needed.
00142  */
00143 
00144 #if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
00145 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
00146 #else
00147 #define MULTIPLY(var,const)  ((var) * (const))
00148 #endif
00149 
00150 
00151 static always_inline void row_fdct(DCTELEM * data){
00152   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00153   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00154   int_fast32_t z1, z2, z3, z4, z5;
00155   DCTELEM *dataptr;
00156   int ctr;
00157   SHIFT_TEMPS
00158 
00159   /* Pass 1: process rows. */
00160   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00161   /* furthermore, we scale the results by 2**PASS1_BITS. */
00162 
00163   dataptr = data;
00164   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00165     tmp0 = dataptr[0] + dataptr[7];
00166     tmp7 = dataptr[0] - dataptr[7];
00167     tmp1 = dataptr[1] + dataptr[6];
00168     tmp6 = dataptr[1] - dataptr[6];
00169     tmp2 = dataptr[2] + dataptr[5];
00170     tmp5 = dataptr[2] - dataptr[5];
00171     tmp3 = dataptr[3] + dataptr[4];
00172     tmp4 = dataptr[3] - dataptr[4];
00173     
00174     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00175      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00176      */
00177     
00178     tmp10 = tmp0 + tmp3;
00179     tmp13 = tmp0 - tmp3;
00180     tmp11 = tmp1 + tmp2;
00181     tmp12 = tmp1 - tmp2;
00182     
00183     dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
00184     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
00185     
00186     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00187     dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00188                                    CONST_BITS-PASS1_BITS);
00189     dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00190                                    CONST_BITS-PASS1_BITS);
00191     
00192     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00193      * cK represents cos(K*pi/16).
00194      * i0..i3 in the paper are tmp4..tmp7 here.
00195      */
00196     
00197     z1 = tmp4 + tmp7;
00198     z2 = tmp5 + tmp6;
00199     z3 = tmp4 + tmp6;
00200     z4 = tmp5 + tmp7;
00201     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00202     
00203     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00204     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00205     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00206     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00207     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00208     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00209     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00210     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00211     
00212     z3 += z5;
00213     z4 += z5;
00214     
00215     dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
00216     dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
00217     dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
00218     dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
00219     
00220     dataptr += DCTSIZE;         /* advance pointer to next row */
00221   }
00222 }
00223 
00224 /*
00225  * Perform the forward DCT on one block of samples.
00226  */
00227 
00228 GLOBAL(void)
00229 ff_jpeg_fdct_islow (DCTELEM * data)
00230 {
00231   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00232   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00233   int_fast32_t z1, z2, z3, z4, z5;
00234   DCTELEM *dataptr;
00235   int ctr;
00236   SHIFT_TEMPS
00237 
00238   row_fdct(data);
00239 
00240   /* Pass 2: process columns.
00241    * We remove the PASS1_BITS scaling, but leave the results scaled up
00242    * by an overall factor of 8.
00243    */
00244 
00245   dataptr = data;
00246   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00247     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00248     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00249     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00250     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00251     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00252     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00253     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00254     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00255     
00256     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00257      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00258      */
00259     
00260     tmp10 = tmp0 + tmp3;
00261     tmp13 = tmp0 - tmp3;
00262     tmp11 = tmp1 + tmp2;
00263     tmp12 = tmp1 - tmp2;
00264     
00265     dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00266     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00267     
00268     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00269     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00270                                            CONST_BITS+PASS1_BITS);
00271     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00272                                            CONST_BITS+PASS1_BITS);
00273     
00274     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00275      * cK represents cos(K*pi/16).
00276      * i0..i3 in the paper are tmp4..tmp7 here.
00277      */
00278     
00279     z1 = tmp4 + tmp7;
00280     z2 = tmp5 + tmp6;
00281     z3 = tmp4 + tmp6;
00282     z4 = tmp5 + tmp7;
00283     z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
00284     
00285     tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
00286     tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
00287     tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
00288     tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
00289     z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
00290     z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
00291     z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
00292     z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
00293     
00294     z3 += z5;
00295     z4 += z5;
00296     
00297     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
00298                                            CONST_BITS+PASS1_BITS);
00299     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
00300                                            CONST_BITS+PASS1_BITS);
00301     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
00302                                            CONST_BITS+PASS1_BITS);
00303     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
00304                                            CONST_BITS+PASS1_BITS);
00305     
00306     dataptr++;                  /* advance pointer to next column */
00307   }
00308 }
00309 
00310 /*
00311  * The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
00312  * on the rows and then, instead of doing even and odd, part on the colums
00313  * you do even part two times.
00314  */
00315 GLOBAL(void)
00316 ff_fdct248_islow (DCTELEM * data)
00317 {
00318   int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00319   int_fast32_t tmp10, tmp11, tmp12, tmp13;
00320   int_fast32_t z1;
00321   DCTELEM *dataptr;
00322   int ctr;
00323   SHIFT_TEMPS
00324 
00325   row_fdct(data);
00326 
00327   /* Pass 2: process columns.
00328    * We remove the PASS1_BITS scaling, but leave the results scaled up
00329    * by an overall factor of 8.
00330    */
00331 
00332   dataptr = data;
00333   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00334      tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
00335      tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00336      tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
00337      tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
00338      tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
00339      tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00340      tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
00341      tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
00342       
00343      tmp10 = tmp0 + tmp3;
00344      tmp11 = tmp1 + tmp2;
00345      tmp12 = tmp1 - tmp2;
00346      tmp13 = tmp0 - tmp3;
00347      
00348      dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00349      dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00350      
00351      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00352      dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00353                                             CONST_BITS+PASS1_BITS);
00354      dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00355                                             CONST_BITS+PASS1_BITS);
00356 
00357      tmp10 = tmp4 + tmp7;
00358      tmp11 = tmp5 + tmp6;
00359      tmp12 = tmp5 - tmp6;
00360      tmp13 = tmp4 - tmp7;
00361 
00362      dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
00363      dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
00364      
00365      z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00366      dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
00367                                             CONST_BITS+PASS1_BITS);
00368      dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
00369                                             CONST_BITS+PASS1_BITS);
00370     
00371      dataptr++;                 /* advance pointer to next column */
00372   }
00373 }