libs/jpeg/jfdctflt.c

   1 /*
   2  * jfdctflt.c
   3  *
   4  * Copyright (C) 1994-1996, Thomas G. Lane.
   5  * Modified 2003-2017 by Guido Vollbeding.
   6  * This file is part of the Independent JPEG Group's software.
   7  * For conditions of distribution and use, see the accompanying README file.
   8  *
   9  * This file contains a floating-point implementation of the
  10  * forward DCT (Discrete Cosine Transform).
  11  *
  12  * This implementation should be more accurate than either of the integer
  13  * DCT implementations.  However, it may not give the same results on all
  14  * machines because of differences in roundoff behavior.  Speed will depend
  15  * on the hardware's floating point capacity.
  16  *
  17  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
  18  * on each column.  Direct algorithms are also available, but they are
  19  * much more complex and seem not to be any faster when reduced to code.
  20  *
  21  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
  22  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
  23  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
  24  * JPEG textbook (see REFERENCES section in file README).  The following code
  25  * is based directly on figure 4-8 in P&M.
  26  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
  27  * possible to arrange the computation so that many of the multiplies are
  28  * simple scalings of the final outputs.  These multiplies can then be
  29  * folded into the multiplications or divisions by the JPEG quantization
  30  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
  31  * to be done in the DCT itself.
  32  * The primary disadvantage of this method is that with a fixed-point
  33  * implementation, accuracy is lost due to imprecise representation of the
  34  * scaled quantization values.  However, that problem does not arise if
  35  * we use floating point arithmetic.
  36  */
  37
  38 #define JPEG_INTERNALS
  39 #include "jinclude.h"
  40 #include "jpeglib.h"
  41 #include "jdct.h"               /* Private declarations for DCT subsystem */
  42
  43 #ifdef DCT_FLOAT_SUPPORTED
  44
  45
  46 /*
  47  * This module is specialized to the case DCTSIZE = 8.
  48  */
  49
  50 #if DCTSIZE != 8
  51   Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */
  52 #endif
  53
  54
  55 /*
  56  * Perform the forward DCT on one block of samples.
  57  *
  58  * cK represents cos(K*pi/16).
  59  */
  60
  61 GLOBAL(void)
  62 jpeg_fdct_float (FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col)
  63 {
  64   FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
  65   FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
  66   FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
  67   FAST_FLOAT *dataptr;
  68   JSAMPROW elemptr;
  69   int ctr;
  70
  71   /* Pass 1: process rows. */
  72
  73   dataptr = data;
  74   for (ctr = 0; ctr < DCTSIZE; ctr++) {
  75     elemptr = sample_data[ctr] + start_col;
  76
  77     /* Load data into workspace */
  78     tmp0 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]));
  79     tmp7 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]));
  80     tmp1 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]));
  81     tmp6 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]));
  82     tmp2 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]));
  83     tmp5 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]));
  84     tmp3 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]));
  85     tmp4 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]));
  86
  87     /* Even part */
  88
  89     tmp10 = tmp0 + tmp3;        /* phase 2 */
  90     tmp13 = tmp0 - tmp3;
  91     tmp11 = tmp1 + tmp2;
  92     tmp12 = tmp1 - tmp2;
  93
  94     /* Apply unsigned->signed conversion. */
  95     dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
  96     dataptr[4] = tmp10 - tmp11;
  97
  98     z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
  99     dataptr[2] = tmp13 + z1;    /* phase 5 */
 100     dataptr[6] = tmp13 - z1;
 101
 102     /* Odd part */
 103
 104     tmp10 = tmp4 + tmp5;        /* phase 2 */
 105     tmp11 = tmp5 + tmp6;
 106     tmp12 = tmp6 + tmp7;
 107
 108     /* The rotator is modified from fig 4-8 to avoid extra negations. */
 109     z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
 110     z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
 111     z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
 112     z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
 113
 114     z11 = tmp7 + z3;            /* phase 5 */
 115     z13 = tmp7 - z3;
 116
 117     dataptr[5] = z13 + z2;      /* phase 6 */
 118     dataptr[3] = z13 - z2;
 119     dataptr[1] = z11 + z4;
 120     dataptr[7] = z11 - z4;
 121
 122     dataptr += DCTSIZE;         /* advance pointer to next row */
 123   }
 124
 125   /* Pass 2: process columns. */
 126
 127   dataptr = data;
 128   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
 129     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
 130     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
 131     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
 132     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
 133     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
 134     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
 135     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
 136     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
 137
 138     /* Even part */
 139
 140     tmp10 = tmp0 + tmp3;        /* phase 2 */
 141     tmp13 = tmp0 - tmp3;
 142     tmp11 = tmp1 + tmp2;
 143     tmp12 = tmp1 - tmp2;
 144
 145     dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
 146     dataptr[DCTSIZE*4] = tmp10 - tmp11;
 147
 148     z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
 149     dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
 150     dataptr[DCTSIZE*6] = tmp13 - z1;
 151
 152     /* Odd part */
 153
 154     tmp10 = tmp4 + tmp5;        /* phase 2 */
 155     tmp11 = tmp5 + tmp6;
 156     tmp12 = tmp6 + tmp7;
 157
 158     /* The rotator is modified from fig 4-8 to avoid extra negations. */
 159     z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
 160     z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
 161     z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
 162     z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
 163
 164     z11 = tmp7 + z3;            /* phase 5 */
 165     z13 = tmp7 - z3;
 166
 167     dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
 168     dataptr[DCTSIZE*3] = z13 - z2;
 169     dataptr[DCTSIZE*1] = z11 + z4;
 170     dataptr[DCTSIZE*7] = z11 - z4;
 171
 172     dataptr++;                  /* advance pointer to next column */
 173   }
 174 }
 175
 176 #endif /* DCT_FLOAT_SUPPORTED */