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