4 * Copyright (C) 1994-1996, Thomas G. Lane.
\r
5 * Modified 2003-2009 by Guido Vollbeding.
\r
6 * This file is part of the Independent JPEG Group's software.
\r
7 * For conditions of distribution and use, see the accompanying README file.
\r
9 * This file contains a floating-point implementation of the
\r
10 * forward DCT (Discrete Cosine Transform).
\r
12 * This implementation should be more accurate than either of the integer
\r
13 * DCT implementations. However, it may not give the same results on all
\r
14 * machines because of differences in roundoff behavior. Speed will depend
\r
15 * on the hardware's floating point capacity.
\r
17 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
\r
18 * on each column. Direct algorithms are also available, but they are
\r
19 * much more complex and seem not to be any faster when reduced to code.
\r
21 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
\r
22 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
\r
23 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
\r
24 * JPEG textbook (see REFERENCES section in file README). The following code
\r
25 * is based directly on figure 4-8 in P&M.
\r
26 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
\r
27 * possible to arrange the computation so that many of the multiplies are
\r
28 * simple scalings of the final outputs. These multiplies can then be
\r
29 * folded into the multiplications or divisions by the JPEG quantization
\r
30 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
\r
31 * to be done in the DCT itself.
\r
32 * The primary disadvantage of this method is that with a fixed-point
\r
33 * implementation, accuracy is lost due to imprecise representation of the
\r
34 * scaled quantization values. However, that problem does not arise if
\r
35 * we use floating point arithmetic.
\r
38 #define JPEG_INTERNALS
\r
39 #include "jinclude.h"
\r
40 #include "jpeglib.h"
\r
41 #include "jdct.h" /* Private declarations for DCT subsystem */
\r
43 #ifdef DCT_FLOAT_SUPPORTED
\r
47 * This module is specialized to the case DCTSIZE = 8.
\r
51 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
\r
56 * Perform the forward DCT on one block of samples.
\r
60 jpeg_fdct_float (FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col)
\r
62 FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
\r
63 FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
\r
64 FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
\r
65 FAST_FLOAT *dataptr;
\r
69 /* Pass 1: process rows. */
\r
72 for (ctr = 0; ctr < DCTSIZE; ctr++) {
\r
73 elemptr = sample_data[ctr] + start_col;
\r
75 /* Load data into workspace */
\r
76 tmp0 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]));
\r
77 tmp7 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]));
\r
78 tmp1 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]));
\r
79 tmp6 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]));
\r
80 tmp2 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]));
\r
81 tmp5 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]));
\r
82 tmp3 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]));
\r
83 tmp4 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]));
\r
87 tmp10 = tmp0 + tmp3; /* phase 2 */
\r
88 tmp13 = tmp0 - tmp3;
\r
89 tmp11 = tmp1 + tmp2;
\r
90 tmp12 = tmp1 - tmp2;
\r
92 /* Apply unsigned->signed conversion */
\r
93 dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
\r
94 dataptr[4] = tmp10 - tmp11;
\r
96 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
\r
97 dataptr[2] = tmp13 + z1; /* phase 5 */
\r
98 dataptr[6] = tmp13 - z1;
\r
102 tmp10 = tmp4 + tmp5; /* phase 2 */
\r
103 tmp11 = tmp5 + tmp6;
\r
104 tmp12 = tmp6 + tmp7;
\r
106 /* The rotator is modified from fig 4-8 to avoid extra negations. */
\r
107 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
\r
108 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
\r
109 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
\r
110 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
\r
112 z11 = tmp7 + z3; /* phase 5 */
\r
115 dataptr[5] = z13 + z2; /* phase 6 */
\r
116 dataptr[3] = z13 - z2;
\r
117 dataptr[1] = z11 + z4;
\r
118 dataptr[7] = z11 - z4;
\r
120 dataptr += DCTSIZE; /* advance pointer to next row */
\r
123 /* Pass 2: process columns. */
\r
126 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
\r
127 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
\r
128 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
\r
129 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
\r
130 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
\r
131 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
\r
132 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
\r
133 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
\r
134 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
\r
138 tmp10 = tmp0 + tmp3; /* phase 2 */
\r
139 tmp13 = tmp0 - tmp3;
\r
140 tmp11 = tmp1 + tmp2;
\r
141 tmp12 = tmp1 - tmp2;
\r
143 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
\r
144 dataptr[DCTSIZE*4] = tmp10 - tmp11;
\r
146 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
\r
147 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
\r
148 dataptr[DCTSIZE*6] = tmp13 - z1;
\r
152 tmp10 = tmp4 + tmp5; /* phase 2 */
\r
153 tmp11 = tmp5 + tmp6;
\r
154 tmp12 = tmp6 + tmp7;
\r
156 /* The rotator is modified from fig 4-8 to avoid extra negations. */
\r
157 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
\r
158 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
\r
159 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
\r
160 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
\r
162 z11 = tmp7 + z3; /* phase 5 */
\r
165 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
\r
166 dataptr[DCTSIZE*3] = z13 - z2;
\r
167 dataptr[DCTSIZE*1] = z11 + z4;
\r
168 dataptr[DCTSIZE*7] = z11 - z4;
\r
170 dataptr++; /* advance pointer to next column */
\r
174 #endif /* DCT_FLOAT_SUPPORTED */
\r