1 | /* |
2 | * jfdctflt.c |
3 | * |
4 | * Copyright (C) 1994-1996, Thomas G. Lane. |
5 | * Modified 2003-2009 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 DCTs. /* deliberate syntax err */ |
52 | #endif |
53 | |
54 | |
55 | /* |
56 | * Perform the forward DCT on one block of samples. |
57 | */ |
58 | |
59 | GLOBAL(void) |
60 | jpeg_fdct_float (FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col) |
61 | { |
62 | FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
63 | FAST_FLOAT tmp10, tmp11, tmp12, tmp13; |
64 | FAST_FLOAT z1, z2, z3, z4, z5, z11, z13; |
65 | FAST_FLOAT *dataptr; |
66 | JSAMPROW elemptr; |
67 | int ctr; |
68 | |
69 | /* Pass 1: process rows. */ |
70 | |
71 | dataptr = data; |
72 | for (ctr = 0; ctr < DCTSIZE; ctr++) { |
73 | elemptr = sample_data[ctr] + start_col; |
74 | |
75 | /* Load data into workspace */ |
76 | tmp0 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7])); |
77 | tmp7 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7])); |
78 | tmp1 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6])); |
79 | tmp6 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6])); |
80 | tmp2 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5])); |
81 | tmp5 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5])); |
82 | tmp3 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4])); |
83 | tmp4 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4])); |
84 | |
85 | /* Even part */ |
86 | |
87 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
88 | tmp13 = tmp0 - tmp3; |
89 | tmp11 = tmp1 + tmp2; |
90 | tmp12 = tmp1 - tmp2; |
91 | |
92 | /* Apply unsigned->signed conversion */ |
93 | dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */ |
94 | dataptr[4] = tmp10 - tmp11; |
95 | |
96 | z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */ |
97 | dataptr[2] = tmp13 + z1; /* phase 5 */ |
98 | dataptr[6] = tmp13 - z1; |
99 | |
100 | /* Odd part */ |
101 | |
102 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
103 | tmp11 = tmp5 + tmp6; |
104 | tmp12 = tmp6 + tmp7; |
105 | |
106 | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
107 | z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */ |
108 | z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */ |
109 | z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */ |
110 | z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */ |
111 | |
112 | z11 = tmp7 + z3; /* phase 5 */ |
113 | z13 = tmp7 - z3; |
114 | |
115 | dataptr[5] = z13 + z2; /* phase 6 */ |
116 | dataptr[3] = z13 - z2; |
117 | dataptr[1] = z11 + z4; |
118 | dataptr[7] = z11 - z4; |
119 | |
120 | dataptr += DCTSIZE; /* advance pointer to next row */ |
121 | } |
122 | |
123 | /* Pass 2: process columns. */ |
124 | |
125 | dataptr = data; |
126 | for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { |
127 | tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; |
128 | tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; |
129 | tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; |
130 | tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; |
131 | tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; |
132 | tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; |
133 | tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; |
134 | tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; |
135 | |
136 | /* Even part */ |
137 | |
138 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
139 | tmp13 = tmp0 - tmp3; |
140 | tmp11 = tmp1 + tmp2; |
141 | tmp12 = tmp1 - tmp2; |
142 | |
143 | dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ |
144 | dataptr[DCTSIZE*4] = tmp10 - tmp11; |
145 | |
146 | z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */ |
147 | dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ |
148 | dataptr[DCTSIZE*6] = tmp13 - z1; |
149 | |
150 | /* Odd part */ |
151 | |
152 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
153 | tmp11 = tmp5 + tmp6; |
154 | tmp12 = tmp6 + tmp7; |
155 | |
156 | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
157 | z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */ |
158 | z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */ |
159 | z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */ |
160 | z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */ |
161 | |
162 | z11 = tmp7 + z3; /* phase 5 */ |
163 | z13 = tmp7 - z3; |
164 | |
165 | dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ |
166 | dataptr[DCTSIZE*3] = z13 - z2; |
167 | dataptr[DCTSIZE*1] = z11 + z4; |
168 | dataptr[DCTSIZE*7] = z11 - z4; |
169 | |
170 | dataptr++; /* advance pointer to next column */ |
171 | } |
172 | } |
173 | |
174 | #endif /* DCT_FLOAT_SUPPORTED */ |
175 | |