/src/freeimage-svn/FreeImage/trunk/Source/LibJPEG/jfdctfst.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * jfdctfst.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 fast, not so accurate integer implementation of the |
10 | | * forward DCT (Discrete Cosine Transform). |
11 | | * |
12 | | * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT |
13 | | * on each column. Direct algorithms are also available, but they are |
14 | | * much more complex and seem not to be any faster when reduced to code. |
15 | | * |
16 | | * This implementation is based on Arai, Agui, and Nakajima's algorithm for |
17 | | * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in |
18 | | * Japanese, but the algorithm is described in the Pennebaker & Mitchell |
19 | | * JPEG textbook (see REFERENCES section in file README). The following code |
20 | | * is based directly on figure 4-8 in P&M. |
21 | | * While an 8-point DCT cannot be done in less than 11 multiplies, it is |
22 | | * possible to arrange the computation so that many of the multiplies are |
23 | | * simple scalings of the final outputs. These multiplies can then be |
24 | | * folded into the multiplications or divisions by the JPEG quantization |
25 | | * table entries. The AA&N method leaves only 5 multiplies and 29 adds |
26 | | * to be done in the DCT itself. |
27 | | * The primary disadvantage of this method is that with fixed-point math, |
28 | | * accuracy is lost due to imprecise representation of the scaled |
29 | | * quantization values. The smaller the quantization table entry, the less |
30 | | * precise the scaled value, so this implementation does worse with high- |
31 | | * quality-setting files than with low-quality ones. |
32 | | */ |
33 | | |
34 | | #define JPEG_INTERNALS |
35 | | #include "jinclude.h" |
36 | | #include "jpeglib.h" |
37 | | #include "jdct.h" /* Private declarations for DCT subsystem */ |
38 | | |
39 | | #ifdef DCT_IFAST_SUPPORTED |
40 | | |
41 | | |
42 | | /* |
43 | | * This module is specialized to the case DCTSIZE = 8. |
44 | | */ |
45 | | |
46 | | #if DCTSIZE != 8 |
47 | | Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */ |
48 | | #endif |
49 | | |
50 | | |
51 | | /* Scaling decisions are generally the same as in the LL&M algorithm; |
52 | | * see jfdctint.c for more details. However, we choose to descale |
53 | | * (right shift) multiplication products as soon as they are formed, |
54 | | * rather than carrying additional fractional bits into subsequent additions. |
55 | | * This compromises accuracy slightly, but it lets us save a few shifts. |
56 | | * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) |
57 | | * everywhere except in the multiplications proper; this saves a good deal |
58 | | * of work on 16-bit-int machines. |
59 | | * |
60 | | * Again to save a few shifts, the intermediate results between pass 1 and |
61 | | * pass 2 are not upscaled, but are represented only to integral precision. |
62 | | * |
63 | | * A final compromise is to represent the multiplicative constants to only |
64 | | * 8 fractional bits, rather than 13. This saves some shifting work on some |
65 | | * machines, and may also reduce the cost of multiplication (since there |
66 | | * are fewer one-bits in the constants). |
67 | | */ |
68 | | |
69 | | #define CONST_BITS 8 |
70 | | |
71 | | |
72 | | /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus |
73 | | * causing a lot of useless floating-point operations at run time. |
74 | | * To get around this we use the following pre-calculated constants. |
75 | | * If you change CONST_BITS you may want to add appropriate values. |
76 | | * (With a reasonable C compiler, you can just rely on the FIX() macro...) |
77 | | */ |
78 | | |
79 | | #if CONST_BITS == 8 |
80 | | #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */ |
81 | | #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */ |
82 | | #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */ |
83 | | #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */ |
84 | | #else |
85 | | #define FIX_0_382683433 FIX(0.382683433) |
86 | | #define FIX_0_541196100 FIX(0.541196100) |
87 | | #define FIX_0_707106781 FIX(0.707106781) |
88 | | #define FIX_1_306562965 FIX(1.306562965) |
89 | | #endif |
90 | | |
91 | | |
92 | | /* We can gain a little more speed, with a further compromise in accuracy, |
93 | | * by omitting the addition in a descaling shift. This yields an incorrectly |
94 | | * rounded result half the time... |
95 | | */ |
96 | | |
97 | | #ifndef USE_ACCURATE_ROUNDING |
98 | | #undef DESCALE |
99 | 0 | #define DESCALE(x,n) RIGHT_SHIFT(x, n) |
100 | | #endif |
101 | | |
102 | | |
103 | | /* Multiply a DCTELEM variable by an INT32 constant, and immediately |
104 | | * descale to yield a DCTELEM result. |
105 | | */ |
106 | | |
107 | 0 | #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) |
108 | | |
109 | | |
110 | | /* |
111 | | * Perform the forward DCT on one block of samples. |
112 | | * |
113 | | * cK represents cos(K*pi/16). |
114 | | */ |
115 | | |
116 | | GLOBAL(void) |
117 | | jpeg_fdct_ifast (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col) |
118 | 0 | { |
119 | 0 | DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
120 | 0 | DCTELEM tmp10, tmp11, tmp12, tmp13; |
121 | 0 | DCTELEM z1, z2, z3, z4, z5, z11, z13; |
122 | 0 | DCTELEM *dataptr; |
123 | 0 | JSAMPROW elemptr; |
124 | 0 | int ctr; |
125 | 0 | SHIFT_TEMPS |
126 | | |
127 | | /* Pass 1: process rows. */ |
128 | |
|
129 | 0 | dataptr = data; |
130 | 0 | for (ctr = 0; ctr < DCTSIZE; ctr++) { |
131 | 0 | elemptr = sample_data[ctr] + start_col; |
132 | | |
133 | | /* Load data into workspace */ |
134 | 0 | tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]); |
135 | 0 | tmp7 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]); |
136 | 0 | tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]); |
137 | 0 | tmp6 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]); |
138 | 0 | tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]); |
139 | 0 | tmp5 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]); |
140 | 0 | tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]); |
141 | 0 | tmp4 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]); |
142 | | |
143 | | /* Even part */ |
144 | |
|
145 | 0 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
146 | 0 | tmp13 = tmp0 - tmp3; |
147 | 0 | tmp11 = tmp1 + tmp2; |
148 | 0 | tmp12 = tmp1 - tmp2; |
149 | | |
150 | | /* Apply unsigned->signed conversion. */ |
151 | 0 | dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */ |
152 | 0 | dataptr[4] = tmp10 - tmp11; |
153 | |
|
154 | 0 | z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ |
155 | 0 | dataptr[2] = tmp13 + z1; /* phase 5 */ |
156 | 0 | dataptr[6] = tmp13 - z1; |
157 | | |
158 | | /* Odd part */ |
159 | |
|
160 | 0 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
161 | 0 | tmp11 = tmp5 + tmp6; |
162 | 0 | tmp12 = tmp6 + tmp7; |
163 | | |
164 | | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
165 | 0 | z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ |
166 | 0 | z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ |
167 | 0 | z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ |
168 | 0 | z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ |
169 | |
|
170 | 0 | z11 = tmp7 + z3; /* phase 5 */ |
171 | 0 | z13 = tmp7 - z3; |
172 | |
|
173 | 0 | dataptr[5] = z13 + z2; /* phase 6 */ |
174 | 0 | dataptr[3] = z13 - z2; |
175 | 0 | dataptr[1] = z11 + z4; |
176 | 0 | dataptr[7] = z11 - z4; |
177 | |
|
178 | 0 | dataptr += DCTSIZE; /* advance pointer to next row */ |
179 | 0 | } |
180 | | |
181 | | /* Pass 2: process columns. */ |
182 | |
|
183 | 0 | dataptr = data; |
184 | 0 | for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { |
185 | 0 | tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; |
186 | 0 | tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; |
187 | 0 | tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; |
188 | 0 | tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; |
189 | 0 | tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; |
190 | 0 | tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; |
191 | 0 | tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; |
192 | 0 | tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; |
193 | | |
194 | | /* Even part */ |
195 | |
|
196 | 0 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
197 | 0 | tmp13 = tmp0 - tmp3; |
198 | 0 | tmp11 = tmp1 + tmp2; |
199 | 0 | tmp12 = tmp1 - tmp2; |
200 | |
|
201 | 0 | dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ |
202 | 0 | dataptr[DCTSIZE*4] = tmp10 - tmp11; |
203 | |
|
204 | 0 | z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ |
205 | 0 | dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ |
206 | 0 | dataptr[DCTSIZE*6] = tmp13 - z1; |
207 | | |
208 | | /* Odd part */ |
209 | |
|
210 | 0 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
211 | 0 | tmp11 = tmp5 + tmp6; |
212 | 0 | tmp12 = tmp6 + tmp7; |
213 | | |
214 | | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
215 | 0 | z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ |
216 | 0 | z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ |
217 | 0 | z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ |
218 | 0 | z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ |
219 | |
|
220 | 0 | z11 = tmp7 + z3; /* phase 5 */ |
221 | 0 | z13 = tmp7 - z3; |
222 | |
|
223 | 0 | dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ |
224 | 0 | dataptr[DCTSIZE*3] = z13 - z2; |
225 | 0 | dataptr[DCTSIZE*1] = z11 + z4; |
226 | 0 | dataptr[DCTSIZE*7] = z11 - z4; |
227 | |
|
228 | 0 | dataptr++; /* advance pointer to next column */ |
229 | 0 | } |
230 | 0 | } |
231 | | |
232 | | #endif /* DCT_IFAST_SUPPORTED */ |