/src/freeimage-svn/FreeImage/trunk/Source/LibJPEG/jidctflt.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * jidctflt.c |
3 | | * |
4 | | * Copyright (C) 1994-1998, Thomas G. Lane. |
5 | | * Modified 2010-2017 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 | | * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine |
11 | | * must also perform dequantization of the input coefficients. |
12 | | * |
13 | | * This implementation should be more accurate than either of the integer |
14 | | * IDCT implementations. However, it may not give the same results on all |
15 | | * machines because of differences in roundoff behavior. Speed will depend |
16 | | * on the hardware's floating point capacity. |
17 | | * |
18 | | * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT |
19 | | * on each row (or vice versa, but it's more convenient to emit a row at |
20 | | * a time). Direct algorithms are also available, but they are much more |
21 | | * complex and seem not to be any faster when reduced to code. |
22 | | * |
23 | | * This implementation is based on Arai, Agui, and Nakajima's algorithm for |
24 | | * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in |
25 | | * Japanese, but the algorithm is described in the Pennebaker & Mitchell |
26 | | * JPEG textbook (see REFERENCES section in file README). The following code |
27 | | * is based directly on figure 4-8 in P&M. |
28 | | * While an 8-point DCT cannot be done in less than 11 multiplies, it is |
29 | | * possible to arrange the computation so that many of the multiplies are |
30 | | * simple scalings of the final outputs. These multiplies can then be |
31 | | * folded into the multiplications or divisions by the JPEG quantization |
32 | | * table entries. The AA&N method leaves only 5 multiplies and 29 adds |
33 | | * to be done in the DCT itself. |
34 | | * The primary disadvantage of this method is that with a fixed-point |
35 | | * implementation, accuracy is lost due to imprecise representation of the |
36 | | * scaled quantization values. However, that problem does not arise if |
37 | | * we use floating point arithmetic. |
38 | | */ |
39 | | |
40 | | #define JPEG_INTERNALS |
41 | | #include "jinclude.h" |
42 | | #include "jpeglib.h" |
43 | | #include "jdct.h" /* Private declarations for DCT subsystem */ |
44 | | |
45 | | #ifdef DCT_FLOAT_SUPPORTED |
46 | | |
47 | | |
48 | | /* |
49 | | * This module is specialized to the case DCTSIZE = 8. |
50 | | */ |
51 | | |
52 | | #if DCTSIZE != 8 |
53 | | Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */ |
54 | | #endif |
55 | | |
56 | | |
57 | | /* Dequantize a coefficient by multiplying it by the multiplier-table |
58 | | * entry; produce a float result. |
59 | | */ |
60 | | |
61 | 0 | #define DEQUANTIZE(coef,quantval) (((FAST_FLOAT) (coef)) * (quantval)) |
62 | | |
63 | | |
64 | | /* |
65 | | * Perform dequantization and inverse DCT on one block of coefficients. |
66 | | * |
67 | | * cK represents cos(K*pi/16). |
68 | | */ |
69 | | |
70 | | GLOBAL(void) |
71 | | jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
72 | | JCOEFPTR coef_block, |
73 | | JSAMPARRAY output_buf, JDIMENSION output_col) |
74 | 0 | { |
75 | 0 | FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
76 | 0 | FAST_FLOAT tmp10, tmp11, tmp12, tmp13; |
77 | 0 | FAST_FLOAT z5, z10, z11, z12, z13; |
78 | 0 | JCOEFPTR inptr; |
79 | 0 | FLOAT_MULT_TYPE * quantptr; |
80 | 0 | FAST_FLOAT * wsptr; |
81 | 0 | JSAMPROW outptr; |
82 | 0 | JSAMPLE *range_limit = IDCT_range_limit(cinfo); |
83 | 0 | int ctr; |
84 | 0 | FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */ |
85 | | |
86 | | /* Pass 1: process columns from input, store into work array. */ |
87 | |
|
88 | 0 | inptr = coef_block; |
89 | 0 | quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table; |
90 | 0 | wsptr = workspace; |
91 | 0 | for (ctr = DCTSIZE; ctr > 0; ctr--) { |
92 | | /* Due to quantization, we will usually find that many of the input |
93 | | * coefficients are zero, especially the AC terms. We can exploit this |
94 | | * by short-circuiting the IDCT calculation for any column in which all |
95 | | * the AC terms are zero. In that case each output is equal to the |
96 | | * DC coefficient (with scale factor as needed). |
97 | | * With typical images and quantization tables, half or more of the |
98 | | * column DCT calculations can be simplified this way. |
99 | | */ |
100 | |
|
101 | 0 | if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 && |
102 | 0 | inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 && |
103 | 0 | inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 && |
104 | 0 | inptr[DCTSIZE*7] == 0) { |
105 | | /* AC terms all zero */ |
106 | 0 | FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); |
107 | |
|
108 | 0 | wsptr[DCTSIZE*0] = dcval; |
109 | 0 | wsptr[DCTSIZE*1] = dcval; |
110 | 0 | wsptr[DCTSIZE*2] = dcval; |
111 | 0 | wsptr[DCTSIZE*3] = dcval; |
112 | 0 | wsptr[DCTSIZE*4] = dcval; |
113 | 0 | wsptr[DCTSIZE*5] = dcval; |
114 | 0 | wsptr[DCTSIZE*6] = dcval; |
115 | 0 | wsptr[DCTSIZE*7] = dcval; |
116 | |
|
117 | 0 | inptr++; /* advance pointers to next column */ |
118 | 0 | quantptr++; |
119 | 0 | wsptr++; |
120 | 0 | continue; |
121 | 0 | } |
122 | | |
123 | | /* Even part */ |
124 | | |
125 | 0 | tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); |
126 | 0 | tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]); |
127 | 0 | tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]); |
128 | 0 | tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]); |
129 | |
|
130 | 0 | tmp10 = tmp0 + tmp2; /* phase 3 */ |
131 | 0 | tmp11 = tmp0 - tmp2; |
132 | |
|
133 | 0 | tmp13 = tmp1 + tmp3; /* phases 5-3 */ |
134 | 0 | tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */ |
135 | |
|
136 | 0 | tmp0 = tmp10 + tmp13; /* phase 2 */ |
137 | 0 | tmp3 = tmp10 - tmp13; |
138 | 0 | tmp1 = tmp11 + tmp12; |
139 | 0 | tmp2 = tmp11 - tmp12; |
140 | | |
141 | | /* Odd part */ |
142 | |
|
143 | 0 | tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]); |
144 | 0 | tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]); |
145 | 0 | tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]); |
146 | 0 | tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]); |
147 | |
|
148 | 0 | z13 = tmp6 + tmp5; /* phase 6 */ |
149 | 0 | z10 = tmp6 - tmp5; |
150 | 0 | z11 = tmp4 + tmp7; |
151 | 0 | z12 = tmp4 - tmp7; |
152 | |
|
153 | 0 | tmp7 = z11 + z13; /* phase 5 */ |
154 | 0 | tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */ |
155 | |
|
156 | 0 | z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */ |
157 | 0 | tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */ |
158 | 0 | tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */ |
159 | |
|
160 | 0 | tmp6 = tmp12 - tmp7; /* phase 2 */ |
161 | 0 | tmp5 = tmp11 - tmp6; |
162 | 0 | tmp4 = tmp10 - tmp5; |
163 | |
|
164 | 0 | wsptr[DCTSIZE*0] = tmp0 + tmp7; |
165 | 0 | wsptr[DCTSIZE*7] = tmp0 - tmp7; |
166 | 0 | wsptr[DCTSIZE*1] = tmp1 + tmp6; |
167 | 0 | wsptr[DCTSIZE*6] = tmp1 - tmp6; |
168 | 0 | wsptr[DCTSIZE*2] = tmp2 + tmp5; |
169 | 0 | wsptr[DCTSIZE*5] = tmp2 - tmp5; |
170 | 0 | wsptr[DCTSIZE*3] = tmp3 + tmp4; |
171 | 0 | wsptr[DCTSIZE*4] = tmp3 - tmp4; |
172 | |
|
173 | 0 | inptr++; /* advance pointers to next column */ |
174 | 0 | quantptr++; |
175 | 0 | wsptr++; |
176 | 0 | } |
177 | | |
178 | | /* Pass 2: process rows from work array, store into output array. */ |
179 | |
|
180 | 0 | wsptr = workspace; |
181 | 0 | for (ctr = 0; ctr < DCTSIZE; ctr++) { |
182 | 0 | outptr = output_buf[ctr] + output_col; |
183 | | /* Rows of zeroes can be exploited in the same way as we did with columns. |
184 | | * However, the column calculation has created many nonzero AC terms, so |
185 | | * the simplification applies less often (typically 5% to 10% of the time). |
186 | | * And testing floats for zero is relatively expensive, so we don't bother. |
187 | | */ |
188 | | |
189 | | /* Even part */ |
190 | | |
191 | | /* Prepare range-limit and float->int conversion */ |
192 | 0 | z5 = wsptr[0] + (((FAST_FLOAT) RANGE_CENTER) + ((FAST_FLOAT) 0.5)); |
193 | 0 | tmp10 = z5 + wsptr[4]; |
194 | 0 | tmp11 = z5 - wsptr[4]; |
195 | |
|
196 | 0 | tmp13 = wsptr[2] + wsptr[6]; |
197 | 0 | tmp12 = (wsptr[2] - wsptr[6]) * |
198 | 0 | ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */ |
199 | |
|
200 | 0 | tmp0 = tmp10 + tmp13; |
201 | 0 | tmp3 = tmp10 - tmp13; |
202 | 0 | tmp1 = tmp11 + tmp12; |
203 | 0 | tmp2 = tmp11 - tmp12; |
204 | | |
205 | | /* Odd part */ |
206 | |
|
207 | 0 | z13 = wsptr[5] + wsptr[3]; |
208 | 0 | z10 = wsptr[5] - wsptr[3]; |
209 | 0 | z11 = wsptr[1] + wsptr[7]; |
210 | 0 | z12 = wsptr[1] - wsptr[7]; |
211 | |
|
212 | 0 | tmp7 = z11 + z13; /* phase 5 */ |
213 | 0 | tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */ |
214 | |
|
215 | 0 | z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */ |
216 | 0 | tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */ |
217 | 0 | tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */ |
218 | |
|
219 | 0 | tmp6 = tmp12 - tmp7; /* phase 2 */ |
220 | 0 | tmp5 = tmp11 - tmp6; |
221 | 0 | tmp4 = tmp10 - tmp5; |
222 | | |
223 | | /* Final output stage: float->int conversion and range-limit */ |
224 | |
|
225 | 0 | outptr[0] = range_limit[(int) (tmp0 + tmp7) & RANGE_MASK]; |
226 | 0 | outptr[7] = range_limit[(int) (tmp0 - tmp7) & RANGE_MASK]; |
227 | 0 | outptr[1] = range_limit[(int) (tmp1 + tmp6) & RANGE_MASK]; |
228 | 0 | outptr[6] = range_limit[(int) (tmp1 - tmp6) & RANGE_MASK]; |
229 | 0 | outptr[2] = range_limit[(int) (tmp2 + tmp5) & RANGE_MASK]; |
230 | 0 | outptr[5] = range_limit[(int) (tmp2 - tmp5) & RANGE_MASK]; |
231 | 0 | outptr[3] = range_limit[(int) (tmp3 + tmp4) & RANGE_MASK]; |
232 | 0 | outptr[4] = range_limit[(int) (tmp3 - tmp4) & RANGE_MASK]; |
233 | |
|
234 | 0 | wsptr += DCTSIZE; /* advance pointer to next row */ |
235 | 0 | } |
236 | 0 | } |
237 | | |
238 | | #endif /* DCT_FLOAT_SUPPORTED */ |