4 * Copyright (C) 1994-1998, Thomas G. Lane.
\r
5 * Modified 2010 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 * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
\r
11 * must also perform dequantization of the input coefficients.
\r
13 * This implementation should be more accurate than either of the integer
\r
14 * IDCT implementations. However, it may not give the same results on all
\r
15 * machines because of differences in roundoff behavior. Speed will depend
\r
16 * on the hardware's floating point capacity.
\r
18 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
\r
19 * on each row (or vice versa, but it's more convenient to emit a row at
\r
20 * a time). Direct algorithms are also available, but they are much more
\r
21 * complex and seem not to be any faster when reduced to code.
\r
23 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
\r
24 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
\r
25 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
\r
26 * JPEG textbook (see REFERENCES section in file README). The following code
\r
27 * is based directly on figure 4-8 in P&M.
\r
28 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
\r
29 * possible to arrange the computation so that many of the multiplies are
\r
30 * simple scalings of the final outputs. These multiplies can then be
\r
31 * folded into the multiplications or divisions by the JPEG quantization
\r
32 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
\r
33 * to be done in the DCT itself.
\r
34 * The primary disadvantage of this method is that with a fixed-point
\r
35 * implementation, accuracy is lost due to imprecise representation of the
\r
36 * scaled quantization values. However, that problem does not arise if
\r
37 * we use floating point arithmetic.
\r
40 #define JPEG_INTERNALS
\r
41 #include "jinclude.h"
\r
42 #include "jpeglib.h"
\r
43 #include "jdct.h" /* Private declarations for DCT subsystem */
\r
45 #ifdef DCT_FLOAT_SUPPORTED
\r
49 * This module is specialized to the case DCTSIZE = 8.
\r
53 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
\r
57 /* Dequantize a coefficient by multiplying it by the multiplier-table
\r
58 * entry; produce a float result.
\r
61 #define DEQUANTIZE(coef,quantval) (((FAST_FLOAT) (coef)) * (quantval))
\r
65 * Perform dequantization and inverse DCT on one block of coefficients.
\r
69 jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
\r
70 JCOEFPTR coef_block,
\r
71 JSAMPARRAY output_buf, JDIMENSION output_col)
\r
73 FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
\r
74 FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
\r
75 FAST_FLOAT z5, z10, z11, z12, z13;
\r
77 FLOAT_MULT_TYPE * quantptr;
\r
80 JSAMPLE *range_limit = cinfo->sample_range_limit;
\r
82 FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
\r
84 /* Pass 1: process columns from input, store into work array. */
\r
87 quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
\r
89 for (ctr = DCTSIZE; ctr > 0; ctr--) {
\r
90 /* Due to quantization, we will usually find that many of the input
\r
91 * coefficients are zero, especially the AC terms. We can exploit this
\r
92 * by short-circuiting the IDCT calculation for any column in which all
\r
93 * the AC terms are zero. In that case each output is equal to the
\r
94 * DC coefficient (with scale factor as needed).
\r
95 * With typical images and quantization tables, half or more of the
\r
96 * column DCT calculations can be simplified this way.
\r
99 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
\r
100 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
\r
101 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
\r
102 inptr[DCTSIZE*7] == 0) {
\r
103 /* AC terms all zero */
\r
104 FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
\r
106 wsptr[DCTSIZE*0] = dcval;
\r
107 wsptr[DCTSIZE*1] = dcval;
\r
108 wsptr[DCTSIZE*2] = dcval;
\r
109 wsptr[DCTSIZE*3] = dcval;
\r
110 wsptr[DCTSIZE*4] = dcval;
\r
111 wsptr[DCTSIZE*5] = dcval;
\r
112 wsptr[DCTSIZE*6] = dcval;
\r
113 wsptr[DCTSIZE*7] = dcval;
\r
115 inptr++; /* advance pointers to next column */
\r
123 tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
\r
124 tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
\r
125 tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
\r
126 tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
\r
128 tmp10 = tmp0 + tmp2; /* phase 3 */
\r
129 tmp11 = tmp0 - tmp2;
\r
131 tmp13 = tmp1 + tmp3; /* phases 5-3 */
\r
132 tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
\r
134 tmp0 = tmp10 + tmp13; /* phase 2 */
\r
135 tmp3 = tmp10 - tmp13;
\r
136 tmp1 = tmp11 + tmp12;
\r
137 tmp2 = tmp11 - tmp12;
\r
141 tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
\r
142 tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
\r
143 tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
\r
144 tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
\r
146 z13 = tmp6 + tmp5; /* phase 6 */
\r
151 tmp7 = z11 + z13; /* phase 5 */
\r
152 tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
\r
154 z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
\r
155 tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
\r
156 tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
\r
158 tmp6 = tmp12 - tmp7; /* phase 2 */
\r
159 tmp5 = tmp11 - tmp6;
\r
160 tmp4 = tmp10 - tmp5;
\r
162 wsptr[DCTSIZE*0] = tmp0 + tmp7;
\r
163 wsptr[DCTSIZE*7] = tmp0 - tmp7;
\r
164 wsptr[DCTSIZE*1] = tmp1 + tmp6;
\r
165 wsptr[DCTSIZE*6] = tmp1 - tmp6;
\r
166 wsptr[DCTSIZE*2] = tmp2 + tmp5;
\r
167 wsptr[DCTSIZE*5] = tmp2 - tmp5;
\r
168 wsptr[DCTSIZE*3] = tmp3 + tmp4;
\r
169 wsptr[DCTSIZE*4] = tmp3 - tmp4;
\r
171 inptr++; /* advance pointers to next column */
\r
176 /* Pass 2: process rows from work array, store into output array. */
\r
179 for (ctr = 0; ctr < DCTSIZE; ctr++) {
\r
180 outptr = output_buf[ctr] + output_col;
\r
181 /* Rows of zeroes can be exploited in the same way as we did with columns.
\r
182 * However, the column calculation has created many nonzero AC terms, so
\r
183 * the simplification applies less often (typically 5% to 10% of the time).
\r
184 * And testing floats for zero is relatively expensive, so we don't bother.
\r
189 /* Apply signed->unsigned and prepare float->int conversion */
\r
190 z5 = wsptr[0] + ((FAST_FLOAT) CENTERJSAMPLE + (FAST_FLOAT) 0.5);
\r
191 tmp10 = z5 + wsptr[4];
\r
192 tmp11 = z5 - wsptr[4];
\r
194 tmp13 = wsptr[2] + wsptr[6];
\r
195 tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
\r
197 tmp0 = tmp10 + tmp13;
\r
198 tmp3 = tmp10 - tmp13;
\r
199 tmp1 = tmp11 + tmp12;
\r
200 tmp2 = tmp11 - tmp12;
\r
204 z13 = wsptr[5] + wsptr[3];
\r
205 z10 = wsptr[5] - wsptr[3];
\r
206 z11 = wsptr[1] + wsptr[7];
\r
207 z12 = wsptr[1] - wsptr[7];
\r
210 tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
\r
212 z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
\r
213 tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
\r
214 tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
\r
216 tmp6 = tmp12 - tmp7;
\r
217 tmp5 = tmp11 - tmp6;
\r
218 tmp4 = tmp10 - tmp5;
\r
220 /* Final output stage: float->int conversion and range-limit */
\r
222 outptr[0] = range_limit[((int) (tmp0 + tmp7)) & RANGE_MASK];
\r
223 outptr[7] = range_limit[((int) (tmp0 - tmp7)) & RANGE_MASK];
\r
224 outptr[1] = range_limit[((int) (tmp1 + tmp6)) & RANGE_MASK];
\r
225 outptr[6] = range_limit[((int) (tmp1 - tmp6)) & RANGE_MASK];
\r
226 outptr[2] = range_limit[((int) (tmp2 + tmp5)) & RANGE_MASK];
\r
227 outptr[5] = range_limit[((int) (tmp2 - tmp5)) & RANGE_MASK];
\r
228 outptr[3] = range_limit[((int) (tmp3 + tmp4)) & RANGE_MASK];
\r
229 outptr[4] = range_limit[((int) (tmp3 - tmp4)) & RANGE_MASK];
\r
231 wsptr += DCTSIZE; /* advance pointer to next row */
\r
235 #endif /* DCT_FLOAT_SUPPORTED */
\r