rev |
line source |
nuclear@0
|
1 /*
|
nuclear@0
|
2 * jidctfst.c
|
nuclear@0
|
3 *
|
nuclear@0
|
4 * Copyright (C) 1994-1998, Thomas G. Lane.
|
nuclear@0
|
5 * This file is part of the Independent JPEG Group's software.
|
nuclear@0
|
6 * For conditions of distribution and use, see the accompanying README file.
|
nuclear@0
|
7 *
|
nuclear@0
|
8 * This file contains a fast, not so accurate integer implementation of the
|
nuclear@0
|
9 * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
|
nuclear@0
|
10 * must also perform dequantization of the input coefficients.
|
nuclear@0
|
11 *
|
nuclear@0
|
12 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
|
nuclear@0
|
13 * on each row (or vice versa, but it's more convenient to emit a row at
|
nuclear@0
|
14 * a time). Direct algorithms are also available, but they are much more
|
nuclear@0
|
15 * complex and seem not to be any faster when reduced to code.
|
nuclear@0
|
16 *
|
nuclear@0
|
17 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
|
nuclear@0
|
18 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
|
nuclear@0
|
19 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
|
nuclear@0
|
20 * JPEG textbook (see REFERENCES section in file README). The following code
|
nuclear@0
|
21 * is based directly on figure 4-8 in P&M.
|
nuclear@0
|
22 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
|
nuclear@0
|
23 * possible to arrange the computation so that many of the multiplies are
|
nuclear@0
|
24 * simple scalings of the final outputs. These multiplies can then be
|
nuclear@0
|
25 * folded into the multiplications or divisions by the JPEG quantization
|
nuclear@0
|
26 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
|
nuclear@0
|
27 * to be done in the DCT itself.
|
nuclear@0
|
28 * The primary disadvantage of this method is that with fixed-point math,
|
nuclear@0
|
29 * accuracy is lost due to imprecise representation of the scaled
|
nuclear@0
|
30 * quantization values. The smaller the quantization table entry, the less
|
nuclear@0
|
31 * precise the scaled value, so this implementation does worse with high-
|
nuclear@0
|
32 * quality-setting files than with low-quality ones.
|
nuclear@0
|
33 */
|
nuclear@0
|
34
|
nuclear@0
|
35 #define JPEG_INTERNALS
|
nuclear@0
|
36 #include "jinclude.h"
|
nuclear@0
|
37 #include "jpeglib.h"
|
nuclear@0
|
38 #include "jdct.h" /* Private declarations for DCT subsystem */
|
nuclear@0
|
39
|
nuclear@0
|
40 #ifdef DCT_IFAST_SUPPORTED
|
nuclear@0
|
41
|
nuclear@0
|
42
|
nuclear@0
|
43 /*
|
nuclear@0
|
44 * This module is specialized to the case DCTSIZE = 8.
|
nuclear@0
|
45 */
|
nuclear@0
|
46
|
nuclear@0
|
47 #if DCTSIZE != 8
|
nuclear@0
|
48 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
|
nuclear@0
|
49 #endif
|
nuclear@0
|
50
|
nuclear@0
|
51
|
nuclear@0
|
52 /* Scaling decisions are generally the same as in the LL&M algorithm;
|
nuclear@0
|
53 * see jidctint.c for more details. However, we choose to descale
|
nuclear@0
|
54 * (right shift) multiplication products as soon as they are formed,
|
nuclear@0
|
55 * rather than carrying additional fractional bits into subsequent additions.
|
nuclear@0
|
56 * This compromises accuracy slightly, but it lets us save a few shifts.
|
nuclear@0
|
57 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
|
nuclear@0
|
58 * everywhere except in the multiplications proper; this saves a good deal
|
nuclear@0
|
59 * of work on 16-bit-int machines.
|
nuclear@0
|
60 *
|
nuclear@0
|
61 * The dequantized coefficients are not integers because the AA&N scaling
|
nuclear@0
|
62 * factors have been incorporated. We represent them scaled up by PASS1_BITS,
|
nuclear@0
|
63 * so that the first and second IDCT rounds have the same input scaling.
|
nuclear@0
|
64 * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
|
nuclear@0
|
65 * avoid a descaling shift; this compromises accuracy rather drastically
|
nuclear@0
|
66 * for small quantization table entries, but it saves a lot of shifts.
|
nuclear@0
|
67 * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
|
nuclear@0
|
68 * so we use a much larger scaling factor to preserve accuracy.
|
nuclear@0
|
69 *
|
nuclear@0
|
70 * A final compromise is to represent the multiplicative constants to only
|
nuclear@0
|
71 * 8 fractional bits, rather than 13. This saves some shifting work on some
|
nuclear@0
|
72 * machines, and may also reduce the cost of multiplication (since there
|
nuclear@0
|
73 * are fewer one-bits in the constants).
|
nuclear@0
|
74 */
|
nuclear@0
|
75
|
nuclear@0
|
76 #if BITS_IN_JSAMPLE == 8
|
nuclear@0
|
77 #define CONST_BITS 8
|
nuclear@0
|
78 #define PASS1_BITS 2
|
nuclear@0
|
79 #else
|
nuclear@0
|
80 #define CONST_BITS 8
|
nuclear@0
|
81 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */
|
nuclear@0
|
82 #endif
|
nuclear@0
|
83
|
nuclear@0
|
84 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
|
nuclear@0
|
85 * causing a lot of useless floating-point operations at run time.
|
nuclear@0
|
86 * To get around this we use the following pre-calculated constants.
|
nuclear@0
|
87 * If you change CONST_BITS you may want to add appropriate values.
|
nuclear@0
|
88 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
|
nuclear@0
|
89 */
|
nuclear@0
|
90
|
nuclear@0
|
91 #if CONST_BITS == 8
|
nuclear@0
|
92 #define FIX_1_082392200 ((INT32) 277) /* FIX(1.082392200) */
|
nuclear@0
|
93 #define FIX_1_414213562 ((INT32) 362) /* FIX(1.414213562) */
|
nuclear@0
|
94 #define FIX_1_847759065 ((INT32) 473) /* FIX(1.847759065) */
|
nuclear@0
|
95 #define FIX_2_613125930 ((INT32) 669) /* FIX(2.613125930) */
|
nuclear@0
|
96 #else
|
nuclear@0
|
97 #define FIX_1_082392200 FIX(1.082392200)
|
nuclear@0
|
98 #define FIX_1_414213562 FIX(1.414213562)
|
nuclear@0
|
99 #define FIX_1_847759065 FIX(1.847759065)
|
nuclear@0
|
100 #define FIX_2_613125930 FIX(2.613125930)
|
nuclear@0
|
101 #endif
|
nuclear@0
|
102
|
nuclear@0
|
103
|
nuclear@0
|
104 /* We can gain a little more speed, with a further compromise in accuracy,
|
nuclear@0
|
105 * by omitting the addition in a descaling shift. This yields an incorrectly
|
nuclear@0
|
106 * rounded result half the time...
|
nuclear@0
|
107 */
|
nuclear@0
|
108
|
nuclear@0
|
109 #ifndef USE_ACCURATE_ROUNDING
|
nuclear@0
|
110 #undef DESCALE
|
nuclear@0
|
111 #define DESCALE(x,n) RIGHT_SHIFT(x, n)
|
nuclear@0
|
112 #endif
|
nuclear@0
|
113
|
nuclear@0
|
114
|
nuclear@0
|
115 /* Multiply a DCTELEM variable by an INT32 constant, and immediately
|
nuclear@0
|
116 * descale to yield a DCTELEM result.
|
nuclear@0
|
117 */
|
nuclear@0
|
118
|
nuclear@0
|
119 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
|
nuclear@0
|
120
|
nuclear@0
|
121
|
nuclear@0
|
122 /* Dequantize a coefficient by multiplying it by the multiplier-table
|
nuclear@0
|
123 * entry; produce a DCTELEM result. For 8-bit data a 16x16->16
|
nuclear@0
|
124 * multiplication will do. For 12-bit data, the multiplier table is
|
nuclear@0
|
125 * declared INT32, so a 32-bit multiply will be used.
|
nuclear@0
|
126 */
|
nuclear@0
|
127
|
nuclear@0
|
128 #if BITS_IN_JSAMPLE == 8
|
nuclear@0
|
129 #define DEQUANTIZE(coef,quantval) (((IFAST_MULT_TYPE) (coef)) * (quantval))
|
nuclear@0
|
130 #else
|
nuclear@0
|
131 #define DEQUANTIZE(coef,quantval) \
|
nuclear@0
|
132 DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
|
nuclear@0
|
133 #endif
|
nuclear@0
|
134
|
nuclear@0
|
135
|
nuclear@0
|
136 /* Like DESCALE, but applies to a DCTELEM and produces an int.
|
nuclear@0
|
137 * We assume that int right shift is unsigned if INT32 right shift is.
|
nuclear@0
|
138 */
|
nuclear@0
|
139
|
nuclear@0
|
140 #ifdef RIGHT_SHIFT_IS_UNSIGNED
|
nuclear@0
|
141 #define ISHIFT_TEMPS DCTELEM ishift_temp;
|
nuclear@0
|
142 #if BITS_IN_JSAMPLE == 8
|
nuclear@0
|
143 #define DCTELEMBITS 16 /* DCTELEM may be 16 or 32 bits */
|
nuclear@0
|
144 #else
|
nuclear@0
|
145 #define DCTELEMBITS 32 /* DCTELEM must be 32 bits */
|
nuclear@0
|
146 #endif
|
nuclear@0
|
147 #define IRIGHT_SHIFT(x,shft) \
|
nuclear@0
|
148 ((ishift_temp = (x)) < 0 ? \
|
nuclear@0
|
149 (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
|
nuclear@0
|
150 (ishift_temp >> (shft)))
|
nuclear@0
|
151 #else
|
nuclear@0
|
152 #define ISHIFT_TEMPS
|
nuclear@0
|
153 #define IRIGHT_SHIFT(x,shft) ((x) >> (shft))
|
nuclear@0
|
154 #endif
|
nuclear@0
|
155
|
nuclear@0
|
156 #ifdef USE_ACCURATE_ROUNDING
|
nuclear@0
|
157 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
|
nuclear@0
|
158 #else
|
nuclear@0
|
159 #define IDESCALE(x,n) ((int) IRIGHT_SHIFT(x, n))
|
nuclear@0
|
160 #endif
|
nuclear@0
|
161
|
nuclear@0
|
162
|
nuclear@0
|
163 /*
|
nuclear@0
|
164 * Perform dequantization and inverse DCT on one block of coefficients.
|
nuclear@0
|
165 */
|
nuclear@0
|
166
|
nuclear@0
|
167 GLOBAL(void)
|
nuclear@0
|
168 jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
|
nuclear@0
|
169 JCOEFPTR coef_block,
|
nuclear@0
|
170 JSAMPARRAY output_buf, JDIMENSION output_col)
|
nuclear@0
|
171 {
|
nuclear@0
|
172 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
|
nuclear@0
|
173 DCTELEM tmp10, tmp11, tmp12, tmp13;
|
nuclear@0
|
174 DCTELEM z5, z10, z11, z12, z13;
|
nuclear@0
|
175 JCOEFPTR inptr;
|
nuclear@0
|
176 IFAST_MULT_TYPE * quantptr;
|
nuclear@0
|
177 int * wsptr;
|
nuclear@0
|
178 JSAMPROW outptr;
|
nuclear@0
|
179 JSAMPLE *range_limit = IDCT_range_limit(cinfo);
|
nuclear@0
|
180 int ctr;
|
nuclear@0
|
181 int workspace[DCTSIZE2]; /* buffers data between passes */
|
nuclear@0
|
182 SHIFT_TEMPS /* for DESCALE */
|
nuclear@0
|
183 ISHIFT_TEMPS /* for IDESCALE */
|
nuclear@0
|
184
|
nuclear@0
|
185 /* Pass 1: process columns from input, store into work array. */
|
nuclear@0
|
186
|
nuclear@0
|
187 inptr = coef_block;
|
nuclear@0
|
188 quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
|
nuclear@0
|
189 wsptr = workspace;
|
nuclear@0
|
190 for (ctr = DCTSIZE; ctr > 0; ctr--) {
|
nuclear@0
|
191 /* Due to quantization, we will usually find that many of the input
|
nuclear@0
|
192 * coefficients are zero, especially the AC terms. We can exploit this
|
nuclear@0
|
193 * by short-circuiting the IDCT calculation for any column in which all
|
nuclear@0
|
194 * the AC terms are zero. In that case each output is equal to the
|
nuclear@0
|
195 * DC coefficient (with scale factor as needed).
|
nuclear@0
|
196 * With typical images and quantization tables, half or more of the
|
nuclear@0
|
197 * column DCT calculations can be simplified this way.
|
nuclear@0
|
198 */
|
nuclear@0
|
199
|
nuclear@0
|
200 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
|
nuclear@0
|
201 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
|
nuclear@0
|
202 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
|
nuclear@0
|
203 inptr[DCTSIZE*7] == 0) {
|
nuclear@0
|
204 /* AC terms all zero */
|
nuclear@0
|
205 int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
|
nuclear@0
|
206
|
nuclear@0
|
207 wsptr[DCTSIZE*0] = dcval;
|
nuclear@0
|
208 wsptr[DCTSIZE*1] = dcval;
|
nuclear@0
|
209 wsptr[DCTSIZE*2] = dcval;
|
nuclear@0
|
210 wsptr[DCTSIZE*3] = dcval;
|
nuclear@0
|
211 wsptr[DCTSIZE*4] = dcval;
|
nuclear@0
|
212 wsptr[DCTSIZE*5] = dcval;
|
nuclear@0
|
213 wsptr[DCTSIZE*6] = dcval;
|
nuclear@0
|
214 wsptr[DCTSIZE*7] = dcval;
|
nuclear@0
|
215
|
nuclear@0
|
216 inptr++; /* advance pointers to next column */
|
nuclear@0
|
217 quantptr++;
|
nuclear@0
|
218 wsptr++;
|
nuclear@0
|
219 continue;
|
nuclear@0
|
220 }
|
nuclear@0
|
221
|
nuclear@0
|
222 /* Even part */
|
nuclear@0
|
223
|
nuclear@0
|
224 tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
|
nuclear@0
|
225 tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
|
nuclear@0
|
226 tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
|
nuclear@0
|
227 tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
|
nuclear@0
|
228
|
nuclear@0
|
229 tmp10 = tmp0 + tmp2; /* phase 3 */
|
nuclear@0
|
230 tmp11 = tmp0 - tmp2;
|
nuclear@0
|
231
|
nuclear@0
|
232 tmp13 = tmp1 + tmp3; /* phases 5-3 */
|
nuclear@0
|
233 tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
|
nuclear@0
|
234
|
nuclear@0
|
235 tmp0 = tmp10 + tmp13; /* phase 2 */
|
nuclear@0
|
236 tmp3 = tmp10 - tmp13;
|
nuclear@0
|
237 tmp1 = tmp11 + tmp12;
|
nuclear@0
|
238 tmp2 = tmp11 - tmp12;
|
nuclear@0
|
239
|
nuclear@0
|
240 /* Odd part */
|
nuclear@0
|
241
|
nuclear@0
|
242 tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
|
nuclear@0
|
243 tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
|
nuclear@0
|
244 tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
|
nuclear@0
|
245 tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
|
nuclear@0
|
246
|
nuclear@0
|
247 z13 = tmp6 + tmp5; /* phase 6 */
|
nuclear@0
|
248 z10 = tmp6 - tmp5;
|
nuclear@0
|
249 z11 = tmp4 + tmp7;
|
nuclear@0
|
250 z12 = tmp4 - tmp7;
|
nuclear@0
|
251
|
nuclear@0
|
252 tmp7 = z11 + z13; /* phase 5 */
|
nuclear@0
|
253 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
|
nuclear@0
|
254
|
nuclear@0
|
255 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
|
nuclear@0
|
256 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
|
nuclear@0
|
257 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
|
nuclear@0
|
258
|
nuclear@0
|
259 tmp6 = tmp12 - tmp7; /* phase 2 */
|
nuclear@0
|
260 tmp5 = tmp11 - tmp6;
|
nuclear@0
|
261 tmp4 = tmp10 + tmp5;
|
nuclear@0
|
262
|
nuclear@0
|
263 wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
|
nuclear@0
|
264 wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
|
nuclear@0
|
265 wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
|
nuclear@0
|
266 wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
|
nuclear@0
|
267 wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
|
nuclear@0
|
268 wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
|
nuclear@0
|
269 wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
|
nuclear@0
|
270 wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
|
nuclear@0
|
271
|
nuclear@0
|
272 inptr++; /* advance pointers to next column */
|
nuclear@0
|
273 quantptr++;
|
nuclear@0
|
274 wsptr++;
|
nuclear@0
|
275 }
|
nuclear@0
|
276
|
nuclear@0
|
277 /* Pass 2: process rows from work array, store into output array. */
|
nuclear@0
|
278 /* Note that we must descale the results by a factor of 8 == 2**3, */
|
nuclear@0
|
279 /* and also undo the PASS1_BITS scaling. */
|
nuclear@0
|
280
|
nuclear@0
|
281 wsptr = workspace;
|
nuclear@0
|
282 for (ctr = 0; ctr < DCTSIZE; ctr++) {
|
nuclear@0
|
283 outptr = output_buf[ctr] + output_col;
|
nuclear@0
|
284 /* Rows of zeroes can be exploited in the same way as we did with columns.
|
nuclear@0
|
285 * However, the column calculation has created many nonzero AC terms, so
|
nuclear@0
|
286 * the simplification applies less often (typically 5% to 10% of the time).
|
nuclear@0
|
287 * On machines with very fast multiplication, it's possible that the
|
nuclear@0
|
288 * test takes more time than it's worth. In that case this section
|
nuclear@0
|
289 * may be commented out.
|
nuclear@0
|
290 */
|
nuclear@0
|
291
|
nuclear@0
|
292 #ifndef NO_ZERO_ROW_TEST
|
nuclear@0
|
293 if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
|
nuclear@0
|
294 wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
|
nuclear@0
|
295 /* AC terms all zero */
|
nuclear@0
|
296 JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
|
nuclear@0
|
297 & RANGE_MASK];
|
nuclear@0
|
298
|
nuclear@0
|
299 outptr[0] = dcval;
|
nuclear@0
|
300 outptr[1] = dcval;
|
nuclear@0
|
301 outptr[2] = dcval;
|
nuclear@0
|
302 outptr[3] = dcval;
|
nuclear@0
|
303 outptr[4] = dcval;
|
nuclear@0
|
304 outptr[5] = dcval;
|
nuclear@0
|
305 outptr[6] = dcval;
|
nuclear@0
|
306 outptr[7] = dcval;
|
nuclear@0
|
307
|
nuclear@0
|
308 wsptr += DCTSIZE; /* advance pointer to next row */
|
nuclear@0
|
309 continue;
|
nuclear@0
|
310 }
|
nuclear@0
|
311 #endif
|
nuclear@0
|
312
|
nuclear@0
|
313 /* Even part */
|
nuclear@0
|
314
|
nuclear@0
|
315 tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
|
nuclear@0
|
316 tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
|
nuclear@0
|
317
|
nuclear@0
|
318 tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
|
nuclear@0
|
319 tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
|
nuclear@0
|
320 - tmp13;
|
nuclear@0
|
321
|
nuclear@0
|
322 tmp0 = tmp10 + tmp13;
|
nuclear@0
|
323 tmp3 = tmp10 - tmp13;
|
nuclear@0
|
324 tmp1 = tmp11 + tmp12;
|
nuclear@0
|
325 tmp2 = tmp11 - tmp12;
|
nuclear@0
|
326
|
nuclear@0
|
327 /* Odd part */
|
nuclear@0
|
328
|
nuclear@0
|
329 z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
|
nuclear@0
|
330 z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
|
nuclear@0
|
331 z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
|
nuclear@0
|
332 z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
|
nuclear@0
|
333
|
nuclear@0
|
334 tmp7 = z11 + z13; /* phase 5 */
|
nuclear@0
|
335 tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
|
nuclear@0
|
336
|
nuclear@0
|
337 z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
|
nuclear@0
|
338 tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
|
nuclear@0
|
339 tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
|
nuclear@0
|
340
|
nuclear@0
|
341 tmp6 = tmp12 - tmp7; /* phase 2 */
|
nuclear@0
|
342 tmp5 = tmp11 - tmp6;
|
nuclear@0
|
343 tmp4 = tmp10 + tmp5;
|
nuclear@0
|
344
|
nuclear@0
|
345 /* Final output stage: scale down by a factor of 8 and range-limit */
|
nuclear@0
|
346
|
nuclear@0
|
347 outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
|
nuclear@0
|
348 & RANGE_MASK];
|
nuclear@0
|
349 outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
|
nuclear@0
|
350 & RANGE_MASK];
|
nuclear@0
|
351 outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
|
nuclear@0
|
352 & RANGE_MASK];
|
nuclear@0
|
353 outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
|
nuclear@0
|
354 & RANGE_MASK];
|
nuclear@0
|
355 outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
|
nuclear@0
|
356 & RANGE_MASK];
|
nuclear@0
|
357 outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
|
nuclear@0
|
358 & RANGE_MASK];
|
nuclear@0
|
359 outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
|
nuclear@0
|
360 & RANGE_MASK];
|
nuclear@0
|
361 outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
|
nuclear@0
|
362 & RANGE_MASK];
|
nuclear@0
|
363
|
nuclear@0
|
364 wsptr += DCTSIZE; /* advance pointer to next row */
|
nuclear@0
|
365 }
|
nuclear@0
|
366 }
|
nuclear@0
|
367
|
nuclear@0
|
368 #endif /* DCT_IFAST_SUPPORTED */
|