FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
jfdctfst.c
Go to the documentation of this file.
1 /*
2  * This file is part of the Independent JPEG Group's software.
3  *
4  * The authors make NO WARRANTY or representation, either express or implied,
5  * with respect to this software, its quality, accuracy, merchantability, or
6  * fitness for a particular purpose. This software is provided "AS IS", and
7  * you, its user, assume the entire risk as to its quality and accuracy.
8  *
9  * This software is copyright (C) 1994-1996, Thomas G. Lane.
10  * All Rights Reserved except as specified below.
11  *
12  * Permission is hereby granted to use, copy, modify, and distribute this
13  * software (or portions thereof) for any purpose, without fee, subject to
14  * these conditions:
15  * (1) If any part of the source code for this software is distributed, then
16  * this README file must be included, with this copyright and no-warranty
17  * notice unaltered; and any additions, deletions, or changes to the original
18  * files must be clearly indicated in accompanying documentation.
19  * (2) If only executable code is distributed, then the accompanying
20  * documentation must state that "this software is based in part on the work
21  * of the Independent JPEG Group".
22  * (3) Permission for use of this software is granted only if the user accepts
23  * full responsibility for any undesirable consequences; the authors accept
24  * NO LIABILITY for damages of any kind.
25  *
26  * These conditions apply to any software derived from or based on the IJG
27  * code, not just to the unmodified library. If you use our work, you ought
28  * to acknowledge us.
29  *
30  * Permission is NOT granted for the use of any IJG author's name or company
31  * name in advertising or publicity relating to this software or products
32  * derived from it. This software may be referred to only as "the Independent
33  * JPEG Group's software".
34  *
35  * We specifically permit and encourage the use of this software as the basis
36  * of commercial products, provided that all warranty or liability claims are
37  * assumed by the product vendor.
38  *
39  * This file contains a fast, not so accurate integer implementation of the
40  * forward DCT (Discrete Cosine Transform).
41  *
42  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
43  * on each column. Direct algorithms are also available, but they are
44  * much more complex and seem not to be any faster when reduced to code.
45  *
46  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
47  * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
48  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
49  * JPEG textbook (see REFERENCES section in file README). The following code
50  * is based directly on figure 4-8 in P&M.
51  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
52  * possible to arrange the computation so that many of the multiplies are
53  * simple scalings of the final outputs. These multiplies can then be
54  * folded into the multiplications or divisions by the JPEG quantization
55  * table entries. The AA&N method leaves only 5 multiplies and 29 adds
56  * to be done in the DCT itself.
57  * The primary disadvantage of this method is that with fixed-point math,
58  * accuracy is lost due to imprecise representation of the scaled
59  * quantization values. The smaller the quantization table entry, the less
60  * precise the scaled value, so this implementation does worse with high-
61  * quality-setting files than with low-quality ones.
62  */
63 
64 /**
65  * @file
66  * Independent JPEG Group's fast AAN dct.
67  */
68 
69 #include <stdlib.h>
70 #include <stdio.h>
71 #include "libavutil/common.h"
72 #include "dct.h"
73 
74 #define DCTSIZE 8
75 #define GLOBAL(x) x
76 #define RIGHT_SHIFT(x, n) ((x) >> (n))
77 
78 /*
79  * This module is specialized to the case DCTSIZE = 8.
80  */
81 
82 #if DCTSIZE != 8
83  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
84 #endif
85 
86 
87 /* Scaling decisions are generally the same as in the LL&M algorithm;
88  * see jfdctint.c for more details. However, we choose to descale
89  * (right shift) multiplication products as soon as they are formed,
90  * rather than carrying additional fractional bits into subsequent additions.
91  * This compromises accuracy slightly, but it lets us save a few shifts.
92  * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
93  * everywhere except in the multiplications proper; this saves a good deal
94  * of work on 16-bit-int machines.
95  *
96  * Again to save a few shifts, the intermediate results between pass 1 and
97  * pass 2 are not upscaled, but are represented only to integral precision.
98  *
99  * A final compromise is to represent the multiplicative constants to only
100  * 8 fractional bits, rather than 13. This saves some shifting work on some
101  * machines, and may also reduce the cost of multiplication (since there
102  * are fewer one-bits in the constants).
103  */
104 
105 #define CONST_BITS 8
106 
107 
108 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
109  * causing a lot of useless floating-point operations at run time.
110  * To get around this we use the following pre-calculated constants.
111  * If you change CONST_BITS you may want to add appropriate values.
112  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
113  */
114 
115 #if CONST_BITS == 8
116 #define FIX_0_382683433 ((int32_t) 98) /* FIX(0.382683433) */
117 #define FIX_0_541196100 ((int32_t) 139) /* FIX(0.541196100) */
118 #define FIX_0_707106781 ((int32_t) 181) /* FIX(0.707106781) */
119 #define FIX_1_306562965 ((int32_t) 334) /* FIX(1.306562965) */
120 #else
121 #define FIX_0_382683433 FIX(0.382683433)
122 #define FIX_0_541196100 FIX(0.541196100)
123 #define FIX_0_707106781 FIX(0.707106781)
124 #define FIX_1_306562965 FIX(1.306562965)
125 #endif
126 
127 
128 /* We can gain a little more speed, with a further compromise in accuracy,
129  * by omitting the addition in a descaling shift. This yields an incorrectly
130  * rounded result half the time...
131  */
132 
133 #ifndef USE_ACCURATE_ROUNDING
134 #undef DESCALE
135 #define DESCALE(x,n) RIGHT_SHIFT(x, n)
136 #endif
137 
138 
139 /* Multiply a int16_t variable by an int32_t constant, and immediately
140  * descale to yield a int16_t result.
141  */
142 
143 #define MULTIPLY(var,const) ((int16_t) DESCALE((var) * (const), CONST_BITS))
144 
145 static av_always_inline void row_fdct(int16_t * data){
146  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
147  int tmp10, tmp11, tmp12, tmp13;
148  int z1, z2, z3, z4, z5, z11, z13;
149  int16_t *dataptr;
150  int ctr;
151 
152  /* Pass 1: process rows. */
153 
154  dataptr = data;
155  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
156  tmp0 = dataptr[0] + dataptr[7];
157  tmp7 = dataptr[0] - dataptr[7];
158  tmp1 = dataptr[1] + dataptr[6];
159  tmp6 = dataptr[1] - dataptr[6];
160  tmp2 = dataptr[2] + dataptr[5];
161  tmp5 = dataptr[2] - dataptr[5];
162  tmp3 = dataptr[3] + dataptr[4];
163  tmp4 = dataptr[3] - dataptr[4];
164 
165  /* Even part */
166 
167  tmp10 = tmp0 + tmp3; /* phase 2 */
168  tmp13 = tmp0 - tmp3;
169  tmp11 = tmp1 + tmp2;
170  tmp12 = tmp1 - tmp2;
171 
172  dataptr[0] = tmp10 + tmp11; /* phase 3 */
173  dataptr[4] = tmp10 - tmp11;
174 
175  z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
176  dataptr[2] = tmp13 + z1; /* phase 5 */
177  dataptr[6] = tmp13 - z1;
178 
179  /* Odd part */
180 
181  tmp10 = tmp4 + tmp5; /* phase 2 */
182  tmp11 = tmp5 + tmp6;
183  tmp12 = tmp6 + tmp7;
184 
185  /* The rotator is modified from fig 4-8 to avoid extra negations. */
186  z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
187  z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
188  z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
189  z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
190 
191  z11 = tmp7 + z3; /* phase 5 */
192  z13 = tmp7 - z3;
193 
194  dataptr[5] = z13 + z2; /* phase 6 */
195  dataptr[3] = z13 - z2;
196  dataptr[1] = z11 + z4;
197  dataptr[7] = z11 - z4;
198 
199  dataptr += DCTSIZE; /* advance pointer to next row */
200  }
201 }
202 
203 /*
204  * Perform the forward DCT on one block of samples.
205  */
206 
207 GLOBAL(void)
208 ff_fdct_ifast (int16_t * data)
209 {
210  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
211  int tmp10, tmp11, tmp12, tmp13;
212  int z1, z2, z3, z4, z5, z11, z13;
213  int16_t *dataptr;
214  int ctr;
215 
216  row_fdct(data);
217 
218  /* Pass 2: process columns. */
219 
220  dataptr = data;
221  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
222  tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
223  tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
224  tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
225  tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
226  tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
227  tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
228  tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
229  tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
230 
231  /* Even part */
232 
233  tmp10 = tmp0 + tmp3; /* phase 2 */
234  tmp13 = tmp0 - tmp3;
235  tmp11 = tmp1 + tmp2;
236  tmp12 = tmp1 - tmp2;
237 
238  dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
239  dataptr[DCTSIZE*4] = tmp10 - tmp11;
240 
241  z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
242  dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
243  dataptr[DCTSIZE*6] = tmp13 - z1;
244 
245  /* Odd part */
246 
247  tmp10 = tmp4 + tmp5; /* phase 2 */
248  tmp11 = tmp5 + tmp6;
249  tmp12 = tmp6 + tmp7;
250 
251  /* The rotator is modified from fig 4-8 to avoid extra negations. */
252  z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
253  z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
254  z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
255  z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
256 
257  z11 = tmp7 + z3; /* phase 5 */
258  z13 = tmp7 - z3;
259 
260  dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
261  dataptr[DCTSIZE*3] = z13 - z2;
262  dataptr[DCTSIZE*1] = z11 + z4;
263  dataptr[DCTSIZE*7] = z11 - z4;
264 
265  dataptr++; /* advance pointer to next column */
266  }
267 }
268 
269 /*
270  * Perform the forward 2-4-8 DCT on one block of samples.
271  */
272 
273 GLOBAL(void)
275 {
276  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
277  int tmp10, tmp11, tmp12, tmp13;
278  int z1;
279  int16_t *dataptr;
280  int ctr;
281 
282  row_fdct(data);
283 
284  /* Pass 2: process columns. */
285 
286  dataptr = data;
287  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
288  tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
289  tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
290  tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
291  tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
292  tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
293  tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
294  tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
295  tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
296 
297  /* Even part */
298 
299  tmp10 = tmp0 + tmp3;
300  tmp11 = tmp1 + tmp2;
301  tmp12 = tmp1 - tmp2;
302  tmp13 = tmp0 - tmp3;
303 
304  dataptr[DCTSIZE*0] = tmp10 + tmp11;
305  dataptr[DCTSIZE*4] = tmp10 - tmp11;
306 
307  z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
308  dataptr[DCTSIZE*2] = tmp13 + z1;
309  dataptr[DCTSIZE*6] = tmp13 - z1;
310 
311  tmp10 = tmp4 + tmp7;
312  tmp11 = tmp5 + tmp6;
313  tmp12 = tmp5 - tmp6;
314  tmp13 = tmp4 - tmp7;
315 
316  dataptr[DCTSIZE*1] = tmp10 + tmp11;
317  dataptr[DCTSIZE*5] = tmp10 - tmp11;
318 
319  z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
320  dataptr[DCTSIZE*3] = tmp13 + z1;
321  dataptr[DCTSIZE*7] = tmp13 - z1;
322 
323  dataptr++; /* advance pointer to next column */
324  }
325 }
326 
327 
328 #undef GLOBAL
329 #undef CONST_BITS
330 #undef DESCALE
331 #undef FIX_0_541196100
332 #undef FIX_1_306562965