xref: /third_party/ffmpeg/libavcodec/jfdctfst.c (revision cabdff1a)
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 <stdint.h>
70#include "libavutil/attributes.h"
71#include "dct.h"
72
73#define DCTSIZE 8
74#define GLOBAL(x) x
75#define RIGHT_SHIFT(x, n) ((x) >> (n))
76
77/*
78 * This module is specialized to the case DCTSIZE = 8.
79 */
80
81#if DCTSIZE != 8
82  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
83#endif
84
85
86/* Scaling decisions are generally the same as in the LL&M algorithm;
87 * see jfdctint.c for more details.  However, we choose to descale
88 * (right shift) multiplication products as soon as they are formed,
89 * rather than carrying additional fractional bits into subsequent additions.
90 * This compromises accuracy slightly, but it lets us save a few shifts.
91 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
92 * everywhere except in the multiplications proper; this saves a good deal
93 * of work on 16-bit-int machines.
94 *
95 * Again to save a few shifts, the intermediate results between pass 1 and
96 * pass 2 are not upscaled, but are represented only to integral precision.
97 *
98 * A final compromise is to represent the multiplicative constants to only
99 * 8 fractional bits, rather than 13.  This saves some shifting work on some
100 * machines, and may also reduce the cost of multiplication (since there
101 * are fewer one-bits in the constants).
102 */
103
104#define CONST_BITS  8
105
106
107/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
108 * causing a lot of useless floating-point operations at run time.
109 * To get around this we use the following pre-calculated constants.
110 * If you change CONST_BITS you may want to add appropriate values.
111 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
112 */
113
114#if CONST_BITS == 8
115#define FIX_0_382683433  ((int32_t)   98)       /* FIX(0.382683433) */
116#define FIX_0_541196100  ((int32_t)  139)       /* FIX(0.541196100) */
117#define FIX_0_707106781  ((int32_t)  181)       /* FIX(0.707106781) */
118#define FIX_1_306562965  ((int32_t)  334)       /* FIX(1.306562965) */
119#else
120#define FIX_0_382683433  FIX(0.382683433)
121#define FIX_0_541196100  FIX(0.541196100)
122#define FIX_0_707106781  FIX(0.707106781)
123#define FIX_1_306562965  FIX(1.306562965)
124#endif
125
126
127/* We can gain a little more speed, with a further compromise in accuracy,
128 * by omitting the addition in a descaling shift.  This yields an incorrectly
129 * rounded result half the time...
130 */
131
132#ifndef USE_ACCURATE_ROUNDING
133#undef DESCALE
134#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
135#endif
136
137
138/* Multiply a int16_t variable by an int32_t constant, and immediately
139 * descale to yield a int16_t result.
140 */
141
142#define MULTIPLY(var,const)  ((int16_t) DESCALE((var) * (const), CONST_BITS))
143
144static av_always_inline void row_fdct(int16_t * data){
145  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
146  int tmp10, tmp11, tmp12, tmp13;
147  int z1, z2, z3, z4, z5, z11, z13;
148  int16_t *dataptr;
149  int ctr;
150
151  /* Pass 1: process rows. */
152
153  dataptr = data;
154  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
155    tmp0 = dataptr[0] + dataptr[7];
156    tmp7 = dataptr[0] - dataptr[7];
157    tmp1 = dataptr[1] + dataptr[6];
158    tmp6 = dataptr[1] - dataptr[6];
159    tmp2 = dataptr[2] + dataptr[5];
160    tmp5 = dataptr[2] - dataptr[5];
161    tmp3 = dataptr[3] + dataptr[4];
162    tmp4 = dataptr[3] - dataptr[4];
163
164    /* Even part */
165
166    tmp10 = tmp0 + tmp3;        /* phase 2 */
167    tmp13 = tmp0 - tmp3;
168    tmp11 = tmp1 + tmp2;
169    tmp12 = tmp1 - tmp2;
170
171    dataptr[0] = tmp10 + tmp11; /* phase 3 */
172    dataptr[4] = tmp10 - tmp11;
173
174    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
175    dataptr[2] = tmp13 + z1;    /* phase 5 */
176    dataptr[6] = tmp13 - z1;
177
178    /* Odd part */
179
180    tmp10 = tmp4 + tmp5;        /* phase 2 */
181    tmp11 = tmp5 + tmp6;
182    tmp12 = tmp6 + tmp7;
183
184    /* The rotator is modified from fig 4-8 to avoid extra negations. */
185    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
186    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5;    /* c2-c6 */
187    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5;    /* c2+c6 */
188    z3 = MULTIPLY(tmp11, FIX_0_707106781);         /* c4 */
189
190    z11 = tmp7 + z3;            /* phase 5 */
191    z13 = tmp7 - z3;
192
193    dataptr[5] = z13 + z2;      /* phase 6 */
194    dataptr[3] = z13 - z2;
195    dataptr[1] = z11 + z4;
196    dataptr[7] = z11 - z4;
197
198    dataptr += DCTSIZE;         /* advance pointer to next row */
199  }
200}
201
202/*
203 * Perform the forward DCT on one block of samples.
204 */
205
206GLOBAL(void)
207ff_fdct_ifast (int16_t * data)
208{
209  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
210  int tmp10, tmp11, tmp12, tmp13;
211  int z1, z2, z3, z4, z5, z11, z13;
212  int16_t *dataptr;
213  int ctr;
214
215  row_fdct(data);
216
217  /* Pass 2: process columns. */
218
219  dataptr = data;
220  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
221    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
222    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
223    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
224    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
225    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
226    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
227    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
228    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
229
230    /* Even part */
231
232    tmp10 = tmp0 + tmp3;        /* phase 2 */
233    tmp13 = tmp0 - tmp3;
234    tmp11 = tmp1 + tmp2;
235    tmp12 = tmp1 - tmp2;
236
237    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
238    dataptr[DCTSIZE*4] = tmp10 - tmp11;
239
240    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
241    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
242    dataptr[DCTSIZE*6] = tmp13 - z1;
243
244    /* Odd part */
245
246    tmp10 = tmp4 + tmp5;        /* phase 2 */
247    tmp11 = tmp5 + tmp6;
248    tmp12 = tmp6 + tmp7;
249
250    /* The rotator is modified from fig 4-8 to avoid extra negations. */
251    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
252    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
253    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
254    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
255
256    z11 = tmp7 + z3;            /* phase 5 */
257    z13 = tmp7 - z3;
258
259    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
260    dataptr[DCTSIZE*3] = z13 - z2;
261    dataptr[DCTSIZE*1] = z11 + z4;
262    dataptr[DCTSIZE*7] = z11 - z4;
263
264    dataptr++;                  /* advance pointer to next column */
265  }
266}
267
268/*
269 * Perform the forward 2-4-8 DCT on one block of samples.
270 */
271
272GLOBAL(void)
273ff_fdct_ifast248 (int16_t * data)
274{
275  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
276  int tmp10, tmp11, tmp12, tmp13;
277  int z1;
278  int16_t *dataptr;
279  int ctr;
280
281  row_fdct(data);
282
283  /* Pass 2: process columns. */
284
285  dataptr = data;
286  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
287    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
288    tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
289    tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
290    tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
291    tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
292    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
293    tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
294    tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
295
296    /* Even part */
297
298    tmp10 = tmp0 + tmp3;
299    tmp11 = tmp1 + tmp2;
300    tmp12 = tmp1 - tmp2;
301    tmp13 = tmp0 - tmp3;
302
303    dataptr[DCTSIZE*0] = tmp10 + tmp11;
304    dataptr[DCTSIZE*4] = tmp10 - tmp11;
305
306    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
307    dataptr[DCTSIZE*2] = tmp13 + z1;
308    dataptr[DCTSIZE*6] = tmp13 - z1;
309
310    tmp10 = tmp4 + tmp7;
311    tmp11 = tmp5 + tmp6;
312    tmp12 = tmp5 - tmp6;
313    tmp13 = tmp4 - tmp7;
314
315    dataptr[DCTSIZE*1] = tmp10 + tmp11;
316    dataptr[DCTSIZE*5] = tmp10 - tmp11;
317
318    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
319    dataptr[DCTSIZE*3] = tmp13 + z1;
320    dataptr[DCTSIZE*7] = tmp13 - z1;
321
322    dataptr++;                        /* advance pointer to next column */
323  }
324}
325
326
327#undef GLOBAL
328#undef CONST_BITS
329#undef DESCALE
330#undef FIX_0_541196100
331#undef FIX_1_306562965
332