4
* This file is part of the Independent JPEG Group's software.
6
* The authors make NO WARRANTY or representation, either express or implied,
7
* with respect to this software, its quality, accuracy, merchantability, or
8
* fitness for a particular purpose. This software is provided "AS IS", and
9
* you, its user, assume the entire risk as to its quality and accuracy.
11
* This software is copyright (C) 1994-1996, Thomas G. Lane.
12
* All Rights Reserved except as specified below.
14
* Permission is hereby granted to use, copy, modify, and distribute this
15
* software (or portions thereof) for any purpose, without fee, subject to
17
* (1) If any part of the source code for this software is distributed, then
18
* this README file must be included, with this copyright and no-warranty
19
* notice unaltered; and any additions, deletions, or changes to the original
20
* files must be clearly indicated in accompanying documentation.
21
* (2) If only executable code is distributed, then the accompanying
22
* documentation must state that "this software is based in part on the work
23
* of the Independent JPEG Group".
24
* (3) Permission for use of this software is granted only if the user accepts
25
* full responsibility for any undesirable consequences; the authors accept
26
* NO LIABILITY for damages of any kind.
28
* These conditions apply to any software derived from or based on the IJG
29
* code, not just to the unmodified library. If you use our work, you ought
32
* Permission is NOT granted for the use of any IJG author's name or company
33
* name in advertising or publicity relating to this software or products
34
* derived from it. This software may be referred to only as "the Independent
35
* JPEG Group's software".
37
* We specifically permit and encourage the use of this software as the basis
38
* of commercial products, provided that all warranty or liability claims are
39
* assumed by the product vendor.
41
* This file contains a fast, not so accurate integer implementation of the
42
* forward DCT (Discrete Cosine Transform).
44
* A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
45
* on each column. Direct algorithms are also available, but they are
46
* much more complex and seem not to be any faster when reduced to code.
48
* This implementation is based on Arai, Agui, and Nakajima's algorithm for
49
* scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
50
* Japanese, but the algorithm is described in the Pennebaker & Mitchell
51
* JPEG textbook (see REFERENCES section in file README). The following code
52
* is based directly on figure 4-8 in P&M.
53
* While an 8-point DCT cannot be done in less than 11 multiplies, it is
54
* possible to arrange the computation so that many of the multiplies are
55
* simple scalings of the final outputs. These multiplies can then be
56
* folded into the multiplications or divisions by the JPEG quantization
57
* table entries. The AA&N method leaves only 5 multiplies and 29 adds
58
* to be done in the DCT itself.
59
* The primary disadvantage of this method is that with fixed-point math,
60
* accuracy is lost due to imprecise representation of the scaled
61
* quantization values. The smaller the quantization table entry, the less
62
* precise the scaled value, so this implementation does worse with high-
63
* quality-setting files than with low-quality ones.
68
* Independent JPEG Group's fast AAN dct.
73
#include "libavutil/common.h"
78
#define RIGHT_SHIFT(x, n) ((x) >> (n))
81
* This module is specialized to the case DCTSIZE = 8.
85
Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
89
/* Scaling decisions are generally the same as in the LL&M algorithm;
90
* see jfdctint.c for more details. However, we choose to descale
91
* (right shift) multiplication products as soon as they are formed,
92
* rather than carrying additional fractional bits into subsequent additions.
93
* This compromises accuracy slightly, but it lets us save a few shifts.
94
* More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
95
* everywhere except in the multiplications proper; this saves a good deal
96
* of work on 16-bit-int machines.
98
* Again to save a few shifts, the intermediate results between pass 1 and
99
* pass 2 are not upscaled, but are represented only to integral precision.
101
* A final compromise is to represent the multiplicative constants to only
102
* 8 fractional bits, rather than 13. This saves some shifting work on some
103
* machines, and may also reduce the cost of multiplication (since there
104
* are fewer one-bits in the constants).
110
/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
111
* causing a lot of useless floating-point operations at run time.
112
* To get around this we use the following pre-calculated constants.
113
* If you change CONST_BITS you may want to add appropriate values.
114
* (With a reasonable C compiler, you can just rely on the FIX() macro...)
118
#define FIX_0_382683433 ((int32_t) 98) /* FIX(0.382683433) */
119
#define FIX_0_541196100 ((int32_t) 139) /* FIX(0.541196100) */
120
#define FIX_0_707106781 ((int32_t) 181) /* FIX(0.707106781) */
121
#define FIX_1_306562965 ((int32_t) 334) /* FIX(1.306562965) */
123
#define FIX_0_382683433 FIX(0.382683433)
124
#define FIX_0_541196100 FIX(0.541196100)
125
#define FIX_0_707106781 FIX(0.707106781)
126
#define FIX_1_306562965 FIX(1.306562965)
130
/* We can gain a little more speed, with a further compromise in accuracy,
131
* by omitting the addition in a descaling shift. This yields an incorrectly
132
* rounded result half the time...
135
#ifndef USE_ACCURATE_ROUNDING
137
#define DESCALE(x,n) RIGHT_SHIFT(x, n)
141
/* Multiply a DCTELEM variable by an int32_t constant, and immediately
142
* descale to yield a DCTELEM result.
145
#define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
147
static av_always_inline void row_fdct(DCTELEM * data){
148
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
149
int_fast16_t tmp10, tmp11, tmp12, tmp13;
150
int_fast16_t z1, z2, z3, z4, z5, z11, z13;
154
/* Pass 1: process rows. */
157
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
158
tmp0 = dataptr[0] + dataptr[7];
159
tmp7 = dataptr[0] - dataptr[7];
160
tmp1 = dataptr[1] + dataptr[6];
161
tmp6 = dataptr[1] - dataptr[6];
162
tmp2 = dataptr[2] + dataptr[5];
163
tmp5 = dataptr[2] - dataptr[5];
164
tmp3 = dataptr[3] + dataptr[4];
165
tmp4 = dataptr[3] - dataptr[4];
169
tmp10 = tmp0 + tmp3; /* phase 2 */
174
dataptr[0] = tmp10 + tmp11; /* phase 3 */
175
dataptr[4] = tmp10 - tmp11;
177
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
178
dataptr[2] = tmp13 + z1; /* phase 5 */
179
dataptr[6] = tmp13 - z1;
183
tmp10 = tmp4 + tmp5; /* phase 2 */
187
/* The rotator is modified from fig 4-8 to avoid extra negations. */
188
z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
189
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
190
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
191
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
193
z11 = tmp7 + z3; /* phase 5 */
196
dataptr[5] = z13 + z2; /* phase 6 */
197
dataptr[3] = z13 - z2;
198
dataptr[1] = z11 + z4;
199
dataptr[7] = z11 - z4;
201
dataptr += DCTSIZE; /* advance pointer to next row */
206
* Perform the forward DCT on one block of samples.
210
fdct_ifast (DCTELEM * data)
212
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
213
int_fast16_t tmp10, tmp11, tmp12, tmp13;
214
int_fast16_t z1, z2, z3, z4, z5, z11, z13;
220
/* Pass 2: process columns. */
223
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
224
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
225
tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
226
tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
227
tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
228
tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
229
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
230
tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
231
tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
235
tmp10 = tmp0 + tmp3; /* phase 2 */
240
dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
241
dataptr[DCTSIZE*4] = tmp10 - tmp11;
243
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
244
dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
245
dataptr[DCTSIZE*6] = tmp13 - z1;
249
tmp10 = tmp4 + tmp5; /* phase 2 */
253
/* The rotator is modified from fig 4-8 to avoid extra negations. */
254
z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
255
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
256
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
257
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
259
z11 = tmp7 + z3; /* phase 5 */
262
dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
263
dataptr[DCTSIZE*3] = z13 - z2;
264
dataptr[DCTSIZE*1] = z11 + z4;
265
dataptr[DCTSIZE*7] = z11 - z4;
267
dataptr++; /* advance pointer to next column */
272
* Perform the forward 2-4-8 DCT on one block of samples.
276
fdct_ifast248 (DCTELEM * data)
278
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
279
int_fast16_t tmp10, tmp11, tmp12, tmp13;
286
/* Pass 2: process columns. */
289
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
290
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
291
tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
292
tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
293
tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
294
tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
295
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
296
tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
297
tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
306
dataptr[DCTSIZE*0] = tmp10 + tmp11;
307
dataptr[DCTSIZE*4] = tmp10 - tmp11;
309
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
310
dataptr[DCTSIZE*2] = tmp13 + z1;
311
dataptr[DCTSIZE*6] = tmp13 - z1;
318
dataptr[DCTSIZE*1] = tmp10 + tmp11;
319
dataptr[DCTSIZE*5] = tmp10 - tmp11;
321
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781);
322
dataptr[DCTSIZE*3] = tmp13 + z1;
323
dataptr[DCTSIZE*7] = tmp13 - z1;
325
dataptr++; /* advance pointer to next column */
333
#undef FIX_0_541196100
334
#undef FIX_1_306562965