ffmpeg / libavcodec / jfdctfst.c @ 849f1035
History  View  Annotate  Download (11.1 KB)
1 
/*


2 
* jfdctfst.c

3 
*

4 
* This file is part of the Independent JPEG Group's software.

5 
*

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.

10 
*

11 
* This software is copyright (C) 19941996, Thomas G. Lane.

12 
* All Rights Reserved except as specified below.

13 
*

14 
* Permission is hereby granted to use, copy, modify, and distribute this

15 
* software (or portions thereof) for any purpose, without fee, subject to

16 
* these conditions:

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 nowarranty

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.

27 
*

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

30 
* to acknowledge us.

31 
*

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".

36 
*

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.

40 
*

41 
* This file contains a fast, not so accurate integer implementation of the

42 
* forward DCT (Discrete Cosine Transform).

43 
*

44 
* A 2D DCT can be done by 1D DCT on each row followed by 1D 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.

47 
*

48 
* This implementation is based on Arai, Agui, and Nakajima's algorithm for

49 
* scaled DCT. Their original paper (Trans. IEICE E71(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 48 in P&M.

53 
* While an 8point 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 fixedpoint 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 
* qualitysetting files than with lowquality ones.

64 
*/

65  
66 
/**

67 
* @file jfdctfst.c

68 
* Independent JPEG Group's fast AAN dct.

69 
*/

70  
71 
#include <stdlib.h> 
72 
#include <stdio.h> 
73 
#include "common.h" 
74 
#include "dsputil.h" 
75  
76 
#define DCTSIZE 8 
77 
#define GLOBAL(x) x

78 
#define RIGHT_SHIFT(x, n) ((x) >> (n))

79 
#define SHIFT_TEMPS

80  
81 
/*

82 
* This module is specialized to the case DCTSIZE = 8.

83 
*/

84  
85 
#if DCTSIZE != 8 
86 
Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ 
87 
#endif

88  
89  
90 
/* Scaling decisions are generally the same as in the LL&M algorithm;

91 
* see jfdctint.c for more details. However, we choose to descale

92 
* (right shift) multiplication products as soon as they are formed,

93 
* rather than carrying additional fractional bits into subsequent additions.

94 
* This compromises accuracy slightly, but it lets us save a few shifts.

95 
* More importantly, 16bit arithmetic is then adequate (for 8bit samples)

96 
* everywhere except in the multiplications proper; this saves a good deal

97 
* of work on 16bitint machines.

98 
*

99 
* Again to save a few shifts, the intermediate results between pass 1 and

100 
* pass 2 are not upscaled, but are represented only to integral precision.

101 
*

102 
* A final compromise is to represent the multiplicative constants to only

103 
* 8 fractional bits, rather than 13. This saves some shifting work on some

104 
* machines, and may also reduce the cost of multiplication (since there

105 
* are fewer onebits in the constants).

106 
*/

107  
108 
#define CONST_BITS 8 
109  
110  
111 
/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus

112 
* causing a lot of useless floatingpoint operations at run time.

113 
* To get around this we use the following precalculated constants.

114 
* If you change CONST_BITS you may want to add appropriate values.

115 
* (With a reasonable C compiler, you can just rely on the FIX() macro...)

116 
*/

117  
118 
#if CONST_BITS == 8 
119 
#define FIX_0_382683433 ((int32_t) 98) /* FIX(0.382683433) */ 
120 
#define FIX_0_541196100 ((int32_t) 139) /* FIX(0.541196100) */ 
121 
#define FIX_0_707106781 ((int32_t) 181) /* FIX(0.707106781) */ 
122 
#define FIX_1_306562965 ((int32_t) 334) /* FIX(1.306562965) */ 
123 
#else

124 
#define FIX_0_382683433 FIX(0.382683433) 
125 
#define FIX_0_541196100 FIX(0.541196100) 
126 
#define FIX_0_707106781 FIX(0.707106781) 
127 
#define FIX_1_306562965 FIX(1.306562965) 
128 
#endif

129  
130  
131 
/* We can gain a little more speed, with a further compromise in accuracy,

132 
* by omitting the addition in a descaling shift. This yields an incorrectly

133 
* rounded result half the time...

134 
*/

135  
136 
#ifndef USE_ACCURATE_ROUNDING

137 
#undef DESCALE

138 
#define DESCALE(x,n) RIGHT_SHIFT(x, n)

139 
#endif

140  
141  
142 
/* Multiply a DCTELEM variable by an int32_t constant, and immediately

143 
* descale to yield a DCTELEM result.

144 
*/

145  
146 
#define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) 
147  
148 
static av_always_inline void row_fdct(DCTELEM * data){ 
149 
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 
150 
int_fast16_t tmp10, tmp11, tmp12, tmp13; 
151 
int_fast16_t z1, z2, z3, z4, z5, z11, z13; 
152 
DCTELEM *dataptr; 
153 
int ctr;

154 
SHIFT_TEMPS 
155  
156 
/* Pass 1: process rows. */

157  
158 
dataptr = data; 
159 
for (ctr = DCTSIZE1; ctr >= 0; ctr) { 
160 
tmp0 = dataptr[0] + dataptr[7]; 
161 
tmp7 = dataptr[0]  dataptr[7]; 
162 
tmp1 = dataptr[1] + dataptr[6]; 
163 
tmp6 = dataptr[1]  dataptr[6]; 
164 
tmp2 = dataptr[2] + dataptr[5]; 
165 
tmp5 = dataptr[2]  dataptr[5]; 
166 
tmp3 = dataptr[3] + dataptr[4]; 
167 
tmp4 = dataptr[3]  dataptr[4]; 
168  
169 
/* Even part */

170  
171 
tmp10 = tmp0 + tmp3; /* phase 2 */

172 
tmp13 = tmp0  tmp3; 
173 
tmp11 = tmp1 + tmp2; 
174 
tmp12 = tmp1  tmp2; 
175  
176 
dataptr[0] = tmp10 + tmp11; /* phase 3 */ 
177 
dataptr[4] = tmp10  tmp11;

178  
179 
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */

180 
dataptr[2] = tmp13 + z1; /* phase 5 */ 
181 
dataptr[6] = tmp13  z1;

182  
183 
/* Odd part */

184  
185 
tmp10 = tmp4 + tmp5; /* phase 2 */

186 
tmp11 = tmp5 + tmp6; 
187 
tmp12 = tmp6 + tmp7; 
188  
189 
/* The rotator is modified from fig 48 to avoid extra negations. */

190 
z5 = MULTIPLY(tmp10  tmp12, FIX_0_382683433); /* c6 */

191 
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2c6 */

192 
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */

193 
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */

194  
195 
z11 = tmp7 + z3; /* phase 5 */

196 
z13 = tmp7  z3; 
197  
198 
dataptr[5] = z13 + z2; /* phase 6 */ 
199 
dataptr[3] = z13  z2;

200 
dataptr[1] = z11 + z4;

201 
dataptr[7] = z11  z4;

202  
203 
dataptr += DCTSIZE; /* advance pointer to next row */

204 
} 
205 
} 
206  
207 
/*

208 
* Perform the forward DCT on one block of samples.

209 
*/

210  
211 
GLOBAL(void)

212 
fdct_ifast (DCTELEM * data) 
213 
{ 
214 
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 
215 
int_fast16_t tmp10, tmp11, tmp12, tmp13; 
216 
int_fast16_t z1, z2, z3, z4, z5, z11, z13; 
217 
DCTELEM *dataptr; 
218 
int ctr;

219 
SHIFT_TEMPS 
220  
221 
row_fdct(data); 
222  
223 
/* Pass 2: process columns. */

224  
225 
dataptr = data; 
226 
for (ctr = DCTSIZE1; ctr >= 0; ctr) { 
227 
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; 
228 
tmp7 = dataptr[DCTSIZE*0]  dataptr[DCTSIZE*7]; 
229 
tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; 
230 
tmp6 = dataptr[DCTSIZE*1]  dataptr[DCTSIZE*6]; 
231 
tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; 
232 
tmp5 = dataptr[DCTSIZE*2]  dataptr[DCTSIZE*5]; 
233 
tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; 
234 
tmp4 = dataptr[DCTSIZE*3]  dataptr[DCTSIZE*4]; 
235  
236 
/* Even part */

237  
238 
tmp10 = tmp0 + tmp3; /* phase 2 */

239 
tmp13 = tmp0  tmp3; 
240 
tmp11 = tmp1 + tmp2; 
241 
tmp12 = tmp1  tmp2; 
242  
243 
dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ 
244 
dataptr[DCTSIZE*4] = tmp10  tmp11;

245  
246 
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */

247 
dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ 
248 
dataptr[DCTSIZE*6] = tmp13  z1;

249  
250 
/* Odd part */

251  
252 
tmp10 = tmp4 + tmp5; /* phase 2 */

253 
tmp11 = tmp5 + tmp6; 
254 
tmp12 = tmp6 + tmp7; 
255  
256 
/* The rotator is modified from fig 48 to avoid extra negations. */

257 
z5 = MULTIPLY(tmp10  tmp12, FIX_0_382683433); /* c6 */

258 
z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2c6 */

259 
z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */

260 
z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */

261  
262 
z11 = tmp7 + z3; /* phase 5 */

263 
z13 = tmp7  z3; 
264  
265 
dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ 
266 
dataptr[DCTSIZE*3] = z13  z2;

267 
dataptr[DCTSIZE*1] = z11 + z4;

268 
dataptr[DCTSIZE*7] = z11  z4;

269  
270 
dataptr++; /* advance pointer to next column */

271 
} 
272 
} 
273  
274 
/*

275 
* Perform the forward 248 DCT on one block of samples.

276 
*/

277  
278 
GLOBAL(void)

279 
fdct_ifast248 (DCTELEM * data) 
280 
{ 
281 
int_fast16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 
282 
int_fast16_t tmp10, tmp11, tmp12, tmp13; 
283 
int_fast16_t z1; 
284 
DCTELEM *dataptr; 
285 
int ctr;

286 
SHIFT_TEMPS 
287  
288 
row_fdct(data); 
289  
290 
/* Pass 2: process columns. */

291  
292 
dataptr = data; 
293 
for (ctr = DCTSIZE1; ctr >= 0; ctr) { 
294 
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1]; 
295 
tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3]; 
296 
tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5]; 
297 
tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7]; 
298 
tmp4 = dataptr[DCTSIZE*0]  dataptr[DCTSIZE*1]; 
299 
tmp5 = dataptr[DCTSIZE*2]  dataptr[DCTSIZE*3]; 
300 
tmp6 = dataptr[DCTSIZE*4]  dataptr[DCTSIZE*5]; 
301 
tmp7 = dataptr[DCTSIZE*6]  dataptr[DCTSIZE*7]; 
302  
303 
/* Even part */

304  
305 
tmp10 = tmp0 + tmp3; 
306 
tmp11 = tmp1 + tmp2; 
307 
tmp12 = tmp1  tmp2; 
308 
tmp13 = tmp0  tmp3; 
309  
310 
dataptr[DCTSIZE*0] = tmp10 + tmp11;

311 
dataptr[DCTSIZE*4] = tmp10  tmp11;

312  
313 
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); 
314 
dataptr[DCTSIZE*2] = tmp13 + z1;

315 
dataptr[DCTSIZE*6] = tmp13  z1;

316  
317 
tmp10 = tmp4 + tmp7; 
318 
tmp11 = tmp5 + tmp6; 
319 
tmp12 = tmp5  tmp6; 
320 
tmp13 = tmp4  tmp7; 
321  
322 
dataptr[DCTSIZE*1] = tmp10 + tmp11;

323 
dataptr[DCTSIZE*5] = tmp10  tmp11;

324  
325 
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); 
326 
dataptr[DCTSIZE*3] = tmp13 + z1;

327 
dataptr[DCTSIZE*7] = tmp13  z1;

328  
329 
dataptr++; /* advance pointer to next column */

330 
} 
331 
} 
332  
333  
334 
#undef GLOBAL

335 
#undef CONST_BITS

336 
#undef DESCALE

337 
#undef FIX_0_541196100

338 
#undef FIX_1_306562965
