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) 1991, 1992, 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 the basic inverse-DCT transformation subroutine.
40 *
41 * This implementation is based on an algorithm described in
42 * C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
43 * Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
44 * Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
45 * The primary algorithm described there uses 11 multiplies and 29 adds.
46 * We use their alternate method with 12 multiplies and 32 adds.
47 * The advantage of this method is that no data path contains more than one
48 * multiplication; this allows a very simple and accurate implementation in
49 * scaled fixed-point arithmetic, with a minimal number of shifts.
50 *
51 * I've made lots of modifications to attempt to take advantage of the
52 * sparse nature of the DCT matrices we're getting. Although the logic
53 * is cumbersome, it's straightforward and the resulting code is much
54 * faster.
55 *
56 * A better way to do this would be to pass in the DCT block as a sparse
57 * matrix, perhaps with the difference cases encoded.
58 */
59
60 /**
61 * @file
62 * Independent JPEG Group's LLM idct.
63 */
64
65 #include <stddef.h>
66 #include <stdint.h>
67
68 #include "libavutil/intreadwrite.h"
69
70 #include "dct.h"
71 #include "idctdsp.h"
72
73 #define EIGHT_BIT_SAMPLES
74
75 #define DCTSIZE 8
76 #define DCTSIZE2 64
77
78 #define GLOBAL
79
80 #define RIGHT_SHIFT(x, n) ((x) >> (n))
81
82 typedef int16_t DCTBLOCK[DCTSIZE2];
83
84 #define CONST_BITS 13
85
86 /*
87 * This routine is specialized to the case DCTSIZE = 8.
88 */
89
90 #if DCTSIZE != 8
91 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
92 #endif
93
94
95 /*
96 * A 2-D IDCT can be done by 1-D IDCT on each row followed by 1-D IDCT
97 * on each column. Direct algorithms are also available, but they are
98 * much more complex and seem not to be any faster when reduced to code.
99 *
100 * The poop on this scaling stuff is as follows:
101 *
102 * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
103 * larger than the true IDCT outputs. The final outputs are therefore
104 * a factor of N larger than desired; since N=8 this can be cured by
105 * a simple right shift at the end of the algorithm. The advantage of
106 * this arrangement is that we save two multiplications per 1-D IDCT,
107 * because the y0 and y4 inputs need not be divided by sqrt(N).
108 *
109 * We have to do addition and subtraction of the integer inputs, which
110 * is no problem, and multiplication by fractional constants, which is
111 * a problem to do in integer arithmetic. We multiply all the constants
112 * by CONST_SCALE and convert them to integer constants (thus retaining
113 * CONST_BITS bits of precision in the constants). After doing a
114 * multiplication we have to divide the product by CONST_SCALE, with proper
115 * rounding, to produce the correct output. This division can be done
116 * cheaply as a right shift of CONST_BITS bits. We postpone shifting
117 * as long as possible so that partial sums can be added together with
118 * full fractional precision.
119 *
120 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
121 * they are represented to better-than-integral precision. These outputs
122 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
123 * with the recommended scaling. (To scale up 12-bit sample data further, an
124 * intermediate int32 array would be needed.)
125 *
126 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
127 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
128 * shows that the values given below are the most effective.
129 */
130
131 #ifdef EIGHT_BIT_SAMPLES
132 #define PASS1_BITS 2
133 #else
134 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */
135 #endif
136
137 #define ONE ((int32_t) 1)
138
139 #define CONST_SCALE (ONE << CONST_BITS)
140
141 /* Convert a positive real constant to an integer scaled by CONST_SCALE.
142 * IMPORTANT: if your compiler doesn't do this arithmetic at compile time,
143 * you will pay a significant penalty in run time. In that case, figure
144 * the correct integer constant values and insert them by hand.
145 */
146
147 /* Actually FIX is no longer used, we precomputed them all */
148 #define FIX(x) ((int32_t) ((x) * CONST_SCALE + 0.5))
149
150 /* Descale and correctly round an int32_t value that's scaled by N bits.
151 * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
152 * the fudge factor is correct for either sign of X.
153 */
154
155 #define DESCALE(x,n) RIGHT_SHIFT((x) + (ONE << ((n)-1)), n)
156
157 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
158 * For 8-bit samples with the recommended scaling, all the variable
159 * and constant values involved are no more than 16 bits wide, so a
160 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply;
161 * this provides a useful speedup on many machines.
162 * There is no way to specify a 16x16->32 multiply in portable C, but
163 * some C compilers will do the right thing if you provide the correct
164 * combination of casts.
165 * NB: for 12-bit samples, a full 32-bit multiplication will be needed.
166 */
167
168 #ifdef EIGHT_BIT_SAMPLES
169 #ifdef SHORTxSHORT_32 /* may work if 'int' is 32 bits */
170 #define MULTIPLY(var,const) (((int16_t) (var)) * ((int16_t) (const)))
171 #endif
172 #ifdef SHORTxLCONST_32 /* known to work with Microsoft C 6.0 */
173 #define MULTIPLY(var,const) (((int16_t) (var)) * ((int32_t) (const)))
174 #endif
175 #endif
176
177 #ifndef MULTIPLY /* default definition */
178 #define MULTIPLY(var,const) ((var) * (const))
179 #endif
180
181
182 /*
183 Unlike our decoder where we approximate the FIXes, we need to use exact
184 ones here or successive P-frames will drift too much with Reference frame coding
185 */
186 #define FIX_0_211164243 1730
187 #define FIX_0_275899380 2260
188 #define FIX_0_298631336 2446
189 #define FIX_0_390180644 3196
190 #define FIX_0_509795579 4176
191 #define FIX_0_541196100 4433
192 #define FIX_0_601344887 4926
193 #define FIX_0_765366865 6270
194 #define FIX_0_785694958 6436
195 #define FIX_0_899976223 7373
196 #define FIX_1_061594337 8697
197 #define FIX_1_111140466 9102
198 #define FIX_1_175875602 9633
199 #define FIX_1_306562965 10703
200 #define FIX_1_387039845 11363
201 #define FIX_1_451774981 11893
202 #define FIX_1_501321110 12299
203 #define FIX_1_662939225 13623
204 #define FIX_1_847759065 15137
205 #define FIX_1_961570560 16069
206 #define FIX_2_053119869 16819
207 #define FIX_2_172734803 17799
208 #define FIX_2_562915447 20995
209 #define FIX_3_072711026 25172
210
211 /*
212 * Perform the inverse DCT on one block of coefficients.
213 */
214
215 void ff_j_rev_dct(DCTBLOCK data)
216 {
217 int32_t tmp0, tmp1, tmp2, tmp3;
218 int32_t tmp10, tmp11, tmp12, tmp13;
219 int32_t z1, z2, z3, z4, z5;
220 int32_t d0, d1, d2, d3, d4, d5, d6, d7;
221 register int16_t *dataptr;
222 int rowctr;
223
224 /* Pass 1: process rows. */
225 /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
226 /* furthermore, we scale the results by 2**PASS1_BITS. */
227
228 dataptr = data;
229
230 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
231 /* Due to quantization, we will usually find that many of the input
232 * coefficients are zero, especially the AC terms. We can exploit this
233 * by short-circuiting the IDCT calculation for any row in which all
234 * the AC terms are zero. In that case each output is equal to the
235 * DC coefficient (with scale factor as needed).
236 * With typical images and quantization tables, half or more of the
237 * row DCT calculations can be simplified this way.
238 */
239
240 register uint8_t *idataptr = (uint8_t*)dataptr;
241
242 /* WARNING: we do the same permutation as MMX idct to simplify the
243 video core */
244 d0 = dataptr[0];
245 d2 = dataptr[1];
246 d4 = dataptr[2];
247 d6 = dataptr[3];
248 d1 = dataptr[4];
249 d3 = dataptr[5];
250 d5 = dataptr[6];
251 d7 = dataptr[7];
252
253 if ((d1 | d2 | d3 | d4 | d5 | d6 | d7) == 0) {
254 /* AC terms all zero */
255 if (d0) {
256 /* Compute a 32 bit value to assign. */
257 int16_t dcval = (int16_t) (d0 * (1 << PASS1_BITS));
258 register int v = (dcval & 0xffff) | ((dcval * (1 << 16)) & 0xffff0000);
259
260 AV_WN32A(&idataptr[ 0], v);
261 AV_WN32A(&idataptr[ 4], v);
262 AV_WN32A(&idataptr[ 8], v);
263 AV_WN32A(&idataptr[12], v);
264 }
265
266 dataptr += DCTSIZE; /* advance pointer to next row */
267 continue;
268 }
269
270 /* Even part: reverse the even part of the forward DCT. */
271 /* The rotator is sqrt(2)*c(-6). */
272 {
273 if (d6) {
274 if (d2) {
275 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
276 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
277 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
278 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
279
280 tmp0 = (d0 + d4) * CONST_SCALE;
281 tmp1 = (d0 - d4) * CONST_SCALE;
282
283 tmp10 = tmp0 + tmp3;
284 tmp13 = tmp0 - tmp3;
285 tmp11 = tmp1 + tmp2;
286 tmp12 = tmp1 - tmp2;
287 } else {
288 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
289 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
290 tmp3 = MULTIPLY(d6, FIX_0_541196100);
291
292 tmp0 = (d0 + d4) * CONST_SCALE;
293 tmp1 = (d0 - d4) * CONST_SCALE;
294
295 tmp10 = tmp0 + tmp3;
296 tmp13 = tmp0 - tmp3;
297 tmp11 = tmp1 + tmp2;
298 tmp12 = tmp1 - tmp2;
299 }
300 } else {
301 if (d2) {
302 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
303 tmp2 = MULTIPLY(d2, FIX_0_541196100);
304 tmp3 = MULTIPLY(d2, FIX_1_306562965);
305
306 tmp0 = (d0 + d4) * CONST_SCALE;
307 tmp1 = (d0 - d4) * CONST_SCALE;
308
309 tmp10 = tmp0 + tmp3;
310 tmp13 = tmp0 - tmp3;
311 tmp11 = tmp1 + tmp2;
312 tmp12 = tmp1 - tmp2;
313 } else {
314 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
315 tmp10 = tmp13 = (d0 + d4) * CONST_SCALE;
316 tmp11 = tmp12 = (d0 - d4) * CONST_SCALE;
317 }
318 }
319
320 /* Odd part per figure 8; the matrix is unitary and hence its
321 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
322 */
323
324 if (d7) {
325 if (d5) {
326 if (d3) {
327 if (d1) {
328 /* d1 != 0, d3 != 0, d5 != 0, d7 != 0 */
329 z1 = d7 + d1;
330 z2 = d5 + d3;
331 z3 = d7 + d3;
332 z4 = d5 + d1;
333 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
334
335 tmp0 = MULTIPLY(d7, FIX_0_298631336);
336 tmp1 = MULTIPLY(d5, FIX_2_053119869);
337 tmp2 = MULTIPLY(d3, FIX_3_072711026);
338 tmp3 = MULTIPLY(d1, FIX_1_501321110);
339 z1 = MULTIPLY(-z1, FIX_0_899976223);
340 z2 = MULTIPLY(-z2, FIX_2_562915447);
341 z3 = MULTIPLY(-z3, FIX_1_961570560);
342 z4 = MULTIPLY(-z4, FIX_0_390180644);
343
344 z3 += z5;
345 z4 += z5;
346
347 tmp0 += z1 + z3;
348 tmp1 += z2 + z4;
349 tmp2 += z2 + z3;
350 tmp3 += z1 + z4;
351 } else {
352 /* d1 == 0, d3 != 0, d5 != 0, d7 != 0 */
353 z2 = d5 + d3;
354 z3 = d7 + d3;
355 z5 = MULTIPLY(z3 + d5, FIX_1_175875602);
356
357 tmp0 = MULTIPLY(d7, FIX_0_298631336);
358 tmp1 = MULTIPLY(d5, FIX_2_053119869);
359 tmp2 = MULTIPLY(d3, FIX_3_072711026);
360 z1 = MULTIPLY(-d7, FIX_0_899976223);
361 z2 = MULTIPLY(-z2, FIX_2_562915447);
362 z3 = MULTIPLY(-z3, FIX_1_961570560);
363 z4 = MULTIPLY(-d5, FIX_0_390180644);
364
365 z3 += z5;
366 z4 += z5;
367
368 tmp0 += z1 + z3;
369 tmp1 += z2 + z4;
370 tmp2 += z2 + z3;
371 tmp3 = z1 + z4;
372 }
373 } else {
374 if (d1) {
375 /* d1 != 0, d3 == 0, d5 != 0, d7 != 0 */
376 z1 = d7 + d1;
377 z4 = d5 + d1;
378 z5 = MULTIPLY(d7 + z4, FIX_1_175875602);
379
380 tmp0 = MULTIPLY(d7, FIX_0_298631336);
381 tmp1 = MULTIPLY(d5, FIX_2_053119869);
382 tmp3 = MULTIPLY(d1, FIX_1_501321110);
383 z1 = MULTIPLY(-z1, FIX_0_899976223);
384 z2 = MULTIPLY(-d5, FIX_2_562915447);
385 z3 = MULTIPLY(-d7, FIX_1_961570560);
386 z4 = MULTIPLY(-z4, FIX_0_390180644);
387
388 z3 += z5;
389 z4 += z5;
390
391 tmp0 += z1 + z3;
392 tmp1 += z2 + z4;
393 tmp2 = z2 + z3;
394 tmp3 += z1 + z4;
395 } else {
396 /* d1 == 0, d3 == 0, d5 != 0, d7 != 0 */
397 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
398 z1 = MULTIPLY(-d7, FIX_0_899976223);
399 z3 = MULTIPLY(-d7, FIX_1_961570560);
400 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
401 z2 = MULTIPLY(-d5, FIX_2_562915447);
402 z4 = MULTIPLY(-d5, FIX_0_390180644);
403 z5 = MULTIPLY(d5 + d7, FIX_1_175875602);
404
405 z3 += z5;
406 z4 += z5;
407
408 tmp0 += z3;
409 tmp1 += z4;
410 tmp2 = z2 + z3;
411 tmp3 = z1 + z4;
412 }
413 }
414 } else {
415 if (d3) {
416 if (d1) {
417 /* d1 != 0, d3 != 0, d5 == 0, d7 != 0 */
418 z1 = d7 + d1;
419 z3 = d7 + d3;
420 z5 = MULTIPLY(z3 + d1, FIX_1_175875602);
421
422 tmp0 = MULTIPLY(d7, FIX_0_298631336);
423 tmp2 = MULTIPLY(d3, FIX_3_072711026);
424 tmp3 = MULTIPLY(d1, FIX_1_501321110);
425 z1 = MULTIPLY(-z1, FIX_0_899976223);
426 z2 = MULTIPLY(-d3, FIX_2_562915447);
427 z3 = MULTIPLY(-z3, FIX_1_961570560);
428 z4 = MULTIPLY(-d1, FIX_0_390180644);
429
430 z3 += z5;
431 z4 += z5;
432
433 tmp0 += z1 + z3;
434 tmp1 = z2 + z4;
435 tmp2 += z2 + z3;
436 tmp3 += z1 + z4;
437 } else {
438 /* d1 == 0, d3 != 0, d5 == 0, d7 != 0 */
439 z3 = d7 + d3;
440
441 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
442 z1 = MULTIPLY(-d7, FIX_0_899976223);
443 tmp2 = MULTIPLY(d3, FIX_0_509795579);
444 z2 = MULTIPLY(-d3, FIX_2_562915447);
445 z5 = MULTIPLY(z3, FIX_1_175875602);
446 z3 = MULTIPLY(-z3, FIX_0_785694958);
447
448 tmp0 += z3;
449 tmp1 = z2 + z5;
450 tmp2 += z3;
451 tmp3 = z1 + z5;
452 }
453 } else {
454 if (d1) {
455 /* d1 != 0, d3 == 0, d5 == 0, d7 != 0 */
456 z1 = d7 + d1;
457 z5 = MULTIPLY(z1, FIX_1_175875602);
458
459 z1 = MULTIPLY(z1, FIX_0_275899380);
460 z3 = MULTIPLY(-d7, FIX_1_961570560);
461 tmp0 = MULTIPLY(-d7, FIX_1_662939225);
462 z4 = MULTIPLY(-d1, FIX_0_390180644);
463 tmp3 = MULTIPLY(d1, FIX_1_111140466);
464
465 tmp0 += z1;
466 tmp1 = z4 + z5;
467 tmp2 = z3 + z5;
468 tmp3 += z1;
469 } else {
470 /* d1 == 0, d3 == 0, d5 == 0, d7 != 0 */
471 tmp0 = MULTIPLY(-d7, FIX_1_387039845);
472 tmp1 = MULTIPLY(d7, FIX_1_175875602);
473 tmp2 = MULTIPLY(-d7, FIX_0_785694958);
474 tmp3 = MULTIPLY(d7, FIX_0_275899380);
475 }
476 }
477 }
478 } else {
479 if (d5) {
480 if (d3) {
481 if (d1) {
482 /* d1 != 0, d3 != 0, d5 != 0, d7 == 0 */
483 z2 = d5 + d3;
484 z4 = d5 + d1;
485 z5 = MULTIPLY(d3 + z4, FIX_1_175875602);
486
487 tmp1 = MULTIPLY(d5, FIX_2_053119869);
488 tmp2 = MULTIPLY(d3, FIX_3_072711026);
489 tmp3 = MULTIPLY(d1, FIX_1_501321110);
490 z1 = MULTIPLY(-d1, FIX_0_899976223);
491 z2 = MULTIPLY(-z2, FIX_2_562915447);
492 z3 = MULTIPLY(-d3, FIX_1_961570560);
493 z4 = MULTIPLY(-z4, FIX_0_390180644);
494
495 z3 += z5;
496 z4 += z5;
497
498 tmp0 = z1 + z3;
499 tmp1 += z2 + z4;
500 tmp2 += z2 + z3;
501 tmp3 += z1 + z4;
502 } else {
503 /* d1 == 0, d3 != 0, d5 != 0, d7 == 0 */
504 z2 = d5 + d3;
505
506 z5 = MULTIPLY(z2, FIX_1_175875602);
507 tmp1 = MULTIPLY(d5, FIX_1_662939225);
508 z4 = MULTIPLY(-d5, FIX_0_390180644);
509 z2 = MULTIPLY(-z2, FIX_1_387039845);
510 tmp2 = MULTIPLY(d3, FIX_1_111140466);
511 z3 = MULTIPLY(-d3, FIX_1_961570560);
512
513 tmp0 = z3 + z5;
514 tmp1 += z2;
515 tmp2 += z2;
516 tmp3 = z4 + z5;
517 }
518 } else {
519 if (d1) {
520 /* d1 != 0, d3 == 0, d5 != 0, d7 == 0 */
521 z4 = d5 + d1;
522
523 z5 = MULTIPLY(z4, FIX_1_175875602);
524 z1 = MULTIPLY(-d1, FIX_0_899976223);
525 tmp3 = MULTIPLY(d1, FIX_0_601344887);
526 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
527 z2 = MULTIPLY(-d5, FIX_2_562915447);
528 z4 = MULTIPLY(z4, FIX_0_785694958);
529
530 tmp0 = z1 + z5;
531 tmp1 += z4;
532 tmp2 = z2 + z5;
533 tmp3 += z4;
534 } else {
535 /* d1 == 0, d3 == 0, d5 != 0, d7 == 0 */
536 tmp0 = MULTIPLY(d5, FIX_1_175875602);
537 tmp1 = MULTIPLY(d5, FIX_0_275899380);
538 tmp2 = MULTIPLY(-d5, FIX_1_387039845);
539 tmp3 = MULTIPLY(d5, FIX_0_785694958);
540 }
541 }
542 } else {
543 if (d3) {
544 if (d1) {
545 /* d1 != 0, d3 != 0, d5 == 0, d7 == 0 */
546 z5 = d1 + d3;
547 tmp3 = MULTIPLY(d1, FIX_0_211164243);
548 tmp2 = MULTIPLY(-d3, FIX_1_451774981);
549 z1 = MULTIPLY(d1, FIX_1_061594337);
550 z2 = MULTIPLY(-d3, FIX_2_172734803);
551 z4 = MULTIPLY(z5, FIX_0_785694958);
552 z5 = MULTIPLY(z5, FIX_1_175875602);
553
554 tmp0 = z1 - z4;
555 tmp1 = z2 + z4;
556 tmp2 += z5;
557 tmp3 += z5;
558 } else {
559 /* d1 == 0, d3 != 0, d5 == 0, d7 == 0 */
560 tmp0 = MULTIPLY(-d3, FIX_0_785694958);
561 tmp1 = MULTIPLY(-d3, FIX_1_387039845);
562 tmp2 = MULTIPLY(-d3, FIX_0_275899380);
563 tmp3 = MULTIPLY(d3, FIX_1_175875602);
564 }
565 } else {
566 if (d1) {
567 /* d1 != 0, d3 == 0, d5 == 0, d7 == 0 */
568 tmp0 = MULTIPLY(d1, FIX_0_275899380);
569 tmp1 = MULTIPLY(d1, FIX_0_785694958);
570 tmp2 = MULTIPLY(d1, FIX_1_175875602);
571 tmp3 = MULTIPLY(d1, FIX_1_387039845);
572 } else {
573 /* d1 == 0, d3 == 0, d5 == 0, d7 == 0 */
574 tmp0 = tmp1 = tmp2 = tmp3 = 0;
575 }
576 }
577 }
578 }
579 }
580 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
581
582 dataptr[0] = (int16_t) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
583 dataptr[7] = (int16_t) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
584 dataptr[1] = (int16_t) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
585 dataptr[6] = (int16_t) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
586 dataptr[2] = (int16_t) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
587 dataptr[5] = (int16_t) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
588 dataptr[3] = (int16_t) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
589 dataptr[4] = (int16_t) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
590
591 dataptr += DCTSIZE; /* advance pointer to next row */
592 }
593
594 /* Pass 2: process columns. */
595 /* Note that we must descale the results by a factor of 8 == 2**3, */
596 /* and also undo the PASS1_BITS scaling. */
597
598 dataptr = data;
599 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
600 /* Columns of zeroes can be exploited in the same way as we did with rows.
601 * However, the row calculation has created many nonzero AC terms, so the
602 * simplification applies less often (typically 5% to 10% of the time).
603 * On machines with very fast multiplication, it's possible that the
604 * test takes more time than it's worth. In that case this section
605 * may be commented out.
606 */
607
608 d0 = dataptr[DCTSIZE*0];
609 d1 = dataptr[DCTSIZE*1];
610 d2 = dataptr[DCTSIZE*2];
611 d3 = dataptr[DCTSIZE*3];
612 d4 = dataptr[DCTSIZE*4];
613 d5 = dataptr[DCTSIZE*5];
614 d6 = dataptr[DCTSIZE*6];
615 d7 = dataptr[DCTSIZE*7];
616
617 /* Even part: reverse the even part of the forward DCT. */
618 /* The rotator is sqrt(2)*c(-6). */
619 if (d6) {
620 if (d2) {
621 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
622 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
623 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
624 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
625
626 tmp0 = (d0 + d4) * CONST_SCALE;
627 tmp1 = (d0 - d4) * CONST_SCALE;
628
629 tmp10 = tmp0 + tmp3;
630 tmp13 = tmp0 - tmp3;
631 tmp11 = tmp1 + tmp2;
632 tmp12 = tmp1 - tmp2;
633 } else {
634 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
635 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
636 tmp3 = MULTIPLY(d6, FIX_0_541196100);
637
638 tmp0 = (d0 + d4) * CONST_SCALE;
639 tmp1 = (d0 - d4) * CONST_SCALE;
640
641 tmp10 = tmp0 + tmp3;
642 tmp13 = tmp0 - tmp3;
643 tmp11 = tmp1 + tmp2;
644 tmp12 = tmp1 - tmp2;
645 }
646 } else {
647 if (d2) {
648 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
649 tmp2 = MULTIPLY(d2, FIX_0_541196100);
650 tmp3 = MULTIPLY(d2, FIX_1_306562965);
651
652 tmp0 = (d0 + d4) * CONST_SCALE;
653 tmp1 = (d0 - d4) * CONST_SCALE;
654
655 tmp10 = tmp0 + tmp3;
656 tmp13 = tmp0 - tmp3;
657 tmp11 = tmp1 + tmp2;
658 tmp12 = tmp1 - tmp2;
659 } else {
660 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
661 tmp10 = tmp13 = (d0 + d4) * CONST_SCALE;
662 tmp11 = tmp12 = (d0 - d4) * CONST_SCALE;
663 }
664 }
665
666 /* Odd part per figure 8; the matrix is unitary and hence its
667 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
668 */
669 if (d7) {
670 if (d5) {
671 if (d3) {
672 if (d1) {
673 /* d1 != 0, d3 != 0, d5 != 0, d7 != 0 */
674 z1 = d7 + d1;
675 z2 = d5 + d3;
676 z3 = d7 + d3;
677 z4 = d5 + d1;
678 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
679
680 tmp0 = MULTIPLY(d7, FIX_0_298631336);
681 tmp1 = MULTIPLY(d5, FIX_2_053119869);
682 tmp2 = MULTIPLY(d3, FIX_3_072711026);
683 tmp3 = MULTIPLY(d1, FIX_1_501321110);
684 z1 = MULTIPLY(-z1, FIX_0_899976223);
685 z2 = MULTIPLY(-z2, FIX_2_562915447);
686 z3 = MULTIPLY(-z3, FIX_1_961570560);
687 z4 = MULTIPLY(-z4, FIX_0_390180644);
688
689 z3 += z5;
690 z4 += z5;
691
692 tmp0 += z1 + z3;
693 tmp1 += z2 + z4;
694 tmp2 += z2 + z3;
695 tmp3 += z1 + z4;
696 } else {
697 /* d1 == 0, d3 != 0, d5 != 0, d7 != 0 */
698 z2 = d5 + d3;
699 z3 = d7 + d3;
700 z5 = MULTIPLY(z3 + d5, FIX_1_175875602);
701
702 tmp0 = MULTIPLY(d7, FIX_0_298631336);
703 tmp1 = MULTIPLY(d5, FIX_2_053119869);
704 tmp2 = MULTIPLY(d3, FIX_3_072711026);
705 z1 = MULTIPLY(-d7, FIX_0_899976223);
706 z2 = MULTIPLY(-z2, FIX_2_562915447);
707 z3 = MULTIPLY(-z3, FIX_1_961570560);
708 z4 = MULTIPLY(-d5, FIX_0_390180644);
709
710 z3 += z5;
711 z4 += z5;
712
713 tmp0 += z1 + z3;
714 tmp1 += z2 + z4;
715 tmp2 += z2 + z3;
716 tmp3 = z1 + z4;
717 }
718 } else {
719 if (d1) {
720 /* d1 != 0, d3 == 0, d5 != 0, d7 != 0 */
721 z1 = d7 + d1;
722 z3 = d7;
723 z4 = d5 + d1;
724 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
725
726 tmp0 = MULTIPLY(d7, FIX_0_298631336);
727 tmp1 = MULTIPLY(d5, FIX_2_053119869);
728 tmp3 = MULTIPLY(d1, FIX_1_501321110);
729 z1 = MULTIPLY(-z1, FIX_0_899976223);
730 z2 = MULTIPLY(-d5, FIX_2_562915447);
731 z3 = MULTIPLY(-d7, FIX_1_961570560);
732 z4 = MULTIPLY(-z4, FIX_0_390180644);
733
734 z3 += z5;
735 z4 += z5;
736
737 tmp0 += z1 + z3;
738 tmp1 += z2 + z4;
739 tmp2 = z2 + z3;
740 tmp3 += z1 + z4;
741 } else {
742 /* d1 == 0, d3 == 0, d5 != 0, d7 != 0 */
743 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
744 z1 = MULTIPLY(-d7, FIX_0_899976223);
745 z3 = MULTIPLY(-d7, FIX_1_961570560);
746 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
747 z2 = MULTIPLY(-d5, FIX_2_562915447);
748 z4 = MULTIPLY(-d5, FIX_0_390180644);
749 z5 = MULTIPLY(d5 + d7, FIX_1_175875602);
750
751 z3 += z5;
752 z4 += z5;
753
754 tmp0 += z3;
755 tmp1 += z4;
756 tmp2 = z2 + z3;
757 tmp3 = z1 + z4;
758 }
759 }
760 } else {
761 if (d3) {
762 if (d1) {
763 /* d1 != 0, d3 != 0, d5 == 0, d7 != 0 */
764 z1 = d7 + d1;
765 z3 = d7 + d3;
766 z5 = MULTIPLY(z3 + d1, FIX_1_175875602);
767
768 tmp0 = MULTIPLY(d7, FIX_0_298631336);
769 tmp2 = MULTIPLY(d3, FIX_3_072711026);
770 tmp3 = MULTIPLY(d1, FIX_1_501321110);
771 z1 = MULTIPLY(-z1, FIX_0_899976223);
772 z2 = MULTIPLY(-d3, FIX_2_562915447);
773 z3 = MULTIPLY(-z3, FIX_1_961570560);
774 z4 = MULTIPLY(-d1, FIX_0_390180644);
775
776 z3 += z5;
777 z4 += z5;
778
779 tmp0 += z1 + z3;
780 tmp1 = z2 + z4;
781 tmp2 += z2 + z3;
782 tmp3 += z1 + z4;
783 } else {
784 /* d1 == 0, d3 != 0, d5 == 0, d7 != 0 */
785 z3 = d7 + d3;
786
787 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
788 z1 = MULTIPLY(-d7, FIX_0_899976223);
789 tmp2 = MULTIPLY(d3, FIX_0_509795579);
790 z2 = MULTIPLY(-d3, FIX_2_562915447);
791 z5 = MULTIPLY(z3, FIX_1_175875602);
792 z3 = MULTIPLY(-z3, FIX_0_785694958);
793
794 tmp0 += z3;
795 tmp1 = z2 + z5;
796 tmp2 += z3;
797 tmp3 = z1 + z5;
798 }
799 } else {
800 if (d1) {
801 /* d1 != 0, d3 == 0, d5 == 0, d7 != 0 */
802 z1 = d7 + d1;
803 z5 = MULTIPLY(z1, FIX_1_175875602);
804
805 z1 = MULTIPLY(z1, FIX_0_275899380);
806 z3 = MULTIPLY(-d7, FIX_1_961570560);
807 tmp0 = MULTIPLY(-d7, FIX_1_662939225);
808 z4 = MULTIPLY(-d1, FIX_0_390180644);
809 tmp3 = MULTIPLY(d1, FIX_1_111140466);
810
811 tmp0 += z1;
812 tmp1 = z4 + z5;
813 tmp2 = z3 + z5;
814 tmp3 += z1;
815 } else {
816 /* d1 == 0, d3 == 0, d5 == 0, d7 != 0 */
817 tmp0 = MULTIPLY(-d7, FIX_1_387039845);
818 tmp1 = MULTIPLY(d7, FIX_1_175875602);
819 tmp2 = MULTIPLY(-d7, FIX_0_785694958);
820 tmp3 = MULTIPLY(d7, FIX_0_275899380);
821 }
822 }
823 }
824 } else {
825 if (d5) {
826 if (d3) {
827 if (d1) {
828 /* d1 != 0, d3 != 0, d5 != 0, d7 == 0 */
829 z2 = d5 + d3;
830 z4 = d5 + d1;
831 z5 = MULTIPLY(d3 + z4, FIX_1_175875602);
832
833 tmp1 = MULTIPLY(d5, FIX_2_053119869);
834 tmp2 = MULTIPLY(d3, FIX_3_072711026);
835 tmp3 = MULTIPLY(d1, FIX_1_501321110);
836 z1 = MULTIPLY(-d1, FIX_0_899976223);
837 z2 = MULTIPLY(-z2, FIX_2_562915447);
838 z3 = MULTIPLY(-d3, FIX_1_961570560);
839 z4 = MULTIPLY(-z4, FIX_0_390180644);
840
841 z3 += z5;
842 z4 += z5;
843
844 tmp0 = z1 + z3;
845 tmp1 += z2 + z4;
846 tmp2 += z2 + z3;
847 tmp3 += z1 + z4;
848 } else {
849 /* d1 == 0, d3 != 0, d5 != 0, d7 == 0 */
850 z2 = d5 + d3;
851
852 z5 = MULTIPLY(z2, FIX_1_175875602);
853 tmp1 = MULTIPLY(d5, FIX_1_662939225);
854 z4 = MULTIPLY(-d5, FIX_0_390180644);
855 z2 = MULTIPLY(-z2, FIX_1_387039845);
856 tmp2 = MULTIPLY(d3, FIX_1_111140466);
857 z3 = MULTIPLY(-d3, FIX_1_961570560);
858
859 tmp0 = z3 + z5;
860 tmp1 += z2;
861 tmp2 += z2;
862 tmp3 = z4 + z5;
863 }
864 } else {
865 if (d1) {
866 /* d1 != 0, d3 == 0, d5 != 0, d7 == 0 */
867 z4 = d5 + d1;
868
869 z5 = MULTIPLY(z4, FIX_1_175875602);
870 z1 = MULTIPLY(-d1, FIX_0_899976223);
871 tmp3 = MULTIPLY(d1, FIX_0_601344887);
872 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
873 z2 = MULTIPLY(-d5, FIX_2_562915447);
874 z4 = MULTIPLY(z4, FIX_0_785694958);
875
876 tmp0 = z1 + z5;
877 tmp1 += z4;
878 tmp2 = z2 + z5;
879 tmp3 += z4;
880 } else {
881 /* d1 == 0, d3 == 0, d5 != 0, d7 == 0 */
882 tmp0 = MULTIPLY(d5, FIX_1_175875602);
883 tmp1 = MULTIPLY(d5, FIX_0_275899380);
884 tmp2 = MULTIPLY(-d5, FIX_1_387039845);
885 tmp3 = MULTIPLY(d5, FIX_0_785694958);
886 }
887 }
888 } else {
889 if (d3) {
890 if (d1) {
891 /* d1 != 0, d3 != 0, d5 == 0, d7 == 0 */
892 z5 = d1 + d3;
893 tmp3 = MULTIPLY(d1, FIX_0_211164243);
894 tmp2 = MULTIPLY(-d3, FIX_1_451774981);
895 z1 = MULTIPLY(d1, FIX_1_061594337);
896 z2 = MULTIPLY(-d3, FIX_2_172734803);
897 z4 = MULTIPLY(z5, FIX_0_785694958);
898 z5 = MULTIPLY(z5, FIX_1_175875602);
899
900 tmp0 = z1 - z4;
901 tmp1 = z2 + z4;
902 tmp2 += z5;
903 tmp3 += z5;
904 } else {
905 /* d1 == 0, d3 != 0, d5 == 0, d7 == 0 */
906 tmp0 = MULTIPLY(-d3, FIX_0_785694958);
907 tmp1 = MULTIPLY(-d3, FIX_1_387039845);
908 tmp2 = MULTIPLY(-d3, FIX_0_275899380);
909 tmp3 = MULTIPLY(d3, FIX_1_175875602);
910 }
911 } else {
912 if (d1) {
913 /* d1 != 0, d3 == 0, d5 == 0, d7 == 0 */
914 tmp0 = MULTIPLY(d1, FIX_0_275899380);
915 tmp1 = MULTIPLY(d1, FIX_0_785694958);
916 tmp2 = MULTIPLY(d1, FIX_1_175875602);
917 tmp3 = MULTIPLY(d1, FIX_1_387039845);
918 } else {
919 /* d1 == 0, d3 == 0, d5 == 0, d7 == 0 */
920 tmp0 = tmp1 = tmp2 = tmp3 = 0;
921 }
922 }
923 }
924 }
925
926 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
927
928 dataptr[DCTSIZE*0] = (int16_t) DESCALE(tmp10 + tmp3,
929 CONST_BITS+PASS1_BITS+3);
930 dataptr[DCTSIZE*7] = (int16_t) DESCALE(tmp10 - tmp3,
931 CONST_BITS+PASS1_BITS+3);
932 dataptr[DCTSIZE*1] = (int16_t) DESCALE(tmp11 + tmp2,
933 CONST_BITS+PASS1_BITS+3);
934 dataptr[DCTSIZE*6] = (int16_t) DESCALE(tmp11 - tmp2,
935 CONST_BITS+PASS1_BITS+3);
936 dataptr[DCTSIZE*2] = (int16_t) DESCALE(tmp12 + tmp1,
937 CONST_BITS+PASS1_BITS+3);
938 dataptr[DCTSIZE*5] = (int16_t) DESCALE(tmp12 - tmp1,
939 CONST_BITS+PASS1_BITS+3);
940 dataptr[DCTSIZE*3] = (int16_t) DESCALE(tmp13 + tmp0,
941 CONST_BITS+PASS1_BITS+3);
942 dataptr[DCTSIZE*4] = (int16_t) DESCALE(tmp13 - tmp0,
943 CONST_BITS+PASS1_BITS+3);
944
945 dataptr++; /* advance pointer to next column */
946 }
947 }
948
949 #undef DCTSIZE
950 #define DCTSIZE 4
951 #define DCTSTRIDE 8
952
ff_j_rev_dct4(DCTBLOCK data)953 void ff_j_rev_dct4(DCTBLOCK data)
954 {
955 int32_t tmp0, tmp1, tmp2, tmp3;
956 int32_t tmp10, tmp11, tmp12, tmp13;
957 int32_t z1;
958 int32_t d0, d2, d4, d6;
959 register int16_t *dataptr;
960 int rowctr;
961
962 /* Pass 1: process rows. */
963 /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
964 /* furthermore, we scale the results by 2**PASS1_BITS. */
965
966 data[0] += 4;
967
968 dataptr = data;
969
970 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
971 /* Due to quantization, we will usually find that many of the input
972 * coefficients are zero, especially the AC terms. We can exploit this
973 * by short-circuiting the IDCT calculation for any row in which all
974 * the AC terms are zero. In that case each output is equal to the
975 * DC coefficient (with scale factor as needed).
976 * With typical images and quantization tables, half or more of the
977 * row DCT calculations can be simplified this way.
978 */
979
980 register uint8_t *idataptr = (uint8_t*)dataptr;
981
982 d0 = dataptr[0];
983 d2 = dataptr[1];
984 d4 = dataptr[2];
985 d6 = dataptr[3];
986
987 if ((d2 | d4 | d6) == 0) {
988 /* AC terms all zero */
989 if (d0) {
990 /* Compute a 32 bit value to assign. */
991 int16_t dcval = (int16_t) (d0 << PASS1_BITS);
992 register int v = (dcval & 0xffff) | ((dcval << 16) & 0xffff0000);
993
994 AV_WN32A(&idataptr[0], v);
995 AV_WN32A(&idataptr[4], v);
996 }
997
998 dataptr += DCTSTRIDE; /* advance pointer to next row */
999 continue;
1000 }
1001
1002 /* Even part: reverse the even part of the forward DCT. */
1003 /* The rotator is sqrt(2)*c(-6). */
1004 if (d6) {
1005 if (d2) {
1006 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
1007 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
1008 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
1009 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
1010
1011 tmp0 = (d0 + d4) << CONST_BITS;
1012 tmp1 = (d0 - d4) << CONST_BITS;
1013
1014 tmp10 = tmp0 + tmp3;
1015 tmp13 = tmp0 - tmp3;
1016 tmp11 = tmp1 + tmp2;
1017 tmp12 = tmp1 - tmp2;
1018 } else {
1019 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
1020 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
1021 tmp3 = MULTIPLY(d6, FIX_0_541196100);
1022
1023 tmp0 = (d0 + d4) << CONST_BITS;
1024 tmp1 = (d0 - d4) << CONST_BITS;
1025
1026 tmp10 = tmp0 + tmp3;
1027 tmp13 = tmp0 - tmp3;
1028 tmp11 = tmp1 + tmp2;
1029 tmp12 = tmp1 - tmp2;
1030 }
1031 } else {
1032 if (d2) {
1033 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
1034 tmp2 = MULTIPLY(d2, FIX_0_541196100);
1035 tmp3 = MULTIPLY(d2, FIX_1_306562965);
1036
1037 tmp0 = (d0 + d4) << CONST_BITS;
1038 tmp1 = (d0 - d4) << CONST_BITS;
1039
1040 tmp10 = tmp0 + tmp3;
1041 tmp13 = tmp0 - tmp3;
1042 tmp11 = tmp1 + tmp2;
1043 tmp12 = tmp1 - tmp2;
1044 } else {
1045 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
1046 tmp10 = tmp13 = (d0 + d4) << CONST_BITS;
1047 tmp11 = tmp12 = (d0 - d4) << CONST_BITS;
1048 }
1049 }
1050
1051 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
1052
1053 dataptr[0] = (int16_t) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
1054 dataptr[1] = (int16_t) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
1055 dataptr[2] = (int16_t) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
1056 dataptr[3] = (int16_t) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
1057
1058 dataptr += DCTSTRIDE; /* advance pointer to next row */
1059 }
1060
1061 /* Pass 2: process columns. */
1062 /* Note that we must descale the results by a factor of 8 == 2**3, */
1063 /* and also undo the PASS1_BITS scaling. */
1064
1065 dataptr = data;
1066 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
1067 /* Columns of zeroes can be exploited in the same way as we did with rows.
1068 * However, the row calculation has created many nonzero AC terms, so the
1069 * simplification applies less often (typically 5% to 10% of the time).
1070 * On machines with very fast multiplication, it's possible that the
1071 * test takes more time than it's worth. In that case this section
1072 * may be commented out.
1073 */
1074
1075 d0 = dataptr[DCTSTRIDE*0];
1076 d2 = dataptr[DCTSTRIDE*1];
1077 d4 = dataptr[DCTSTRIDE*2];
1078 d6 = dataptr[DCTSTRIDE*3];
1079
1080 /* Even part: reverse the even part of the forward DCT. */
1081 /* The rotator is sqrt(2)*c(-6). */
1082 if (d6) {
1083 if (d2) {
1084 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
1085 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
1086 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
1087 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
1088
1089 tmp0 = (d0 + d4) << CONST_BITS;
1090 tmp1 = (d0 - d4) << CONST_BITS;
1091
1092 tmp10 = tmp0 + tmp3;
1093 tmp13 = tmp0 - tmp3;
1094 tmp11 = tmp1 + tmp2;
1095 tmp12 = tmp1 - tmp2;
1096 } else {
1097 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
1098 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
1099 tmp3 = MULTIPLY(d6, FIX_0_541196100);
1100
1101 tmp0 = (d0 + d4) << CONST_BITS;
1102 tmp1 = (d0 - d4) << CONST_BITS;
1103
1104 tmp10 = tmp0 + tmp3;
1105 tmp13 = tmp0 - tmp3;
1106 tmp11 = tmp1 + tmp2;
1107 tmp12 = tmp1 - tmp2;
1108 }
1109 } else {
1110 if (d2) {
1111 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
1112 tmp2 = MULTIPLY(d2, FIX_0_541196100);
1113 tmp3 = MULTIPLY(d2, FIX_1_306562965);
1114
1115 tmp0 = (d0 + d4) << CONST_BITS;
1116 tmp1 = (d0 - d4) << CONST_BITS;
1117
1118 tmp10 = tmp0 + tmp3;
1119 tmp13 = tmp0 - tmp3;
1120 tmp11 = tmp1 + tmp2;
1121 tmp12 = tmp1 - tmp2;
1122 } else {
1123 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
1124 tmp10 = tmp13 = (d0 + d4) << CONST_BITS;
1125 tmp11 = tmp12 = (d0 - d4) << CONST_BITS;
1126 }
1127 }
1128
1129 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
1130
1131 dataptr[DCTSTRIDE*0] = tmp10 >> (CONST_BITS+PASS1_BITS+3);
1132 dataptr[DCTSTRIDE*1] = tmp11 >> (CONST_BITS+PASS1_BITS+3);
1133 dataptr[DCTSTRIDE*2] = tmp12 >> (CONST_BITS+PASS1_BITS+3);
1134 dataptr[DCTSTRIDE*3] = tmp13 >> (CONST_BITS+PASS1_BITS+3);
1135
1136 dataptr++; /* advance pointer to next column */
1137 }
1138 }
1139
ff_j_rev_dct2(DCTBLOCK data)1140 void ff_j_rev_dct2(DCTBLOCK data){
1141 int d00, d01, d10, d11;
1142
1143 data[0] += 4;
1144 d00 = data[0+0*DCTSTRIDE] + data[1+0*DCTSTRIDE];
1145 d01 = data[0+0*DCTSTRIDE] - data[1+0*DCTSTRIDE];
1146 d10 = data[0+1*DCTSTRIDE] + data[1+1*DCTSTRIDE];
1147 d11 = data[0+1*DCTSTRIDE] - data[1+1*DCTSTRIDE];
1148
1149 data[0+0*DCTSTRIDE]= (d00 + d10)>>3;
1150 data[1+0*DCTSTRIDE]= (d01 + d11)>>3;
1151 data[0+1*DCTSTRIDE]= (d00 - d10)>>3;
1152 data[1+1*DCTSTRIDE]= (d01 - d11)>>3;
1153 }
1154
ff_j_rev_dct1(DCTBLOCK data)1155 void ff_j_rev_dct1(DCTBLOCK data){
1156 data[0] = (data[0] + 4)>>3;
1157 }
1158
1159 #undef FIX
1160 #undef CONST_BITS
1161
ff_jref_idct_put(uint8_t * dest,ptrdiff_t line_size,int16_t * block)1162 void ff_jref_idct_put(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
1163 {
1164 ff_j_rev_dct(block);
1165 ff_put_pixels_clamped_c(block, dest, line_size);
1166 }
1167
ff_jref_idct_add(uint8_t * dest,ptrdiff_t line_size,int16_t * block)1168 void ff_jref_idct_add(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
1169 {
1170 ff_j_rev_dct(block);
1171 ff_add_pixels_clamped_c(block, dest, line_size);
1172 }
1173