2 * This file is part of the Independent JPEG Group's software.
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.
9 * This software is copyright (C) 1991, 1992, Thomas G. Lane.
10 * All Rights Reserved except as specified below.
12 * Permission is hereby granted to use, copy, modify, and distribute this
13 * software (or portions thereof) for any purpose, without fee, subject to
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.
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
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".
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.
39 * This file contains the basic inverse-DCT transformation subroutine.
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.
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
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.
62 * Independent JPEG Group's LLM idct.
65 #include "libavutil/common.h"
70 #define EIGHT_BIT_SAMPLES
77 #define RIGHT_SHIFT(x, n) ((x) >> (n))
79 typedef int16_t DCTBLOCK[DCTSIZE2];
84 * This routine is specialized to the case DCTSIZE = 8.
88 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
93 * A 2-D IDCT can be done by 1-D IDCT on each row followed by 1-D IDCT
94 * on each column. Direct algorithms are also available, but they are
95 * much more complex and seem not to be any faster when reduced to code.
97 * The poop on this scaling stuff is as follows:
99 * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
100 * larger than the true IDCT outputs. The final outputs are therefore
101 * a factor of N larger than desired; since N=8 this can be cured by
102 * a simple right shift at the end of the algorithm. The advantage of
103 * this arrangement is that we save two multiplications per 1-D IDCT,
104 * because the y0 and y4 inputs need not be divided by sqrt(N).
106 * We have to do addition and subtraction of the integer inputs, which
107 * is no problem, and multiplication by fractional constants, which is
108 * a problem to do in integer arithmetic. We multiply all the constants
109 * by CONST_SCALE and convert them to integer constants (thus retaining
110 * CONST_BITS bits of precision in the constants). After doing a
111 * multiplication we have to divide the product by CONST_SCALE, with proper
112 * rounding, to produce the correct output. This division can be done
113 * cheaply as a right shift of CONST_BITS bits. We postpone shifting
114 * as long as possible so that partial sums can be added together with
115 * full fractional precision.
117 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
118 * they are represented to better-than-integral precision. These outputs
119 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
120 * with the recommended scaling. (To scale up 12-bit sample data further, an
121 * intermediate int32 array would be needed.)
123 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
124 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
125 * shows that the values given below are the most effective.
128 #ifdef EIGHT_BIT_SAMPLES
131 #define PASS1_BITS 1 /* lose a little precision to avoid overflow */
134 #define ONE ((int32_t) 1)
136 #define CONST_SCALE (ONE << CONST_BITS)
138 /* Convert a positive real constant to an integer scaled by CONST_SCALE.
139 * IMPORTANT: if your compiler doesn't do this arithmetic at compile time,
140 * you will pay a significant penalty in run time. In that case, figure
141 * the correct integer constant values and insert them by hand.
144 /* Actually FIX is no longer used, we precomputed them all */
145 #define FIX(x) ((int32_t) ((x) * CONST_SCALE + 0.5))
147 /* Descale and correctly round an int32_t value that's scaled by N bits.
148 * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
149 * the fudge factor is correct for either sign of X.
152 #define DESCALE(x,n) RIGHT_SHIFT((x) + (ONE << ((n)-1)), n)
154 /* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
155 * For 8-bit samples with the recommended scaling, all the variable
156 * and constant values involved are no more than 16 bits wide, so a
157 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply;
158 * this provides a useful speedup on many machines.
159 * There is no way to specify a 16x16->32 multiply in portable C, but
160 * some C compilers will do the right thing if you provide the correct
161 * combination of casts.
162 * NB: for 12-bit samples, a full 32-bit multiplication will be needed.
165 #ifdef EIGHT_BIT_SAMPLES
166 #ifdef SHORTxSHORT_32 /* may work if 'int' is 32 bits */
167 #define MULTIPLY(var,const) (((int16_t) (var)) * ((int16_t) (const)))
169 #ifdef SHORTxLCONST_32 /* known to work with Microsoft C 6.0 */
170 #define MULTIPLY(var,const) (((int16_t) (var)) * ((int32_t) (const)))
174 #ifndef MULTIPLY /* default definition */
175 #define MULTIPLY(var,const) ((var) * (const))
180 Unlike our decoder where we approximate the FIXes, we need to use exact
181 ones here or successive P-frames will drift too much with Reference frame coding
183 #define FIX_0_211164243 1730
184 #define FIX_0_275899380 2260
185 #define FIX_0_298631336 2446
186 #define FIX_0_390180644 3196
187 #define FIX_0_509795579 4176
188 #define FIX_0_541196100 4433
189 #define FIX_0_601344887 4926
190 #define FIX_0_765366865 6270
191 #define FIX_0_785694958 6436
192 #define FIX_0_899976223 7373
193 #define FIX_1_061594337 8697
194 #define FIX_1_111140466 9102
195 #define FIX_1_175875602 9633
196 #define FIX_1_306562965 10703
197 #define FIX_1_387039845 11363
198 #define FIX_1_451774981 11893
199 #define FIX_1_501321110 12299
200 #define FIX_1_662939225 13623
201 #define FIX_1_847759065 15137
202 #define FIX_1_961570560 16069
203 #define FIX_2_053119869 16819
204 #define FIX_2_172734803 17799
205 #define FIX_2_562915447 20995
206 #define FIX_3_072711026 25172
209 * Perform the inverse DCT on one block of coefficients.
212 void ff_j_rev_dct(DCTBLOCK data)
214 int32_t tmp0, tmp1, tmp2, tmp3;
215 int32_t tmp10, tmp11, tmp12, tmp13;
216 int32_t z1, z2, z3, z4, z5;
217 int32_t d0, d1, d2, d3, d4, d5, d6, d7;
218 register int16_t *dataptr;
221 /* Pass 1: process rows. */
222 /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
223 /* furthermore, we scale the results by 2**PASS1_BITS. */
227 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
228 /* Due to quantization, we will usually find that many of the input
229 * coefficients are zero, especially the AC terms. We can exploit this
230 * by short-circuiting the IDCT calculation for any row in which all
231 * the AC terms are zero. In that case each output is equal to the
232 * DC coefficient (with scale factor as needed).
233 * With typical images and quantization tables, half or more of the
234 * row DCT calculations can be simplified this way.
237 register int *idataptr = (int*)dataptr;
239 /* WARNING: we do the same permutation as MMX idct to simplify the
250 if ((d1 | d2 | d3 | d4 | d5 | d6 | d7) == 0) {
251 /* AC terms all zero */
253 /* Compute a 32 bit value to assign. */
254 int16_t dcval = (int16_t) (d0 * (1 << PASS1_BITS));
255 register int v = (dcval & 0xffff) | ((dcval * (1 << 16)) & 0xffff0000);
263 dataptr += DCTSIZE; /* advance pointer to next row */
267 /* Even part: reverse the even part of the forward DCT. */
268 /* The rotator is sqrt(2)*c(-6). */
272 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
273 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
274 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
275 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
277 tmp0 = (d0 + d4) * CONST_SCALE;
278 tmp1 = (d0 - d4) * CONST_SCALE;
285 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
286 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
287 tmp3 = MULTIPLY(d6, FIX_0_541196100);
289 tmp0 = (d0 + d4) * CONST_SCALE;
290 tmp1 = (d0 - d4) * CONST_SCALE;
299 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
300 tmp2 = MULTIPLY(d2, FIX_0_541196100);
301 tmp3 = MULTIPLY(d2, FIX_1_306562965);
303 tmp0 = (d0 + d4) * CONST_SCALE;
304 tmp1 = (d0 - d4) * CONST_SCALE;
311 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
312 tmp10 = tmp13 = (d0 + d4) * CONST_SCALE;
313 tmp11 = tmp12 = (d0 - d4) * CONST_SCALE;
317 /* Odd part per figure 8; the matrix is unitary and hence its
318 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
325 /* d1 != 0, d3 != 0, d5 != 0, d7 != 0 */
330 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
332 tmp0 = MULTIPLY(d7, FIX_0_298631336);
333 tmp1 = MULTIPLY(d5, FIX_2_053119869);
334 tmp2 = MULTIPLY(d3, FIX_3_072711026);
335 tmp3 = MULTIPLY(d1, FIX_1_501321110);
336 z1 = MULTIPLY(-z1, FIX_0_899976223);
337 z2 = MULTIPLY(-z2, FIX_2_562915447);
338 z3 = MULTIPLY(-z3, FIX_1_961570560);
339 z4 = MULTIPLY(-z4, FIX_0_390180644);
349 /* d1 == 0, d3 != 0, d5 != 0, d7 != 0 */
352 z5 = MULTIPLY(z3 + d5, FIX_1_175875602);
354 tmp0 = MULTIPLY(d7, FIX_0_298631336);
355 tmp1 = MULTIPLY(d5, FIX_2_053119869);
356 tmp2 = MULTIPLY(d3, FIX_3_072711026);
357 z1 = MULTIPLY(-d7, FIX_0_899976223);
358 z2 = MULTIPLY(-z2, FIX_2_562915447);
359 z3 = MULTIPLY(-z3, FIX_1_961570560);
360 z4 = MULTIPLY(-d5, FIX_0_390180644);
372 /* d1 != 0, d3 == 0, d5 != 0, d7 != 0 */
375 z5 = MULTIPLY(d7 + z4, FIX_1_175875602);
377 tmp0 = MULTIPLY(d7, FIX_0_298631336);
378 tmp1 = MULTIPLY(d5, FIX_2_053119869);
379 tmp3 = MULTIPLY(d1, FIX_1_501321110);
380 z1 = MULTIPLY(-z1, FIX_0_899976223);
381 z2 = MULTIPLY(-d5, FIX_2_562915447);
382 z3 = MULTIPLY(-d7, FIX_1_961570560);
383 z4 = MULTIPLY(-z4, FIX_0_390180644);
393 /* d1 == 0, d3 == 0, d5 != 0, d7 != 0 */
394 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
395 z1 = MULTIPLY(-d7, FIX_0_899976223);
396 z3 = MULTIPLY(-d7, FIX_1_961570560);
397 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
398 z2 = MULTIPLY(-d5, FIX_2_562915447);
399 z4 = MULTIPLY(-d5, FIX_0_390180644);
400 z5 = MULTIPLY(d5 + d7, FIX_1_175875602);
414 /* d1 != 0, d3 != 0, d5 == 0, d7 != 0 */
417 z5 = MULTIPLY(z3 + d1, FIX_1_175875602);
419 tmp0 = MULTIPLY(d7, FIX_0_298631336);
420 tmp2 = MULTIPLY(d3, FIX_3_072711026);
421 tmp3 = MULTIPLY(d1, FIX_1_501321110);
422 z1 = MULTIPLY(-z1, FIX_0_899976223);
423 z2 = MULTIPLY(-d3, FIX_2_562915447);
424 z3 = MULTIPLY(-z3, FIX_1_961570560);
425 z4 = MULTIPLY(-d1, FIX_0_390180644);
435 /* d1 == 0, d3 != 0, d5 == 0, d7 != 0 */
438 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
439 z1 = MULTIPLY(-d7, FIX_0_899976223);
440 tmp2 = MULTIPLY(d3, FIX_0_509795579);
441 z2 = MULTIPLY(-d3, FIX_2_562915447);
442 z5 = MULTIPLY(z3, FIX_1_175875602);
443 z3 = MULTIPLY(-z3, FIX_0_785694958);
452 /* d1 != 0, d3 == 0, d5 == 0, d7 != 0 */
454 z5 = MULTIPLY(z1, FIX_1_175875602);
456 z1 = MULTIPLY(z1, FIX_0_275899380);
457 z3 = MULTIPLY(-d7, FIX_1_961570560);
458 tmp0 = MULTIPLY(-d7, FIX_1_662939225);
459 z4 = MULTIPLY(-d1, FIX_0_390180644);
460 tmp3 = MULTIPLY(d1, FIX_1_111140466);
467 /* d1 == 0, d3 == 0, d5 == 0, d7 != 0 */
468 tmp0 = MULTIPLY(-d7, FIX_1_387039845);
469 tmp1 = MULTIPLY(d7, FIX_1_175875602);
470 tmp2 = MULTIPLY(-d7, FIX_0_785694958);
471 tmp3 = MULTIPLY(d7, FIX_0_275899380);
479 /* d1 != 0, d3 != 0, d5 != 0, d7 == 0 */
482 z5 = MULTIPLY(d3 + z4, FIX_1_175875602);
484 tmp1 = MULTIPLY(d5, FIX_2_053119869);
485 tmp2 = MULTIPLY(d3, FIX_3_072711026);
486 tmp3 = MULTIPLY(d1, FIX_1_501321110);
487 z1 = MULTIPLY(-d1, FIX_0_899976223);
488 z2 = MULTIPLY(-z2, FIX_2_562915447);
489 z3 = MULTIPLY(-d3, FIX_1_961570560);
490 z4 = MULTIPLY(-z4, FIX_0_390180644);
500 /* d1 == 0, d3 != 0, d5 != 0, d7 == 0 */
503 z5 = MULTIPLY(z2, FIX_1_175875602);
504 tmp1 = MULTIPLY(d5, FIX_1_662939225);
505 z4 = MULTIPLY(-d5, FIX_0_390180644);
506 z2 = MULTIPLY(-z2, FIX_1_387039845);
507 tmp2 = MULTIPLY(d3, FIX_1_111140466);
508 z3 = MULTIPLY(-d3, FIX_1_961570560);
517 /* d1 != 0, d3 == 0, d5 != 0, d7 == 0 */
520 z5 = MULTIPLY(z4, FIX_1_175875602);
521 z1 = MULTIPLY(-d1, FIX_0_899976223);
522 tmp3 = MULTIPLY(d1, FIX_0_601344887);
523 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
524 z2 = MULTIPLY(-d5, FIX_2_562915447);
525 z4 = MULTIPLY(z4, FIX_0_785694958);
532 /* d1 == 0, d3 == 0, d5 != 0, d7 == 0 */
533 tmp0 = MULTIPLY(d5, FIX_1_175875602);
534 tmp1 = MULTIPLY(d5, FIX_0_275899380);
535 tmp2 = MULTIPLY(-d5, FIX_1_387039845);
536 tmp3 = MULTIPLY(d5, FIX_0_785694958);
542 /* d1 != 0, d3 != 0, d5 == 0, d7 == 0 */
544 tmp3 = MULTIPLY(d1, FIX_0_211164243);
545 tmp2 = MULTIPLY(-d3, FIX_1_451774981);
546 z1 = MULTIPLY(d1, FIX_1_061594337);
547 z2 = MULTIPLY(-d3, FIX_2_172734803);
548 z4 = MULTIPLY(z5, FIX_0_785694958);
549 z5 = MULTIPLY(z5, FIX_1_175875602);
556 /* d1 == 0, d3 != 0, d5 == 0, d7 == 0 */
557 tmp0 = MULTIPLY(-d3, FIX_0_785694958);
558 tmp1 = MULTIPLY(-d3, FIX_1_387039845);
559 tmp2 = MULTIPLY(-d3, FIX_0_275899380);
560 tmp3 = MULTIPLY(d3, FIX_1_175875602);
564 /* d1 != 0, d3 == 0, d5 == 0, d7 == 0 */
565 tmp0 = MULTIPLY(d1, FIX_0_275899380);
566 tmp1 = MULTIPLY(d1, FIX_0_785694958);
567 tmp2 = MULTIPLY(d1, FIX_1_175875602);
568 tmp3 = MULTIPLY(d1, FIX_1_387039845);
570 /* d1 == 0, d3 == 0, d5 == 0, d7 == 0 */
571 tmp0 = tmp1 = tmp2 = tmp3 = 0;
577 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
579 dataptr[0] = (int16_t) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
580 dataptr[7] = (int16_t) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
581 dataptr[1] = (int16_t) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
582 dataptr[6] = (int16_t) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
583 dataptr[2] = (int16_t) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
584 dataptr[5] = (int16_t) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
585 dataptr[3] = (int16_t) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
586 dataptr[4] = (int16_t) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
588 dataptr += DCTSIZE; /* advance pointer to next row */
591 /* Pass 2: process columns. */
592 /* Note that we must descale the results by a factor of 8 == 2**3, */
593 /* and also undo the PASS1_BITS scaling. */
596 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
597 /* Columns of zeroes can be exploited in the same way as we did with rows.
598 * However, the row calculation has created many nonzero AC terms, so the
599 * simplification applies less often (typically 5% to 10% of the time).
600 * On machines with very fast multiplication, it's possible that the
601 * test takes more time than it's worth. In that case this section
602 * may be commented out.
605 d0 = dataptr[DCTSIZE*0];
606 d1 = dataptr[DCTSIZE*1];
607 d2 = dataptr[DCTSIZE*2];
608 d3 = dataptr[DCTSIZE*3];
609 d4 = dataptr[DCTSIZE*4];
610 d5 = dataptr[DCTSIZE*5];
611 d6 = dataptr[DCTSIZE*6];
612 d7 = dataptr[DCTSIZE*7];
614 /* Even part: reverse the even part of the forward DCT. */
615 /* The rotator is sqrt(2)*c(-6). */
618 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
619 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
620 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
621 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
623 tmp0 = (d0 + d4) * CONST_SCALE;
624 tmp1 = (d0 - d4) * CONST_SCALE;
631 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
632 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
633 tmp3 = MULTIPLY(d6, FIX_0_541196100);
635 tmp0 = (d0 + d4) * CONST_SCALE;
636 tmp1 = (d0 - d4) * CONST_SCALE;
645 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
646 tmp2 = MULTIPLY(d2, FIX_0_541196100);
647 tmp3 = MULTIPLY(d2, FIX_1_306562965);
649 tmp0 = (d0 + d4) * CONST_SCALE;
650 tmp1 = (d0 - d4) * CONST_SCALE;
657 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
658 tmp10 = tmp13 = (d0 + d4) * CONST_SCALE;
659 tmp11 = tmp12 = (d0 - d4) * CONST_SCALE;
663 /* Odd part per figure 8; the matrix is unitary and hence its
664 * transpose is its inverse. i0..i3 are y7,y5,y3,y1 respectively.
670 /* d1 != 0, d3 != 0, d5 != 0, d7 != 0 */
675 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
677 tmp0 = MULTIPLY(d7, FIX_0_298631336);
678 tmp1 = MULTIPLY(d5, FIX_2_053119869);
679 tmp2 = MULTIPLY(d3, FIX_3_072711026);
680 tmp3 = MULTIPLY(d1, FIX_1_501321110);
681 z1 = MULTIPLY(-z1, FIX_0_899976223);
682 z2 = MULTIPLY(-z2, FIX_2_562915447);
683 z3 = MULTIPLY(-z3, FIX_1_961570560);
684 z4 = MULTIPLY(-z4, FIX_0_390180644);
694 /* d1 == 0, d3 != 0, d5 != 0, d7 != 0 */
697 z5 = MULTIPLY(z3 + d5, FIX_1_175875602);
699 tmp0 = MULTIPLY(d7, FIX_0_298631336);
700 tmp1 = MULTIPLY(d5, FIX_2_053119869);
701 tmp2 = MULTIPLY(d3, FIX_3_072711026);
702 z1 = MULTIPLY(-d7, FIX_0_899976223);
703 z2 = MULTIPLY(-z2, FIX_2_562915447);
704 z3 = MULTIPLY(-z3, FIX_1_961570560);
705 z4 = MULTIPLY(-d5, FIX_0_390180644);
717 /* d1 != 0, d3 == 0, d5 != 0, d7 != 0 */
721 z5 = MULTIPLY(z3 + z4, FIX_1_175875602);
723 tmp0 = MULTIPLY(d7, FIX_0_298631336);
724 tmp1 = MULTIPLY(d5, FIX_2_053119869);
725 tmp3 = MULTIPLY(d1, FIX_1_501321110);
726 z1 = MULTIPLY(-z1, FIX_0_899976223);
727 z2 = MULTIPLY(-d5, FIX_2_562915447);
728 z3 = MULTIPLY(-d7, FIX_1_961570560);
729 z4 = MULTIPLY(-z4, FIX_0_390180644);
739 /* d1 == 0, d3 == 0, d5 != 0, d7 != 0 */
740 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
741 z1 = MULTIPLY(-d7, FIX_0_899976223);
742 z3 = MULTIPLY(-d7, FIX_1_961570560);
743 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
744 z2 = MULTIPLY(-d5, FIX_2_562915447);
745 z4 = MULTIPLY(-d5, FIX_0_390180644);
746 z5 = MULTIPLY(d5 + d7, FIX_1_175875602);
760 /* d1 != 0, d3 != 0, d5 == 0, d7 != 0 */
763 z5 = MULTIPLY(z3 + d1, FIX_1_175875602);
765 tmp0 = MULTIPLY(d7, FIX_0_298631336);
766 tmp2 = MULTIPLY(d3, FIX_3_072711026);
767 tmp3 = MULTIPLY(d1, FIX_1_501321110);
768 z1 = MULTIPLY(-z1, FIX_0_899976223);
769 z2 = MULTIPLY(-d3, FIX_2_562915447);
770 z3 = MULTIPLY(-z3, FIX_1_961570560);
771 z4 = MULTIPLY(-d1, FIX_0_390180644);
781 /* d1 == 0, d3 != 0, d5 == 0, d7 != 0 */
784 tmp0 = MULTIPLY(-d7, FIX_0_601344887);
785 z1 = MULTIPLY(-d7, FIX_0_899976223);
786 tmp2 = MULTIPLY(d3, FIX_0_509795579);
787 z2 = MULTIPLY(-d3, FIX_2_562915447);
788 z5 = MULTIPLY(z3, FIX_1_175875602);
789 z3 = MULTIPLY(-z3, FIX_0_785694958);
798 /* d1 != 0, d3 == 0, d5 == 0, d7 != 0 */
800 z5 = MULTIPLY(z1, FIX_1_175875602);
802 z1 = MULTIPLY(z1, FIX_0_275899380);
803 z3 = MULTIPLY(-d7, FIX_1_961570560);
804 tmp0 = MULTIPLY(-d7, FIX_1_662939225);
805 z4 = MULTIPLY(-d1, FIX_0_390180644);
806 tmp3 = MULTIPLY(d1, FIX_1_111140466);
813 /* d1 == 0, d3 == 0, d5 == 0, d7 != 0 */
814 tmp0 = MULTIPLY(-d7, FIX_1_387039845);
815 tmp1 = MULTIPLY(d7, FIX_1_175875602);
816 tmp2 = MULTIPLY(-d7, FIX_0_785694958);
817 tmp3 = MULTIPLY(d7, FIX_0_275899380);
825 /* d1 != 0, d3 != 0, d5 != 0, d7 == 0 */
828 z5 = MULTIPLY(d3 + z4, FIX_1_175875602);
830 tmp1 = MULTIPLY(d5, FIX_2_053119869);
831 tmp2 = MULTIPLY(d3, FIX_3_072711026);
832 tmp3 = MULTIPLY(d1, FIX_1_501321110);
833 z1 = MULTIPLY(-d1, FIX_0_899976223);
834 z2 = MULTIPLY(-z2, FIX_2_562915447);
835 z3 = MULTIPLY(-d3, FIX_1_961570560);
836 z4 = MULTIPLY(-z4, FIX_0_390180644);
846 /* d1 == 0, d3 != 0, d5 != 0, d7 == 0 */
849 z5 = MULTIPLY(z2, FIX_1_175875602);
850 tmp1 = MULTIPLY(d5, FIX_1_662939225);
851 z4 = MULTIPLY(-d5, FIX_0_390180644);
852 z2 = MULTIPLY(-z2, FIX_1_387039845);
853 tmp2 = MULTIPLY(d3, FIX_1_111140466);
854 z3 = MULTIPLY(-d3, FIX_1_961570560);
863 /* d1 != 0, d3 == 0, d5 != 0, d7 == 0 */
866 z5 = MULTIPLY(z4, FIX_1_175875602);
867 z1 = MULTIPLY(-d1, FIX_0_899976223);
868 tmp3 = MULTIPLY(d1, FIX_0_601344887);
869 tmp1 = MULTIPLY(-d5, FIX_0_509795579);
870 z2 = MULTIPLY(-d5, FIX_2_562915447);
871 z4 = MULTIPLY(z4, FIX_0_785694958);
878 /* d1 == 0, d3 == 0, d5 != 0, d7 == 0 */
879 tmp0 = MULTIPLY(d5, FIX_1_175875602);
880 tmp1 = MULTIPLY(d5, FIX_0_275899380);
881 tmp2 = MULTIPLY(-d5, FIX_1_387039845);
882 tmp3 = MULTIPLY(d5, FIX_0_785694958);
888 /* d1 != 0, d3 != 0, d5 == 0, d7 == 0 */
890 tmp3 = MULTIPLY(d1, FIX_0_211164243);
891 tmp2 = MULTIPLY(-d3, FIX_1_451774981);
892 z1 = MULTIPLY(d1, FIX_1_061594337);
893 z2 = MULTIPLY(-d3, FIX_2_172734803);
894 z4 = MULTIPLY(z5, FIX_0_785694958);
895 z5 = MULTIPLY(z5, FIX_1_175875602);
902 /* d1 == 0, d3 != 0, d5 == 0, d7 == 0 */
903 tmp0 = MULTIPLY(-d3, FIX_0_785694958);
904 tmp1 = MULTIPLY(-d3, FIX_1_387039845);
905 tmp2 = MULTIPLY(-d3, FIX_0_275899380);
906 tmp3 = MULTIPLY(d3, FIX_1_175875602);
910 /* d1 != 0, d3 == 0, d5 == 0, d7 == 0 */
911 tmp0 = MULTIPLY(d1, FIX_0_275899380);
912 tmp1 = MULTIPLY(d1, FIX_0_785694958);
913 tmp2 = MULTIPLY(d1, FIX_1_175875602);
914 tmp3 = MULTIPLY(d1, FIX_1_387039845);
916 /* d1 == 0, d3 == 0, d5 == 0, d7 == 0 */
917 tmp0 = tmp1 = tmp2 = tmp3 = 0;
923 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
925 dataptr[DCTSIZE*0] = (int16_t) DESCALE(tmp10 + tmp3,
926 CONST_BITS+PASS1_BITS+3);
927 dataptr[DCTSIZE*7] = (int16_t) DESCALE(tmp10 - tmp3,
928 CONST_BITS+PASS1_BITS+3);
929 dataptr[DCTSIZE*1] = (int16_t) DESCALE(tmp11 + tmp2,
930 CONST_BITS+PASS1_BITS+3);
931 dataptr[DCTSIZE*6] = (int16_t) DESCALE(tmp11 - tmp2,
932 CONST_BITS+PASS1_BITS+3);
933 dataptr[DCTSIZE*2] = (int16_t) DESCALE(tmp12 + tmp1,
934 CONST_BITS+PASS1_BITS+3);
935 dataptr[DCTSIZE*5] = (int16_t) DESCALE(tmp12 - tmp1,
936 CONST_BITS+PASS1_BITS+3);
937 dataptr[DCTSIZE*3] = (int16_t) DESCALE(tmp13 + tmp0,
938 CONST_BITS+PASS1_BITS+3);
939 dataptr[DCTSIZE*4] = (int16_t) DESCALE(tmp13 - tmp0,
940 CONST_BITS+PASS1_BITS+3);
942 dataptr++; /* advance pointer to next column */
950 void ff_j_rev_dct4(DCTBLOCK data)
952 int32_t tmp0, tmp1, tmp2, tmp3;
953 int32_t tmp10, tmp11, tmp12, tmp13;
955 int32_t d0, d2, d4, d6;
956 register int16_t *dataptr;
959 /* Pass 1: process rows. */
960 /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
961 /* furthermore, we scale the results by 2**PASS1_BITS. */
967 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
968 /* Due to quantization, we will usually find that many of the input
969 * coefficients are zero, especially the AC terms. We can exploit this
970 * by short-circuiting the IDCT calculation for any row in which all
971 * the AC terms are zero. In that case each output is equal to the
972 * DC coefficient (with scale factor as needed).
973 * With typical images and quantization tables, half or more of the
974 * row DCT calculations can be simplified this way.
977 register int *idataptr = (int*)dataptr;
984 if ((d2 | d4 | d6) == 0) {
985 /* AC terms all zero */
987 /* Compute a 32 bit value to assign. */
988 int16_t dcval = (int16_t) (d0 << PASS1_BITS);
989 register int v = (dcval & 0xffff) | ((dcval << 16) & 0xffff0000);
995 dataptr += DCTSTRIDE; /* advance pointer to next row */
999 /* Even part: reverse the even part of the forward DCT. */
1000 /* The rotator is sqrt(2)*c(-6). */
1003 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
1004 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
1005 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
1006 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
1008 tmp0 = (d0 + d4) << CONST_BITS;
1009 tmp1 = (d0 - d4) << CONST_BITS;
1011 tmp10 = tmp0 + tmp3;
1012 tmp13 = tmp0 - tmp3;
1013 tmp11 = tmp1 + tmp2;
1014 tmp12 = tmp1 - tmp2;
1016 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
1017 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
1018 tmp3 = MULTIPLY(d6, FIX_0_541196100);
1020 tmp0 = (d0 + d4) << CONST_BITS;
1021 tmp1 = (d0 - d4) << CONST_BITS;
1023 tmp10 = tmp0 + tmp3;
1024 tmp13 = tmp0 - tmp3;
1025 tmp11 = tmp1 + tmp2;
1026 tmp12 = tmp1 - tmp2;
1030 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
1031 tmp2 = MULTIPLY(d2, FIX_0_541196100);
1032 tmp3 = MULTIPLY(d2, FIX_1_306562965);
1034 tmp0 = (d0 + d4) << CONST_BITS;
1035 tmp1 = (d0 - d4) << CONST_BITS;
1037 tmp10 = tmp0 + tmp3;
1038 tmp13 = tmp0 - tmp3;
1039 tmp11 = tmp1 + tmp2;
1040 tmp12 = tmp1 - tmp2;
1042 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
1043 tmp10 = tmp13 = (d0 + d4) << CONST_BITS;
1044 tmp11 = tmp12 = (d0 - d4) << CONST_BITS;
1048 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
1050 dataptr[0] = (int16_t) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
1051 dataptr[1] = (int16_t) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
1052 dataptr[2] = (int16_t) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
1053 dataptr[3] = (int16_t) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
1055 dataptr += DCTSTRIDE; /* advance pointer to next row */
1058 /* Pass 2: process columns. */
1059 /* Note that we must descale the results by a factor of 8 == 2**3, */
1060 /* and also undo the PASS1_BITS scaling. */
1063 for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) {
1064 /* Columns of zeroes can be exploited in the same way as we did with rows.
1065 * However, the row calculation has created many nonzero AC terms, so the
1066 * simplification applies less often (typically 5% to 10% of the time).
1067 * On machines with very fast multiplication, it's possible that the
1068 * test takes more time than it's worth. In that case this section
1069 * may be commented out.
1072 d0 = dataptr[DCTSTRIDE*0];
1073 d2 = dataptr[DCTSTRIDE*1];
1074 d4 = dataptr[DCTSTRIDE*2];
1075 d6 = dataptr[DCTSTRIDE*3];
1077 /* Even part: reverse the even part of the forward DCT. */
1078 /* The rotator is sqrt(2)*c(-6). */
1081 /* d0 != 0, d2 != 0, d4 != 0, d6 != 0 */
1082 z1 = MULTIPLY(d2 + d6, FIX_0_541196100);
1083 tmp2 = z1 + MULTIPLY(-d6, FIX_1_847759065);
1084 tmp3 = z1 + MULTIPLY(d2, FIX_0_765366865);
1086 tmp0 = (d0 + d4) << CONST_BITS;
1087 tmp1 = (d0 - d4) << CONST_BITS;
1089 tmp10 = tmp0 + tmp3;
1090 tmp13 = tmp0 - tmp3;
1091 tmp11 = tmp1 + tmp2;
1092 tmp12 = tmp1 - tmp2;
1094 /* d0 != 0, d2 == 0, d4 != 0, d6 != 0 */
1095 tmp2 = MULTIPLY(-d6, FIX_1_306562965);
1096 tmp3 = MULTIPLY(d6, FIX_0_541196100);
1098 tmp0 = (d0 + d4) << CONST_BITS;
1099 tmp1 = (d0 - d4) << CONST_BITS;
1101 tmp10 = tmp0 + tmp3;
1102 tmp13 = tmp0 - tmp3;
1103 tmp11 = tmp1 + tmp2;
1104 tmp12 = tmp1 - tmp2;
1108 /* d0 != 0, d2 != 0, d4 != 0, d6 == 0 */
1109 tmp2 = MULTIPLY(d2, FIX_0_541196100);
1110 tmp3 = MULTIPLY(d2, FIX_1_306562965);
1112 tmp0 = (d0 + d4) << CONST_BITS;
1113 tmp1 = (d0 - d4) << CONST_BITS;
1115 tmp10 = tmp0 + tmp3;
1116 tmp13 = tmp0 - tmp3;
1117 tmp11 = tmp1 + tmp2;
1118 tmp12 = tmp1 - tmp2;
1120 /* d0 != 0, d2 == 0, d4 != 0, d6 == 0 */
1121 tmp10 = tmp13 = (d0 + d4) << CONST_BITS;
1122 tmp11 = tmp12 = (d0 - d4) << CONST_BITS;
1126 /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
1128 dataptr[DCTSTRIDE*0] = tmp10 >> (CONST_BITS+PASS1_BITS+3);
1129 dataptr[DCTSTRIDE*1] = tmp11 >> (CONST_BITS+PASS1_BITS+3);
1130 dataptr[DCTSTRIDE*2] = tmp12 >> (CONST_BITS+PASS1_BITS+3);
1131 dataptr[DCTSTRIDE*3] = tmp13 >> (CONST_BITS+PASS1_BITS+3);
1133 dataptr++; /* advance pointer to next column */
1137 void ff_j_rev_dct2(DCTBLOCK data){
1138 int d00, d01, d10, d11;
1141 d00 = data[0+0*DCTSTRIDE] + data[1+0*DCTSTRIDE];
1142 d01 = data[0+0*DCTSTRIDE] - data[1+0*DCTSTRIDE];
1143 d10 = data[0+1*DCTSTRIDE] + data[1+1*DCTSTRIDE];
1144 d11 = data[0+1*DCTSTRIDE] - data[1+1*DCTSTRIDE];
1146 data[0+0*DCTSTRIDE]= (d00 + d10)>>3;
1147 data[1+0*DCTSTRIDE]= (d01 + d11)>>3;
1148 data[0+1*DCTSTRIDE]= (d00 - d10)>>3;
1149 data[1+1*DCTSTRIDE]= (d01 - d11)>>3;
1152 void ff_j_rev_dct1(DCTBLOCK data){
1153 data[0] = (data[0] + 4)>>3;
1159 void ff_jref_idct_put(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
1161 ff_j_rev_dct(block);
1162 ff_put_pixels_clamped_c(block, dest, line_size);
1165 void ff_jref_idct_add(uint8_t *dest, ptrdiff_t line_size, int16_t *block)
1167 ff_j_rev_dct(block);
1168 ff_add_pixels_clamped_c(block, dest, line_size);