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) 1991-1996, 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 slow-but-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 an algorithm described in
47 *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
48 *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
49 *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
50 * The primary algorithm described there uses 11 multiplies and 29 adds.
51 * We use their alternate method with 12 multiplies and 32 adds.
52 * The advantage of this method is that no data path contains more than one
53 * multiplication; this allows a very simple and accurate implementation in
54 * scaled fixed-point arithmetic, with a minimal number of shifts.
55 */
56
57/**
58 * @file
59 * Independent JPEG Group's slow & accurate dct.
60 */
61
62#include "libavutil/common.h"
63#include "dct.h"
64
65#include "bit_depth_template.c"
66
67#define DCTSIZE 8
68#define BITS_IN_JSAMPLE BIT_DEPTH
69#define GLOBAL(x) x
70#define RIGHT_SHIFT(x, n) ((x) >> (n))
71#define MULTIPLY16C16(var,const) ((var)*(const))
72#define DESCALE(x,n)  RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
73
74
75/*
76 * This module is specialized to the case DCTSIZE = 8.
77 */
78
79#if DCTSIZE != 8
80#error  "Sorry, this code only copes with 8x8 DCTs."
81#endif
82
83
84/*
85 * The poop on this scaling stuff is as follows:
86 *
87 * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
88 * larger than the true DCT outputs.  The final outputs are therefore
89 * a factor of N larger than desired; since N=8 this can be cured by
90 * a simple right shift at the end of the algorithm.  The advantage of
91 * this arrangement is that we save two multiplications per 1-D DCT,
92 * because the y0 and y4 outputs need not be divided by sqrt(N).
93 * In the IJG code, this factor of 8 is removed by the quantization step
94 * (in jcdctmgr.c), NOT in this module.
95 *
96 * We have to do addition and subtraction of the integer inputs, which
97 * is no problem, and multiplication by fractional constants, which is
98 * a problem to do in integer arithmetic.  We multiply all the constants
99 * by CONST_SCALE and convert them to integer constants (thus retaining
100 * CONST_BITS bits of precision in the constants).  After doing a
101 * multiplication we have to divide the product by CONST_SCALE, with proper
102 * rounding, to produce the correct output.  This division can be done
103 * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
104 * as long as possible so that partial sums can be added together with
105 * full fractional precision.
106 *
107 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
108 * they are represented to better-than-integral precision.  These outputs
109 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
110 * with the recommended scaling.  (For 12-bit sample data, the intermediate
111 * array is int32_t anyway.)
112 *
113 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
114 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
115 * shows that the values given below are the most effective.
116 */
117
118#undef CONST_BITS
119#undef PASS1_BITS
120#undef OUT_SHIFT
121
122#if BITS_IN_JSAMPLE == 8
123#define CONST_BITS  13
124#define PASS1_BITS  4   /* set this to 2 if 16x16 multiplies are faster */
125#define OUT_SHIFT   PASS1_BITS
126#else
127#define CONST_BITS  13
128#define PASS1_BITS  1   /* lose a little precision to avoid overflow */
129#define OUT_SHIFT   (PASS1_BITS + 1)
130#endif
131
132/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
133 * causing a lot of useless floating-point operations at run time.
134 * To get around this we use the following pre-calculated constants.
135 * If you change CONST_BITS you may want to add appropriate values.
136 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
137 */
138
139#if CONST_BITS == 13
140#define FIX_0_298631336  ((int32_t)  2446)      /* FIX(0.298631336) */
141#define FIX_0_390180644  ((int32_t)  3196)      /* FIX(0.390180644) */
142#define FIX_0_541196100  ((int32_t)  4433)      /* FIX(0.541196100) */
143#define FIX_0_765366865  ((int32_t)  6270)      /* FIX(0.765366865) */
144#define FIX_0_899976223  ((int32_t)  7373)      /* FIX(0.899976223) */
145#define FIX_1_175875602  ((int32_t)  9633)      /* FIX(1.175875602) */
146#define FIX_1_501321110  ((int32_t)  12299)     /* FIX(1.501321110) */
147#define FIX_1_847759065  ((int32_t)  15137)     /* FIX(1.847759065) */
148#define FIX_1_961570560  ((int32_t)  16069)     /* FIX(1.961570560) */
149#define FIX_2_053119869  ((int32_t)  16819)     /* FIX(2.053119869) */
150#define FIX_2_562915447  ((int32_t)  20995)     /* FIX(2.562915447) */
151#define FIX_3_072711026  ((int32_t)  25172)     /* FIX(3.072711026) */
152#else
153#define FIX_0_298631336  FIX(0.298631336)
154#define FIX_0_390180644  FIX(0.390180644)
155#define FIX_0_541196100  FIX(0.541196100)
156#define FIX_0_765366865  FIX(0.765366865)
157#define FIX_0_899976223  FIX(0.899976223)
158#define FIX_1_175875602  FIX(1.175875602)
159#define FIX_1_501321110  FIX(1.501321110)
160#define FIX_1_847759065  FIX(1.847759065)
161#define FIX_1_961570560  FIX(1.961570560)
162#define FIX_2_053119869  FIX(2.053119869)
163#define FIX_2_562915447  FIX(2.562915447)
164#define FIX_3_072711026  FIX(3.072711026)
165#endif
166
167
168/* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
169 * For 8-bit samples with the recommended scaling, all the variable
170 * and constant values involved are no more than 16 bits wide, so a
171 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
172 * For 12-bit samples, a full 32-bit multiplication will be needed.
173 */
174
175#if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
176#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
177#else
178#define MULTIPLY(var,const)  ((var) * (const))
179#endif
180
181
182static av_always_inline void FUNC(row_fdct)(int16_t *data)
183{
184  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
185  int tmp10, tmp11, tmp12, tmp13;
186  int z1, z2, z3, z4, z5;
187  int16_t *dataptr;
188  int ctr;
189
190  /* Pass 1: process rows. */
191  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
192  /* furthermore, we scale the results by 2**PASS1_BITS. */
193
194  dataptr = data;
195  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
196    tmp0 = dataptr[0] + dataptr[7];
197    tmp7 = dataptr[0] - dataptr[7];
198    tmp1 = dataptr[1] + dataptr[6];
199    tmp6 = dataptr[1] - dataptr[6];
200    tmp2 = dataptr[2] + dataptr[5];
201    tmp5 = dataptr[2] - dataptr[5];
202    tmp3 = dataptr[3] + dataptr[4];
203    tmp4 = dataptr[3] - dataptr[4];
204
205    /* Even part per LL&M figure 1 --- note that published figure is faulty;
206     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
207     */
208
209    tmp10 = tmp0 + tmp3;
210    tmp13 = tmp0 - tmp3;
211    tmp11 = tmp1 + tmp2;
212    tmp12 = tmp1 - tmp2;
213
214    dataptr[0] = (int16_t) ((tmp10 + tmp11) * (1 << PASS1_BITS));
215    dataptr[4] = (int16_t) ((tmp10 - tmp11) * (1 << PASS1_BITS));
216
217    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
218    dataptr[2] = (int16_t) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
219                                   CONST_BITS-PASS1_BITS);
220    dataptr[6] = (int16_t) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
221                                   CONST_BITS-PASS1_BITS);
222
223    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
224     * cK represents cos(K*pi/16).
225     * i0..i3 in the paper are tmp4..tmp7 here.
226     */
227
228    z1 = tmp4 + tmp7;
229    z2 = tmp5 + tmp6;
230    z3 = tmp4 + tmp6;
231    z4 = tmp5 + tmp7;
232    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
233
234    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
235    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
236    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
237    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
238    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
239    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
240    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
241    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
242
243    z3 += z5;
244    z4 += z5;
245
246    dataptr[7] = (int16_t) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
247    dataptr[5] = (int16_t) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
248    dataptr[3] = (int16_t) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
249    dataptr[1] = (int16_t) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
250
251    dataptr += DCTSIZE;         /* advance pointer to next row */
252  }
253}
254
255/*
256 * Perform the forward DCT on one block of samples.
257 */
258
259GLOBAL(void)
260FUNC(ff_jpeg_fdct_islow)(int16_t *data)
261{
262  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
263  int tmp10, tmp11, tmp12, tmp13;
264  int z1, z2, z3, z4, z5;
265  int16_t *dataptr;
266  int ctr;
267
268  FUNC(row_fdct)(data);
269
270  /* Pass 2: process columns.
271   * We remove the PASS1_BITS scaling, but leave the results scaled up
272   * by an overall factor of 8.
273   */
274
275  dataptr = data;
276  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
277    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
278    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
279    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
280    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
281    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
282    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
283    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
284    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
285
286    /* Even part per LL&M figure 1 --- note that published figure is faulty;
287     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
288     */
289
290    tmp10 = tmp0 + tmp3;
291    tmp13 = tmp0 - tmp3;
292    tmp11 = tmp1 + tmp2;
293    tmp12 = tmp1 - tmp2;
294
295    dataptr[DCTSIZE*0] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
296    dataptr[DCTSIZE*4] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
297
298    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
299    dataptr[DCTSIZE*2] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
300                                 CONST_BITS + OUT_SHIFT);
301    dataptr[DCTSIZE*6] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
302                                 CONST_BITS + OUT_SHIFT);
303
304    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
305     * cK represents cos(K*pi/16).
306     * i0..i3 in the paper are tmp4..tmp7 here.
307     */
308
309    z1 = tmp4 + tmp7;
310    z2 = tmp5 + tmp6;
311    z3 = tmp4 + tmp6;
312    z4 = tmp5 + tmp7;
313    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
314
315    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
316    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
317    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
318    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
319    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
320    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
321    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
322    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
323
324    z3 += z5;
325    z4 += z5;
326
327    dataptr[DCTSIZE*7] = DESCALE(tmp4 + z1 + z3, CONST_BITS + OUT_SHIFT);
328    dataptr[DCTSIZE*5] = DESCALE(tmp5 + z2 + z4, CONST_BITS + OUT_SHIFT);
329    dataptr[DCTSIZE*3] = DESCALE(tmp6 + z2 + z3, CONST_BITS + OUT_SHIFT);
330    dataptr[DCTSIZE*1] = DESCALE(tmp7 + z1 + z4, CONST_BITS + OUT_SHIFT);
331
332    dataptr++;                  /* advance pointer to next column */
333  }
334}
335
336/*
337 * The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
338 * on the rows and then, instead of doing even and odd, part on the columns
339 * you do even part two times.
340 */
341GLOBAL(void)
342FUNC(ff_fdct248_islow)(int16_t *data)
343{
344  int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
345  int tmp10, tmp11, tmp12, tmp13;
346  int z1;
347  int16_t *dataptr;
348  int ctr;
349
350  FUNC(row_fdct)(data);
351
352  /* Pass 2: process columns.
353   * We remove the PASS1_BITS scaling, but leave the results scaled up
354   * by an overall factor of 8.
355   */
356
357  dataptr = data;
358  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
359     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
360     tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
361     tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
362     tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
363     tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
364     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
365     tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
366     tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
367
368     tmp10 = tmp0 + tmp3;
369     tmp11 = tmp1 + tmp2;
370     tmp12 = tmp1 - tmp2;
371     tmp13 = tmp0 - tmp3;
372
373     dataptr[DCTSIZE*0] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
374     dataptr[DCTSIZE*4] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
375
376     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
377     dataptr[DCTSIZE*2] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
378                                  CONST_BITS+OUT_SHIFT);
379     dataptr[DCTSIZE*6] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
380                                  CONST_BITS+OUT_SHIFT);
381
382     tmp10 = tmp4 + tmp7;
383     tmp11 = tmp5 + tmp6;
384     tmp12 = tmp5 - tmp6;
385     tmp13 = tmp4 - tmp7;
386
387     dataptr[DCTSIZE*1] = DESCALE(tmp10 + tmp11, OUT_SHIFT);
388     dataptr[DCTSIZE*5] = DESCALE(tmp10 - tmp11, OUT_SHIFT);
389
390     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
391     dataptr[DCTSIZE*3] = DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
392                                  CONST_BITS + OUT_SHIFT);
393     dataptr[DCTSIZE*7] = DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
394                                  CONST_BITS + OUT_SHIFT);
395
396     dataptr++;                 /* advance pointer to next column */
397  }
398}
399