4
* This file is part of the Independent JPEG Group's software.
6
* The authors make NO WARRANTY or representation, either express or implied,
7
* with respect to this software, its quality, accuracy, merchantability, or
8
* fitness for a particular purpose. This software is provided "AS IS", and
9
* you, its user, assume the entire risk as to its quality and accuracy.
11
* This software is copyright (C) 1991-1996, Thomas G. Lane.
12
* All Rights Reserved except as specified below.
14
* Permission is hereby granted to use, copy, modify, and distribute this
15
* software (or portions thereof) for any purpose, without fee, subject to
17
* (1) If any part of the source code for this software is distributed, then
18
* this README file must be included, with this copyright and no-warranty
19
* notice unaltered; and any additions, deletions, or changes to the original
20
* files must be clearly indicated in accompanying documentation.
21
* (2) If only executable code is distributed, then the accompanying
22
* documentation must state that "this software is based in part on the work
23
* of the Independent JPEG Group".
24
* (3) Permission for use of this software is granted only if the user accepts
25
* full responsibility for any undesirable consequences; the authors accept
26
* NO LIABILITY for damages of any kind.
28
* These conditions apply to any software derived from or based on the IJG
29
* code, not just to the unmodified library. If you use our work, you ought
32
* Permission is NOT granted for the use of any IJG author's name or company
33
* name in advertising or publicity relating to this software or products
34
* derived from it. This software may be referred to only as "the Independent
35
* JPEG Group's software".
37
* We specifically permit and encourage the use of this software as the basis
38
* of commercial products, provided that all warranty or liability claims are
39
* assumed by the product vendor.
41
* This file contains a slow-but-accurate integer implementation of the
42
* forward DCT (Discrete Cosine Transform).
44
* A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
45
* on each column. Direct algorithms are also available, but they are
46
* much more complex and seem not to be any faster when reduced to code.
48
* This implementation is based on an algorithm described in
49
* C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
50
* Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
51
* Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
52
* The primary algorithm described there uses 11 multiplies and 29 adds.
53
* We use their alternate method with 12 multiplies and 32 adds.
54
* The advantage of this method is that no data path contains more than one
55
* multiplication; this allows a very simple and accurate implementation in
56
* scaled fixed-point arithmetic, with a minimal number of shifts.
61
* Independent JPEG Group's slow & accurate dct.
71
#define BITS_IN_JSAMPLE 8
73
#define RIGHT_SHIFT(x, n) ((x) >> (n))
74
#define MULTIPLY16C16(var,const) ((var)*(const))
76
#if 1 //def USE_ACCURATE_ROUNDING
77
#define DESCALE(x,n) RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
79
#define DESCALE(x,n) RIGHT_SHIFT(x, n)
84
* This module is specialized to the case DCTSIZE = 8.
88
Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
93
* The poop on this scaling stuff is as follows:
95
* Each 1-D DCT step produces outputs which are a factor of sqrt(N)
96
* larger than the true DCT outputs. The final outputs are therefore
97
* a factor of N larger than desired; since N=8 this can be cured by
98
* a simple right shift at the end of the algorithm. The advantage of
99
* this arrangement is that we save two multiplications per 1-D DCT,
100
* because the y0 and y4 outputs need not be divided by sqrt(N).
101
* In the IJG code, this factor of 8 is removed by the quantization step
102
* (in jcdctmgr.c), NOT in this module.
104
* We have to do addition and subtraction of the integer inputs, which
105
* is no problem, and multiplication by fractional constants, which is
106
* a problem to do in integer arithmetic. We multiply all the constants
107
* by CONST_SCALE and convert them to integer constants (thus retaining
108
* CONST_BITS bits of precision in the constants). After doing a
109
* multiplication we have to divide the product by CONST_SCALE, with proper
110
* rounding, to produce the correct output. This division can be done
111
* cheaply as a right shift of CONST_BITS bits. We postpone shifting
112
* as long as possible so that partial sums can be added together with
113
* full fractional precision.
115
* The outputs of the first pass are scaled up by PASS1_BITS bits so that
116
* they are represented to better-than-integral precision. These outputs
117
* require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
118
* with the recommended scaling. (For 12-bit sample data, the intermediate
119
* array is int32_t anyway.)
121
* To avoid overflow of the 32-bit intermediate results in pass 2, we must
122
* have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
123
* shows that the values given below are the most effective.
126
#if BITS_IN_JSAMPLE == 8
127
#define CONST_BITS 13
128
#define PASS1_BITS 4 /* set this to 2 if 16x16 multiplies are faster */
130
#define CONST_BITS 13
131
#define PASS1_BITS 1 /* lose a little precision to avoid overflow */
134
/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
135
* causing a lot of useless floating-point operations at run time.
136
* To get around this we use the following pre-calculated constants.
137
* If you change CONST_BITS you may want to add appropriate values.
138
* (With a reasonable C compiler, you can just rely on the FIX() macro...)
142
#define FIX_0_298631336 ((int32_t) 2446) /* FIX(0.298631336) */
143
#define FIX_0_390180644 ((int32_t) 3196) /* FIX(0.390180644) */
144
#define FIX_0_541196100 ((int32_t) 4433) /* FIX(0.541196100) */
145
#define FIX_0_765366865 ((int32_t) 6270) /* FIX(0.765366865) */
146
#define FIX_0_899976223 ((int32_t) 7373) /* FIX(0.899976223) */
147
#define FIX_1_175875602 ((int32_t) 9633) /* FIX(1.175875602) */
148
#define FIX_1_501321110 ((int32_t) 12299) /* FIX(1.501321110) */
149
#define FIX_1_847759065 ((int32_t) 15137) /* FIX(1.847759065) */
150
#define FIX_1_961570560 ((int32_t) 16069) /* FIX(1.961570560) */
151
#define FIX_2_053119869 ((int32_t) 16819) /* FIX(2.053119869) */
152
#define FIX_2_562915447 ((int32_t) 20995) /* FIX(2.562915447) */
153
#define FIX_3_072711026 ((int32_t) 25172) /* FIX(3.072711026) */
155
#define FIX_0_298631336 FIX(0.298631336)
156
#define FIX_0_390180644 FIX(0.390180644)
157
#define FIX_0_541196100 FIX(0.541196100)
158
#define FIX_0_765366865 FIX(0.765366865)
159
#define FIX_0_899976223 FIX(0.899976223)
160
#define FIX_1_175875602 FIX(1.175875602)
161
#define FIX_1_501321110 FIX(1.501321110)
162
#define FIX_1_847759065 FIX(1.847759065)
163
#define FIX_1_961570560 FIX(1.961570560)
164
#define FIX_2_053119869 FIX(2.053119869)
165
#define FIX_2_562915447 FIX(2.562915447)
166
#define FIX_3_072711026 FIX(3.072711026)
170
/* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
171
* For 8-bit samples with the recommended scaling, all the variable
172
* and constant values involved are no more than 16 bits wide, so a
173
* 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
174
* For 12-bit samples, a full 32-bit multiplication will be needed.
177
#if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
178
#define MULTIPLY(var,const) MULTIPLY16C16(var,const)
180
#define MULTIPLY(var,const) ((var) * (const))
184
static av_always_inline void row_fdct(DCTELEM * data){
185
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
186
int_fast32_t tmp10, tmp11, tmp12, tmp13;
187
int_fast32_t z1, z2, z3, z4, z5;
192
/* Pass 1: process rows. */
193
/* Note results are scaled up by sqrt(8) compared to a true DCT; */
194
/* furthermore, we scale the results by 2**PASS1_BITS. */
197
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
198
tmp0 = dataptr[0] + dataptr[7];
199
tmp7 = dataptr[0] - dataptr[7];
200
tmp1 = dataptr[1] + dataptr[6];
201
tmp6 = dataptr[1] - dataptr[6];
202
tmp2 = dataptr[2] + dataptr[5];
203
tmp5 = dataptr[2] - dataptr[5];
204
tmp3 = dataptr[3] + dataptr[4];
205
tmp4 = dataptr[3] - dataptr[4];
207
/* Even part per LL&M figure 1 --- note that published figure is faulty;
208
* rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
216
dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
217
dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
219
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
220
dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
221
CONST_BITS-PASS1_BITS);
222
dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
223
CONST_BITS-PASS1_BITS);
225
/* Odd part per figure 8 --- note paper omits factor of sqrt(2).
226
* cK represents cos(K*pi/16).
227
* i0..i3 in the paper are tmp4..tmp7 here.
234
z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
236
tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
237
tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
238
tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
239
tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
240
z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
241
z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
242
z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
243
z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
248
dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
249
dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
250
dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
251
dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
253
dataptr += DCTSIZE; /* advance pointer to next row */
258
* Perform the forward DCT on one block of samples.
262
ff_jpeg_fdct_islow (DCTELEM * data)
264
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
265
int_fast32_t tmp10, tmp11, tmp12, tmp13;
266
int_fast32_t z1, z2, z3, z4, z5;
273
/* Pass 2: process columns.
274
* We remove the PASS1_BITS scaling, but leave the results scaled up
275
* by an overall factor of 8.
279
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
280
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
281
tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
282
tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
283
tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
284
tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
285
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
286
tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
287
tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
289
/* Even part per LL&M figure 1 --- note that published figure is faulty;
290
* rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
298
dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
299
dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
301
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
302
dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
303
CONST_BITS+PASS1_BITS);
304
dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
305
CONST_BITS+PASS1_BITS);
307
/* Odd part per figure 8 --- note paper omits factor of sqrt(2).
308
* cK represents cos(K*pi/16).
309
* i0..i3 in the paper are tmp4..tmp7 here.
316
z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
318
tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
319
tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
320
tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
321
tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
322
z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
323
z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
324
z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
325
z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
330
dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
331
CONST_BITS+PASS1_BITS);
332
dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
333
CONST_BITS+PASS1_BITS);
334
dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
335
CONST_BITS+PASS1_BITS);
336
dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
337
CONST_BITS+PASS1_BITS);
339
dataptr++; /* advance pointer to next column */
344
* The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
345
* on the rows and then, instead of doing even and odd, part on the colums
346
* you do even part two times.
349
ff_fdct248_islow (DCTELEM * data)
351
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
352
int_fast32_t tmp10, tmp11, tmp12, tmp13;
360
/* Pass 2: process columns.
361
* We remove the PASS1_BITS scaling, but leave the results scaled up
362
* by an overall factor of 8.
366
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
367
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
368
tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
369
tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
370
tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
371
tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
372
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
373
tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
374
tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
381
dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
382
dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
384
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
385
dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
386
CONST_BITS+PASS1_BITS);
387
dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
388
CONST_BITS+PASS1_BITS);
395
dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
396
dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
398
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
399
dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
400
CONST_BITS+PASS1_BITS);
401
dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
402
CONST_BITS+PASS1_BITS);
404
dataptr++; /* advance pointer to next column */