4
* Copyright (C) 1991-1996, Thomas G. Lane.
5
* This file is part of the Independent JPEG Group's software.
6
* For conditions of distribution and use, see the accompanying README file.
8
* This file contains a slow-but-accurate integer implementation of the
9
* forward DCT (Discrete Cosine Transform).
11
* A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
12
* on each column. Direct algorithms are also available, but they are
13
* much more complex and seem not to be any faster when reduced to code.
15
* This implementation is based on an algorithm described in
16
* C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
17
* Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
18
* Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
19
* The primary algorithm described there uses 11 multiplies and 29 adds.
20
* We use their alternate method with 12 multiplies and 32 adds.
21
* The advantage of this method is that no data path contains more than one
22
* multiplication; this allows a very simple and accurate implementation in
23
* scaled fixed-point arithmetic, with a minimal number of shifts.
28
* Independent JPEG Group's slow & accurate dct.
38
#define BITS_IN_JSAMPLE 8
40
#define RIGHT_SHIFT(x, n) ((x) >> (n))
41
#define MULTIPLY16C16(var,const) ((var)*(const))
43
#if 1 //def USE_ACCURATE_ROUNDING
44
#define DESCALE(x,n) RIGHT_SHIFT((x) + (1 << ((n) - 1)), n)
46
#define DESCALE(x,n) RIGHT_SHIFT(x, n)
51
* This module is specialized to the case DCTSIZE = 8.
55
Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
60
* The poop on this scaling stuff is as follows:
62
* Each 1-D DCT step produces outputs which are a factor of sqrt(N)
63
* larger than the true DCT outputs. The final outputs are therefore
64
* a factor of N larger than desired; since N=8 this can be cured by
65
* a simple right shift at the end of the algorithm. The advantage of
66
* this arrangement is that we save two multiplications per 1-D DCT,
67
* because the y0 and y4 outputs need not be divided by sqrt(N).
68
* In the IJG code, this factor of 8 is removed by the quantization step
69
* (in jcdctmgr.c), NOT in this module.
71
* We have to do addition and subtraction of the integer inputs, which
72
* is no problem, and multiplication by fractional constants, which is
73
* a problem to do in integer arithmetic. We multiply all the constants
74
* by CONST_SCALE and convert them to integer constants (thus retaining
75
* CONST_BITS bits of precision in the constants). After doing a
76
* multiplication we have to divide the product by CONST_SCALE, with proper
77
* rounding, to produce the correct output. This division can be done
78
* cheaply as a right shift of CONST_BITS bits. We postpone shifting
79
* as long as possible so that partial sums can be added together with
80
* full fractional precision.
82
* The outputs of the first pass are scaled up by PASS1_BITS bits so that
83
* they are represented to better-than-integral precision. These outputs
84
* require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
85
* with the recommended scaling. (For 12-bit sample data, the intermediate
86
* array is int32_t anyway.)
88
* To avoid overflow of the 32-bit intermediate results in pass 2, we must
89
* have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26. Error analysis
90
* shows that the values given below are the most effective.
93
#if BITS_IN_JSAMPLE == 8
95
#define PASS1_BITS 4 /* set this to 2 if 16x16 multiplies are faster */
98
#define PASS1_BITS 1 /* lose a little precision to avoid overflow */
101
/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
102
* causing a lot of useless floating-point operations at run time.
103
* To get around this we use the following pre-calculated constants.
104
* If you change CONST_BITS you may want to add appropriate values.
105
* (With a reasonable C compiler, you can just rely on the FIX() macro...)
109
#define FIX_0_298631336 ((int32_t) 2446) /* FIX(0.298631336) */
110
#define FIX_0_390180644 ((int32_t) 3196) /* FIX(0.390180644) */
111
#define FIX_0_541196100 ((int32_t) 4433) /* FIX(0.541196100) */
112
#define FIX_0_765366865 ((int32_t) 6270) /* FIX(0.765366865) */
113
#define FIX_0_899976223 ((int32_t) 7373) /* FIX(0.899976223) */
114
#define FIX_1_175875602 ((int32_t) 9633) /* FIX(1.175875602) */
115
#define FIX_1_501321110 ((int32_t) 12299) /* FIX(1.501321110) */
116
#define FIX_1_847759065 ((int32_t) 15137) /* FIX(1.847759065) */
117
#define FIX_1_961570560 ((int32_t) 16069) /* FIX(1.961570560) */
118
#define FIX_2_053119869 ((int32_t) 16819) /* FIX(2.053119869) */
119
#define FIX_2_562915447 ((int32_t) 20995) /* FIX(2.562915447) */
120
#define FIX_3_072711026 ((int32_t) 25172) /* FIX(3.072711026) */
122
#define FIX_0_298631336 FIX(0.298631336)
123
#define FIX_0_390180644 FIX(0.390180644)
124
#define FIX_0_541196100 FIX(0.541196100)
125
#define FIX_0_765366865 FIX(0.765366865)
126
#define FIX_0_899976223 FIX(0.899976223)
127
#define FIX_1_175875602 FIX(1.175875602)
128
#define FIX_1_501321110 FIX(1.501321110)
129
#define FIX_1_847759065 FIX(1.847759065)
130
#define FIX_1_961570560 FIX(1.961570560)
131
#define FIX_2_053119869 FIX(2.053119869)
132
#define FIX_2_562915447 FIX(2.562915447)
133
#define FIX_3_072711026 FIX(3.072711026)
137
/* Multiply an int32_t variable by an int32_t constant to yield an int32_t result.
138
* For 8-bit samples with the recommended scaling, all the variable
139
* and constant values involved are no more than 16 bits wide, so a
140
* 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
141
* For 12-bit samples, a full 32-bit multiplication will be needed.
144
#if BITS_IN_JSAMPLE == 8 && CONST_BITS<=13 && PASS1_BITS<=2
145
#define MULTIPLY(var,const) MULTIPLY16C16(var,const)
147
#define MULTIPLY(var,const) ((var) * (const))
151
static always_inline void row_fdct(DCTELEM * data){
152
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
153
int_fast32_t tmp10, tmp11, tmp12, tmp13;
154
int_fast32_t z1, z2, z3, z4, z5;
159
/* Pass 1: process rows. */
160
/* Note results are scaled up by sqrt(8) compared to a true DCT; */
161
/* furthermore, we scale the results by 2**PASS1_BITS. */
164
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
165
tmp0 = dataptr[0] + dataptr[7];
166
tmp7 = dataptr[0] - dataptr[7];
167
tmp1 = dataptr[1] + dataptr[6];
168
tmp6 = dataptr[1] - dataptr[6];
169
tmp2 = dataptr[2] + dataptr[5];
170
tmp5 = dataptr[2] - dataptr[5];
171
tmp3 = dataptr[3] + dataptr[4];
172
tmp4 = dataptr[3] - dataptr[4];
174
/* Even part per LL&M figure 1 --- note that published figure is faulty;
175
* rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
183
dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
184
dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
186
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
187
dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
188
CONST_BITS-PASS1_BITS);
189
dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
190
CONST_BITS-PASS1_BITS);
192
/* Odd part per figure 8 --- note paper omits factor of sqrt(2).
193
* cK represents cos(K*pi/16).
194
* i0..i3 in the paper are tmp4..tmp7 here.
201
z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
203
tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
204
tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
205
tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
206
tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
207
z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
208
z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
209
z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
210
z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
215
dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
216
dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
217
dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
218
dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
220
dataptr += DCTSIZE; /* advance pointer to next row */
225
* Perform the forward DCT on one block of samples.
229
ff_jpeg_fdct_islow (DCTELEM * data)
231
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
232
int_fast32_t tmp10, tmp11, tmp12, tmp13;
233
int_fast32_t z1, z2, z3, z4, z5;
240
/* Pass 2: process columns.
241
* We remove the PASS1_BITS scaling, but leave the results scaled up
242
* by an overall factor of 8.
246
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
247
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
248
tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
249
tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
250
tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
251
tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
252
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
253
tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
254
tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
256
/* Even part per LL&M figure 1 --- note that published figure is faulty;
257
* rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
265
dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
266
dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
268
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
269
dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
270
CONST_BITS+PASS1_BITS);
271
dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
272
CONST_BITS+PASS1_BITS);
274
/* Odd part per figure 8 --- note paper omits factor of sqrt(2).
275
* cK represents cos(K*pi/16).
276
* i0..i3 in the paper are tmp4..tmp7 here.
283
z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
285
tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
286
tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
287
tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
288
tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
289
z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
290
z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
291
z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
292
z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
297
dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
298
CONST_BITS+PASS1_BITS);
299
dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
300
CONST_BITS+PASS1_BITS);
301
dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
302
CONST_BITS+PASS1_BITS);
303
dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
304
CONST_BITS+PASS1_BITS);
306
dataptr++; /* advance pointer to next column */
311
* The secret of DCT2-4-8 is really simple -- you do the usual 1-DCT
312
* on the rows and then, instead of doing even and odd, part on the colums
313
* you do even part two times.
316
ff_fdct248_islow (DCTELEM * data)
318
int_fast32_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
319
int_fast32_t tmp10, tmp11, tmp12, tmp13;
327
/* Pass 2: process columns.
328
* We remove the PASS1_BITS scaling, but leave the results scaled up
329
* by an overall factor of 8.
333
for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
334
tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1];
335
tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
336
tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
337
tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
338
tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1];
339
tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
340
tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
341
tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
348
dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
349
dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
351
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
352
dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
353
CONST_BITS+PASS1_BITS);
354
dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
355
CONST_BITS+PASS1_BITS);
362
dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
363
dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
365
z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
366
dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
367
CONST_BITS+PASS1_BITS);
368
dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
369
CONST_BITS+PASS1_BITS);
371
dataptr++; /* advance pointer to next column */