1/*
2 * jfdctint.c
3 *
4 * This file was part of the Independent JPEG Group's software.
5 * Copyright (C) 1991-1996, Thomas G. Lane.
6 * libjpeg-turbo Modifications:
7 * Copyright (C) 2015, D. R. Commander
8 * For conditions of distribution and use, see the accompanying README file.
9 *
10 * This file contains a slow-but-accurate integer implementation of the
11 * forward DCT (Discrete Cosine Transform).
12 *
13 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
14 * on each column.  Direct algorithms are also available, but they are
15 * much more complex and seem not to be any faster when reduced to code.
16 *
17 * This implementation is based on an algorithm described in
18 *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
19 *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
20 *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
21 * The primary algorithm described there uses 11 multiplies and 29 adds.
22 * We use their alternate method with 12 multiplies and 32 adds.
23 * The advantage of this method is that no data path contains more than one
24 * multiplication; this allows a very simple and accurate implementation in
25 * scaled fixed-point arithmetic, with a minimal number of shifts.
26 */
27
28#define JPEG_INTERNALS
29#include "jinclude.h"
30#include "jpeglib.h"
31#include "jdct.h"               /* Private declarations for DCT subsystem */
32
33#ifdef DCT_ISLOW_SUPPORTED
34
35
36/*
37 * This module is specialized to the case DCTSIZE = 8.
38 */
39
40#if DCTSIZE != 8
41  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
42#endif
43
44
45/*
46 * The poop on this scaling stuff is as follows:
47 *
48 * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
49 * larger than the true DCT outputs.  The final outputs are therefore
50 * a factor of N larger than desired; since N=8 this can be cured by
51 * a simple right shift at the end of the algorithm.  The advantage of
52 * this arrangement is that we save two multiplications per 1-D DCT,
53 * because the y0 and y4 outputs need not be divided by sqrt(N).
54 * In the IJG code, this factor of 8 is removed by the quantization step
55 * (in jcdctmgr.c), NOT in this module.
56 *
57 * We have to do addition and subtraction of the integer inputs, which
58 * is no problem, and multiplication by fractional constants, which is
59 * a problem to do in integer arithmetic.  We multiply all the constants
60 * by CONST_SCALE and convert them to integer constants (thus retaining
61 * CONST_BITS bits of precision in the constants).  After doing a
62 * multiplication we have to divide the product by CONST_SCALE, with proper
63 * rounding, to produce the correct output.  This division can be done
64 * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
65 * as long as possible so that partial sums can be added together with
66 * full fractional precision.
67 *
68 * The outputs of the first pass are scaled up by PASS1_BITS bits so that
69 * they are represented to better-than-integral precision.  These outputs
70 * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
71 * with the recommended scaling.  (For 12-bit sample data, the intermediate
72 * array is INT32 anyway.)
73 *
74 * To avoid overflow of the 32-bit intermediate results in pass 2, we must
75 * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
76 * shows that the values given below are the most effective.
77 */
78
79#if BITS_IN_JSAMPLE == 8
80#define CONST_BITS  13
81#define PASS1_BITS  2
82#else
83#define CONST_BITS  13
84#define PASS1_BITS  1           /* lose a little precision to avoid overflow */
85#endif
86
87/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
88 * causing a lot of useless floating-point operations at run time.
89 * To get around this we use the following pre-calculated constants.
90 * If you change CONST_BITS you may want to add appropriate values.
91 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
92 */
93
94#if CONST_BITS == 13
95#define FIX_0_298631336  ((INT32)  2446)        /* FIX(0.298631336) */
96#define FIX_0_390180644  ((INT32)  3196)        /* FIX(0.390180644) */
97#define FIX_0_541196100  ((INT32)  4433)        /* FIX(0.541196100) */
98#define FIX_0_765366865  ((INT32)  6270)        /* FIX(0.765366865) */
99#define FIX_0_899976223  ((INT32)  7373)        /* FIX(0.899976223) */
100#define FIX_1_175875602  ((INT32)  9633)        /* FIX(1.175875602) */
101#define FIX_1_501321110  ((INT32)  12299)       /* FIX(1.501321110) */
102#define FIX_1_847759065  ((INT32)  15137)       /* FIX(1.847759065) */
103#define FIX_1_961570560  ((INT32)  16069)       /* FIX(1.961570560) */
104#define FIX_2_053119869  ((INT32)  16819)       /* FIX(2.053119869) */
105#define FIX_2_562915447  ((INT32)  20995)       /* FIX(2.562915447) */
106#define FIX_3_072711026  ((INT32)  25172)       /* FIX(3.072711026) */
107#else
108#define FIX_0_298631336  FIX(0.298631336)
109#define FIX_0_390180644  FIX(0.390180644)
110#define FIX_0_541196100  FIX(0.541196100)
111#define FIX_0_765366865  FIX(0.765366865)
112#define FIX_0_899976223  FIX(0.899976223)
113#define FIX_1_175875602  FIX(1.175875602)
114#define FIX_1_501321110  FIX(1.501321110)
115#define FIX_1_847759065  FIX(1.847759065)
116#define FIX_1_961570560  FIX(1.961570560)
117#define FIX_2_053119869  FIX(2.053119869)
118#define FIX_2_562915447  FIX(2.562915447)
119#define FIX_3_072711026  FIX(3.072711026)
120#endif
121
122
123/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
124 * For 8-bit samples with the recommended scaling, all the variable
125 * and constant values involved are no more than 16 bits wide, so a
126 * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
127 * For 12-bit samples, a full 32-bit multiplication will be needed.
128 */
129
130#if BITS_IN_JSAMPLE == 8
131#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
132#else
133#define MULTIPLY(var,const)  ((var) * (const))
134#endif
135
136
137/*
138 * Perform the forward DCT on one block of samples.
139 */
140
141GLOBAL(void)
142jpeg_fdct_islow (DCTELEM * data)
143{
144  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
145  INT32 tmp10, tmp11, tmp12, tmp13;
146  INT32 z1, z2, z3, z4, z5;
147  DCTELEM *dataptr;
148  int ctr;
149  SHIFT_TEMPS
150
151  /* Pass 1: process rows. */
152  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
153  /* furthermore, we scale the results by 2**PASS1_BITS. */
154
155  dataptr = data;
156  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
157    tmp0 = dataptr[0] + dataptr[7];
158    tmp7 = dataptr[0] - dataptr[7];
159    tmp1 = dataptr[1] + dataptr[6];
160    tmp6 = dataptr[1] - dataptr[6];
161    tmp2 = dataptr[2] + dataptr[5];
162    tmp5 = dataptr[2] - dataptr[5];
163    tmp3 = dataptr[3] + dataptr[4];
164    tmp4 = dataptr[3] - dataptr[4];
165
166    /* Even part per LL&M figure 1 --- note that published figure is faulty;
167     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
168     */
169
170    tmp10 = tmp0 + tmp3;
171    tmp13 = tmp0 - tmp3;
172    tmp11 = tmp1 + tmp2;
173    tmp12 = tmp1 - tmp2;
174
175    dataptr[0] = (DCTELEM) LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS);
176    dataptr[4] = (DCTELEM) LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS);
177
178    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
179    dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
180                                   CONST_BITS-PASS1_BITS);
181    dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
182                                   CONST_BITS-PASS1_BITS);
183
184    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
185     * cK represents cos(K*pi/16).
186     * i0..i3 in the paper are tmp4..tmp7 here.
187     */
188
189    z1 = tmp4 + tmp7;
190    z2 = tmp5 + tmp6;
191    z3 = tmp4 + tmp6;
192    z4 = tmp5 + tmp7;
193    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
194
195    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
196    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
197    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
198    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
199    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
200    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
201    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
202    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
203
204    z3 += z5;
205    z4 += z5;
206
207    dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
208    dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
209    dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
210    dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
211
212    dataptr += DCTSIZE;         /* advance pointer to next row */
213  }
214
215  /* Pass 2: process columns.
216   * We remove the PASS1_BITS scaling, but leave the results scaled up
217   * by an overall factor of 8.
218   */
219
220  dataptr = data;
221  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
222    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
223    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
224    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
225    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
226    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
227    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
228    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
229    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
230
231    /* Even part per LL&M figure 1 --- note that published figure is faulty;
232     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
233     */
234
235    tmp10 = tmp0 + tmp3;
236    tmp13 = tmp0 - tmp3;
237    tmp11 = tmp1 + tmp2;
238    tmp12 = tmp1 - tmp2;
239
240    dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
241    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
242
243    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
244    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
245                                           CONST_BITS+PASS1_BITS);
246    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
247                                           CONST_BITS+PASS1_BITS);
248
249    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
250     * cK represents cos(K*pi/16).
251     * i0..i3 in the paper are tmp4..tmp7 here.
252     */
253
254    z1 = tmp4 + tmp7;
255    z2 = tmp5 + tmp6;
256    z3 = tmp4 + tmp6;
257    z4 = tmp5 + tmp7;
258    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
259
260    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
261    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
262    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
263    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
264    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
265    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
266    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
267    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
268
269    z3 += z5;
270    z4 += z5;
271
272    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
273                                           CONST_BITS+PASS1_BITS);
274    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
275                                           CONST_BITS+PASS1_BITS);
276    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
277                                           CONST_BITS+PASS1_BITS);
278    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
279                                           CONST_BITS+PASS1_BITS);
280
281    dataptr++;                  /* advance pointer to next column */
282  }
283}
284
285#endif /* DCT_ISLOW_SUPPORTED */
286