1/*
2 * jfdctfst.c
3 *
4 * Copyright (C) 1994-1996, Thomas G. Lane.
5 * Modified 2003-2009 by Guido Vollbeding.
6 * This file is part of the Independent JPEG Group's software.
7 * For conditions of distribution and use, see the accompanying README file.
8 *
9 * This file contains a fast, not so accurate integer implementation of the
10 * forward DCT (Discrete Cosine Transform).
11 *
12 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
13 * on each column.  Direct algorithms are also available, but they are
14 * much more complex and seem not to be any faster when reduced to code.
15 *
16 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
17 * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
18 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
19 * JPEG textbook (see REFERENCES section in file README).  The following code
20 * is based directly on figure 4-8 in P&M.
21 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
22 * possible to arrange the computation so that many of the multiplies are
23 * simple scalings of the final outputs.  These multiplies can then be
24 * folded into the multiplications or divisions by the JPEG quantization
25 * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
26 * to be done in the DCT itself.
27 * The primary disadvantage of this method is that with fixed-point math,
28 * accuracy is lost due to imprecise representation of the scaled
29 * quantization values.  The smaller the quantization table entry, the less
30 * precise the scaled value, so this implementation does worse with high-
31 * quality-setting files than with low-quality ones.
32 */
33
34#define JPEG_INTERNALS
35#include "jinclude.h"
36#include "jpeglib.h"
37#include "jdct.h"		/* Private declarations for DCT subsystem */
38
39#ifdef DCT_IFAST_SUPPORTED
40
41
42/*
43 * This module is specialized to the case DCTSIZE = 8.
44 */
45
46#if DCTSIZE != 8
47  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
48#endif
49
50
51/* Scaling decisions are generally the same as in the LL&M algorithm;
52 * see jfdctint.c for more details.  However, we choose to descale
53 * (right shift) multiplication products as soon as they are formed,
54 * rather than carrying additional fractional bits into subsequent additions.
55 * This compromises accuracy slightly, but it lets us save a few shifts.
56 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
57 * everywhere except in the multiplications proper; this saves a good deal
58 * of work on 16-bit-int machines.
59 *
60 * Again to save a few shifts, the intermediate results between pass 1 and
61 * pass 2 are not upscaled, but are represented only to integral precision.
62 *
63 * A final compromise is to represent the multiplicative constants to only
64 * 8 fractional bits, rather than 13.  This saves some shifting work on some
65 * machines, and may also reduce the cost of multiplication (since there
66 * are fewer one-bits in the constants).
67 */
68
69#define CONST_BITS  8
70
71
72/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
73 * causing a lot of useless floating-point operations at run time.
74 * To get around this we use the following pre-calculated constants.
75 * If you change CONST_BITS you may want to add appropriate values.
76 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
77 */
78
79#if CONST_BITS == 8
80#define FIX_0_382683433  ((INT32)   98)		/* FIX(0.382683433) */
81#define FIX_0_541196100  ((INT32)  139)		/* FIX(0.541196100) */
82#define FIX_0_707106781  ((INT32)  181)		/* FIX(0.707106781) */
83#define FIX_1_306562965  ((INT32)  334)		/* FIX(1.306562965) */
84#else
85#define FIX_0_382683433  FIX(0.382683433)
86#define FIX_0_541196100  FIX(0.541196100)
87#define FIX_0_707106781  FIX(0.707106781)
88#define FIX_1_306562965  FIX(1.306562965)
89#endif
90
91
92/* We can gain a little more speed, with a further compromise in accuracy,
93 * by omitting the addition in a descaling shift.  This yields an incorrectly
94 * rounded result half the time...
95 */
96
97#ifndef USE_ACCURATE_ROUNDING
98#undef DESCALE
99#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
100#endif
101
102
103/* Multiply a DCTELEM variable by an INT32 constant, and immediately
104 * descale to yield a DCTELEM result.
105 */
106
107#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
108
109
110/*
111 * Perform the forward DCT on one block of samples.
112 */
113
114GLOBAL(void)
115jpeg_fdct_ifast (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
116{
117  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
118  DCTELEM tmp10, tmp11, tmp12, tmp13;
119  DCTELEM z1, z2, z3, z4, z5, z11, z13;
120  DCTELEM *dataptr;
121  JSAMPROW elemptr;
122  int ctr;
123  SHIFT_TEMPS
124
125  /* Pass 1: process rows. */
126
127  dataptr = data;
128  for (ctr = 0; ctr < DCTSIZE; ctr++) {
129    elemptr = sample_data[ctr] + start_col;
130
131    /* Load data into workspace */
132    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
133    tmp7 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
134    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
135    tmp6 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
136    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
137    tmp5 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
138    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
139    tmp4 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
140
141    /* Even part */
142
143    tmp10 = tmp0 + tmp3;	/* phase 2 */
144    tmp13 = tmp0 - tmp3;
145    tmp11 = tmp1 + tmp2;
146    tmp12 = tmp1 - tmp2;
147
148    /* Apply unsigned->signed conversion */
149    dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
150    dataptr[4] = tmp10 - tmp11;
151
152    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
153    dataptr[2] = tmp13 + z1;	/* phase 5 */
154    dataptr[6] = tmp13 - z1;
155
156    /* Odd part */
157
158    tmp10 = tmp4 + tmp5;	/* phase 2 */
159    tmp11 = tmp5 + tmp6;
160    tmp12 = tmp6 + tmp7;
161
162    /* The rotator is modified from fig 4-8 to avoid extra negations. */
163    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
164    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
165    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
166    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
167
168    z11 = tmp7 + z3;		/* phase 5 */
169    z13 = tmp7 - z3;
170
171    dataptr[5] = z13 + z2;	/* phase 6 */
172    dataptr[3] = z13 - z2;
173    dataptr[1] = z11 + z4;
174    dataptr[7] = z11 - z4;
175
176    dataptr += DCTSIZE;		/* advance pointer to next row */
177  }
178
179  /* Pass 2: process columns. */
180
181  dataptr = data;
182  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
183    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
184    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
185    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
186    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
187    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
188    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
189    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
190    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
191
192    /* Even part */
193
194    tmp10 = tmp0 + tmp3;	/* phase 2 */
195    tmp13 = tmp0 - tmp3;
196    tmp11 = tmp1 + tmp2;
197    tmp12 = tmp1 - tmp2;
198
199    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
200    dataptr[DCTSIZE*4] = tmp10 - tmp11;
201
202    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
203    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
204    dataptr[DCTSIZE*6] = tmp13 - z1;
205
206    /* Odd part */
207
208    tmp10 = tmp4 + tmp5;	/* phase 2 */
209    tmp11 = tmp5 + tmp6;
210    tmp12 = tmp6 + tmp7;
211
212    /* The rotator is modified from fig 4-8 to avoid extra negations. */
213    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
214    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
215    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
216    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
217
218    z11 = tmp7 + z3;		/* phase 5 */
219    z13 = tmp7 - z3;
220
221    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
222    dataptr[DCTSIZE*3] = z13 - z2;
223    dataptr[DCTSIZE*1] = z11 + z4;
224    dataptr[DCTSIZE*7] = z11 - z4;
225
226    dataptr++;			/* advance pointer to next column */
227  }
228}
229
230#endif /* DCT_IFAST_SUPPORTED */
231