4 * Copyright (C) 1994-1996, 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 fast, not so 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 Arai, Agui, and Nakajima's algorithm for
16 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
17 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
18 * JPEG textbook (see REFERENCES section in file README). The following code
19 * is based directly on figure 4-8 in P&M.
20 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
21 * possible to arrange the computation so that many of the multiplies are
22 * simple scalings of the final outputs. These multiplies can then be
23 * folded into the multiplications or divisions by the JPEG quantization
24 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
25 * to be done in the DCT itself.
26 * The primary disadvantage of this method is that with fixed-point math,
27 * accuracy is lost due to imprecise representation of the scaled
28 * quantization values. The smaller the quantization table entry, the less
29 * precise the scaled value, so this implementation does worse with high-
30 * quality-setting files than with low-quality ones.
40 #define RIGHT_SHIFT(x, n) ((x) >> (n))
44 * This module is specialized to the case DCTSIZE = 8.
48 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
52 /* Scaling decisions are generally the same as in the LL&M algorithm;
53 * see jfdctint.c for more details. However, we choose to descale
54 * (right shift) multiplication products as soon as they are formed,
55 * rather than carrying additional fractional bits into subsequent additions.
56 * This compromises accuracy slightly, but it lets us save a few shifts.
57 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
58 * everywhere except in the multiplications proper; this saves a good deal
59 * of work on 16-bit-int machines.
61 * Again to save a few shifts, the intermediate results between pass 1 and
62 * pass 2 are not upscaled, but are represented only to integral precision.
64 * A final compromise is to represent the multiplicative constants to only
65 * 8 fractional bits, rather than 13. This saves some shifting work on some
66 * machines, and may also reduce the cost of multiplication (since there
67 * are fewer one-bits in the constants).
73 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
74 * causing a lot of useless floating-point operations at run time.
75 * To get around this we use the following pre-calculated constants.
76 * If you change CONST_BITS you may want to add appropriate values.
77 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
81 #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */
82 #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */
83 #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */
84 #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */
86 #define FIX_0_382683433 FIX(0.382683433)
87 #define FIX_0_541196100 FIX(0.541196100)
88 #define FIX_0_707106781 FIX(0.707106781)
89 #define FIX_1_306562965 FIX(1.306562965)
93 /* We can gain a little more speed, with a further compromise in accuracy,
94 * by omitting the addition in a descaling shift. This yields an incorrectly
95 * rounded result half the time...
98 #ifndef USE_ACCURATE_ROUNDING
100 #define DESCALE(x,n) RIGHT_SHIFT(x, n)
104 /* Multiply a DCTELEM variable by an INT32 constant, and immediately
105 * descale to yield a DCTELEM result.
108 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
112 * Perform the forward DCT on one block of samples.
116 jpeg_fdct_ifast (DCTELEM * data)
118 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
119 DCTELEM tmp10, tmp11, tmp12, tmp13;
120 DCTELEM z1, z2, z3, z4, z5, z11, z13;
125 /* Pass 1: process rows. */
128 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
129 tmp0 = dataptr[0] + dataptr[7];
130 tmp7 = dataptr[0] - dataptr[7];
131 tmp1 = dataptr[1] + dataptr[6];
132 tmp6 = dataptr[1] - dataptr[6];
133 tmp2 = dataptr[2] + dataptr[5];
134 tmp5 = dataptr[2] - dataptr[5];
135 tmp3 = dataptr[3] + dataptr[4];
136 tmp4 = dataptr[3] - dataptr[4];
140 tmp10 = tmp0 + tmp3; /* phase 2 */
145 dataptr[0] = tmp10 + tmp11; /* phase 3 */
146 dataptr[4] = tmp10 - tmp11;
148 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
149 dataptr[2] = tmp13 + z1; /* phase 5 */
150 dataptr[6] = tmp13 - z1;
154 tmp10 = tmp4 + tmp5; /* phase 2 */
158 /* The rotator is modified from fig 4-8 to avoid extra negations. */
159 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
160 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
161 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
162 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
164 z11 = tmp7 + z3; /* phase 5 */
167 dataptr[5] = z13 + z2; /* phase 6 */
168 dataptr[3] = z13 - z2;
169 dataptr[1] = z11 + z4;
170 dataptr[7] = z11 - z4;
172 dataptr += DCTSIZE; /* advance pointer to next row */
175 /* Pass 2: process columns. */
178 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
179 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
180 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
181 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
182 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
183 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
184 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
185 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
186 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
190 tmp10 = tmp0 + tmp3; /* phase 2 */
195 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
196 dataptr[DCTSIZE*4] = tmp10 - tmp11;
198 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
199 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
200 dataptr[DCTSIZE*6] = tmp13 - z1;
204 tmp10 = tmp4 + tmp5; /* phase 2 */
208 /* The rotator is modified from fig 4-8 to avoid extra negations. */
209 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
210 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
211 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
212 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
214 z11 = tmp7 + z3; /* phase 5 */
217 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
218 dataptr[DCTSIZE*3] = z13 - z2;
219 dataptr[DCTSIZE*1] = z11 + z4;
220 dataptr[DCTSIZE*7] = z11 - z4;
222 dataptr++; /* advance pointer to next column */