istereo2

annotate libs/libjpeg/jfdctfst.c @ 13:ea928c313344

find changesets by author, revision, files, or words in the commit message
foo
author John Tsiombikas <nuclear@member.fsf.org>
date Mon, 28 Sep 2015 19:04:50 +0300
parents
children
rev   line source
nuclear@2 1 /*
nuclear@2 2 * jfdctfst.c
nuclear@2 3 *
nuclear@2 4 * Copyright (C) 1994-1996, Thomas G. Lane.
nuclear@2 5 * This file is part of the Independent JPEG Group's software.
nuclear@2 6 * For conditions of distribution and use, see the accompanying README file.
nuclear@2 7 *
nuclear@2 8 * This file contains a fast, not so accurate integer implementation of the
nuclear@2 9 * forward DCT (Discrete Cosine Transform).
nuclear@2 10 *
nuclear@2 11 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
nuclear@2 12 * on each column. Direct algorithms are also available, but they are
nuclear@2 13 * much more complex and seem not to be any faster when reduced to code.
nuclear@2 14 *
nuclear@2 15 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
nuclear@2 16 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
nuclear@2 17 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
nuclear@2 18 * JPEG textbook (see REFERENCES section in file README). The following code
nuclear@2 19 * is based directly on figure 4-8 in P&M.
nuclear@2 20 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
nuclear@2 21 * possible to arrange the computation so that many of the multiplies are
nuclear@2 22 * simple scalings of the final outputs. These multiplies can then be
nuclear@2 23 * folded into the multiplications or divisions by the JPEG quantization
nuclear@2 24 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
nuclear@2 25 * to be done in the DCT itself.
nuclear@2 26 * The primary disadvantage of this method is that with fixed-point math,
nuclear@2 27 * accuracy is lost due to imprecise representation of the scaled
nuclear@2 28 * quantization values. The smaller the quantization table entry, the less
nuclear@2 29 * precise the scaled value, so this implementation does worse with high-
nuclear@2 30 * quality-setting files than with low-quality ones.
nuclear@2 31 */
nuclear@2 32
nuclear@2 33 #define JPEG_INTERNALS
nuclear@2 34 #include "jinclude.h"
nuclear@2 35 #include "jpeglib.h"
nuclear@2 36 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@2 37
nuclear@2 38 #ifdef DCT_IFAST_SUPPORTED
nuclear@2 39
nuclear@2 40
nuclear@2 41 /*
nuclear@2 42 * This module is specialized to the case DCTSIZE = 8.
nuclear@2 43 */
nuclear@2 44
nuclear@2 45 #if DCTSIZE != 8
nuclear@2 46 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@2 47 #endif
nuclear@2 48
nuclear@2 49
nuclear@2 50 /* Scaling decisions are generally the same as in the LL&M algorithm;
nuclear@2 51 * see jfdctint.c for more details. However, we choose to descale
nuclear@2 52 * (right shift) multiplication products as soon as they are formed,
nuclear@2 53 * rather than carrying additional fractional bits into subsequent additions.
nuclear@2 54 * This compromises accuracy slightly, but it lets us save a few shifts.
nuclear@2 55 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
nuclear@2 56 * everywhere except in the multiplications proper; this saves a good deal
nuclear@2 57 * of work on 16-bit-int machines.
nuclear@2 58 *
nuclear@2 59 * Again to save a few shifts, the intermediate results between pass 1 and
nuclear@2 60 * pass 2 are not upscaled, but are represented only to integral precision.
nuclear@2 61 *
nuclear@2 62 * A final compromise is to represent the multiplicative constants to only
nuclear@2 63 * 8 fractional bits, rather than 13. This saves some shifting work on some
nuclear@2 64 * machines, and may also reduce the cost of multiplication (since there
nuclear@2 65 * are fewer one-bits in the constants).
nuclear@2 66 */
nuclear@2 67
nuclear@2 68 #define CONST_BITS 8
nuclear@2 69
nuclear@2 70
nuclear@2 71 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
nuclear@2 72 * causing a lot of useless floating-point operations at run time.
nuclear@2 73 * To get around this we use the following pre-calculated constants.
nuclear@2 74 * If you change CONST_BITS you may want to add appropriate values.
nuclear@2 75 * (With a reasonable C compiler, you can just rely on the FIX() macro...)
nuclear@2 76 */
nuclear@2 77
nuclear@2 78 #if CONST_BITS == 8
nuclear@2 79 #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */
nuclear@2 80 #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */
nuclear@2 81 #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */
nuclear@2 82 #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */
nuclear@2 83 #else
nuclear@2 84 #define FIX_0_382683433 FIX(0.382683433)
nuclear@2 85 #define FIX_0_541196100 FIX(0.541196100)
nuclear@2 86 #define FIX_0_707106781 FIX(0.707106781)
nuclear@2 87 #define FIX_1_306562965 FIX(1.306562965)
nuclear@2 88 #endif
nuclear@2 89
nuclear@2 90
nuclear@2 91 /* We can gain a little more speed, with a further compromise in accuracy,
nuclear@2 92 * by omitting the addition in a descaling shift. This yields an incorrectly
nuclear@2 93 * rounded result half the time...
nuclear@2 94 */
nuclear@2 95
nuclear@2 96 #ifndef USE_ACCURATE_ROUNDING
nuclear@2 97 #undef DESCALE
nuclear@2 98 #define DESCALE(x,n) RIGHT_SHIFT(x, n)
nuclear@2 99 #endif
nuclear@2 100
nuclear@2 101
nuclear@2 102 /* Multiply a DCTELEM variable by an INT32 constant, and immediately
nuclear@2 103 * descale to yield a DCTELEM result.
nuclear@2 104 */
nuclear@2 105
nuclear@2 106 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
nuclear@2 107
nuclear@2 108
nuclear@2 109 /*
nuclear@2 110 * Perform the forward DCT on one block of samples.
nuclear@2 111 */
nuclear@2 112
nuclear@2 113 GLOBAL(void)
nuclear@2 114 jpeg_fdct_ifast (DCTELEM * data)
nuclear@2 115 {
nuclear@2 116 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
nuclear@2 117 DCTELEM tmp10, tmp11, tmp12, tmp13;
nuclear@2 118 DCTELEM z1, z2, z3, z4, z5, z11, z13;
nuclear@2 119 DCTELEM *dataptr;
nuclear@2 120 int ctr;
nuclear@2 121 SHIFT_TEMPS
nuclear@2 122
nuclear@2 123 /* Pass 1: process rows. */
nuclear@2 124
nuclear@2 125 dataptr = data;
nuclear@2 126 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@2 127 tmp0 = dataptr[0] + dataptr[7];
nuclear@2 128 tmp7 = dataptr[0] - dataptr[7];
nuclear@2 129 tmp1 = dataptr[1] + dataptr[6];
nuclear@2 130 tmp6 = dataptr[1] - dataptr[6];
nuclear@2 131 tmp2 = dataptr[2] + dataptr[5];
nuclear@2 132 tmp5 = dataptr[2] - dataptr[5];
nuclear@2 133 tmp3 = dataptr[3] + dataptr[4];
nuclear@2 134 tmp4 = dataptr[3] - dataptr[4];
nuclear@2 135
nuclear@2 136 /* Even part */
nuclear@2 137
nuclear@2 138 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@2 139 tmp13 = tmp0 - tmp3;
nuclear@2 140 tmp11 = tmp1 + tmp2;
nuclear@2 141 tmp12 = tmp1 - tmp2;
nuclear@2 142
nuclear@2 143 dataptr[0] = tmp10 + tmp11; /* phase 3 */
nuclear@2 144 dataptr[4] = tmp10 - tmp11;
nuclear@2 145
nuclear@2 146 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
nuclear@2 147 dataptr[2] = tmp13 + z1; /* phase 5 */
nuclear@2 148 dataptr[6] = tmp13 - z1;
nuclear@2 149
nuclear@2 150 /* Odd part */
nuclear@2 151
nuclear@2 152 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@2 153 tmp11 = tmp5 + tmp6;
nuclear@2 154 tmp12 = tmp6 + tmp7;
nuclear@2 155
nuclear@2 156 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@2 157 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
nuclear@2 158 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
nuclear@2 159 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
nuclear@2 160 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
nuclear@2 161
nuclear@2 162 z11 = tmp7 + z3; /* phase 5 */
nuclear@2 163 z13 = tmp7 - z3;
nuclear@2 164
nuclear@2 165 dataptr[5] = z13 + z2; /* phase 6 */
nuclear@2 166 dataptr[3] = z13 - z2;
nuclear@2 167 dataptr[1] = z11 + z4;
nuclear@2 168 dataptr[7] = z11 - z4;
nuclear@2 169
nuclear@2 170 dataptr += DCTSIZE; /* advance pointer to next row */
nuclear@2 171 }
nuclear@2 172
nuclear@2 173 /* Pass 2: process columns. */
nuclear@2 174
nuclear@2 175 dataptr = data;
nuclear@2 176 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@2 177 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
nuclear@2 178 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
nuclear@2 179 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
nuclear@2 180 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
nuclear@2 181 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
nuclear@2 182 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
nuclear@2 183 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
nuclear@2 184 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
nuclear@2 185
nuclear@2 186 /* Even part */
nuclear@2 187
nuclear@2 188 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@2 189 tmp13 = tmp0 - tmp3;
nuclear@2 190 tmp11 = tmp1 + tmp2;
nuclear@2 191 tmp12 = tmp1 - tmp2;
nuclear@2 192
nuclear@2 193 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
nuclear@2 194 dataptr[DCTSIZE*4] = tmp10 - tmp11;
nuclear@2 195
nuclear@2 196 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
nuclear@2 197 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
nuclear@2 198 dataptr[DCTSIZE*6] = tmp13 - z1;
nuclear@2 199
nuclear@2 200 /* Odd part */
nuclear@2 201
nuclear@2 202 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@2 203 tmp11 = tmp5 + tmp6;
nuclear@2 204 tmp12 = tmp6 + tmp7;
nuclear@2 205
nuclear@2 206 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@2 207 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
nuclear@2 208 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
nuclear@2 209 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
nuclear@2 210 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
nuclear@2 211
nuclear@2 212 z11 = tmp7 + z3; /* phase 5 */
nuclear@2 213 z13 = tmp7 - z3;
nuclear@2 214
nuclear@2 215 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
nuclear@2 216 dataptr[DCTSIZE*3] = z13 - z2;
nuclear@2 217 dataptr[DCTSIZE*1] = z11 + z4;
nuclear@2 218 dataptr[DCTSIZE*7] = z11 - z4;
nuclear@2 219
nuclear@2 220 dataptr++; /* advance pointer to next column */
nuclear@2 221 }
nuclear@2 222 }
nuclear@2 223
nuclear@2 224 #endif /* DCT_IFAST_SUPPORTED */