istereo

annotate libs/libjpeg/jfdctflt.c @ 28:c0ae8e668447

find changesets by author, revision, files, or words in the commit message
added vmath library
author John Tsiombikas <nuclear@mutantstargoat.com>
date Thu, 08 Sep 2011 08:30:42 +0300
parents
children
rev   line source
nuclear@26 1 /*
nuclear@26 2 * jfdctflt.c
nuclear@26 3 *
nuclear@26 4 * Copyright (C) 1994-1996, Thomas G. Lane.
nuclear@26 5 * This file is part of the Independent JPEG Group's software.
nuclear@26 6 * For conditions of distribution and use, see the accompanying README file.
nuclear@26 7 *
nuclear@26 8 * This file contains a floating-point implementation of the
nuclear@26 9 * forward DCT (Discrete Cosine Transform).
nuclear@26 10 *
nuclear@26 11 * This implementation should be more accurate than either of the integer
nuclear@26 12 * DCT implementations. However, it may not give the same results on all
nuclear@26 13 * machines because of differences in roundoff behavior. Speed will depend
nuclear@26 14 * on the hardware's floating point capacity.
nuclear@26 15 *
nuclear@26 16 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
nuclear@26 17 * on each column. Direct algorithms are also available, but they are
nuclear@26 18 * much more complex and seem not to be any faster when reduced to code.
nuclear@26 19 *
nuclear@26 20 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
nuclear@26 21 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
nuclear@26 22 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
nuclear@26 23 * JPEG textbook (see REFERENCES section in file README). The following code
nuclear@26 24 * is based directly on figure 4-8 in P&M.
nuclear@26 25 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
nuclear@26 26 * possible to arrange the computation so that many of the multiplies are
nuclear@26 27 * simple scalings of the final outputs. These multiplies can then be
nuclear@26 28 * folded into the multiplications or divisions by the JPEG quantization
nuclear@26 29 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
nuclear@26 30 * to be done in the DCT itself.
nuclear@26 31 * The primary disadvantage of this method is that with a fixed-point
nuclear@26 32 * implementation, accuracy is lost due to imprecise representation of the
nuclear@26 33 * scaled quantization values. However, that problem does not arise if
nuclear@26 34 * we use floating point arithmetic.
nuclear@26 35 */
nuclear@26 36
nuclear@26 37 #define JPEG_INTERNALS
nuclear@26 38 #include "jinclude.h"
nuclear@26 39 #include "jpeglib.h"
nuclear@26 40 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@26 41
nuclear@26 42 #ifdef DCT_FLOAT_SUPPORTED
nuclear@26 43
nuclear@26 44
nuclear@26 45 /*
nuclear@26 46 * This module is specialized to the case DCTSIZE = 8.
nuclear@26 47 */
nuclear@26 48
nuclear@26 49 #if DCTSIZE != 8
nuclear@26 50 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@26 51 #endif
nuclear@26 52
nuclear@26 53
nuclear@26 54 /*
nuclear@26 55 * Perform the forward DCT on one block of samples.
nuclear@26 56 */
nuclear@26 57
nuclear@26 58 GLOBAL(void)
nuclear@26 59 jpeg_fdct_float (FAST_FLOAT * data)
nuclear@26 60 {
nuclear@26 61 FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
nuclear@26 62 FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
nuclear@26 63 FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
nuclear@26 64 FAST_FLOAT *dataptr;
nuclear@26 65 int ctr;
nuclear@26 66
nuclear@26 67 /* Pass 1: process rows. */
nuclear@26 68
nuclear@26 69 dataptr = data;
nuclear@26 70 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@26 71 tmp0 = dataptr[0] + dataptr[7];
nuclear@26 72 tmp7 = dataptr[0] - dataptr[7];
nuclear@26 73 tmp1 = dataptr[1] + dataptr[6];
nuclear@26 74 tmp6 = dataptr[1] - dataptr[6];
nuclear@26 75 tmp2 = dataptr[2] + dataptr[5];
nuclear@26 76 tmp5 = dataptr[2] - dataptr[5];
nuclear@26 77 tmp3 = dataptr[3] + dataptr[4];
nuclear@26 78 tmp4 = dataptr[3] - dataptr[4];
nuclear@26 79
nuclear@26 80 /* Even part */
nuclear@26 81
nuclear@26 82 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@26 83 tmp13 = tmp0 - tmp3;
nuclear@26 84 tmp11 = tmp1 + tmp2;
nuclear@26 85 tmp12 = tmp1 - tmp2;
nuclear@26 86
nuclear@26 87 dataptr[0] = tmp10 + tmp11; /* phase 3 */
nuclear@26 88 dataptr[4] = tmp10 - tmp11;
nuclear@26 89
nuclear@26 90 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@26 91 dataptr[2] = tmp13 + z1; /* phase 5 */
nuclear@26 92 dataptr[6] = tmp13 - z1;
nuclear@26 93
nuclear@26 94 /* Odd part */
nuclear@26 95
nuclear@26 96 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@26 97 tmp11 = tmp5 + tmp6;
nuclear@26 98 tmp12 = tmp6 + tmp7;
nuclear@26 99
nuclear@26 100 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@26 101 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
nuclear@26 102 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
nuclear@26 103 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
nuclear@26 104 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@26 105
nuclear@26 106 z11 = tmp7 + z3; /* phase 5 */
nuclear@26 107 z13 = tmp7 - z3;
nuclear@26 108
nuclear@26 109 dataptr[5] = z13 + z2; /* phase 6 */
nuclear@26 110 dataptr[3] = z13 - z2;
nuclear@26 111 dataptr[1] = z11 + z4;
nuclear@26 112 dataptr[7] = z11 - z4;
nuclear@26 113
nuclear@26 114 dataptr += DCTSIZE; /* advance pointer to next row */
nuclear@26 115 }
nuclear@26 116
nuclear@26 117 /* Pass 2: process columns. */
nuclear@26 118
nuclear@26 119 dataptr = data;
nuclear@26 120 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@26 121 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
nuclear@26 122 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
nuclear@26 123 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
nuclear@26 124 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
nuclear@26 125 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
nuclear@26 126 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
nuclear@26 127 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
nuclear@26 128 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
nuclear@26 129
nuclear@26 130 /* Even part */
nuclear@26 131
nuclear@26 132 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@26 133 tmp13 = tmp0 - tmp3;
nuclear@26 134 tmp11 = tmp1 + tmp2;
nuclear@26 135 tmp12 = tmp1 - tmp2;
nuclear@26 136
nuclear@26 137 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
nuclear@26 138 dataptr[DCTSIZE*4] = tmp10 - tmp11;
nuclear@26 139
nuclear@26 140 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@26 141 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
nuclear@26 142 dataptr[DCTSIZE*6] = tmp13 - z1;
nuclear@26 143
nuclear@26 144 /* Odd part */
nuclear@26 145
nuclear@26 146 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@26 147 tmp11 = tmp5 + tmp6;
nuclear@26 148 tmp12 = tmp6 + tmp7;
nuclear@26 149
nuclear@26 150 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@26 151 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
nuclear@26 152 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
nuclear@26 153 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
nuclear@26 154 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@26 155
nuclear@26 156 z11 = tmp7 + z3; /* phase 5 */
nuclear@26 157 z13 = tmp7 - z3;
nuclear@26 158
nuclear@26 159 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
nuclear@26 160 dataptr[DCTSIZE*3] = z13 - z2;
nuclear@26 161 dataptr[DCTSIZE*1] = z11 + z4;
nuclear@26 162 dataptr[DCTSIZE*7] = z11 - z4;
nuclear@26 163
nuclear@26 164 dataptr++; /* advance pointer to next column */
nuclear@26 165 }
nuclear@26 166 }
nuclear@26 167
nuclear@26 168 #endif /* DCT_FLOAT_SUPPORTED */