vrshoot

annotate libs/libjpeg/jfdctflt.c @ 0:b2f14e535253

find changesets by author, revision, files, or words in the commit message
initial commit
author John Tsiombikas <nuclear@member.fsf.org>
date Sat, 01 Feb 2014 19:58:19 +0200
parents
children
rev   line source
nuclear@0 1 /*
nuclear@0 2 * jfdctflt.c
nuclear@0 3 *
nuclear@0 4 * Copyright (C) 1994-1996, Thomas G. Lane.
nuclear@0 5 * This file is part of the Independent JPEG Group's software.
nuclear@0 6 * For conditions of distribution and use, see the accompanying README file.
nuclear@0 7 *
nuclear@0 8 * This file contains a floating-point implementation of the
nuclear@0 9 * forward DCT (Discrete Cosine Transform).
nuclear@0 10 *
nuclear@0 11 * This implementation should be more accurate than either of the integer
nuclear@0 12 * DCT implementations. However, it may not give the same results on all
nuclear@0 13 * machines because of differences in roundoff behavior. Speed will depend
nuclear@0 14 * on the hardware's floating point capacity.
nuclear@0 15 *
nuclear@0 16 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
nuclear@0 17 * on each column. Direct algorithms are also available, but they are
nuclear@0 18 * much more complex and seem not to be any faster when reduced to code.
nuclear@0 19 *
nuclear@0 20 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
nuclear@0 21 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
nuclear@0 22 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
nuclear@0 23 * JPEG textbook (see REFERENCES section in file README). The following code
nuclear@0 24 * is based directly on figure 4-8 in P&M.
nuclear@0 25 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
nuclear@0 26 * possible to arrange the computation so that many of the multiplies are
nuclear@0 27 * simple scalings of the final outputs. These multiplies can then be
nuclear@0 28 * folded into the multiplications or divisions by the JPEG quantization
nuclear@0 29 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
nuclear@0 30 * to be done in the DCT itself.
nuclear@0 31 * The primary disadvantage of this method is that with a fixed-point
nuclear@0 32 * implementation, accuracy is lost due to imprecise representation of the
nuclear@0 33 * scaled quantization values. However, that problem does not arise if
nuclear@0 34 * we use floating point arithmetic.
nuclear@0 35 */
nuclear@0 36
nuclear@0 37 #define JPEG_INTERNALS
nuclear@0 38 #include "jinclude.h"
nuclear@0 39 #include "jpeglib.h"
nuclear@0 40 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@0 41
nuclear@0 42 #ifdef DCT_FLOAT_SUPPORTED
nuclear@0 43
nuclear@0 44
nuclear@0 45 /*
nuclear@0 46 * This module is specialized to the case DCTSIZE = 8.
nuclear@0 47 */
nuclear@0 48
nuclear@0 49 #if DCTSIZE != 8
nuclear@0 50 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@0 51 #endif
nuclear@0 52
nuclear@0 53
nuclear@0 54 /*
nuclear@0 55 * Perform the forward DCT on one block of samples.
nuclear@0 56 */
nuclear@0 57
nuclear@0 58 GLOBAL(void)
nuclear@0 59 jpeg_fdct_float (FAST_FLOAT * data)
nuclear@0 60 {
nuclear@0 61 FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
nuclear@0 62 FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
nuclear@0 63 FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
nuclear@0 64 FAST_FLOAT *dataptr;
nuclear@0 65 int ctr;
nuclear@0 66
nuclear@0 67 /* Pass 1: process rows. */
nuclear@0 68
nuclear@0 69 dataptr = data;
nuclear@0 70 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@0 71 tmp0 = dataptr[0] + dataptr[7];
nuclear@0 72 tmp7 = dataptr[0] - dataptr[7];
nuclear@0 73 tmp1 = dataptr[1] + dataptr[6];
nuclear@0 74 tmp6 = dataptr[1] - dataptr[6];
nuclear@0 75 tmp2 = dataptr[2] + dataptr[5];
nuclear@0 76 tmp5 = dataptr[2] - dataptr[5];
nuclear@0 77 tmp3 = dataptr[3] + dataptr[4];
nuclear@0 78 tmp4 = dataptr[3] - dataptr[4];
nuclear@0 79
nuclear@0 80 /* Even part */
nuclear@0 81
nuclear@0 82 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@0 83 tmp13 = tmp0 - tmp3;
nuclear@0 84 tmp11 = tmp1 + tmp2;
nuclear@0 85 tmp12 = tmp1 - tmp2;
nuclear@0 86
nuclear@0 87 dataptr[0] = tmp10 + tmp11; /* phase 3 */
nuclear@0 88 dataptr[4] = tmp10 - tmp11;
nuclear@0 89
nuclear@0 90 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@0 91 dataptr[2] = tmp13 + z1; /* phase 5 */
nuclear@0 92 dataptr[6] = tmp13 - z1;
nuclear@0 93
nuclear@0 94 /* Odd part */
nuclear@0 95
nuclear@0 96 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@0 97 tmp11 = tmp5 + tmp6;
nuclear@0 98 tmp12 = tmp6 + tmp7;
nuclear@0 99
nuclear@0 100 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@0 101 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
nuclear@0 102 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
nuclear@0 103 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
nuclear@0 104 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@0 105
nuclear@0 106 z11 = tmp7 + z3; /* phase 5 */
nuclear@0 107 z13 = tmp7 - z3;
nuclear@0 108
nuclear@0 109 dataptr[5] = z13 + z2; /* phase 6 */
nuclear@0 110 dataptr[3] = z13 - z2;
nuclear@0 111 dataptr[1] = z11 + z4;
nuclear@0 112 dataptr[7] = z11 - z4;
nuclear@0 113
nuclear@0 114 dataptr += DCTSIZE; /* advance pointer to next row */
nuclear@0 115 }
nuclear@0 116
nuclear@0 117 /* Pass 2: process columns. */
nuclear@0 118
nuclear@0 119 dataptr = data;
nuclear@0 120 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
nuclear@0 121 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
nuclear@0 122 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
nuclear@0 123 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
nuclear@0 124 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
nuclear@0 125 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
nuclear@0 126 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
nuclear@0 127 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
nuclear@0 128 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
nuclear@0 129
nuclear@0 130 /* Even part */
nuclear@0 131
nuclear@0 132 tmp10 = tmp0 + tmp3; /* phase 2 */
nuclear@0 133 tmp13 = tmp0 - tmp3;
nuclear@0 134 tmp11 = tmp1 + tmp2;
nuclear@0 135 tmp12 = tmp1 - tmp2;
nuclear@0 136
nuclear@0 137 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
nuclear@0 138 dataptr[DCTSIZE*4] = tmp10 - tmp11;
nuclear@0 139
nuclear@0 140 z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@0 141 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
nuclear@0 142 dataptr[DCTSIZE*6] = tmp13 - z1;
nuclear@0 143
nuclear@0 144 /* Odd part */
nuclear@0 145
nuclear@0 146 tmp10 = tmp4 + tmp5; /* phase 2 */
nuclear@0 147 tmp11 = tmp5 + tmp6;
nuclear@0 148 tmp12 = tmp6 + tmp7;
nuclear@0 149
nuclear@0 150 /* The rotator is modified from fig 4-8 to avoid extra negations. */
nuclear@0 151 z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
nuclear@0 152 z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
nuclear@0 153 z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
nuclear@0 154 z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
nuclear@0 155
nuclear@0 156 z11 = tmp7 + z3; /* phase 5 */
nuclear@0 157 z13 = tmp7 - z3;
nuclear@0 158
nuclear@0 159 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
nuclear@0 160 dataptr[DCTSIZE*3] = z13 - z2;
nuclear@0 161 dataptr[DCTSIZE*1] = z11 + z4;
nuclear@0 162 dataptr[DCTSIZE*7] = z11 - z4;
nuclear@0 163
nuclear@0 164 dataptr++; /* advance pointer to next column */
nuclear@0 165 }
nuclear@0 166 }
nuclear@0 167
nuclear@0 168 #endif /* DCT_FLOAT_SUPPORTED */