3dphotoshoot

annotate libs/libjpeg/jidctflt.c @ 14:06dc8b9b4f89

added libimago, libjpeg and libpng
author John Tsiombikas <nuclear@member.fsf.org>
date Sun, 07 Jun 2015 17:25:49 +0300
parents
children
rev   line source
nuclear@14 1 /*
nuclear@14 2 * jidctflt.c
nuclear@14 3 *
nuclear@14 4 * Copyright (C) 1994-1998, Thomas G. Lane.
nuclear@14 5 * This file is part of the Independent JPEG Group's software.
nuclear@14 6 * For conditions of distribution and use, see the accompanying README file.
nuclear@14 7 *
nuclear@14 8 * This file contains a floating-point implementation of the
nuclear@14 9 * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
nuclear@14 10 * must also perform dequantization of the input coefficients.
nuclear@14 11 *
nuclear@14 12 * This implementation should be more accurate than either of the integer
nuclear@14 13 * IDCT implementations. However, it may not give the same results on all
nuclear@14 14 * machines because of differences in roundoff behavior. Speed will depend
nuclear@14 15 * on the hardware's floating point capacity.
nuclear@14 16 *
nuclear@14 17 * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
nuclear@14 18 * on each row (or vice versa, but it's more convenient to emit a row at
nuclear@14 19 * a time). Direct algorithms are also available, but they are much more
nuclear@14 20 * complex and seem not to be any faster when reduced to code.
nuclear@14 21 *
nuclear@14 22 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
nuclear@14 23 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
nuclear@14 24 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
nuclear@14 25 * JPEG textbook (see REFERENCES section in file README). The following code
nuclear@14 26 * is based directly on figure 4-8 in P&M.
nuclear@14 27 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
nuclear@14 28 * possible to arrange the computation so that many of the multiplies are
nuclear@14 29 * simple scalings of the final outputs. These multiplies can then be
nuclear@14 30 * folded into the multiplications or divisions by the JPEG quantization
nuclear@14 31 * table entries. The AA&N method leaves only 5 multiplies and 29 adds
nuclear@14 32 * to be done in the DCT itself.
nuclear@14 33 * The primary disadvantage of this method is that with a fixed-point
nuclear@14 34 * implementation, accuracy is lost due to imprecise representation of the
nuclear@14 35 * scaled quantization values. However, that problem does not arise if
nuclear@14 36 * we use floating point arithmetic.
nuclear@14 37 */
nuclear@14 38
nuclear@14 39 #define JPEG_INTERNALS
nuclear@14 40 #include "jinclude.h"
nuclear@14 41 #include "jpeglib.h"
nuclear@14 42 #include "jdct.h" /* Private declarations for DCT subsystem */
nuclear@14 43
nuclear@14 44 #ifdef DCT_FLOAT_SUPPORTED
nuclear@14 45
nuclear@14 46
nuclear@14 47 /*
nuclear@14 48 * This module is specialized to the case DCTSIZE = 8.
nuclear@14 49 */
nuclear@14 50
nuclear@14 51 #if DCTSIZE != 8
nuclear@14 52 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
nuclear@14 53 #endif
nuclear@14 54
nuclear@14 55
nuclear@14 56 /* Dequantize a coefficient by multiplying it by the multiplier-table
nuclear@14 57 * entry; produce a float result.
nuclear@14 58 */
nuclear@14 59
nuclear@14 60 #define DEQUANTIZE(coef,quantval) (((FAST_FLOAT) (coef)) * (quantval))
nuclear@14 61
nuclear@14 62
nuclear@14 63 /*
nuclear@14 64 * Perform dequantization and inverse DCT on one block of coefficients.
nuclear@14 65 */
nuclear@14 66
nuclear@14 67 GLOBAL(void)
nuclear@14 68 jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
nuclear@14 69 JCOEFPTR coef_block,
nuclear@14 70 JSAMPARRAY output_buf, JDIMENSION output_col)
nuclear@14 71 {
nuclear@14 72 FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
nuclear@14 73 FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
nuclear@14 74 FAST_FLOAT z5, z10, z11, z12, z13;
nuclear@14 75 JCOEFPTR inptr;
nuclear@14 76 FLOAT_MULT_TYPE * quantptr;
nuclear@14 77 FAST_FLOAT * wsptr;
nuclear@14 78 JSAMPROW outptr;
nuclear@14 79 JSAMPLE *range_limit = IDCT_range_limit(cinfo);
nuclear@14 80 int ctr;
nuclear@14 81 FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
nuclear@14 82 SHIFT_TEMPS
nuclear@14 83
nuclear@14 84 /* Pass 1: process columns from input, store into work array. */
nuclear@14 85
nuclear@14 86 inptr = coef_block;
nuclear@14 87 quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
nuclear@14 88 wsptr = workspace;
nuclear@14 89 for (ctr = DCTSIZE; ctr > 0; ctr--) {
nuclear@14 90 /* Due to quantization, we will usually find that many of the input
nuclear@14 91 * coefficients are zero, especially the AC terms. We can exploit this
nuclear@14 92 * by short-circuiting the IDCT calculation for any column in which all
nuclear@14 93 * the AC terms are zero. In that case each output is equal to the
nuclear@14 94 * DC coefficient (with scale factor as needed).
nuclear@14 95 * With typical images and quantization tables, half or more of the
nuclear@14 96 * column DCT calculations can be simplified this way.
nuclear@14 97 */
nuclear@14 98
nuclear@14 99 if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
nuclear@14 100 inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
nuclear@14 101 inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
nuclear@14 102 inptr[DCTSIZE*7] == 0) {
nuclear@14 103 /* AC terms all zero */
nuclear@14 104 FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
nuclear@14 105
nuclear@14 106 wsptr[DCTSIZE*0] = dcval;
nuclear@14 107 wsptr[DCTSIZE*1] = dcval;
nuclear@14 108 wsptr[DCTSIZE*2] = dcval;
nuclear@14 109 wsptr[DCTSIZE*3] = dcval;
nuclear@14 110 wsptr[DCTSIZE*4] = dcval;
nuclear@14 111 wsptr[DCTSIZE*5] = dcval;
nuclear@14 112 wsptr[DCTSIZE*6] = dcval;
nuclear@14 113 wsptr[DCTSIZE*7] = dcval;
nuclear@14 114
nuclear@14 115 inptr++; /* advance pointers to next column */
nuclear@14 116 quantptr++;
nuclear@14 117 wsptr++;
nuclear@14 118 continue;
nuclear@14 119 }
nuclear@14 120
nuclear@14 121 /* Even part */
nuclear@14 122
nuclear@14 123 tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
nuclear@14 124 tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
nuclear@14 125 tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
nuclear@14 126 tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
nuclear@14 127
nuclear@14 128 tmp10 = tmp0 + tmp2; /* phase 3 */
nuclear@14 129 tmp11 = tmp0 - tmp2;
nuclear@14 130
nuclear@14 131 tmp13 = tmp1 + tmp3; /* phases 5-3 */
nuclear@14 132 tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
nuclear@14 133
nuclear@14 134 tmp0 = tmp10 + tmp13; /* phase 2 */
nuclear@14 135 tmp3 = tmp10 - tmp13;
nuclear@14 136 tmp1 = tmp11 + tmp12;
nuclear@14 137 tmp2 = tmp11 - tmp12;
nuclear@14 138
nuclear@14 139 /* Odd part */
nuclear@14 140
nuclear@14 141 tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
nuclear@14 142 tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
nuclear@14 143 tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
nuclear@14 144 tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
nuclear@14 145
nuclear@14 146 z13 = tmp6 + tmp5; /* phase 6 */
nuclear@14 147 z10 = tmp6 - tmp5;
nuclear@14 148 z11 = tmp4 + tmp7;
nuclear@14 149 z12 = tmp4 - tmp7;
nuclear@14 150
nuclear@14 151 tmp7 = z11 + z13; /* phase 5 */
nuclear@14 152 tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
nuclear@14 153
nuclear@14 154 z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
nuclear@14 155 tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
nuclear@14 156 tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
nuclear@14 157
nuclear@14 158 tmp6 = tmp12 - tmp7; /* phase 2 */
nuclear@14 159 tmp5 = tmp11 - tmp6;
nuclear@14 160 tmp4 = tmp10 + tmp5;
nuclear@14 161
nuclear@14 162 wsptr[DCTSIZE*0] = tmp0 + tmp7;
nuclear@14 163 wsptr[DCTSIZE*7] = tmp0 - tmp7;
nuclear@14 164 wsptr[DCTSIZE*1] = tmp1 + tmp6;
nuclear@14 165 wsptr[DCTSIZE*6] = tmp1 - tmp6;
nuclear@14 166 wsptr[DCTSIZE*2] = tmp2 + tmp5;
nuclear@14 167 wsptr[DCTSIZE*5] = tmp2 - tmp5;
nuclear@14 168 wsptr[DCTSIZE*4] = tmp3 + tmp4;
nuclear@14 169 wsptr[DCTSIZE*3] = tmp3 - tmp4;
nuclear@14 170
nuclear@14 171 inptr++; /* advance pointers to next column */
nuclear@14 172 quantptr++;
nuclear@14 173 wsptr++;
nuclear@14 174 }
nuclear@14 175
nuclear@14 176 /* Pass 2: process rows from work array, store into output array. */
nuclear@14 177 /* Note that we must descale the results by a factor of 8 == 2**3. */
nuclear@14 178
nuclear@14 179 wsptr = workspace;
nuclear@14 180 for (ctr = 0; ctr < DCTSIZE; ctr++) {
nuclear@14 181 outptr = output_buf[ctr] + output_col;
nuclear@14 182 /* Rows of zeroes can be exploited in the same way as we did with columns.
nuclear@14 183 * However, the column calculation has created many nonzero AC terms, so
nuclear@14 184 * the simplification applies less often (typically 5% to 10% of the time).
nuclear@14 185 * And testing floats for zero is relatively expensive, so we don't bother.
nuclear@14 186 */
nuclear@14 187
nuclear@14 188 /* Even part */
nuclear@14 189
nuclear@14 190 tmp10 = wsptr[0] + wsptr[4];
nuclear@14 191 tmp11 = wsptr[0] - wsptr[4];
nuclear@14 192
nuclear@14 193 tmp13 = wsptr[2] + wsptr[6];
nuclear@14 194 tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
nuclear@14 195
nuclear@14 196 tmp0 = tmp10 + tmp13;
nuclear@14 197 tmp3 = tmp10 - tmp13;
nuclear@14 198 tmp1 = tmp11 + tmp12;
nuclear@14 199 tmp2 = tmp11 - tmp12;
nuclear@14 200
nuclear@14 201 /* Odd part */
nuclear@14 202
nuclear@14 203 z13 = wsptr[5] + wsptr[3];
nuclear@14 204 z10 = wsptr[5] - wsptr[3];
nuclear@14 205 z11 = wsptr[1] + wsptr[7];
nuclear@14 206 z12 = wsptr[1] - wsptr[7];
nuclear@14 207
nuclear@14 208 tmp7 = z11 + z13;
nuclear@14 209 tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
nuclear@14 210
nuclear@14 211 z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
nuclear@14 212 tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
nuclear@14 213 tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
nuclear@14 214
nuclear@14 215 tmp6 = tmp12 - tmp7;
nuclear@14 216 tmp5 = tmp11 - tmp6;
nuclear@14 217 tmp4 = tmp10 + tmp5;
nuclear@14 218
nuclear@14 219 /* Final output stage: scale down by a factor of 8 and range-limit */
nuclear@14 220
nuclear@14 221 outptr[0] = range_limit[(int) DESCALE((INT32) (tmp0 + tmp7), 3)
nuclear@14 222 & RANGE_MASK];
nuclear@14 223 outptr[7] = range_limit[(int) DESCALE((INT32) (tmp0 - tmp7), 3)
nuclear@14 224 & RANGE_MASK];
nuclear@14 225 outptr[1] = range_limit[(int) DESCALE((INT32) (tmp1 + tmp6), 3)
nuclear@14 226 & RANGE_MASK];
nuclear@14 227 outptr[6] = range_limit[(int) DESCALE((INT32) (tmp1 - tmp6), 3)
nuclear@14 228 & RANGE_MASK];
nuclear@14 229 outptr[2] = range_limit[(int) DESCALE((INT32) (tmp2 + tmp5), 3)
nuclear@14 230 & RANGE_MASK];
nuclear@14 231 outptr[5] = range_limit[(int) DESCALE((INT32) (tmp2 - tmp5), 3)
nuclear@14 232 & RANGE_MASK];
nuclear@14 233 outptr[4] = range_limit[(int) DESCALE((INT32) (tmp3 + tmp4), 3)
nuclear@14 234 & RANGE_MASK];
nuclear@14 235 outptr[3] = range_limit[(int) DESCALE((INT32) (tmp3 - tmp4), 3)
nuclear@14 236 & RANGE_MASK];
nuclear@14 237
nuclear@14 238 wsptr += DCTSIZE; /* advance pointer to next row */
nuclear@14 239 }
nuclear@14 240 }
nuclear@14 241
nuclear@14 242 #endif /* DCT_FLOAT_SUPPORTED */