1 #if !defined(_FX_JPEG_TURBO_) 2 /* 3 * jfdctfst.c 4 * 5 * Copyright (C) 1994-1996, Thomas G. Lane. 6 * This file is part of the Independent JPEG Group's software. 7 * For conditions of distribution and use, see the accompanying README file. 8 * 9 * This file contains a fast, not so accurate integer implementation of the 10 * forward DCT (Discrete Cosine Transform). 11 * 12 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT 13 * on each column. Direct algorithms are also available, but they are 14 * much more complex and seem not to be any faster when reduced to code. 15 * 16 * This implementation is based on Arai, Agui, and Nakajima's algorithm for 17 * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in 18 * Japanese, but the algorithm is described in the Pennebaker & Mitchell 19 * JPEG textbook (see REFERENCES section in file README). The following code 20 * is based directly on figure 4-8 in P&M. 21 * While an 8-point DCT cannot be done in less than 11 multiplies, it is 22 * possible to arrange the computation so that many of the multiplies are 23 * simple scalings of the final outputs. These multiplies can then be 24 * folded into the multiplications or divisions by the JPEG quantization 25 * table entries. The AA&N method leaves only 5 multiplies and 29 adds 26 * to be done in the DCT itself. 27 * The primary disadvantage of this method is that with fixed-point math, 28 * accuracy is lost due to imprecise representation of the scaled 29 * quantization values. The smaller the quantization table entry, the less 30 * precise the scaled value, so this implementation does worse with high- 31 * quality-setting files than with low-quality ones. 32 */ 33 34 #define JPEG_INTERNALS 35 #include "jinclude.h" 36 #include "jpeglib.h" 37 #include "jdct.h" /* Private declarations for DCT subsystem */ 38 39 #ifdef DCT_IFAST_SUPPORTED 40 41 42 /* 43 * This module is specialized to the case DCTSIZE = 8. 44 */ 45 46 #if DCTSIZE != 8 47 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ 48 #endif 49 50 51 /* Scaling decisions are generally the same as in the LL&M algorithm; 52 * see jfdctint.c for more details. However, we choose to descale 53 * (right shift) multiplication products as soon as they are formed, 54 * rather than carrying additional fractional bits into subsequent additions. 55 * This compromises accuracy slightly, but it lets us save a few shifts. 56 * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) 57 * everywhere except in the multiplications proper; this saves a good deal 58 * of work on 16-bit-int machines. 59 * 60 * Again to save a few shifts, the intermediate results between pass 1 and 61 * pass 2 are not upscaled, but are represented only to integral precision. 62 * 63 * A final compromise is to represent the multiplicative constants to only 64 * 8 fractional bits, rather than 13. This saves some shifting work on some 65 * machines, and may also reduce the cost of multiplication (since there 66 * are fewer one-bits in the constants). 67 */ 68 69 #define CONST_BITS 8 70 71 72 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus 73 * causing a lot of useless floating-point operations at run time. 74 * To get around this we use the following pre-calculated constants. 75 * If you change CONST_BITS you may want to add appropriate values. 76 * (With a reasonable C compiler, you can just rely on the FIX() macro...) 77 */ 78 79 #if CONST_BITS == 8 80 #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */ 81 #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */ 82 #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */ 83 #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */ 84 #else 85 #define FIX_0_382683433 FIX(0.382683433) 86 #define FIX_0_541196100 FIX(0.541196100) 87 #define FIX_0_707106781 FIX(0.707106781) 88 #define FIX_1_306562965 FIX(1.306562965) 89 #endif 90 91 92 /* We can gain a little more speed, with a further compromise in accuracy, 93 * by omitting the addition in a descaling shift. This yields an incorrectly 94 * rounded result half the time... 95 */ 96 97 #ifndef USE_ACCURATE_ROUNDING 98 #undef DESCALE 99 #define DESCALE(x,n) RIGHT_SHIFT(x, n) 100 #endif 101 102 103 /* Multiply a DCTELEM variable by an INT32 constant, and immediately 104 * descale to yield a DCTELEM result. 105 */ 106 107 #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) 108 109 110 /* 111 * Perform the forward DCT on one block of samples. 112 */ 113 114 GLOBAL(void) 115 jpeg_fdct_ifast (DCTELEM * data) 116 { 117 DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 118 DCTELEM tmp10, tmp11, tmp12, tmp13; 119 DCTELEM z1, z2, z3, z4, z5, z11, z13; 120 DCTELEM *dataptr; 121 int ctr; 122 SHIFT_TEMPS 123 124 /* Pass 1: process rows. */ 125 126 dataptr = data; 127 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { 128 tmp0 = dataptr[0] + dataptr[7]; 129 tmp7 = dataptr[0] - dataptr[7]; 130 tmp1 = dataptr[1] + dataptr[6]; 131 tmp6 = dataptr[1] - dataptr[6]; 132 tmp2 = dataptr[2] + dataptr[5]; 133 tmp5 = dataptr[2] - dataptr[5]; 134 tmp3 = dataptr[3] + dataptr[4]; 135 tmp4 = dataptr[3] - dataptr[4]; 136 137 /* Even part */ 138 139 tmp10 = tmp0 + tmp3; /* phase 2 */ 140 tmp13 = tmp0 - tmp3; 141 tmp11 = tmp1 + tmp2; 142 tmp12 = tmp1 - tmp2; 143 144 dataptr[0] = tmp10 + tmp11; /* phase 3 */ 145 dataptr[4] = tmp10 - tmp11; 146 147 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ 148 dataptr[2] = tmp13 + z1; /* phase 5 */ 149 dataptr[6] = tmp13 - z1; 150 151 /* Odd part */ 152 153 tmp10 = tmp4 + tmp5; /* phase 2 */ 154 tmp11 = tmp5 + tmp6; 155 tmp12 = tmp6 + tmp7; 156 157 /* The rotator is modified from fig 4-8 to avoid extra negations. */ 158 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ 159 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ 160 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ 161 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ 162 163 z11 = tmp7 + z3; /* phase 5 */ 164 z13 = tmp7 - z3; 165 166 dataptr[5] = z13 + z2; /* phase 6 */ 167 dataptr[3] = z13 - z2; 168 dataptr[1] = z11 + z4; 169 dataptr[7] = z11 - z4; 170 171 dataptr += DCTSIZE; /* advance pointer to next row */ 172 } 173 174 /* Pass 2: process columns. */ 175 176 dataptr = data; 177 for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { 178 tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; 179 tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; 180 tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; 181 tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; 182 tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; 183 tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; 184 tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; 185 tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; 186 187 /* Even part */ 188 189 tmp10 = tmp0 + tmp3; /* phase 2 */ 190 tmp13 = tmp0 - tmp3; 191 tmp11 = tmp1 + tmp2; 192 tmp12 = tmp1 - tmp2; 193 194 dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ 195 dataptr[DCTSIZE*4] = tmp10 - tmp11; 196 197 z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ 198 dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ 199 dataptr[DCTSIZE*6] = tmp13 - z1; 200 201 /* Odd part */ 202 203 tmp10 = tmp4 + tmp5; /* phase 2 */ 204 tmp11 = tmp5 + tmp6; 205 tmp12 = tmp6 + tmp7; 206 207 /* The rotator is modified from fig 4-8 to avoid extra negations. */ 208 z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ 209 z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ 210 z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ 211 z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ 212 213 z11 = tmp7 + z3; /* phase 5 */ 214 z13 = tmp7 - z3; 215 216 dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ 217 dataptr[DCTSIZE*3] = z13 - z2; 218 dataptr[DCTSIZE*1] = z11 + z4; 219 dataptr[DCTSIZE*7] = z11 - z4; 220 221 dataptr++; /* advance pointer to next column */ 222 } 223 } 224 225 #endif /* DCT_IFAST_SUPPORTED */ 226 227 #endif //_FX_JPEG_TURBO_ 228