vrshoot

diff libs/libjpeg/jfdctint.c @ 0:b2f14e535253

initial commit
author John Tsiombikas <nuclear@member.fsf.org>
date Sat, 01 Feb 2014 19:58:19 +0200
parents
children
line diff
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/libs/libjpeg/jfdctint.c	Sat Feb 01 19:58:19 2014 +0200
     1.3 @@ -0,0 +1,283 @@
     1.4 +/*
     1.5 + * jfdctint.c
     1.6 + *
     1.7 + * Copyright (C) 1991-1996, Thomas G. Lane.
     1.8 + * This file is part of the Independent JPEG Group's software.
     1.9 + * For conditions of distribution and use, see the accompanying README file.
    1.10 + *
    1.11 + * This file contains a slow-but-accurate integer implementation of the
    1.12 + * forward DCT (Discrete Cosine Transform).
    1.13 + *
    1.14 + * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
    1.15 + * on each column.  Direct algorithms are also available, but they are
    1.16 + * much more complex and seem not to be any faster when reduced to code.
    1.17 + *
    1.18 + * This implementation is based on an algorithm described in
    1.19 + *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
    1.20 + *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
    1.21 + *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
    1.22 + * The primary algorithm described there uses 11 multiplies and 29 adds.
    1.23 + * We use their alternate method with 12 multiplies and 32 adds.
    1.24 + * The advantage of this method is that no data path contains more than one
    1.25 + * multiplication; this allows a very simple and accurate implementation in
    1.26 + * scaled fixed-point arithmetic, with a minimal number of shifts.
    1.27 + */
    1.28 +
    1.29 +#define JPEG_INTERNALS
    1.30 +#include "jinclude.h"
    1.31 +#include "jpeglib.h"
    1.32 +#include "jdct.h"		/* Private declarations for DCT subsystem */
    1.33 +
    1.34 +#ifdef DCT_ISLOW_SUPPORTED
    1.35 +
    1.36 +
    1.37 +/*
    1.38 + * This module is specialized to the case DCTSIZE = 8.
    1.39 + */
    1.40 +
    1.41 +#if DCTSIZE != 8
    1.42 +  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
    1.43 +#endif
    1.44 +
    1.45 +
    1.46 +/*
    1.47 + * The poop on this scaling stuff is as follows:
    1.48 + *
    1.49 + * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
    1.50 + * larger than the true DCT outputs.  The final outputs are therefore
    1.51 + * a factor of N larger than desired; since N=8 this can be cured by
    1.52 + * a simple right shift at the end of the algorithm.  The advantage of
    1.53 + * this arrangement is that we save two multiplications per 1-D DCT,
    1.54 + * because the y0 and y4 outputs need not be divided by sqrt(N).
    1.55 + * In the IJG code, this factor of 8 is removed by the quantization step
    1.56 + * (in jcdctmgr.c), NOT in this module.
    1.57 + *
    1.58 + * We have to do addition and subtraction of the integer inputs, which
    1.59 + * is no problem, and multiplication by fractional constants, which is
    1.60 + * a problem to do in integer arithmetic.  We multiply all the constants
    1.61 + * by CONST_SCALE and convert them to integer constants (thus retaining
    1.62 + * CONST_BITS bits of precision in the constants).  After doing a
    1.63 + * multiplication we have to divide the product by CONST_SCALE, with proper
    1.64 + * rounding, to produce the correct output.  This division can be done
    1.65 + * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
    1.66 + * as long as possible so that partial sums can be added together with
    1.67 + * full fractional precision.
    1.68 + *
    1.69 + * The outputs of the first pass are scaled up by PASS1_BITS bits so that
    1.70 + * they are represented to better-than-integral precision.  These outputs
    1.71 + * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
    1.72 + * with the recommended scaling.  (For 12-bit sample data, the intermediate
    1.73 + * array is INT32 anyway.)
    1.74 + *
    1.75 + * To avoid overflow of the 32-bit intermediate results in pass 2, we must
    1.76 + * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
    1.77 + * shows that the values given below are the most effective.
    1.78 + */
    1.79 +
    1.80 +#if BITS_IN_JSAMPLE == 8
    1.81 +#define CONST_BITS  13
    1.82 +#define PASS1_BITS  2
    1.83 +#else
    1.84 +#define CONST_BITS  13
    1.85 +#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
    1.86 +#endif
    1.87 +
    1.88 +/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
    1.89 + * causing a lot of useless floating-point operations at run time.
    1.90 + * To get around this we use the following pre-calculated constants.
    1.91 + * If you change CONST_BITS you may want to add appropriate values.
    1.92 + * (With a reasonable C compiler, you can just rely on the FIX() macro...)
    1.93 + */
    1.94 +
    1.95 +#if CONST_BITS == 13
    1.96 +#define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
    1.97 +#define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
    1.98 +#define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
    1.99 +#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
   1.100 +#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
   1.101 +#define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
   1.102 +#define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
   1.103 +#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
   1.104 +#define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
   1.105 +#define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
   1.106 +#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
   1.107 +#define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
   1.108 +#else
   1.109 +#define FIX_0_298631336  FIX(0.298631336)
   1.110 +#define FIX_0_390180644  FIX(0.390180644)
   1.111 +#define FIX_0_541196100  FIX(0.541196100)
   1.112 +#define FIX_0_765366865  FIX(0.765366865)
   1.113 +#define FIX_0_899976223  FIX(0.899976223)
   1.114 +#define FIX_1_175875602  FIX(1.175875602)
   1.115 +#define FIX_1_501321110  FIX(1.501321110)
   1.116 +#define FIX_1_847759065  FIX(1.847759065)
   1.117 +#define FIX_1_961570560  FIX(1.961570560)
   1.118 +#define FIX_2_053119869  FIX(2.053119869)
   1.119 +#define FIX_2_562915447  FIX(2.562915447)
   1.120 +#define FIX_3_072711026  FIX(3.072711026)
   1.121 +#endif
   1.122 +
   1.123 +
   1.124 +/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
   1.125 + * For 8-bit samples with the recommended scaling, all the variable
   1.126 + * and constant values involved are no more than 16 bits wide, so a
   1.127 + * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
   1.128 + * For 12-bit samples, a full 32-bit multiplication will be needed.
   1.129 + */
   1.130 +
   1.131 +#if BITS_IN_JSAMPLE == 8
   1.132 +#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
   1.133 +#else
   1.134 +#define MULTIPLY(var,const)  ((var) * (const))
   1.135 +#endif
   1.136 +
   1.137 +
   1.138 +/*
   1.139 + * Perform the forward DCT on one block of samples.
   1.140 + */
   1.141 +
   1.142 +GLOBAL(void)
   1.143 +jpeg_fdct_islow (DCTELEM * data)
   1.144 +{
   1.145 +  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
   1.146 +  INT32 tmp10, tmp11, tmp12, tmp13;
   1.147 +  INT32 z1, z2, z3, z4, z5;
   1.148 +  DCTELEM *dataptr;
   1.149 +  int ctr;
   1.150 +  SHIFT_TEMPS
   1.151 +
   1.152 +  /* Pass 1: process rows. */
   1.153 +  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
   1.154 +  /* furthermore, we scale the results by 2**PASS1_BITS. */
   1.155 +
   1.156 +  dataptr = data;
   1.157 +  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
   1.158 +    tmp0 = dataptr[0] + dataptr[7];
   1.159 +    tmp7 = dataptr[0] - dataptr[7];
   1.160 +    tmp1 = dataptr[1] + dataptr[6];
   1.161 +    tmp6 = dataptr[1] - dataptr[6];
   1.162 +    tmp2 = dataptr[2] + dataptr[5];
   1.163 +    tmp5 = dataptr[2] - dataptr[5];
   1.164 +    tmp3 = dataptr[3] + dataptr[4];
   1.165 +    tmp4 = dataptr[3] - dataptr[4];
   1.166 +    
   1.167 +    /* Even part per LL&M figure 1 --- note that published figure is faulty;
   1.168 +     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
   1.169 +     */
   1.170 +    
   1.171 +    tmp10 = tmp0 + tmp3;
   1.172 +    tmp13 = tmp0 - tmp3;
   1.173 +    tmp11 = tmp1 + tmp2;
   1.174 +    tmp12 = tmp1 - tmp2;
   1.175 +    
   1.176 +    dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
   1.177 +    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
   1.178 +    
   1.179 +    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
   1.180 +    dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
   1.181 +				   CONST_BITS-PASS1_BITS);
   1.182 +    dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
   1.183 +				   CONST_BITS-PASS1_BITS);
   1.184 +    
   1.185 +    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
   1.186 +     * cK represents cos(K*pi/16).
   1.187 +     * i0..i3 in the paper are tmp4..tmp7 here.
   1.188 +     */
   1.189 +    
   1.190 +    z1 = tmp4 + tmp7;
   1.191 +    z2 = tmp5 + tmp6;
   1.192 +    z3 = tmp4 + tmp6;
   1.193 +    z4 = tmp5 + tmp7;
   1.194 +    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
   1.195 +    
   1.196 +    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
   1.197 +    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
   1.198 +    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
   1.199 +    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
   1.200 +    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
   1.201 +    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
   1.202 +    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
   1.203 +    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
   1.204 +    
   1.205 +    z3 += z5;
   1.206 +    z4 += z5;
   1.207 +    
   1.208 +    dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
   1.209 +    dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
   1.210 +    dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
   1.211 +    dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
   1.212 +    
   1.213 +    dataptr += DCTSIZE;		/* advance pointer to next row */
   1.214 +  }
   1.215 +
   1.216 +  /* Pass 2: process columns.
   1.217 +   * We remove the PASS1_BITS scaling, but leave the results scaled up
   1.218 +   * by an overall factor of 8.
   1.219 +   */
   1.220 +
   1.221 +  dataptr = data;
   1.222 +  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
   1.223 +    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
   1.224 +    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
   1.225 +    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
   1.226 +    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
   1.227 +    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
   1.228 +    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
   1.229 +    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
   1.230 +    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
   1.231 +    
   1.232 +    /* Even part per LL&M figure 1 --- note that published figure is faulty;
   1.233 +     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
   1.234 +     */
   1.235 +    
   1.236 +    tmp10 = tmp0 + tmp3;
   1.237 +    tmp13 = tmp0 - tmp3;
   1.238 +    tmp11 = tmp1 + tmp2;
   1.239 +    tmp12 = tmp1 - tmp2;
   1.240 +    
   1.241 +    dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
   1.242 +    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
   1.243 +    
   1.244 +    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
   1.245 +    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
   1.246 +					   CONST_BITS+PASS1_BITS);
   1.247 +    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
   1.248 +					   CONST_BITS+PASS1_BITS);
   1.249 +    
   1.250 +    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
   1.251 +     * cK represents cos(K*pi/16).
   1.252 +     * i0..i3 in the paper are tmp4..tmp7 here.
   1.253 +     */
   1.254 +    
   1.255 +    z1 = tmp4 + tmp7;
   1.256 +    z2 = tmp5 + tmp6;
   1.257 +    z3 = tmp4 + tmp6;
   1.258 +    z4 = tmp5 + tmp7;
   1.259 +    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
   1.260 +    
   1.261 +    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
   1.262 +    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
   1.263 +    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
   1.264 +    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
   1.265 +    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
   1.266 +    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
   1.267 +    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
   1.268 +    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
   1.269 +    
   1.270 +    z3 += z5;
   1.271 +    z4 += z5;
   1.272 +    
   1.273 +    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
   1.274 +					   CONST_BITS+PASS1_BITS);
   1.275 +    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
   1.276 +					   CONST_BITS+PASS1_BITS);
   1.277 +    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
   1.278 +					   CONST_BITS+PASS1_BITS);
   1.279 +    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
   1.280 +					   CONST_BITS+PASS1_BITS);
   1.281 +    
   1.282 +    dataptr++;			/* advance pointer to next column */
   1.283 +  }
   1.284 +}
   1.285 +
   1.286 +#endif /* DCT_ISLOW_SUPPORTED */