1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/libs/libjpeg/jidctflt.c	Sat Feb 01 19:58:19 2014 +0200
     1.3 @@ -0,0 +1,242 @@
     1.4 +/*
     1.5 + * jidctflt.c
     1.6 + *
     1.7 + * Copyright (C) 1994-1998, 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 floating-point implementation of the
    1.12 + * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
    1.13 + * must also perform dequantization of the input coefficients.
    1.14 + *
    1.15 + * This implementation should be more accurate than either of the integer
    1.16 + * IDCT implementations.  However, it may not give the same results on all
    1.17 + * machines because of differences in roundoff behavior.  Speed will depend
    1.18 + * on the hardware's floating point capacity.
    1.19 + *
    1.20 + * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
    1.21 + * on each row (or vice versa, but it's more convenient to emit a row at
    1.22 + * a time).  Direct algorithms are also available, but they are much more
    1.23 + * complex and seem not to be any faster when reduced to code.
    1.24 + *
    1.25 + * This implementation is based on Arai, Agui, and Nakajima's algorithm for
    1.26 + * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
    1.27 + * Japanese, but the algorithm is described in the Pennebaker & Mitchell
    1.28 + * JPEG textbook (see REFERENCES section in file README).  The following code
    1.29 + * is based directly on figure 4-8 in P&M.
    1.30 + * While an 8-point DCT cannot be done in less than 11 multiplies, it is
    1.31 + * possible to arrange the computation so that many of the multiplies are
    1.32 + * simple scalings of the final outputs.  These multiplies can then be
    1.33 + * folded into the multiplications or divisions by the JPEG quantization
    1.34 + * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
    1.35 + * to be done in the DCT itself.
    1.36 + * The primary disadvantage of this method is that with a fixed-point
    1.37 + * implementation, accuracy is lost due to imprecise representation of the
    1.38 + * scaled quantization values.  However, that problem does not arise if
    1.39 + * we use floating point arithmetic.
    1.40 + */
    1.41 +
    1.42 +#define JPEG_INTERNALS
    1.43 +#include "jinclude.h"
    1.44 +#include "jpeglib.h"
    1.45 +#include "jdct.h"		/* Private declarations for DCT subsystem */
    1.46 +
    1.47 +#ifdef DCT_FLOAT_SUPPORTED
    1.48 +
    1.49 +
    1.50 +/*
    1.51 + * This module is specialized to the case DCTSIZE = 8.
    1.52 + */
    1.53 +
    1.54 +#if DCTSIZE != 8
    1.55 +  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
    1.56 +#endif
    1.57 +
    1.58 +
    1.59 +/* Dequantize a coefficient by multiplying it by the multiplier-table
    1.60 + * entry; produce a float result.
    1.61 + */
    1.62 +
    1.63 +#define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
    1.64 +
    1.65 +
    1.66 +/*
    1.67 + * Perform dequantization and inverse DCT on one block of coefficients.
    1.68 + */
    1.69 +
    1.70 +GLOBAL(void)
    1.71 +jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
    1.72 +		 JCOEFPTR coef_block,
    1.73 +		 JSAMPARRAY output_buf, JDIMENSION output_col)
    1.74 +{
    1.75 +  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    1.76 +  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
    1.77 +  FAST_FLOAT z5, z10, z11, z12, z13;
    1.78 +  JCOEFPTR inptr;
    1.79 +  FLOAT_MULT_TYPE * quantptr;
    1.80 +  FAST_FLOAT * wsptr;
    1.81 +  JSAMPROW outptr;
    1.82 +  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
    1.83 +  int ctr;
    1.84 +  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
    1.85 +  SHIFT_TEMPS
    1.86 +
    1.87 +  /* Pass 1: process columns from input, store into work array. */
    1.88 +
    1.89 +  inptr = coef_block;
    1.90 +  quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
    1.91 +  wsptr = workspace;
    1.92 +  for (ctr = DCTSIZE; ctr > 0; ctr--) {
    1.93 +    /* Due to quantization, we will usually find that many of the input
    1.94 +     * coefficients are zero, especially the AC terms.  We can exploit this
    1.95 +     * by short-circuiting the IDCT calculation for any column in which all
    1.96 +     * the AC terms are zero.  In that case each output is equal to the
    1.97 +     * DC coefficient (with scale factor as needed).
    1.98 +     * With typical images and quantization tables, half or more of the
    1.99 +     * column DCT calculations can be simplified this way.
   1.100 +     */
   1.101 +    
   1.102 +    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
   1.103 +	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
   1.104 +	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
   1.105 +	inptr[DCTSIZE*7] == 0) {
   1.106 +      /* AC terms all zero */
   1.107 +      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
   1.108 +      
   1.109 +      wsptr[DCTSIZE*0] = dcval;
   1.110 +      wsptr[DCTSIZE*1] = dcval;
   1.111 +      wsptr[DCTSIZE*2] = dcval;
   1.112 +      wsptr[DCTSIZE*3] = dcval;
   1.113 +      wsptr[DCTSIZE*4] = dcval;
   1.114 +      wsptr[DCTSIZE*5] = dcval;
   1.115 +      wsptr[DCTSIZE*6] = dcval;
   1.116 +      wsptr[DCTSIZE*7] = dcval;
   1.117 +      
   1.118 +      inptr++;			/* advance pointers to next column */
   1.119 +      quantptr++;
   1.120 +      wsptr++;
   1.121 +      continue;
   1.122 +    }
   1.123 +    
   1.124 +    /* Even part */
   1.125 +
   1.126 +    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
   1.127 +    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
   1.128 +    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
   1.129 +    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
   1.130 +
   1.131 +    tmp10 = tmp0 + tmp2;	/* phase 3 */
   1.132 +    tmp11 = tmp0 - tmp2;
   1.133 +
   1.134 +    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
   1.135 +    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
   1.136 +
   1.137 +    tmp0 = tmp10 + tmp13;	/* phase 2 */
   1.138 +    tmp3 = tmp10 - tmp13;
   1.139 +    tmp1 = tmp11 + tmp12;
   1.140 +    tmp2 = tmp11 - tmp12;
   1.141 +    
   1.142 +    /* Odd part */
   1.143 +
   1.144 +    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
   1.145 +    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
   1.146 +    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
   1.147 +    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
   1.148 +
   1.149 +    z13 = tmp6 + tmp5;		/* phase 6 */
   1.150 +    z10 = tmp6 - tmp5;
   1.151 +    z11 = tmp4 + tmp7;
   1.152 +    z12 = tmp4 - tmp7;
   1.153 +
   1.154 +    tmp7 = z11 + z13;		/* phase 5 */
   1.155 +    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
   1.156 +
   1.157 +    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
   1.158 +    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
   1.159 +    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
   1.160 +
   1.161 +    tmp6 = tmp12 - tmp7;	/* phase 2 */
   1.162 +    tmp5 = tmp11 - tmp6;
   1.163 +    tmp4 = tmp10 + tmp5;
   1.164 +
   1.165 +    wsptr[DCTSIZE*0] = tmp0 + tmp7;
   1.166 +    wsptr[DCTSIZE*7] = tmp0 - tmp7;
   1.167 +    wsptr[DCTSIZE*1] = tmp1 + tmp6;
   1.168 +    wsptr[DCTSIZE*6] = tmp1 - tmp6;
   1.169 +    wsptr[DCTSIZE*2] = tmp2 + tmp5;
   1.170 +    wsptr[DCTSIZE*5] = tmp2 - tmp5;
   1.171 +    wsptr[DCTSIZE*4] = tmp3 + tmp4;
   1.172 +    wsptr[DCTSIZE*3] = tmp3 - tmp4;
   1.173 +
   1.174 +    inptr++;			/* advance pointers to next column */
   1.175 +    quantptr++;
   1.176 +    wsptr++;
   1.177 +  }
   1.178 +  
   1.179 +  /* Pass 2: process rows from work array, store into output array. */
   1.180 +  /* Note that we must descale the results by a factor of 8 == 2**3. */
   1.181 +
   1.182 +  wsptr = workspace;
   1.183 +  for (ctr = 0; ctr < DCTSIZE; ctr++) {
   1.184 +    outptr = output_buf[ctr] + output_col;
   1.185 +    /* Rows of zeroes can be exploited in the same way as we did with columns.
   1.186 +     * However, the column calculation has created many nonzero AC terms, so
   1.187 +     * the simplification applies less often (typically 5% to 10% of the time).
   1.188 +     * And testing floats for zero is relatively expensive, so we don't bother.
   1.189 +     */
   1.190 +    
   1.191 +    /* Even part */
   1.192 +
   1.193 +    tmp10 = wsptr[0] + wsptr[4];
   1.194 +    tmp11 = wsptr[0] - wsptr[4];
   1.195 +
   1.196 +    tmp13 = wsptr[2] + wsptr[6];
   1.197 +    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
   1.198 +
   1.199 +    tmp0 = tmp10 + tmp13;
   1.200 +    tmp3 = tmp10 - tmp13;
   1.201 +    tmp1 = tmp11 + tmp12;
   1.202 +    tmp2 = tmp11 - tmp12;
   1.203 +
   1.204 +    /* Odd part */
   1.205 +
   1.206 +    z13 = wsptr[5] + wsptr[3];
   1.207 +    z10 = wsptr[5] - wsptr[3];
   1.208 +    z11 = wsptr[1] + wsptr[7];
   1.209 +    z12 = wsptr[1] - wsptr[7];
   1.210 +
   1.211 +    tmp7 = z11 + z13;
   1.212 +    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
   1.213 +
   1.214 +    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
   1.215 +    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
   1.216 +    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
   1.217 +
   1.218 +    tmp6 = tmp12 - tmp7;
   1.219 +    tmp5 = tmp11 - tmp6;
   1.220 +    tmp4 = tmp10 + tmp5;
   1.221 +
   1.222 +    /* Final output stage: scale down by a factor of 8 and range-limit */
   1.223 +
   1.224 +    outptr[0] = range_limit[(int) DESCALE((INT32) (tmp0 + tmp7), 3)
   1.225 +			    & RANGE_MASK];
   1.226 +    outptr[7] = range_limit[(int) DESCALE((INT32) (tmp0 - tmp7), 3)
   1.227 +			    & RANGE_MASK];
   1.228 +    outptr[1] = range_limit[(int) DESCALE((INT32) (tmp1 + tmp6), 3)
   1.229 +			    & RANGE_MASK];
   1.230 +    outptr[6] = range_limit[(int) DESCALE((INT32) (tmp1 - tmp6), 3)
   1.231 +			    & RANGE_MASK];
   1.232 +    outptr[2] = range_limit[(int) DESCALE((INT32) (tmp2 + tmp5), 3)
   1.233 +			    & RANGE_MASK];
   1.234 +    outptr[5] = range_limit[(int) DESCALE((INT32) (tmp2 - tmp5), 3)
   1.235 +			    & RANGE_MASK];
   1.236 +    outptr[4] = range_limit[(int) DESCALE((INT32) (tmp3 + tmp4), 3)
   1.237 +			    & RANGE_MASK];
   1.238 +    outptr[3] = range_limit[(int) DESCALE((INT32) (tmp3 - tmp4), 3)
   1.239 +			    & RANGE_MASK];
   1.240 +    
   1.241 +    wsptr += DCTSIZE;		/* advance pointer to next row */
   1.242 +  }
   1.243 +}
   1.244 +
   1.245 +#endif /* DCT_FLOAT_SUPPORTED */