jfdctflt.c

00001 /*
00002  * jfdctflt.c
00003  *
00004  * Copyright (C) 1994-1996, Thomas G. Lane.
00005  * Modified 2003-2015 by Guido Vollbeding.
00006  * This file is part of the Independent JPEG Group's software.
00007  * For conditions of distribution and use, see the accompanying README file.
00008  *
00009  * This file contains a floating-point implementation of the
00010  * forward DCT (Discrete Cosine Transform).
00011  *
00012  * This implementation should be more accurate than either of the integer
00013  * DCT implementations.  However, it may not give the same results on all
00014  * machines because of differences in roundoff behavior.  Speed will depend
00015  * on the hardware's floating point capacity.
00016  *
00017  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
00018  * on each column.  Direct algorithms are also available, but they are
00019  * much more complex and seem not to be any faster when reduced to code.
00020  *
00021  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
00022  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
00023  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
00024  * JPEG textbook (see REFERENCES section in file README).  The following code
00025  * is based directly on figure 4-8 in P&M.
00026  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
00027  * possible to arrange the computation so that many of the multiplies are
00028  * simple scalings of the final outputs.  These multiplies can then be
00029  * folded into the multiplications or divisions by the JPEG quantization
00030  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
00031  * to be done in the DCT itself.
00032  * The primary disadvantage of this method is that with a fixed-point
00033  * implementation, accuracy is lost due to imprecise representation of the
00034  * scaled quantization values.  However, that problem does not arise if
00035  * we use floating point arithmetic.
00036  */
00037 
00038 #define JPEG_INTERNALS
00039 #include "jinclude.h"
00040 #include "jpeglib.h"
00041 #include "jdct.h"       /* Private declarations for DCT subsystem */
00042 
00043 #ifdef DCT_FLOAT_SUPPORTED
00044 
00045 
00046 /*
00047  * This module is specialized to the case DCTSIZE = 8.
00048  */
00049 
00050 #if DCTSIZE != 8
00051   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00052 #endif
00053 
00054 
00055 /*
00056  * Perform the forward DCT on one block of samples.
00057  *
00058  * cK represents cos(K*pi/16).
00059  */
00060 
00061 GLOBAL(void)
00062 jpeg_fdct_float (FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00063 {
00064   FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00065   FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
00066   FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
00067   FAST_FLOAT *dataptr;
00068   JSAMPROW elemptr;
00069   int ctr;
00070 
00071   /* Pass 1: process rows. */
00072 
00073   dataptr = data;
00074   for (ctr = 0; ctr < DCTSIZE; ctr++) {
00075     elemptr = sample_data[ctr] + start_col;
00076 
00077     /* Load data into workspace */
00078     tmp0 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]));
00079     tmp7 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]));
00080     tmp1 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]));
00081     tmp6 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]));
00082     tmp2 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]));
00083     tmp5 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]));
00084     tmp3 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]));
00085     tmp4 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]));
00086 
00087     /* Even part */
00088 
00089     tmp10 = tmp0 + tmp3;    /* phase 2 */
00090     tmp13 = tmp0 - tmp3;
00091     tmp11 = tmp1 + tmp2;
00092     tmp12 = tmp1 - tmp2;
00093 
00094     /* Apply unsigned->signed conversion. */
00095     dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
00096     dataptr[4] = tmp10 - tmp11;
00097 
00098     z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
00099     dataptr[2] = tmp13 + z1;    /* phase 5 */
00100     dataptr[6] = tmp13 - z1;
00101 
00102     /* Odd part */
00103 
00104     tmp10 = tmp4 + tmp5;    /* phase 2 */
00105     tmp11 = tmp5 + tmp6;
00106     tmp12 = tmp6 + tmp7;
00107 
00108     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00109     z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
00110     z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
00111     z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
00112     z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
00113 
00114     z11 = tmp7 + z3;        /* phase 5 */
00115     z13 = tmp7 - z3;
00116 
00117     dataptr[5] = z13 + z2;  /* phase 6 */
00118     dataptr[3] = z13 - z2;
00119     dataptr[1] = z11 + z4;
00120     dataptr[7] = z11 - z4;
00121 
00122     dataptr += DCTSIZE;     /* advance pointer to next row */
00123   }
00124 
00125   /* Pass 2: process columns. */
00126 
00127   dataptr = data;
00128   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00129     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00130     tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00131     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00132     tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00133     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00134     tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00135     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00136     tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00137 
00138     /* Even part */
00139 
00140     tmp10 = tmp0 + tmp3;    /* phase 2 */
00141     tmp13 = tmp0 - tmp3;
00142     tmp11 = tmp1 + tmp2;
00143     tmp12 = tmp1 - tmp2;
00144 
00145     dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
00146     dataptr[DCTSIZE*4] = tmp10 - tmp11;
00147 
00148     z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
00149     dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
00150     dataptr[DCTSIZE*6] = tmp13 - z1;
00151 
00152     /* Odd part */
00153 
00154     tmp10 = tmp4 + tmp5;    /* phase 2 */
00155     tmp11 = tmp5 + tmp6;
00156     tmp12 = tmp6 + tmp7;
00157 
00158     /* The rotator is modified from fig 4-8 to avoid extra negations. */
00159     z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
00160     z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
00161     z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
00162     z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
00163 
00164     z11 = tmp7 + z3;        /* phase 5 */
00165     z13 = tmp7 - z3;
00166 
00167     dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
00168     dataptr[DCTSIZE*3] = z13 - z2;
00169     dataptr[DCTSIZE*1] = z11 + z4;
00170     dataptr[DCTSIZE*7] = z11 - z4;
00171 
00172     dataptr++;          /* advance pointer to next column */
00173   }
00174 }
00175 
00176 #endif /* DCT_FLOAT_SUPPORTED */