jidctflt.c

00001 /*
00002  * jidctflt.c
00003  *
00004  * Copyright (C) 1994-1998, Thomas G. Lane.
00005  * Modified 2010-2015 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  * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
00011  * must also perform dequantization of the input coefficients.
00012  *
00013  * This implementation should be more accurate than either of the integer
00014  * IDCT implementations.  However, it may not give the same results on all
00015  * machines because of differences in roundoff behavior.  Speed will depend
00016  * on the hardware's floating point capacity.
00017  *
00018  * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
00019  * on each row (or vice versa, but it's more convenient to emit a row at
00020  * a time).  Direct algorithms are also available, but they are much more
00021  * complex and seem not to be any faster when reduced to code.
00022  *
00023  * This implementation is based on Arai, Agui, and Nakajima's algorithm for
00024  * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
00025  * Japanese, but the algorithm is described in the Pennebaker & Mitchell
00026  * JPEG textbook (see REFERENCES section in file README).  The following code
00027  * is based directly on figure 4-8 in P&M.
00028  * While an 8-point DCT cannot be done in less than 11 multiplies, it is
00029  * possible to arrange the computation so that many of the multiplies are
00030  * simple scalings of the final outputs.  These multiplies can then be
00031  * folded into the multiplications or divisions by the JPEG quantization
00032  * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
00033  * to be done in the DCT itself.
00034  * The primary disadvantage of this method is that with a fixed-point
00035  * implementation, accuracy is lost due to imprecise representation of the
00036  * scaled quantization values.  However, that problem does not arise if
00037  * we use floating point arithmetic.
00038  */
00039 
00040 #define JPEG_INTERNALS
00041 #include "jinclude.h"
00042 #include "jpeglib.h"
00043 #include "jdct.h"       /* Private declarations for DCT subsystem */
00044 
00045 #ifdef DCT_FLOAT_SUPPORTED
00046 
00047 
00048 /*
00049  * This module is specialized to the case DCTSIZE = 8.
00050  */
00051 
00052 #if DCTSIZE != 8
00053   Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
00054 #endif
00055 
00056 
00057 /* Dequantize a coefficient by multiplying it by the multiplier-table
00058  * entry; produce a float result.
00059  */
00060 
00061 #define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
00062 
00063 
00064 /*
00065  * Perform dequantization and inverse DCT on one block of coefficients.
00066  *
00067  * cK represents cos(K*pi/16).
00068  */
00069 
00070 GLOBAL(void)
00071 jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
00072          JCOEFPTR coef_block,
00073          JSAMPARRAY output_buf, JDIMENSION output_col)
00074 {
00075   FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
00076   FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
00077   FAST_FLOAT z5, z10, z11, z12, z13;
00078   JCOEFPTR inptr;
00079   FLOAT_MULT_TYPE * quantptr;
00080   FAST_FLOAT * wsptr;
00081   JSAMPROW outptr;
00082   JSAMPLE *range_limit = IDCT_range_limit(cinfo);
00083   int ctr;
00084   FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
00085 
00086   /* Pass 1: process columns from input, store into work array. */
00087 
00088   inptr = coef_block;
00089   quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
00090   wsptr = workspace;
00091   for (ctr = DCTSIZE; ctr > 0; ctr--) {
00092     /* Due to quantization, we will usually find that many of the input
00093      * coefficients are zero, especially the AC terms.  We can exploit this
00094      * by short-circuiting the IDCT calculation for any column in which all
00095      * the AC terms are zero.  In that case each output is equal to the
00096      * DC coefficient (with scale factor as needed).
00097      * With typical images and quantization tables, half or more of the
00098      * column DCT calculations can be simplified this way.
00099      */
00100 
00101     if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
00102     inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
00103     inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
00104     inptr[DCTSIZE*7] == 0) {
00105       /* AC terms all zero */
00106       FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
00107 
00108       wsptr[DCTSIZE*0] = dcval;
00109       wsptr[DCTSIZE*1] = dcval;
00110       wsptr[DCTSIZE*2] = dcval;
00111       wsptr[DCTSIZE*3] = dcval;
00112       wsptr[DCTSIZE*4] = dcval;
00113       wsptr[DCTSIZE*5] = dcval;
00114       wsptr[DCTSIZE*6] = dcval;
00115       wsptr[DCTSIZE*7] = dcval;
00116 
00117       inptr++;          /* advance pointers to next column */
00118       quantptr++;
00119       wsptr++;
00120       continue;
00121     }
00122 
00123     /* Even part */
00124 
00125     tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
00126     tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
00127     tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
00128     tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
00129 
00130     tmp10 = tmp0 + tmp2;    /* phase 3 */
00131     tmp11 = tmp0 - tmp2;
00132 
00133     tmp13 = tmp1 + tmp3;    /* phases 5-3 */
00134     tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
00135 
00136     tmp0 = tmp10 + tmp13;   /* phase 2 */
00137     tmp3 = tmp10 - tmp13;
00138     tmp1 = tmp11 + tmp12;
00139     tmp2 = tmp11 - tmp12;
00140 
00141     /* Odd part */
00142 
00143     tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
00144     tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
00145     tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
00146     tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
00147 
00148     z13 = tmp6 + tmp5;      /* phase 6 */
00149     z10 = tmp6 - tmp5;
00150     z11 = tmp4 + tmp7;
00151     z12 = tmp4 - tmp7;
00152 
00153     tmp7 = z11 + z13;       /* phase 5 */
00154     tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
00155 
00156     z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
00157     tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
00158     tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
00159 
00160     tmp6 = tmp12 - tmp7;    /* phase 2 */
00161     tmp5 = tmp11 - tmp6;
00162     tmp4 = tmp10 - tmp5;
00163 
00164     wsptr[DCTSIZE*0] = tmp0 + tmp7;
00165     wsptr[DCTSIZE*7] = tmp0 - tmp7;
00166     wsptr[DCTSIZE*1] = tmp1 + tmp6;
00167     wsptr[DCTSIZE*6] = tmp1 - tmp6;
00168     wsptr[DCTSIZE*2] = tmp2 + tmp5;
00169     wsptr[DCTSIZE*5] = tmp2 - tmp5;
00170     wsptr[DCTSIZE*3] = tmp3 + tmp4;
00171     wsptr[DCTSIZE*4] = tmp3 - tmp4;
00172 
00173     inptr++;            /* advance pointers to next column */
00174     quantptr++;
00175     wsptr++;
00176   }
00177 
00178   /* Pass 2: process rows from work array, store into output array. */
00179 
00180   wsptr = workspace;
00181   for (ctr = 0; ctr < DCTSIZE; ctr++) {
00182     outptr = output_buf[ctr] + output_col;
00183     /* Rows of zeroes can be exploited in the same way as we did with columns.
00184      * However, the column calculation has created many nonzero AC terms, so
00185      * the simplification applies less often (typically 5% to 10% of the time).
00186      * And testing floats for zero is relatively expensive, so we don't bother.
00187      */
00188 
00189     /* Even part */
00190 
00191     /* Prepare range-limit and float->int conversion */
00192     z5 = wsptr[0] + (((FAST_FLOAT) RANGE_CENTER) + ((FAST_FLOAT) 0.5));
00193     tmp10 = z5 + wsptr[4];
00194     tmp11 = z5 - wsptr[4];
00195 
00196     tmp13 = wsptr[2] + wsptr[6];
00197     tmp12 = (wsptr[2] - wsptr[6]) *
00198           ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
00199 
00200     tmp0 = tmp10 + tmp13;
00201     tmp3 = tmp10 - tmp13;
00202     tmp1 = tmp11 + tmp12;
00203     tmp2 = tmp11 - tmp12;
00204 
00205     /* Odd part */
00206 
00207     z13 = wsptr[5] + wsptr[3];
00208     z10 = wsptr[5] - wsptr[3];
00209     z11 = wsptr[1] + wsptr[7];
00210     z12 = wsptr[1] - wsptr[7];
00211 
00212     tmp7 = z11 + z13;       /* phase 5 */
00213     tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
00214 
00215     z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
00216     tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
00217     tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
00218 
00219     tmp6 = tmp12 - tmp7;    /* phase 2 */
00220     tmp5 = tmp11 - tmp6;
00221     tmp4 = tmp10 - tmp5;
00222 
00223     /* Final output stage: float->int conversion and range-limit */
00224 
00225     outptr[0] = range_limit[(int) (tmp0 + tmp7) & RANGE_MASK];
00226     outptr[7] = range_limit[(int) (tmp0 - tmp7) & RANGE_MASK];
00227     outptr[1] = range_limit[(int) (tmp1 + tmp6) & RANGE_MASK];
00228     outptr[6] = range_limit[(int) (tmp1 - tmp6) & RANGE_MASK];
00229     outptr[2] = range_limit[(int) (tmp2 + tmp5) & RANGE_MASK];
00230     outptr[5] = range_limit[(int) (tmp2 - tmp5) & RANGE_MASK];
00231     outptr[3] = range_limit[(int) (tmp3 + tmp4) & RANGE_MASK];
00232     outptr[4] = range_limit[(int) (tmp3 - tmp4) & RANGE_MASK];
00233 
00234     wsptr += DCTSIZE;       /* advance pointer to next row */
00235   }
00236 }
00237 
00238 #endif /* DCT_FLOAT_SUPPORTED */