graf2d » asimage » src » libAfterImage » libjpeg

jfdctint.c

Go to the documentation of this file.
00001 /*
00002  * jfdctint.c
00003  *
00004  * Copyright (C) 1991-1996, Thomas G. Lane.
00005  * Modification developed 2003-2009 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 slow-but-accurate integer implementation of the
00010  * forward DCT (Discrete Cosine Transform).
00011  *
00012  * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
00013  * on each column.  Direct algorithms are also available, but they are
00014  * much more complex and seem not to be any faster when reduced to code.
00015  *
00016  * This implementation is based on an algorithm described in
00017  *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
00018  *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
00019  *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
00020  * The primary algorithm described there uses 11 multiplies and 29 adds.
00021  * We use their alternate method with 12 multiplies and 32 adds.
00022  * The advantage of this method is that no data path contains more than one
00023  * multiplication; this allows a very simple and accurate implementation in
00024  * scaled fixed-point arithmetic, with a minimal number of shifts.
00025  *
00026  * We also provide FDCT routines with various input sample block sizes for
00027  * direct resolution reduction or enlargement and for direct resolving the
00028  * common 2x1 and 1x2 subsampling cases without additional resampling: NxN
00029  * (N=1...16), 2NxN, and Nx2N (N=1...8) pixels for one 8x8 output DCT block.
00030  *
00031  * For N<8 we fill the remaining block coefficients with zero.
00032  * For N>8 we apply a partial N-point FDCT on the input samples, computing
00033  * just the lower 8 frequency coefficients and discarding the rest.
00034  *
00035  * We must scale the output coefficients of the N-point FDCT appropriately
00036  * to the standard 8-point FDCT level by 8/N per 1-D pass.  This scaling
00037  * is folded into the constant multipliers (pass 2) and/or final/initial
00038  * shifting.
00039  *
00040  * CAUTION: We rely on the FIX() macro except for the N=1,2,4,8 cases
00041  * since there would be too many additional constants to pre-calculate.
00042  */
00043 
00044 #define JPEG_INTERNALS
00045 #include "jinclude.h"
00046 #include "jpeglib.h"
00047 #include "jdct.h"               /* Private declarations for DCT subsystem */
00048 
00049 #ifdef DCT_ISLOW_SUPPORTED
00050 
00051 
00052 /*
00053  * This module is specialized to the case DCTSIZE = 8.
00054  */
00055 
00056 #if DCTSIZE != 8
00057   Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */
00058 #endif
00059 
00060 
00061 /*
00062  * The poop on this scaling stuff is as follows:
00063  *
00064  * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
00065  * larger than the true DCT outputs.  The final outputs are therefore
00066  * a factor of N larger than desired; since N=8 this can be cured by
00067  * a simple right shift at the end of the algorithm.  The advantage of
00068  * this arrangement is that we save two multiplications per 1-D DCT,
00069  * because the y0 and y4 outputs need not be divided by sqrt(N).
00070  * In the IJG code, this factor of 8 is removed by the quantization step
00071  * (in jcdctmgr.c), NOT in this module.
00072  *
00073  * We have to do addition and subtraction of the integer inputs, which
00074  * is no problem, and multiplication by fractional constants, which is
00075  * a problem to do in integer arithmetic.  We multiply all the constants
00076  * by CONST_SCALE and convert them to integer constants (thus retaining
00077  * CONST_BITS bits of precision in the constants).  After doing a
00078  * multiplication we have to divide the product by CONST_SCALE, with proper
00079  * rounding, to produce the correct output.  This division can be done
00080  * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
00081  * as long as possible so that partial sums can be added together with
00082  * full fractional precision.
00083  *
00084  * The outputs of the first pass are scaled up by PASS1_BITS bits so that
00085  * they are represented to better-than-integral precision.  These outputs
00086  * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
00087  * with the recommended scaling.  (For 12-bit sample data, the intermediate
00088  * array is INT32 anyway.)
00089  *
00090  * To avoid overflow of the 32-bit intermediate results in pass 2, we must
00091  * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
00092  * shows that the values given below are the most effective.
00093  */
00094 
00095 #if BITS_IN_JSAMPLE == 8
00096 #define CONST_BITS  13
00097 #define PASS1_BITS  2
00098 #else
00099 #define CONST_BITS  13
00100 #define PASS1_BITS  1           /* lose a little precision to avoid overflow */
00101 #endif
00102 
00103 /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
00104  * causing a lot of useless floating-point operations at run time.
00105  * To get around this we use the following pre-calculated constants.
00106  * If you change CONST_BITS you may want to add appropriate values.
00107  * (With a reasonable C compiler, you can just rely on the FIX() macro...)
00108  */
00109 
00110 #if CONST_BITS == 13
00111 #define FIX_0_298631336  ((INT32)  2446)        /* FIX(0.298631336) */
00112 #define FIX_0_390180644  ((INT32)  3196)        /* FIX(0.390180644) */
00113 #define FIX_0_541196100  ((INT32)  4433)        /* FIX(0.541196100) */
00114 #define FIX_0_765366865  ((INT32)  6270)        /* FIX(0.765366865) */
00115 #define FIX_0_899976223  ((INT32)  7373)        /* FIX(0.899976223) */
00116 #define FIX_1_175875602  ((INT32)  9633)        /* FIX(1.175875602) */
00117 #define FIX_1_501321110  ((INT32)  12299)       /* FIX(1.501321110) */
00118 #define FIX_1_847759065  ((INT32)  15137)       /* FIX(1.847759065) */
00119 #define FIX_1_961570560  ((INT32)  16069)       /* FIX(1.961570560) */
00120 #define FIX_2_053119869  ((INT32)  16819)       /* FIX(2.053119869) */
00121 #define FIX_2_562915447  ((INT32)  20995)       /* FIX(2.562915447) */
00122 #define FIX_3_072711026  ((INT32)  25172)       /* FIX(3.072711026) */
00123 #else
00124 #define FIX_0_298631336  FIX(0.298631336)
00125 #define FIX_0_390180644  FIX(0.390180644)
00126 #define FIX_0_541196100  FIX(0.541196100)
00127 #define FIX_0_765366865  FIX(0.765366865)
00128 #define FIX_0_899976223  FIX(0.899976223)
00129 #define FIX_1_175875602  FIX(1.175875602)
00130 #define FIX_1_501321110  FIX(1.501321110)
00131 #define FIX_1_847759065  FIX(1.847759065)
00132 #define FIX_1_961570560  FIX(1.961570560)
00133 #define FIX_2_053119869  FIX(2.053119869)
00134 #define FIX_2_562915447  FIX(2.562915447)
00135 #define FIX_3_072711026  FIX(3.072711026)
00136 #endif
00137 
00138 
00139 /* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
00140  * For 8-bit samples with the recommended scaling, all the variable
00141  * and constant values involved are no more than 16 bits wide, so a
00142  * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
00143  * For 12-bit samples, a full 32-bit multiplication will be needed.
00144  */
00145 
00146 #if BITS_IN_JSAMPLE == 8
00147 #define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
00148 #else
00149 #define MULTIPLY(var,const)  ((var) * (const))
00150 #endif
00151 
00152 
00153 /*
00154  * Perform the forward DCT on one block of samples.
00155  */
00156 
00157 GLOBAL(void)
00158 jpeg_fdct_islow (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00159 {
00160   INT32 tmp0, tmp1, tmp2, tmp3;
00161   INT32 tmp10, tmp11, tmp12, tmp13;
00162   INT32 z1;
00163   DCTELEM *dataptr;
00164   JSAMPROW elemptr;
00165   int ctr;
00166   SHIFT_TEMPS
00167 
00168   /* Pass 1: process rows. */
00169   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00170   /* furthermore, we scale the results by 2**PASS1_BITS. */
00171 
00172   dataptr = data;
00173   for (ctr = 0; ctr < DCTSIZE; ctr++) {
00174     elemptr = sample_data[ctr] + start_col;
00175 
00176     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00177      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00178      */
00179 
00180     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
00181     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
00182     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
00183     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
00184 
00185     tmp10 = tmp0 + tmp3;
00186     tmp12 = tmp0 - tmp3;
00187     tmp11 = tmp1 + tmp2;
00188     tmp13 = tmp1 - tmp2;
00189 
00190     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
00191     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
00192     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
00193     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
00194 
00195     /* Apply unsigned->signed conversion */
00196     dataptr[0] = (DCTELEM) ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << PASS1_BITS);
00197     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
00198 
00199     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00200     /* Add fudge factor here for final descale. */
00201     z1 += ONE << (CONST_BITS-PASS1_BITS-1);
00202     dataptr[2] = (DCTELEM) RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865),
00203                                        CONST_BITS-PASS1_BITS);
00204     dataptr[6] = (DCTELEM) RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065),
00205                                        CONST_BITS-PASS1_BITS);
00206 
00207     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00208      * cK represents sqrt(2) * cos(K*pi/16).
00209      * i0..i3 in the paper are tmp0..tmp3 here.
00210      */
00211 
00212     tmp10 = tmp0 + tmp3;
00213     tmp11 = tmp1 + tmp2;
00214     tmp12 = tmp0 + tmp2;
00215     tmp13 = tmp1 + tmp3;
00216     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
00217     /* Add fudge factor here for final descale. */
00218     z1 += ONE << (CONST_BITS-PASS1_BITS-1);
00219 
00220     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
00221     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
00222     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
00223     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
00224     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
00225     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
00226     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
00227     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
00228 
00229     tmp12 += z1;
00230     tmp13 += z1;
00231 
00232     dataptr[1] = (DCTELEM)
00233       RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS);
00234     dataptr[3] = (DCTELEM)
00235       RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS);
00236     dataptr[5] = (DCTELEM)
00237       RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS);
00238     dataptr[7] = (DCTELEM)
00239       RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS);
00240 
00241     dataptr += DCTSIZE;         /* advance pointer to next row */
00242   }
00243 
00244   /* Pass 2: process columns.
00245    * We remove the PASS1_BITS scaling, but leave the results scaled up
00246    * by an overall factor of 8.
00247    */
00248 
00249   dataptr = data;
00250   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
00251     /* Even part per LL&M figure 1 --- note that published figure is faulty;
00252      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
00253      */
00254 
00255     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
00256     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
00257     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
00258     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
00259 
00260     /* Add fudge factor here for final descale. */
00261     tmp10 = tmp0 + tmp3 + (ONE << (PASS1_BITS-1));
00262     tmp12 = tmp0 - tmp3;
00263     tmp11 = tmp1 + tmp2;
00264     tmp13 = tmp1 - tmp2;
00265 
00266     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
00267     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
00268     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
00269     tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
00270 
00271     dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp10 + tmp11, PASS1_BITS);
00272     dataptr[DCTSIZE*4] = (DCTELEM) RIGHT_SHIFT(tmp10 - tmp11, PASS1_BITS);
00273 
00274     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
00275     /* Add fudge factor here for final descale. */
00276     z1 += ONE << (CONST_BITS+PASS1_BITS-1);
00277     dataptr[DCTSIZE*2] = (DCTELEM)
00278       RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865), CONST_BITS+PASS1_BITS);
00279     dataptr[DCTSIZE*6] = (DCTELEM)
00280       RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065), CONST_BITS+PASS1_BITS);
00281 
00282     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
00283      * cK represents sqrt(2) * cos(K*pi/16).
00284      * i0..i3 in the paper are tmp0..tmp3 here.
00285      */
00286 
00287     tmp10 = tmp0 + tmp3;
00288     tmp11 = tmp1 + tmp2;
00289     tmp12 = tmp0 + tmp2;
00290     tmp13 = tmp1 + tmp3;
00291     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
00292     /* Add fudge factor here for final descale. */
00293     z1 += ONE << (CONST_BITS+PASS1_BITS-1);
00294 
00295     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
00296     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
00297     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
00298     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
00299     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
00300     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
00301     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
00302     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
00303 
00304     tmp12 += z1;
00305     tmp13 += z1;
00306 
00307     dataptr[DCTSIZE*1] = (DCTELEM)
00308       RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS+PASS1_BITS);
00309     dataptr[DCTSIZE*3] = (DCTELEM)
00310       RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS+PASS1_BITS);
00311     dataptr[DCTSIZE*5] = (DCTELEM)
00312       RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS+PASS1_BITS);
00313     dataptr[DCTSIZE*7] = (DCTELEM)
00314       RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS+PASS1_BITS);
00315 
00316     dataptr++;                  /* advance pointer to next column */
00317   }
00318 }
00319 
00320 #ifdef DCT_SCALING_SUPPORTED
00321 
00322 
00323 /*
00324  * Perform the forward DCT on a 7x7 sample block.
00325  */
00326 
00327 GLOBAL(void)
00328 jpeg_fdct_7x7 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00329 {
00330   INT32 tmp0, tmp1, tmp2, tmp3;
00331   INT32 tmp10, tmp11, tmp12;
00332   INT32 z1, z2, z3;
00333   DCTELEM *dataptr;
00334   JSAMPROW elemptr;
00335   int ctr;
00336   SHIFT_TEMPS
00337 
00338   /* Pre-zero output coefficient block. */
00339   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00340 
00341   /* Pass 1: process rows. */
00342   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00343   /* furthermore, we scale the results by 2**PASS1_BITS. */
00344   /* cK represents sqrt(2) * cos(K*pi/14). */
00345 
00346   dataptr = data;
00347   for (ctr = 0; ctr < 7; ctr++) {
00348     elemptr = sample_data[ctr] + start_col;
00349 
00350     /* Even part */
00351 
00352     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[6]);
00353     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[5]);
00354     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[4]);
00355     tmp3 = GETJSAMPLE(elemptr[3]);
00356 
00357     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[6]);
00358     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[5]);
00359     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[4]);
00360 
00361     z1 = tmp0 + tmp2;
00362     /* Apply unsigned->signed conversion */
00363     dataptr[0] = (DCTELEM)
00364       ((z1 + tmp1 + tmp3 - 7 * CENTERJSAMPLE) << PASS1_BITS);
00365     tmp3 += tmp3;
00366     z1 -= tmp3;
00367     z1 -= tmp3;
00368     z1 = MULTIPLY(z1, FIX(0.353553391));                /* (c2+c6-c4)/2 */
00369     z2 = MULTIPLY(tmp0 - tmp2, FIX(0.920609002));       /* (c2+c4-c6)/2 */
00370     z3 = MULTIPLY(tmp1 - tmp2, FIX(0.314692123));       /* c6 */
00371     dataptr[2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS-PASS1_BITS);
00372     z1 -= z2;
00373     z2 = MULTIPLY(tmp0 - tmp1, FIX(0.881747734));       /* c4 */
00374     dataptr[4] = (DCTELEM)
00375       DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.707106781)), /* c2+c6-c4 */
00376               CONST_BITS-PASS1_BITS);
00377     dataptr[6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS-PASS1_BITS);
00378 
00379     /* Odd part */
00380 
00381     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(0.935414347));   /* (c3+c1-c5)/2 */
00382     tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.170262339));   /* (c3+c5-c1)/2 */
00383     tmp0 = tmp1 - tmp2;
00384     tmp1 += tmp2;
00385     tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.378756276)); /* -c1 */
00386     tmp1 += tmp2;
00387     tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.613604268));   /* c5 */
00388     tmp0 += tmp3;
00389     tmp2 += tmp3 + MULTIPLY(tmp12, FIX(1.870828693));   /* c3+c1-c5 */
00390 
00391     dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-PASS1_BITS);
00392     dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-PASS1_BITS);
00393     dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-PASS1_BITS);
00394 
00395     dataptr += DCTSIZE;         /* advance pointer to next row */
00396   }
00397 
00398   /* Pass 2: process columns.
00399    * We remove the PASS1_BITS scaling, but leave the results scaled up
00400    * by an overall factor of 8.
00401    * We must also scale the output by (8/7)**2 = 64/49, which we fold
00402    * into the constant multipliers:
00403    * cK now represents sqrt(2) * cos(K*pi/14) * 64/49.
00404    */
00405 
00406   dataptr = data;
00407   for (ctr = 0; ctr < 7; ctr++) {
00408     /* Even part */
00409 
00410     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*6];
00411     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*5];
00412     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*4];
00413     tmp3 = dataptr[DCTSIZE*3];
00414 
00415     tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*6];
00416     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*5];
00417     tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*4];
00418 
00419     z1 = tmp0 + tmp2;
00420     dataptr[DCTSIZE*0] = (DCTELEM)
00421       DESCALE(MULTIPLY(z1 + tmp1 + tmp3, FIX(1.306122449)), /* 64/49 */
00422               CONST_BITS+PASS1_BITS);
00423     tmp3 += tmp3;
00424     z1 -= tmp3;
00425     z1 -= tmp3;
00426     z1 = MULTIPLY(z1, FIX(0.461784020));                /* (c2+c6-c4)/2 */
00427     z2 = MULTIPLY(tmp0 - tmp2, FIX(1.202428084));       /* (c2+c4-c6)/2 */
00428     z3 = MULTIPLY(tmp1 - tmp2, FIX(0.411026446));       /* c6 */
00429     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS+PASS1_BITS);
00430     z1 -= z2;
00431     z2 = MULTIPLY(tmp0 - tmp1, FIX(1.151670509));       /* c4 */
00432     dataptr[DCTSIZE*4] = (DCTELEM)
00433       DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.923568041)), /* c2+c6-c4 */
00434               CONST_BITS+PASS1_BITS);
00435     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+PASS1_BITS);
00436 
00437     /* Odd part */
00438 
00439     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.221765677));   /* (c3+c1-c5)/2 */
00440     tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.222383464));   /* (c3+c5-c1)/2 */
00441     tmp0 = tmp1 - tmp2;
00442     tmp1 += tmp2;
00443     tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.800824523)); /* -c1 */
00444     tmp1 += tmp2;
00445     tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.801442310));   /* c5 */
00446     tmp0 += tmp3;
00447     tmp2 += tmp3 + MULTIPLY(tmp12, FIX(2.443531355));   /* c3+c1-c5 */
00448 
00449     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+PASS1_BITS);
00450     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+PASS1_BITS);
00451     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+PASS1_BITS);
00452 
00453     dataptr++;                  /* advance pointer to next column */
00454   }
00455 }
00456 
00457 
00458 /*
00459  * Perform the forward DCT on a 6x6 sample block.
00460  */
00461 
00462 GLOBAL(void)
00463 jpeg_fdct_6x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00464 {
00465   INT32 tmp0, tmp1, tmp2;
00466   INT32 tmp10, tmp11, tmp12;
00467   DCTELEM *dataptr;
00468   JSAMPROW elemptr;
00469   int ctr;
00470   SHIFT_TEMPS
00471 
00472   /* Pre-zero output coefficient block. */
00473   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00474 
00475   /* Pass 1: process rows. */
00476   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00477   /* furthermore, we scale the results by 2**PASS1_BITS. */
00478   /* cK represents sqrt(2) * cos(K*pi/12). */
00479 
00480   dataptr = data;
00481   for (ctr = 0; ctr < 6; ctr++) {
00482     elemptr = sample_data[ctr] + start_col;
00483 
00484     /* Even part */
00485 
00486     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
00487     tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
00488     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
00489 
00490     tmp10 = tmp0 + tmp2;
00491     tmp12 = tmp0 - tmp2;
00492 
00493     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
00494     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
00495     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
00496 
00497     /* Apply unsigned->signed conversion */
00498     dataptr[0] = (DCTELEM)
00499       ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << PASS1_BITS);
00500     dataptr[2] = (DCTELEM)
00501       DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
00502               CONST_BITS-PASS1_BITS);
00503     dataptr[4] = (DCTELEM)
00504       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
00505               CONST_BITS-PASS1_BITS);
00506 
00507     /* Odd part */
00508 
00509     tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
00510                     CONST_BITS-PASS1_BITS);
00511 
00512     dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << PASS1_BITS));
00513     dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << PASS1_BITS);
00514     dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << PASS1_BITS));
00515 
00516     dataptr += DCTSIZE;         /* advance pointer to next row */
00517   }
00518 
00519   /* Pass 2: process columns.
00520    * We remove the PASS1_BITS scaling, but leave the results scaled up
00521    * by an overall factor of 8.
00522    * We must also scale the output by (8/6)**2 = 16/9, which we fold
00523    * into the constant multipliers:
00524    * cK now represents sqrt(2) * cos(K*pi/12) * 16/9.
00525    */
00526 
00527   dataptr = data;
00528   for (ctr = 0; ctr < 6; ctr++) {
00529     /* Even part */
00530 
00531     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
00532     tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
00533     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
00534 
00535     tmp10 = tmp0 + tmp2;
00536     tmp12 = tmp0 - tmp2;
00537 
00538     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
00539     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
00540     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
00541 
00542     dataptr[DCTSIZE*0] = (DCTELEM)
00543       DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
00544               CONST_BITS+PASS1_BITS);
00545     dataptr[DCTSIZE*2] = (DCTELEM)
00546       DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
00547               CONST_BITS+PASS1_BITS);
00548     dataptr[DCTSIZE*4] = (DCTELEM)
00549       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
00550               CONST_BITS+PASS1_BITS);
00551 
00552     /* Odd part */
00553 
00554     tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
00555 
00556     dataptr[DCTSIZE*1] = (DCTELEM)
00557       DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
00558               CONST_BITS+PASS1_BITS);
00559     dataptr[DCTSIZE*3] = (DCTELEM)
00560       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
00561               CONST_BITS+PASS1_BITS);
00562     dataptr[DCTSIZE*5] = (DCTELEM)
00563       DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
00564               CONST_BITS+PASS1_BITS);
00565 
00566     dataptr++;                  /* advance pointer to next column */
00567   }
00568 }
00569 
00570 
00571 /*
00572  * Perform the forward DCT on a 5x5 sample block.
00573  */
00574 
00575 GLOBAL(void)
00576 jpeg_fdct_5x5 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00577 {
00578   INT32 tmp0, tmp1, tmp2;
00579   INT32 tmp10, tmp11;
00580   DCTELEM *dataptr;
00581   JSAMPROW elemptr;
00582   int ctr;
00583   SHIFT_TEMPS
00584 
00585   /* Pre-zero output coefficient block. */
00586   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00587 
00588   /* Pass 1: process rows. */
00589   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00590   /* furthermore, we scale the results by 2**PASS1_BITS. */
00591   /* We scale the results further by 2 as part of output adaption */
00592   /* scaling for different DCT size. */
00593   /* cK represents sqrt(2) * cos(K*pi/10). */
00594 
00595   dataptr = data;
00596   for (ctr = 0; ctr < 5; ctr++) {
00597     elemptr = sample_data[ctr] + start_col;
00598 
00599     /* Even part */
00600 
00601     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[4]);
00602     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[3]);
00603     tmp2 = GETJSAMPLE(elemptr[2]);
00604 
00605     tmp10 = tmp0 + tmp1;
00606     tmp11 = tmp0 - tmp1;
00607 
00608     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[4]);
00609     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[3]);
00610 
00611     /* Apply unsigned->signed conversion */
00612     dataptr[0] = (DCTELEM)
00613       ((tmp10 + tmp2 - 5 * CENTERJSAMPLE) << (PASS1_BITS+1));
00614     tmp11 = MULTIPLY(tmp11, FIX(0.790569415));          /* (c2+c4)/2 */
00615     tmp10 -= tmp2 << 2;
00616     tmp10 = MULTIPLY(tmp10, FIX(0.353553391));          /* (c2-c4)/2 */
00617     dataptr[2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS-PASS1_BITS-1);
00618     dataptr[4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS-PASS1_BITS-1);
00619 
00620     /* Odd part */
00621 
00622     tmp10 = MULTIPLY(tmp0 + tmp1, FIX(0.831253876));    /* c3 */
00623 
00624     dataptr[1] = (DCTELEM)
00625       DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.513743148)), /* c1-c3 */
00626               CONST_BITS-PASS1_BITS-1);
00627     dataptr[3] = (DCTELEM)
00628       DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.176250899)), /* c1+c3 */
00629               CONST_BITS-PASS1_BITS-1);
00630 
00631     dataptr += DCTSIZE;         /* advance pointer to next row */
00632   }
00633 
00634   /* Pass 2: process columns.
00635    * We remove the PASS1_BITS scaling, but leave the results scaled up
00636    * by an overall factor of 8.
00637    * We must also scale the output by (8/5)**2 = 64/25, which we partially
00638    * fold into the constant multipliers (other part was done in pass 1):
00639    * cK now represents sqrt(2) * cos(K*pi/10) * 32/25.
00640    */
00641 
00642   dataptr = data;
00643   for (ctr = 0; ctr < 5; ctr++) {
00644     /* Even part */
00645 
00646     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*4];
00647     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*3];
00648     tmp2 = dataptr[DCTSIZE*2];
00649 
00650     tmp10 = tmp0 + tmp1;
00651     tmp11 = tmp0 - tmp1;
00652 
00653     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*4];
00654     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*3];
00655 
00656     dataptr[DCTSIZE*0] = (DCTELEM)
00657       DESCALE(MULTIPLY(tmp10 + tmp2, FIX(1.28)),        /* 32/25 */
00658               CONST_BITS+PASS1_BITS);
00659     tmp11 = MULTIPLY(tmp11, FIX(1.011928851));          /* (c2+c4)/2 */
00660     tmp10 -= tmp2 << 2;
00661     tmp10 = MULTIPLY(tmp10, FIX(0.452548340));          /* (c2-c4)/2 */
00662     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS+PASS1_BITS);
00663     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS+PASS1_BITS);
00664 
00665     /* Odd part */
00666 
00667     tmp10 = MULTIPLY(tmp0 + tmp1, FIX(1.064004961));    /* c3 */
00668 
00669     dataptr[DCTSIZE*1] = (DCTELEM)
00670       DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.657591230)), /* c1-c3 */
00671               CONST_BITS+PASS1_BITS);
00672     dataptr[DCTSIZE*3] = (DCTELEM)
00673       DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.785601151)), /* c1+c3 */
00674               CONST_BITS+PASS1_BITS);
00675 
00676     dataptr++;                  /* advance pointer to next column */
00677   }
00678 }
00679 
00680 
00681 /*
00682  * Perform the forward DCT on a 4x4 sample block.
00683  */
00684 
00685 GLOBAL(void)
00686 jpeg_fdct_4x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00687 {
00688   INT32 tmp0, tmp1;
00689   INT32 tmp10, tmp11;
00690   DCTELEM *dataptr;
00691   JSAMPROW elemptr;
00692   int ctr;
00693   SHIFT_TEMPS
00694 
00695   /* Pre-zero output coefficient block. */
00696   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00697 
00698   /* Pass 1: process rows. */
00699   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00700   /* furthermore, we scale the results by 2**PASS1_BITS. */
00701   /* We must also scale the output by (8/4)**2 = 2**2, which we add here. */
00702   /* cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT]. */
00703 
00704   dataptr = data;
00705   for (ctr = 0; ctr < 4; ctr++) {
00706     elemptr = sample_data[ctr] + start_col;
00707 
00708     /* Even part */
00709 
00710     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
00711     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
00712 
00713     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
00714     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
00715 
00716     /* Apply unsigned->signed conversion */
00717     dataptr[0] = (DCTELEM)
00718       ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+2));
00719     dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+2));
00720 
00721     /* Odd part */
00722 
00723     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
00724     /* Add fudge factor here for final descale. */
00725     tmp0 += ONE << (CONST_BITS-PASS1_BITS-3);
00726 
00727     dataptr[1] = (DCTELEM)
00728       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
00729                   CONST_BITS-PASS1_BITS-2);
00730     dataptr[3] = (DCTELEM)
00731       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
00732                   CONST_BITS-PASS1_BITS-2);
00733 
00734     dataptr += DCTSIZE;         /* advance pointer to next row */
00735   }
00736 
00737   /* Pass 2: process columns.
00738    * We remove the PASS1_BITS scaling, but leave the results scaled up
00739    * by an overall factor of 8.
00740    */
00741 
00742   dataptr = data;
00743   for (ctr = 0; ctr < 4; ctr++) {
00744     /* Even part */
00745 
00746     /* Add fudge factor here for final descale. */
00747     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3] + (ONE << (PASS1_BITS-1));
00748     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
00749 
00750     tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
00751     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
00752 
00753     dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
00754     dataptr[DCTSIZE*2] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
00755 
00756     /* Odd part */
00757 
00758     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
00759     /* Add fudge factor here for final descale. */
00760     tmp0 += ONE << (CONST_BITS+PASS1_BITS-1);
00761 
00762     dataptr[DCTSIZE*1] = (DCTELEM)
00763       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
00764                   CONST_BITS+PASS1_BITS);
00765     dataptr[DCTSIZE*3] = (DCTELEM)
00766       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
00767                   CONST_BITS+PASS1_BITS);
00768 
00769     dataptr++;                  /* advance pointer to next column */
00770   }
00771 }
00772 
00773 
00774 /*
00775  * Perform the forward DCT on a 3x3 sample block.
00776  */
00777 
00778 GLOBAL(void)
00779 jpeg_fdct_3x3 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00780 {
00781   INT32 tmp0, tmp1, tmp2;
00782   DCTELEM *dataptr;
00783   JSAMPROW elemptr;
00784   int ctr;
00785   SHIFT_TEMPS
00786 
00787   /* Pre-zero output coefficient block. */
00788   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00789 
00790   /* Pass 1: process rows. */
00791   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00792   /* furthermore, we scale the results by 2**PASS1_BITS. */
00793   /* We scale the results further by 2**2 as part of output adaption */
00794   /* scaling for different DCT size. */
00795   /* cK represents sqrt(2) * cos(K*pi/6). */
00796 
00797   dataptr = data;
00798   for (ctr = 0; ctr < 3; ctr++) {
00799     elemptr = sample_data[ctr] + start_col;
00800 
00801     /* Even part */
00802 
00803     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[2]);
00804     tmp1 = GETJSAMPLE(elemptr[1]);
00805 
00806     tmp2 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[2]);
00807 
00808     /* Apply unsigned->signed conversion */
00809     dataptr[0] = (DCTELEM)
00810       ((tmp0 + tmp1 - 3 * CENTERJSAMPLE) << (PASS1_BITS+2));
00811     dataptr[2] = (DCTELEM)
00812       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(0.707106781)), /* c2 */
00813               CONST_BITS-PASS1_BITS-2);
00814 
00815     /* Odd part */
00816 
00817     dataptr[1] = (DCTELEM)
00818       DESCALE(MULTIPLY(tmp2, FIX(1.224744871)),               /* c1 */
00819               CONST_BITS-PASS1_BITS-2);
00820 
00821     dataptr += DCTSIZE;         /* advance pointer to next row */
00822   }
00823 
00824   /* Pass 2: process columns.
00825    * We remove the PASS1_BITS scaling, but leave the results scaled up
00826    * by an overall factor of 8.
00827    * We must also scale the output by (8/3)**2 = 64/9, which we partially
00828    * fold into the constant multipliers (other part was done in pass 1):
00829    * cK now represents sqrt(2) * cos(K*pi/6) * 16/9.
00830    */
00831 
00832   dataptr = data;
00833   for (ctr = 0; ctr < 3; ctr++) {
00834     /* Even part */
00835 
00836     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*2];
00837     tmp1 = dataptr[DCTSIZE*1];
00838 
00839     tmp2 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*2];
00840 
00841     dataptr[DCTSIZE*0] = (DCTELEM)
00842       DESCALE(MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),        /* 16/9 */
00843               CONST_BITS+PASS1_BITS);
00844     dataptr[DCTSIZE*2] = (DCTELEM)
00845       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(1.257078722)), /* c2 */
00846               CONST_BITS+PASS1_BITS);
00847 
00848     /* Odd part */
00849 
00850     dataptr[DCTSIZE*1] = (DCTELEM)
00851       DESCALE(MULTIPLY(tmp2, FIX(2.177324216)),               /* c1 */
00852               CONST_BITS+PASS1_BITS);
00853 
00854     dataptr++;                  /* advance pointer to next column */
00855   }
00856 }
00857 
00858 
00859 /*
00860  * Perform the forward DCT on a 2x2 sample block.
00861  */
00862 
00863 GLOBAL(void)
00864 jpeg_fdct_2x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00865 {
00866   INT32 tmp0, tmp1, tmp2, tmp3;
00867   JSAMPROW elemptr;
00868 
00869   /* Pre-zero output coefficient block. */
00870   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00871 
00872   /* Pass 1: process rows. */
00873   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
00874 
00875   /* Row 0 */
00876   elemptr = sample_data[0] + start_col;
00877 
00878   tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[1]);
00879   tmp1 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[1]);
00880 
00881   /* Row 1 */
00882   elemptr = sample_data[1] + start_col;
00883 
00884   tmp2 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[1]);
00885   tmp3 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[1]);
00886 
00887   /* Pass 2: process columns.
00888    * We leave the results scaled up by an overall factor of 8.
00889    * We must also scale the output by (8/2)**2 = 2**4.
00890    */
00891 
00892   /* Column 0 */
00893   /* Apply unsigned->signed conversion */
00894   data[DCTSIZE*0] = (DCTELEM) ((tmp0 + tmp2 - 4 * CENTERJSAMPLE) << 4);
00895   data[DCTSIZE*1] = (DCTELEM) ((tmp0 - tmp2) << 4);
00896 
00897   /* Column 1 */
00898   data[DCTSIZE*0+1] = (DCTELEM) ((tmp1 + tmp3) << 4);
00899   data[DCTSIZE*1+1] = (DCTELEM) ((tmp1 - tmp3) << 4);
00900 }
00901 
00902 
00903 /*
00904  * Perform the forward DCT on a 1x1 sample block.
00905  */
00906 
00907 GLOBAL(void)
00908 jpeg_fdct_1x1 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00909 {
00910   /* Pre-zero output coefficient block. */
00911   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
00912 
00913   /* We leave the result scaled up by an overall factor of 8. */
00914   /* We must also scale the output by (8/1)**2 = 2**6. */
00915   /* Apply unsigned->signed conversion */
00916   data[0] = (DCTELEM)
00917     ((GETJSAMPLE(sample_data[0][start_col]) - CENTERJSAMPLE) << 6);
00918 }
00919 
00920 
00921 /*
00922  * Perform the forward DCT on a 9x9 sample block.
00923  */
00924 
00925 GLOBAL(void)
00926 jpeg_fdct_9x9 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
00927 {
00928   INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
00929   INT32 tmp10, tmp11, tmp12, tmp13;
00930   INT32 z1, z2;
00931   DCTELEM workspace[8];
00932   DCTELEM *dataptr;
00933   DCTELEM *wsptr;
00934   JSAMPROW elemptr;
00935   int ctr;
00936   SHIFT_TEMPS
00937 
00938   /* Pass 1: process rows. */
00939   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
00940   /* we scale the results further by 2 as part of output adaption */
00941   /* scaling for different DCT size. */
00942   /* cK represents sqrt(2) * cos(K*pi/18). */
00943 
00944   dataptr = data;
00945   ctr = 0;
00946   for (;;) {
00947     elemptr = sample_data[ctr] + start_col;
00948 
00949     /* Even part */
00950 
00951     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[8]);
00952     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[7]);
00953     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[6]);
00954     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[5]);
00955     tmp4 = GETJSAMPLE(elemptr[4]);
00956 
00957     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[8]);
00958     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[7]);
00959     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[6]);
00960     tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[5]);
00961 
00962     z1 = tmp0 + tmp2 + tmp3;
00963     z2 = tmp1 + tmp4;
00964     /* Apply unsigned->signed conversion */
00965     dataptr[0] = (DCTELEM) ((z1 + z2 - 9 * CENTERJSAMPLE) << 1);
00966     dataptr[6] = (DCTELEM)
00967       DESCALE(MULTIPLY(z1 - z2 - z2, FIX(0.707106781)),  /* c6 */
00968               CONST_BITS-1);
00969     z1 = MULTIPLY(tmp0 - tmp2, FIX(1.328926049));        /* c2 */
00970     z2 = MULTIPLY(tmp1 - tmp4 - tmp4, FIX(0.707106781)); /* c6 */
00971     dataptr[2] = (DCTELEM)
00972       DESCALE(MULTIPLY(tmp2 - tmp3, FIX(1.083350441))    /* c4 */
00973               + z1 + z2, CONST_BITS-1);
00974     dataptr[4] = (DCTELEM)
00975       DESCALE(MULTIPLY(tmp3 - tmp0, FIX(0.245575608))    /* c8 */
00976               + z1 - z2, CONST_BITS-1);
00977 
00978     /* Odd part */
00979 
00980     dataptr[3] = (DCTELEM)
00981       DESCALE(MULTIPLY(tmp10 - tmp12 - tmp13, FIX(1.224744871)), /* c3 */
00982               CONST_BITS-1);
00983 
00984     tmp11 = MULTIPLY(tmp11, FIX(1.224744871));        /* c3 */
00985     tmp0 = MULTIPLY(tmp10 + tmp12, FIX(0.909038955)); /* c5 */
00986     tmp1 = MULTIPLY(tmp10 + tmp13, FIX(0.483689525)); /* c7 */
00987 
00988     dataptr[1] = (DCTELEM) DESCALE(tmp11 + tmp0 + tmp1, CONST_BITS-1);
00989 
00990     tmp2 = MULTIPLY(tmp12 - tmp13, FIX(1.392728481)); /* c1 */
00991 
00992     dataptr[5] = (DCTELEM) DESCALE(tmp0 - tmp11 - tmp2, CONST_BITS-1);
00993     dataptr[7] = (DCTELEM) DESCALE(tmp1 - tmp11 + tmp2, CONST_BITS-1);
00994 
00995     ctr++;
00996 
00997     if (ctr != DCTSIZE) {
00998       if (ctr == 9)
00999         break;                  /* Done. */
01000       dataptr += DCTSIZE;       /* advance pointer to next row */
01001     } else
01002       dataptr = workspace;      /* switch pointer to extended workspace */
01003   }
01004 
01005   /* Pass 2: process columns.
01006    * We leave the results scaled up by an overall factor of 8.
01007    * We must also scale the output by (8/9)**2 = 64/81, which we partially
01008    * fold into the constant multipliers and final/initial shifting:
01009    * cK now represents sqrt(2) * cos(K*pi/18) * 128/81.
01010    */
01011 
01012   dataptr = data;
01013   wsptr = workspace;
01014   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01015     /* Even part */
01016 
01017     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*0];
01018     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*7];
01019     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*6];
01020     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*5];
01021     tmp4 = dataptr[DCTSIZE*4];
01022 
01023     tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*0];
01024     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*7];
01025     tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*6];
01026     tmp13 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*5];
01027 
01028     z1 = tmp0 + tmp2 + tmp3;
01029     z2 = tmp1 + tmp4;
01030     dataptr[DCTSIZE*0] = (DCTELEM)
01031       DESCALE(MULTIPLY(z1 + z2, FIX(1.580246914)),       /* 128/81 */
01032               CONST_BITS+2);
01033     dataptr[DCTSIZE*6] = (DCTELEM)
01034       DESCALE(MULTIPLY(z1 - z2 - z2, FIX(1.117403309)),  /* c6 */
01035               CONST_BITS+2);
01036     z1 = MULTIPLY(tmp0 - tmp2, FIX(2.100031287));        /* c2 */
01037     z2 = MULTIPLY(tmp1 - tmp4 - tmp4, FIX(1.117403309)); /* c6 */
01038     dataptr[DCTSIZE*2] = (DCTELEM)
01039       DESCALE(MULTIPLY(tmp2 - tmp3, FIX(1.711961190))    /* c4 */
01040               + z1 + z2, CONST_BITS+2);
01041     dataptr[DCTSIZE*4] = (DCTELEM)
01042       DESCALE(MULTIPLY(tmp3 - tmp0, FIX(0.388070096))    /* c8 */
01043               + z1 - z2, CONST_BITS+2);
01044 
01045     /* Odd part */
01046 
01047     dataptr[DCTSIZE*3] = (DCTELEM)
01048       DESCALE(MULTIPLY(tmp10 - tmp12 - tmp13, FIX(1.935399303)), /* c3 */
01049               CONST_BITS+2);
01050 
01051     tmp11 = MULTIPLY(tmp11, FIX(1.935399303));        /* c3 */
01052     tmp0 = MULTIPLY(tmp10 + tmp12, FIX(1.436506004)); /* c5 */
01053     tmp1 = MULTIPLY(tmp10 + tmp13, FIX(0.764348879)); /* c7 */
01054 
01055     dataptr[DCTSIZE*1] = (DCTELEM)
01056       DESCALE(tmp11 + tmp0 + tmp1, CONST_BITS+2);
01057 
01058     tmp2 = MULTIPLY(tmp12 - tmp13, FIX(2.200854883)); /* c1 */
01059 
01060     dataptr[DCTSIZE*5] = (DCTELEM)
01061       DESCALE(tmp0 - tmp11 - tmp2, CONST_BITS+2);
01062     dataptr[DCTSIZE*7] = (DCTELEM)
01063       DESCALE(tmp1 - tmp11 + tmp2, CONST_BITS+2);
01064 
01065     dataptr++;                  /* advance pointer to next column */
01066     wsptr++;                    /* advance pointer to next column */
01067   }
01068 }
01069 
01070 
01071 /*
01072  * Perform the forward DCT on a 10x10 sample block.
01073  */
01074 
01075 GLOBAL(void)
01076 jpeg_fdct_10x10 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01077 {
01078   INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
01079   INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
01080   DCTELEM workspace[8*2];
01081   DCTELEM *dataptr;
01082   DCTELEM *wsptr;
01083   JSAMPROW elemptr;
01084   int ctr;
01085   SHIFT_TEMPS
01086 
01087   /* Pass 1: process rows. */
01088   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
01089   /* we scale the results further by 2 as part of output adaption */
01090   /* scaling for different DCT size. */
01091   /* cK represents sqrt(2) * cos(K*pi/20). */
01092 
01093   dataptr = data;
01094   ctr = 0;
01095   for (;;) {
01096     elemptr = sample_data[ctr] + start_col;
01097 
01098     /* Even part */
01099 
01100     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[9]);
01101     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[8]);
01102     tmp12 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[7]);
01103     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[6]);
01104     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[5]);
01105 
01106     tmp10 = tmp0 + tmp4;
01107     tmp13 = tmp0 - tmp4;
01108     tmp11 = tmp1 + tmp3;
01109     tmp14 = tmp1 - tmp3;
01110 
01111     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[9]);
01112     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[8]);
01113     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[7]);
01114     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[6]);
01115     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[5]);
01116 
01117     /* Apply unsigned->signed conversion */
01118     dataptr[0] = (DCTELEM)
01119       ((tmp10 + tmp11 + tmp12 - 10 * CENTERJSAMPLE) << 1);
01120     tmp12 += tmp12;
01121     dataptr[4] = (DCTELEM)
01122       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.144122806)) - /* c4 */
01123               MULTIPLY(tmp11 - tmp12, FIX(0.437016024)),  /* c8 */
01124               CONST_BITS-1);
01125     tmp10 = MULTIPLY(tmp13 + tmp14, FIX(0.831253876));    /* c6 */
01126     dataptr[2] = (DCTELEM)
01127       DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.513743148)),  /* c2-c6 */
01128               CONST_BITS-1);
01129     dataptr[6] = (DCTELEM)
01130       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.176250899)),  /* c2+c6 */
01131               CONST_BITS-1);
01132 
01133     /* Odd part */
01134 
01135     tmp10 = tmp0 + tmp4;
01136     tmp11 = tmp1 - tmp3;
01137     dataptr[5] = (DCTELEM) ((tmp10 - tmp11 - tmp2) << 1);
01138     tmp2 <<= CONST_BITS;
01139     dataptr[1] = (DCTELEM)
01140       DESCALE(MULTIPLY(tmp0, FIX(1.396802247)) +          /* c1 */
01141               MULTIPLY(tmp1, FIX(1.260073511)) + tmp2 +   /* c3 */
01142               MULTIPLY(tmp3, FIX(0.642039522)) +          /* c7 */
01143               MULTIPLY(tmp4, FIX(0.221231742)),           /* c9 */
01144               CONST_BITS-1);
01145     tmp12 = MULTIPLY(tmp0 - tmp4, FIX(0.951056516)) -     /* (c3+c7)/2 */
01146             MULTIPLY(tmp1 + tmp3, FIX(0.587785252));      /* (c1-c9)/2 */
01147     tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.309016994)) +   /* (c3-c7)/2 */
01148             (tmp11 << (CONST_BITS - 1)) - tmp2;
01149     dataptr[3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS-1);
01150     dataptr[7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS-1);
01151 
01152     ctr++;
01153 
01154     if (ctr != DCTSIZE) {
01155       if (ctr == 10)
01156         break;                  /* Done. */
01157       dataptr += DCTSIZE;       /* advance pointer to next row */
01158     } else
01159       dataptr = workspace;      /* switch pointer to extended workspace */
01160   }
01161 
01162   /* Pass 2: process columns.
01163    * We leave the results scaled up by an overall factor of 8.
01164    * We must also scale the output by (8/10)**2 = 16/25, which we partially
01165    * fold into the constant multipliers and final/initial shifting:
01166    * cK now represents sqrt(2) * cos(K*pi/20) * 32/25.
01167    */
01168 
01169   dataptr = data;
01170   wsptr = workspace;
01171   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01172     /* Even part */
01173 
01174     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*1];
01175     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*0];
01176     tmp12 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*7];
01177     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*6];
01178     tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
01179 
01180     tmp10 = tmp0 + tmp4;
01181     tmp13 = tmp0 - tmp4;
01182     tmp11 = tmp1 + tmp3;
01183     tmp14 = tmp1 - tmp3;
01184 
01185     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*1];
01186     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*0];
01187     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*7];
01188     tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*6];
01189     tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
01190 
01191     dataptr[DCTSIZE*0] = (DCTELEM)
01192       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(1.28)), /* 32/25 */
01193               CONST_BITS+2);
01194     tmp12 += tmp12;
01195     dataptr[DCTSIZE*4] = (DCTELEM)
01196       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.464477191)) - /* c4 */
01197               MULTIPLY(tmp11 - tmp12, FIX(0.559380511)),  /* c8 */
01198               CONST_BITS+2);
01199     tmp10 = MULTIPLY(tmp13 + tmp14, FIX(1.064004961));    /* c6 */
01200     dataptr[DCTSIZE*2] = (DCTELEM)
01201       DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.657591230)),  /* c2-c6 */
01202               CONST_BITS+2);
01203     dataptr[DCTSIZE*6] = (DCTELEM)
01204       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.785601151)),  /* c2+c6 */
01205               CONST_BITS+2);
01206 
01207     /* Odd part */
01208 
01209     tmp10 = tmp0 + tmp4;
01210     tmp11 = tmp1 - tmp3;
01211     dataptr[DCTSIZE*5] = (DCTELEM)
01212       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp2, FIX(1.28)),  /* 32/25 */
01213               CONST_BITS+2);
01214     tmp2 = MULTIPLY(tmp2, FIX(1.28));                     /* 32/25 */
01215     dataptr[DCTSIZE*1] = (DCTELEM)
01216       DESCALE(MULTIPLY(tmp0, FIX(1.787906876)) +          /* c1 */
01217               MULTIPLY(tmp1, FIX(1.612894094)) + tmp2 +   /* c3 */
01218               MULTIPLY(tmp3, FIX(0.821810588)) +          /* c7 */
01219               MULTIPLY(tmp4, FIX(0.283176630)),           /* c9 */
01220               CONST_BITS+2);
01221     tmp12 = MULTIPLY(tmp0 - tmp4, FIX(1.217352341)) -     /* (c3+c7)/2 */
01222             MULTIPLY(tmp1 + tmp3, FIX(0.752365123));      /* (c1-c9)/2 */
01223     tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.395541753)) +   /* (c3-c7)/2 */
01224             MULTIPLY(tmp11, FIX(0.64)) - tmp2;            /* 16/25 */
01225     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS+2);
01226     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS+2);
01227 
01228     dataptr++;                  /* advance pointer to next column */
01229     wsptr++;                    /* advance pointer to next column */
01230   }
01231 }
01232 
01233 
01234 /*
01235  * Perform the forward DCT on an 11x11 sample block.
01236  */
01237 
01238 GLOBAL(void)
01239 jpeg_fdct_11x11 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01240 {
01241   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
01242   INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
01243   INT32 z1, z2, z3;
01244   DCTELEM workspace[8*3];
01245   DCTELEM *dataptr;
01246   DCTELEM *wsptr;
01247   JSAMPROW elemptr;
01248   int ctr;
01249   SHIFT_TEMPS
01250 
01251   /* Pass 1: process rows. */
01252   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
01253   /* we scale the results further by 2 as part of output adaption */
01254   /* scaling for different DCT size. */
01255   /* cK represents sqrt(2) * cos(K*pi/22). */
01256 
01257   dataptr = data;
01258   ctr = 0;
01259   for (;;) {
01260     elemptr = sample_data[ctr] + start_col;
01261 
01262     /* Even part */
01263 
01264     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[10]);
01265     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[9]);
01266     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[8]);
01267     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[7]);
01268     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[6]);
01269     tmp5 = GETJSAMPLE(elemptr[5]);
01270 
01271     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[10]);
01272     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[9]);
01273     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[8]);
01274     tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[7]);
01275     tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[6]);
01276 
01277     /* Apply unsigned->signed conversion */
01278     dataptr[0] = (DCTELEM)
01279       ((tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 - 11 * CENTERJSAMPLE) << 1);
01280     tmp5 += tmp5;
01281     tmp0 -= tmp5;
01282     tmp1 -= tmp5;
01283     tmp2 -= tmp5;
01284     tmp3 -= tmp5;
01285     tmp4 -= tmp5;
01286     z1 = MULTIPLY(tmp0 + tmp3, FIX(1.356927976)) +       /* c2 */
01287          MULTIPLY(tmp2 + tmp4, FIX(0.201263574));        /* c10 */
01288     z2 = MULTIPLY(tmp1 - tmp3, FIX(0.926112931));        /* c6 */
01289     z3 = MULTIPLY(tmp0 - tmp1, FIX(1.189712156));        /* c4 */
01290     dataptr[2] = (DCTELEM)
01291       DESCALE(z1 + z2 - MULTIPLY(tmp3, FIX(1.018300590)) /* c2+c8-c6 */
01292               - MULTIPLY(tmp4, FIX(1.390975730)),        /* c4+c10 */
01293               CONST_BITS-1);
01294     dataptr[4] = (DCTELEM)
01295       DESCALE(z2 + z3 + MULTIPLY(tmp1, FIX(0.062335650)) /* c4-c6-c10 */
01296               - MULTIPLY(tmp2, FIX(1.356927976))         /* c2 */
01297               + MULTIPLY(tmp4, FIX(0.587485545)),        /* c8 */
01298               CONST_BITS-1);
01299     dataptr[6] = (DCTELEM)
01300       DESCALE(z1 + z3 - MULTIPLY(tmp0, FIX(1.620527200)) /* c2+c4-c6 */
01301               - MULTIPLY(tmp2, FIX(0.788749120)),        /* c8+c10 */
01302               CONST_BITS-1);
01303 
01304     /* Odd part */
01305 
01306     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.286413905));    /* c3 */
01307     tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.068791298));    /* c5 */
01308     tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.764581576));    /* c7 */
01309     tmp0 = tmp1 + tmp2 + tmp3 - MULTIPLY(tmp10, FIX(1.719967871)) /* c7+c5+c3-c1 */
01310            + MULTIPLY(tmp14, FIX(0.398430003));          /* c9 */
01311     tmp4 = MULTIPLY(tmp11 + tmp12, - FIX(0.764581576));  /* -c7 */
01312     tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.399818907));  /* -c1 */
01313     tmp1 += tmp4 + tmp5 + MULTIPLY(tmp11, FIX(1.276416582)) /* c9+c7+c1-c3 */
01314             - MULTIPLY(tmp14, FIX(1.068791298));         /* c5 */
01315     tmp10 = MULTIPLY(tmp12 + tmp13, FIX(0.398430003));   /* c9 */
01316     tmp2 += tmp4 + tmp10 - MULTIPLY(tmp12, FIX(1.989053629)) /* c9+c5+c3-c7 */
01317             + MULTIPLY(tmp14, FIX(1.399818907));         /* c1 */
01318     tmp3 += tmp5 + tmp10 + MULTIPLY(tmp13, FIX(1.305598626)) /* c1+c5-c9-c7 */
01319             - MULTIPLY(tmp14, FIX(1.286413905));         /* c3 */
01320 
01321     dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-1);
01322     dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-1);
01323     dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-1);
01324     dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS-1);
01325 
01326     ctr++;
01327 
01328     if (ctr != DCTSIZE) {
01329       if (ctr == 11)
01330         break;                  /* Done. */
01331       dataptr += DCTSIZE;       /* advance pointer to next row */
01332     } else
01333       dataptr = workspace;      /* switch pointer to extended workspace */
01334   }
01335 
01336   /* Pass 2: process columns.
01337    * We leave the results scaled up by an overall factor of 8.
01338    * We must also scale the output by (8/11)**2 = 64/121, which we partially
01339    * fold into the constant multipliers and final/initial shifting:
01340    * cK now represents sqrt(2) * cos(K*pi/22) * 128/121.
01341    */
01342 
01343   dataptr = data;
01344   wsptr = workspace;
01345   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01346     /* Even part */
01347 
01348     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*2];
01349     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*1];
01350     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*0];
01351     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*7];
01352     tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*6];
01353     tmp5 = dataptr[DCTSIZE*5];
01354 
01355     tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*2];
01356     tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*1];
01357     tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*0];
01358     tmp13 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*7];
01359     tmp14 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*6];
01360 
01361     dataptr[DCTSIZE*0] = (DCTELEM)
01362       DESCALE(MULTIPLY(tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5,
01363                        FIX(1.057851240)),                /* 128/121 */
01364               CONST_BITS+2);
01365     tmp5 += tmp5;
01366     tmp0 -= tmp5;
01367     tmp1 -= tmp5;
01368     tmp2 -= tmp5;
01369     tmp3 -= tmp5;
01370     tmp4 -= tmp5;
01371     z1 = MULTIPLY(tmp0 + tmp3, FIX(1.435427942)) +       /* c2 */
01372          MULTIPLY(tmp2 + tmp4, FIX(0.212906922));        /* c10 */
01373     z2 = MULTIPLY(tmp1 - tmp3, FIX(0.979689713));        /* c6 */
01374     z3 = MULTIPLY(tmp0 - tmp1, FIX(1.258538479));        /* c4 */
01375     dataptr[DCTSIZE*2] = (DCTELEM)
01376       DESCALE(z1 + z2 - MULTIPLY(tmp3, FIX(1.077210542)) /* c2+c8-c6 */
01377               - MULTIPLY(tmp4, FIX(1.471445400)),        /* c4+c10 */
01378               CONST_BITS+2);
01379     dataptr[DCTSIZE*4] = (DCTELEM)
01380       DESCALE(z2 + z3 + MULTIPLY(tmp1, FIX(0.065941844)) /* c4-c6-c10 */
01381               - MULTIPLY(tmp2, FIX(1.435427942))         /* c2 */
01382               + MULTIPLY(tmp4, FIX(0.621472312)),        /* c8 */
01383               CONST_BITS+2);
01384     dataptr[DCTSIZE*6] = (DCTELEM)
01385       DESCALE(z1 + z3 - MULTIPLY(tmp0, FIX(1.714276708)) /* c2+c4-c6 */
01386               - MULTIPLY(tmp2, FIX(0.834379234)),        /* c8+c10 */
01387               CONST_BITS+2);
01388 
01389     /* Odd part */
01390 
01391     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.360834544));    /* c3 */
01392     tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.130622199));    /* c5 */
01393     tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.808813568));    /* c7 */
01394     tmp0 = tmp1 + tmp2 + tmp3 - MULTIPLY(tmp10, FIX(1.819470145)) /* c7+c5+c3-c1 */
01395            + MULTIPLY(tmp14, FIX(0.421479672));          /* c9 */
01396     tmp4 = MULTIPLY(tmp11 + tmp12, - FIX(0.808813568));  /* -c7 */
01397     tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.480800167));  /* -c1 */
01398     tmp1 += tmp4 + tmp5 + MULTIPLY(tmp11, FIX(1.350258864)) /* c9+c7+c1-c3 */
01399             - MULTIPLY(tmp14, FIX(1.130622199));         /* c5 */
01400     tmp10 = MULTIPLY(tmp12 + tmp13, FIX(0.421479672));   /* c9 */
01401     tmp2 += tmp4 + tmp10 - MULTIPLY(tmp12, FIX(2.104122847)) /* c9+c5+c3-c7 */
01402             + MULTIPLY(tmp14, FIX(1.480800167));         /* c1 */
01403     tmp3 += tmp5 + tmp10 + MULTIPLY(tmp13, FIX(1.381129125)) /* c1+c5-c9-c7 */
01404             - MULTIPLY(tmp14, FIX(1.360834544));         /* c3 */
01405 
01406     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+2);
01407     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+2);
01408     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+2);
01409     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+2);
01410 
01411     dataptr++;                  /* advance pointer to next column */
01412     wsptr++;                    /* advance pointer to next column */
01413   }
01414 }
01415 
01416 
01417 /*
01418  * Perform the forward DCT on a 12x12 sample block.
01419  */
01420 
01421 GLOBAL(void)
01422 jpeg_fdct_12x12 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01423 {
01424   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
01425   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
01426   DCTELEM workspace[8*4];
01427   DCTELEM *dataptr;
01428   DCTELEM *wsptr;
01429   JSAMPROW elemptr;
01430   int ctr;
01431   SHIFT_TEMPS
01432 
01433   /* Pass 1: process rows. */
01434   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
01435   /* cK represents sqrt(2) * cos(K*pi/24). */
01436 
01437   dataptr = data;
01438   ctr = 0;
01439   for (;;) {
01440     elemptr = sample_data[ctr] + start_col;
01441 
01442     /* Even part */
01443 
01444     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[11]);
01445     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[10]);
01446     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[9]);
01447     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[8]);
01448     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[7]);
01449     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[6]);
01450 
01451     tmp10 = tmp0 + tmp5;
01452     tmp13 = tmp0 - tmp5;
01453     tmp11 = tmp1 + tmp4;
01454     tmp14 = tmp1 - tmp4;
01455     tmp12 = tmp2 + tmp3;
01456     tmp15 = tmp2 - tmp3;
01457 
01458     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[11]);
01459     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[10]);
01460     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[9]);
01461     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[8]);
01462     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[7]);
01463     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[6]);
01464 
01465     /* Apply unsigned->signed conversion */
01466     dataptr[0] = (DCTELEM) (tmp10 + tmp11 + tmp12 - 12 * CENTERJSAMPLE);
01467     dataptr[6] = (DCTELEM) (tmp13 - tmp14 - tmp15);
01468     dataptr[4] = (DCTELEM)
01469       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.224744871)), /* c4 */
01470               CONST_BITS);
01471     dataptr[2] = (DCTELEM)
01472       DESCALE(tmp14 - tmp15 + MULTIPLY(tmp13 + tmp15, FIX(1.366025404)), /* c2 */
01473               CONST_BITS);
01474 
01475     /* Odd part */
01476 
01477     tmp10 = MULTIPLY(tmp1 + tmp4, FIX_0_541196100);    /* c9 */
01478     tmp14 = tmp10 + MULTIPLY(tmp1, FIX_0_765366865);   /* c3-c9 */
01479     tmp15 = tmp10 - MULTIPLY(tmp4, FIX_1_847759065);   /* c3+c9 */
01480     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.121971054));   /* c5 */
01481     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.860918669));   /* c7 */
01482     tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.580774953)) /* c5+c7-c1 */
01483             + MULTIPLY(tmp5, FIX(0.184591911));        /* c11 */
01484     tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.184591911)); /* -c11 */
01485     tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.339493912)) /* c1+c5-c11 */
01486             + MULTIPLY(tmp5, FIX(0.860918669));        /* c7 */
01487     tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.725788011)) /* c1+c11-c7 */
01488             - MULTIPLY(tmp5, FIX(1.121971054));        /* c5 */
01489     tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.306562965)) /* c3 */
01490             - MULTIPLY(tmp2 + tmp5, FIX_0_541196100);  /* c9 */
01491 
01492     dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS);
01493     dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS);
01494     dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS);
01495     dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS);
01496 
01497     ctr++;
01498 
01499     if (ctr != DCTSIZE) {
01500       if (ctr == 12)
01501         break;                  /* Done. */
01502       dataptr += DCTSIZE;       /* advance pointer to next row */
01503     } else
01504       dataptr = workspace;      /* switch pointer to extended workspace */
01505   }
01506 
01507   /* Pass 2: process columns.
01508    * We leave the results scaled up by an overall factor of 8.
01509    * We must also scale the output by (8/12)**2 = 4/9, which we partially
01510    * fold into the constant multipliers and final shifting:
01511    * cK now represents sqrt(2) * cos(K*pi/24) * 8/9.
01512    */
01513 
01514   dataptr = data;
01515   wsptr = workspace;
01516   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01517     /* Even part */
01518 
01519     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*3];
01520     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*2];
01521     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*1];
01522     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*0];
01523     tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*7];
01524     tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*6];
01525 
01526     tmp10 = tmp0 + tmp5;
01527     tmp13 = tmp0 - tmp5;
01528     tmp11 = tmp1 + tmp4;
01529     tmp14 = tmp1 - tmp4;
01530     tmp12 = tmp2 + tmp3;
01531     tmp15 = tmp2 - tmp3;
01532 
01533     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*3];
01534     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*2];
01535     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*1];
01536     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*0];
01537     tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*7];
01538     tmp5 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*6];
01539 
01540     dataptr[DCTSIZE*0] = (DCTELEM)
01541       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(0.888888889)), /* 8/9 */
01542               CONST_BITS+1);
01543     dataptr[DCTSIZE*6] = (DCTELEM)
01544       DESCALE(MULTIPLY(tmp13 - tmp14 - tmp15, FIX(0.888888889)), /* 8/9 */
01545               CONST_BITS+1);
01546     dataptr[DCTSIZE*4] = (DCTELEM)
01547       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.088662108)),         /* c4 */
01548               CONST_BITS+1);
01549     dataptr[DCTSIZE*2] = (DCTELEM)
01550       DESCALE(MULTIPLY(tmp14 - tmp15, FIX(0.888888889)) +        /* 8/9 */
01551               MULTIPLY(tmp13 + tmp15, FIX(1.214244803)),         /* c2 */
01552               CONST_BITS+1);
01553 
01554     /* Odd part */
01555 
01556     tmp10 = MULTIPLY(tmp1 + tmp4, FIX(0.481063200));   /* c9 */
01557     tmp14 = tmp10 + MULTIPLY(tmp1, FIX(0.680326102));  /* c3-c9 */
01558     tmp15 = tmp10 - MULTIPLY(tmp4, FIX(1.642452502));  /* c3+c9 */
01559     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(0.997307603));   /* c5 */
01560     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.765261039));   /* c7 */
01561     tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.516244403)) /* c5+c7-c1 */
01562             + MULTIPLY(tmp5, FIX(0.164081699));        /* c11 */
01563     tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.164081699)); /* -c11 */
01564     tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.079550144)) /* c1+c5-c11 */
01565             + MULTIPLY(tmp5, FIX(0.765261039));        /* c7 */
01566     tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.645144899)) /* c1+c11-c7 */
01567             - MULTIPLY(tmp5, FIX(0.997307603));        /* c5 */
01568     tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.161389302)) /* c3 */
01569             - MULTIPLY(tmp2 + tmp5, FIX(0.481063200)); /* c9 */
01570 
01571     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+1);
01572     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+1);
01573     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+1);
01574     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+1);
01575 
01576     dataptr++;                  /* advance pointer to next column */
01577     wsptr++;                    /* advance pointer to next column */
01578   }
01579 }
01580 
01581 
01582 /*
01583  * Perform the forward DCT on a 13x13 sample block.
01584  */
01585 
01586 GLOBAL(void)
01587 jpeg_fdct_13x13 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01588 {
01589   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
01590   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
01591   INT32 z1, z2;
01592   DCTELEM workspace[8*5];
01593   DCTELEM *dataptr;
01594   DCTELEM *wsptr;
01595   JSAMPROW elemptr;
01596   int ctr;
01597   SHIFT_TEMPS
01598 
01599   /* Pass 1: process rows. */
01600   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
01601   /* cK represents sqrt(2) * cos(K*pi/26). */
01602 
01603   dataptr = data;
01604   ctr = 0;
01605   for (;;) {
01606     elemptr = sample_data[ctr] + start_col;
01607 
01608     /* Even part */
01609 
01610     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[12]);
01611     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[11]);
01612     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[10]);
01613     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[9]);
01614     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[8]);
01615     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[7]);
01616     tmp6 = GETJSAMPLE(elemptr[6]);
01617 
01618     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[12]);
01619     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[11]);
01620     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[10]);
01621     tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[9]);
01622     tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[8]);
01623     tmp15 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[7]);
01624 
01625     /* Apply unsigned->signed conversion */
01626     dataptr[0] = (DCTELEM)
01627       (tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 + tmp6 - 13 * CENTERJSAMPLE);
01628     tmp6 += tmp6;
01629     tmp0 -= tmp6;
01630     tmp1 -= tmp6;
01631     tmp2 -= tmp6;
01632     tmp3 -= tmp6;
01633     tmp4 -= tmp6;
01634     tmp5 -= tmp6;
01635     dataptr[2] = (DCTELEM)
01636       DESCALE(MULTIPLY(tmp0, FIX(1.373119086)) +   /* c2 */
01637               MULTIPLY(tmp1, FIX(1.058554052)) +   /* c6 */
01638               MULTIPLY(tmp2, FIX(0.501487041)) -   /* c10 */
01639               MULTIPLY(tmp3, FIX(0.170464608)) -   /* c12 */
01640               MULTIPLY(tmp4, FIX(0.803364869)) -   /* c8 */
01641               MULTIPLY(tmp5, FIX(1.252223920)),    /* c4 */
01642               CONST_BITS);
01643     z1 = MULTIPLY(tmp0 - tmp2, FIX(1.155388986)) - /* (c4+c6)/2 */
01644          MULTIPLY(tmp3 - tmp4, FIX(0.435816023)) - /* (c2-c10)/2 */
01645          MULTIPLY(tmp1 - tmp5, FIX(0.316450131));  /* (c8-c12)/2 */
01646     z2 = MULTIPLY(tmp0 + tmp2, FIX(0.096834934)) - /* (c4-c6)/2 */
01647          MULTIPLY(tmp3 + tmp4, FIX(0.937303064)) + /* (c2+c10)/2 */
01648          MULTIPLY(tmp1 + tmp5, FIX(0.486914739));  /* (c8+c12)/2 */
01649 
01650     dataptr[4] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS);
01651     dataptr[6] = (DCTELEM) DESCALE(z1 - z2, CONST_BITS);
01652 
01653     /* Odd part */
01654 
01655     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.322312651));   /* c3 */
01656     tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.163874945));   /* c5 */
01657     tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.937797057)) +  /* c7 */
01658            MULTIPLY(tmp14 + tmp15, FIX(0.338443458));   /* c11 */
01659     tmp0 = tmp1 + tmp2 + tmp3 -
01660            MULTIPLY(tmp10, FIX(2.020082300)) +          /* c3+c5+c7-c1 */
01661            MULTIPLY(tmp14, FIX(0.318774355));           /* c9-c11 */
01662     tmp4 = MULTIPLY(tmp14 - tmp15, FIX(0.937797057)) -  /* c7 */
01663            MULTIPLY(tmp11 + tmp12, FIX(0.338443458));   /* c11 */
01664     tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.163874945)); /* -c5 */
01665     tmp1 += tmp4 + tmp5 +
01666             MULTIPLY(tmp11, FIX(0.837223564)) -         /* c5+c9+c11-c3 */
01667             MULTIPLY(tmp14, FIX(2.341699410));          /* c1+c7 */
01668     tmp6 = MULTIPLY(tmp12 + tmp13, - FIX(0.657217813)); /* -c9 */
01669     tmp2 += tmp4 + tmp6 -
01670             MULTIPLY(tmp12, FIX(1.572116027)) +         /* c1+c5-c9-c11 */
01671             MULTIPLY(tmp15, FIX(2.260109708));          /* c3+c7 */
01672     tmp3 += tmp5 + tmp6 +
01673             MULTIPLY(tmp13, FIX(2.205608352)) -         /* c3+c5+c9-c7 */
01674             MULTIPLY(tmp15, FIX(1.742345811));          /* c1+c11 */
01675 
01676     dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS);
01677     dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS);
01678     dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS);
01679     dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS);
01680 
01681     ctr++;
01682 
01683     if (ctr != DCTSIZE) {
01684       if (ctr == 13)
01685         break;                  /* Done. */
01686       dataptr += DCTSIZE;       /* advance pointer to next row */
01687     } else
01688       dataptr = workspace;      /* switch pointer to extended workspace */
01689   }
01690 
01691   /* Pass 2: process columns.
01692    * We leave the results scaled up by an overall factor of 8.
01693    * We must also scale the output by (8/13)**2 = 64/169, which we partially
01694    * fold into the constant multipliers and final shifting:
01695    * cK now represents sqrt(2) * cos(K*pi/26) * 128/169.
01696    */
01697 
01698   dataptr = data;
01699   wsptr = workspace;
01700   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01701     /* Even part */
01702 
01703     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*4];
01704     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*3];
01705     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*2];
01706     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*1];
01707     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*0];
01708     tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*7];
01709     tmp6 = dataptr[DCTSIZE*6];
01710 
01711     tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*4];
01712     tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*3];
01713     tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*2];
01714     tmp13 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*1];
01715     tmp14 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*0];
01716     tmp15 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*7];
01717 
01718     dataptr[DCTSIZE*0] = (DCTELEM)
01719       DESCALE(MULTIPLY(tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 + tmp6,
01720                        FIX(0.757396450)),          /* 128/169 */
01721               CONST_BITS+1);
01722     tmp6 += tmp6;
01723     tmp0 -= tmp6;
01724     tmp1 -= tmp6;
01725     tmp2 -= tmp6;
01726     tmp3 -= tmp6;
01727     tmp4 -= tmp6;
01728     tmp5 -= tmp6;
01729     dataptr[DCTSIZE*2] = (DCTELEM)
01730       DESCALE(MULTIPLY(tmp0, FIX(1.039995521)) +   /* c2 */
01731               MULTIPLY(tmp1, FIX(0.801745081)) +   /* c6 */
01732               MULTIPLY(tmp2, FIX(0.379824504)) -   /* c10 */
01733               MULTIPLY(tmp3, FIX(0.129109289)) -   /* c12 */
01734               MULTIPLY(tmp4, FIX(0.608465700)) -   /* c8 */
01735               MULTIPLY(tmp5, FIX(0.948429952)),    /* c4 */
01736               CONST_BITS+1);
01737     z1 = MULTIPLY(tmp0 - tmp2, FIX(0.875087516)) - /* (c4+c6)/2 */
01738          MULTIPLY(tmp3 - tmp4, FIX(0.330085509)) - /* (c2-c10)/2 */
01739          MULTIPLY(tmp1 - tmp5, FIX(0.239678205));  /* (c8-c12)/2 */
01740     z2 = MULTIPLY(tmp0 + tmp2, FIX(0.073342435)) - /* (c4-c6)/2 */
01741          MULTIPLY(tmp3 + tmp4, FIX(0.709910013)) + /* (c2+c10)/2 */
01742          MULTIPLY(tmp1 + tmp5, FIX(0.368787494));  /* (c8+c12)/2 */
01743 
01744     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+1);
01745     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 - z2, CONST_BITS+1);
01746 
01747     /* Odd part */
01748 
01749     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.001514908));   /* c3 */
01750     tmp2 = MULTIPLY(tmp10 + tmp12, FIX(0.881514751));   /* c5 */
01751     tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.710284161)) +  /* c7 */
01752            MULTIPLY(tmp14 + tmp15, FIX(0.256335874));   /* c11 */
01753     tmp0 = tmp1 + tmp2 + tmp3 -
01754            MULTIPLY(tmp10, FIX(1.530003162)) +          /* c3+c5+c7-c1 */
01755            MULTIPLY(tmp14, FIX(0.241438564));           /* c9-c11 */
01756     tmp4 = MULTIPLY(tmp14 - tmp15, FIX(0.710284161)) -  /* c7 */
01757            MULTIPLY(tmp11 + tmp12, FIX(0.256335874));   /* c11 */
01758     tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(0.881514751)); /* -c5 */
01759     tmp1 += tmp4 + tmp5 +
01760             MULTIPLY(tmp11, FIX(0.634110155)) -         /* c5+c9+c11-c3 */
01761             MULTIPLY(tmp14, FIX(1.773594819));          /* c1+c7 */
01762     tmp6 = MULTIPLY(tmp12 + tmp13, - FIX(0.497774438)); /* -c9 */
01763     tmp2 += tmp4 + tmp6 -
01764             MULTIPLY(tmp12, FIX(1.190715098)) +         /* c1+c5-c9-c11 */
01765             MULTIPLY(tmp15, FIX(1.711799069));          /* c3+c7 */
01766     tmp3 += tmp5 + tmp6 +
01767             MULTIPLY(tmp13, FIX(1.670519935)) -         /* c3+c5+c9-c7 */
01768             MULTIPLY(tmp15, FIX(1.319646532));          /* c1+c11 */
01769 
01770     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+1);
01771     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+1);
01772     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+1);
01773     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+1);
01774 
01775     dataptr++;                  /* advance pointer to next column */
01776     wsptr++;                    /* advance pointer to next column */
01777   }
01778 }
01779 
01780 
01781 /*
01782  * Perform the forward DCT on a 14x14 sample block.
01783  */
01784 
01785 GLOBAL(void)
01786 jpeg_fdct_14x14 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01787 {
01788   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
01789   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
01790   DCTELEM workspace[8*6];
01791   DCTELEM *dataptr;
01792   DCTELEM *wsptr;
01793   JSAMPROW elemptr;
01794   int ctr;
01795   SHIFT_TEMPS
01796 
01797   /* Pass 1: process rows. */
01798   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
01799   /* cK represents sqrt(2) * cos(K*pi/28). */
01800 
01801   dataptr = data;
01802   ctr = 0;
01803   for (;;) {
01804     elemptr = sample_data[ctr] + start_col;
01805 
01806     /* Even part */
01807 
01808     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[13]);
01809     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[12]);
01810     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[11]);
01811     tmp13 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[10]);
01812     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[9]);
01813     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[8]);
01814     tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[7]);
01815 
01816     tmp10 = tmp0 + tmp6;
01817     tmp14 = tmp0 - tmp6;
01818     tmp11 = tmp1 + tmp5;
01819     tmp15 = tmp1 - tmp5;
01820     tmp12 = tmp2 + tmp4;
01821     tmp16 = tmp2 - tmp4;
01822 
01823     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[13]);
01824     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[12]);
01825     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[11]);
01826     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[10]);
01827     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[9]);
01828     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[8]);
01829     tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[7]);
01830 
01831     /* Apply unsigned->signed conversion */
01832     dataptr[0] = (DCTELEM)
01833       (tmp10 + tmp11 + tmp12 + tmp13 - 14 * CENTERJSAMPLE);
01834     tmp13 += tmp13;
01835     dataptr[4] = (DCTELEM)
01836       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.274162392)) + /* c4 */
01837               MULTIPLY(tmp11 - tmp13, FIX(0.314692123)) - /* c12 */
01838               MULTIPLY(tmp12 - tmp13, FIX(0.881747734)),  /* c8 */
01839               CONST_BITS);
01840 
01841     tmp10 = MULTIPLY(tmp14 + tmp15, FIX(1.105676686));    /* c6 */
01842 
01843     dataptr[2] = (DCTELEM)
01844       DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.273079590))   /* c2-c6 */
01845               + MULTIPLY(tmp16, FIX(0.613604268)),        /* c10 */
01846               CONST_BITS);
01847     dataptr[6] = (DCTELEM)
01848       DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.719280954))   /* c6+c10 */
01849               - MULTIPLY(tmp16, FIX(1.378756276)),        /* c2 */
01850               CONST_BITS);
01851 
01852     /* Odd part */
01853 
01854     tmp10 = tmp1 + tmp2;
01855     tmp11 = tmp5 - tmp4;
01856     dataptr[7] = (DCTELEM) (tmp0 - tmp10 + tmp3 - tmp11 - tmp6);
01857     tmp3 <<= CONST_BITS;
01858     tmp10 = MULTIPLY(tmp10, - FIX(0.158341681));          /* -c13 */
01859     tmp11 = MULTIPLY(tmp11, FIX(1.405321284));            /* c1 */
01860     tmp10 += tmp11 - tmp3;
01861     tmp11 = MULTIPLY(tmp0 + tmp2, FIX(1.197448846)) +     /* c5 */
01862             MULTIPLY(tmp4 + tmp6, FIX(0.752406978));      /* c9 */
01863     dataptr[5] = (DCTELEM)
01864       DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(2.373959773)) /* c3+c5-c13 */
01865               + MULTIPLY(tmp4, FIX(1.119999435)),         /* c1+c11-c9 */
01866               CONST_BITS);
01867     tmp12 = MULTIPLY(tmp0 + tmp1, FIX(1.334852607)) +     /* c3 */
01868             MULTIPLY(tmp5 - tmp6, FIX(0.467085129));      /* c11 */
01869     dataptr[3] = (DCTELEM)
01870       DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.424103948)) /* c3-c9-c13 */
01871               - MULTIPLY(tmp5, FIX(3.069855259)),         /* c1+c5+c11 */
01872               CONST_BITS);
01873     dataptr[1] = (DCTELEM)
01874       DESCALE(tmp11 + tmp12 + tmp3 + tmp6 -
01875               MULTIPLY(tmp0 + tmp6, FIX(1.126980169)),    /* c3+c5-c1 */
01876               CONST_BITS);
01877 
01878     ctr++;
01879 
01880     if (ctr != DCTSIZE) {
01881       if (ctr == 14)
01882         break;                  /* Done. */
01883       dataptr += DCTSIZE;       /* advance pointer to next row */
01884     } else
01885       dataptr = workspace;      /* switch pointer to extended workspace */
01886   }
01887 
01888   /* Pass 2: process columns.
01889    * We leave the results scaled up by an overall factor of 8.
01890    * We must also scale the output by (8/14)**2 = 16/49, which we partially
01891    * fold into the constant multipliers and final shifting:
01892    * cK now represents sqrt(2) * cos(K*pi/28) * 32/49.
01893    */
01894 
01895   dataptr = data;
01896   wsptr = workspace;
01897   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
01898     /* Even part */
01899 
01900     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*5];
01901     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*4];
01902     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*3];
01903     tmp13 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*2];
01904     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*1];
01905     tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*0];
01906     tmp6 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
01907 
01908     tmp10 = tmp0 + tmp6;
01909     tmp14 = tmp0 - tmp6;
01910     tmp11 = tmp1 + tmp5;
01911     tmp15 = tmp1 - tmp5;
01912     tmp12 = tmp2 + tmp4;
01913     tmp16 = tmp2 - tmp4;
01914 
01915     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*5];
01916     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*4];
01917     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*3];
01918     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*2];
01919     tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*1];
01920     tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*0];
01921     tmp6 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
01922 
01923     dataptr[DCTSIZE*0] = (DCTELEM)
01924       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12 + tmp13,
01925                        FIX(0.653061224)),                 /* 32/49 */
01926               CONST_BITS+1);
01927     tmp13 += tmp13;
01928     dataptr[DCTSIZE*4] = (DCTELEM)
01929       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(0.832106052)) + /* c4 */
01930               MULTIPLY(tmp11 - tmp13, FIX(0.205513223)) - /* c12 */
01931               MULTIPLY(tmp12 - tmp13, FIX(0.575835255)),  /* c8 */
01932               CONST_BITS+1);
01933 
01934     tmp10 = MULTIPLY(tmp14 + tmp15, FIX(0.722074570));    /* c6 */
01935 
01936     dataptr[DCTSIZE*2] = (DCTELEM)
01937       DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.178337691))   /* c2-c6 */
01938               + MULTIPLY(tmp16, FIX(0.400721155)),        /* c10 */
01939               CONST_BITS+1);
01940     dataptr[DCTSIZE*6] = (DCTELEM)
01941       DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.122795725))   /* c6+c10 */
01942               - MULTIPLY(tmp16, FIX(0.900412262)),        /* c2 */
01943               CONST_BITS+1);
01944 
01945     /* Odd part */
01946 
01947     tmp10 = tmp1 + tmp2;
01948     tmp11 = tmp5 - tmp4;
01949     dataptr[DCTSIZE*7] = (DCTELEM)
01950       DESCALE(MULTIPLY(tmp0 - tmp10 + tmp3 - tmp11 - tmp6,
01951                        FIX(0.653061224)),                 /* 32/49 */
01952               CONST_BITS+1);
01953     tmp3  = MULTIPLY(tmp3 , FIX(0.653061224));            /* 32/49 */
01954     tmp10 = MULTIPLY(tmp10, - FIX(0.103406812));          /* -c13 */
01955     tmp11 = MULTIPLY(tmp11, FIX(0.917760839));            /* c1 */
01956     tmp10 += tmp11 - tmp3;
01957     tmp11 = MULTIPLY(tmp0 + tmp2, FIX(0.782007410)) +     /* c5 */
01958             MULTIPLY(tmp4 + tmp6, FIX(0.491367823));      /* c9 */
01959     dataptr[DCTSIZE*5] = (DCTELEM)
01960       DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(1.550341076)) /* c3+c5-c13 */
01961               + MULTIPLY(tmp4, FIX(0.731428202)),         /* c1+c11-c9 */
01962               CONST_BITS+1);
01963     tmp12 = MULTIPLY(tmp0 + tmp1, FIX(0.871740478)) +     /* c3 */
01964             MULTIPLY(tmp5 - tmp6, FIX(0.305035186));      /* c11 */
01965     dataptr[DCTSIZE*3] = (DCTELEM)
01966       DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.276965844)) /* c3-c9-c13 */
01967               - MULTIPLY(tmp5, FIX(2.004803435)),         /* c1+c5+c11 */
01968               CONST_BITS+1);
01969     dataptr[DCTSIZE*1] = (DCTELEM)
01970       DESCALE(tmp11 + tmp12 + tmp3
01971               - MULTIPLY(tmp0, FIX(0.735987049))          /* c3+c5-c1 */
01972               - MULTIPLY(tmp6, FIX(0.082925825)),         /* c9-c11-c13 */
01973               CONST_BITS+1);
01974 
01975     dataptr++;                  /* advance pointer to next column */
01976     wsptr++;                    /* advance pointer to next column */
01977   }
01978 }
01979 
01980 
01981 /*
01982  * Perform the forward DCT on a 15x15 sample block.
01983  */
01984 
01985 GLOBAL(void)
01986 jpeg_fdct_15x15 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
01987 {
01988   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
01989   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
01990   INT32 z1, z2, z3;
01991   DCTELEM workspace[8*7];
01992   DCTELEM *dataptr;
01993   DCTELEM *wsptr;
01994   JSAMPROW elemptr;
01995   int ctr;
01996   SHIFT_TEMPS
01997 
01998   /* Pass 1: process rows. */
01999   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
02000   /* cK represents sqrt(2) * cos(K*pi/30). */
02001 
02002   dataptr = data;
02003   ctr = 0;
02004   for (;;) {
02005     elemptr = sample_data[ctr] + start_col;
02006 
02007     /* Even part */
02008 
02009     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[14]);
02010     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[13]);
02011     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[12]);
02012     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[11]);
02013     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[10]);
02014     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[9]);
02015     tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[8]);
02016     tmp7 = GETJSAMPLE(elemptr[7]);
02017 
02018     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[14]);
02019     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[13]);
02020     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[12]);
02021     tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[11]);
02022     tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[10]);
02023     tmp15 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[9]);
02024     tmp16 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[8]);
02025 
02026     z1 = tmp0 + tmp4 + tmp5;
02027     z2 = tmp1 + tmp3 + tmp6;
02028     z3 = tmp2 + tmp7;
02029     /* Apply unsigned->signed conversion */
02030     dataptr[0] = (DCTELEM) (z1 + z2 + z3 - 15 * CENTERJSAMPLE);
02031     z3 += z3;
02032     dataptr[6] = (DCTELEM)
02033       DESCALE(MULTIPLY(z1 - z3, FIX(1.144122806)) - /* c6 */
02034               MULTIPLY(z2 - z3, FIX(0.437016024)),  /* c12 */
02035               CONST_BITS);
02036     tmp2 += ((tmp1 + tmp4) >> 1) - tmp7 - tmp7;
02037     z1 = MULTIPLY(tmp3 - tmp2, FIX(1.531135173)) -  /* c2+c14 */
02038          MULTIPLY(tmp6 - tmp2, FIX(2.238241955));   /* c4+c8 */
02039     z2 = MULTIPLY(tmp5 - tmp2, FIX(0.798468008)) -  /* c8-c14 */
02040          MULTIPLY(tmp0 - tmp2, FIX(0.091361227));   /* c2-c4 */
02041     z3 = MULTIPLY(tmp0 - tmp3, FIX(1.383309603)) +  /* c2 */
02042          MULTIPLY(tmp6 - tmp5, FIX(0.946293579)) +  /* c8 */
02043          MULTIPLY(tmp1 - tmp4, FIX(0.790569415));   /* (c6+c12)/2 */
02044 
02045     dataptr[2] = (DCTELEM) DESCALE(z1 + z3, CONST_BITS);
02046     dataptr[4] = (DCTELEM) DESCALE(z2 + z3, CONST_BITS);
02047 
02048     /* Odd part */
02049 
02050     tmp2 = MULTIPLY(tmp10 - tmp12 - tmp13 + tmp15 + tmp16,
02051                     FIX(1.224744871));                         /* c5 */
02052     tmp1 = MULTIPLY(tmp10 - tmp14 - tmp15, FIX(1.344997024)) + /* c3 */
02053            MULTIPLY(tmp11 - tmp13 - tmp16, FIX(0.831253876));  /* c9 */
02054     tmp12 = MULTIPLY(tmp12, FIX(1.224744871));                 /* c5 */
02055     tmp4 = MULTIPLY(tmp10 - tmp16, FIX(1.406466353)) +         /* c1 */
02056            MULTIPLY(tmp11 + tmp14, FIX(1.344997024)) +         /* c3 */
02057            MULTIPLY(tmp13 + tmp15, FIX(0.575212477));          /* c11 */
02058     tmp0 = MULTIPLY(tmp13, FIX(0.475753014)) -                 /* c7-c11 */
02059            MULTIPLY(tmp14, FIX(0.513743148)) +                 /* c3-c9 */
02060            MULTIPLY(tmp16, FIX(1.700497885)) + tmp4 + tmp12;   /* c1+c13 */
02061     tmp3 = MULTIPLY(tmp10, - FIX(0.355500862)) -               /* -(c1-c7) */
02062            MULTIPLY(tmp11, FIX(2.176250899)) -                 /* c3+c9 */
02063            MULTIPLY(tmp15, FIX(0.869244010)) + tmp4 - tmp12;   /* c11+c13 */
02064 
02065     dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS);
02066     dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS);
02067     dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS);
02068     dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS);
02069 
02070     ctr++;
02071 
02072     if (ctr != DCTSIZE) {
02073       if (ctr == 15)
02074         break;                  /* Done. */
02075       dataptr += DCTSIZE;       /* advance pointer to next row */
02076     } else
02077       dataptr = workspace;      /* switch pointer to extended workspace */
02078   }
02079 
02080   /* Pass 2: process columns.
02081    * We leave the results scaled up by an overall factor of 8.
02082    * We must also scale the output by (8/15)**2 = 64/225, which we partially
02083    * fold into the constant multipliers and final shifting:
02084    * cK now represents sqrt(2) * cos(K*pi/30) * 256/225.
02085    */
02086 
02087   dataptr = data;
02088   wsptr = workspace;
02089   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02090     /* Even part */
02091 
02092     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*6];
02093     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*5];
02094     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*4];
02095     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*3];
02096     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*2];
02097     tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*1];
02098     tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*0];
02099     tmp7 = dataptr[DCTSIZE*7];
02100 
02101     tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*6];
02102     tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*5];
02103     tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*4];
02104     tmp13 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*3];
02105     tmp14 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*2];
02106     tmp15 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*1];
02107     tmp16 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*0];
02108 
02109     z1 = tmp0 + tmp4 + tmp5;
02110     z2 = tmp1 + tmp3 + tmp6;
02111     z3 = tmp2 + tmp7;
02112     dataptr[DCTSIZE*0] = (DCTELEM)
02113       DESCALE(MULTIPLY(z1 + z2 + z3, FIX(1.137777778)), /* 256/225 */
02114               CONST_BITS+2);
02115     z3 += z3;
02116     dataptr[DCTSIZE*6] = (DCTELEM)
02117       DESCALE(MULTIPLY(z1 - z3, FIX(1.301757503)) - /* c6 */
02118               MULTIPLY(z2 - z3, FIX(0.497227121)),  /* c12 */
02119               CONST_BITS+2);
02120     tmp2 += ((tmp1 + tmp4) >> 1) - tmp7 - tmp7;
02121     z1 = MULTIPLY(tmp3 - tmp2, FIX(1.742091575)) -  /* c2+c14 */
02122          MULTIPLY(tmp6 - tmp2, FIX(2.546621957));   /* c4+c8 */
02123     z2 = MULTIPLY(tmp5 - tmp2, FIX(0.908479156)) -  /* c8-c14 */
02124          MULTIPLY(tmp0 - tmp2, FIX(0.103948774));   /* c2-c4 */
02125     z3 = MULTIPLY(tmp0 - tmp3, FIX(1.573898926)) +  /* c2 */
02126          MULTIPLY(tmp6 - tmp5, FIX(1.076671805)) +  /* c8 */
02127          MULTIPLY(tmp1 - tmp4, FIX(0.899492312));   /* (c6+c12)/2 */
02128 
02129     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z3, CONST_BITS+2);
02130     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(z2 + z3, CONST_BITS+2);
02131 
02132     /* Odd part */
02133 
02134     tmp2 = MULTIPLY(tmp10 - tmp12 - tmp13 + tmp15 + tmp16,
02135                     FIX(1.393487498));                         /* c5 */
02136     tmp1 = MULTIPLY(tmp10 - tmp14 - tmp15, FIX(1.530307725)) + /* c3 */
02137            MULTIPLY(tmp11 - tmp13 - tmp16, FIX(0.945782187));  /* c9 */
02138     tmp12 = MULTIPLY(tmp12, FIX(1.393487498));                 /* c5 */
02139     tmp4 = MULTIPLY(tmp10 - tmp16, FIX(1.600246161)) +         /* c1 */
02140            MULTIPLY(tmp11 + tmp14, FIX(1.530307725)) +         /* c3 */
02141            MULTIPLY(tmp13 + tmp15, FIX(0.654463974));          /* c11 */
02142     tmp0 = MULTIPLY(tmp13, FIX(0.541301207)) -                 /* c7-c11 */
02143            MULTIPLY(tmp14, FIX(0.584525538)) +                 /* c3-c9 */
02144            MULTIPLY(tmp16, FIX(1.934788705)) + tmp4 + tmp12;   /* c1+c13 */
02145     tmp3 = MULTIPLY(tmp10, - FIX(0.404480980)) -               /* -(c1-c7) */
02146            MULTIPLY(tmp11, FIX(2.476089912)) -                 /* c3+c9 */
02147            MULTIPLY(tmp15, FIX(0.989006518)) + tmp4 - tmp12;   /* c11+c13 */
02148 
02149     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+2);
02150     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+2);
02151     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+2);
02152     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+2);
02153 
02154     dataptr++;                  /* advance pointer to next column */
02155     wsptr++;                    /* advance pointer to next column */
02156   }
02157 }
02158 
02159 
02160 /*
02161  * Perform the forward DCT on a 16x16 sample block.
02162  */
02163 
02164 GLOBAL(void)
02165 jpeg_fdct_16x16 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02166 {
02167   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
02168   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
02169   DCTELEM workspace[DCTSIZE2];
02170   DCTELEM *dataptr;
02171   DCTELEM *wsptr;
02172   JSAMPROW elemptr;
02173   int ctr;
02174   SHIFT_TEMPS
02175 
02176   /* Pass 1: process rows. */
02177   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
02178   /* furthermore, we scale the results by 2**PASS1_BITS. */
02179   /* cK represents sqrt(2) * cos(K*pi/32). */
02180 
02181   dataptr = data;
02182   ctr = 0;
02183   for (;;) {
02184     elemptr = sample_data[ctr] + start_col;
02185 
02186     /* Even part */
02187 
02188     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[15]);
02189     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[14]);
02190     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[13]);
02191     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[12]);
02192     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[11]);
02193     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[10]);
02194     tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[9]);
02195     tmp7 = GETJSAMPLE(elemptr[7]) + GETJSAMPLE(elemptr[8]);
02196 
02197     tmp10 = tmp0 + tmp7;
02198     tmp14 = tmp0 - tmp7;
02199     tmp11 = tmp1 + tmp6;
02200     tmp15 = tmp1 - tmp6;
02201     tmp12 = tmp2 + tmp5;
02202     tmp16 = tmp2 - tmp5;
02203     tmp13 = tmp3 + tmp4;
02204     tmp17 = tmp3 - tmp4;
02205 
02206     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[15]);
02207     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[14]);
02208     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[13]);
02209     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[12]);
02210     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[11]);
02211     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[10]);
02212     tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[9]);
02213     tmp7 = GETJSAMPLE(elemptr[7]) - GETJSAMPLE(elemptr[8]);
02214 
02215     /* Apply unsigned->signed conversion */
02216     dataptr[0] = (DCTELEM)
02217       ((tmp10 + tmp11 + tmp12 + tmp13 - 16 * CENTERJSAMPLE) << PASS1_BITS);
02218     dataptr[4] = (DCTELEM)
02219       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
02220               MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
02221               CONST_BITS-PASS1_BITS);
02222 
02223     tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
02224             MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
02225 
02226     dataptr[2] = (DCTELEM)
02227       DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
02228               + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
02229               CONST_BITS-PASS1_BITS);
02230     dataptr[6] = (DCTELEM)
02231       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
02232               - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
02233               CONST_BITS-PASS1_BITS);
02234 
02235     /* Odd part */
02236 
02237     tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
02238             MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
02239     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
02240             MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
02241     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
02242             MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
02243     tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
02244             MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
02245     tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
02246             MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
02247     tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
02248             MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
02249     tmp10 = tmp11 + tmp12 + tmp13 -
02250             MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
02251             MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
02252     tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
02253              - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
02254     tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
02255              + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
02256     tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
02257              + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
02258 
02259     dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
02260     dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
02261     dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
02262     dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
02263 
02264     ctr++;
02265 
02266     if (ctr != DCTSIZE) {
02267       if (ctr == DCTSIZE * 2)
02268         break;                  /* Done. */
02269       dataptr += DCTSIZE;       /* advance pointer to next row */
02270     } else
02271       dataptr = workspace;      /* switch pointer to extended workspace */
02272   }
02273 
02274   /* Pass 2: process columns.
02275    * We remove the PASS1_BITS scaling, but leave the results scaled up
02276    * by an overall factor of 8.
02277    * We must also scale the output by (8/16)**2 = 1/2**2.
02278    */
02279 
02280   dataptr = data;
02281   wsptr = workspace;
02282   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02283     /* Even part */
02284 
02285     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*7];
02286     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*6];
02287     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*5];
02288     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*4];
02289     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*3];
02290     tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*2];
02291     tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*1];
02292     tmp7 = dataptr[DCTSIZE*7] + wsptr[DCTSIZE*0];
02293 
02294     tmp10 = tmp0 + tmp7;
02295     tmp14 = tmp0 - tmp7;
02296     tmp11 = tmp1 + tmp6;
02297     tmp15 = tmp1 - tmp6;
02298     tmp12 = tmp2 + tmp5;
02299     tmp16 = tmp2 - tmp5;
02300     tmp13 = tmp3 + tmp4;
02301     tmp17 = tmp3 - tmp4;
02302 
02303     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*7];
02304     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*6];
02305     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*5];
02306     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*4];
02307     tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*3];
02308     tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*2];
02309     tmp6 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*1];
02310     tmp7 = dataptr[DCTSIZE*7] - wsptr[DCTSIZE*0];
02311 
02312     dataptr[DCTSIZE*0] = (DCTELEM)
02313       DESCALE(tmp10 + tmp11 + tmp12 + tmp13, PASS1_BITS+2);
02314     dataptr[DCTSIZE*4] = (DCTELEM)
02315       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
02316               MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
02317               CONST_BITS+PASS1_BITS+2);
02318 
02319     tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
02320             MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
02321 
02322     dataptr[DCTSIZE*2] = (DCTELEM)
02323       DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
02324               + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+10 */
02325               CONST_BITS+PASS1_BITS+2);
02326     dataptr[DCTSIZE*6] = (DCTELEM)
02327       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
02328               - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
02329               CONST_BITS+PASS1_BITS+2);
02330 
02331     /* Odd part */
02332 
02333     tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
02334             MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
02335     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
02336             MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
02337     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
02338             MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
02339     tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
02340             MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
02341     tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
02342             MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
02343     tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
02344             MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
02345     tmp10 = tmp11 + tmp12 + tmp13 -
02346             MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
02347             MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
02348     tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
02349              - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
02350     tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
02351              + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
02352     tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
02353              + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
02354 
02355     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS+2);
02356     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS+2);
02357     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS+2);
02358     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS+2);
02359 
02360     dataptr++;                  /* advance pointer to next column */
02361     wsptr++;                    /* advance pointer to next column */
02362   }
02363 }
02364 
02365 
02366 /*
02367  * Perform the forward DCT on a 16x8 sample block.
02368  *
02369  * 16-point FDCT in pass 1 (rows), 8-point in pass 2 (columns).
02370  */
02371 
02372 GLOBAL(void)
02373 jpeg_fdct_16x8 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02374 {
02375   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
02376   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
02377   INT32 z1;
02378   DCTELEM *dataptr;
02379   JSAMPROW elemptr;
02380   int ctr;
02381   SHIFT_TEMPS
02382 
02383   /* Pass 1: process rows. */
02384   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
02385   /* furthermore, we scale the results by 2**PASS1_BITS. */
02386   /* 16-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/32). */
02387 
02388   dataptr = data;
02389   ctr = 0;
02390   for (ctr = 0; ctr < DCTSIZE; ctr++) {
02391     elemptr = sample_data[ctr] + start_col;
02392 
02393     /* Even part */
02394 
02395     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[15]);
02396     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[14]);
02397     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[13]);
02398     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[12]);
02399     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[11]);
02400     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[10]);
02401     tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[9]);
02402     tmp7 = GETJSAMPLE(elemptr[7]) + GETJSAMPLE(elemptr[8]);
02403 
02404     tmp10 = tmp0 + tmp7;
02405     tmp14 = tmp0 - tmp7;
02406     tmp11 = tmp1 + tmp6;
02407     tmp15 = tmp1 - tmp6;
02408     tmp12 = tmp2 + tmp5;
02409     tmp16 = tmp2 - tmp5;
02410     tmp13 = tmp3 + tmp4;
02411     tmp17 = tmp3 - tmp4;
02412 
02413     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[15]);
02414     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[14]);
02415     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[13]);
02416     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[12]);
02417     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[11]);
02418     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[10]);
02419     tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[9]);
02420     tmp7 = GETJSAMPLE(elemptr[7]) - GETJSAMPLE(elemptr[8]);
02421 
02422     /* Apply unsigned->signed conversion */
02423     dataptr[0] = (DCTELEM)
02424       ((tmp10 + tmp11 + tmp12 + tmp13 - 16 * CENTERJSAMPLE) << PASS1_BITS);
02425     dataptr[4] = (DCTELEM)
02426       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
02427               MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
02428               CONST_BITS-PASS1_BITS);
02429 
02430     tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
02431             MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
02432 
02433     dataptr[2] = (DCTELEM)
02434       DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
02435               + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
02436               CONST_BITS-PASS1_BITS);
02437     dataptr[6] = (DCTELEM)
02438       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
02439               - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
02440               CONST_BITS-PASS1_BITS);
02441 
02442     /* Odd part */
02443 
02444     tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
02445             MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
02446     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
02447             MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
02448     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
02449             MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
02450     tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
02451             MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
02452     tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
02453             MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
02454     tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
02455             MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
02456     tmp10 = tmp11 + tmp12 + tmp13 -
02457             MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
02458             MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
02459     tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
02460              - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
02461     tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
02462              + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
02463     tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
02464              + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
02465 
02466     dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
02467     dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
02468     dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
02469     dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
02470 
02471     dataptr += DCTSIZE;         /* advance pointer to next row */
02472   }
02473 
02474   /* Pass 2: process columns.
02475    * We remove the PASS1_BITS scaling, but leave the results scaled up
02476    * by an overall factor of 8.
02477    * We must also scale the output by 8/16 = 1/2.
02478    */
02479 
02480   dataptr = data;
02481   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02482     /* Even part per LL&M figure 1 --- note that published figure is faulty;
02483      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
02484      */
02485 
02486     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
02487     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
02488     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
02489     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
02490 
02491     tmp10 = tmp0 + tmp3;
02492     tmp12 = tmp0 - tmp3;
02493     tmp11 = tmp1 + tmp2;
02494     tmp13 = tmp1 - tmp2;
02495 
02496     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
02497     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
02498     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
02499     tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
02500 
02501     dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS+1);
02502     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS+1);
02503 
02504     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
02505     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, FIX_0_765366865),
02506                                            CONST_BITS+PASS1_BITS+1);
02507     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 - MULTIPLY(tmp13, FIX_1_847759065),
02508                                            CONST_BITS+PASS1_BITS+1);
02509 
02510     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
02511      * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
02512      * i0..i3 in the paper are tmp0..tmp3 here.
02513      */
02514 
02515     tmp10 = tmp0 + tmp3;
02516     tmp11 = tmp1 + tmp2;
02517     tmp12 = tmp0 + tmp2;
02518     tmp13 = tmp1 + tmp3;
02519     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
02520 
02521     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
02522     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
02523     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
02524     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
02525     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
02526     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
02527     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
02528     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
02529 
02530     tmp12 += z1;
02531     tmp13 += z1;
02532 
02533     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0 + tmp10 + tmp12,
02534                                            CONST_BITS+PASS1_BITS+1);
02535     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1 + tmp11 + tmp13,
02536                                            CONST_BITS+PASS1_BITS+1);
02537     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2 + tmp11 + tmp12,
02538                                            CONST_BITS+PASS1_BITS+1);
02539     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3 + tmp10 + tmp13,
02540                                            CONST_BITS+PASS1_BITS+1);
02541 
02542     dataptr++;                  /* advance pointer to next column */
02543   }
02544 }
02545 
02546 
02547 /*
02548  * Perform the forward DCT on a 14x7 sample block.
02549  *
02550  * 14-point FDCT in pass 1 (rows), 7-point in pass 2 (columns).
02551  */
02552 
02553 GLOBAL(void)
02554 jpeg_fdct_14x7 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02555 {
02556   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
02557   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
02558   INT32 z1, z2, z3;
02559   DCTELEM *dataptr;
02560   JSAMPROW elemptr;
02561   int ctr;
02562   SHIFT_TEMPS
02563 
02564   /* Zero bottom row of output coefficient block. */
02565   MEMZERO(&data[DCTSIZE*7], SIZEOF(DCTELEM) * DCTSIZE);
02566 
02567   /* Pass 1: process rows. */
02568   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
02569   /* furthermore, we scale the results by 2**PASS1_BITS. */
02570   /* 14-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/28). */
02571 
02572   dataptr = data;
02573   for (ctr = 0; ctr < 7; ctr++) {
02574     elemptr = sample_data[ctr] + start_col;
02575 
02576     /* Even part */
02577 
02578     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[13]);
02579     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[12]);
02580     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[11]);
02581     tmp13 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[10]);
02582     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[9]);
02583     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[8]);
02584     tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[7]);
02585 
02586     tmp10 = tmp0 + tmp6;
02587     tmp14 = tmp0 - tmp6;
02588     tmp11 = tmp1 + tmp5;
02589     tmp15 = tmp1 - tmp5;
02590     tmp12 = tmp2 + tmp4;
02591     tmp16 = tmp2 - tmp4;
02592 
02593     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[13]);
02594     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[12]);
02595     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[11]);
02596     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[10]);
02597     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[9]);
02598     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[8]);
02599     tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[7]);
02600 
02601     /* Apply unsigned->signed conversion */
02602     dataptr[0] = (DCTELEM)
02603       ((tmp10 + tmp11 + tmp12 + tmp13 - 14 * CENTERJSAMPLE) << PASS1_BITS);
02604     tmp13 += tmp13;
02605     dataptr[4] = (DCTELEM)
02606       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.274162392)) + /* c4 */
02607               MULTIPLY(tmp11 - tmp13, FIX(0.314692123)) - /* c12 */
02608               MULTIPLY(tmp12 - tmp13, FIX(0.881747734)),  /* c8 */
02609               CONST_BITS-PASS1_BITS);
02610 
02611     tmp10 = MULTIPLY(tmp14 + tmp15, FIX(1.105676686));    /* c6 */
02612 
02613     dataptr[2] = (DCTELEM)
02614       DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.273079590))   /* c2-c6 */
02615               + MULTIPLY(tmp16, FIX(0.613604268)),        /* c10 */
02616               CONST_BITS-PASS1_BITS);
02617     dataptr[6] = (DCTELEM)
02618       DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.719280954))   /* c6+c10 */
02619               - MULTIPLY(tmp16, FIX(1.378756276)),        /* c2 */
02620               CONST_BITS-PASS1_BITS);
02621 
02622     /* Odd part */
02623 
02624     tmp10 = tmp1 + tmp2;
02625     tmp11 = tmp5 - tmp4;
02626     dataptr[7] = (DCTELEM) ((tmp0 - tmp10 + tmp3 - tmp11 - tmp6) << PASS1_BITS);
02627     tmp3 <<= CONST_BITS;
02628     tmp10 = MULTIPLY(tmp10, - FIX(0.158341681));          /* -c13 */
02629     tmp11 = MULTIPLY(tmp11, FIX(1.405321284));            /* c1 */
02630     tmp10 += tmp11 - tmp3;
02631     tmp11 = MULTIPLY(tmp0 + tmp2, FIX(1.197448846)) +     /* c5 */
02632             MULTIPLY(tmp4 + tmp6, FIX(0.752406978));      /* c9 */
02633     dataptr[5] = (DCTELEM)
02634       DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(2.373959773)) /* c3+c5-c13 */
02635               + MULTIPLY(tmp4, FIX(1.119999435)),         /* c1+c11-c9 */
02636               CONST_BITS-PASS1_BITS);
02637     tmp12 = MULTIPLY(tmp0 + tmp1, FIX(1.334852607)) +     /* c3 */
02638             MULTIPLY(tmp5 - tmp6, FIX(0.467085129));      /* c11 */
02639     dataptr[3] = (DCTELEM)
02640       DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.424103948)) /* c3-c9-c13 */
02641               - MULTIPLY(tmp5, FIX(3.069855259)),         /* c1+c5+c11 */
02642               CONST_BITS-PASS1_BITS);
02643     dataptr[1] = (DCTELEM)
02644       DESCALE(tmp11 + tmp12 + tmp3 + tmp6 -
02645               MULTIPLY(tmp0 + tmp6, FIX(1.126980169)),    /* c3+c5-c1 */
02646               CONST_BITS-PASS1_BITS);
02647 
02648     dataptr += DCTSIZE;         /* advance pointer to next row */
02649   }
02650 
02651   /* Pass 2: process columns.
02652    * We remove the PASS1_BITS scaling, but leave the results scaled up
02653    * by an overall factor of 8.
02654    * We must also scale the output by (8/14)*(8/7) = 32/49, which we
02655    * partially fold into the constant multipliers and final shifting:
02656    * 7-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/14) * 64/49.
02657    */
02658 
02659   dataptr = data;
02660   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02661     /* Even part */
02662 
02663     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*6];
02664     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*5];
02665     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*4];
02666     tmp3 = dataptr[DCTSIZE*3];
02667 
02668     tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*6];
02669     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*5];
02670     tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*4];
02671 
02672     z1 = tmp0 + tmp2;
02673     dataptr[DCTSIZE*0] = (DCTELEM)
02674       DESCALE(MULTIPLY(z1 + tmp1 + tmp3, FIX(1.306122449)), /* 64/49 */
02675               CONST_BITS+PASS1_BITS+1);
02676     tmp3 += tmp3;
02677     z1 -= tmp3;
02678     z1 -= tmp3;
02679     z1 = MULTIPLY(z1, FIX(0.461784020));                /* (c2+c6-c4)/2 */
02680     z2 = MULTIPLY(tmp0 - tmp2, FIX(1.202428084));       /* (c2+c4-c6)/2 */
02681     z3 = MULTIPLY(tmp1 - tmp2, FIX(0.411026446));       /* c6 */
02682     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS+PASS1_BITS+1);
02683     z1 -= z2;
02684     z2 = MULTIPLY(tmp0 - tmp1, FIX(1.151670509));       /* c4 */
02685     dataptr[DCTSIZE*4] = (DCTELEM)
02686       DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.923568041)), /* c2+c6-c4 */
02687               CONST_BITS+PASS1_BITS+1);
02688     dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+PASS1_BITS+1);
02689 
02690     /* Odd part */
02691 
02692     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.221765677));   /* (c3+c1-c5)/2 */
02693     tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.222383464));   /* (c3+c5-c1)/2 */
02694     tmp0 = tmp1 - tmp2;
02695     tmp1 += tmp2;
02696     tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.800824523)); /* -c1 */
02697     tmp1 += tmp2;
02698     tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.801442310));   /* c5 */
02699     tmp0 += tmp3;
02700     tmp2 += tmp3 + MULTIPLY(tmp12, FIX(2.443531355));   /* c3+c1-c5 */
02701 
02702     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+PASS1_BITS+1);
02703     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+PASS1_BITS+1);
02704     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+PASS1_BITS+1);
02705 
02706     dataptr++;                  /* advance pointer to next column */
02707   }
02708 }
02709 
02710 
02711 /*
02712  * Perform the forward DCT on a 12x6 sample block.
02713  *
02714  * 12-point FDCT in pass 1 (rows), 6-point in pass 2 (columns).
02715  */
02716 
02717 GLOBAL(void)
02718 jpeg_fdct_12x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02719 {
02720   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
02721   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
02722   DCTELEM *dataptr;
02723   JSAMPROW elemptr;
02724   int ctr;
02725   SHIFT_TEMPS
02726 
02727   /* Zero 2 bottom rows of output coefficient block. */
02728   MEMZERO(&data[DCTSIZE*6], SIZEOF(DCTELEM) * DCTSIZE * 2);
02729 
02730   /* Pass 1: process rows. */
02731   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
02732   /* furthermore, we scale the results by 2**PASS1_BITS. */
02733   /* 12-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/24). */
02734 
02735   dataptr = data;
02736   for (ctr = 0; ctr < 6; ctr++) {
02737     elemptr = sample_data[ctr] + start_col;
02738 
02739     /* Even part */
02740 
02741     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[11]);
02742     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[10]);
02743     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[9]);
02744     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[8]);
02745     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[7]);
02746     tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[6]);
02747 
02748     tmp10 = tmp0 + tmp5;
02749     tmp13 = tmp0 - tmp5;
02750     tmp11 = tmp1 + tmp4;
02751     tmp14 = tmp1 - tmp4;
02752     tmp12 = tmp2 + tmp3;
02753     tmp15 = tmp2 - tmp3;
02754 
02755     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[11]);
02756     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[10]);
02757     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[9]);
02758     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[8]);
02759     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[7]);
02760     tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[6]);
02761 
02762     /* Apply unsigned->signed conversion */
02763     dataptr[0] = (DCTELEM)
02764       ((tmp10 + tmp11 + tmp12 - 12 * CENTERJSAMPLE) << PASS1_BITS);
02765     dataptr[6] = (DCTELEM) ((tmp13 - tmp14 - tmp15) << PASS1_BITS);
02766     dataptr[4] = (DCTELEM)
02767       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.224744871)), /* c4 */
02768               CONST_BITS-PASS1_BITS);
02769     dataptr[2] = (DCTELEM)
02770       DESCALE(tmp14 - tmp15 + MULTIPLY(tmp13 + tmp15, FIX(1.366025404)), /* c2 */
02771               CONST_BITS-PASS1_BITS);
02772 
02773     /* Odd part */
02774 
02775     tmp10 = MULTIPLY(tmp1 + tmp4, FIX_0_541196100);    /* c9 */
02776     tmp14 = tmp10 + MULTIPLY(tmp1, FIX_0_765366865);   /* c3-c9 */
02777     tmp15 = tmp10 - MULTIPLY(tmp4, FIX_1_847759065);   /* c3+c9 */
02778     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.121971054));   /* c5 */
02779     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.860918669));   /* c7 */
02780     tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.580774953)) /* c5+c7-c1 */
02781             + MULTIPLY(tmp5, FIX(0.184591911));        /* c11 */
02782     tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.184591911)); /* -c11 */
02783     tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.339493912)) /* c1+c5-c11 */
02784             + MULTIPLY(tmp5, FIX(0.860918669));        /* c7 */
02785     tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.725788011)) /* c1+c11-c7 */
02786             - MULTIPLY(tmp5, FIX(1.121971054));        /* c5 */
02787     tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.306562965)) /* c3 */
02788             - MULTIPLY(tmp2 + tmp5, FIX_0_541196100);  /* c9 */
02789 
02790     dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
02791     dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
02792     dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
02793     dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
02794 
02795     dataptr += DCTSIZE;         /* advance pointer to next row */
02796   }
02797 
02798   /* Pass 2: process columns.
02799    * We remove the PASS1_BITS scaling, but leave the results scaled up
02800    * by an overall factor of 8.
02801    * We must also scale the output by (8/12)*(8/6) = 8/9, which we
02802    * partially fold into the constant multipliers and final shifting:
02803    * 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12) * 16/9.
02804    */
02805 
02806   dataptr = data;
02807   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02808     /* Even part */
02809 
02810     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
02811     tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
02812     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
02813 
02814     tmp10 = tmp0 + tmp2;
02815     tmp12 = tmp0 - tmp2;
02816 
02817     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
02818     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
02819     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
02820 
02821     dataptr[DCTSIZE*0] = (DCTELEM)
02822       DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
02823               CONST_BITS+PASS1_BITS+1);
02824     dataptr[DCTSIZE*2] = (DCTELEM)
02825       DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
02826               CONST_BITS+PASS1_BITS+1);
02827     dataptr[DCTSIZE*4] = (DCTELEM)
02828       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
02829               CONST_BITS+PASS1_BITS+1);
02830 
02831     /* Odd part */
02832 
02833     tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
02834 
02835     dataptr[DCTSIZE*1] = (DCTELEM)
02836       DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
02837               CONST_BITS+PASS1_BITS+1);
02838     dataptr[DCTSIZE*3] = (DCTELEM)
02839       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
02840               CONST_BITS+PASS1_BITS+1);
02841     dataptr[DCTSIZE*5] = (DCTELEM)
02842       DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
02843               CONST_BITS+PASS1_BITS+1);
02844 
02845     dataptr++;                  /* advance pointer to next column */
02846   }
02847 }
02848 
02849 
02850 /*
02851  * Perform the forward DCT on a 10x5 sample block.
02852  *
02853  * 10-point FDCT in pass 1 (rows), 5-point in pass 2 (columns).
02854  */
02855 
02856 GLOBAL(void)
02857 jpeg_fdct_10x5 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02858 {
02859   INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
02860   INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
02861   DCTELEM *dataptr;
02862   JSAMPROW elemptr;
02863   int ctr;
02864   SHIFT_TEMPS
02865 
02866   /* Zero 3 bottom rows of output coefficient block. */
02867   MEMZERO(&data[DCTSIZE*5], SIZEOF(DCTELEM) * DCTSIZE * 3);
02868 
02869   /* Pass 1: process rows. */
02870   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
02871   /* furthermore, we scale the results by 2**PASS1_BITS. */
02872   /* 10-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/20). */
02873 
02874   dataptr = data;
02875   for (ctr = 0; ctr < 5; ctr++) {
02876     elemptr = sample_data[ctr] + start_col;
02877 
02878     /* Even part */
02879 
02880     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[9]);
02881     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[8]);
02882     tmp12 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[7]);
02883     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[6]);
02884     tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[5]);
02885 
02886     tmp10 = tmp0 + tmp4;
02887     tmp13 = tmp0 - tmp4;
02888     tmp11 = tmp1 + tmp3;
02889     tmp14 = tmp1 - tmp3;
02890 
02891     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[9]);
02892     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[8]);
02893     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[7]);
02894     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[6]);
02895     tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[5]);
02896 
02897     /* Apply unsigned->signed conversion */
02898     dataptr[0] = (DCTELEM)
02899       ((tmp10 + tmp11 + tmp12 - 10 * CENTERJSAMPLE) << PASS1_BITS);
02900     tmp12 += tmp12;
02901     dataptr[4] = (DCTELEM)
02902       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.144122806)) - /* c4 */
02903               MULTIPLY(tmp11 - tmp12, FIX(0.437016024)),  /* c8 */
02904               CONST_BITS-PASS1_BITS);
02905     tmp10 = MULTIPLY(tmp13 + tmp14, FIX(0.831253876));    /* c6 */
02906     dataptr[2] = (DCTELEM)
02907       DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.513743148)),  /* c2-c6 */
02908               CONST_BITS-PASS1_BITS);
02909     dataptr[6] = (DCTELEM)
02910       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.176250899)),  /* c2+c6 */
02911               CONST_BITS-PASS1_BITS);
02912 
02913     /* Odd part */
02914 
02915     tmp10 = tmp0 + tmp4;
02916     tmp11 = tmp1 - tmp3;
02917     dataptr[5] = (DCTELEM) ((tmp10 - tmp11 - tmp2) << PASS1_BITS);
02918     tmp2 <<= CONST_BITS;
02919     dataptr[1] = (DCTELEM)
02920       DESCALE(MULTIPLY(tmp0, FIX(1.396802247)) +          /* c1 */
02921               MULTIPLY(tmp1, FIX(1.260073511)) + tmp2 +   /* c3 */
02922               MULTIPLY(tmp3, FIX(0.642039522)) +          /* c7 */
02923               MULTIPLY(tmp4, FIX(0.221231742)),           /* c9 */
02924               CONST_BITS-PASS1_BITS);
02925     tmp12 = MULTIPLY(tmp0 - tmp4, FIX(0.951056516)) -     /* (c3+c7)/2 */
02926             MULTIPLY(tmp1 + tmp3, FIX(0.587785252));      /* (c1-c9)/2 */
02927     tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.309016994)) +   /* (c3-c7)/2 */
02928             (tmp11 << (CONST_BITS - 1)) - tmp2;
02929     dataptr[3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS-PASS1_BITS);
02930     dataptr[7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS-PASS1_BITS);
02931 
02932     dataptr += DCTSIZE;         /* advance pointer to next row */
02933   }
02934 
02935   /* Pass 2: process columns.
02936    * We remove the PASS1_BITS scaling, but leave the results scaled up
02937    * by an overall factor of 8.
02938    * We must also scale the output by (8/10)*(8/5) = 32/25, which we
02939    * fold into the constant multipliers:
02940    * 5-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/10) * 32/25.
02941    */
02942 
02943   dataptr = data;
02944   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
02945     /* Even part */
02946 
02947     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*4];
02948     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*3];
02949     tmp2 = dataptr[DCTSIZE*2];
02950 
02951     tmp10 = tmp0 + tmp1;
02952     tmp11 = tmp0 - tmp1;
02953 
02954     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*4];
02955     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*3];
02956 
02957     dataptr[DCTSIZE*0] = (DCTELEM)
02958       DESCALE(MULTIPLY(tmp10 + tmp2, FIX(1.28)),        /* 32/25 */
02959               CONST_BITS+PASS1_BITS);
02960     tmp11 = MULTIPLY(tmp11, FIX(1.011928851));          /* (c2+c4)/2 */
02961     tmp10 -= tmp2 << 2;
02962     tmp10 = MULTIPLY(tmp10, FIX(0.452548340));          /* (c2-c4)/2 */
02963     dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS+PASS1_BITS);
02964     dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS+PASS1_BITS);
02965 
02966     /* Odd part */
02967 
02968     tmp10 = MULTIPLY(tmp0 + tmp1, FIX(1.064004961));    /* c3 */
02969 
02970     dataptr[DCTSIZE*1] = (DCTELEM)
02971       DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.657591230)), /* c1-c3 */
02972               CONST_BITS+PASS1_BITS);
02973     dataptr[DCTSIZE*3] = (DCTELEM)
02974       DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.785601151)), /* c1+c3 */
02975               CONST_BITS+PASS1_BITS);
02976 
02977     dataptr++;                  /* advance pointer to next column */
02978   }
02979 }
02980 
02981 
02982 /*
02983  * Perform the forward DCT on an 8x4 sample block.
02984  *
02985  * 8-point FDCT in pass 1 (rows), 4-point in pass 2 (columns).
02986  */
02987 
02988 GLOBAL(void)
02989 jpeg_fdct_8x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
02990 {
02991   INT32 tmp0, tmp1, tmp2, tmp3;
02992   INT32 tmp10, tmp11, tmp12, tmp13;
02993   INT32 z1;
02994   DCTELEM *dataptr;
02995   JSAMPROW elemptr;
02996   int ctr;
02997   SHIFT_TEMPS
02998 
02999   /* Zero 4 bottom rows of output coefficient block. */
03000   MEMZERO(&data[DCTSIZE*4], SIZEOF(DCTELEM) * DCTSIZE * 4);
03001 
03002   /* Pass 1: process rows. */
03003   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03004   /* furthermore, we scale the results by 2**PASS1_BITS. */
03005   /* We must also scale the output by 8/4 = 2, which we add here. */
03006 
03007   dataptr = data;
03008   for (ctr = 0; ctr < 4; ctr++) {
03009     elemptr = sample_data[ctr] + start_col;
03010 
03011     /* Even part per LL&M figure 1 --- note that published figure is faulty;
03012      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
03013      */
03014 
03015     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
03016     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
03017     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
03018     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
03019 
03020     tmp10 = tmp0 + tmp3;
03021     tmp12 = tmp0 - tmp3;
03022     tmp11 = tmp1 + tmp2;
03023     tmp13 = tmp1 - tmp2;
03024 
03025     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
03026     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
03027     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
03028     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
03029 
03030     /* Apply unsigned->signed conversion */
03031     dataptr[0] = (DCTELEM)
03032       ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << (PASS1_BITS+1));
03033     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << (PASS1_BITS+1));
03034 
03035     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
03036     /* Add fudge factor here for final descale. */
03037     z1 += ONE << (CONST_BITS-PASS1_BITS-2);
03038     dataptr[2] = (DCTELEM) RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865),
03039                                        CONST_BITS-PASS1_BITS-1);
03040     dataptr[6] = (DCTELEM) RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065),
03041                                        CONST_BITS-PASS1_BITS-1);
03042 
03043     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
03044      * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
03045      * i0..i3 in the paper are tmp0..tmp3 here.
03046      */
03047 
03048     tmp10 = tmp0 + tmp3;
03049     tmp11 = tmp1 + tmp2;
03050     tmp12 = tmp0 + tmp2;
03051     tmp13 = tmp1 + tmp3;
03052     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
03053     /* Add fudge factor here for final descale. */
03054     z1 += ONE << (CONST_BITS-PASS1_BITS-2);
03055 
03056     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
03057     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
03058     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
03059     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
03060     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
03061     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
03062     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
03063     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
03064 
03065     tmp12 += z1;
03066     tmp13 += z1;
03067 
03068     dataptr[1] = (DCTELEM)
03069       RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS-1);
03070     dataptr[3] = (DCTELEM)
03071       RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS-1);
03072     dataptr[5] = (DCTELEM)
03073       RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS-1);
03074     dataptr[7] = (DCTELEM)
03075       RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS-1);
03076 
03077     dataptr += DCTSIZE;         /* advance pointer to next row */
03078   }
03079 
03080   /* Pass 2: process columns.
03081    * We remove the PASS1_BITS scaling, but leave the results scaled up
03082    * by an overall factor of 8.
03083    * 4-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
03084    */
03085 
03086   dataptr = data;
03087   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
03088     /* Even part */
03089 
03090     /* Add fudge factor here for final descale. */
03091     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3] + (ONE << (PASS1_BITS-1));
03092     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
03093 
03094     tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
03095     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
03096 
03097     dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
03098     dataptr[DCTSIZE*2] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
03099 
03100     /* Odd part */
03101 
03102     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);   /* c6 */
03103     /* Add fudge factor here for final descale. */
03104     tmp0 += ONE << (CONST_BITS+PASS1_BITS-1);
03105 
03106     dataptr[DCTSIZE*1] = (DCTELEM)
03107       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
03108                   CONST_BITS+PASS1_BITS);
03109     dataptr[DCTSIZE*3] = (DCTELEM)
03110       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
03111                   CONST_BITS+PASS1_BITS);
03112 
03113     dataptr++;                  /* advance pointer to next column */
03114   }
03115 }
03116 
03117 
03118 /*
03119  * Perform the forward DCT on a 6x3 sample block.
03120  *
03121  * 6-point FDCT in pass 1 (rows), 3-point in pass 2 (columns).
03122  */
03123 
03124 GLOBAL(void)
03125 jpeg_fdct_6x3 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03126 {
03127   INT32 tmp0, tmp1, tmp2;
03128   INT32 tmp10, tmp11, tmp12;
03129   DCTELEM *dataptr;
03130   JSAMPROW elemptr;
03131   int ctr;
03132   SHIFT_TEMPS
03133 
03134   /* Pre-zero output coefficient block. */
03135   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03136 
03137   /* Pass 1: process rows. */
03138   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03139   /* furthermore, we scale the results by 2**PASS1_BITS. */
03140   /* We scale the results further by 2 as part of output adaption */
03141   /* scaling for different DCT size. */
03142   /* 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12). */
03143 
03144   dataptr = data;
03145   for (ctr = 0; ctr < 3; ctr++) {
03146     elemptr = sample_data[ctr] + start_col;
03147 
03148     /* Even part */
03149 
03150     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
03151     tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
03152     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
03153 
03154     tmp10 = tmp0 + tmp2;
03155     tmp12 = tmp0 - tmp2;
03156 
03157     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
03158     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
03159     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
03160 
03161     /* Apply unsigned->signed conversion */
03162     dataptr[0] = (DCTELEM)
03163       ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << (PASS1_BITS+1));
03164     dataptr[2] = (DCTELEM)
03165       DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
03166               CONST_BITS-PASS1_BITS-1);
03167     dataptr[4] = (DCTELEM)
03168       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
03169               CONST_BITS-PASS1_BITS-1);
03170 
03171     /* Odd part */
03172 
03173     tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
03174                     CONST_BITS-PASS1_BITS-1);
03175 
03176     dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << (PASS1_BITS+1)));
03177     dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << (PASS1_BITS+1));
03178     dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << (PASS1_BITS+1)));
03179 
03180     dataptr += DCTSIZE;         /* advance pointer to next row */
03181   }
03182 
03183   /* Pass 2: process columns.
03184    * We remove the PASS1_BITS scaling, but leave the results scaled up
03185    * by an overall factor of 8.
03186    * We must also scale the output by (8/6)*(8/3) = 32/9, which we partially
03187    * fold into the constant multipliers (other part was done in pass 1):
03188    * 3-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/6) * 16/9.
03189    */
03190 
03191   dataptr = data;
03192   for (ctr = 0; ctr < 6; ctr++) {
03193     /* Even part */
03194 
03195     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*2];
03196     tmp1 = dataptr[DCTSIZE*1];
03197 
03198     tmp2 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*2];
03199 
03200     dataptr[DCTSIZE*0] = (DCTELEM)
03201       DESCALE(MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),        /* 16/9 */
03202               CONST_BITS+PASS1_BITS);
03203     dataptr[DCTSIZE*2] = (DCTELEM)
03204       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(1.257078722)), /* c2 */
03205               CONST_BITS+PASS1_BITS);
03206 
03207     /* Odd part */
03208 
03209     dataptr[DCTSIZE*1] = (DCTELEM)
03210       DESCALE(MULTIPLY(tmp2, FIX(2.177324216)),               /* c1 */
03211               CONST_BITS+PASS1_BITS);
03212 
03213     dataptr++;                  /* advance pointer to next column */
03214   }
03215 }
03216 
03217 
03218 /*
03219  * Perform the forward DCT on a 4x2 sample block.
03220  *
03221  * 4-point FDCT in pass 1 (rows), 2-point in pass 2 (columns).
03222  */
03223 
03224 GLOBAL(void)
03225 jpeg_fdct_4x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03226 {
03227   INT32 tmp0, tmp1;
03228   INT32 tmp10, tmp11;
03229   DCTELEM *dataptr;
03230   JSAMPROW elemptr;
03231   int ctr;
03232   SHIFT_TEMPS
03233 
03234   /* Pre-zero output coefficient block. */
03235   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03236 
03237   /* Pass 1: process rows. */
03238   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03239   /* furthermore, we scale the results by 2**PASS1_BITS. */
03240   /* We must also scale the output by (8/4)*(8/2) = 2**3, which we add here. */
03241   /* 4-point FDCT kernel, */
03242   /* cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT]. */
03243 
03244   dataptr = data;
03245   for (ctr = 0; ctr < 2; ctr++) {
03246     elemptr = sample_data[ctr] + start_col;
03247 
03248     /* Even part */
03249 
03250     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
03251     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
03252 
03253     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
03254     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
03255 
03256     /* Apply unsigned->signed conversion */
03257     dataptr[0] = (DCTELEM)
03258       ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+3));
03259     dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+3));
03260 
03261     /* Odd part */
03262 
03263     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
03264     /* Add fudge factor here for final descale. */
03265     tmp0 += ONE << (CONST_BITS-PASS1_BITS-4);
03266 
03267     dataptr[1] = (DCTELEM)
03268       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
03269                   CONST_BITS-PASS1_BITS-3);
03270     dataptr[3] = (DCTELEM)
03271       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
03272                   CONST_BITS-PASS1_BITS-3);
03273 
03274     dataptr += DCTSIZE;         /* advance pointer to next row */
03275   }
03276 
03277   /* Pass 2: process columns.
03278    * We remove the PASS1_BITS scaling, but leave the results scaled up
03279    * by an overall factor of 8.
03280    */
03281 
03282   dataptr = data;
03283   for (ctr = 0; ctr < 4; ctr++) {
03284     /* Even part */
03285 
03286     /* Add fudge factor here for final descale. */
03287     tmp0 = dataptr[DCTSIZE*0] + (ONE << (PASS1_BITS-1));
03288     tmp1 = dataptr[DCTSIZE*1];
03289 
03290     dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
03291 
03292     /* Odd part */
03293 
03294     dataptr[DCTSIZE*1] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
03295 
03296     dataptr++;                  /* advance pointer to next column */
03297   }
03298 }
03299 
03300 
03301 /*
03302  * Perform the forward DCT on a 2x1 sample block.
03303  *
03304  * 2-point FDCT in pass 1 (rows), 1-point in pass 2 (columns).
03305  */
03306 
03307 GLOBAL(void)
03308 jpeg_fdct_2x1 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03309 {
03310   INT32 tmp0, tmp1;
03311   JSAMPROW elemptr;
03312 
03313   /* Pre-zero output coefficient block. */
03314   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03315 
03316   elemptr = sample_data[0] + start_col;
03317 
03318   tmp0 = GETJSAMPLE(elemptr[0]);
03319   tmp1 = GETJSAMPLE(elemptr[1]);
03320 
03321   /* We leave the results scaled up by an overall factor of 8.
03322    * We must also scale the output by (8/2)*(8/1) = 2**5.
03323    */
03324 
03325   /* Even part */
03326   /* Apply unsigned->signed conversion */
03327   data[0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 5);
03328 
03329   /* Odd part */
03330   data[1] = (DCTELEM) ((tmp0 - tmp1) << 5);
03331 }
03332 
03333 
03334 /*
03335  * Perform the forward DCT on an 8x16 sample block.
03336  *
03337  * 8-point FDCT in pass 1 (rows), 16-point in pass 2 (columns).
03338  */
03339 
03340 GLOBAL(void)
03341 jpeg_fdct_8x16 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03342 {
03343   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
03344   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
03345   INT32 z1;
03346   DCTELEM workspace[DCTSIZE2];
03347   DCTELEM *dataptr;
03348   DCTELEM *wsptr;
03349   JSAMPROW elemptr;
03350   int ctr;
03351   SHIFT_TEMPS
03352 
03353   /* Pass 1: process rows. */
03354   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03355   /* furthermore, we scale the results by 2**PASS1_BITS. */
03356 
03357   dataptr = data;
03358   ctr = 0;
03359   for (;;) {
03360     elemptr = sample_data[ctr] + start_col;
03361 
03362     /* Even part per LL&M figure 1 --- note that published figure is faulty;
03363      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
03364      */
03365 
03366     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
03367     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
03368     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
03369     tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
03370 
03371     tmp10 = tmp0 + tmp3;
03372     tmp12 = tmp0 - tmp3;
03373     tmp11 = tmp1 + tmp2;
03374     tmp13 = tmp1 - tmp2;
03375 
03376     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
03377     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
03378     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
03379     tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
03380 
03381     /* Apply unsigned->signed conversion */
03382     dataptr[0] = (DCTELEM) ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << PASS1_BITS);
03383     dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
03384 
03385     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
03386     dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, FIX_0_765366865),
03387                                    CONST_BITS-PASS1_BITS);
03388     dataptr[6] = (DCTELEM) DESCALE(z1 - MULTIPLY(tmp13, FIX_1_847759065),
03389                                    CONST_BITS-PASS1_BITS);
03390 
03391     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
03392      * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
03393      * i0..i3 in the paper are tmp0..tmp3 here.
03394      */
03395 
03396     tmp10 = tmp0 + tmp3;
03397     tmp11 = tmp1 + tmp2;
03398     tmp12 = tmp0 + tmp2;
03399     tmp13 = tmp1 + tmp3;
03400     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
03401 
03402     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
03403     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
03404     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
03405     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
03406     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
03407     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
03408     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
03409     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
03410 
03411     tmp12 += z1;
03412     tmp13 += z1;
03413 
03414     dataptr[1] = (DCTELEM) DESCALE(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS);
03415     dataptr[3] = (DCTELEM) DESCALE(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS);
03416     dataptr[5] = (DCTELEM) DESCALE(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS);
03417     dataptr[7] = (DCTELEM) DESCALE(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS);
03418 
03419     ctr++;
03420 
03421     if (ctr != DCTSIZE) {
03422       if (ctr == DCTSIZE * 2)
03423         break;                  /* Done. */
03424       dataptr += DCTSIZE;       /* advance pointer to next row */
03425     } else
03426       dataptr = workspace;      /* switch pointer to extended workspace */
03427   }
03428 
03429   /* Pass 2: process columns.
03430    * We remove the PASS1_BITS scaling, but leave the results scaled up
03431    * by an overall factor of 8.
03432    * We must also scale the output by 8/16 = 1/2.
03433    * 16-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/32).
03434    */
03435 
03436   dataptr = data;
03437   wsptr = workspace;
03438   for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
03439     /* Even part */
03440 
03441     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*7];
03442     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*6];
03443     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*5];
03444     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*4];
03445     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*3];
03446     tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*2];
03447     tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*1];
03448     tmp7 = dataptr[DCTSIZE*7] + wsptr[DCTSIZE*0];
03449 
03450     tmp10 = tmp0 + tmp7;
03451     tmp14 = tmp0 - tmp7;
03452     tmp11 = tmp1 + tmp6;
03453     tmp15 = tmp1 - tmp6;
03454     tmp12 = tmp2 + tmp5;
03455     tmp16 = tmp2 - tmp5;
03456     tmp13 = tmp3 + tmp4;
03457     tmp17 = tmp3 - tmp4;
03458 
03459     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*7];
03460     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*6];
03461     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*5];
03462     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*4];
03463     tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*3];
03464     tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*2];
03465     tmp6 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*1];
03466     tmp7 = dataptr[DCTSIZE*7] - wsptr[DCTSIZE*0];
03467 
03468     dataptr[DCTSIZE*0] = (DCTELEM)
03469       DESCALE(tmp10 + tmp11 + tmp12 + tmp13, PASS1_BITS+1);
03470     dataptr[DCTSIZE*4] = (DCTELEM)
03471       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
03472               MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
03473               CONST_BITS+PASS1_BITS+1);
03474 
03475     tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
03476             MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
03477 
03478     dataptr[DCTSIZE*2] = (DCTELEM)
03479       DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
03480               + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
03481               CONST_BITS+PASS1_BITS+1);
03482     dataptr[DCTSIZE*6] = (DCTELEM)
03483       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
03484               - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
03485               CONST_BITS+PASS1_BITS+1);
03486 
03487     /* Odd part */
03488 
03489     tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
03490             MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
03491     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
03492             MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
03493     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
03494             MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
03495     tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
03496             MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
03497     tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
03498             MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
03499     tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
03500             MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
03501     tmp10 = tmp11 + tmp12 + tmp13 -
03502             MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
03503             MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
03504     tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
03505              - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
03506     tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
03507              + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
03508     tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
03509              + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
03510 
03511     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS+1);
03512     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS+1);
03513     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS+1);
03514     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS+1);
03515 
03516     dataptr++;                  /* advance pointer to next column */
03517     wsptr++;                    /* advance pointer to next column */
03518   }
03519 }
03520 
03521 
03522 /*
03523  * Perform the forward DCT on a 7x14 sample block.
03524  *
03525  * 7-point FDCT in pass 1 (rows), 14-point in pass 2 (columns).
03526  */
03527 
03528 GLOBAL(void)
03529 jpeg_fdct_7x14 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03530 {
03531   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
03532   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
03533   INT32 z1, z2, z3;
03534   DCTELEM workspace[8*6];
03535   DCTELEM *dataptr;
03536   DCTELEM *wsptr;
03537   JSAMPROW elemptr;
03538   int ctr;
03539   SHIFT_TEMPS
03540 
03541   /* Pre-zero output coefficient block. */
03542   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03543 
03544   /* Pass 1: process rows. */
03545   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03546   /* furthermore, we scale the results by 2**PASS1_BITS. */
03547   /* 7-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/14). */
03548 
03549   dataptr = data;
03550   ctr = 0;
03551   for (;;) {
03552     elemptr = sample_data[ctr] + start_col;
03553 
03554     /* Even part */
03555 
03556     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[6]);
03557     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[5]);
03558     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[4]);
03559     tmp3 = GETJSAMPLE(elemptr[3]);
03560 
03561     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[6]);
03562     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[5]);
03563     tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[4]);
03564 
03565     z1 = tmp0 + tmp2;
03566     /* Apply unsigned->signed conversion */
03567     dataptr[0] = (DCTELEM)
03568       ((z1 + tmp1 + tmp3 - 7 * CENTERJSAMPLE) << PASS1_BITS);
03569     tmp3 += tmp3;
03570     z1 -= tmp3;
03571     z1 -= tmp3;
03572     z1 = MULTIPLY(z1, FIX(0.353553391));                /* (c2+c6-c4)/2 */
03573     z2 = MULTIPLY(tmp0 - tmp2, FIX(0.920609002));       /* (c2+c4-c6)/2 */
03574     z3 = MULTIPLY(tmp1 - tmp2, FIX(0.314692123));       /* c6 */
03575     dataptr[2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS-PASS1_BITS);
03576     z1 -= z2;
03577     z2 = MULTIPLY(tmp0 - tmp1, FIX(0.881747734));       /* c4 */
03578     dataptr[4] = (DCTELEM)
03579       DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.707106781)), /* c2+c6-c4 */
03580               CONST_BITS-PASS1_BITS);
03581     dataptr[6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS-PASS1_BITS);
03582 
03583     /* Odd part */
03584 
03585     tmp1 = MULTIPLY(tmp10 + tmp11, FIX(0.935414347));   /* (c3+c1-c5)/2 */
03586     tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.170262339));   /* (c3+c5-c1)/2 */
03587     tmp0 = tmp1 - tmp2;
03588     tmp1 += tmp2;
03589     tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.378756276)); /* -c1 */
03590     tmp1 += tmp2;
03591     tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.613604268));   /* c5 */
03592     tmp0 += tmp3;
03593     tmp2 += tmp3 + MULTIPLY(tmp12, FIX(1.870828693));   /* c3+c1-c5 */
03594 
03595     dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-PASS1_BITS);
03596     dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-PASS1_BITS);
03597     dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-PASS1_BITS);
03598 
03599     ctr++;
03600 
03601     if (ctr != DCTSIZE) {
03602       if (ctr == 14)
03603         break;                  /* Done. */
03604       dataptr += DCTSIZE;       /* advance pointer to next row */
03605     } else
03606       dataptr = workspace;      /* switch pointer to extended workspace */
03607   }
03608 
03609   /* Pass 2: process columns.
03610    * We remove the PASS1_BITS scaling, but leave the results scaled up
03611    * by an overall factor of 8.
03612    * We must also scale the output by (8/7)*(8/14) = 32/49, which we
03613    * fold into the constant multipliers:
03614    * 14-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/28) * 32/49.
03615    */
03616 
03617   dataptr = data;
03618   wsptr = workspace;
03619   for (ctr = 0; ctr < 7; ctr++) {
03620     /* Even part */
03621 
03622     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*5];
03623     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*4];
03624     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*3];
03625     tmp13 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*2];
03626     tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*1];
03627     tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*0];
03628     tmp6 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
03629 
03630     tmp10 = tmp0 + tmp6;
03631     tmp14 = tmp0 - tmp6;
03632     tmp11 = tmp1 + tmp5;
03633     tmp15 = tmp1 - tmp5;
03634     tmp12 = tmp2 + tmp4;
03635     tmp16 = tmp2 - tmp4;
03636 
03637     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*5];
03638     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*4];
03639     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*3];
03640     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*2];
03641     tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*1];
03642     tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*0];
03643     tmp6 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
03644 
03645     dataptr[DCTSIZE*0] = (DCTELEM)
03646       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12 + tmp13,
03647                        FIX(0.653061224)),                 /* 32/49 */
03648               CONST_BITS+PASS1_BITS);
03649     tmp13 += tmp13;
03650     dataptr[DCTSIZE*4] = (DCTELEM)
03651       DESCALE(MULTIPLY(tmp10 - tmp13, FIX(0.832106052)) + /* c4 */
03652               MULTIPLY(tmp11 - tmp13, FIX(0.205513223)) - /* c12 */
03653               MULTIPLY(tmp12 - tmp13, FIX(0.575835255)),  /* c8 */
03654               CONST_BITS+PASS1_BITS);
03655 
03656     tmp10 = MULTIPLY(tmp14 + tmp15, FIX(0.722074570));    /* c6 */
03657 
03658     dataptr[DCTSIZE*2] = (DCTELEM)
03659       DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.178337691))   /* c2-c6 */
03660               + MULTIPLY(tmp16, FIX(0.400721155)),        /* c10 */
03661               CONST_BITS+PASS1_BITS);
03662     dataptr[DCTSIZE*6] = (DCTELEM)
03663       DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.122795725))   /* c6+c10 */
03664               - MULTIPLY(tmp16, FIX(0.900412262)),        /* c2 */
03665               CONST_BITS+PASS1_BITS);
03666 
03667     /* Odd part */
03668 
03669     tmp10 = tmp1 + tmp2;
03670     tmp11 = tmp5 - tmp4;
03671     dataptr[DCTSIZE*7] = (DCTELEM)
03672       DESCALE(MULTIPLY(tmp0 - tmp10 + tmp3 - tmp11 - tmp6,
03673                        FIX(0.653061224)),                 /* 32/49 */
03674               CONST_BITS+PASS1_BITS);
03675     tmp3  = MULTIPLY(tmp3 , FIX(0.653061224));            /* 32/49 */
03676     tmp10 = MULTIPLY(tmp10, - FIX(0.103406812));          /* -c13 */
03677     tmp11 = MULTIPLY(tmp11, FIX(0.917760839));            /* c1 */
03678     tmp10 += tmp11 - tmp3;
03679     tmp11 = MULTIPLY(tmp0 + tmp2, FIX(0.782007410)) +     /* c5 */
03680             MULTIPLY(tmp4 + tmp6, FIX(0.491367823));      /* c9 */
03681     dataptr[DCTSIZE*5] = (DCTELEM)
03682       DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(1.550341076)) /* c3+c5-c13 */
03683               + MULTIPLY(tmp4, FIX(0.731428202)),         /* c1+c11-c9 */
03684               CONST_BITS+PASS1_BITS);
03685     tmp12 = MULTIPLY(tmp0 + tmp1, FIX(0.871740478)) +     /* c3 */
03686             MULTIPLY(tmp5 - tmp6, FIX(0.305035186));      /* c11 */
03687     dataptr[DCTSIZE*3] = (DCTELEM)
03688       DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.276965844)) /* c3-c9-c13 */
03689               - MULTIPLY(tmp5, FIX(2.004803435)),         /* c1+c5+c11 */
03690               CONST_BITS+PASS1_BITS);
03691     dataptr[DCTSIZE*1] = (DCTELEM)
03692       DESCALE(tmp11 + tmp12 + tmp3
03693               - MULTIPLY(tmp0, FIX(0.735987049))          /* c3+c5-c1 */
03694               - MULTIPLY(tmp6, FIX(0.082925825)),         /* c9-c11-c13 */
03695               CONST_BITS+PASS1_BITS);
03696 
03697     dataptr++;                  /* advance pointer to next column */
03698     wsptr++;                    /* advance pointer to next column */
03699   }
03700 }
03701 
03702 
03703 /*
03704  * Perform the forward DCT on a 6x12 sample block.
03705  *
03706  * 6-point FDCT in pass 1 (rows), 12-point in pass 2 (columns).
03707  */
03708 
03709 GLOBAL(void)
03710 jpeg_fdct_6x12 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03711 {
03712   INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
03713   INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
03714   DCTELEM workspace[8*4];
03715   DCTELEM *dataptr;
03716   DCTELEM *wsptr;
03717   JSAMPROW elemptr;
03718   int ctr;
03719   SHIFT_TEMPS
03720 
03721   /* Pre-zero output coefficient block. */
03722   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03723 
03724   /* Pass 1: process rows. */
03725   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03726   /* furthermore, we scale the results by 2**PASS1_BITS. */
03727   /* 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12). */
03728 
03729   dataptr = data;
03730   ctr = 0;
03731   for (;;) {
03732     elemptr = sample_data[ctr] + start_col;
03733 
03734     /* Even part */
03735 
03736     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
03737     tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
03738     tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
03739 
03740     tmp10 = tmp0 + tmp2;
03741     tmp12 = tmp0 - tmp2;
03742 
03743     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
03744     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
03745     tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
03746 
03747     /* Apply unsigned->signed conversion */
03748     dataptr[0] = (DCTELEM)
03749       ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << PASS1_BITS);
03750     dataptr[2] = (DCTELEM)
03751       DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
03752               CONST_BITS-PASS1_BITS);
03753     dataptr[4] = (DCTELEM)
03754       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
03755               CONST_BITS-PASS1_BITS);
03756 
03757     /* Odd part */
03758 
03759     tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
03760                     CONST_BITS-PASS1_BITS);
03761 
03762     dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << PASS1_BITS));
03763     dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << PASS1_BITS);
03764     dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << PASS1_BITS));
03765 
03766     ctr++;
03767 
03768     if (ctr != DCTSIZE) {
03769       if (ctr == 12)
03770         break;                  /* Done. */
03771       dataptr += DCTSIZE;       /* advance pointer to next row */
03772     } else
03773       dataptr = workspace;      /* switch pointer to extended workspace */
03774   }
03775 
03776   /* Pass 2: process columns.
03777    * We remove the PASS1_BITS scaling, but leave the results scaled up
03778    * by an overall factor of 8.
03779    * We must also scale the output by (8/6)*(8/12) = 8/9, which we
03780    * fold into the constant multipliers:
03781    * 12-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/24) * 8/9.
03782    */
03783 
03784   dataptr = data;
03785   wsptr = workspace;
03786   for (ctr = 0; ctr < 6; ctr++) {
03787     /* Even part */
03788 
03789     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*3];
03790     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*2];
03791     tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*1];
03792     tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*0];
03793     tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*7];
03794     tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*6];
03795 
03796     tmp10 = tmp0 + tmp5;
03797     tmp13 = tmp0 - tmp5;
03798     tmp11 = tmp1 + tmp4;
03799     tmp14 = tmp1 - tmp4;
03800     tmp12 = tmp2 + tmp3;
03801     tmp15 = tmp2 - tmp3;
03802 
03803     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*3];
03804     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*2];
03805     tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*1];
03806     tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*0];
03807     tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*7];
03808     tmp5 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*6];
03809 
03810     dataptr[DCTSIZE*0] = (DCTELEM)
03811       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(0.888888889)), /* 8/9 */
03812               CONST_BITS+PASS1_BITS);
03813     dataptr[DCTSIZE*6] = (DCTELEM)
03814       DESCALE(MULTIPLY(tmp13 - tmp14 - tmp15, FIX(0.888888889)), /* 8/9 */
03815               CONST_BITS+PASS1_BITS);
03816     dataptr[DCTSIZE*4] = (DCTELEM)
03817       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.088662108)),         /* c4 */
03818               CONST_BITS+PASS1_BITS);
03819     dataptr[DCTSIZE*2] = (DCTELEM)
03820       DESCALE(MULTIPLY(tmp14 - tmp15, FIX(0.888888889)) +        /* 8/9 */
03821               MULTIPLY(tmp13 + tmp15, FIX(1.214244803)),         /* c2 */
03822               CONST_BITS+PASS1_BITS);
03823 
03824     /* Odd part */
03825 
03826     tmp10 = MULTIPLY(tmp1 + tmp4, FIX(0.481063200));   /* c9 */
03827     tmp14 = tmp10 + MULTIPLY(tmp1, FIX(0.680326102));  /* c3-c9 */
03828     tmp15 = tmp10 - MULTIPLY(tmp4, FIX(1.642452502));  /* c3+c9 */
03829     tmp12 = MULTIPLY(tmp0 + tmp2, FIX(0.997307603));   /* c5 */
03830     tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.765261039));   /* c7 */
03831     tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.516244403)) /* c5+c7-c1 */
03832             + MULTIPLY(tmp5, FIX(0.164081699));        /* c11 */
03833     tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.164081699)); /* -c11 */
03834     tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.079550144)) /* c1+c5-c11 */
03835             + MULTIPLY(tmp5, FIX(0.765261039));        /* c7 */
03836     tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.645144899)) /* c1+c11-c7 */
03837             - MULTIPLY(tmp5, FIX(0.997307603));        /* c5 */
03838     tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.161389302)) /* c3 */
03839             - MULTIPLY(tmp2 + tmp5, FIX(0.481063200)); /* c9 */
03840 
03841     dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS);
03842     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS);
03843     dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS);
03844     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS);
03845 
03846     dataptr++;                  /* advance pointer to next column */
03847     wsptr++;                    /* advance pointer to next column */
03848   }
03849 }
03850 
03851 
03852 /*
03853  * Perform the forward DCT on a 5x10 sample block.
03854  *
03855  * 5-point FDCT in pass 1 (rows), 10-point in pass 2 (columns).
03856  */
03857 
03858 GLOBAL(void)
03859 jpeg_fdct_5x10 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
03860 {
03861   INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
03862   INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
03863   DCTELEM workspace[8*2];
03864   DCTELEM *dataptr;
03865   DCTELEM *wsptr;
03866   JSAMPROW elemptr;
03867   int ctr;
03868   SHIFT_TEMPS
03869 
03870   /* Pre-zero output coefficient block. */
03871   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
03872 
03873   /* Pass 1: process rows. */
03874   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
03875   /* furthermore, we scale the results by 2**PASS1_BITS. */
03876   /* 5-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/10). */
03877 
03878   dataptr = data;
03879   ctr = 0;
03880   for (;;) {
03881     elemptr = sample_data[ctr] + start_col;
03882 
03883     /* Even part */
03884 
03885     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[4]);
03886     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[3]);
03887     tmp2 = GETJSAMPLE(elemptr[2]);
03888 
03889     tmp10 = tmp0 + tmp1;
03890     tmp11 = tmp0 - tmp1;
03891 
03892     tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[4]);
03893     tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[3]);
03894 
03895     /* Apply unsigned->signed conversion */
03896     dataptr[0] = (DCTELEM)
03897       ((tmp10 + tmp2 - 5 * CENTERJSAMPLE) << PASS1_BITS);
03898     tmp11 = MULTIPLY(tmp11, FIX(0.790569415));          /* (c2+c4)/2 */
03899     tmp10 -= tmp2 << 2;
03900     tmp10 = MULTIPLY(tmp10, FIX(0.353553391));          /* (c2-c4)/2 */
03901     dataptr[2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS-PASS1_BITS);
03902     dataptr[4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS-PASS1_BITS);
03903 
03904     /* Odd part */
03905 
03906     tmp10 = MULTIPLY(tmp0 + tmp1, FIX(0.831253876));    /* c3 */
03907 
03908     dataptr[1] = (DCTELEM)
03909       DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.513743148)), /* c1-c3 */
03910               CONST_BITS-PASS1_BITS);
03911     dataptr[3] = (DCTELEM)
03912       DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.176250899)), /* c1+c3 */
03913               CONST_BITS-PASS1_BITS);
03914 
03915     ctr++;
03916 
03917     if (ctr != DCTSIZE) {
03918       if (ctr == 10)
03919         break;                  /* Done. */
03920       dataptr += DCTSIZE;       /* advance pointer to next row */
03921     } else
03922       dataptr = workspace;      /* switch pointer to extended workspace */
03923   }
03924 
03925   /* Pass 2: process columns.
03926    * We remove the PASS1_BITS scaling, but leave the results scaled up
03927    * by an overall factor of 8.
03928    * We must also scale the output by (8/5)*(8/10) = 32/25, which we
03929    * fold into the constant multipliers:
03930    * 10-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/20) * 32/25.
03931    */
03932 
03933   dataptr = data;
03934   wsptr = workspace;
03935   for (ctr = 0; ctr < 5; ctr++) {
03936     /* Even part */
03937 
03938     tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*1];
03939     tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*0];
03940     tmp12 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*7];
03941     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*6];
03942     tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
03943 
03944     tmp10 = tmp0 + tmp4;
03945     tmp13 = tmp0 - tmp4;
03946     tmp11 = tmp1 + tmp3;
03947     tmp14 = tmp1 - tmp3;
03948 
03949     tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*1];
03950     tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*0];
03951     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*7];
03952     tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*6];
03953     tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
03954 
03955     dataptr[DCTSIZE*0] = (DCTELEM)
03956       DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(1.28)), /* 32/25 */
03957               CONST_BITS+PASS1_BITS);
03958     tmp12 += tmp12;
03959     dataptr[DCTSIZE*4] = (DCTELEM)
03960       DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.464477191)) - /* c4 */
03961               MULTIPLY(tmp11 - tmp12, FIX(0.559380511)),  /* c8 */
03962               CONST_BITS+PASS1_BITS);
03963     tmp10 = MULTIPLY(tmp13 + tmp14, FIX(1.064004961));    /* c6 */
03964     dataptr[DCTSIZE*2] = (DCTELEM)
03965       DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.657591230)),  /* c2-c6 */
03966               CONST_BITS+PASS1_BITS);
03967     dataptr[DCTSIZE*6] = (DCTELEM)
03968       DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.785601151)),  /* c2+c6 */
03969               CONST_BITS+PASS1_BITS);
03970 
03971     /* Odd part */
03972 
03973     tmp10 = tmp0 + tmp4;
03974     tmp11 = tmp1 - tmp3;
03975     dataptr[DCTSIZE*5] = (DCTELEM)
03976       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp2, FIX(1.28)),  /* 32/25 */
03977               CONST_BITS+PASS1_BITS);
03978     tmp2 = MULTIPLY(tmp2, FIX(1.28));                     /* 32/25 */
03979     dataptr[DCTSIZE*1] = (DCTELEM)
03980       DESCALE(MULTIPLY(tmp0, FIX(1.787906876)) +          /* c1 */
03981               MULTIPLY(tmp1, FIX(1.612894094)) + tmp2 +   /* c3 */
03982               MULTIPLY(tmp3, FIX(0.821810588)) +          /* c7 */
03983               MULTIPLY(tmp4, FIX(0.283176630)),           /* c9 */
03984               CONST_BITS+PASS1_BITS);
03985     tmp12 = MULTIPLY(tmp0 - tmp4, FIX(1.217352341)) -     /* (c3+c7)/2 */
03986             MULTIPLY(tmp1 + tmp3, FIX(0.752365123));      /* (c1-c9)/2 */
03987     tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.395541753)) +   /* (c3-c7)/2 */
03988             MULTIPLY(tmp11, FIX(0.64)) - tmp2;            /* 16/25 */
03989     dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS+PASS1_BITS);
03990     dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS+PASS1_BITS);
03991 
03992     dataptr++;                  /* advance pointer to next column */
03993     wsptr++;                    /* advance pointer to next column */
03994   }
03995 }
03996 
03997 
03998 /*
03999  * Perform the forward DCT on a 4x8 sample block.
04000  *
04001  * 4-point FDCT in pass 1 (rows), 8-point in pass 2 (columns).
04002  */
04003 
04004 GLOBAL(void)
04005 jpeg_fdct_4x8 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
04006 {
04007   INT32 tmp0, tmp1, tmp2, tmp3;
04008   INT32 tmp10, tmp11, tmp12, tmp13;
04009   INT32 z1;
04010   DCTELEM *dataptr;
04011   JSAMPROW elemptr;
04012   int ctr;
04013   SHIFT_TEMPS
04014 
04015   /* Pre-zero output coefficient block. */
04016   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
04017 
04018   /* Pass 1: process rows. */
04019   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
04020   /* furthermore, we scale the results by 2**PASS1_BITS. */
04021   /* We must also scale the output by 8/4 = 2, which we add here. */
04022   /* 4-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16). */
04023 
04024   dataptr = data;
04025   for (ctr = 0; ctr < DCTSIZE; ctr++) {
04026     elemptr = sample_data[ctr] + start_col;
04027 
04028     /* Even part */
04029 
04030     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
04031     tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
04032 
04033     tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
04034     tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
04035 
04036     /* Apply unsigned->signed conversion */
04037     dataptr[0] = (DCTELEM)
04038       ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+1));
04039     dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+1));
04040 
04041     /* Odd part */
04042 
04043     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
04044     /* Add fudge factor here for final descale. */
04045     tmp0 += ONE << (CONST_BITS-PASS1_BITS-2);
04046 
04047     dataptr[1] = (DCTELEM)
04048       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
04049                   CONST_BITS-PASS1_BITS-1);
04050     dataptr[3] = (DCTELEM)
04051       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
04052                   CONST_BITS-PASS1_BITS-1);
04053 
04054     dataptr += DCTSIZE;         /* advance pointer to next row */
04055   }
04056 
04057   /* Pass 2: process columns.
04058    * We remove the PASS1_BITS scaling, but leave the results scaled up
04059    * by an overall factor of 8.
04060    */
04061 
04062   dataptr = data;
04063   for (ctr = 0; ctr < 4; ctr++) {
04064     /* Even part per LL&M figure 1 --- note that published figure is faulty;
04065      * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
04066      */
04067 
04068     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
04069     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
04070     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
04071     tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
04072 
04073     /* Add fudge factor here for final descale. */
04074     tmp10 = tmp0 + tmp3 + (ONE << (PASS1_BITS-1));
04075     tmp12 = tmp0 - tmp3;
04076     tmp11 = tmp1 + tmp2;
04077     tmp13 = tmp1 - tmp2;
04078 
04079     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
04080     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
04081     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
04082     tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
04083 
04084     dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp10 + tmp11, PASS1_BITS);
04085     dataptr[DCTSIZE*4] = (DCTELEM) RIGHT_SHIFT(tmp10 - tmp11, PASS1_BITS);
04086 
04087     z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
04088     /* Add fudge factor here for final descale. */
04089     z1 += ONE << (CONST_BITS+PASS1_BITS-1);
04090     dataptr[DCTSIZE*2] = (DCTELEM)
04091       RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865), CONST_BITS+PASS1_BITS);
04092     dataptr[DCTSIZE*6] = (DCTELEM)
04093       RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065), CONST_BITS+PASS1_BITS);
04094 
04095     /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
04096      * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
04097      * i0..i3 in the paper are tmp0..tmp3 here.
04098      */
04099 
04100     tmp10 = tmp0 + tmp3;
04101     tmp11 = tmp1 + tmp2;
04102     tmp12 = tmp0 + tmp2;
04103     tmp13 = tmp1 + tmp3;
04104     z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
04105     /* Add fudge factor here for final descale. */
04106     z1 += ONE << (CONST_BITS+PASS1_BITS-1);
04107 
04108     tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
04109     tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
04110     tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
04111     tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
04112     tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
04113     tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
04114     tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
04115     tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
04116 
04117     tmp12 += z1;
04118     tmp13 += z1;
04119 
04120     dataptr[DCTSIZE*1] = (DCTELEM)
04121       RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS+PASS1_BITS);
04122     dataptr[DCTSIZE*3] = (DCTELEM)
04123       RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS+PASS1_BITS);
04124     dataptr[DCTSIZE*5] = (DCTELEM)
04125       RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS+PASS1_BITS);
04126     dataptr[DCTSIZE*7] = (DCTELEM)
04127       RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS+PASS1_BITS);
04128 
04129     dataptr++;                  /* advance pointer to next column */
04130   }
04131 }
04132 
04133 
04134 /*
04135  * Perform the forward DCT on a 3x6 sample block.
04136  *
04137  * 3-point FDCT in pass 1 (rows), 6-point in pass 2 (columns).
04138  */
04139 
04140 GLOBAL(void)
04141 jpeg_fdct_3x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
04142 {
04143   INT32 tmp0, tmp1, tmp2;
04144   INT32 tmp10, tmp11, tmp12;
04145   DCTELEM *dataptr;
04146   JSAMPROW elemptr;
04147   int ctr;
04148   SHIFT_TEMPS
04149 
04150   /* Pre-zero output coefficient block. */
04151   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
04152 
04153   /* Pass 1: process rows. */
04154   /* Note results are scaled up by sqrt(8) compared to a true DCT; */
04155   /* furthermore, we scale the results by 2**PASS1_BITS. */
04156   /* We scale the results further by 2 as part of output adaption */
04157   /* scaling for different DCT size. */
04158   /* 3-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/6). */
04159 
04160   dataptr = data;
04161   for (ctr = 0; ctr < 6; ctr++) {
04162     elemptr = sample_data[ctr] + start_col;
04163 
04164     /* Even part */
04165 
04166     tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[2]);
04167     tmp1 = GETJSAMPLE(elemptr[1]);
04168 
04169     tmp2 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[2]);
04170 
04171     /* Apply unsigned->signed conversion */
04172     dataptr[0] = (DCTELEM)
04173       ((tmp0 + tmp1 - 3 * CENTERJSAMPLE) << (PASS1_BITS+1));
04174     dataptr[2] = (DCTELEM)
04175       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(0.707106781)), /* c2 */
04176               CONST_BITS-PASS1_BITS-1);
04177 
04178     /* Odd part */
04179 
04180     dataptr[1] = (DCTELEM)
04181       DESCALE(MULTIPLY(tmp2, FIX(1.224744871)),               /* c1 */
04182               CONST_BITS-PASS1_BITS-1);
04183 
04184     dataptr += DCTSIZE;         /* advance pointer to next row */
04185   }
04186 
04187   /* Pass 2: process columns.
04188    * We remove the PASS1_BITS scaling, but leave the results scaled up
04189    * by an overall factor of 8.
04190    * We must also scale the output by (8/6)*(8/3) = 32/9, which we partially
04191    * fold into the constant multipliers (other part was done in pass 1):
04192    * 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12) * 16/9.
04193    */
04194 
04195   dataptr = data;
04196   for (ctr = 0; ctr < 3; ctr++) {
04197     /* Even part */
04198 
04199     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
04200     tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
04201     tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
04202 
04203     tmp10 = tmp0 + tmp2;
04204     tmp12 = tmp0 - tmp2;
04205 
04206     tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
04207     tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
04208     tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
04209 
04210     dataptr[DCTSIZE*0] = (DCTELEM)
04211       DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
04212               CONST_BITS+PASS1_BITS);
04213     dataptr[DCTSIZE*2] = (DCTELEM)
04214       DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
04215               CONST_BITS+PASS1_BITS);
04216     dataptr[DCTSIZE*4] = (DCTELEM)
04217       DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
04218               CONST_BITS+PASS1_BITS);
04219 
04220     /* Odd part */
04221 
04222     tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
04223 
04224     dataptr[DCTSIZE*1] = (DCTELEM)
04225       DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
04226               CONST_BITS+PASS1_BITS);
04227     dataptr[DCTSIZE*3] = (DCTELEM)
04228       DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
04229               CONST_BITS+PASS1_BITS);
04230     dataptr[DCTSIZE*5] = (DCTELEM)
04231       DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
04232               CONST_BITS+PASS1_BITS);
04233 
04234     dataptr++;                  /* advance pointer to next column */
04235   }
04236 }
04237 
04238 
04239 /*
04240  * Perform the forward DCT on a 2x4 sample block.
04241  *
04242  * 2-point FDCT in pass 1 (rows), 4-point in pass 2 (columns).
04243  */
04244 
04245 GLOBAL(void)
04246 jpeg_fdct_2x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
04247 {
04248   INT32 tmp0, tmp1;
04249   INT32 tmp10, tmp11;
04250   DCTELEM *dataptr;
04251   JSAMPROW elemptr;
04252   int ctr;
04253   SHIFT_TEMPS
04254 
04255   /* Pre-zero output coefficient block. */
04256   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
04257 
04258   /* Pass 1: process rows. */
04259   /* Note results are scaled up by sqrt(8) compared to a true DCT. */
04260   /* We must also scale the output by (8/2)*(8/4) = 2**3, which we add here. */
04261 
04262   dataptr = data;
04263   for (ctr = 0; ctr < 4; ctr++) {
04264     elemptr = sample_data[ctr] + start_col;
04265 
04266     /* Even part */
04267 
04268     tmp0 = GETJSAMPLE(elemptr[0]);
04269     tmp1 = GETJSAMPLE(elemptr[1]);
04270 
04271     /* Apply unsigned->signed conversion */
04272     dataptr[0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 3);
04273 
04274     /* Odd part */
04275 
04276     dataptr[1] = (DCTELEM) ((tmp0 - tmp1) << 3);
04277 
04278     dataptr += DCTSIZE;         /* advance pointer to next row */
04279   }
04280 
04281   /* Pass 2: process columns.
04282    * We leave the results scaled up by an overall factor of 8.
04283    * 4-point FDCT kernel,
04284    * cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT].
04285    */
04286 
04287   dataptr = data;
04288   for (ctr = 0; ctr < 2; ctr++) {
04289     /* Even part */
04290 
04291     tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3];
04292     tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
04293 
04294     tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
04295     tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
04296 
04297     dataptr[DCTSIZE*0] = (DCTELEM) (tmp0 + tmp1);
04298     dataptr[DCTSIZE*2] = (DCTELEM) (tmp0 - tmp1);
04299 
04300     /* Odd part */
04301 
04302     tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
04303     /* Add fudge factor here for final descale. */
04304     tmp0 += ONE << (CONST_BITS-1);
04305 
04306     dataptr[DCTSIZE*1] = (DCTELEM)
04307       RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
04308                   CONST_BITS);
04309     dataptr[DCTSIZE*3] = (DCTELEM)
04310       RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
04311                   CONST_BITS);
04312 
04313     dataptr++;                  /* advance pointer to next column */
04314   }
04315 }
04316 
04317 
04318 /*
04319  * Perform the forward DCT on a 1x2 sample block.
04320  *
04321  * 1-point FDCT in pass 1 (rows), 2-point in pass 2 (columns).
04322  */
04323 
04324 GLOBAL(void)
04325 jpeg_fdct_1x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
04326 {
04327   INT32 tmp0, tmp1;
04328 
04329   /* Pre-zero output coefficient block. */
04330   MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
04331 
04332   tmp0 = GETJSAMPLE(sample_data[0][start_col]);
04333   tmp1 = GETJSAMPLE(sample_data[1][start_col]);
04334 
04335   /* We leave the results scaled up by an overall factor of 8.
04336    * We must also scale the output by (8/1)*(8/2) = 2**5.
04337    */
04338 
04339   /* Even part */
04340   /* Apply unsigned->signed conversion */
04341   data[DCTSIZE*0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 5);
04342 
04343   /* Odd part */
04344   data[DCTSIZE*1] = (DCTELEM) ((tmp0 - tmp1) << 5);
04345 }
04346 
04347 #endif /* DCT_SCALING_SUPPORTED */
04348 #endif /* DCT_ISLOW_SUPPORTED */
Generated on Tue Jul 5 14:13:29 2011 for ROOT_528-00b_version by  doxygen 1.5.1