FFmpeg: libavcodec/mips/mpegaudiodsp_mips

00001 /*
00002  * Copyright (c) 2012
00003  *      MIPS Technologies, Inc., California.
00004  *
00005  * Redistribution and use in source and binary forms, with or without
00006  * modification, are permitted provided that the following conditions
00007  * are met:
00008  * 1. Redistributions of source code must retain the above copyright
00009  *    notice, this list of conditions and the following disclaimer.
00010  * 2. Redistributions in binary form must reproduce the above copyright
00011  *    notice, this list of conditions and the following disclaimer in the
00012  *    documentation and/or other materials provided with the distribution.
00013  * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
00014  *    contributors may be used to endorse or promote products derived from
00015  *    this software without specific prior written permission.
00016  *
00017  * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
00018  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
00019  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
00020  * ARE DISCLAIMED.  IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
00021  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
00022  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
00023  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
00024  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
00025  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
00026  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
00027  * SUCH DAMAGE.
00028  *
00029  * Author:  Bojan Zivkovic (bojan@mips.com)
00030  *
00031  * MPEG Audio decoder optimized for MIPS floating-point architecture
00032  *
00033  * This file is part of FFmpeg.
00034  *
00035  * FFmpeg is free software; you can redistribute it and/or
00036  * modify it under the terms of the GNU Lesser General Public
00037  * License as published by the Free Software Foundation; either
00038  * version 2.1 of the License, or (at your option) any later version.
00039  *
00040  * FFmpeg is distributed in the hope that it will be useful,
00041  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00042  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00043  * Lesser General Public License for more details.
00044  *
00045  * You should have received a copy of the GNU Lesser General Public
00046  * License along with FFmpeg; if not, write to the Free Software
00047  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00048  */
00049 
00056 #include <string.h>
00057 
00058 #include "libavcodec/mpegaudiodsp.h"
00059 
00060 static void ff_mpadsp_apply_window_mips_float(float *synth_buf, float *window,
00061                                int *dither_state, float *samples, int incr)
00062 {
00063     register const float *w, *w2, *p;
00064     int j;
00065     float *samples2;
00066     float sum, sum2;
00067     /* temporary variables */
00068     int incr1 = incr << 2;
00069     int t_sample;
00070     float in1, in2, in3, in4, in5, in6, in7, in8;
00071     float *p2;
00072 
00073     /* copy to avoid wrap */
00074     memcpy(synth_buf + 512, synth_buf, 32 * sizeof(*synth_buf));
00075 
00082     __asm__ volatile (
00083         "lwc1    %[sum],      0(%[dither_state])                            \t\n"
00084         "sll     %[t_sample], %[incr1],     5                               \t\n"
00085         "sub     %[t_sample], %[t_sample],  %[incr1]                        \n\t"
00086         "li      %[j],        4                                             \t\n"
00087         "lwc1    %[in1],      0(%[window])                                  \t\n"
00088         "lwc1    %[in2],      16*4(%[synth_buf])                            \t\n"
00089         "sw      $zero,       0(%[dither_state])                            \t\n"
00090         "lwc1    %[in3],      64*4(%[window])                               \t\n"
00091         "lwc1    %[in4],      80*4(%[synth_buf])                            \t\n"
00092         "addu    %[samples2], %[samples],   %[t_sample]                     \t\n"
00093         "madd.s  %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00094         "lwc1    %[in5],      128*4(%[window])                              \t\n"
00095         "lwc1    %[in6],      144*4(%[synth_buf])                           \t\n"
00096         "lwc1    %[in7],      192*4(%[window])                              \t\n"
00097         "madd.s  %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00098         "lwc1    %[in8],      208*4(%[synth_buf])                           \t\n"
00099         "lwc1    %[in1],      256*4(%[window])                              \t\n"
00100         "lwc1    %[in2],      272*4(%[synth_buf])                           \t\n"
00101         "madd.s  %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00102         "lwc1    %[in3],      320*4(%[window])                              \t\n"
00103         "lwc1    %[in4],      336*4(%[synth_buf])                           \t\n"
00104         "lwc1    %[in5],      384*4(%[window])                              \t\n"
00105         "madd.s  %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00106         "lwc1    %[in6],      400*4(%[synth_buf])                           \t\n"
00107         "lwc1    %[in7],      448*4(%[window])                              \t\n"
00108         "lwc1    %[in8],      464*4(%[synth_buf])                           \t\n"
00109         "madd.s  %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00110         "lwc1    %[in1],      32*4(%[window])                               \t\n"
00111         "lwc1    %[in2],      48*4(%[synth_buf])                            \t\n"
00112         "madd.s  %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00113         "lwc1    %[in3],      96*4(%[window])                               \t\n"
00114         "lwc1    %[in4],      112*4(%[synth_buf])                           \t\n"
00115         "madd.s  %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00116         "lwc1    %[in5],      160*4(%[window])                              \t\n"
00117         "lwc1    %[in6],      176*4(%[synth_buf])                           \t\n"
00118         "madd.s  %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00119         "lwc1    %[in7],      224*4(%[window])                              \t\n"
00120         "lwc1    %[in8],      240*4(%[synth_buf])                           \t\n"
00121         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00122         "lwc1    %[in1],      288*4(%[window])                              \t\n"
00123         "lwc1    %[in2],      304*4(%[synth_buf])                           \t\n"
00124         "nmsub.s %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00125         "lwc1    %[in3],      352*4(%[window])                              \t\n"
00126         "lwc1    %[in4],      368*4(%[synth_buf])                           \t\n"
00127         "nmsub.s %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00128         "lwc1    %[in5],      416*4(%[window])                              \t\n"
00129         "lwc1    %[in6],      432*4(%[synth_buf])                           \t\n"
00130         "nmsub.s %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00131         "lwc1    %[in7],      480*4(%[window])                              \t\n"
00132         "lwc1    %[in8],      496*4(%[synth_buf])                           \t\n"
00133         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00134         "addu    %[w],        %[window],    4                               \t\n"
00135         "nmsub.s %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00136         "addu    %[w2],       %[window],    124                             \t\n"
00137         "addiu   %[p],        %[synth_buf], 68                              \t\n"
00138         "addiu   %[p2],       %[synth_buf], 188                             \t\n"
00139         "nmsub.s %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00140         "nmsub.s %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00141         "swc1    %[sum],      0(%[samples])                                 \t\n"
00142         "addu    %[samples],  %[samples],   %[incr1]                        \t\n"
00143 
00144         /* calculate two samples at the same time to avoid one memory
00145            access per two sample */
00146 
00147         "ff_mpadsp_apply_window_loop%=:                                     \t\n"
00148         "lwc1    %[in1],      0(%[w])                                       \t\n"
00149         "lwc1    %[in2],      0(%[p])                                       \t\n"
00150         "lwc1    %[in3],      0(%[w2])                                      \t\n"
00151         "lwc1    %[in4],      64*4(%[w])                                    \t\n"
00152         "lwc1    %[in5],      64*4(%[p])                                    \t\n"
00153         "lwc1    %[in6],      64*4(%[w2])                                   \t\n"
00154         "mul.s   %[sum],      %[in1],       %[in2]                          \t\n"
00155         "mul.s   %[sum2],     %[in2],       %[in3]                          \t\n"
00156         "lwc1    %[in1],      128*4(%[w])                                   \t\n"
00157         "lwc1    %[in2],      128*4(%[p])                                   \t\n"
00158         "madd.s  %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00159         "nmadd.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00160         "lwc1    %[in3],      128*4(%[w2])                                  \t\n"
00161         "lwc1    %[in4],      192*4(%[w])                                   \t\n"
00162         "madd.s  %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00163         "lwc1    %[in5],      192*4(%[p])                                   \t\n"
00164         "lwc1    %[in6],      192*4(%[w2])                                  \t\n"
00165         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00166         "lwc1    %[in1],      256*4(%[w])                                   \t\n"
00167         "lwc1    %[in2],      256*4(%[p])                                   \t\n"
00168         "madd.s  %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00169         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00170         "lwc1    %[in3],      256*4(%[w2])                                  \t\n"
00171         "lwc1    %[in4],      320*4(%[w])                                   \t\n"
00172         "madd.s  %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00173         "lwc1    %[in5],      320*4(%[p])                                   \t\n"
00174         "lwc1    %[in6],      320*4(%[w2])                                  \t\n"
00175         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00176         "lwc1    %[in1],      384*4(%[w])                                   \t\n"
00177         "lwc1    %[in2],      384*4(%[p])                                   \t\n"
00178         "madd.s  %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00179         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00180         "lwc1    %[in3],      384*4(%[w2])                                  \t\n"
00181         "lwc1    %[in4],      448*4(%[w])                                   \t\n"
00182         "madd.s  %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00183         "lwc1    %[in5],      448*4(%[p])                                   \t\n"
00184         "lwc1    %[in6],      448*4(%[w2])                                  \t\n"
00185         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00186         "madd.s  %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00187         "lwc1    %[in1],      32*4(%[w])                                    \t\n"
00188         "lwc1    %[in2],      0(%[p2])                                      \t\n"
00189         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00190         "lwc1    %[in3],      32*4(%[w2])                                   \t\n"
00191         "lwc1    %[in4],      96*4(%[w])                                    \t\n"
00192         "lwc1    %[in5],      64*4(%[p2])                                   \t\n"
00193         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00194         "lwc1    %[in6],      96*4(%[w2])                                   \t\n"
00195         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00196         "lwc1    %[in1],      160*4(%[w])                                   \t\n"
00197         "nmsub.s %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00198         "lwc1    %[in2],      128*4(%[p2])                                  \t\n"
00199         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00200         "lwc1    %[in3],      160*4(%[w2])                                  \t\n"
00201         "lwc1    %[in4],      224*4(%[w])                                   \t\n"
00202         "lwc1    %[in5],      192*4(%[p2])                                  \t\n"
00203         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00204         "lwc1    %[in6],      224*4(%[w2])                                  \t\n"
00205         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00206         "lwc1    %[in1],      288*4(%[w])                                   \t\n"
00207         "nmsub.s %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00208         "lwc1    %[in2],      256*4(%[p2])                                  \t\n"
00209         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00210         "lwc1    %[in3],      288*4(%[w2])                                  \t\n"
00211         "lwc1    %[in4],      352*4(%[w])                                   \t\n"
00212         "lwc1    %[in5],      320*4(%[p2])                                  \t\n"
00213         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00214         "lwc1    %[in6],      352*4(%[w2])                                  \t\n"
00215         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00216         "lwc1    %[in1],      416*4(%[w])                                   \t\n"
00217         "nmsub.s %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00218         "lwc1    %[in2],      384*4(%[p2])                                  \t\n"
00219         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00220         "lwc1    %[in3],      416*4(%[w2])                                  \t\n"
00221         "lwc1    %[in4],      480*4(%[w])                                   \t\n"
00222         "lwc1    %[in5],      448*4(%[p2])                                  \t\n"
00223         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00224         "lwc1    %[in6],      480*4(%[w2])                                  \t\n"
00225         "nmsub.s %[sum2],     %[sum2],      %[in2], %[in3]                  \t\n"
00226         "addiu   %[w],        %[w],         4                               \t\n"
00227         "nmsub.s %[sum],      %[sum],       %[in4], %[in5]                  \t\n"
00228         "addiu   %[w2],       %[w2],        -4                              \t\n"
00229         "nmsub.s %[sum2],     %[sum2],      %[in5], %[in6]                  \t\n"
00230         "addu    %[j],        %[j],         4                               \t\n"
00231         "addiu   %[p],        4                                             \t\n"
00232         "swc1    %[sum],      0(%[samples])                                 \t\n"
00233         "addiu   %[p2],       -4                                            \t\n"
00234         "swc1    %[sum2],     0(%[samples2])                                \t\n"
00235         "addu    %[samples],  %[samples],   %[incr1]                        \t\n"
00236         "subu    %[samples2], %[samples2],  %[incr1]                        \t\n"
00237         "bne     %[j],        64,           ff_mpadsp_apply_window_loop%=   \t\n"
00238 
00239         "lwc1    %[in1],      48*4(%[window])                               \t\n"
00240         "lwc1    %[in2],      32*4(%[synth_buf])                            \t\n"
00241         "lwc1    %[in3],      112*4(%[window])                              \t\n"
00242         "lwc1    %[in4],      96*4(%[synth_buf])                            \t\n"
00243         "lwc1    %[in5],      176*4(%[window])                              \t\n"
00244         "lwc1    %[in6],      160*4(%[synth_buf])                           \t\n"
00245         "mul.s   %[sum],      %[in1],       %[in2]                          \t\n"
00246         "lwc1    %[in7],      240*4(%[window])                              \t\n"
00247         "lwc1    %[in8],      224*4(%[synth_buf])                           \t\n"
00248         "lwc1    %[in1],      304*4(%[window])                              \t\n"
00249         "nmadd.s %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00250         "lwc1    %[in2],      288*4(%[synth_buf])                           \t\n"
00251         "lwc1    %[in3],      368*4(%[window])                              \t\n"
00252         "lwc1    %[in4],      352*4(%[synth_buf])                           \t\n"
00253         "nmsub.s %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00254         "nmsub.s %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00255         "lwc1    %[in5],      432*4(%[window])                              \t\n"
00256         "lwc1    %[in6],      416*4(%[synth_buf])                           \t\n"
00257         "nmsub.s %[sum],      %[sum],       %[in1], %[in2]                  \t\n"
00258         "lwc1    %[in7],      496*4(%[window])                              \t\n"
00259         "lwc1    %[in8],      480*4(%[synth_buf])                           \t\n"
00260         "nmsub.s %[sum],      %[sum],       %[in3], %[in4]                  \t\n"
00261         "nmsub.s %[sum],      %[sum],       %[in5], %[in6]                  \t\n"
00262         "nmsub.s %[sum],      %[sum],       %[in7], %[in8]                  \t\n"
00263         "swc1    %[sum],      0(%[samples])                                 \t\n"
00264 
00265         : [sum] "=&f" (sum), [sum2] "=&f" (sum2),
00266           [w2] "=&r" (w2),   [w] "=&r" (w),
00267           [p] "=&r" (p), [p2] "=&r" (p2), [j] "=&r" (j),
00268           [samples] "+r" (samples), [samples2] "=&r" (samples2),
00269           [in1] "=&f" (in1), [in2] "=&f" (in2),
00270           [in3] "=&f" (in3), [in4] "=&f" (in4),
00271           [in5] "=&f" (in5), [in6] "=&f" (in6),
00272           [in7] "=&f" (in7), [in8] "=&f" (in8),
00273           [t_sample] "=&r" (t_sample)
00274         : [synth_buf] "r" (synth_buf), [window] "r" (window),
00275           [dither_state] "r" (dither_state), [incr1] "r" (incr1)
00276     );
00277 }
00278 
00279 static void ff_dct32_mips_float(float *out, const float *tab)
00280 {
00281     float val0 , val1 , val2 , val3 , val4 , val5 , val6 , val7,
00282           val8 , val9 , val10, val11, val12, val13, val14, val15,
00283           val16, val17, val18, val19, val20, val21, val22, val23,
00284           val24, val25, val26, val27, val28, val29, val30, val31;
00285     float fTmp1, fTmp2, fTmp3, fTmp4, fTmp5, fTmp6, fTmp7, fTmp8,
00286           fTmp9, fTmp10, fTmp11;
00287 
00291     __asm__ volatile (
00292         "lwc1       %[fTmp1],       0*4(%[tab])                             \n\t"
00293         "lwc1       %[fTmp2],       31*4(%[tab])                            \n\t"
00294         "lwc1       %[fTmp3],       15*4(%[tab])                            \n\t"
00295         "lwc1       %[fTmp4],       16*4(%[tab])                            \n\t"
00296         "li.s       %[fTmp7],       0.50241928618815570551                  \n\t"
00297         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00298         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00299         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00300         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00301         "li.s       %[fTmp10],      0.50060299823519630134                  \n\t"
00302         "li.s       %[fTmp11],      10.19000812354805681150                 \n\t"
00303         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00304         "add.s      %[val0],        %[fTmp5],       %[fTmp6]                \n\t"
00305         "sub.s      %[val15],       %[fTmp5],       %[fTmp6]                \n\t"
00306         "lwc1       %[fTmp1],       7*4(%[tab])                             \n\t"
00307         "lwc1       %[fTmp2],       24*4(%[tab])                            \n\t"
00308         "madd.s     %[val16],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00309         "nmsub.s    %[val31],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00310         "mul.s      %[val15],       %[val15],       %[fTmp7]                \n\t"
00311         "lwc1       %[fTmp3],       8*4(%[tab])                             \n\t"
00312         "lwc1       %[fTmp4],       23*4(%[tab])                            \n\t"
00313         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00314         "mul.s      %[val31],       %[val31],       %[fTmp7]                \n\t"
00315         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00316         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00317         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00318         "li.s       %[fTmp7],       5.10114861868916385802                  \n\t"
00319         "li.s       %[fTmp10],      0.67480834145500574602                  \n\t"
00320         "li.s       %[fTmp11],      0.74453627100229844977                  \n\t"
00321         "add.s      %[val7],        %[fTmp5],       %[fTmp6]                \n\t"
00322         "sub.s      %[val8],        %[fTmp5],       %[fTmp6]                \n\t"
00323         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00324         "li.s       %[fTmp1],       0.50979557910415916894                  \n\t"
00325         "sub.s      %[fTmp2],       %[val0],        %[val7]                 \n\t"
00326         "mul.s      %[val8],        %[val8],        %[fTmp7]                \n\t"
00327         "madd.s     %[val23],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00328         "nmsub.s    %[val24],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00329         "add.s      %[val0],        %[val0],        %[val7]                 \n\t"
00330         "mul.s      %[val7],        %[fTmp1],       %[fTmp2]                \n\t"
00331         "sub.s      %[fTmp2],       %[val15],       %[val8]                 \n\t"
00332         "add.s      %[val8],        %[val15],       %[val8]                 \n\t"
00333         "mul.s      %[val24],       %[val24],       %[fTmp7]                \n\t"
00334         "sub.s      %[fTmp3],       %[val16],       %[val23]                \n\t"
00335         "add.s      %[val16],       %[val16],       %[val23]                \n\t"
00336         "mul.s      %[val15],       %[fTmp1],       %[fTmp2]                \n\t"
00337         "sub.s      %[fTmp4],       %[val31],       %[val24]                \n\t"
00338         "mul.s      %[val23],       %[fTmp1],       %[fTmp3]                \n\t"
00339         "add.s      %[val24],       %[val31],       %[val24]                \n\t"
00340         "mul.s      %[val31],       %[fTmp1],       %[fTmp4]                \n\t"
00341 
00342         : [fTmp1]  "=&f" (fTmp1),  [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3),
00343           [fTmp4]  "=&f" (fTmp4),  [fTmp5] "=&f" (fTmp5), [fTmp6] "=&f" (fTmp6),
00344           [fTmp7]  "=&f" (fTmp7),  [fTmp8] "=&f" (fTmp8), [fTmp9] "=&f" (fTmp9),
00345           [fTmp10] "=&f" (fTmp10), [fTmp11] "=&f" (fTmp11),
00346           [val0]  "=f" (val0),  [val7]  "=f" (val7),
00347           [val8]  "=f" (val8),  [val15] "=f" (val15),
00348           [val16] "=f" (val16), [val23] "=f" (val23),
00349           [val24] "=f" (val24), [val31] "=f" (val31)
00350         : [tab] "r" (tab)
00351     );
00352 
00353     __asm__ volatile (
00354         "lwc1       %[fTmp1],       3*4(%[tab])                             \n\t"
00355         "lwc1       %[fTmp2],       28*4(%[tab])                            \n\t"
00356         "lwc1       %[fTmp3],       12*4(%[tab])                            \n\t"
00357         "lwc1       %[fTmp4],       19*4(%[tab])                            \n\t"
00358         "li.s       %[fTmp7],       0.64682178335999012954                  \n\t"
00359         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00360         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00361         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00362         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00363         "li.s       %[fTmp10],      0.53104259108978417447                  \n\t"
00364         "li.s       %[fTmp11],      1.48416461631416627724                  \n\t"
00365         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00366         "add.s      %[val3],        %[fTmp5],       %[fTmp6]                \n\t"
00367         "sub.s      %[val12],       %[fTmp5],       %[fTmp6]                \n\t"
00368         "lwc1       %[fTmp1],       4*4(%[tab])                             \n\t"
00369         "lwc1       %[fTmp2],       27*4(%[tab])                            \n\t"
00370         "madd.s     %[val19],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00371         "nmsub.s    %[val28],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00372         "mul.s      %[val12],       %[val12],       %[fTmp7]                \n\t"
00373         "lwc1       %[fTmp3],       11*4(%[tab])                            \n\t"
00374         "lwc1       %[fTmp4],       20*4(%[tab])                            \n\t"
00375         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00376         "mul.s      %[val28],       %[val28],       %[fTmp7]                \n\t"
00377         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00378         "li.s       %[fTmp7],       0.78815462345125022473                  \n\t"
00379         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00380         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00381         "li.s       %[fTmp10],      0.55310389603444452782                  \n\t"
00382         "li.s       %[fTmp11],      1.16943993343288495515                  \n\t"
00383         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00384         "add.s      %[val4],        %[fTmp5],       %[fTmp6]                \n\t"
00385         "sub.s      %[val11],       %[fTmp5],       %[fTmp6]                \n\t"
00386         "li.s       %[fTmp1],       2.56291544774150617881                  \n\t"
00387         "madd.s     %[val20],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00388         "nmsub.s    %[val27],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00389         "mul.s      %[val11],       %[val11],       %[fTmp7]                \n\t"
00390         "sub.s      %[fTmp2],       %[val3],        %[val4]                 \n\t"
00391         "add.s      %[val3],        %[val3],        %[val4]                 \n\t"
00392         "sub.s      %[fTmp4],       %[val19],       %[val20]                \n\t"
00393         "mul.s      %[val27],       %[val27],       %[fTmp7]                \n\t"
00394         "sub.s      %[fTmp3],       %[val12],       %[val11]                \n\t"
00395         "mul.s      %[val4],        %[fTmp1],       %[fTmp2]                \n\t"
00396         "add.s      %[val11],       %[val12],       %[val11]                \n\t"
00397         "add.s      %[val19],       %[val19],       %[val20]                \n\t"
00398         "mul.s      %[val20],       %[fTmp1],       %[fTmp4]                \n\t"
00399         "mul.s      %[val12],       %[fTmp1],       %[fTmp3]                \n\t"
00400         "sub.s      %[fTmp2],       %[val28],       %[val27]                \n\t"
00401         "add.s      %[val27],       %[val28],       %[val27]                \n\t"
00402         "mul.s      %[val28],       %[fTmp1],       %[fTmp2]                \n\t"
00403 
00404         : [fTmp1]  "=&f" (fTmp1),  [fTmp2]  "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3),
00405           [fTmp4]  "=&f" (fTmp4),  [fTmp5]  "=&f" (fTmp5), [fTmp6] "=&f" (fTmp6),
00406           [fTmp7]  "=&f" (fTmp7),  [fTmp8]  "=&f" (fTmp8), [fTmp9] "=&f" (fTmp9),
00407           [fTmp10] "=&f" (fTmp10), [fTmp11] "=&f" (fTmp11),
00408           [val3]  "=f" (val3),  [val4]  "=f" (val4),
00409           [val11] "=f" (val11), [val12] "=f" (val12),
00410           [val19] "=f" (val19), [val20] "=f" (val20),
00411           [val27] "=f" (val27), [val28] "=f" (val28)
00412         : [tab] "r" (tab)
00413     );
00414 
00415     __asm__ volatile (
00416         "li.s       %[fTmp1],       0.54119610014619698439                  \n\t"
00417         "sub.s      %[fTmp2],       %[val0],        %[val3]                 \n\t"
00418         "add.s      %[val0],        %[val0],        %[val3]                 \n\t"
00419         "sub.s      %[fTmp3],       %[val7],        %[val4]                 \n\t"
00420         "add.s      %[val4],        %[val7],        %[val4]                 \n\t"
00421         "sub.s      %[fTmp4],       %[val8],        %[val11]                \n\t"
00422         "mul.s      %[val3],        %[fTmp1],       %[fTmp2]                \n\t"
00423         "add.s      %[val8],        %[val8],        %[val11]                \n\t"
00424         "mul.s      %[val7],        %[fTmp1],       %[fTmp3]                \n\t"
00425         "sub.s      %[fTmp2],       %[val15],       %[val12]                \n\t"
00426         "mul.s      %[val11],       %[fTmp1],       %[fTmp4]                \n\t"
00427         "add.s      %[val12],       %[val15],       %[val12]                \n\t"
00428         "mul.s      %[val15],       %[fTmp1],       %[fTmp2]                \n\t"
00429 
00430         : [val0]  "+f" (val0),   [val3] "+f" (val3),
00431           [val4]  "+f" (val4),   [val7] "+f" (val7),
00432           [val8]  "+f" (val8),   [val11] "+f" (val11),
00433           [val12] "+f" (val12),  [val15] "+f" (val15),
00434           [fTmp1] "=f"  (fTmp1), [fTmp2] "=&f" (fTmp2),
00435           [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4)
00436         :
00437     );
00438 
00439     __asm__ volatile (
00440         "sub.s      %[fTmp2],       %[val16],       %[val19]                \n\t"
00441         "add.s      %[val16],       %[val16],       %[val19]                \n\t"
00442         "sub.s      %[fTmp3],       %[val23],       %[val20]                \n\t"
00443         "add.s      %[val20],       %[val23],       %[val20]                \n\t"
00444         "sub.s      %[fTmp4],       %[val24],       %[val27]                \n\t"
00445         "mul.s      %[val19],       %[fTmp1],       %[fTmp2]                \n\t"
00446         "add.s      %[val24],       %[val24],       %[val27]                \n\t"
00447         "mul.s      %[val23],       %[fTmp1],       %[fTmp3]                \n\t"
00448         "sub.s      %[fTmp2],       %[val31],       %[val28]                \n\t"
00449         "mul.s      %[val27],       %[fTmp1],       %[fTmp4]                \n\t"
00450         "add.s      %[val28],       %[val31],       %[val28]                \n\t"
00451         "mul.s      %[val31],       %[fTmp1],       %[fTmp2]                \n\t"
00452 
00453         : [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00454           [val16] "+f" (val16), [val19] "+f" (val19), [val20] "+f" (val20),
00455           [val23] "+f" (val23), [val24] "+f" (val24), [val27] "+f" (val27),
00456           [val28] "+f" (val28), [val31] "+f" (val31)
00457         : [fTmp1] "f" (fTmp1)
00458     );
00459 
00460     __asm__ volatile (
00461         "lwc1       %[fTmp1],       1*4(%[tab])                             \n\t"
00462         "lwc1       %[fTmp2],       30*4(%[tab])                            \n\t"
00463         "lwc1       %[fTmp3],       14*4(%[tab])                            \n\t"
00464         "lwc1       %[fTmp4],       17*4(%[tab])                            \n\t"
00465         "li.s       %[fTmp7],       0.52249861493968888062                  \n\t"
00466         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00467         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00468         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00469         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00470         "li.s       %[fTmp10],      0.50547095989754365998                  \n\t"
00471         "li.s       %[fTmp11],      3.40760841846871878570                  \n\t"
00472         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00473         "add.s      %[val1],        %[fTmp5],       %[fTmp6]                \n\t"
00474         "sub.s      %[val14],       %[fTmp5],       %[fTmp6]                \n\t"
00475         "lwc1       %[fTmp1],       6*4(%[tab])                             \n\t"
00476         "lwc1       %[fTmp2],       25*4(%[tab])                            \n\t"
00477         "madd.s     %[val17],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00478         "nmsub.s    %[val30],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00479         "mul.s      %[val14],       %[val14],       %[fTmp7]                \n\t"
00480         "lwc1       %[fTmp3],       9*4(%[tab])                             \n\t"
00481         "lwc1       %[fTmp4],       22*4(%[tab])                            \n\t"
00482         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00483         "mul.s      %[val30],       %[val30],       %[fTmp7]                \n\t"
00484         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00485         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00486         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00487         "li.s       %[fTmp7],       1.72244709823833392782                  \n\t"
00488         "li.s       %[fTmp10],      0.62250412303566481615                  \n\t"
00489         "li.s       %[fTmp11],      0.83934964541552703873                  \n\t"
00490         "add.s      %[val6],        %[fTmp5],       %[fTmp6]                \n\t"
00491         "sub.s      %[val9],        %[fTmp5],       %[fTmp6]                \n\t"
00492         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00493         "li.s       %[fTmp1],       0.60134488693504528054                  \n\t"
00494         "sub.s      %[fTmp2],       %[val1],        %[val6]                 \n\t"
00495         "add.s      %[val1],        %[val1],        %[val6]                 \n\t"
00496         "mul.s      %[val9],        %[val9],        %[fTmp7]                \n\t"
00497         "madd.s     %[val22],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00498         "nmsub.s    %[val25],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00499         "mul.s      %[val6],        %[fTmp1],       %[fTmp2]                \n\t"
00500         "sub.s      %[fTmp2],       %[val14],       %[val9]                 \n\t"
00501         "add.s      %[val9],        %[val14],       %[val9]                 \n\t"
00502         "mul.s      %[val25],       %[val25],       %[fTmp7]                \n\t"
00503         "sub.s      %[fTmp3],       %[val17],       %[val22]                \n\t"
00504         "add.s      %[val17],       %[val17],       %[val22]                \n\t"
00505         "mul.s      %[val14],       %[fTmp1],       %[fTmp2]                \n\t"
00506         "sub.s      %[fTmp2],       %[val30],       %[val25]                \n\t"
00507         "mul.s      %[val22],       %[fTmp1],       %[fTmp3]                \n\t"
00508         "add.s      %[val25],       %[val30],       %[val25]                \n\t"
00509         "mul.s      %[val30],       %[fTmp1],       %[fTmp2]                \n\t"
00510 
00511         : [fTmp1]  "=&f" (fTmp1),  [fTmp2]  "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3),
00512           [fTmp4]  "=&f" (fTmp4),  [fTmp5]  "=&f" (fTmp5), [fTmp6] "=&f" (fTmp6),
00513           [fTmp7]  "=&f" (fTmp7),  [fTmp8]  "=&f" (fTmp8), [fTmp9] "=&f" (fTmp9),
00514           [fTmp10] "=&f" (fTmp10), [fTmp11] "=&f" (fTmp11),
00515           [val1]  "=f" (val1),  [val6]  "=f" (val6),
00516           [val9]  "=f" (val9),  [val14] "=f" (val14),
00517           [val17] "=f" (val17), [val22] "=f" (val22),
00518           [val25] "=f" (val25), [val30] "=f" (val30)
00519         : [tab] "r" (tab)
00520     );
00521 
00522     __asm__ volatile (
00523         "lwc1       %[fTmp1],       2*4(%[tab])                             \n\t"
00524         "lwc1       %[fTmp2],       29*4(%[tab])                            \n\t"
00525         "lwc1       %[fTmp3],       13*4(%[tab])                            \n\t"
00526         "lwc1       %[fTmp4],       18*4(%[tab])                            \n\t"
00527         "li.s       %[fTmp7],       0.56694403481635770368                  \n\t"
00528         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00529         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00530         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00531         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00532         "li.s       %[fTmp10],      0.51544730992262454697                  \n\t"
00533         "li.s       %[fTmp11],      2.05778100995341155085                  \n\t"
00534         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00535         "add.s      %[val2],        %[fTmp5],       %[fTmp6]                \n\t"
00536         "sub.s      %[val13],       %[fTmp5],       %[fTmp6]                \n\t"
00537         "lwc1       %[fTmp1],       5*4(%[tab])                             \n\t"
00538         "lwc1       %[fTmp2],       26*4(%[tab])                            \n\t"
00539         "madd.s     %[val18],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00540         "nmsub.s    %[val29],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00541         "mul.s      %[val13],       %[val13],       %[fTmp7]                \n\t"
00542         "lwc1       %[fTmp3],       10*4(%[tab])                            \n\t"
00543         "lwc1       %[fTmp4],       21*4(%[tab])                            \n\t"
00544         "mul.s      %[val29],       %[val29],       %[fTmp7]                \n\t"
00545         "add.s      %[fTmp5],       %[fTmp1],       %[fTmp2]                \n\t"
00546         "sub.s      %[fTmp8],       %[fTmp1],       %[fTmp2]                \n\t"
00547         "add.s      %[fTmp6],       %[fTmp3],       %[fTmp4]                \n\t"
00548         "sub.s      %[fTmp9],       %[fTmp3],       %[fTmp4]                \n\t"
00549         "li.s       %[fTmp7],       1.06067768599034747134                  \n\t"
00550         "li.s       %[fTmp10],      0.58293496820613387367                  \n\t"
00551         "li.s       %[fTmp11],      0.97256823786196069369                  \n\t"
00552         "add.s      %[val5],        %[fTmp5],       %[fTmp6]                \n\t"
00553         "sub.s      %[val10],       %[fTmp5],       %[fTmp6]                \n\t"
00554         "mul.s      %[fTmp8],       %[fTmp8],       %[fTmp10]               \n\t"
00555         "li.s       %[fTmp1],       0.89997622313641570463                  \n\t"
00556         "sub.s      %[fTmp2],       %[val2],        %[val5]                 \n\t"
00557         "mul.s      %[val10],       %[val10],       %[fTmp7]                \n\t"
00558         "madd.s     %[val21],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00559         "nmsub.s    %[val26],       %[fTmp8],       %[fTmp9],   %[fTmp11]   \n\t"
00560         "add.s      %[val2],        %[val2],        %[val5]                 \n\t"
00561         "mul.s      %[val5],        %[fTmp1],       %[fTmp2]                \n\t"
00562         "sub.s      %[fTmp3],       %[val13],       %[val10]                \n\t"
00563         "add.s      %[val10],       %[val13],       %[val10]                \n\t"
00564         "mul.s      %[val26],       %[val26],       %[fTmp7]                \n\t"
00565         "sub.s      %[fTmp4],       %[val18],       %[val21]                \n\t"
00566         "add.s      %[val18],       %[val18],       %[val21]                \n\t"
00567         "mul.s      %[val13],       %[fTmp1],       %[fTmp3]                \n\t"
00568         "sub.s      %[fTmp2],       %[val29],       %[val26]                \n\t"
00569         "add.s      %[val26],       %[val29],       %[val26]                \n\t"
00570         "mul.s      %[val21],       %[fTmp1],       %[fTmp4]                \n\t"
00571         "mul.s      %[val29],       %[fTmp1],       %[fTmp2]                \n\t"
00572 
00573         : [fTmp1]  "=&f" (fTmp1),  [fTmp2]  "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3),
00574           [fTmp4]  "=&f" (fTmp4),  [fTmp5]  "=&f" (fTmp5), [fTmp6] "=&f" (fTmp6),
00575           [fTmp7]  "=&f" (fTmp7),  [fTmp8]  "=&f" (fTmp8), [fTmp9] "=&f" (fTmp9),
00576           [fTmp10] "=&f" (fTmp10), [fTmp11] "=&f" (fTmp11),
00577           [val2]  "=f" (val2),  [val5]  "=f" (val5),
00578           [val10] "=f" (val10), [val13] "=f" (val13),
00579           [val18] "=f" (val18), [val21] "=f" (val21),
00580           [val26] "=f" (val26), [val29] "=f" (val29)
00581         : [tab] "r" (tab)
00582     );
00583 
00584     __asm__ volatile (
00585         "li.s       %[fTmp1],       1.30656296487637652785                  \n\t"
00586         "sub.s      %[fTmp2],       %[val1],        %[val2]                 \n\t"
00587         "add.s      %[val1],        %[val1],        %[val2]                 \n\t"
00588         "sub.s      %[fTmp3],       %[val6],        %[val5]                 \n\t"
00589         "add.s      %[val5],        %[val6],        %[val5]                 \n\t"
00590         "sub.s      %[fTmp4],       %[val9],        %[val10]                \n\t"
00591         "mul.s      %[val2],        %[fTmp1],       %[fTmp2]                \n\t"
00592         "add.s      %[val9],        %[val9],        %[val10]                \n\t"
00593         "mul.s      %[val6],        %[fTmp1],       %[fTmp3]                \n\t"
00594         "sub.s      %[fTmp2],       %[val14],       %[val13]                \n\t"
00595         "mul.s      %[val10],       %[fTmp1],       %[fTmp4]                \n\t"
00596         "add.s      %[val13],       %[val14],       %[val13]                \n\t"
00597         "mul.s      %[val14],       %[fTmp1],       %[fTmp2]                \n\t"
00598 
00599         : [fTmp1] "=f"  (fTmp1), [fTmp2] "=&f" (fTmp2),
00600           [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00601           [val1]  "+f" (val1),  [val2]  "+f" (val2),
00602           [val5]  "+f" (val5),  [val6]  "+f" (val6),
00603           [val9]  "+f" (val9),  [val10] "+f" (val10),
00604           [val13] "+f" (val13), [val14] "+f" (val14)
00605         :
00606     );
00607 
00608     __asm__ volatile (
00609         "sub.s      %[fTmp2],       %[val17],       %[val18]                \n\t"
00610         "add.s      %[val17],       %[val17],       %[val18]                \n\t"
00611         "sub.s      %[fTmp3],       %[val22],       %[val21]                \n\t"
00612         "add.s      %[val21],       %[val22],       %[val21]                \n\t"
00613         "sub.s      %[fTmp4],       %[val25],       %[val26]                \n\t"
00614         "mul.s      %[val18],       %[fTmp1],       %[fTmp2]                \n\t"
00615         "add.s      %[val25],       %[val25],       %[val26]                \n\t"
00616         "mul.s      %[val22],       %[fTmp1],       %[fTmp3]                \n\t"
00617         "sub.s      %[fTmp2],       %[val30],       %[val29]                \n\t"
00618         "mul.s      %[val26],       %[fTmp1],       %[fTmp4]                \n\t"
00619         "add.s      %[val29],       %[val30],       %[val29]                \n\t"
00620         "mul.s      %[val30],       %[fTmp1],       %[fTmp2]                \n\t"
00621 
00622         : [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00623           [val17] "+f" (val17), [val18] "+f" (val18), [val21] "+f" (val21),
00624           [val22] "+f" (val22), [val25] "+f" (val25), [val26] "+f" (val26),
00625           [val29] "+f" (val29), [val30] "+f" (val30)
00626         : [fTmp1] "f" (fTmp1)
00627     );
00628 
00629     __asm__ volatile (
00630         "li.s       %[fTmp1],       0.70710678118654752439                  \n\t"
00631         "sub.s      %[fTmp2],       %[val0],        %[val1]                 \n\t"
00632         "add.s      %[val0],        %[val0],        %[val1]                 \n\t"
00633         "sub.s      %[fTmp3],       %[val3],        %[val2]                 \n\t"
00634         "add.s      %[val2],        %[val3],        %[val2]                 \n\t"
00635         "sub.s      %[fTmp4],       %[val4],        %[val5]                 \n\t"
00636         "mul.s      %[val1],        %[fTmp1],       %[fTmp2]                \n\t"
00637         "swc1       %[val0],        0(%[out])                               \n\t"
00638         "mul.s      %[val3],        %[fTmp3],       %[fTmp1]                \n\t"
00639         "add.s      %[val4],        %[val4],        %[val5]                 \n\t"
00640         "mul.s      %[val5],        %[fTmp1],       %[fTmp4]                \n\t"
00641         "swc1       %[val1],        16*4(%[out])                            \n\t"
00642         "sub.s      %[fTmp2],       %[val7],        %[val6]                 \n\t"
00643         "add.s      %[val2],        %[val2],        %[val3]                 \n\t"
00644         "swc1       %[val3],        24*4(%[out])                            \n\t"
00645         "add.s      %[val6],        %[val7],        %[val6]                 \n\t"
00646         "mul.s      %[val7],        %[fTmp1],       %[fTmp2]                \n\t"
00647         "swc1       %[val2],        8*4(%[out])                             \n\t"
00648         "add.s      %[val6],        %[val6],        %[val7]                 \n\t"
00649         "swc1       %[val7],        28*4(%[out])                            \n\t"
00650         "add.s      %[val4],        %[val4],        %[val6]                 \n\t"
00651         "add.s      %[val6],        %[val6],        %[val5]                 \n\t"
00652         "add.s      %[val5],        %[val5],        %[val7]                 \n\t"
00653         "swc1       %[val4],        4*4(%[out])                             \n\t"
00654         "swc1       %[val5],        20*4(%[out])                            \n\t"
00655         "swc1       %[val6],        12*4(%[out])                            \n\t"
00656 
00657         : [fTmp1] "=f"  (fTmp1), [fTmp2] "=&f" (fTmp2),
00658           [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00659           [val0] "+f" (val0), [val1] "+f" (val1),
00660           [val2] "+f" (val2), [val3] "+f" (val3),
00661           [val4] "+f" (val4), [val5] "+f" (val5),
00662           [val6] "+f" (val6), [val7] "+f" (val7)
00663         : [out] "r" (out)
00664     );
00665 
00666     __asm__ volatile (
00667         "sub.s      %[fTmp2],       %[val8],        %[val9]                 \n\t"
00668         "add.s      %[val8],        %[val8],        %[val9]                 \n\t"
00669         "sub.s      %[fTmp3],       %[val11],       %[val10]                \n\t"
00670         "add.s      %[val10],       %[val11],       %[val10]                \n\t"
00671         "sub.s      %[fTmp4],       %[val12],       %[val13]                \n\t"
00672         "mul.s      %[val9],        %[fTmp1],       %[fTmp2]                \n\t"
00673         "add.s      %[val12],       %[val12],       %[val13]                \n\t"
00674         "mul.s      %[val11],       %[fTmp1],       %[fTmp3]                \n\t"
00675         "sub.s      %[fTmp2],       %[val15],       %[val14]                \n\t"
00676         "mul.s      %[val13],       %[fTmp1],       %[fTmp4]                \n\t"
00677         "add.s      %[val14],       %[val15],       %[val14]                \n\t"
00678         "add.s      %[val10],       %[val10],       %[val11]                \n\t"
00679         "mul.s      %[val15],       %[fTmp1],       %[fTmp2]                \n\t"
00680         "add.s      %[val14],       %[val14],       %[val15]                \n\t"
00681         "add.s      %[val12],       %[val12],       %[val14]                \n\t"
00682         "add.s      %[val14],       %[val14],       %[val13]                \n\t"
00683         "add.s      %[val13],       %[val13],       %[val15]                \n\t"
00684         "add.s      %[val8],        %[val8],        %[val12]                \n\t"
00685         "add.s      %[val12],       %[val12],       %[val10]                \n\t"
00686         "add.s      %[val10],       %[val10],       %[val14]                \n\t"
00687         "add.s      %[val14],       %[val14],       %[val9]                 \n\t"
00688         "add.s      %[val9],        %[val9],        %[val13]                \n\t"
00689         "add.s      %[val13],       %[val13],       %[val11]                \n\t"
00690         "add.s      %[val11],       %[val11],       %[val15]                \n\t"
00691         "swc1       %[val8],         2*4(%[out])                            \n\t"
00692         "swc1       %[val9],        18*4(%[out])                            \n\t"
00693         "swc1       %[val10],       10*4(%[out])                            \n\t"
00694         "swc1       %[val11],       26*4(%[out])                            \n\t"
00695         "swc1       %[val12],        6*4(%[out])                            \n\t"
00696         "swc1       %[val13],       22*4(%[out])                            \n\t"
00697         "swc1       %[val14],       14*4(%[out])                            \n\t"
00698         "swc1       %[val15],       30*4(%[out])                            \n\t"
00699 
00700         : [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00701           [val8]  "+f" (val8),  [val9]  "+f" (val9),  [val10] "+f" (val10),
00702           [val11] "+f" (val11), [val12] "+f" (val12), [val13] "+f" (val13),
00703           [val14] "+f" (val14), [val15] "+f" (val15)
00704         : [fTmp1] "f" (fTmp1), [out] "r" (out)
00705     );
00706 
00707     __asm__ volatile (
00708         "sub.s      %[fTmp2],       %[val16],       %[val17]                \n\t"
00709         "add.s      %[val16],       %[val16],       %[val17]                \n\t"
00710         "sub.s      %[fTmp3],       %[val19],       %[val18]                \n\t"
00711         "add.s      %[val18],       %[val19],       %[val18]                \n\t"
00712         "sub.s      %[fTmp4],       %[val20],       %[val21]                \n\t"
00713         "mul.s      %[val17],       %[fTmp1],       %[fTmp2]                \n\t"
00714         "add.s      %[val20],       %[val20],       %[val21]                \n\t"
00715         "mul.s      %[val19],       %[fTmp1],       %[fTmp3]                \n\t"
00716         "sub.s      %[fTmp2],       %[val23],       %[val22]                \n\t"
00717         "mul.s      %[val21],       %[fTmp1],       %[fTmp4]                \n\t"
00718         "add.s      %[val22],       %[val23],       %[val22]                \n\t"
00719         "add.s      %[val18],       %[val18],       %[val19]                \n\t"
00720         "mul.s      %[val23],       %[fTmp1],       %[fTmp2]                \n\t"
00721         "add.s      %[val22],       %[val22],       %[val23]                \n\t"
00722         "add.s      %[val20],       %[val20],       %[val22]                \n\t"
00723         "add.s      %[val22],       %[val22],       %[val21]                \n\t"
00724         "add.s      %[val21],       %[val21],       %[val23]                \n\t"
00725 
00726         : [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00727           [val16] "+f" (val16), [val17] "+f" (val17), [val18] "+f" (val18),
00728           [val19] "+f" (val19), [val20] "+f" (val20), [val21] "+f" (val21),
00729           [val22] "+f" (val22), [val23] "+f" (val23)
00730         : [fTmp1] "f" (fTmp1)
00731     );
00732 
00733     __asm__ volatile (
00734         "sub.s      %[fTmp2],       %[val24],       %[val25]                \n\t"
00735         "add.s      %[val24],       %[val24],       %[val25]                \n\t"
00736         "sub.s      %[fTmp3],       %[val27],       %[val26]                \n\t"
00737         "add.s      %[val26],       %[val27],       %[val26]                \n\t"
00738         "sub.s      %[fTmp4],       %[val28],       %[val29]                \n\t"
00739         "mul.s      %[val25],       %[fTmp1],       %[fTmp2]                \n\t"
00740         "add.s      %[val28],       %[val28],       %[val29]                \n\t"
00741         "mul.s      %[val27],       %[fTmp1],       %[fTmp3]                \n\t"
00742         "sub.s      %[fTmp2],       %[val31],       %[val30]                \n\t"
00743         "mul.s      %[val29],       %[fTmp1],       %[fTmp4]                \n\t"
00744         "add.s      %[val30],       %[val31],       %[val30]                \n\t"
00745         "add.s      %[val26],       %[val26],       %[val27]                \n\t"
00746         "mul.s      %[val31],       %[fTmp1],       %[fTmp2]                \n\t"
00747         "add.s      %[val30],       %[val30],       %[val31]                \n\t"
00748         "add.s      %[val28],       %[val28],       %[val30]                \n\t"
00749         "add.s      %[val30],       %[val30],       %[val29]                \n\t"
00750         "add.s      %[val29],       %[val29],       %[val31]                \n\t"
00751         "add.s      %[val24],       %[val24],       %[val28]                \n\t"
00752         "add.s      %[val28],       %[val28],       %[val26]                \n\t"
00753         "add.s      %[val26],       %[val26],       %[val30]                \n\t"
00754         "add.s      %[val30],       %[val30],       %[val25]                \n\t"
00755         "add.s      %[val25],       %[val25],       %[val29]                \n\t"
00756         "add.s      %[val29],       %[val29],       %[val27]                \n\t"
00757         "add.s      %[val27],       %[val27],       %[val31]                \n\t"
00758 
00759         : [fTmp2] "=&f" (fTmp2), [fTmp3] "=&f" (fTmp3), [fTmp4] "=&f" (fTmp4),
00760           [val24] "+f" (val24), [val25] "+f" (val25), [val26] "+f" (val26),
00761           [val27] "+f" (val27), [val28] "+f" (val28), [val29] "+f" (val29),
00762           [val30] "+f" (val30), [val31] "+f" (val31)
00763         : [fTmp1] "f" (fTmp1)
00764     );
00765 
00766     out[ 1] = val16 + val24;
00767     out[17] = val17 + val25;
00768     out[ 9] = val18 + val26;
00769     out[25] = val19 + val27;
00770     out[ 5] = val20 + val28;
00771     out[21] = val21 + val29;
00772     out[13] = val22 + val30;
00773     out[29] = val23 + val31;
00774     out[ 3] = val24 + val20;
00775     out[19] = val25 + val21;
00776     out[11] = val26 + val22;
00777     out[27] = val27 + val23;
00778     out[ 7] = val28 + val18;
00779     out[23] = val29 + val19;
00780     out[15] = val30 + val17;
00781     out[31] = val31;
00782 }
00783 
00784 static void imdct36_mips_float(float *out, float *buf, float *in, float *win)
00785 {
00786     float t0, t1, t2, t3, s0, s1, s2, s3;
00787     float tmp[18];
00788     /* temporary variables */
00789     float in1, in2, in3, in4, in5, in6;
00790     float out1, out2, out3, out4, out5;
00791     float c1, c2, c3, c4, c5, c6, c7, c8, c9;
00792 
00801     /* loop 1 and 2 */
00802     __asm__ volatile (
00803         "lwc1   %[in1],  17*4(%[in])                                    \t\n"
00804         "lwc1   %[in2],  16*4(%[in])                                    \t\n"
00805         "lwc1   %[in3],  15*4(%[in])                                    \t\n"
00806         "lwc1   %[in4],  14*4(%[in])                                    \t\n"
00807         "lwc1   %[in5],  13*4(%[in])                                    \t\n"
00808         "lwc1   %[in6],  12*4(%[in])                                    \t\n"
00809         "add.s  %[out1], %[in1],  %[in2]                                \t\n"
00810         "add.s  %[out2], %[in2],  %[in3]                                \t\n"
00811         "add.s  %[out3], %[in3],  %[in4]                                \t\n"
00812         "add.s  %[out4], %[in4],  %[in5]                                \t\n"
00813         "add.s  %[out5], %[in5],  %[in6]                                \t\n"
00814         "lwc1   %[in1],  11*4(%[in])                                    \t\n"
00815         "swc1   %[out2], 16*4(%[in])                                    \t\n"
00816         "add.s  %[out1], %[out1], %[out3]                               \t\n"
00817         "swc1   %[out4], 14*4(%[in])                                    \t\n"
00818         "add.s  %[out3], %[out3], %[out5]                               \t\n"
00819         "lwc1   %[in2],  10*4(%[in])                                    \t\n"
00820         "lwc1   %[in3],  9*4(%[in])                                     \t\n"
00821         "swc1   %[out1], 17*4(%[in])                                    \t\n"
00822         "lwc1   %[in4],  8*4(%[in])                                     \t\n"
00823         "swc1   %[out3], 15*4(%[in])                                    \t\n"
00824         "add.s  %[out1], %[in6],  %[in1]                                \t\n"
00825         "add.s  %[out2], %[in1],  %[in2]                                \t\n"
00826         "add.s  %[out3], %[in2],  %[in3]                                \t\n"
00827         "add.s  %[out4], %[in3],  %[in4]                                \t\n"
00828         "lwc1   %[in5],  7*4(%[in])                                     \t\n"
00829         "swc1   %[out1], 12*4(%[in])                                    \t\n"
00830         "add.s  %[out5], %[out5], %[out2]                               \t\n"
00831         "swc1   %[out3], 10*4(%[in])                                    \t\n"
00832         "add.s  %[out2], %[out2], %[out4]                               \t\n"
00833         "lwc1   %[in6],  6*4(%[in])                                     \t\n"
00834         "lwc1   %[in1],  5*4(%[in])                                     \t\n"
00835         "swc1   %[out5], 13*4(%[in])                                    \t\n"
00836         "lwc1   %[in2],  4*4(%[in])                                     \t\n"
00837         "swc1   %[out2], 11*4(%[in])                                    \t\n"
00838         "add.s  %[out5], %[in4],  %[in5]                                \t\n"
00839         "add.s  %[out1], %[in5],  %[in6]                                \t\n"
00840         "add.s  %[out2], %[in6],  %[in1]                                \t\n"
00841         "add.s  %[out3], %[in1],  %[in2]                                \t\n"
00842         "lwc1   %[in3],  3*4(%[in])                                     \t\n"
00843         "swc1   %[out5], 8*4(%[in])                                     \t\n"
00844         "add.s  %[out4], %[out4], %[out1]                               \t\n"
00845         "swc1   %[out2], 6*4(%[in])                                     \t\n"
00846         "add.s  %[out1], %[out1], %[out3]                               \t\n"
00847         "lwc1   %[in4],  2*4(%[in])                                     \t\n"
00848         "lwc1   %[in5],  1*4(%[in])                                     \t\n"
00849         "swc1   %[out4], 9*4(%[in])                                     \t\n"
00850         "lwc1   %[in6],  0(%[in])                                       \t\n"
00851         "swc1   %[out1], 7*4(%[in])                                     \t\n"
00852         "add.s  %[out4], %[in2],  %[in3]                                \t\n"
00853         "add.s  %[out5], %[in3],  %[in4]                                \t\n"
00854         "add.s  %[out1], %[in4],  %[in5]                                \t\n"
00855         "add.s  %[out2], %[in5],  %[in6]                                \t\n"
00856         "swc1   %[out4], 4*4(%[in])                                     \t\n"
00857         "add.s  %[out3], %[out3], %[out5]                               \t\n"
00858         "swc1   %[out1], 2*4(%[in])                                     \t\n"
00859         "add.s  %[out5], %[out5], %[out2]                               \t\n"
00860         "swc1   %[out2], 1*4(%[in])                                     \t\n"
00861         "swc1   %[out3], 5*4(%[in])                                     \t\n"
00862         "swc1   %[out5], 3*4(%[in])                                     \t\n"
00863 
00864         : [in1] "=&f" (in1), [in2] "=&f" (in2),
00865           [in3] "=&f" (in3), [in4] "=&f" (in4),
00866           [in5] "=&f" (in5), [in6] "=&f" (in6),
00867           [out1] "=&f" (out1), [out2] "=&f" (out2),
00868           [out3] "=&f" (out3), [out4] "=&f" (out4),
00869           [out5] "=&f" (out5)
00870         : [in] "r" (in)
00871     );
00872 
00873     /* loop 3 */
00874     __asm__ volatile (
00875         "li.s    %[c1],   0.5                                           \t\n"
00876         "lwc1    %[in1],  8*4(%[in])                                    \t\n"
00877         "lwc1    %[in2],  16*4(%[in])                                   \t\n"
00878         "lwc1    %[in3],  4*4(%[in])                                    \t\n"
00879         "lwc1    %[in4],  0(%[in])                                      \t\n"
00880         "lwc1    %[in5],  12*4(%[in])                                   \t\n"
00881         "li.s    %[c2],   0.93969262078590838405                        \t\n"
00882         "add.s   %[t2],   %[in1],  %[in2]                               \t\n"
00883         "add.s   %[t0],   %[in1],  %[in3]                               \t\n"
00884         "li.s    %[c3],   -0.76604444311897803520                       \t\n"
00885         "madd.s  %[t3],   %[in4],  %[in5], %[c1]                        \t\n"
00886         "sub.s   %[t1],   %[in4],  %[in5]                               \t\n"
00887         "sub.s   %[t2],   %[t2],   %[in3]                               \t\n"
00888         "mul.s   %[t0],   %[t0],   %[c2]                                \t\n"
00889         "li.s    %[c4],   -0.17364817766693034885                       \t\n"
00890         "li.s    %[c5],   -0.86602540378443864676                       \t\n"
00891         "li.s    %[c6],   0.98480775301220805936                        \t\n"
00892         "nmsub.s %[out1], %[t1],   %[t2],  %[c1]                        \t\n"
00893         "add.s   %[out2], %[t1],   %[t2]                                \t\n"
00894         "add.s   %[t2],   %[in2],  %[in3]                               \t\n"
00895         "sub.s   %[t1],   %[in1],  %[in2]                               \t\n"
00896         "sub.s   %[out3], %[t3],   %[t0]                                \t\n"
00897         "swc1    %[out1], 6*4(%[tmp])                                   \t\n"
00898         "swc1    %[out2], 16*4(%[tmp])                                  \t\n"
00899         "mul.s   %[t2],   %[t2],   %[c3]                                \t\n"
00900         "mul.s   %[t1],   %[t1],   %[c4]                                \t\n"
00901         "add.s   %[out1], %[t3],   %[t0]                                \t\n"
00902         "lwc1    %[in1],  10*4(%[in])                                   \t\n"
00903         "lwc1    %[in2],  14*4(%[in])                                   \t\n"
00904         "sub.s   %[out3], %[out3], %[t2]                                \t\n"
00905         "add.s   %[out2], %[t3],   %[t2]                                \t\n"
00906         "add.s   %[out1], %[out1], %[t1]                                \t\n"
00907         "lwc1    %[in3],  2*4(%[in])                                    \t\n"
00908         "lwc1    %[in4],  6*4(%[in])                                    \t\n"
00909         "swc1    %[out3], 10*4(%[tmp])                                  \t\n"
00910         "sub.s   %[out2], %[out2], %[t1]                                \t\n"
00911         "swc1    %[out1], 2*4(%[tmp])                                   \t\n"
00912         "add.s   %[out1], %[in1],  %[in2]                               \t\n"
00913         "add.s   %[t2],   %[in1],  %[in3]                               \t\n"
00914         "sub.s   %[t3],   %[in1],  %[in2]                               \t\n"
00915         "swc1    %[out2], 14*4(%[tmp])                                  \t\n"
00916         "li.s    %[c7],   -0.34202014332566873304                       \t\n"
00917         "sub.s   %[out1], %[out1], %[in3]                               \t\n"
00918         "mul.s   %[t2],   %[t2],   %[c6]                                \t\n"
00919         "mul.s   %[t3],   %[t3],   %[c7]                                \t\n"
00920         "li.s    %[c8],   0.86602540378443864676                        \t\n"
00921         "mul.s   %[t0],   %[in4],  %[c8]                                \t\n"
00922         "mul.s   %[out1], %[out1], %[c5]                                \t\n"
00923         "add.s   %[t1],   %[in2],  %[in3]                               \t\n"
00924         "li.s    %[c9],   -0.64278760968653932632                       \t\n"
00925         "add.s   %[out2], %[t2],   %[t3]                                \t\n"
00926         "lwc1    %[in1],  9*4(%[in])                                    \t\n"
00927         "swc1    %[out1], 4*4(%[tmp])                                   \t\n"
00928         "mul.s   %[t1],   %[t1],   %[c9]                                \t\n"
00929         "lwc1    %[in2],  17*4(%[in])                                   \t\n"
00930         "add.s   %[out2], %[out2], %[t0]                                \t\n"
00931         "lwc1    %[in3],  5*4(%[in])                                    \t\n"
00932         "lwc1    %[in4],  1*4(%[in])                                    \t\n"
00933         "add.s   %[out3], %[t2],   %[t1]                                \t\n"
00934         "sub.s   %[out1], %[t3],   %[t1]                                \t\n"
00935         "swc1    %[out2], 0(%[tmp])                                     \t\n"
00936         "lwc1    %[in5],  13*4(%[in])                                   \t\n"
00937         "add.s   %[t2],   %[in1],  %[in2]                               \t\n"
00938         "sub.s   %[out3], %[out3], %[t0]                                \t\n"
00939         "sub.s   %[out1], %[out1], %[t0]                                \t\n"
00940         "add.s   %[t0],   %[in1],  %[in3]                               \t\n"
00941         "madd.s  %[t3],   %[in4],  %[in5], %[c1]                        \t\n"
00942         "sub.s   %[t2],   %[t2],   %[in3]                               \t\n"
00943         "swc1    %[out3], 12*4(%[tmp])                                  \t\n"
00944         "swc1    %[out1], 8*4(%[tmp])                                   \t\n"
00945         "sub.s   %[t1],   %[in4],  %[in5]                               \t\n"
00946         "mul.s   %[t0],   %[t0],   %[c2]                                \t\n"
00947         "nmsub.s %[out1], %[t1],   %[t2],  %[c1]                        \t\n"
00948         "add.s   %[out2], %[t1],   %[t2]                                \t\n"
00949         "add.s   %[t2],   %[in2],  %[in3]                               \t\n"
00950         "sub.s   %[t1],   %[in1],  %[in2]                               \t\n"
00951         "sub.s   %[out3], %[t3],   %[t0]                                \t\n"
00952         "swc1    %[out1], 7*4(%[tmp])                                   \t\n"
00953         "swc1    %[out2], 17*4(%[tmp])                                  \t\n"
00954         "mul.s   %[t2],   %[t2],   %[c3]                                \t\n"
00955         "mul.s   %[t1],   %[t1],   %[c4]                                \t\n"
00956         "add.s   %[out1], %[t3],   %[t0]                                \t\n"
00957         "lwc1    %[in1],  11*4(%[in])                                   \t\n"
00958         "lwc1    %[in2],  15*4(%[in])                                   \t\n"
00959         "sub.s   %[out3], %[out3], %[t2]                                \t\n"
00960         "add.s   %[out2], %[t3],   %[t2]                                \t\n"
00961         "add.s   %[out1], %[out1], %[t1]                                \t\n"
00962         "lwc1    %[in3],  3*4(%[in])                                    \t\n"
00963         "lwc1    %[in4],  7*4(%[in])                                    \t\n"
00964         "swc1    %[out3], 11*4(%[tmp])                                  \t\n"
00965         "sub.s   %[out2], %[out2], %[t1]                                \t\n"
00966         "swc1    %[out1], 3*4(%[tmp])                                   \t\n"
00967         "add.s   %[out3], %[in1],  %[in2]                               \t\n"
00968         "add.s   %[t2],   %[in1],  %[in3]                               \t\n"
00969         "sub.s   %[t3],   %[in1],  %[in2]                               \t\n"
00970         "swc1    %[out2], 15*4(%[tmp])                                  \t\n"
00971         "mul.s   %[t0],   %[in4],  %[c8]                                \t\n"
00972         "sub.s   %[out3], %[out3], %[in3]                               \t\n"
00973         "mul.s   %[t2],   %[t2],   %[c6]                                \t\n"
00974         "mul.s   %[t3],   %[t3],   %[c7]                                \t\n"
00975         "add.s   %[t1],   %[in2],  %[in3]                               \t\n"
00976         "mul.s   %[out3], %[out3], %[c5]                                \t\n"
00977         "add.s   %[out1], %[t2],   %[t3]                                \t\n"
00978         "mul.s   %[t1],   %[t1],   %[c9]                                \t\n"
00979         "swc1    %[out3], 5*4(%[tmp])                                   \t\n"
00980         "add.s   %[out1], %[out1], %[t0]                                \t\n"
00981         "add.s   %[out2], %[t2],   %[t1]                                \t\n"
00982         "sub.s   %[out3], %[t3],   %[t1]                                \t\n"
00983         "swc1    %[out1], 1*4(%[tmp])                                   \t\n"
00984         "sub.s   %[out2], %[out2], %[t0]                                \t\n"
00985         "sub.s   %[out3], %[out3], %[t0]                                \t\n"
00986         "swc1    %[out2], 13*4(%[tmp])                                  \t\n"
00987         "swc1    %[out3], 9*4(%[tmp])                                   \t\n"
00988 
00989         : [t0] "=&f" (t0), [t1] "=&f" (t1),
00990           [t2] "=&f" (t2), [t3] "=&f" (t3),
00991           [in1] "=&f" (in1), [in2] "=&f" (in2),
00992           [in3] "=&f" (in3), [in4] "=&f" (in4),
00993           [in5] "=&f" (in5),
00994           [out1] "=&f" (out1), [out2] "=&f" (out2),
00995           [out3] "=&f" (out3),
00996           [c1] "=&f" (c1), [c2] "=&f" (c2),
00997           [c3] "=&f" (c3), [c4] "=&f" (c4),
00998           [c5] "=&f" (c5), [c6] "=&f" (c6),
00999           [c7] "=&f" (c7), [c8] "=&f" (c8),
01000           [c9] "=&f" (c9)
01001         : [in] "r" (in), [tmp] "r" (tmp)
01002     );
01003 
01004     /* loop 4 */
01005     __asm__ volatile (
01006         "lwc1   %[in1],  2*4(%[tmp])                                    \t\n"
01007         "lwc1   %[in2],  0(%[tmp])                                      \t\n"
01008         "lwc1   %[in3],  3*4(%[tmp])                                    \t\n"
01009         "lwc1   %[in4],  1*4(%[tmp])                                    \t\n"
01010         "li.s   %[c1],   0.50190991877167369479                         \t\n"
01011         "li.s   %[c2],   5.73685662283492756461                         \t\n"
01012         "add.s  %[s0],   %[in1], %[in2]                                 \t\n"
01013         "sub.s  %[s2],   %[in1], %[in2]                                 \t\n"
01014         "add.s  %[s1],   %[in3], %[in4]                                 \t\n"
01015         "sub.s  %[s3],   %[in3], %[in4]                                 \t\n"
01016         "lwc1   %[in1],  9*4(%[win])                                    \t\n"
01017         "lwc1   %[in2],  4*9*4(%[buf])                                  \t\n"
01018         "lwc1   %[in3],  8*4(%[win])                                    \t\n"
01019         "mul.s  %[s1],   %[s1],  %[c1]                                  \t\n"
01020         "mul.s  %[s3],   %[s3],  %[c2]                                  \t\n"
01021         "lwc1   %[in4],  4*8*4(%[buf])                                  \t\n"
01022         "lwc1   %[in5],  29*4(%[win])                                   \t\n"
01023         "lwc1   %[in6],  28*4(%[win])                                   \t\n"
01024         "add.s  %[t0],   %[s0],  %[s1]                                  \t\n"
01025         "sub.s  %[t1],   %[s0],  %[s1]                                  \t\n"
01026         "li.s   %[c1],   0.51763809020504152469                         \t\n"
01027         "li.s   %[c2],   1.93185165257813657349                         \t\n"
01028         "mul.s  %[out3], %[in5], %[t0]                                  \t\n"
01029         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01030         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01031         "mul.s  %[out4], %[in6], %[t0]                                  \t\n"
01032         "add.s  %[t0],   %[s2],  %[s3]                                  \t\n"
01033         "swc1   %[out3], 4*9*4(%[buf])                                  \t\n"
01034         "swc1   %[out1], 288*4(%[out])                                  \t\n"
01035         "swc1   %[out2], 256*4(%[out])                                  \t\n"
01036         "swc1   %[out4], 4*8*4(%[buf])                                  \t\n"
01037         "sub.s  %[t1],   %[s2],  %[s3]                                  \t\n"
01038         "lwc1   %[in1],  17*4(%[win])                                   \t\n"
01039         "lwc1   %[in2],  4*17*4(%[buf])                                 \t\n"
01040         "lwc1   %[in3],  0(%[win])                                      \t\n"
01041         "lwc1   %[in4],  0(%[buf])                                      \t\n"
01042         "lwc1   %[in5],  37*4(%[win])                                   \t\n"
01043         "lwc1   %[in6],  20*4(%[win])                                   \t\n"
01044         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01045         "lwc1   %[in1],  6*4(%[tmp])                                    \t\n"
01046         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01047         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01048         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01049         "swc1   %[out1], 544*4(%[out])                                  \t\n"
01050         "lwc1   %[in2],  4*4(%[tmp])                                    \t\n"
01051         "swc1   %[out2], 0(%[out])                                      \t\n"
01052         "swc1   %[out3], 4*17*4(%[buf])                                 \t\n"
01053         "swc1   %[out4], 0(%[buf])                                      \t\n"
01054         "lwc1   %[in3],  7*4(%[tmp])                                    \t\n"
01055         "add.s  %[s0],   %[in1], %[in2]                                 \t\n"
01056         "sub.s  %[s2],   %[in1], %[in2]                                 \t\n"
01057         "lwc1   %[in4],  5*4(%[tmp])                                    \t\n"
01058         "add.s  %[s1],   %[in3], %[in4]                                 \t\n"
01059         "sub.s  %[s3],   %[in3], %[in4]                                 \t\n"
01060         "lwc1   %[in1],  10*4(%[win])                                   \t\n"
01061         "lwc1   %[in2],  4*10*4(%[buf])                                 \t\n"
01062         "lwc1   %[in3],  7*4(%[win])                                    \t\n"
01063         "mul.s  %[s1],   %[s1],  %[c1]                                  \t\n"
01064         "mul.s  %[s3],   %[s3],  %[c2]                                  \t\n"
01065         "add.s  %[t0],   %[s0],  %[s1]                                  \t\n"
01066         "sub.s  %[t1],   %[s0],  %[s1]                                  \t\n"
01067         "lwc1   %[in4],  4*7*4(%[buf])                                  \t\n"
01068         "lwc1   %[in5],  30*4(%[win])                                   \t\n"
01069         "lwc1   %[in6],  27*4(%[win])                                   \t\n"
01070         "li.s   %[c1],   0.55168895948124587824                         \t\n"
01071         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01072         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01073         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01074         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01075         "add.s  %[t0],   %[s2],  %[s3]                                  \t\n"
01076         "swc1   %[out1], 320*4(%[out])                                  \t\n"
01077         "swc1   %[out2], 224*4(%[out])                                  \t\n"
01078         "swc1   %[out3], 4*10*4(%[buf])                                 \t\n"
01079         "swc1   %[out4], 4*7*4(%[buf])                                  \t\n"
01080         "sub.s  %[t1],   %[s2],  %[s3]                                  \t\n"
01081         "lwc1   %[in1],  16*4(%[win])                                   \t\n"
01082         "lwc1   %[in2],  4*16*4(%[buf])                                 \t\n"
01083         "lwc1   %[in3],  1*4(%[win])                                    \t\n"
01084         "lwc1   %[in4],  4*1*4(%[buf])                                  \t\n"
01085         "lwc1   %[in5],  36*4(%[win])                                   \t\n"
01086         "lwc1   %[in6],  21*4(%[win])                                   \t\n"
01087         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01088         "lwc1   %[in1],  10*4(%[tmp])                                   \t\n"
01089         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01090         "mul.s  %[out3], %[in5], %[t0]                                  \t\n"
01091         "mul.s  %[out4], %[in6], %[t0]                                  \t\n"
01092         "swc1   %[out1], 512*4(%[out])                                  \t\n"
01093         "lwc1   %[in2],  8*4(%[tmp])                                    \t\n"
01094         "swc1   %[out2], 32*4(%[out])                                   \t\n"
01095         "swc1   %[out3], 4*16*4(%[buf])                                 \t\n"
01096         "swc1   %[out4], 4*1*4(%[buf])                                  \t\n"
01097         "li.s   %[c2],   1.18310079157624925896                         \t\n"
01098         "add.s  %[s0],   %[in1], %[in2]                                 \t\n"
01099         "sub.s  %[s2],   %[in1], %[in2]                                 \t\n"
01100         "lwc1   %[in3],  11*4(%[tmp])                                   \t\n"
01101         "lwc1   %[in4],  9*4(%[tmp])                                    \t\n"
01102         "add.s  %[s1],   %[in3], %[in4]                                 \t\n"
01103         "sub.s  %[s3],   %[in3], %[in4]                                 \t\n"
01104         "lwc1   %[in1],  11*4(%[win])                                   \t\n"
01105         "lwc1   %[in2],  4*11*4(%[buf])                                 \t\n"
01106         "lwc1   %[in3],  6*4(%[win])                                    \t\n"
01107         "mul.s  %[s1],   %[s1],  %[c1]                                  \t\n"
01108         "mul.s  %[s3],   %[s3],  %[c2]                                  \t\n"
01109         "lwc1   %[in4],  4*6*4(%[buf])                                  \t\n"
01110         "lwc1   %[in5],  31*4(%[win])                                   \t\n"
01111         "lwc1   %[in6],  26*4(%[win])                                   \t\n"
01112         "add.s  %[t0],   %[s0],  %[s1]                                  \t\n"
01113         "sub.s  %[t1],   %[s0],  %[s1]                                  \t\n"
01114         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01115         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01116         "add.s  %[t0],   %[s2],  %[s3]                                  \t\n"
01117         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01118         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01119         "swc1   %[out3], 4*11*4(%[buf])                                 \t\n"
01120         "swc1   %[out4], 4*6*4(%[buf])                                  \t\n"
01121         "sub.s  %[t1],   %[s2],  %[s3]                                  \t\n"
01122         "swc1   %[out1], 352*4(%[out])                                  \t\n"
01123         "swc1   %[out2], 192*4(%[out])                                  \t\n"
01124         "lwc1   %[in1],  15*4(%[win])                                   \t\n"
01125         "lwc1   %[in2],  4*15*4(%[buf])                                 \t\n"
01126         "lwc1   %[in3],  2*4(%[win])                                    \t\n"
01127         "lwc1   %[in4],  4*2*4(%[buf])                                  \t\n"
01128         "lwc1   %[in5],  35*4(%[win])                                   \t\n"
01129         "lwc1   %[in6],  22*4(%[win])                                   \t\n"
01130         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01131         "lwc1   %[in1],  14*4(%[tmp])                                   \t\n"
01132         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01133         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01134         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01135         "swc1   %[out1], 480*4(%[out])                                  \t\n"
01136         "lwc1   %[in2],  12*4(%[tmp])                                   \t\n"
01137         "swc1   %[out2], 64*4(%[out])                                   \t\n"
01138         "swc1   %[out3], 4*15*4(%[buf])                                 \t\n"
01139         "swc1   %[out4], 4*2*4(%[buf])                                  \t\n"
01140         "lwc1   %[in3],  15*4(%[tmp])                                   \t\n"
01141         "add.s  %[s0],   %[in1], %[in2]                                 \t\n"
01142         "sub.s  %[s2],   %[in1], %[in2]                                 \t\n"
01143         "lwc1   %[in4],  13*4(%[tmp])                                   \t\n"
01144         "li.s   %[c1],   0.61038729438072803416                         \t\n"
01145         "li.s   %[c2],   0.87172339781054900991                         \t\n"
01146         "add.s  %[s1],   %[in3], %[in4]                                 \t\n"
01147         "sub.s  %[s3],   %[in3], %[in4]                                 \t\n"
01148         "lwc1   %[in1],  12*4(%[win])                                   \t\n"
01149         "lwc1   %[in2],  4*12*4(%[buf])                                 \t\n"
01150         "lwc1   %[in3],  5*4(%[win])                                    \t\n"
01151         "mul.s  %[s1],   %[s1],  %[c1]                                  \t\n"
01152         "mul.s  %[s3],   %[s3],  %[c2]                                  \t\n"
01153         "lwc1   %[in4],  4*5*4(%[buf])                                  \t\n"
01154         "lwc1   %[in5],  32*4(%[win])                                   \t\n"
01155         "lwc1   %[in6],  25*4(%[win])                                   \t\n"
01156         "add.s  %[t0],   %[s0],  %[s1]                                  \t\n"
01157         "sub.s  %[t1],   %[s0],  %[s1]                                  \t\n"
01158         "lwc1   %[s0],   16*4(%[tmp])                                   \t\n"
01159         "lwc1   %[s1],   17*4(%[tmp])                                   \t\n"
01160         "li.s   %[c1],   0.70710678118654752439                         \t\n"
01161         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01162         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01163         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01164         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01165         "add.s  %[t0],   %[s2],  %[s3]                                  \t\n"
01166         "swc1   %[out3], 4*12*4(%[buf])                                 \t\n"
01167         "swc1   %[out1], 384*4(%[out])                                  \t\n"
01168         "swc1   %[out2], 160*4(%[out])                                  \t\n"
01169         "swc1   %[out4], 4*5*4(%[buf])                                  \t\n"
01170         "sub.s  %[t1],   %[s2],  %[s3]                                  \t\n"
01171         "lwc1   %[in1],  14*4(%[win])                                   \t\n"
01172         "lwc1   %[in2],  4*14*4(%[buf])                                 \t\n"
01173         "lwc1   %[in3],  3*4(%[win])                                    \t\n"
01174         "lwc1   %[in4],  4*3*4(%[buf])                                  \t\n"
01175         "lwc1   %[in5],  34*4(%[win])                                   \t\n"
01176         "lwc1   %[in6],  23*4(%[win])                                   \t\n"
01177         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01178         "mul.s  %[s1],   %[s1],  %[c1]                                  \t\n"
01179         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01180         "mul.s  %[out3], %[in5], %[t0]                                  \t\n"
01181         "mul.s  %[out4], %[in6], %[t0]                                  \t\n"
01182         "swc1   %[out1], 448*4(%[out])                                  \t\n"
01183         "add.s  %[t0],   %[s0],  %[s1]                                  \t\n"
01184         "swc1   %[out2], 96*4(%[out])                                   \t\n"
01185         "swc1   %[out3], 4*14*4(%[buf])                                 \t\n"
01186         "swc1   %[out4], 4*3*4(%[buf])                                  \t\n"
01187         "sub.s  %[t1],   %[s0],  %[s1]                                  \t\n"
01188         "lwc1   %[in1],  13*4(%[win])                                   \t\n"
01189         "lwc1   %[in2],  4*13*4(%[buf])                                 \t\n"
01190         "lwc1   %[in3],  4*4(%[win])                                    \t\n"
01191         "lwc1   %[in4],  4*4*4(%[buf])                                  \t\n"
01192         "lwc1   %[in5],  33*4(%[win])                                   \t\n"
01193         "lwc1   %[in6],  24*4(%[win])                                   \t\n"
01194         "madd.s %[out1], %[in2], %[in1], %[t1]                          \t\n"
01195         "madd.s %[out2], %[in4], %[in3], %[t1]                          \t\n"
01196         "mul.s  %[out3], %[t0],  %[in5]                                 \t\n"
01197         "mul.s  %[out4], %[t0],  %[in6]                                 \t\n"
01198         "swc1   %[out1], 416*4(%[out])                                  \t\n"
01199         "swc1   %[out2], 128*4(%[out])                                  \t\n"
01200         "swc1   %[out3], 4*13*4(%[buf])                                 \t\n"
01201         "swc1   %[out4], 4*4*4(%[buf])                                  \t\n"
01202 
01203         : [c1] "=&f" (c1), [c2] "=&f" (c2),
01204           [in1] "=&f" (in1), [in2] "=&f" (in2),
01205           [in3] "=&f" (in3), [in4] "=&f" (in4),
01206           [in5] "=&f" (in5), [in6] "=&f" (in6),
01207           [out1] "=&f" (out1), [out2] "=&f" (out2),
01208           [out3] "=&f" (out3), [out4] "=&f" (out4),
01209           [t0] "=&f" (t0), [t1] "=&f" (t1),
01210           [t2] "=&f" (t2), [t3] "=&f" (t3),
01211           [s0] "=&f" (s0), [s1] "=&f" (s1),
01212           [s2] "=&f" (s2), [s3] "=&f" (s3)
01213         : [tmp] "r" (tmp), [win] "r" (win),
01214           [buf] "r" (buf), [out] "r" (out)
01215     );
01216 }
01217 
01218 static void ff_imdct36_blocks_mips_float(float *out, float *buf, float *in,
01219                                int count, int switch_point, int block_type)
01220 {
01221     int j;
01222     for (j=0 ; j < count; j++) {
01223         /* apply window & overlap with previous buffer */
01224 
01225         /* select window */
01226         int win_idx = (switch_point && j < 2) ? 0 : block_type;
01227         float *win = ff_mdct_win_float[win_idx + (4 & -(j & 1))];
01228 
01229         imdct36_mips_float(out, buf, in, win);
01230 
01231         in  += 18;
01232         buf += ((j&3) != 3 ? 1 : (72-3));
01233         out++;
01234     }
01235 }
01236 
01237 void ff_mpadsp_init_mipsfpu(MPADSPContext *s)
01238 {
01239     s->apply_window_float   = ff_mpadsp_apply_window_mips_float;
01240     s->imdct36_blocks_float = ff_imdct36_blocks_mips_float;
01241     s->dct32_float          = ff_dct32_mips_float;
01242 }
libavcodec/mips/mpegaudiodsp_mips_float.c
