git.sesse.net Git - ffmpeg/blob - libavcodec/mips/celp_filters_mips.c

   1 /*
   2  * Copyright (c) 2012
   3  *      MIPS Technologies, Inc., California.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
  14  *    contributors may be used to endorse or promote products derived from
  15  *    this software without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  * Author:  Nedeljko Babic (nbabic@mips.com)
  30  *
  31  * various filters for CELP-based codecs optimized for MIPS
  32  *
  33  * This file is part of FFmpeg.
  34  *
  35  * FFmpeg is free software; you can redistribute it and/or
  36  * modify it under the terms of the GNU Lesser General Public
  37  * License as published by the Free Software Foundation; either
  38  * version 2.1 of the License, or (at your option) any later version.
  39  *
  40  * FFmpeg is distributed in the hope that it will be useful,
  41  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  42  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  43  * Lesser General Public License for more details.
  44  *
  45  * You should have received a copy of the GNU Lesser General Public
  46  * License along with FFmpeg; if not, write to the Free Software
  47  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  48  */
  49
  50 /**
  51  * @file
  52  * Reference: libavcodec/celp_filters.c
  53  */
  54 #include "config.h"
  55 #include "libavutil/attributes.h"
  56 #include "libavutil/common.h"
  57 #include "libavcodec/celp_filters.h"
  58 #include "libavutil/mips/asmdefs.h"
  59
  60 #if HAVE_INLINE_ASM
  61 static void ff_celp_lp_synthesis_filterf_mips(float *out,
  62                                   const float *filter_coeffs,
  63                                   const float* in, int buffer_length,
  64                                   int filter_length)
  65 {
  66     int i,n;
  67
  68     float out0, out1, out2, out3;
  69     float old_out0, old_out1, old_out2, old_out3;
  70     float a,b,c;
  71     const float *p_filter_coeffs;
  72     float *p_out;
  73
  74     a = filter_coeffs[0];
  75     b = filter_coeffs[1];
  76     c = filter_coeffs[2];
  77     b -= filter_coeffs[0] * filter_coeffs[0];
  78     c -= filter_coeffs[1] * filter_coeffs[0];
  79     c -= filter_coeffs[0] * b;
  80
  81     old_out0 = out[-4];
  82     old_out1 = out[-3];
  83     old_out2 = out[-2];
  84     old_out3 = out[-1];
  85     for (n = 0; n <= buffer_length - 4; n+=4) {
  86         p_filter_coeffs = filter_coeffs;
  87         p_out = out;
  88
  89         out0 = in[0];
  90         out1 = in[1];
  91         out2 = in[2];
  92         out3 = in[3];
  93
  94         __asm__ volatile(
  95             "lwc1       $f2,     8(%[filter_coeffs])                        \n\t"
  96             "lwc1       $f1,     4(%[filter_coeffs])                        \n\t"
  97             "lwc1       $f0,     0(%[filter_coeffs])                        \n\t"
  98             "nmsub.s    %[out0], %[out0],             $f2, %[old_out1]      \n\t"
  99             "nmsub.s    %[out1], %[out1],             $f2, %[old_out2]      \n\t"
 100             "nmsub.s    %[out2], %[out2],             $f2, %[old_out3]      \n\t"
 101             "lwc1       $f3,     12(%[filter_coeffs])                       \n\t"
 102             "nmsub.s    %[out0], %[out0],             $f1, %[old_out2]      \n\t"
 103             "nmsub.s    %[out1], %[out1],             $f1, %[old_out3]      \n\t"
 104             "nmsub.s    %[out2], %[out2],             $f3, %[old_out2]      \n\t"
 105             "nmsub.s    %[out0], %[out0],             $f0, %[old_out3]      \n\t"
 106             "nmsub.s    %[out3], %[out3],             $f3, %[old_out3]      \n\t"
 107             "nmsub.s    %[out1], %[out1],             $f3, %[old_out1]      \n\t"
 108             "nmsub.s    %[out0], %[out0],             $f3, %[old_out0]      \n\t"
 109
 110             : [out0]"+f"(out0), [out1]"+f"(out1),
 111               [out2]"+f"(out2), [out3]"+f"(out3)
 112             : [old_out0]"f"(old_out0), [old_out1]"f"(old_out1),
 113               [old_out2]"f"(old_out2), [old_out3]"f"(old_out3),
 114               [filter_coeffs]"r"(filter_coeffs)
 115             : "$f0", "$f1", "$f2", "$f3", "$f4", "memory"
 116         );
 117
 118         for (i = 5; i <= filter_length; i += 2) {
 119             __asm__ volatile(
 120                 "lwc1    %[old_out3], -20(%[p_out])                         \n\t"
 121                 "lwc1    $f5,         16(%[p_filter_coeffs])                \n\t"
 122                 PTR_ADDIU "%[p_out],  -8                                    \n\t"
 123                 PTR_ADDIU "%[p_filter_coeffs], 8                            \n\t"
 124                 "nmsub.s %[out1],     %[out1],      $f5, %[old_out0]        \n\t"
 125                 "nmsub.s %[out3],     %[out3],      $f5, %[old_out2]        \n\t"
 126                 "lwc1    $f4,         12(%[p_filter_coeffs])                \n\t"
 127                 "lwc1    %[old_out2], -16(%[p_out])                         \n\t"
 128                 "nmsub.s %[out0],     %[out0],      $f5, %[old_out3]        \n\t"
 129                 "nmsub.s %[out2],     %[out2],      $f5, %[old_out1]        \n\t"
 130                 "nmsub.s %[out1],     %[out1],      $f4, %[old_out3]        \n\t"
 131                 "nmsub.s %[out3],     %[out3],      $f4, %[old_out1]        \n\t"
 132                 "mov.s   %[old_out1], %[old_out3]                           \n\t"
 133                 "nmsub.s %[out0],     %[out0],      $f4, %[old_out2]        \n\t"
 134                 "nmsub.s %[out2],     %[out2],      $f4, %[old_out0]        \n\t"
 135
 136                 : [out0]"+f"(out0), [out1]"+f"(out1),
 137                   [out2]"+f"(out2), [out3]"+f"(out3), [old_out0]"+f"(old_out0),
 138                   [old_out1]"+f"(old_out1), [old_out2]"+f"(old_out2),
 139                   [old_out3]"+f"(old_out3),[p_filter_coeffs]"+r"(p_filter_coeffs),
 140                   [p_out]"+r"(p_out)
 141                 :
 142                 : "$f4", "$f5", "memory"
 143             );
 144             FFSWAP(float, old_out0, old_out2);
 145         }
 146
 147         __asm__ volatile(
 148             "nmsub.s    %[out3], %[out3], %[a], %[out2]                     \n\t"
 149             "nmsub.s    %[out2], %[out2], %[a], %[out1]                     \n\t"
 150             "nmsub.s    %[out3], %[out3], %[b], %[out1]                     \n\t"
 151             "nmsub.s    %[out1], %[out1], %[a], %[out0]                     \n\t"
 152             "nmsub.s    %[out2], %[out2], %[b], %[out0]                     \n\t"
 153             "nmsub.s    %[out3], %[out3], %[c], %[out0]                     \n\t"
 154
 155             : [out0]"+f"(out0), [out1]"+f"(out1),
 156               [out2]"+f"(out2), [out3]"+f"(out3)
 157             : [a]"f"(a), [b]"f"(b), [c]"f"(c)
 158         );
 159
 160         out[0] = out0;
 161         out[1] = out1;
 162         out[2] = out2;
 163         out[3] = out3;
 164
 165         old_out0 = out0;
 166         old_out1 = out1;
 167         old_out2 = out2;
 168         old_out3 = out3;
 169
 170         out += 4;
 171         in  += 4;
 172     }
 173
 174     out -= n;
 175     in -= n;
 176     for (; n < buffer_length; n++) {
 177         float out_val, out_val_i, fc_val;
 178         p_filter_coeffs = filter_coeffs;
 179         p_out = &out[n];
 180         out_val = in[n];
 181         for (i = 1; i <= filter_length; i++) {
 182             __asm__ volatile(
 183                 "lwc1    %[fc_val],          0(%[p_filter_coeffs])                        \n\t"
 184                 "lwc1    %[out_val_i],       -4(%[p_out])                                 \n\t"
 185                 PTR_ADDIU "%[p_filter_coeffs], 4                                          \n\t"
 186                 PTR_ADDIU "%[p_out],         -4                                           \n\t"
 187                 "nmsub.s %[out_val],         %[out_val],          %[fc_val], %[out_val_i] \n\t"
 188
 189                 : [fc_val]"=&f"(fc_val), [out_val]"+f"(out_val),
 190                   [out_val_i]"=&f"(out_val_i), [p_out]"+r"(p_out),
 191                   [p_filter_coeffs]"+r"(p_filter_coeffs)
 192                 :
 193                 : "memory"
 194             );
 195         }
 196         out[n] = out_val;
 197     }
 198 }
 199
 200 static void ff_celp_lp_zero_synthesis_filterf_mips(float *out,
 201                                        const float *filter_coeffs,
 202                                        const float *in, int buffer_length,
 203                                        int filter_length)
 204 {
 205     int i,n;
 206     float sum_out8, sum_out7, sum_out6, sum_out5, sum_out4, fc_val;
 207     float sum_out3, sum_out2, sum_out1;
 208     const float *p_filter_coeffs, *p_in;
 209
 210     for (n = 0; n < buffer_length; n+=8) {
 211         p_in = &in[n];
 212         p_filter_coeffs = filter_coeffs;
 213         sum_out8 = in[n+7];
 214         sum_out7 = in[n+6];
 215         sum_out6 = in[n+5];
 216         sum_out5 = in[n+4];
 217         sum_out4 = in[n+3];
 218         sum_out3 = in[n+2];
 219         sum_out2 = in[n+1];
 220         sum_out1 = in[n];
 221         i = filter_length;
 222
 223         /* i is always greater than 0
 224         * outer loop is unrolled eight times so there is less memory access
 225         * inner loop is unrolled two times
 226         */
 227         __asm__ volatile(
 228             "filt_lp_inner%=:                                               \n\t"
 229             "lwc1   %[fc_val],   0(%[p_filter_coeffs])                      \n\t"
 230             "lwc1   $f7,         6*4(%[p_in])                               \n\t"
 231             "lwc1   $f6,         5*4(%[p_in])                               \n\t"
 232             "lwc1   $f5,         4*4(%[p_in])                               \n\t"
 233             "lwc1   $f4,         3*4(%[p_in])                               \n\t"
 234             "lwc1   $f3,         2*4(%[p_in])                               \n\t"
 235             "lwc1   $f2,         4(%[p_in])                                 \n\t"
 236             "lwc1   $f1,         0(%[p_in])                                 \n\t"
 237             "lwc1   $f0,         -4(%[p_in])                                \n\t"
 238             "addiu  %[i],        -2                                         \n\t"
 239             "madd.s %[sum_out8], %[sum_out8],          %[fc_val], $f7       \n\t"
 240             "madd.s %[sum_out7], %[sum_out7],          %[fc_val], $f6       \n\t"
 241             "madd.s %[sum_out6], %[sum_out6],          %[fc_val], $f5       \n\t"
 242             "madd.s %[sum_out5], %[sum_out5],          %[fc_val], $f4       \n\t"
 243             "madd.s %[sum_out4], %[sum_out4],          %[fc_val], $f3       \n\t"
 244             "madd.s %[sum_out3], %[sum_out3],          %[fc_val], $f2       \n\t"
 245             "madd.s %[sum_out2], %[sum_out2],          %[fc_val], $f1       \n\t"
 246             "madd.s %[sum_out1], %[sum_out1],          %[fc_val], $f0       \n\t"
 247             "lwc1   %[fc_val],   4(%[p_filter_coeffs])                      \n\t"
 248             "lwc1   $f7,         -8(%[p_in])                                \n\t"
 249             PTR_ADDIU "%[p_filter_coeffs], 8                                \n\t"
 250             PTR_ADDIU "%[p_in],  -8                                         \n\t"
 251             "madd.s %[sum_out8], %[sum_out8],          %[fc_val], $f6       \n\t"
 252             "madd.s %[sum_out7], %[sum_out7],          %[fc_val], $f5       \n\t"
 253             "madd.s %[sum_out6], %[sum_out6],          %[fc_val], $f4       \n\t"
 254             "madd.s %[sum_out5], %[sum_out5],          %[fc_val], $f3       \n\t"
 255             "madd.s %[sum_out4], %[sum_out4],          %[fc_val], $f2       \n\t"
 256             "madd.s %[sum_out3], %[sum_out3],          %[fc_val], $f1       \n\t"
 257             "madd.s %[sum_out2], %[sum_out2],          %[fc_val], $f0       \n\t"
 258             "madd.s %[sum_out1], %[sum_out1],          %[fc_val], $f7       \n\t"
 259             "bgtz   %[i],        filt_lp_inner%=                            \n\t"
 260
 261             : [sum_out8]"+f"(sum_out8), [sum_out7]"+f"(sum_out7),
 262               [sum_out6]"+f"(sum_out6), [sum_out5]"+f"(sum_out5),
 263               [sum_out4]"+f"(sum_out4), [sum_out3]"+f"(sum_out3),
 264               [sum_out2]"+f"(sum_out2), [sum_out1]"+f"(sum_out1),
 265               [fc_val]"=&f"(fc_val), [p_filter_coeffs]"+r"(p_filter_coeffs),
 266               [p_in]"+r"(p_in), [i]"+r"(i)
 267             :
 268             : "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "memory"
 269         );
 270
 271         out[n+7] = sum_out8;
 272         out[n+6] = sum_out7;
 273         out[n+5] = sum_out6;
 274         out[n+4] = sum_out5;
 275         out[n+3] = sum_out4;
 276         out[n+2] = sum_out3;
 277         out[n+1] = sum_out2;
 278         out[n] = sum_out1;
 279     }
 280 }
 281 #endif /* HAVE_INLINE_ASM */
 282
 283 void ff_celp_filter_init_mips(CELPFContext *c)
 284 {
 285 #if HAVE_INLINE_ASM
 286     c->celp_lp_synthesis_filterf        = ff_celp_lp_synthesis_filterf_mips;
 287     c->celp_lp_zero_synthesis_filterf   = ff_celp_lp_zero_synthesis_filterf_mips;
 288 #endif
 289 }