drivers/staging/echo/fir.h

   1 /*
   2  * SpanDSP - a series of DSP components for telephony
   3  *
   4  * fir.h - General telephony FIR routines
   5  *
   6  * Written by Steve Underwood <steveu@coppice.org>
   7  *
   8  * Copyright (C) 2002 Steve Underwood
   9  *
  10  * All rights reserved.
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2, as
  14  * published by the Free Software Foundation.
  15  *
  16  * This program is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  * GNU General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU General Public License
  22  * along with this program; if not, write to the Free Software
  23  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 /*! \page fir_page FIR filtering
  27 \section fir_page_sec_1 What does it do?
  28 ???.
  29
  30 \section fir_page_sec_2 How does it work?
  31 ???.
  32 */
  33
  34 #if !defined(_FIR_H_)
  35 #define _FIR_H_
  36
  37 /*
  38    Blackfin NOTES & IDEAS:
  39
  40    A simple dot product function is used to implement the filter.  This performs
  41    just one MAC/cycle which is inefficient but was easy to implement as a first
  42    pass.  The current Blackfin code also uses an unrolled form of the filter
  43    history to avoid 0 length hardware loop issues.  This is wasteful of
  44    memory.
  45
  46    Ideas for improvement:
  47
  48    1/ Rewrite filter for dual MAC inner loop.  The issue here is handling
  49    history sample offsets that are 16 bit aligned - the dual MAC needs
  50    32 bit aligmnent.  There are some good examples in libbfdsp.
  51
  52    2/ Use the hardware circular buffer facility tohalve memory usage.
  53
  54    3/ Consider using internal memory.
  55
  56    Using less memory might also improve speed as cache misses will be
  57    reduced. A drop in MIPs and memory approaching 50% should be
  58    possible.
  59
  60    The foreground and background filters currenlty use a total of
  61    about 10 MIPs/ch as measured with speedtest.c on a 256 TAP echo
  62    can.
  63 */
  64
  65 /*!
  66     16 bit integer FIR descriptor. This defines the working state for a single
  67     instance of an FIR filter using 16 bit integer coefficients.
  68 */
  69 struct fir16_state_t {
  70         int taps;
  71         int curr_pos;
  72         const int16_t *coeffs;
  73         int16_t *history;
  74 };
  75
  76 /*!
  77     32 bit integer FIR descriptor. This defines the working state for a single
  78     instance of an FIR filter using 32 bit integer coefficients, and filtering
  79     16 bit integer data.
  80 */
  81 struct fir32_state_t {
  82         int taps;
  83         int curr_pos;
  84         const int32_t *coeffs;
  85         int16_t *history;
  86 };
  87
  88 /*!
  89     Floating point FIR descriptor. This defines the working state for a single
  90     instance of an FIR filter using floating point coefficients and data.
  91 */
  92 struct fir_float_state_t {
  93         int taps;
  94         int curr_pos;
  95         const float *coeffs;
  96         float *history;
  97 };
  98
  99 static inline const int16_t *fir16_create(struct fir16_state_t *fir,
 100                                               const int16_t *coeffs, int taps)
 101 {
 102         fir->taps = taps;
 103         fir->curr_pos = taps - 1;
 104         fir->coeffs = coeffs;
 105 #if defined(USE_SSE2) || defined(__bfin__)
 106         fir->history = kcalloc(2 * taps, sizeof(int16_t), GFP_KERNEL);
 107 #else
 108         fir->history = kcalloc(taps, sizeof(int16_t), GFP_KERNEL);
 109 #endif
 110         return fir->history;
 111 }
 112
 113 static inline void fir16_flush(struct fir16_state_t *fir)
 114 {
 115 #if defined(USE_SSE2) || defined(__bfin__)
 116         memset(fir->history, 0, 2 * fir->taps * sizeof(int16_t));
 117 #else
 118         memset(fir->history, 0, fir->taps * sizeof(int16_t));
 119 #endif
 120 }
 121
 122 static inline void fir16_free(struct fir16_state_t *fir)
 123 {
 124         kfree(fir->history);
 125 }
 126
 127 #ifdef __bfin__
 128 static inline int32_t dot_asm(short *x, short *y, int len)
 129 {
 130         int dot;
 131
 132         len--;
 133
 134         __asm__("I0 = %1;\n\t"
 135                 "I1 = %2;\n\t"
 136                 "A0 = 0;\n\t"
 137                 "R0.L = W[I0++] || R1.L = W[I1++];\n\t"
 138                 "LOOP dot%= LC0 = %3;\n\t"
 139                 "LOOP_BEGIN dot%=;\n\t"
 140                 "A0 += R0.L * R1.L (IS) || R0.L = W[I0++] || R1.L = W[I1++];\n\t"
 141                 "LOOP_END dot%=;\n\t"
 142                 "A0 += R0.L*R1.L (IS);\n\t"
 143                 "R0 = A0;\n\t"
 144                 "%0 = R0;\n\t"
 145                 : "=&d"(dot)
 146                 : "a"(x), "a"(y), "a"(len)
 147                 : "I0", "I1", "A1", "A0", "R0", "R1"
 148         );
 149
 150         return dot;
 151 }
 152 #endif
 153
 154 static inline int16_t fir16(struct fir16_state_t *fir, int16_t sample)
 155 {
 156         int32_t y;
 157 #if defined(USE_SSE2)
 158         int i;
 159         union xmm_t *xmm_coeffs;
 160         union xmm_t *xmm_hist;
 161
 162         fir->history[fir->curr_pos] = sample;
 163         fir->history[fir->curr_pos + fir->taps] = sample;
 164
 165         xmm_coeffs = (union xmm_t *)fir->coeffs;
 166         xmm_hist = (union xmm_t *)&fir->history[fir->curr_pos];
 167         i = fir->taps;
 168         pxor_r2r(xmm4, xmm4);
 169         /* 16 samples per iteration, so the filter must be a multiple of 16 long. */
 170         while (i > 0) {
 171                 movdqu_m2r(xmm_coeffs[0], xmm0);
 172                 movdqu_m2r(xmm_coeffs[1], xmm2);
 173                 movdqu_m2r(xmm_hist[0], xmm1);
 174                 movdqu_m2r(xmm_hist[1], xmm3);
 175                 xmm_coeffs += 2;
 176                 xmm_hist += 2;
 177                 pmaddwd_r2r(xmm1, xmm0);
 178                 pmaddwd_r2r(xmm3, xmm2);
 179                 paddd_r2r(xmm0, xmm4);
 180                 paddd_r2r(xmm2, xmm4);
 181                 i -= 16;
 182         }
 183         movdqa_r2r(xmm4, xmm0);
 184         psrldq_i2r(8, xmm0);
 185         paddd_r2r(xmm0, xmm4);
 186         movdqa_r2r(xmm4, xmm0);
 187         psrldq_i2r(4, xmm0);
 188         paddd_r2r(xmm0, xmm4);
 189         movd_r2m(xmm4, y);
 190 #elif defined(__bfin__)
 191         fir->history[fir->curr_pos] = sample;
 192         fir->history[fir->curr_pos + fir->taps] = sample;
 193         y = dot_asm((int16_t *) fir->coeffs, &fir->history[fir->curr_pos],
 194                     fir->taps);
 195 #else
 196         int i;
 197         int offset1;
 198         int offset2;
 199
 200         fir->history[fir->curr_pos] = sample;
 201
 202         offset2 = fir->curr_pos;
 203         offset1 = fir->taps - offset2;
 204         y = 0;
 205         for (i = fir->taps - 1; i >= offset1; i--)
 206                 y += fir->coeffs[i] * fir->history[i - offset1];
 207         for (; i >= 0; i--)
 208                 y += fir->coeffs[i] * fir->history[i + offset2];
 209 #endif
 210         if (fir->curr_pos <= 0)
 211                 fir->curr_pos = fir->taps;
 212         fir->curr_pos--;
 213         return (int16_t) (y >> 15);
 214 }
 215
 216 static inline const int16_t *fir32_create(struct fir32_state_t *fir,
 217                                               const int32_t *coeffs, int taps)
 218 {
 219         fir->taps = taps;
 220         fir->curr_pos = taps - 1;
 221         fir->coeffs = coeffs;
 222         fir->history = kcalloc(taps, sizeof(int16_t), GFP_KERNEL);
 223         return fir->history;
 224 }
 225
 226 static inline void fir32_flush(struct fir32_state_t *fir)
 227 {
 228         memset(fir->history, 0, fir->taps * sizeof(int16_t));
 229 }
 230
 231 static inline void fir32_free(struct fir32_state_t *fir)
 232 {
 233         kfree(fir->history);
 234 }
 235
 236 static inline int16_t fir32(struct fir32_state_t *fir, int16_t sample)
 237 {
 238         int i;
 239         int32_t y;
 240         int offset1;
 241         int offset2;
 242
 243         fir->history[fir->curr_pos] = sample;
 244         offset2 = fir->curr_pos;
 245         offset1 = fir->taps - offset2;
 246         y = 0;
 247         for (i = fir->taps - 1; i >= offset1; i--)
 248                 y += fir->coeffs[i] * fir->history[i - offset1];
 249         for (; i >= 0; i--)
 250                 y += fir->coeffs[i] * fir->history[i + offset2];
 251         if (fir->curr_pos <= 0)
 252                 fir->curr_pos = fir->taps;
 253         fir->curr_pos--;
 254         return (int16_t) (y >> 15);
 255 }
 256
 257 #endif
 258 /*- End of file ------------------------------------------------------------*/