2 * SpanDSP - a series of DSP components for telephony
4 * fir.h - General telephony FIR routines
6 * Written by Steve Underwood <steveu@coppice.org>
8 * Copyright (C) 2002 Steve Underwood
10 * All rights reserved.
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.
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.
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.
26 /*! \page fir_page FIR filtering
27 \section fir_page_sec_1 What does it do?
30 \section fir_page_sec_2 How does it work?
38 Blackfin NOTES & IDEAS:
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
46 Ideas for improvement:
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.
52 2/ Use the hardware circular buffer facility tohalve memory usage.
54 3/ Consider using internal memory.
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
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
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.
69 struct fir16_state_t
{
72 const int16_t *coeffs
;
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
81 struct fir32_state_t
{
84 const int32_t *coeffs
;
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.
92 struct fir_float_state_t
{
99 static inline const int16_t *fir16_create(struct fir16_state_t
*fir
,
100 const int16_t *coeffs
, int 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
);
108 fir
->history
= kcalloc(taps
, sizeof(int16_t), GFP_KERNEL
);
113 static inline void fir16_flush(struct fir16_state_t
*fir
)
115 #if defined(USE_SSE2) || defined(__bfin__)
116 memset(fir
->history
, 0, 2 * fir
->taps
* sizeof(int16_t));
118 memset(fir
->history
, 0, fir
->taps
* sizeof(int16_t));
122 static inline void fir16_free(struct fir16_state_t
*fir
)
128 static inline int32_t dot_asm(short *x
, short *y
, int len
)
134 __asm__("I0 = %1;\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"
146 : "a"(x
), "a"(y
), "a"(len
)
147 : "I0", "I1", "A1", "A0", "R0", "R1"
154 static inline int16_t fir16(struct fir16_state_t
*fir
, int16_t sample
)
157 #if defined(USE_SSE2)
159 union xmm_t
*xmm_coeffs
;
160 union xmm_t
*xmm_hist
;
162 fir
->history
[fir
->curr_pos
] = sample
;
163 fir
->history
[fir
->curr_pos
+ fir
->taps
] = sample
;
165 xmm_coeffs
= (union xmm_t
*)fir
->coeffs
;
166 xmm_hist
= (union xmm_t
*)&fir
->history
[fir
->curr_pos
];
168 pxor_r2r(xmm4
, xmm4
);
169 /* 16 samples per iteration, so the filter must be a multiple of 16 long. */
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
);
177 pmaddwd_r2r(xmm1
, xmm0
);
178 pmaddwd_r2r(xmm3
, xmm2
);
179 paddd_r2r(xmm0
, xmm4
);
180 paddd_r2r(xmm2
, xmm4
);
183 movdqa_r2r(xmm4
, xmm0
);
185 paddd_r2r(xmm0
, xmm4
);
186 movdqa_r2r(xmm4
, xmm0
);
188 paddd_r2r(xmm0
, xmm4
);
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
],
200 fir
->history
[fir
->curr_pos
] = sample
;
202 offset2
= fir
->curr_pos
;
203 offset1
= fir
->taps
- offset2
;
205 for (i
= fir
->taps
- 1; i
>= offset1
; i
--)
206 y
+= fir
->coeffs
[i
] * fir
->history
[i
- offset1
];
208 y
+= fir
->coeffs
[i
] * fir
->history
[i
+ offset2
];
210 if (fir
->curr_pos
<= 0)
211 fir
->curr_pos
= fir
->taps
;
213 return (int16_t) (y
>> 15);
216 static inline const int16_t *fir32_create(struct fir32_state_t
*fir
,
217 const int32_t *coeffs
, int taps
)
220 fir
->curr_pos
= taps
- 1;
221 fir
->coeffs
= coeffs
;
222 fir
->history
= kcalloc(taps
, sizeof(int16_t), GFP_KERNEL
);
226 static inline void fir32_flush(struct fir32_state_t
*fir
)
228 memset(fir
->history
, 0, fir
->taps
* sizeof(int16_t));
231 static inline void fir32_free(struct fir32_state_t
*fir
)
236 static inline int16_t fir32(struct fir32_state_t
*fir
, int16_t sample
)
243 fir
->history
[fir
->curr_pos
] = sample
;
244 offset2
= fir
->curr_pos
;
245 offset1
= fir
->taps
- offset2
;
247 for (i
= fir
->taps
- 1; i
>= offset1
; i
--)
248 y
+= fir
->coeffs
[i
] * fir
->history
[i
- offset1
];
250 y
+= fir
->coeffs
[i
] * fir
->history
[i
+ offset2
];
251 if (fir
->curr_pos
<= 0)
252 fir
->curr_pos
= fir
->taps
;
254 return (int16_t) (y
>> 15);
258 /*- End of file ------------------------------------------------------------*/