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.
25 * $Id: fir.h,v 1.8 2006/10/24 13:45:28 steveu Exp $
28 /*! \page fir_page FIR filtering
29 \section fir_page_sec_1 What does it do?
32 \section fir_page_sec_2 How does it work?
40 Blackfin NOTES & IDEAS:
42 A simple dot product function is used to implement the filter. This performs
43 just one MAC/cycle which is inefficient but was easy to implement as a first
44 pass. The current Blackfin code also uses an unrolled form of the filter
45 history to avoid 0 length hardware loop issues. This is wasteful of
48 Ideas for improvement:
50 1/ Rewrite filter for dual MAC inner loop. The issue here is handling
51 history sample offsets that are 16 bit aligned - the dual MAC needs
52 32 bit aligmnent. There are some good examples in libbfdsp.
54 2/ Use the hardware circular buffer facility tohalve memory usage.
56 3/ Consider using internal memory.
58 Using less memory might also improve speed as cache misses will be
59 reduced. A drop in MIPs and memory approaching 50% should be
62 The foreground and background filters currenlty use a total of
63 about 10 MIPs/ch as measured with speedtest.c on a 256 TAP echo
67 #if defined(USE_MMX) || defined(USE_SSE2)
72 16 bit integer FIR descriptor. This defines the working state for a single
73 instance of an FIR filter using 16 bit integer coefficients.
79 const int16_t *coeffs;
84 32 bit integer FIR descriptor. This defines the working state for a single
85 instance of an FIR filter using 32 bit integer coefficients, and filtering
92 const int32_t *coeffs;
97 Floating point FIR descriptor. This defines the working state for a single
98 instance of an FIR filter using floating point coefficients and data.
112 static __inline__ const int16_t *fir16_create(fir16_state_t *fir,
113 const int16_t *coeffs,
117 fir->curr_pos = taps - 1;
118 fir->coeffs = coeffs;
119 #if defined(USE_MMX) || defined(USE_SSE2) || defined(__BLACKFIN_ASM__)
120 if ((fir->history = malloc(2*taps*sizeof(int16_t))))
121 memset(fir->history, 0, 2*taps*sizeof(int16_t));
123 if ((fir->history = (int16_t *) malloc(taps*sizeof(int16_t))))
124 memset(fir->history, 0, taps*sizeof(int16_t));
128 /*- End of function --------------------------------------------------------*/
130 static __inline__ void fir16_flush(fir16_state_t *fir)
132 #if defined(USE_MMX) || defined(USE_SSE2) || defined(__BLACKFIN_ASM__)
133 memset(fir->history, 0, 2*fir->taps*sizeof(int16_t));
135 memset(fir->history, 0, fir->taps*sizeof(int16_t));
138 /*- End of function --------------------------------------------------------*/
140 static __inline__ void fir16_free(fir16_state_t *fir)
144 /*- End of function --------------------------------------------------------*/
146 #ifdef __BLACKFIN_ASM__
147 static inline int32_t dot_asm(short *x, short *y, int len)
158 "R0.L = W[I0++] || R1.L = W[I1++];\n\t"
159 "LOOP dot%= LC0 = %3;\n\t"
160 "LOOP_BEGIN dot%=;\n\t"
161 "A0 += R0.L * R1.L (IS) || R0.L = W[I0++] || R1.L = W[I1++];\n\t"
162 "LOOP_END dot%=;\n\t"
163 "A0 += R0.L*R1.L (IS);\n\t"
167 : "a" (x), "a" (y), "a" (len)
168 : "I0", "I1", "A1", "A0", "R0", "R1"
174 /*- End of function --------------------------------------------------------*/
176 static __inline__ int16_t fir16(fir16_state_t *fir, int16_t sample)
184 fir->history[fir->curr_pos] = sample;
185 fir->history[fir->curr_pos + fir->taps] = sample;
187 mmx_coeffs = (mmx_t *) fir->coeffs;
188 mmx_hist = (mmx_t *) &fir->history[fir->curr_pos];
191 /* 8 samples per iteration, so the filter must be a multiple of 8 long. */
194 movq_m2r(mmx_coeffs[0], mm0);
195 movq_m2r(mmx_coeffs[1], mm2);
196 movq_m2r(mmx_hist[0], mm1);
197 movq_m2r(mmx_hist[1], mm3);
200 pmaddwd_r2r(mm1, mm0);
201 pmaddwd_r2r(mm3, mm2);
211 #elif defined(USE_SSE2)
216 fir->history[fir->curr_pos] = sample;
217 fir->history[fir->curr_pos + fir->taps] = sample;
219 xmm_coeffs = (xmm_t *) fir->coeffs;
220 xmm_hist = (xmm_t *) &fir->history[fir->curr_pos];
222 pxor_r2r(xmm4, xmm4);
223 /* 16 samples per iteration, so the filter must be a multiple of 16 long. */
226 movdqu_m2r(xmm_coeffs[0], xmm0);
227 movdqu_m2r(xmm_coeffs[1], xmm2);
228 movdqu_m2r(xmm_hist[0], xmm1);
229 movdqu_m2r(xmm_hist[1], xmm3);
232 pmaddwd_r2r(xmm1, xmm0);
233 pmaddwd_r2r(xmm3, xmm2);
234 paddd_r2r(xmm0, xmm4);
235 paddd_r2r(xmm2, xmm4);
238 movdqa_r2r(xmm4, xmm0);
240 paddd_r2r(xmm0, xmm4);
241 movdqa_r2r(xmm4, xmm0);
243 paddd_r2r(xmm0, xmm4);
245 #elif defined(__BLACKFIN_ASM__)
246 fir->history[fir->curr_pos] = sample;
247 fir->history[fir->curr_pos + fir->taps] = sample;
248 y = dot_asm((int16_t*)fir->coeffs, &fir->history[fir->curr_pos], fir->taps);
254 fir->history[fir->curr_pos] = sample;
256 offset2 = fir->curr_pos;
257 offset1 = fir->taps - offset2;
259 for (i = fir->taps - 1; i >= offset1; i--)
260 y += fir->coeffs[i]*fir->history[i - offset1];
262 y += fir->coeffs[i]*fir->history[i + offset2];
264 if (fir->curr_pos <= 0)
265 fir->curr_pos = fir->taps;
267 return (int16_t) (y >> 15);
269 /*- End of function --------------------------------------------------------*/
271 static __inline__ const int16_t *fir32_create(fir32_state_t *fir,
272 const int32_t *coeffs,
276 fir->curr_pos = taps - 1;
277 fir->coeffs = coeffs;
278 fir->history = (int16_t *) malloc(taps*sizeof(int16_t));
280 memset(fir->history, '\0', taps*sizeof(int16_t));
283 /*- End of function --------------------------------------------------------*/
285 static __inline__ void fir32_flush(fir32_state_t *fir)
287 memset(fir->history, 0, fir->taps*sizeof(int16_t));
289 /*- End of function --------------------------------------------------------*/
291 static __inline__ void fir32_free(fir32_state_t *fir)
295 /*- End of function --------------------------------------------------------*/
297 static __inline__ int16_t fir32(fir32_state_t *fir, int16_t sample)
304 fir->history[fir->curr_pos] = sample;
305 offset2 = fir->curr_pos;
306 offset1 = fir->taps - offset2;
308 for (i = fir->taps - 1; i >= offset1; i--)
309 y += fir->coeffs[i]*fir->history[i - offset1];
311 y += fir->coeffs[i]*fir->history[i + offset2];
312 if (fir->curr_pos <= 0)
313 fir->curr_pos = fir->taps;
315 return (int16_t) (y >> 15);
317 /*- End of function --------------------------------------------------------*/
320 static __inline__ const float *fir_float_create(fir_float_state_t *fir,
325 fir->curr_pos = taps - 1;
326 fir->coeffs = coeffs;
327 fir->history = (float *) malloc(taps*sizeof(float));
329 memset(fir->history, '\0', taps*sizeof(float));
332 /*- End of function --------------------------------------------------------*/
334 static __inline__ void fir_float_free(fir_float_state_t *fir)
338 /*- End of function --------------------------------------------------------*/
340 static __inline__ int16_t fir_float(fir_float_state_t *fir, int16_t sample)
347 fir->history[fir->curr_pos] = sample;
349 offset2 = fir->curr_pos;
350 offset1 = fir->taps - offset2;
352 for (i = fir->taps - 1; i >= offset1; i--)
353 y += fir->coeffs[i]*fir->history[i - offset1];
355 y += fir->coeffs[i]*fir->history[i + offset2];
356 if (fir->curr_pos <= 0)
357 fir->curr_pos = fir->taps;
361 /*- End of function --------------------------------------------------------*/
369 /*- End of file ------------------------------------------------------------*/