1 /* ----------------------------------------------------------------------
2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
4 * $Date: 17. January 2013
7 * Project: CMSIS DSP Library
8 * Title: arm_fir_fast_q15.c
10 * Description: Q15 Fast FIR filter processing function.
12 * Target Processor: Cortex-M4/Cortex-M3
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * - Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * - Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in
21 * the documentation and/or other materials provided with the
23 * - Neither the name of ARM LIMITED nor the names of its contributors
24 * may be used to endorse or promote products derived from this
25 * software without specific prior written permission.
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
38 * POSSIBILITY OF SUCH DAMAGE.
39 * -------------------------------------------------------------------- */
44 * @ingroup groupFilters
53 * @param[in] *S points to an instance of the Q15 FIR filter structure.
54 * @param[in] *pSrc points to the block of input data.
55 * @param[out] *pDst points to the block of output data.
56 * @param[in] blockSize number of samples to process per call.
59 * <b>Scaling and Overflow Behavior:</b>
61 * This fast version uses a 32-bit accumulator with 2.30 format.
62 * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit.
63 * Thus, if the accumulator result overflows it wraps around and distorts the result.
64 * In order to avoid overflows completely the input signal must be scaled down by log2(numTaps) bits.
65 * The 2.30 accumulator is then truncated to 2.15 format and saturated to yield the 1.15 result.
68 * Refer to the function <code>arm_fir_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion. Both the slow and the fast versions use the same instance structure.
69 * Use the function <code>arm_fir_init_q15()</code> to initialize the filter structure.
72 void arm_fir_fast_q15(
73 const arm_fir_instance_q15 * S,
78 q15_t *pState = S->pState; /* State pointer */
79 q15_t *pCoeffs = S->pCoeffs; /* Coefficient pointer */
80 q15_t *pStateCurnt; /* Points to the current sample of the state */
81 q31_t acc0, acc1, acc2, acc3; /* Accumulators */
82 q15_t *pb; /* Temporary pointer for coefficient buffer */
83 q15_t *px; /* Temporary q31 pointer for SIMD state buffer accesses */
84 q31_t x0, x1, x2, c0; /* Temporary variables to hold SIMD state and coefficient values */
85 uint32_t numTaps = S->numTaps; /* Number of taps in the filter */
86 uint32_t tapCnt, blkCnt; /* Loop counters */
89 /* S->pState points to state array which contains previous frame (numTaps - 1) samples */
90 /* pStateCurnt points to the location where the new input data should be written */
91 pStateCurnt = &(S->pState[(numTaps - 1u)]);
93 /* Apply loop unrolling and compute 4 output values simultaneously.
94 * The variables acc0 ... acc3 hold output values that are being computed:
96 * acc0 = b[numTaps-1] * x[n-numTaps-1] + b[numTaps-2] * x[n-numTaps-2] + b[numTaps-3] * x[n-numTaps-3] +...+ b[0] * x[0]
97 * acc1 = b[numTaps-1] * x[n-numTaps] + b[numTaps-2] * x[n-numTaps-1] + b[numTaps-3] * x[n-numTaps-2] +...+ b[0] * x[1]
98 * acc2 = b[numTaps-1] * x[n-numTaps+1] + b[numTaps-2] * x[n-numTaps] + b[numTaps-3] * x[n-numTaps-1] +...+ b[0] * x[2]
99 * acc3 = b[numTaps-1] * x[n-numTaps+2] + b[numTaps-2] * x[n-numTaps+1] + b[numTaps-3] * x[n-numTaps] +...+ b[0] * x[3]
102 blkCnt = blockSize >> 2;
104 /* First part of the processing with loop unrolling. Compute 4 outputs at a time.
105 ** a second loop below computes the remaining 1 to 3 samples. */
108 /* Copy four new input samples into the state buffer.
109 ** Use 32-bit SIMD to move the 16-bit data. Only requires two copies. */
110 *pStateCurnt++ = *pSrc++;
111 *pStateCurnt++ = *pSrc++;
112 *pStateCurnt++ = *pSrc++;
113 *pStateCurnt++ = *pSrc++;
116 /* Set all accumulators to zero */
122 /* Typecast q15_t pointer to q31_t pointer for state reading in q31_t */
125 /* Typecast q15_t pointer to q31_t pointer for coefficient reading in q31_t */
128 /* Read the first two samples from the state buffer: x[n-N], x[n-N-1] */
129 x0 = *__SIMD32(px)++;
131 /* Read the third and forth samples from the state buffer: x[n-N-2], x[n-N-3] */
132 x2 = *__SIMD32(px)++;
134 /* Loop over the number of taps. Unroll by a factor of 4.
135 ** Repeat until we've computed numTaps-(numTaps%4) coefficients. */
136 tapCnt = numTaps >> 2;
140 /* Read the first two coefficients using SIMD: b[N] and b[N-1] coefficients */
141 c0 = *__SIMD32(pb)++;
143 /* acc0 += b[N] * x[n-N] + b[N-1] * x[n-N-1] */
144 acc0 = __SMLAD(x0, c0, acc0);
146 /* acc2 += b[N] * x[n-N-2] + b[N-1] * x[n-N-3] */
147 acc2 = __SMLAD(x2, c0, acc2);
149 /* pack x[n-N-1] and x[n-N-2] */
150 #ifndef ARM_MATH_BIG_ENDIAN
151 x1 = __PKHBT(x2, x0, 0);
153 x1 = __PKHBT(x0, x2, 0);
156 /* Read state x[n-N-4], x[n-N-5] */
157 x0 = _SIMD32_OFFSET(px);
159 /* acc1 += b[N] * x[n-N-1] + b[N-1] * x[n-N-2] */
160 acc1 = __SMLADX(x1, c0, acc1);
162 /* pack x[n-N-3] and x[n-N-4] */
163 #ifndef ARM_MATH_BIG_ENDIAN
164 x1 = __PKHBT(x0, x2, 0);
166 x1 = __PKHBT(x2, x0, 0);
169 /* acc3 += b[N] * x[n-N-3] + b[N-1] * x[n-N-4] */
170 acc3 = __SMLADX(x1, c0, acc3);
172 /* Read coefficients b[N-2], b[N-3] */
173 c0 = *__SIMD32(pb)++;
175 /* acc0 += b[N-2] * x[n-N-2] + b[N-3] * x[n-N-3] */
176 acc0 = __SMLAD(x2, c0, acc0);
178 /* Read state x[n-N-6], x[n-N-7] with offset */
179 x2 = _SIMD32_OFFSET(px + 2u);
181 /* acc2 += b[N-2] * x[n-N-4] + b[N-3] * x[n-N-5] */
182 acc2 = __SMLAD(x0, c0, acc2);
184 /* acc1 += b[N-2] * x[n-N-3] + b[N-3] * x[n-N-4] */
185 acc1 = __SMLADX(x1, c0, acc1);
187 /* pack x[n-N-5] and x[n-N-6] */
188 #ifndef ARM_MATH_BIG_ENDIAN
189 x1 = __PKHBT(x2, x0, 0);
191 x1 = __PKHBT(x0, x2, 0);
194 /* acc3 += b[N-2] * x[n-N-5] + b[N-3] * x[n-N-6] */
195 acc3 = __SMLADX(x1, c0, acc3);
197 /* Update state pointer for next state reading */
200 /* Decrement tap count */
205 /* If the filter length is not a multiple of 4, compute the remaining filter taps.
206 ** This is always be 2 taps since the filter length is even. */
207 if((numTaps & 0x3u) != 0u)
210 /* Read last two coefficients */
211 c0 = *__SIMD32(pb)++;
213 /* Perform the multiply-accumulates */
214 acc0 = __SMLAD(x0, c0, acc0);
215 acc2 = __SMLAD(x2, c0, acc2);
217 /* pack state variables */
218 #ifndef ARM_MATH_BIG_ENDIAN
219 x1 = __PKHBT(x2, x0, 0);
221 x1 = __PKHBT(x0, x2, 0);
224 /* Read last state variables */
227 /* Perform the multiply-accumulates */
228 acc1 = __SMLADX(x1, c0, acc1);
230 /* pack state variables */
231 #ifndef ARM_MATH_BIG_ENDIAN
232 x1 = __PKHBT(x0, x2, 0);
234 x1 = __PKHBT(x2, x0, 0);
237 /* Perform the multiply-accumulates */
238 acc3 = __SMLADX(x1, c0, acc3);
241 /* The results in the 4 accumulators are in 2.30 format. Convert to 1.15 with saturation.
242 ** Then store the 4 outputs in the destination buffer. */
244 #ifndef ARM_MATH_BIG_ENDIAN
247 __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
250 __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
255 __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16);
258 __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16);
261 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
263 /* Advance the state pointer by 4 to process the next group of 4 samples */
264 pState = pState + 4u;
266 /* Decrement the loop counter */
270 /* If the blockSize is not a multiple of 4, compute any remaining output samples here.
271 ** No loop unrolling is used. */
272 blkCnt = blockSize % 0x4u;
275 /* Copy two samples into state buffer */
276 *pStateCurnt++ = *pSrc++;
278 /* Set the accumulator to zero */
281 /* Use SIMD to hold states and coefficients */
285 tapCnt = numTaps >> 1u;
290 acc0 += (q31_t) * px++ * *pb++;
291 acc0 += (q31_t) * px++ * *pb++;
297 /* The result is in 2.30 format. Convert to 1.15 with saturation.
298 ** Then store the output in the destination buffer. */
299 *pDst++ = (q15_t) (__SSAT((acc0 >> 15), 16));
301 /* Advance state pointer by 1 for the next sample */
302 pState = pState + 1u;
304 /* Decrement the loop counter */
308 /* Processing is complete.
309 ** Now copy the last numTaps - 1 samples to the satrt of the state buffer.
310 ** This prepares the state buffer for the next function call. */
312 /* Points to the start of the state buffer */
313 pStateCurnt = S->pState;
315 /* Calculation of count for copying integer writes */
316 tapCnt = (numTaps - 1u) >> 2;
320 *pStateCurnt++ = *pState++;
321 *pStateCurnt++ = *pState++;
322 *pStateCurnt++ = *pState++;
323 *pStateCurnt++ = *pState++;
329 /* Calculation of count for remaining q15_t data */
330 tapCnt = (numTaps - 1u) % 0x4u;
332 /* copy remaining data */
335 *pStateCurnt++ = *pState++;
337 /* Decrement the loop counter */
344 * @} end of FIR group