00001 /* ---------------------------------------------------------------------- 00002 * Copyright (C) 2010 ARM Limited. All rights reserved. 00003 * 00004 * $Date: 15. July 2011 00005 * $Revision: V1.0.10 00006 * 00007 * Project: CMSIS DSP Library 00008 * Title: arm_fir_fast_q15.c 00009 * 00010 * Description: Q15 Fast FIR filter processing function. 00011 * 00012 * Target Processor: Cortex-M4/Cortex-M3 00013 * 00014 * Version 1.0.10 2011/7/15 00015 * Big Endian support added and Merged M0 and M3/M4 Source code. 00016 * 00017 * Version 1.0.3 2010/11/29 00018 * Re-organized the CMSIS folders and updated documentation. 00019 * 00020 * Version 1.0.2 2010/11/11 00021 * Documentation updated. 00022 * 00023 * Version 1.0.1 2010/10/05 00024 * Production release and review comments incorporated. 00025 * 00026 * Version 1.0.0 2010/09/20 00027 * Production release and review comments incorporated. 00028 * 00029 * Version 0.0.9 2010/08/16 00030 * Initial version 00031 * 00032 * -------------------------------------------------------------------- */ 00033 00034 #include "arm_math.h" 00035 00065 void arm_fir_fast_q15( 00066 const arm_fir_instance_q15 * S, 00067 q15_t * pSrc, 00068 q15_t * pDst, 00069 uint32_t blockSize) 00070 { 00071 q15_t *pState = S->pState; /* State pointer */ 00072 q15_t *pCoeffs = S->pCoeffs; /* Coefficient pointer */ 00073 q15_t *pStateCurnt; /* Points to the current sample of the state */ 00074 q15_t *px1; /* Temporary q15 pointer for state buffer */ 00075 q31_t *pb; /* Temporary pointer for coefficient buffer */ 00076 q31_t *px2; /* Temporary q31 pointer for SIMD state buffer accesses */ 00077 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold SIMD state and coefficient values */ 00078 q31_t acc0, acc1, acc2, acc3; /* Accumulators */ 00079 uint32_t numTaps = S->numTaps; /* Number of taps in the filter */ 00080 uint32_t tapCnt, blkCnt; /* Loop counters */ 00081 00082 /* S->pState points to buffer which contains previous frame (numTaps - 1) samples */ 00083 /* pStateCurnt points to the location where the new input data should be written */ 00084 pStateCurnt = &(S->pState[(numTaps - 1u)]); 00085 00086 /* Apply loop unrolling and compute 4 output values simultaneously. 00087 * The variables acc0 ... acc3 hold output values that are being computed: 00088 * 00089 * 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] 00090 * 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] 00091 * 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] 00092 * 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] 00093 */ 00094 blkCnt = blockSize >> 2; 00095 00096 /* First part of the processing with loop unrolling. Compute 4 outputs at a time. 00097 ** a second loop below computes the remaining 1 to 3 samples. */ 00098 while(blkCnt > 0u) 00099 { 00100 /* Copy four new input samples into the state buffer. 00101 ** Use 32-bit SIMD to move the 16-bit data. Only requires two copies. */ 00102 *__SIMD32(pStateCurnt)++ = *__SIMD32(pSrc)++; 00103 *__SIMD32(pStateCurnt)++ = *__SIMD32(pSrc)++; 00104 00105 /* Set all accumulators to zero */ 00106 acc0 = 0; 00107 acc1 = 0; 00108 acc2 = 0; 00109 acc3 = 0; 00110 00111 /* Initialize state pointer of type q15 */ 00112 px1 = pState; 00113 00114 /* Initialize coeff pointer of type q31 */ 00115 pb = (q31_t *) (pCoeffs); 00116 00117 /* Read the first two samples from the state buffer: x[n-N], x[n-N-1] */ 00118 x0 = *(q31_t *) (px1++); 00119 00120 /* Read the third and forth samples from the state buffer: x[n-N-1], x[n-N-2] */ 00121 x1 = *(q31_t *) (px1++); 00122 00123 /* Loop over the number of taps. Unroll by a factor of 4. 00124 ** Repeat until we've computed numTaps-4 coefficients. */ 00125 tapCnt = numTaps >> 2; 00126 do 00127 { 00128 /* Read the first two coefficients using SIMD: b[N] and b[N-1] coefficients */ 00129 c0 = *(pb++); 00130 00131 /* acc0 += b[N] * x[n-N] + b[N-1] * x[n-N-1] */ 00132 acc0 = __SMLAD(x0, c0, acc0); 00133 00134 /* acc1 += b[N] * x[n-N-1] + b[N-1] * x[n-N-2] */ 00135 acc1 = __SMLAD(x1, c0, acc1); 00136 00137 /* Read state x[n-N-2], x[n-N-3] */ 00138 x2 = *(q31_t *) (px1++); 00139 00140 /* Read state x[n-N-3], x[n-N-4] */ 00141 x3 = *(q31_t *) (px1++); 00142 00143 /* acc2 += b[N] * x[n-N-2] + b[N-1] * x[n-N-3] */ 00144 acc2 = __SMLAD(x2, c0, acc2); 00145 00146 /* acc3 += b[N] * x[n-N-3] + b[N-1] * x[n-N-4] */ 00147 acc3 = __SMLAD(x3, c0, acc3); 00148 00149 /* Read coefficients b[N-2], b[N-3] */ 00150 c0 = *(pb++); 00151 00152 /* acc0 += b[N-2] * x[n-N-2] + b[N-3] * x[n-N-3] */ 00153 acc0 = __SMLAD(x2, c0, acc0); 00154 00155 /* acc1 += b[N-2] * x[n-N-3] + b[N-3] * x[n-N-4] */ 00156 acc1 = __SMLAD(x3, c0, acc1); 00157 00158 /* Read state x[n-N-4], x[n-N-5] */ 00159 x0 = *(q31_t *) (px1++); 00160 00161 /* Read state x[n-N-5], x[n-N-6] */ 00162 x1 = *(q31_t *) (px1++); 00163 00164 /* acc2 += b[N-2] * x[n-N-4] + b[N-3] * x[n-N-5] */ 00165 acc2 = __SMLAD(x0, c0, acc2); 00166 00167 /* acc3 += b[N-2] * x[n-N-5] + b[N-3] * x[n-N-6] */ 00168 acc3 = __SMLAD(x1, c0, acc3); 00169 tapCnt--; 00170 00171 } 00172 while(tapCnt > 0u); 00173 00174 /* If the filter length is not a multiple of 4, compute the remaining filter taps. 00175 ** This is always 2 taps since the filter length is always even. */ 00176 if((numTaps & 0x3u) != 0u) 00177 { 00178 /* Read 2 coefficients */ 00179 c0 = *(pb++); 00180 /* Fetch 4 state variables */ 00181 x2 = *(q31_t *) (px1++); 00182 x3 = *(q31_t *) (px1++); 00183 00184 /* Perform the multiply-accumulates */ 00185 acc0 = __SMLAD(x0, c0, acc0); 00186 acc1 = __SMLAD(x1, c0, acc1); 00187 acc2 = __SMLAD(x2, c0, acc2); 00188 acc3 = __SMLAD(x3, c0, acc3); 00189 } 00190 00191 /* The results in the 4 accumulators are in 2.30 format. Convert to 1.15 with saturation. 00192 ** Then store the 4 outputs in the destination buffer. */ 00193 00194 #ifndef ARM_MATH_BIG_ENDIAN 00195 00196 *__SIMD32(pDst)++ = __PKHBT((acc0 >> 15), (acc1 >> 15), 16u); 00197 *__SIMD32(pDst)++ = __PKHBT((acc2 >> 15), (acc3 >> 15), 16u); 00198 00199 #else 00200 00201 *__SIMD32(pDst)++ = __PKHBT((acc1 >> 15), (acc0 >> 15), 16u); 00202 *__SIMD32(pDst)++ = __PKHBT((acc3 >> 15), (acc2 >> 15), 16u); 00203 00204 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00205 00206 /* Advance the state pointer by 4 to process the next group of 4 samples */ 00207 pState = pState + 4; 00208 00209 /* Decrement the loop counter */ 00210 blkCnt--; 00211 } 00212 00213 /* If the blockSize is not a multiple of 4, compute any remaining output samples here. 00214 ** No loop unrolling is used. */ 00215 blkCnt = blockSize % 0x4u; 00216 while(blkCnt > 0u) 00217 { 00218 /* Copy two samples into state buffer */ 00219 *pStateCurnt++ = *pSrc++; 00220 00221 /* Set the accumulator to zero */ 00222 acc0 = 0; 00223 00224 /* Use SIMD to hold states and coefficients */ 00225 px2 = (q31_t *) pState; 00226 pb = (q31_t *) (pCoeffs); 00227 tapCnt = numTaps >> 1; 00228 00229 do 00230 { 00231 acc0 = __SMLAD(*px2++, *(pb++), acc0); 00232 tapCnt--; 00233 } 00234 while(tapCnt > 0u); 00235 00236 /* The result is in 2.30 format. Convert to 1.15 with saturation. 00237 ** Then store the output in the destination buffer. */ 00238 *pDst++ = (q15_t) ((acc0 >> 15)); 00239 00240 /* Advance state pointer by 1 for the next sample */ 00241 pState = pState + 1; 00242 00243 /* Decrement the loop counter */ 00244 blkCnt--; 00245 } 00246 00247 /* Processing is complete. 00248 ** Now copy the last numTaps - 1 samples to the satrt of the state buffer. 00249 ** This prepares the state buffer for the next function call. */ 00250 00251 /* Points to the start of the state buffer */ 00252 pStateCurnt = S->pState; 00253 /* Calculation of count for copying integer writes */ 00254 tapCnt = (numTaps - 1u) >> 2; 00255 00256 while(tapCnt > 0u) 00257 { 00258 *__SIMD32(pStateCurnt)++ = *__SIMD32(pState)++; 00259 *__SIMD32(pStateCurnt)++ = *__SIMD32(pState)++; 00260 00261 tapCnt--; 00262 } 00263 00264 /* Calculation of count for remaining q15_t data */ 00265 tapCnt = (numTaps - 1u) % 0x4u; 00266 00267 /* copy remaining data */ 00268 while(tapCnt > 0u) 00269 { 00270 *pStateCurnt++ = *pState++; 00271 00272 /* Decrement the loop counter */ 00273 tapCnt--; 00274 } 00275 } 00276