arm_fir_fast_q31.c

00001 /* ----------------------------------------------------------------------  
00002 * Copyright (C) 2010 ARM Limited. All rights reserved.  
00003 *  
00004 * $Date:        29. November 2010  
00005 * $Revision:    V1.0.3  
00006 *  
00007 * Project:      CMSIS DSP Library  
00008 * Title:        arm_fir_fast_q31.c  
00009 *  
00010 * Description:  Processing function for the Q31 Fast FIR filter.  
00011 *  
00012 * Target Processor: Cortex-M4/Cortex-M3
00013 *  
00014 * Version 1.0.3 2010/11/29 
00015 *    Re-organized the CMSIS folders and updated documentation.  
00016 *   
00017 * Version 1.0.2 2010/11/11  
00018 *    Documentation updated.   
00019 *  
00020 * Version 1.0.1 2010/10/05   
00021 *    Production release and review comments incorporated.  
00022 *  
00023 * Version 1.0.0 2010/09/20   
00024 *    Production release and review comments incorporated.  
00025 *  
00026 * Version 0.0.9  2010/08/27   
00027 *    Initial version  
00028 * -------------------------------------------------------------------- */ 
00029  
00030 #include "arm_math.h" 
00031  
00032 /**  
00033  * @ingroup groupFilters  
00034  */ 
00035  
00036 /**  
00037  * @addtogroup FIR  
00038  * @{  
00039  */ 
00040  
00041 /**  
00042  * @param[in] *S points to an instance of the Q31 structure.  
00043  * @param[in] *pSrc points to the block of input data.  
00044  * @param[out] *pDst points to the block output data.  
00045  * @param[in] blockSize number of samples to process per call.  
00046  * @return none.  
00047  *  
00048  * <b>Scaling and Overflow Behavior:</b>  
00049  *  
00050  * \par  
00051  * This function is optimized for speed at the expense of fixed-point precision and overflow protection.  
00052  * The result of each 1.31 x 1.31 multiplication is truncated to 2.30 format.  
00053  * These intermediate results are added to a 2.30 accumulator.  
00054  * Finally, the accumulator is saturated and converted to a 1.31 result.  
00055  * The fast version has the same overflow behavior as the standard version and provides less precision since it discards the low 32 bits of each multiplication result.  
00056  * In order to avoid overflows completely the input signal must be scaled down by log2(numTaps) bits.  
00057  *  
00058  * \par  
00059  * Refer to the function <code>arm_fir_q31()</code> for a slower implementation of this function which uses a 64-bit accumulator to provide higher precision.  Both the slow and the fast versions use the same instance structure.  
00060  * Use the function <code>arm_fir_init_q31()</code> to initialize the filter structure.  
00061  */ 
00062  
00063 void arm_fir_fast_q31( 
00064   const arm_fir_instance_q31 * S, 
00065   q31_t * pSrc, 
00066   q31_t * pDst, 
00067   uint32_t blockSize) 
00068 { 
00069   q31_t *pState = S->pState;                     /* State pointer */ 
00070   q31_t *pCoeffs = S->pCoeffs;                   /* Coefficient pointer */ 
00071   q31_t *pStateCurnt;                            /* Points to the current sample of the state */ 
00072   q31_t x0, x1, x2, x3;                          /* Temporary variables to hold state */ 
00073   q31_t c0;                                      /* Temporary variable to hold coefficient value */ 
00074   q31_t *px;                                     /* Temporary pointer for state */ 
00075   q31_t *pb;                                     /* Temporary pointer for coefficient buffer */ 
00076   q63_t acc0, acc1, acc2, acc3;                  /* Accumulators */ 
00077   uint32_t numTaps = S->numTaps;                 /* Number of filter coefficients in the filter */ 
00078   uint32_t i, tapCnt, blkCnt;                    /* Loop counters */ 
00079  
00080   /* S->pState points to buffer which contains previous frame (numTaps - 1) samples */ 
00081   /* pStateCurnt points to the location where the new input data should be written */ 
00082   pStateCurnt = &(S->pState[(numTaps - 1u)]); 
00083  
00084   /* Apply loop unrolling and compute 4 output values simultaneously.  
00085    * The variables acc0 ... acc3 hold output values that are being computed:  
00086    *  
00087    *    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]  
00088    *    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]  
00089    *    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]  
00090    *    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]  
00091    */ 
00092   blkCnt = blockSize >> 2; 
00093  
00094   /* First part of the processing with loop unrolling.  Compute 4 outputs at a time.  
00095    ** a second loop below computes the remaining 1 to 3 samples. */ 
00096   while(blkCnt > 0u) 
00097   { 
00098     /* Copy four new input samples into the state buffer */ 
00099     *pStateCurnt++ = *pSrc++; 
00100     *pStateCurnt++ = *pSrc++; 
00101     *pStateCurnt++ = *pSrc++; 
00102     *pStateCurnt++ = *pSrc++; 
00103  
00104     /* Set all accumulators to zero */ 
00105     acc0 = 0; 
00106     acc1 = 0; 
00107     acc2 = 0; 
00108     acc3 = 0; 
00109  
00110     /* Initialize state pointer */ 
00111     px = pState; 
00112  
00113     /* Initialize coefficient pointer */ 
00114     pb = pCoeffs; 
00115  
00116     /* Read the first three samples from the state buffer:  
00117      *  x[n-numTaps], x[n-numTaps-1], x[n-numTaps-2] */ 
00118     x0 = *(px++); 
00119     x1 = *(px++); 
00120     x2 = *(px++); 
00121  
00122     /* Loop unrolling.  Process 4 taps at a time. */ 
00123     tapCnt = numTaps >> 2; 
00124     i = tapCnt; 
00125  
00126     while(i > 0u) 
00127     { 
00128       /* Read the b[numTaps] coefficient */ 
00129       c0 = *(pb++); 
00130  
00131       /* Read x[n-numTaps-3] sample */ 
00132       x3 = *(px++); 
00133  
00134       /* acc0 +=  b[numTaps] * x[n-numTaps] */ 
00135       acc0 = (q31_t) ((((q63_t) x0 * c0) + (acc0 << 32)) >> 32); 
00136  
00137       /* acc1 +=  b[numTaps] * x[n-numTaps-1] */ 
00138       acc1 = (q31_t) ((((q63_t) x1 * c0) + (acc1 << 32)) >> 32); 
00139  
00140       /* acc2 +=  b[numTaps] * x[n-numTaps-2] */ 
00141       acc2 = (q31_t) ((((q63_t) x2 * c0) + (acc2 << 32)) >> 32); 
00142  
00143       /* acc3 +=  b[numTaps] * x[n-numTaps-3] */ 
00144       acc3 = (q31_t) ((((q63_t) x3 * c0) + (acc3 << 32)) >> 32); 
00145  
00146       /* Read the b[numTaps-1] coefficient */ 
00147       c0 = *(pb++); 
00148  
00149       /* Read x[n-numTaps-4] sample */ 
00150       x0 = *(px++); 
00151  
00152       /* Perform the multiply-accumulates */ 
00153       acc0 = (q31_t) ((((q63_t) x1 * c0) + (acc0 << 32)) >> 32); 
00154       acc1 = (q31_t) ((((q63_t) x2 * c0) + (acc1 << 32)) >> 32); 
00155       acc2 = (q31_t) ((((q63_t) x3 * c0) + (acc2 << 32)) >> 32); 
00156       acc3 = (q31_t) ((((q63_t) x0 * c0) + (acc3 << 32)) >> 32); 
00157  
00158       /* Read the b[numTaps-2] coefficient */ 
00159       c0 = *(pb++); 
00160  
00161       /* Read x[n-numTaps-5] sample */ 
00162       x1 = *(px++); 
00163  
00164       /* Perform the multiply-accumulates */ 
00165       acc0 = (q31_t) ((((q63_t) x2 * c0) + (acc0 << 32)) >> 32); 
00166       acc1 = (q31_t) ((((q63_t) x3 * c0) + (acc1 << 32)) >> 32); 
00167       acc2 = (q31_t) ((((q63_t) x0 * c0) + (acc2 << 32)) >> 32); 
00168       acc3 = (q31_t) ((((q63_t) x1 * c0) + (acc3 << 32)) >> 32); 
00169  
00170       /* Read the b[numTaps-3] coefficients */ 
00171       c0 = *(pb++); 
00172  
00173       /* Read x[n-numTaps-6] sample */ 
00174       x2 = *(px++); 
00175  
00176       /* Perform the multiply-accumulates */ 
00177       acc0 = (q31_t) ((((q63_t) x3 * c0) + (acc0 << 32)) >> 32); 
00178       acc1 = (q31_t) ((((q63_t) x0 * c0) + (acc1 << 32)) >> 32); 
00179       acc2 = (q31_t) ((((q63_t) x1 * c0) + (acc2 << 32)) >> 32); 
00180       acc3 = (q31_t) ((((q63_t) x2 * c0) + (acc3 << 32)) >> 32); 
00181       i--; 
00182     } 
00183  
00184     /* If the filter length is not a multiple of 4, compute the remaining filter taps */ 
00185  
00186     i = numTaps - (tapCnt * 4u); 
00187     while(i > 0u) 
00188     { 
00189       /* Read coefficients */ 
00190       c0 = *(pb++); 
00191  
00192       /* Fetch 1 state variable */ 
00193       x3 = *(px++); 
00194  
00195       /* Perform the multiply-accumulates */ 
00196       acc0 = (q31_t) ((((q63_t) x0 * c0) + (acc0 << 32)) >> 32); 
00197       acc1 = (q31_t) ((((q63_t) x1 * c0) + (acc1 << 32)) >> 32); 
00198       acc2 = (q31_t) ((((q63_t) x2 * c0) + (acc2 << 32)) >> 32); 
00199       acc3 = (q31_t) ((((q63_t) x3 * c0) + (acc3 << 32)) >> 32); 
00200  
00201       /* Reuse the present sample states for next sample */ 
00202       x0 = x1; 
00203       x1 = x2; 
00204       x2 = x3; 
00205  
00206       /* Decrement the loop counter */ 
00207       i--; 
00208     } 
00209  
00210     /* Advance the state pointer by 4 to process the next group of 4 samples */ 
00211     pState = pState + 4; 
00212  
00213     /* The results in the 4 accumulators are in 2.30 format.  Convert to 1.31  
00214      ** Then store the 4 outputs in the destination buffer. */ 
00215     *pDst++ = (q31_t) (acc0 << 1); 
00216     *pDst++ = (q31_t) (acc1 << 1); 
00217     *pDst++ = (q31_t) (acc2 << 1); 
00218     *pDst++ = (q31_t) (acc3 << 1); 
00219  
00220     /* Decrement the samples loop counter */ 
00221     blkCnt--; 
00222   } 
00223  
00224  
00225   /* If the blockSize is not a multiple of 4, compute any remaining output samples here.  
00226    ** No loop unrolling is used. */ 
00227   blkCnt = blockSize % 4u; 
00228  
00229   while(blkCnt > 0u) 
00230   { 
00231     /* Copy one sample at a time into state buffer */ 
00232     *pStateCurnt++ = *pSrc++; 
00233  
00234     /* Set the accumulator to zero */ 
00235     acc0 = 0; 
00236  
00237     /* Initialize state pointer */ 
00238     px = pState; 
00239  
00240     /* Initialize Coefficient pointer */ 
00241     pb = (pCoeffs); 
00242  
00243     i = numTaps; 
00244  
00245     /* Perform the multiply-accumulates */ 
00246     do 
00247     { 
00248       acc0 = (q31_t) ((((q63_t) * (px++) * (*(pb++))) + (acc0 << 32)) >> 32); 
00249       i--; 
00250     } while(i > 0u); 
00251  
00252     /* The result is in 2.30 format.  Convert to 1.31  
00253      ** Then store the output in the destination buffer. */ 
00254     *pDst++ = (q31_t) (acc0 << 1); 
00255  
00256     /* Advance state pointer by 1 for the next sample */ 
00257     pState = pState + 1; 
00258  
00259     /* Decrement the samples loop counter */ 
00260     blkCnt--; 
00261   } 
00262  
00263   /* Processing is complete.  
00264    ** Now copy the last numTaps - 1 samples to the satrt of the state buffer.  
00265    ** This prepares the state buffer for the next function call. */ 
00266  
00267   /* Points to the start of the state buffer */ 
00268   pStateCurnt = S->pState; 
00269  
00270   tapCnt = (numTaps - 1u) >> 2u; 
00271  
00272   /* copy data */ 
00273   while(tapCnt > 0u) 
00274   { 
00275     *pStateCurnt++ = *pState++; 
00276     *pStateCurnt++ = *pState++; 
00277     *pStateCurnt++ = *pState++; 
00278     *pStateCurnt++ = *pState++; 
00279  
00280     /* Decrement the loop counter */ 
00281     tapCnt--; 
00282   } 
00283  
00284   /* Calculate remaining number of copies */ 
00285   tapCnt = (numTaps - 1u) % 0x4u; 
00286  
00287   /* Copy the remaining q31_t data */ 
00288   while(tapCnt > 0u) 
00289   { 
00290     *pStateCurnt++ = *pState++; 
00291  
00292     /* Decrement the loop counter */ 
00293     tapCnt--; 
00294   } 
00295  
00296 } 
00297  
00298 /**  
00299  * @} end of FIR group  
00300  */