arm_lms_norm_q15.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_lms_norm_q15.c  
00009 *  
00010 * Description:  Q15 NLMS 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 0.0.7  2010/06/10   
00027 *    Misra-C changes done  
00028 * -------------------------------------------------------------------- */ 
00029  
00030 #include "arm_math.h" 
00031  
00032 /**  
00033  * @ingroup groupFilters  
00034  */ 
00035  
00036 /**  
00037  * @addtogroup LMS_NORM  
00038  * @{  
00039  */ 
00040  
00041 /**  
00042 * @brief Processing function for Q15 normalized LMS filter.  
00043 * @param[in] *S points to an instance of the Q15 normalized LMS filter structure.  
00044 * @param[in] *pSrc points to the block of input data.  
00045 * @param[in] *pRef points to the block of reference data.  
00046 * @param[out] *pOut points to the block of output data.  
00047 * @param[out] *pErr points to the block of error data.  
00048 * @param[in] blockSize number of samples to process.  
00049 * @return none.  
00050 *  
00051 * <b>Scaling and Overflow Behavior:</b>  
00052 * \par  
00053 * The function is implemented using a 64-bit internal accumulator.  
00054 * Both coefficients and state variables are represented in 1.15 format and multiplications yield a 2.30 result.  
00055 * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format.  
00056 * There is no risk of internal overflow with this approach and the full precision of intermediate multiplications is preserved.  
00057 * After all additions have been performed, the accumulator is truncated to 34.15 format by discarding low 15 bits.  
00058 * Lastly, the accumulator is saturated to yield a result in 1.15 format. 
00059 * 
00060 * \par 
00061 *   In this filter, filter coefficients are updated for each sample and the updation of filter cofficients are saturted.  
00062 *  
00063  */ 
00064  
00065 void arm_lms_norm_q15( 
00066   arm_lms_norm_instance_q15 * S, 
00067   q15_t * pSrc, 
00068   q15_t * pRef, 
00069   q15_t * pOut, 
00070   q15_t * pErr, 
00071   uint32_t blockSize) 
00072 { 
00073   q15_t *pState = S->pState;                     /* State pointer */ 
00074   q15_t *pCoeffs = S->pCoeffs;                   /* Coefficient pointer */ 
00075   q15_t *pStateCurnt;                            /* Points to the current sample of the state */ 
00076   q15_t *px, *pb;                                /* Temporary pointers for state and coefficient buffers */ 
00077   q15_t mu = S->mu;                              /* Adaptive factor */ 
00078   uint32_t numTaps = S->numTaps;                 /* Number of filter coefficients in the filter */ 
00079   uint32_t tapCnt, blkCnt;                       /* Loop counters */ 
00080   q31_t energy;                                  /* Energy of the input */ 
00081   q63_t acc;                                     /* Accumulator */ 
00082   q15_t e = 0, d = 0;                            /* error, reference data sample */ 
00083   q15_t w = 0, in;                               /* weight factor and state */ 
00084   q15_t x0;                                      /* temporary variable to hold input sample */ 
00085   uint32_t shift = (uint32_t) S->postShift + 1u; /* Shift to be applied to the output */ 
00086   q15_t errorXmu, oneByEnergy;                   /* Temporary variables to store error and mu product and reciprocal of energy */ 
00087   q15_t postShift;                               /* Post shift to be applied to weight after reciprocal calculation */ 
00088   q31_t coef;                                    /* Teporary variable for coefficient */ 
00089  
00090   energy = S->energy; 
00091   x0 = S->x0; 
00092  
00093   /* S->pState points to buffer which contains previous frame (numTaps - 1) samples */ 
00094   /* pStateCurnt points to the location where the new input data should be written */ 
00095   pStateCurnt = &(S->pState[(numTaps - 1u)]); 
00096  
00097   blkCnt = blockSize; 
00098  
00099   while(blkCnt > 0u) 
00100   { 
00101     /* Copy the new input sample into the state buffer */ 
00102     *pStateCurnt++ = *pSrc; 
00103  
00104     /* Initialize pState pointer */ 
00105     px = pState; 
00106  
00107     /* Initialize coeff pointer */ 
00108     pb = (pCoeffs); 
00109  
00110     /* Read the sample from input buffer */ 
00111     in = *pSrc++; 
00112  
00113     /* Update the energy calculation */ 
00114     energy -= (((q31_t) x0 * (x0)) >> 15); 
00115     energy += (((q31_t) in * (in)) >> 15); 
00116  
00117     /* Set the accumulator to zero */ 
00118     acc = 0; 
00119  
00120     /* Loop unrolling.  Process 4 taps at a time. */ 
00121     tapCnt = numTaps >> 2; 
00122  
00123     while(tapCnt > 0u) 
00124     { 
00125  
00126       /* Perform the multiply-accumulate */ 
00127       acc = __SMLALD(*__SIMD32(px)++, (*__SIMD32(pb)++), acc); 
00128       acc = __SMLALD(*__SIMD32(px)++, (*__SIMD32(pb)++), acc); 
00129  
00130       /* Decrement the loop counter */ 
00131       tapCnt--; 
00132     } 
00133  
00134     /* If the filter length is not a multiple of 4, compute the remaining filter taps */ 
00135     tapCnt = numTaps % 0x4u; 
00136  
00137     while(tapCnt > 0u) 
00138     { 
00139       /* Perform the multiply-accumulate */ 
00140       acc += (((q31_t) * px++ * (*pb++))); 
00141  
00142       /* Decrement the loop counter */ 
00143       tapCnt--; 
00144     } 
00145  
00146     /* Converting the result to 1.15 format */ 
00147     acc = __SSAT((acc >> (16u - shift)), 16u); 
00148  
00149     /* Store the result from accumulator into the destination buffer. */ 
00150     *pOut++ = (q15_t) acc; 
00151  
00152     /* Compute and store error */ 
00153     d = *pRef++; 
00154     e = d - (q15_t) acc; 
00155     *pErr++ = e; 
00156  
00157     /* Calculation of 1/energy */ 
00158     postShift = arm_recip_q15((q15_t) energy + DELTA_Q15, 
00159                               &oneByEnergy, S->recipTable); 
00160  
00161     /* Calculation of e * mu value */ 
00162     errorXmu = (q15_t) (((q31_t) e * mu) >> 15); 
00163  
00164     /* Calculation of (e * mu) * (1/energy) value */ 
00165     acc = (((q31_t) errorXmu * oneByEnergy) >> (15 - postShift)); 
00166  
00167     /* Weighting factor for the normalized version */ 
00168     w = (q15_t) __SSAT((q31_t) acc, 16); 
00169  
00170     /* Initialize pState pointer */ 
00171     px = pState; 
00172  
00173     /* Initialize coeff pointer */ 
00174     pb = (pCoeffs); 
00175  
00176     /* Loop unrolling.  Process 4 taps at a time. */ 
00177     tapCnt = numTaps >> 2; 
00178  
00179     /* Update filter coefficients */ 
00180     while(tapCnt > 0u) 
00181     { 
00182       coef = *pb + (((q31_t) w * (*px++)) >> 15); 
00183       *pb++ = (q15_t) __SSAT((coef), 16); 
00184       coef = *pb + (((q31_t) w * (*px++)) >> 15); 
00185       *pb++ = (q15_t) __SSAT((coef), 16); 
00186       coef = *pb + (((q31_t) w * (*px++)) >> 15); 
00187       *pb++ = (q15_t) __SSAT((coef), 16); 
00188       coef = *pb + (((q31_t) w * (*px++)) >> 15); 
00189       *pb++ = (q15_t) __SSAT((coef), 16); 
00190  
00191       /* Decrement the loop counter */ 
00192       tapCnt--; 
00193     } 
00194  
00195     /* If the filter length is not a multiple of 4, compute the remaining filter taps */ 
00196     tapCnt = numTaps % 0x4u; 
00197  
00198     while(tapCnt > 0u) 
00199     { 
00200       /* Perform the multiply-accumulate */ 
00201       coef = *pb + (((q31_t) w * (*px++)) >> 15); 
00202       *pb++ = (q15_t) __SSAT((coef), 16); 
00203  
00204       /* Decrement the loop counter */ 
00205       tapCnt--; 
00206     } 
00207  
00208     /* Read the sample from state buffer */ 
00209     x0 = *pState; 
00210  
00211     /* Advance state pointer by 1 for the next sample */ 
00212     pState = pState + 1u; 
00213  
00214     /* Decrement the loop counter */ 
00215     blkCnt--; 
00216   } 
00217  
00218   /* Save energy and x0 values for the next frame */ 
00219   S->energy = (q15_t) energy; 
00220   S->x0 = x0; 
00221  
00222   /* Processing is complete. Now copy the last numTaps - 1 samples to the  
00223      satrt of the state buffer. This prepares the state buffer for the  
00224      next function call. */ 
00225  
00226   /* Points to the start of the pState buffer */ 
00227   pStateCurnt = S->pState; 
00228  
00229   /* Calculation of count for copying integer writes */ 
00230   tapCnt = (numTaps - 1u) >> 2; 
00231  
00232   while(tapCnt > 0u) 
00233   { 
00234  
00235     *__SIMD32(pStateCurnt)++ = *__SIMD32(pState)++; 
00236     *__SIMD32(pStateCurnt)++ = *__SIMD32(pState)++; 
00237  
00238     tapCnt--; 
00239  
00240   } 
00241  
00242   /* Calculation of count for remaining q15_t data */ 
00243   tapCnt = (numTaps - 1u) % 0x4u; 
00244  
00245   /* copy data */ 
00246   while(tapCnt > 0u) 
00247   { 
00248     *pStateCurnt++ = *pState++; 
00249  
00250     /* Decrement the loop counter */ 
00251     tapCnt--; 
00252   } 
00253  
00254  
00255 } 
00256  
00257  
00258 /**  
00259    * @} end of LMS_NORM group  
00260    */