arm_mat_mult_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_mat_mult_fast_q31.c  
00009 *  
00010 * Description:   Q31 matrix multiplication (fast variant).  
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  
00027 #include "arm_math.h" 
00028  
00029 /**  
00030  * @ingroup groupMatrix  
00031  */ 
00032  
00033 /**  
00034  * @addtogroup MatrixMult  
00035  * @{  
00036  */ 
00037  
00038 /**  
00039  * @brief Q31 matrix multiplication (fast variant)  
00040  * @param[in]       *pSrcA points to the first input matrix structure  
00041  * @param[in]       *pSrcB points to the second input matrix structure  
00042  * @param[out]      *pDst points to output matrix structure  
00043  * @return          The function returns either  
00044  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.  
00045  *  
00046  * @details  
00047  * <b>Scaling and Overflow Behavior:</b>  
00048  *  
00049  * \par  
00050  * The difference between the function arm_mat_mult_q31() and this fast variant is that  
00051  * the fast variant use a 32-bit rather than a 64-bit accumulator.  
00052  * The result of each 1.31 x 1.31 multiplication is truncated to  
00053  * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30  
00054  * format. Finally, the accumulator is saturated and converted to a 1.31 result.  
00055  *  
00056  * \par  
00057  * The fast version has the same overflow behavior as the standard version but provides  
00058  * less precision since it discards the low 32 bits of each multiplication result.  
00059  * In order to avoid overflows completely the input signals must be scaled down.  
00060  * Scale down one of the input matrices by log2(numColsA) bits to  
00061  * avoid overflows, as a total of numColsA additions are computed internally for each  
00062  * output element.  
00063  *  
00064  * \par  
00065  * See <code>arm_mat_mult_q31()</code> for a slower implementation of this function  
00066  * which uses 64-bit accumulation to provide higher precision.  
00067  */ 
00068  
00069 arm_status arm_mat_mult_fast_q31( 
00070   const arm_matrix_instance_q31 * pSrcA, 
00071   const arm_matrix_instance_q31 * pSrcB, 
00072   arm_matrix_instance_q31 * pDst) 
00073 { 
00074   q31_t *pIn1 = pSrcA->pData;                    /* input data matrix pointer A */ 
00075   q31_t *pIn2 = pSrcB->pData;                    /* input data matrix pointer B */ 
00076   q31_t *pInA = pSrcA->pData;                    /* input data matrix pointer A */ 
00077 //  q31_t *pSrcB = pSrcB->pData;                    /* input data matrix pointer B */  
00078   q31_t *pOut = pDst->pData;                     /* output data matrix pointer */ 
00079   q31_t *px;                                     /* Temporary output data matrix pointer */ 
00080   q31_t sum;                                     /* Accumulator */ 
00081   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */ 
00082   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */ 
00083   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */ 
00084   uint16_t col, i = 0u, j, row = numRowsA, colCnt;      /* loop counters */ 
00085   arm_status status;                             /* status of matrix multiplication */ 
00086  
00087  
00088 #ifdef ARM_MATH_MATRIX_CHECK 
00089   /* Check for matrix mismatch condition */ 
00090   if((pSrcA->numCols != pSrcB->numRows) || 
00091      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols)) 
00092   { 
00093     /* Set status as ARM_MATH_SIZE_MISMATCH */ 
00094     status = ARM_MATH_SIZE_MISMATCH; 
00095   } 
00096   else 
00097 #endif 
00098   { 
00099     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */ 
00100     /* row loop */ 
00101     do 
00102     { 
00103       /* Output pointer is set to starting address of the row being processed */ 
00104       px = pOut + i; 
00105  
00106       /* For every row wise process, the column loop counter is to be initiated */ 
00107       col = numColsB; 
00108  
00109       /* For every row wise process, the pIn2 pointer is set  
00110        ** to the starting address of the pSrcB data */ 
00111       pIn2 = pSrcB->pData; 
00112  
00113       j = 0u; 
00114  
00115       /* column loop */ 
00116       do 
00117       { 
00118         /* Set the variable sum, that acts as accumulator, to zero */ 
00119         sum = 0; 
00120  
00121         /* Initiate the pointer pIn1 to point to the starting address of pInA */ 
00122         pIn1 = pInA; 
00123  
00124         /* Apply loop unrolling and compute 4 MACs simultaneously. */ 
00125         colCnt = numColsA >> 2; 
00126  
00127  
00128         /* matrix multiplication */ 
00129         while(colCnt > 0u) 
00130         { 
00131           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ 
00132           /* Perform the multiply-accumulates */ 
00133           sum = (q31_t) ((((q63_t) sum << 32) + 
00134                           ((q63_t) * pIn1++ * (*pIn2))) >> 32); 
00135           pIn2 += numColsB; 
00136           sum = (q31_t) ((((q63_t) sum << 32) + 
00137                           ((q63_t) * pIn1++ * (*pIn2))) >> 32); 
00138           pIn2 += numColsB; 
00139           sum = (q31_t) ((((q63_t) sum << 32) + 
00140                           ((q63_t) * pIn1++ * (*pIn2))) >> 32); 
00141           pIn2 += numColsB; 
00142           sum = (q31_t) ((((q63_t) sum << 32) + 
00143                           ((q63_t) * pIn1++ * (*pIn2))) >> 32); 
00144           pIn2 += numColsB; 
00145  
00146           /* Decrement the loop counter */ 
00147           colCnt--; 
00148         } 
00149  
00150         /* If the columns of pSrcA is not a multiple of 4, compute any remaining output samples here.  
00151          ** No loop unrolling is used. */ 
00152         colCnt = numColsA % 0x4u; 
00153  
00154         while(colCnt > 0u) 
00155         { 
00156           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ 
00157           /* Perform the multiply-accumulates */ 
00158           sum = (q31_t) ((((q63_t) sum << 32) + 
00159                           ((q63_t) * pIn1++ * (*pIn2))) >> 32); 
00160           pIn2 += numColsB; 
00161  
00162           /* Decrement the loop counter */ 
00163           colCnt--; 
00164         } 
00165  
00166         /* Convert the result from 2.30 to 1.31 format and store in destination buffer */ 
00167         *px++ = sum << 1; 
00168  
00169         /* Update the pointer pIn2 to point to the  starting address of the next column */ 
00170         j++; 
00171         pIn2 = pSrcB->pData + j; 
00172  
00173         /* Decrement the column loop counter */ 
00174         col--; 
00175  
00176       } while(col > 0u); 
00177  
00178       /* Update the pointer pInA to point to the  starting address of the next row */ 
00179       i = i + numColsB; 
00180       pInA = pInA + numColsA; 
00181  
00182       /* Decrement the row loop counter */ 
00183       row--; 
00184  
00185     } while(row > 0u); 
00186  
00187     /* set status as ARM_MATH_SUCCESS */ 
00188     status = ARM_MATH_SUCCESS; 
00189   } 
00190   /* Return to application */ 
00191   return (status); 
00192 } 
00193  
00194 /**  
00195  * @} end of MatrixMult group  
00196  */