arm_mat_mult_q31.c

00001 /* ----------------------------------------------------------------------    
00002 * Copyright (C) 2010-2014 ARM Limited. All rights reserved.    
00003 *    
00004 * $Date:        19. March 2015
00005 * $Revision:    V.1.4.5
00006 *    
00007 * Project:      CMSIS DSP Library    
00008 * Title:        arm_mat_mult_q31.c    
00009 *    
00010 * Description:   Q31 matrix multiplication.    
00011 *    
00012 * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
00013 *  
00014 * Redistribution and use in source and binary forms, with or without 
00015 * modification, are permitted provided that the following conditions
00016 * are met:
00017 *   - Redistributions of source code must retain the above copyright
00018 *     notice, this list of conditions and the following disclaimer.
00019 *   - Redistributions in binary form must reproduce the above copyright
00020 *     notice, this list of conditions and the following disclaimer in
00021 *     the documentation and/or other materials provided with the 
00022 *     distribution.
00023 *   - Neither the name of ARM LIMITED nor the names of its contributors
00024 *     may be used to endorse or promote products derived from this
00025 *     software without specific prior written permission.
00026 *
00027 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
00028 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
00029 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
00030 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 
00031 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
00032 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
00033 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
00034 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
00035 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
00036 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
00037 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
00038 * POSSIBILITY OF SUCH DAMAGE.   
00039 * -------------------------------------------------------------------- */
00040 
00041 #include "arm_math.h"
00042 
00043 /**    
00044  * @ingroup groupMatrix    
00045  */
00046 
00047 /**    
00048  * @addtogroup MatrixMult    
00049  * @{    
00050  */
00051 
00052 /**    
00053  * @brief Q31 matrix multiplication    
00054  * @param[in]       *pSrcA points to the first input matrix structure    
00055  * @param[in]       *pSrcB points to the second input matrix structure    
00056  * @param[out]      *pDst points to output matrix structure    
00057  * @return          The function returns either    
00058  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.    
00059  *    
00060  * @details    
00061  * <b>Scaling and Overflow Behavior:</b>    
00062  *    
00063  * \par    
00064  * The function is implemented using an internal 64-bit accumulator.    
00065  * The accumulator has a 2.62 format and maintains full precision of the intermediate    
00066  * multiplication results but provides only a single guard bit. There is no saturation    
00067  * on intermediate additions. Thus, if the accumulator overflows it wraps around and    
00068  * distorts the result. The input signals should be scaled down to avoid intermediate    
00069  * overflows. The input is thus scaled down by log2(numColsA) bits    
00070  * to avoid overflows, as a total of numColsA additions are performed internally.    
00071  * The 2.62 accumulator is right shifted by 31 bits and saturated to 1.31 format to yield the final result.    
00072  *    
00073  * \par    
00074  * See <code>arm_mat_mult_fast_q31()</code> for a faster but less precise implementation of this function for Cortex-M3 and Cortex-M4.    
00075  *    
00076  */
00077 
00078 arm_status arm_mat_mult_q31(
00079   const arm_matrix_instance_q31 * pSrcA,
00080   const arm_matrix_instance_q31 * pSrcB,
00081   arm_matrix_instance_q31 * pDst)
00082 {
00083   q31_t *pIn1 = pSrcA->pData;                    /* input data matrix pointer A */
00084   q31_t *pIn2 = pSrcB->pData;                    /* input data matrix pointer B */
00085   q31_t *pInA = pSrcA->pData;                    /* input data matrix pointer A */
00086   q31_t *pOut = pDst->pData;                     /* output data matrix pointer */
00087   q31_t *px;                                     /* Temporary output data matrix pointer */
00088   q63_t sum;                                     /* Accumulator */
00089   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
00090   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
00091   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
00092 
00093 #ifndef ARM_MATH_CM0_FAMILY
00094 
00095   /* Run the below code for Cortex-M4 and Cortex-M3 */
00096 
00097   uint16_t col, i = 0u, j, row = numRowsA, colCnt;      /* loop counters */
00098   arm_status status;                             /* status of matrix multiplication */
00099   q31_t a0, a1, a2, a3, b0, b1, b2, b3;
00100 
00101 #ifdef ARM_MATH_MATRIX_CHECK
00102 
00103 
00104   /* Check for matrix mismatch condition */
00105   if((pSrcA->numCols != pSrcB->numRows) ||
00106      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00107   {
00108     /* Set status as ARM_MATH_SIZE_MISMATCH */
00109     status = ARM_MATH_SIZE_MISMATCH;
00110   }
00111   else
00112 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
00113 
00114   {
00115     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00116     /* row loop */
00117     do
00118     {
00119       /* Output pointer is set to starting address of the row being processed */
00120       px = pOut + i;
00121 
00122       /* For every row wise process, the column loop counter is to be initiated */
00123       col = numColsB;
00124 
00125       /* For every row wise process, the pIn2 pointer is set    
00126        ** to the starting address of the pSrcB data */
00127       pIn2 = pSrcB->pData;
00128 
00129       j = 0u;
00130 
00131       /* column loop */
00132       do
00133       {
00134         /* Set the variable sum, that acts as accumulator, to zero */
00135         sum = 0;
00136 
00137         /* Initiate the pointer pIn1 to point to the starting address of pInA */
00138         pIn1 = pInA;
00139 
00140         /* Apply loop unrolling and compute 4 MACs simultaneously. */
00141         colCnt = numColsA >> 2;
00142 
00143 
00144         /* matrix multiplication */
00145         while(colCnt > 0u)
00146         {
00147           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00148           /* Perform the multiply-accumulates */
00149           b0 = *pIn2;
00150           pIn2 += numColsB;
00151 
00152           a0 = *pIn1++;
00153           a1 = *pIn1++;
00154 
00155           b1 = *pIn2;
00156           pIn2 += numColsB;
00157           b2 = *pIn2;
00158           pIn2 += numColsB;
00159 
00160           sum += (q63_t) a0 *b0;
00161           sum += (q63_t) a1 *b1;
00162 
00163           a2 = *pIn1++;
00164           a3 = *pIn1++;
00165 
00166           b3 = *pIn2;
00167           pIn2 += numColsB;
00168 
00169           sum += (q63_t) a2 *b2;
00170           sum += (q63_t) a3 *b3;
00171 
00172           /* Decrement the loop counter */
00173           colCnt--;
00174         }
00175 
00176         /* If the columns of pSrcA is not a multiple of 4, compute any remaining output samples here.    
00177          ** No loop unrolling is used. */
00178         colCnt = numColsA % 0x4u;
00179 
00180         while(colCnt > 0u)
00181         {
00182           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00183           /* Perform the multiply-accumulates */
00184           sum += (q63_t) * pIn1++ * *pIn2;
00185           pIn2 += numColsB;
00186 
00187           /* Decrement the loop counter */
00188           colCnt--;
00189         }
00190 
00191         /* Convert the result from 2.62 to 1.31 format and store in destination buffer */
00192         *px++ = (q31_t) (sum >> 31);
00193 
00194         /* Update the pointer pIn2 to point to the  starting address of the next column */
00195         j++;
00196         pIn2 = (pSrcB->pData) + j;
00197 
00198         /* Decrement the column loop counter */
00199         col--;
00200 
00201       } while(col > 0u);
00202 
00203 #else
00204 
00205   /* Run the below code for Cortex-M0 */
00206 
00207   q31_t *pInB = pSrcB->pData;                    /* input data matrix pointer B */
00208   uint16_t col, i = 0u, row = numRowsA, colCnt;  /* loop counters */
00209   arm_status status;                             /* status of matrix multiplication */
00210 
00211 
00212 #ifdef ARM_MATH_MATRIX_CHECK
00213 
00214   /* Check for matrix mismatch condition */
00215   if((pSrcA->numCols != pSrcB->numRows) ||
00216      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00217   {
00218     /* Set status as ARM_MATH_SIZE_MISMATCH */
00219     status = ARM_MATH_SIZE_MISMATCH;
00220   }
00221   else
00222 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
00223 
00224   {
00225     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00226     /* row loop */
00227     do
00228     {
00229       /* Output pointer is set to starting address of the row being processed */
00230       px = pOut + i;
00231 
00232       /* For every row wise process, the column loop counter is to be initiated */
00233       col = numColsB;
00234 
00235       /* For every row wise process, the pIn2 pointer is set          
00236        ** to the starting address of the pSrcB data */
00237       pIn2 = pSrcB->pData;
00238 
00239       /* column loop */
00240       do
00241       {
00242         /* Set the variable sum, that acts as accumulator, to zero */
00243         sum = 0;
00244 
00245         /* Initiate the pointer pIn1 to point to the starting address of pInA */
00246         pIn1 = pInA;
00247 
00248         /* Matrix A columns number of MAC operations are to be performed */
00249         colCnt = numColsA;
00250 
00251         /* matrix multiplication */
00252         while(colCnt > 0u)
00253         {
00254           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00255           /* Perform the multiply-accumulates */
00256           sum += (q63_t) * pIn1++ * *pIn2;
00257           pIn2 += numColsB;
00258 
00259           /* Decrement the loop counter */
00260           colCnt--;
00261         }
00262 
00263         /* Convert the result from 2.62 to 1.31 format and store in destination buffer */
00264         *px++ = (q31_t) clip_q63_to_q31(sum >> 31);
00265 
00266         /* Decrement the column loop counter */
00267         col--;
00268 
00269         /* Update the pointer pIn2 to point to the  starting address of the next column */
00270         pIn2 = pInB + (numColsB - col);
00271 
00272       } while(col > 0u);
00273 
00274 #endif
00275 
00276       /* Update the pointer pInA to point to the  starting address of the next row */
00277       i = i + numColsB;
00278       pInA = pInA + numColsA;
00279 
00280       /* Decrement the row loop counter */
00281       row--;
00282 
00283     } while(row > 0u);
00284 
00285     /* set status as ARM_MATH_SUCCESS */
00286     status = ARM_MATH_SUCCESS;
00287   }
00288   /* Return to application */
00289   return (status);
00290 }
00291 
00292 /**    
00293  * @} end of MatrixMult group    
00294  */