arm_mat_mult_f32.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_f32.c    
00009 *    
00010 * Description:  Floating-point 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  * @defgroup MatrixMult Matrix Multiplication    
00049  *    
00050  * Multiplies two matrices.    
00051  *    
00052  * \image html MatrixMultiplication.gif "Multiplication of two 3 x 3 matrices"    
00053     
00054  * Matrix multiplication is only defined if the number of columns of the    
00055  * first matrix equals the number of rows of the second matrix.    
00056  * Multiplying an <code>M x N</code> matrix with an <code>N x P</code> matrix results    
00057  * in an <code>M x P</code> matrix.    
00058  * When matrix size checking is enabled, the functions check: (1) that the inner dimensions of    
00059  * <code>pSrcA</code> and <code>pSrcB</code> are equal; and (2) that the size of the output    
00060  * matrix equals the outer dimensions of <code>pSrcA</code> and <code>pSrcB</code>.    
00061  */
00062 
00063 
00064 /**    
00065  * @addtogroup MatrixMult    
00066  * @{    
00067  */
00068 
00069 /**    
00070  * @brief Floating-point matrix multiplication.    
00071  * @param[in]       *pSrcA points to the first input matrix structure    
00072  * @param[in]       *pSrcB points to the second input matrix structure    
00073  * @param[out]      *pDst points to output matrix structure    
00074  * @return          The function returns either    
00075  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.    
00076  */
00077 
00078 arm_status arm_mat_mult_f32(
00079   const arm_matrix_instance_f32 * pSrcA,
00080   const arm_matrix_instance_f32 * pSrcB,
00081   arm_matrix_instance_f32 * pDst)
00082 {
00083   float32_t *pIn1 = pSrcA->pData;                /* input data matrix pointer A */
00084   float32_t *pIn2 = pSrcB->pData;                /* input data matrix pointer B */
00085   float32_t *pInA = pSrcA->pData;                /* input data matrix pointer A  */
00086   float32_t *pOut = pDst->pData;                 /* output data matrix pointer */
00087   float32_t *px;                                 /* Temporary output data matrix pointer */
00088   float32_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   float32_t in1, in2, in3, in4;
00098   uint16_t col, i = 0u, j, row = numRowsA, colCnt;      /* loop counters */
00099   arm_status status;                             /* status of matrix multiplication */
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 
00109     /* Set status as ARM_MATH_SIZE_MISMATCH */
00110     status = ARM_MATH_SIZE_MISMATCH;
00111   }
00112   else
00113 #endif /*      #ifdef ARM_MATH_MATRIX_CHECK    */
00114 
00115   {
00116     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00117     /* row loop */
00118     do
00119     {
00120       /* Output pointer is set to starting address of the row being processed */
00121       px = pOut + i;
00122 
00123       /* For every row wise process, the column loop counter is to be initiated */
00124       col = numColsB;
00125 
00126       /* For every row wise process, the pIn2 pointer is set    
00127        ** to the starting address of the pSrcB data */
00128       pIn2 = pSrcB->pData;
00129 
00130       j = 0u;
00131 
00132       /* column loop */
00133       do
00134       {
00135         /* Set the variable sum, that acts as accumulator, to zero */
00136         sum = 0.0f;
00137 
00138         /* Initiate the pointer pIn1 to point to the starting address of the column being processed */
00139         pIn1 = pInA;
00140 
00141         /* Apply loop unrolling and compute 4 MACs simultaneously. */
00142         colCnt = numColsA >> 2u;
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           in3 = *pIn2;
00149           pIn2 += numColsB;
00150           in1 = pIn1[0];
00151           in2 = pIn1[1];
00152           sum += in1 * in3;
00153           in4 = *pIn2;
00154           pIn2 += numColsB;
00155           sum += in2 * in4;
00156 
00157           in3 = *pIn2;
00158           pIn2 += numColsB;
00159           in1 = pIn1[2];
00160           in2 = pIn1[3];
00161           sum += in1 * in3;
00162           in4 = *pIn2;
00163           pIn2 += numColsB;
00164           sum += in2 * in4;
00165           pIn1 += 4u;
00166 
00167           /* Decrement the loop count */
00168           colCnt--;
00169         }
00170 
00171         /* If the columns of pSrcA is not a multiple of 4, compute any remaining MACs here.    
00172          ** No loop unrolling is used. */
00173         colCnt = numColsA % 0x4u;
00174 
00175         while(colCnt > 0u)
00176         {
00177           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00178           sum += *pIn1++ * (*pIn2);
00179           pIn2 += numColsB;
00180 
00181           /* Decrement the loop counter */
00182           colCnt--;
00183         }
00184 
00185         /* Store the result in the destination buffer */
00186         *px++ = sum;
00187 
00188         /* Update the pointer pIn2 to point to the  starting address of the next column */
00189         j++;
00190         pIn2 = pSrcB->pData + j;
00191 
00192         /* Decrement the column loop counter */
00193         col--;
00194 
00195       } while(col > 0u);
00196 
00197 #else
00198 
00199   /* Run the below code for Cortex-M0 */
00200 
00201   float32_t *pInB = pSrcB->pData;                /* input data matrix pointer B */
00202   uint16_t col, i = 0u, row = numRowsA, colCnt;  /* loop counters */
00203   arm_status status;                             /* status of matrix multiplication */
00204 
00205 #ifdef ARM_MATH_MATRIX_CHECK
00206 
00207   /* Check for matrix mismatch condition */
00208   if((pSrcA->numCols != pSrcB->numRows) ||
00209      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00210   {
00211 
00212     /* Set status as ARM_MATH_SIZE_MISMATCH */
00213     status = ARM_MATH_SIZE_MISMATCH;
00214   }
00215   else
00216 #endif /*      #ifdef ARM_MATH_MATRIX_CHECK    */
00217 
00218   {
00219     /* The following loop performs the dot-product of each row in pInA with each column in pInB */
00220     /* row loop */
00221     do
00222     {
00223       /* Output pointer is set to starting address of the row being processed */
00224       px = pOut + i;
00225 
00226       /* For every row wise process, the column loop counter is to be initiated */
00227       col = numColsB;
00228 
00229       /* For every row wise process, the pIn2 pointer is set     
00230        ** to the starting address of the pSrcB data */
00231       pIn2 = pSrcB->pData;
00232 
00233       /* column loop */
00234       do
00235       {
00236         /* Set the variable sum, that acts as accumulator, to zero */
00237         sum = 0.0f;
00238 
00239         /* Initialize the pointer pIn1 to point to the starting address of the row being processed */
00240         pIn1 = pInA;
00241 
00242         /* Matrix A columns number of MAC operations are to be performed */
00243         colCnt = numColsA;
00244 
00245         while(colCnt > 0u)
00246         {
00247           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00248           sum += *pIn1++ * (*pIn2);
00249           pIn2 += numColsB;
00250 
00251           /* Decrement the loop counter */
00252           colCnt--;
00253         }
00254 
00255         /* Store the result in the destination buffer */
00256         *px++ = sum;
00257 
00258         /* Decrement the column loop counter */
00259         col--;
00260 
00261         /* Update the pointer pIn2 to point to the  starting address of the next column */
00262         pIn2 = pInB + (numColsB - col);
00263 
00264       } while(col > 0u);
00265 
00266 #endif /* #ifndef ARM_MATH_CM0_FAMILY */
00267 
00268       /* Update the pointer pInA to point to the  starting address of the next row */
00269       i = i + numColsB;
00270       pInA = pInA + numColsA;
00271 
00272       /* Decrement the row loop counter */
00273       row--;
00274 
00275     } while(row > 0u);
00276     /* Set status as ARM_MATH_SUCCESS */
00277     status = ARM_MATH_SUCCESS;
00278   }
00279 
00280   /* Return to application */
00281   return (status);
00282 }
00283 
00284 /**    
00285  * @} end of MatrixMult group    
00286  */