arm_mat_mult_f32.c

00001 /* ----------------------------------------------------------------------
00002  * Project:      CMSIS DSP Library
00003  * Title:        arm_mat_mult_f32.c
00004  * Description:  Floating-point matrix multiplication
00005  *
00006  * $Date:        27. January 2017
00007  * $Revision:    V.1.5.1
00008  *
00009  * Target Processor: Cortex-M cores
00010  * -------------------------------------------------------------------- */
00011 /*
00012  * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved.
00013  *
00014  * SPDX-License-Identifier: Apache-2.0
00015  *
00016  * Licensed under the Apache License, Version 2.0 (the License); you may
00017  * not use this file except in compliance with the License.
00018  * You may obtain a copy of the License at
00019  *
00020  * www.apache.org/licenses/LICENSE-2.0
00021  *
00022  * Unless required by applicable law or agreed to in writing, software
00023  * distributed under the License is distributed on an AS IS BASIS, WITHOUT
00024  * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
00025  * See the License for the specific language governing permissions and
00026  * limitations under the License.
00027  */
00028 
00029 #include "arm_math.h"
00030 
00031 /**
00032  * @ingroup groupMatrix
00033  */
00034 
00035 /**
00036  * @defgroup MatrixMult Matrix Multiplication
00037  *
00038  * Multiplies two matrices.
00039  *
00040  * \image html MatrixMultiplication.gif "Multiplication of two 3 x 3 matrices"
00041 
00042  * Matrix multiplication is only defined if the number of columns of the
00043  * first matrix equals the number of rows of the second matrix.
00044  * Multiplying an <code>M x N</code> matrix with an <code>N x P</code> matrix results
00045  * in an <code>M x P</code> matrix.
00046  * When matrix size checking is enabled, the functions check: (1) that the inner dimensions of
00047  * <code>pSrcA</code> and <code>pSrcB</code> are equal; and (2) that the size of the output
00048  * matrix equals the outer dimensions of <code>pSrcA</code> and <code>pSrcB</code>.
00049  */
00050 
00051 
00052 /**
00053  * @addtogroup MatrixMult
00054  * @{
00055  */
00056 
00057 /**
00058  * @brief Floating-point matrix multiplication.
00059  * @param[in]       *pSrcA points to the first input matrix structure
00060  * @param[in]       *pSrcB points to the second input matrix structure
00061  * @param[out]      *pDst points to output matrix structure
00062  * @return          The function returns either
00063  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
00064  */
00065 
00066 arm_status arm_mat_mult_f32(
00067   const arm_matrix_instance_f32 * pSrcA,
00068   const arm_matrix_instance_f32 * pSrcB,
00069   arm_matrix_instance_f32 * pDst)
00070 {
00071   float32_t *pIn1 = pSrcA->pData;                /* input data matrix pointer A */
00072   float32_t *pIn2 = pSrcB->pData;                /* input data matrix pointer B */
00073   float32_t *pInA = pSrcA->pData;                /* input data matrix pointer A  */
00074   float32_t *pOut = pDst->pData;                 /* output data matrix pointer */
00075   float32_t *px;                                 /* Temporary output data matrix pointer */
00076   float32_t sum;                                 /* Accumulator */
00077   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A */
00078   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
00079   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
00080 
00081 #if defined (ARM_MATH_DSP)
00082 
00083   /* Run the below code for Cortex-M4 and Cortex-M3 */
00084 
00085   float32_t in1, in2, in3, in4;
00086   uint16_t col, i = 0U, j, row = numRowsA, colCnt;      /* loop counters */
00087   arm_status status;                             /* status of matrix multiplication */
00088 
00089 #ifdef ARM_MATH_MATRIX_CHECK
00090 
00091 
00092   /* Check for matrix mismatch condition */
00093   if ((pSrcA->numCols != pSrcB->numRows) ||
00094      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00095   {
00096 
00097     /* Set status as ARM_MATH_SIZE_MISMATCH */
00098     status = ARM_MATH_SIZE_MISMATCH;
00099   }
00100   else
00101 #endif /*      #ifdef ARM_MATH_MATRIX_CHECK    */
00102 
00103   {
00104     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
00105     /* row loop */
00106     do
00107     {
00108       /* Output pointer is set to starting address of the row being processed */
00109       px = pOut + i;
00110 
00111       /* For every row wise process, the column loop counter is to be initiated */
00112       col = numColsB;
00113 
00114       /* For every row wise process, the pIn2 pointer is set
00115        ** to the starting address of the pSrcB data */
00116       pIn2 = pSrcB->pData;
00117 
00118       j = 0U;
00119 
00120       /* column loop */
00121       do
00122       {
00123         /* Set the variable sum, that acts as accumulator, to zero */
00124         sum = 0.0f;
00125 
00126         /* Initiate the pointer pIn1 to point to the starting address of the column being processed */
00127         pIn1 = pInA;
00128 
00129         /* Apply loop unrolling and compute 4 MACs simultaneously. */
00130         colCnt = numColsA >> 2U;
00131 
00132         /* matrix multiplication        */
00133         while (colCnt > 0U)
00134         {
00135           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00136           in3 = *pIn2;
00137           pIn2 += numColsB;
00138           in1 = pIn1[0];
00139           in2 = pIn1[1];
00140           sum += in1 * in3;
00141           in4 = *pIn2;
00142           pIn2 += numColsB;
00143           sum += in2 * in4;
00144 
00145           in3 = *pIn2;
00146           pIn2 += numColsB;
00147           in1 = pIn1[2];
00148           in2 = pIn1[3];
00149           sum += in1 * in3;
00150           in4 = *pIn2;
00151           pIn2 += numColsB;
00152           sum += in2 * in4;
00153           pIn1 += 4U;
00154 
00155           /* Decrement the loop count */
00156           colCnt--;
00157         }
00158 
00159         /* If the columns of pSrcA is not a multiple of 4, compute any remaining MACs here.
00160          ** No loop unrolling is used. */
00161         colCnt = numColsA % 0x4U;
00162 
00163         while (colCnt > 0U)
00164         {
00165           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00166           sum += *pIn1++ * (*pIn2);
00167           pIn2 += numColsB;
00168 
00169           /* Decrement the loop counter */
00170           colCnt--;
00171         }
00172 
00173         /* Store the result in the destination buffer */
00174         *px++ = sum;
00175 
00176         /* Update the pointer pIn2 to point to the  starting address of the next column */
00177         j++;
00178         pIn2 = pSrcB->pData + j;
00179 
00180         /* Decrement the column loop counter */
00181         col--;
00182 
00183       } while (col > 0U);
00184 
00185 #else
00186 
00187   /* Run the below code for Cortex-M0 */
00188 
00189   float32_t *pInB = pSrcB->pData;                /* input data matrix pointer B */
00190   uint16_t col, i = 0U, row = numRowsA, colCnt;  /* loop counters */
00191   arm_status status;                             /* status of matrix multiplication */
00192 
00193 #ifdef ARM_MATH_MATRIX_CHECK
00194 
00195   /* Check for matrix mismatch condition */
00196   if ((pSrcA->numCols != pSrcB->numRows) ||
00197      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
00198   {
00199 
00200     /* Set status as ARM_MATH_SIZE_MISMATCH */
00201     status = ARM_MATH_SIZE_MISMATCH;
00202   }
00203   else
00204 #endif /*      #ifdef ARM_MATH_MATRIX_CHECK    */
00205 
00206   {
00207     /* The following loop performs the dot-product of each row in pInA with each column in pInB */
00208     /* row loop */
00209     do
00210     {
00211       /* Output pointer is set to starting address of the row being processed */
00212       px = pOut + i;
00213 
00214       /* For every row wise process, the column loop counter is to be initiated */
00215       col = numColsB;
00216 
00217       /* For every row wise process, the pIn2 pointer is set
00218        ** to the starting address of the pSrcB data */
00219       pIn2 = pSrcB->pData;
00220 
00221       /* column loop */
00222       do
00223       {
00224         /* Set the variable sum, that acts as accumulator, to zero */
00225         sum = 0.0f;
00226 
00227         /* Initialize the pointer pIn1 to point to the starting address of the row being processed */
00228         pIn1 = pInA;
00229 
00230         /* Matrix A columns number of MAC operations are to be performed */
00231         colCnt = numColsA;
00232 
00233         while (colCnt > 0U)
00234         {
00235           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
00236           sum += *pIn1++ * (*pIn2);
00237           pIn2 += numColsB;
00238 
00239           /* Decrement the loop counter */
00240           colCnt--;
00241         }
00242 
00243         /* Store the result in the destination buffer */
00244         *px++ = sum;
00245 
00246         /* Decrement the column loop counter */
00247         col--;
00248 
00249         /* Update the pointer pIn2 to point to the  starting address of the next column */
00250         pIn2 = pInB + (numColsB - col);
00251 
00252       } while (col > 0U);
00253 
00254 #endif /* #if defined (ARM_MATH_DSP) */
00255 
00256       /* Update the pointer pInA to point to the  starting address of the next row */
00257       i = i + numColsB;
00258       pInA = pInA + numColsA;
00259 
00260       /* Decrement the row loop counter */
00261       row--;
00262 
00263     } while (row > 0U);
00264     /* Set status as ARM_MATH_SUCCESS */
00265     status = ARM_MATH_SUCCESS;
00266   }
00267 
00268   /* Return to application */
00269   return (status);
00270 }
00271 
00272 /**
00273  * @} end of MatrixMult group
00274  */
00275