lib/main/CMSIS/DSP/Source/MatrixFunctions/arm_mat_mult_q31.c

   1 /* ----------------------------------------------------------------------
   2  * Project:      CMSIS DSP Library
   3  * Title:        arm_mat_mult_q31.c
   4  * Description:  Q31 matrix multiplication
   5  *
   6  * $Date:        27. January 2017
   7  * $Revision:    V.1.5.1
   8  *
   9  * Target Processor: Cortex-M cores
  10  * -------------------------------------------------------------------- */
  11 /*
  12  * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved.
  13  *
  14  * SPDX-License-Identifier: Apache-2.0
  15  *
  16  * Licensed under the Apache License, Version 2.0 (the License); you may
  17  * not use this file except in compliance with the License.
  18  * You may obtain a copy of the License at
  19  *
  20  * www.apache.org/licenses/LICENSE-2.0
  21  *
  22  * Unless required by applicable law or agreed to in writing, software
  23  * distributed under the License is distributed on an AS IS BASIS, WITHOUT
  24  * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  25  * See the License for the specific language governing permissions and
  26  * limitations under the License.
  27  */
  28
  29 #include "arm_math.h"
  30
  31 /**
  32  * @ingroup groupMatrix
  33  */
  34
  35 /**
  36  * @addtogroup MatrixMult
  37  * @{
  38  */
  39
  40 /**
  41  * @brief Q31 matrix multiplication
  42  * @param[in]       *pSrcA points to the first input matrix structure
  43  * @param[in]       *pSrcB points to the second input matrix structure
  44  * @param[out]      *pDst points to output matrix structure
  45  * @return              The function returns either
  46  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
  47  *
  48  * @details
  49  * <b>Scaling and Overflow Behavior:</b>
  50  *
  51  * \par
  52  * The function is implemented using an internal 64-bit accumulator.
  53  * The accumulator has a 2.62 format and maintains full precision of the intermediate
  54  * multiplication results but provides only a single guard bit. There is no saturation
  55  * on intermediate additions. Thus, if the accumulator overflows it wraps around and
  56  * distorts the result. The input signals should be scaled down to avoid intermediate
  57  * overflows. The input is thus scaled down by log2(numColsA) bits
  58  * to avoid overflows, as a total of numColsA additions are performed internally.
  59  * The 2.62 accumulator is right shifted by 31 bits and saturated to 1.31 format to yield the final result.
  60  *
  61  * \par
  62  * See <code>arm_mat_mult_fast_q31()</code> for a faster but less precise implementation of this function for Cortex-M3 and Cortex-M4.
  63  *
  64  */
  65
  66 arm_status arm_mat_mult_q31(
  67   const arm_matrix_instance_q31 * pSrcA,
  68   const arm_matrix_instance_q31 * pSrcB,
  69   arm_matrix_instance_q31 * pDst)
  70 {
  71   q31_t *pIn1 = pSrcA->pData;                    /* input data matrix pointer A */
  72   q31_t *pIn2 = pSrcB->pData;                    /* input data matrix pointer B */
  73   q31_t *pInA = pSrcA->pData;                    /* input data matrix pointer A */
  74   q31_t *pOut = pDst->pData;                     /* output data matrix pointer */
  75   q31_t *px;                                     /* Temporary output data matrix pointer */
  76   q63_t sum;                                     /* Accumulator */
  77   uint16_t numRowsA = pSrcA->numRows;            /* number of rows of input matrix A    */
  78   uint16_t numColsB = pSrcB->numCols;            /* number of columns of input matrix B */
  79   uint16_t numColsA = pSrcA->numCols;            /* number of columns of input matrix A */
  80
  81 #if defined (ARM_MATH_DSP)
  82
  83   /* Run the below code for Cortex-M4 and Cortex-M3 */
  84
  85   uint16_t col, i = 0U, j, row = numRowsA, colCnt;      /* loop counters */
  86   arm_status status;                             /* status of matrix multiplication */
  87   q31_t a0, a1, a2, a3, b0, b1, b2, b3;
  88
  89 #ifdef ARM_MATH_MATRIX_CHECK
  90
  91
  92   /* Check for matrix mismatch condition */
  93   if ((pSrcA->numCols != pSrcB->numRows) ||
  94      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
  95   {
  96     /* Set status as ARM_MATH_SIZE_MISMATCH */
  97     status = ARM_MATH_SIZE_MISMATCH;
  98   }
  99   else
 100 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
 101
 102   {
 103     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
 104     /* row loop */
 105     do
 106     {
 107       /* Output pointer is set to starting address of the row being processed */
 108       px = pOut + i;
 109
 110       /* For every row wise process, the column loop counter is to be initiated */
 111       col = numColsB;
 112
 113       /* For every row wise process, the pIn2 pointer is set
 114        ** to the starting address of the pSrcB data */
 115       pIn2 = pSrcB->pData;
 116
 117       j = 0U;
 118
 119       /* column loop */
 120       do
 121       {
 122         /* Set the variable sum, that acts as accumulator, to zero */
 123         sum = 0;
 124
 125         /* Initiate the pointer pIn1 to point to the starting address of pInA */
 126         pIn1 = pInA;
 127
 128         /* Apply loop unrolling and compute 4 MACs simultaneously. */
 129         colCnt = numColsA >> 2;
 130
 131
 132         /* matrix multiplication */
 133         while (colCnt > 0U)
 134         {
 135           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 136           /* Perform the multiply-accumulates */
 137           b0 = *pIn2;
 138           pIn2 += numColsB;
 139
 140           a0 = *pIn1++;
 141           a1 = *pIn1++;
 142
 143           b1 = *pIn2;
 144           pIn2 += numColsB;
 145           b2 = *pIn2;
 146           pIn2 += numColsB;
 147
 148           sum += (q63_t) a0 *b0;
 149           sum += (q63_t) a1 *b1;
 150
 151           a2 = *pIn1++;
 152           a3 = *pIn1++;
 153
 154           b3 = *pIn2;
 155           pIn2 += numColsB;
 156
 157           sum += (q63_t) a2 *b2;
 158           sum += (q63_t) a3 *b3;
 159
 160           /* Decrement the loop counter */
 161           colCnt--;
 162         }
 163
 164         /* If the columns of pSrcA is not a multiple of 4, compute any remaining output samples here.
 165          ** No loop unrolling is used. */
 166         colCnt = numColsA % 0x4U;
 167
 168         while (colCnt > 0U)
 169         {
 170           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 171           /* Perform the multiply-accumulates */
 172           sum += (q63_t) * pIn1++ * *pIn2;
 173           pIn2 += numColsB;
 174
 175           /* Decrement the loop counter */
 176           colCnt--;
 177         }
 178
 179         /* Convert the result from 2.62 to 1.31 format and store in destination buffer */
 180         *px++ = (q31_t) (sum >> 31);
 181
 182         /* Update the pointer pIn2 to point to the  starting address of the next column */
 183         j++;
 184         pIn2 = (pSrcB->pData) + j;
 185
 186         /* Decrement the column loop counter */
 187         col--;
 188
 189       } while (col > 0U);
 190
 191 #else
 192
 193   /* Run the below code for Cortex-M0 */
 194
 195   q31_t *pInB = pSrcB->pData;                    /* input data matrix pointer B */
 196   uint16_t col, i = 0U, row = numRowsA, colCnt;  /* loop counters */
 197   arm_status status;                             /* status of matrix multiplication */
 198
 199
 200 #ifdef ARM_MATH_MATRIX_CHECK
 201
 202   /* Check for matrix mismatch condition */
 203   if ((pSrcA->numCols != pSrcB->numRows) ||
 204      (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
 205   {
 206     /* Set status as ARM_MATH_SIZE_MISMATCH */
 207     status = ARM_MATH_SIZE_MISMATCH;
 208   }
 209   else
 210 #endif /*    #ifdef ARM_MATH_MATRIX_CHECK    */
 211
 212   {
 213     /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
 214     /* row loop */
 215     do
 216     {
 217       /* Output pointer is set to starting address of the row being processed */
 218       px = pOut + i;
 219
 220       /* For every row wise process, the column loop counter is to be initiated */
 221       col = numColsB;
 222
 223       /* For every row wise process, the pIn2 pointer is set
 224        ** to the starting address of the pSrcB data */
 225       pIn2 = pSrcB->pData;
 226
 227       /* column loop */
 228       do
 229       {
 230         /* Set the variable sum, that acts as accumulator, to zero */
 231         sum = 0;
 232
 233         /* Initiate the pointer pIn1 to point to the starting address of pInA */
 234         pIn1 = pInA;
 235
 236         /* Matrix A columns number of MAC operations are to be performed */
 237         colCnt = numColsA;
 238
 239         /* matrix multiplication */
 240         while (colCnt > 0U)
 241         {
 242           /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
 243           /* Perform the multiply-accumulates */
 244           sum += (q63_t) * pIn1++ * *pIn2;
 245           pIn2 += numColsB;
 246
 247           /* Decrement the loop counter */
 248           colCnt--;
 249         }
 250
 251         /* Convert the result from 2.62 to 1.31 format and store in destination buffer */
 252         *px++ = (q31_t) clip_q63_to_q31(sum >> 31);
 253
 254         /* Decrement the column loop counter */
 255         col--;
 256
 257         /* Update the pointer pIn2 to point to the  starting address of the next column */
 258         pIn2 = pInB + (numColsB - col);
 259
 260       } while (col > 0U);
 261
 262 #endif
 263
 264       /* Update the pointer pInA to point to the  starting address of the next row */
 265       i = i + numColsB;
 266       pInA = pInA + numColsA;
 267
 268       /* Decrement the row loop counter */
 269       row--;
 270
 271     } while (row > 0U);
 272
 273     /* set status as ARM_MATH_SUCCESS */
 274     status = ARM_MATH_SUCCESS;
 275   }
 276   /* Return to application */
 277   return (status);
 278 }
 279
 280 /**
 281  * @} end of MatrixMult group
 282  */