Laxmi Kant Tiwari / mbed-dsp

Important changes to repositories hosted on mbed.com

Mbed hosted mercurial repositories are deprecated and are due to be permanently deleted in July 2026.

To keep a copy of this software download the repository Zip archive or clone locally using Mercurial.

It is also possible to export all your personal repositories from the account settings page.

Fork of mbed-dsp by mbed official

cmsis_dsp/MatrixFunctions/arm_mat_mult_fast_q31.c@3:7a284390b0ce, 2013-11-08 (annotated)

Committer:
mbed_official
Date:
Fri Nov 08 13:45:10 2013 +0000
Revision:
3:7a284390b0ce
Parent:
2:da51fb522205 
Synchronized with git revision e69956aba2f68a2a26ac26b051f8d349deaa1ce8

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emilmont 1:fdd22bb7aa52 1 /* ----------------------------------------------------------------------
mbed_official 3:7a284390b0ce 2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
emilmont 1:fdd22bb7aa52 3 *
mbed_official 3:7a284390b0ce 4 * $Date: 17. January 2013
mbed_official 3:7a284390b0ce 5 * $Revision: V1.4.1
emilmont 1:fdd22bb7aa52 6 *
emilmont 2:da51fb522205 7 * Project: CMSIS DSP Library
emilmont 2:da51fb522205 8 * Title: arm_mat_mult_fast_q31.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 2:da51fb522205 10 * Description: Q31 matrix multiplication (fast variant).
emilmont 1:fdd22bb7aa52 11 *
emilmont 1:fdd22bb7aa52 12 * Target Processor: Cortex-M4/Cortex-M3
emilmont 1:fdd22bb7aa52 13 *
mbed_official 3:7a284390b0ce 14 * Redistribution and use in source and binary forms, with or without
mbed_official 3:7a284390b0ce 15 * modification, are permitted provided that the following conditions
mbed_official 3:7a284390b0ce 16 * are met:
mbed_official 3:7a284390b0ce 17 * - Redistributions of source code must retain the above copyright
mbed_official 3:7a284390b0ce 18 * notice, this list of conditions and the following disclaimer.
mbed_official 3:7a284390b0ce 19 * - Redistributions in binary form must reproduce the above copyright
mbed_official 3:7a284390b0ce 20 * notice, this list of conditions and the following disclaimer in
mbed_official 3:7a284390b0ce 21 * the documentation and/or other materials provided with the
mbed_official 3:7a284390b0ce 22 * distribution.
mbed_official 3:7a284390b0ce 23 * - Neither the name of ARM LIMITED nor the names of its contributors
mbed_official 3:7a284390b0ce 24 * may be used to endorse or promote products derived from this
mbed_official 3:7a284390b0ce 25 * software without specific prior written permission.
mbed_official 3:7a284390b0ce 26 *
mbed_official 3:7a284390b0ce 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
mbed_official 3:7a284390b0ce 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
mbed_official 3:7a284390b0ce 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
mbed_official 3:7a284390b0ce 30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
mbed_official 3:7a284390b0ce 31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
mbed_official 3:7a284390b0ce 32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
mbed_official 3:7a284390b0ce 33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
mbed_official 3:7a284390b0ce 34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
mbed_official 3:7a284390b0ce 35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
mbed_official 3:7a284390b0ce 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
mbed_official 3:7a284390b0ce 37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
mbed_official 3:7a284390b0ce 38 * POSSIBILITY OF SUCH DAMAGE.
emilmont 1:fdd22bb7aa52 39 * -------------------------------------------------------------------- */
emilmont 1:fdd22bb7aa52 40
emilmont 1:fdd22bb7aa52 41 #include "arm_math.h"
emilmont 1:fdd22bb7aa52 42
emilmont 1:fdd22bb7aa52 43 /**
emilmont 1:fdd22bb7aa52 44 * @ingroup groupMatrix
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @addtogroup MatrixMult
emilmont 1:fdd22bb7aa52 49 * @{
emilmont 1:fdd22bb7aa52 50 */
emilmont 1:fdd22bb7aa52 51
emilmont 1:fdd22bb7aa52 52 /**
emilmont 1:fdd22bb7aa52 53 * @brief Q31 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4
emilmont 1:fdd22bb7aa52 54 * @param[in] *pSrcA points to the first input matrix structure
emilmont 1:fdd22bb7aa52 55 * @param[in] *pSrcB points to the second input matrix structure
emilmont 1:fdd22bb7aa52 56 * @param[out] *pDst points to output matrix structure
emilmont 2:da51fb522205 57 * @return The function returns either
emilmont 1:fdd22bb7aa52 58 * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
emilmont 1:fdd22bb7aa52 59 *
emilmont 1:fdd22bb7aa52 60 * @details
emilmont 1:fdd22bb7aa52 61 * <b>Scaling and Overflow Behavior:</b>
emilmont 1:fdd22bb7aa52 62 *
emilmont 1:fdd22bb7aa52 63 * \par
emilmont 1:fdd22bb7aa52 64 * The difference between the function arm_mat_mult_q31() and this fast variant is that
emilmont 1:fdd22bb7aa52 65 * the fast variant use a 32-bit rather than a 64-bit accumulator.
emilmont 1:fdd22bb7aa52 66 * The result of each 1.31 x 1.31 multiplication is truncated to
emilmont 1:fdd22bb7aa52 67 * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30
emilmont 1:fdd22bb7aa52 68 * format. Finally, the accumulator is saturated and converted to a 1.31 result.
emilmont 1:fdd22bb7aa52 69 *
emilmont 1:fdd22bb7aa52 70 * \par
emilmont 1:fdd22bb7aa52 71 * The fast version has the same overflow behavior as the standard version but provides
emilmont 1:fdd22bb7aa52 72 * less precision since it discards the low 32 bits of each multiplication result.
emilmont 1:fdd22bb7aa52 73 * In order to avoid overflows completely the input signals must be scaled down.
emilmont 1:fdd22bb7aa52 74 * Scale down one of the input matrices by log2(numColsA) bits to
emilmont 1:fdd22bb7aa52 75 * avoid overflows, as a total of numColsA additions are computed internally for each
emilmont 1:fdd22bb7aa52 76 * output element.
emilmont 1:fdd22bb7aa52 77 *
emilmont 1:fdd22bb7aa52 78 * \par
emilmont 1:fdd22bb7aa52 79 * See <code>arm_mat_mult_q31()</code> for a slower implementation of this function
emilmont 1:fdd22bb7aa52 80 * which uses 64-bit accumulation to provide higher precision.
emilmont 1:fdd22bb7aa52 81 */
emilmont 1:fdd22bb7aa52 82
emilmont 1:fdd22bb7aa52 83 arm_status arm_mat_mult_fast_q31(
emilmont 1:fdd22bb7aa52 84 const arm_matrix_instance_q31 * pSrcA,
emilmont 1:fdd22bb7aa52 85 const arm_matrix_instance_q31 * pSrcB,
emilmont 1:fdd22bb7aa52 86 arm_matrix_instance_q31 * pDst)
emilmont 1:fdd22bb7aa52 87 {
emilmont 1:fdd22bb7aa52 88 q31_t *pIn1 = pSrcA->pData; /* input data matrix pointer A */
emilmont 1:fdd22bb7aa52 89 q31_t *pIn2 = pSrcB->pData; /* input data matrix pointer B */
emilmont 1:fdd22bb7aa52 90 q31_t *pInA = pSrcA->pData; /* input data matrix pointer A */
emilmont 1:fdd22bb7aa52 91 // q31_t *pSrcB = pSrcB->pData; /* input data matrix pointer B */
emilmont 1:fdd22bb7aa52 92 q31_t *pOut = pDst->pData; /* output data matrix pointer */
emilmont 1:fdd22bb7aa52 93 q31_t *px; /* Temporary output data matrix pointer */
emilmont 1:fdd22bb7aa52 94 q31_t sum; /* Accumulator */
emilmont 1:fdd22bb7aa52 95 uint16_t numRowsA = pSrcA->numRows; /* number of rows of input matrix A */
emilmont 1:fdd22bb7aa52 96 uint16_t numColsB = pSrcB->numCols; /* number of columns of input matrix B */
emilmont 1:fdd22bb7aa52 97 uint16_t numColsA = pSrcA->numCols; /* number of columns of input matrix A */
emilmont 1:fdd22bb7aa52 98 uint16_t col, i = 0u, j, row = numRowsA, colCnt; /* loop counters */
emilmont 1:fdd22bb7aa52 99 arm_status status; /* status of matrix multiplication */
emilmont 1:fdd22bb7aa52 100 q31_t inA1, inA2, inA3, inA4, inB1, inB2, inB3, inB4;
emilmont 1:fdd22bb7aa52 101
emilmont 1:fdd22bb7aa52 102 #ifdef ARM_MATH_MATRIX_CHECK
emilmont 1:fdd22bb7aa52 103
emilmont 1:fdd22bb7aa52 104
emilmont 1:fdd22bb7aa52 105 /* Check for matrix mismatch condition */
emilmont 1:fdd22bb7aa52 106 if((pSrcA->numCols != pSrcB->numRows) ||
emilmont 1:fdd22bb7aa52 107 (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols))
emilmont 1:fdd22bb7aa52 108 {
emilmont 1:fdd22bb7aa52 109 /* Set status as ARM_MATH_SIZE_MISMATCH */
emilmont 1:fdd22bb7aa52 110 status = ARM_MATH_SIZE_MISMATCH;
emilmont 1:fdd22bb7aa52 111 }
emilmont 1:fdd22bb7aa52 112 else
emilmont 1:fdd22bb7aa52 113 #endif /* #ifdef ARM_MATH_MATRIX_CHECK */
emilmont 1:fdd22bb7aa52 114
emilmont 1:fdd22bb7aa52 115 {
emilmont 1:fdd22bb7aa52 116 /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */
emilmont 1:fdd22bb7aa52 117 /* row loop */
emilmont 1:fdd22bb7aa52 118 do
emilmont 1:fdd22bb7aa52 119 {
emilmont 1:fdd22bb7aa52 120 /* Output pointer is set to starting address of the row being processed */
emilmont 1:fdd22bb7aa52 121 px = pOut + i;
emilmont 1:fdd22bb7aa52 122
emilmont 1:fdd22bb7aa52 123 /* For every row wise process, the column loop counter is to be initiated */
emilmont 1:fdd22bb7aa52 124 col = numColsB;
emilmont 1:fdd22bb7aa52 125
emilmont 1:fdd22bb7aa52 126 /* For every row wise process, the pIn2 pointer is set
emilmont 1:fdd22bb7aa52 127 ** to the starting address of the pSrcB data */
emilmont 1:fdd22bb7aa52 128 pIn2 = pSrcB->pData;
emilmont 1:fdd22bb7aa52 129
emilmont 1:fdd22bb7aa52 130 j = 0u;
emilmont 1:fdd22bb7aa52 131
emilmont 1:fdd22bb7aa52 132 /* column loop */
emilmont 1:fdd22bb7aa52 133 do
emilmont 1:fdd22bb7aa52 134 {
emilmont 1:fdd22bb7aa52 135 /* Set the variable sum, that acts as accumulator, to zero */
emilmont 1:fdd22bb7aa52 136 sum = 0;
emilmont 1:fdd22bb7aa52 137
emilmont 1:fdd22bb7aa52 138 /* Initiate the pointer pIn1 to point to the starting address of pInA */
emilmont 1:fdd22bb7aa52 139 pIn1 = pInA;
emilmont 1:fdd22bb7aa52 140
emilmont 1:fdd22bb7aa52 141 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 142 colCnt = numColsA >> 2;
emilmont 1:fdd22bb7aa52 143
emilmont 1:fdd22bb7aa52 144
emilmont 1:fdd22bb7aa52 145 /* matrix multiplication */
emilmont 1:fdd22bb7aa52 146 while(colCnt > 0u)
emilmont 1:fdd22bb7aa52 147 {
emilmont 1:fdd22bb7aa52 148 /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
emilmont 1:fdd22bb7aa52 149 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 150 inB1 = *pIn2;
emilmont 1:fdd22bb7aa52 151 pIn2 += numColsB;
emilmont 1:fdd22bb7aa52 152
emilmont 1:fdd22bb7aa52 153 inA1 = pIn1[0];
emilmont 1:fdd22bb7aa52 154 inA2 = pIn1[1];
emilmont 1:fdd22bb7aa52 155
emilmont 1:fdd22bb7aa52 156 inB2 = *pIn2;
emilmont 1:fdd22bb7aa52 157 pIn2 += numColsB;
emilmont 1:fdd22bb7aa52 158
emilmont 1:fdd22bb7aa52 159 inB3 = *pIn2;
emilmont 1:fdd22bb7aa52 160 pIn2 += numColsB;
emilmont 1:fdd22bb7aa52 161
emilmont 1:fdd22bb7aa52 162 sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA1 * inB1)) >> 32);
emilmont 1:fdd22bb7aa52 163 sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA2 * inB2)) >> 32);
emilmont 1:fdd22bb7aa52 164
emilmont 1:fdd22bb7aa52 165 inA3 = pIn1[2];
emilmont 1:fdd22bb7aa52 166 inA4 = pIn1[3];
emilmont 1:fdd22bb7aa52 167
emilmont 1:fdd22bb7aa52 168 inB4 = *pIn2;
emilmont 1:fdd22bb7aa52 169 pIn2 += numColsB;
emilmont 1:fdd22bb7aa52 170
emilmont 1:fdd22bb7aa52 171 sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA3 * inB3)) >> 32);
emilmont 1:fdd22bb7aa52 172 sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA4 * inB4)) >> 32);
emilmont 1:fdd22bb7aa52 173
emilmont 1:fdd22bb7aa52 174 pIn1 += 4u;
emilmont 1:fdd22bb7aa52 175
emilmont 1:fdd22bb7aa52 176 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 177 colCnt--;
emilmont 1:fdd22bb7aa52 178 }
emilmont 1:fdd22bb7aa52 179
emilmont 1:fdd22bb7aa52 180 /* If the columns of pSrcA is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 181 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 182 colCnt = numColsA % 0x4u;
emilmont 1:fdd22bb7aa52 183
emilmont 1:fdd22bb7aa52 184 while(colCnt > 0u)
emilmont 1:fdd22bb7aa52 185 {
emilmont 1:fdd22bb7aa52 186 /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */
emilmont 1:fdd22bb7aa52 187 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 188 sum = (q31_t) ((((q63_t) sum << 32) +
emilmont 1:fdd22bb7aa52 189 ((q63_t) * pIn1++ * (*pIn2))) >> 32);
emilmont 1:fdd22bb7aa52 190 pIn2 += numColsB;
emilmont 1:fdd22bb7aa52 191
emilmont 1:fdd22bb7aa52 192 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 193 colCnt--;
emilmont 1:fdd22bb7aa52 194 }
emilmont 1:fdd22bb7aa52 195
emilmont 1:fdd22bb7aa52 196 /* Convert the result from 2.30 to 1.31 format and store in destination buffer */
emilmont 1:fdd22bb7aa52 197 *px++ = sum << 1;
emilmont 1:fdd22bb7aa52 198
emilmont 1:fdd22bb7aa52 199 /* Update the pointer pIn2 to point to the starting address of the next column */
emilmont 1:fdd22bb7aa52 200 j++;
emilmont 1:fdd22bb7aa52 201 pIn2 = pSrcB->pData + j;
emilmont 1:fdd22bb7aa52 202
emilmont 1:fdd22bb7aa52 203 /* Decrement the column loop counter */
emilmont 1:fdd22bb7aa52 204 col--;
emilmont 1:fdd22bb7aa52 205
emilmont 1:fdd22bb7aa52 206 } while(col > 0u);
emilmont 1:fdd22bb7aa52 207
emilmont 1:fdd22bb7aa52 208 /* Update the pointer pInA to point to the starting address of the next row */
emilmont 1:fdd22bb7aa52 209 i = i + numColsB;
emilmont 1:fdd22bb7aa52 210 pInA = pInA + numColsA;
emilmont 1:fdd22bb7aa52 211
emilmont 1:fdd22bb7aa52 212 /* Decrement the row loop counter */
emilmont 1:fdd22bb7aa52 213 row--;
emilmont 1:fdd22bb7aa52 214
emilmont 1:fdd22bb7aa52 215 } while(row > 0u);
emilmont 1:fdd22bb7aa52 216
emilmont 1:fdd22bb7aa52 217 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 218 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 219 }
emilmont 1:fdd22bb7aa52 220 /* Return to application */
emilmont 1:fdd22bb7aa52 221 return (status);
emilmont 1:fdd22bb7aa52 222 }
emilmont 1:fdd22bb7aa52 223
emilmont 1:fdd22bb7aa52 224 /**
emilmont 1:fdd22bb7aa52 225 * @} end of MatrixMult group
emilmont 1:fdd22bb7aa52 226 */