| 提交 | 用户 | age | ||
| bfc108 | 1 | /* ---------------------------------------------------------------------- |
| Q | 2 | * Copyright (C) 2010-2014 ARM Limited. All rights reserved. |
| 3 | * | |
| 4 | * $Date: 19. March 2015 | |
| 5 | * $Revision: V.1.4.5 | |
| 6 | * | |
| 7 | * Project: CMSIS DSP Library | |
| 8 | * Title: arm_mat_mult_fast_q31.c | |
| 9 | * | |
| 10 | * Description: Q31 matrix multiplication (fast variant). | |
| 11 | * | |
| 12 | * Target Processor: Cortex-M4/Cortex-M3 | |
| 13 | * | |
| 14 | * Redistribution and use in source and binary forms, with or without | |
| 15 | * modification, are permitted provided that the following conditions | |
| 16 | * are met: | |
| 17 | * - Redistributions of source code must retain the above copyright | |
| 18 | * notice, this list of conditions and the following disclaimer. | |
| 19 | * - Redistributions in binary form must reproduce the above copyright | |
| 20 | * notice, this list of conditions and the following disclaimer in | |
| 21 | * the documentation and/or other materials provided with the | |
| 22 | * distribution. | |
| 23 | * - Neither the name of ARM LIMITED nor the names of its contributors | |
| 24 | * may be used to endorse or promote products derived from this | |
| 25 | * software without specific prior written permission. | |
| 26 | * | |
| 27 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
| 28 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
| 29 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS | |
| 30 | * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE | |
| 31 | * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, | |
| 32 | * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, | |
| 33 | * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; | |
| 34 | * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER | |
| 35 | * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT | |
| 36 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN | |
| 37 | * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE | |
| 38 | * POSSIBILITY OF SUCH DAMAGE. | |
| 39 | * -------------------------------------------------------------------- */ | |
| 40 | ||
| 41 | #include "arm_math.h" | |
| 42 | ||
| 43 | /** | |
| 44 | * @ingroup groupMatrix | |
| 45 | */ | |
| 46 | ||
| 47 | /** | |
| 48 | * @addtogroup MatrixMult | |
| 49 | * @{ | |
| 50 | */ | |
| 51 | ||
| 52 | /** | |
| 53 | * @brief Q31 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4 | |
| 54 | * @param[in] *pSrcA points to the first input matrix structure | |
| 55 | * @param[in] *pSrcB points to the second input matrix structure | |
| 56 | * @param[out] *pDst points to output matrix structure | |
| 57 | * @return The function returns either | |
| 58 | * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking. | |
| 59 | * | |
| 60 | * @details | |
| 61 | * <b>Scaling and Overflow Behavior:</b> | |
| 62 | * | |
| 63 | * \par | |
| 64 | * The difference between the function arm_mat_mult_q31() and this fast variant is that | |
| 65 | * the fast variant use a 32-bit rather than a 64-bit accumulator. | |
| 66 | * The result of each 1.31 x 1.31 multiplication is truncated to | |
| 67 | * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30 | |
| 68 | * format. Finally, the accumulator is saturated and converted to a 1.31 result. | |
| 69 | * | |
| 70 | * \par | |
| 71 | * The fast version has the same overflow behavior as the standard version but provides | |
| 72 | * less precision since it discards the low 32 bits of each multiplication result. | |
| 73 | * In order to avoid overflows completely the input signals must be scaled down. | |
| 74 | * Scale down one of the input matrices by log2(numColsA) bits to | |
| 75 | * avoid overflows, as a total of numColsA additions are computed internally for each | |
| 76 | * output element. | |
| 77 | * | |
| 78 | * \par | |
| 79 | * See <code>arm_mat_mult_q31()</code> for a slower implementation of this function | |
| 80 | * which uses 64-bit accumulation to provide higher precision. | |
| 81 | */ | |
| 82 | ||
| 83 | arm_status arm_mat_mult_fast_q31( | |
| 84 | const arm_matrix_instance_q31 * pSrcA, | |
| 85 | const arm_matrix_instance_q31 * pSrcB, | |
| 86 | arm_matrix_instance_q31 * pDst) | |
| 87 | { | |
| 88 | q31_t *pIn1 = pSrcA->pData; /* input data matrix pointer A */ | |
| 89 | q31_t *pIn2 = pSrcB->pData; /* input data matrix pointer B */ | |
| 90 | q31_t *pInA = pSrcA->pData; /* input data matrix pointer A */ | |
| 91 | // q31_t *pSrcB = pSrcB->pData; /* input data matrix pointer B */ | |
| 92 | q31_t *pOut = pDst->pData; /* output data matrix pointer */ | |
| 93 | q31_t *px; /* Temporary output data matrix pointer */ | |
| 94 | q31_t sum; /* Accumulator */ | |
| 95 | uint16_t numRowsA = pSrcA->numRows; /* number of rows of input matrix A */ | |
| 96 | uint16_t numColsB = pSrcB->numCols; /* number of columns of input matrix B */ | |
| 97 | uint16_t numColsA = pSrcA->numCols; /* number of columns of input matrix A */ | |
| 98 | uint16_t col, i = 0u, j, row = numRowsA, colCnt; /* loop counters */ | |
| 99 | arm_status status; /* status of matrix multiplication */ | |
| 100 | q31_t inA1, inA2, inA3, inA4, inB1, inB2, inB3, inB4; | |
| 101 | ||
| 102 | #ifdef ARM_MATH_MATRIX_CHECK | |
| 103 | ||
| 104 | ||
| 105 | /* Check for matrix mismatch condition */ | |
| 106 | if((pSrcA->numCols != pSrcB->numRows) || | |
| 107 | (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols)) | |
| 108 | { | |
| 109 | /* Set status as ARM_MATH_SIZE_MISMATCH */ | |
| 110 | status = ARM_MATH_SIZE_MISMATCH; | |
| 111 | } | |
| 112 | else | |
| 113 | #endif /* #ifdef ARM_MATH_MATRIX_CHECK */ | |
| 114 | ||
| 115 | { | |
| 116 | /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */ | |
| 117 | /* row loop */ | |
| 118 | do | |
| 119 | { | |
| 120 | /* Output pointer is set to starting address of the row being processed */ | |
| 121 | px = pOut + i; | |
| 122 | ||
| 123 | /* For every row wise process, the column loop counter is to be initiated */ | |
| 124 | col = numColsB; | |
| 125 | ||
| 126 | /* For every row wise process, the pIn2 pointer is set | |
| 127 | ** to the starting address of the pSrcB data */ | |
| 128 | pIn2 = pSrcB->pData; | |
| 129 | ||
| 130 | j = 0u; | |
| 131 | ||
| 132 | /* column loop */ | |
| 133 | do | |
| 134 | { | |
| 135 | /* Set the variable sum, that acts as accumulator, to zero */ | |
| 136 | sum = 0; | |
| 137 | ||
| 138 | /* Initiate the pointer pIn1 to point to the starting address of pInA */ | |
| 139 | pIn1 = pInA; | |
| 140 | ||
| 141 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ | |
| 142 | colCnt = numColsA >> 2; | |
| 143 | ||
| 144 | ||
| 145 | /* matrix multiplication */ | |
| 146 | while(colCnt > 0u) | |
| 147 | { | |
| 148 | /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ | |
| 149 | /* Perform the multiply-accumulates */ | |
| 150 | inB1 = *pIn2; | |
| 151 | pIn2 += numColsB; | |
| 152 | ||
| 153 | inA1 = pIn1[0]; | |
| 154 | inA2 = pIn1[1]; | |
| 155 | ||
| 156 | inB2 = *pIn2; | |
| 157 | pIn2 += numColsB; | |
| 158 | ||
| 159 | inB3 = *pIn2; | |
| 160 | pIn2 += numColsB; | |
| 161 | ||
| 162 | sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA1 * inB1)) >> 32); | |
| 163 | sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA2 * inB2)) >> 32); | |
| 164 | ||
| 165 | inA3 = pIn1[2]; | |
| 166 | inA4 = pIn1[3]; | |
| 167 | ||
| 168 | inB4 = *pIn2; | |
| 169 | pIn2 += numColsB; | |
| 170 | ||
| 171 | sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA3 * inB3)) >> 32); | |
| 172 | sum = (q31_t) ((((q63_t) sum << 32) + ((q63_t) inA4 * inB4)) >> 32); | |
| 173 | ||
| 174 | pIn1 += 4u; | |
| 175 | ||
| 176 | /* Decrement the loop counter */ | |
| 177 | colCnt--; | |
| 178 | } | |
| 179 | ||
| 180 | /* If the columns of pSrcA is not a multiple of 4, compute any remaining output samples here. | |
| 181 | ** No loop unrolling is used. */ | |
| 182 | colCnt = numColsA % 0x4u; | |
| 183 | ||
| 184 | while(colCnt > 0u) | |
| 185 | { | |
| 186 | /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ | |
| 187 | /* Perform the multiply-accumulates */ | |
| 188 | sum = (q31_t) ((((q63_t) sum << 32) + | |
| 189 | ((q63_t) * pIn1++ * (*pIn2))) >> 32); | |
| 190 | pIn2 += numColsB; | |
| 191 | ||
| 192 | /* Decrement the loop counter */ | |
| 193 | colCnt--; | |
| 194 | } | |
| 195 | ||
| 196 | /* Convert the result from 2.30 to 1.31 format and store in destination buffer */ | |
| 197 | *px++ = sum << 1; | |
| 198 | ||
| 199 | /* Update the pointer pIn2 to point to the starting address of the next column */ | |
| 200 | j++; | |
| 201 | pIn2 = pSrcB->pData + j; | |
| 202 | ||
| 203 | /* Decrement the column loop counter */ | |
| 204 | col--; | |
| 205 | ||
| 206 | } while(col > 0u); | |
| 207 | ||
| 208 | /* Update the pointer pInA to point to the starting address of the next row */ | |
| 209 | i = i + numColsB; | |
| 210 | pInA = pInA + numColsA; | |
| 211 | ||
| 212 | /* Decrement the row loop counter */ | |
| 213 | row--; | |
| 214 | ||
| 215 | } while(row > 0u); | |
| 216 | ||
| 217 | /* set status as ARM_MATH_SUCCESS */ | |
| 218 | status = ARM_MATH_SUCCESS; | |
| 219 | } | |
| 220 | /* Return to application */ | |
| 221 | return (status); | |
| 222 | } | |
| 223 | ||
| 224 | /** | |
| 225 | * @} end of MatrixMult group | |
| 226 | */ | |
