The CMSIS DSP 5 library
Dependents: Nucleo-Heart-Rate ejercicioVrms2 PROYECTOFINAL ejercicioVrms ... more
functions/MatrixFunctions/arm_mat_mult_fast_q15.c@3:4098b9d3d571, 2018-06-21 (annotated)
- Committer:
- xorjoep
- Date:
- Thu Jun 21 11:56:27 2018 +0000
- Revision:
- 3:4098b9d3d571
- Parent:
- 1:24714b45cd1b
headers is a folder not a library
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
xorjoep | 1:24714b45cd1b | 1 | /* ---------------------------------------------------------------------- |
xorjoep | 1:24714b45cd1b | 2 | * Project: CMSIS DSP Library |
xorjoep | 1:24714b45cd1b | 3 | * Title: arm_mat_mult_fast_q15.c |
xorjoep | 1:24714b45cd1b | 4 | * Description: Q15 matrix multiplication (fast variant) |
xorjoep | 1:24714b45cd1b | 5 | * |
xorjoep | 1:24714b45cd1b | 6 | * $Date: 27. January 2017 |
xorjoep | 1:24714b45cd1b | 7 | * $Revision: V.1.5.1 |
xorjoep | 1:24714b45cd1b | 8 | * |
xorjoep | 1:24714b45cd1b | 9 | * Target Processor: Cortex-M cores |
xorjoep | 1:24714b45cd1b | 10 | * -------------------------------------------------------------------- */ |
xorjoep | 1:24714b45cd1b | 11 | /* |
xorjoep | 1:24714b45cd1b | 12 | * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved. |
xorjoep | 1:24714b45cd1b | 13 | * |
xorjoep | 1:24714b45cd1b | 14 | * SPDX-License-Identifier: Apache-2.0 |
xorjoep | 1:24714b45cd1b | 15 | * |
xorjoep | 1:24714b45cd1b | 16 | * Licensed under the Apache License, Version 2.0 (the License); you may |
xorjoep | 1:24714b45cd1b | 17 | * not use this file except in compliance with the License. |
xorjoep | 1:24714b45cd1b | 18 | * You may obtain a copy of the License at |
xorjoep | 1:24714b45cd1b | 19 | * |
xorjoep | 1:24714b45cd1b | 20 | * www.apache.org/licenses/LICENSE-2.0 |
xorjoep | 1:24714b45cd1b | 21 | * |
xorjoep | 1:24714b45cd1b | 22 | * Unless required by applicable law or agreed to in writing, software |
xorjoep | 1:24714b45cd1b | 23 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT |
xorjoep | 1:24714b45cd1b | 24 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
xorjoep | 1:24714b45cd1b | 25 | * See the License for the specific language governing permissions and |
xorjoep | 1:24714b45cd1b | 26 | * limitations under the License. |
xorjoep | 1:24714b45cd1b | 27 | */ |
xorjoep | 1:24714b45cd1b | 28 | |
xorjoep | 1:24714b45cd1b | 29 | #include "arm_math.h" |
xorjoep | 1:24714b45cd1b | 30 | |
xorjoep | 1:24714b45cd1b | 31 | /** |
xorjoep | 1:24714b45cd1b | 32 | * @ingroup groupMatrix |
xorjoep | 1:24714b45cd1b | 33 | */ |
xorjoep | 1:24714b45cd1b | 34 | |
xorjoep | 1:24714b45cd1b | 35 | /** |
xorjoep | 1:24714b45cd1b | 36 | * @addtogroup MatrixMult |
xorjoep | 1:24714b45cd1b | 37 | * @{ |
xorjoep | 1:24714b45cd1b | 38 | */ |
xorjoep | 1:24714b45cd1b | 39 | |
xorjoep | 1:24714b45cd1b | 40 | |
xorjoep | 1:24714b45cd1b | 41 | /** |
xorjoep | 1:24714b45cd1b | 42 | * @brief Q15 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4 |
xorjoep | 1:24714b45cd1b | 43 | * @param[in] *pSrcA points to the first input matrix structure |
xorjoep | 1:24714b45cd1b | 44 | * @param[in] *pSrcB points to the second input matrix structure |
xorjoep | 1:24714b45cd1b | 45 | * @param[out] *pDst points to output matrix structure |
xorjoep | 1:24714b45cd1b | 46 | * @param[in] *pState points to the array for storing intermediate results |
xorjoep | 1:24714b45cd1b | 47 | * @return The function returns either |
xorjoep | 1:24714b45cd1b | 48 | * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking. |
xorjoep | 1:24714b45cd1b | 49 | * |
xorjoep | 1:24714b45cd1b | 50 | * @details |
xorjoep | 1:24714b45cd1b | 51 | * <b>Scaling and Overflow Behavior:</b> |
xorjoep | 1:24714b45cd1b | 52 | * |
xorjoep | 1:24714b45cd1b | 53 | * \par |
xorjoep | 1:24714b45cd1b | 54 | * The difference between the function arm_mat_mult_q15() and this fast variant is that |
xorjoep | 1:24714b45cd1b | 55 | * the fast variant use a 32-bit rather than a 64-bit accumulator. |
xorjoep | 1:24714b45cd1b | 56 | * The result of each 1.15 x 1.15 multiplication is truncated to |
xorjoep | 1:24714b45cd1b | 57 | * 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30 |
xorjoep | 1:24714b45cd1b | 58 | * format. Finally, the accumulator is saturated and converted to a 1.15 result. |
xorjoep | 1:24714b45cd1b | 59 | * |
xorjoep | 1:24714b45cd1b | 60 | * \par |
xorjoep | 1:24714b45cd1b | 61 | * The fast version has the same overflow behavior as the standard version but provides |
xorjoep | 1:24714b45cd1b | 62 | * less precision since it discards the low 16 bits of each multiplication result. |
xorjoep | 1:24714b45cd1b | 63 | * In order to avoid overflows completely the input signals must be scaled down. |
xorjoep | 1:24714b45cd1b | 64 | * Scale down one of the input matrices by log2(numColsA) bits to |
xorjoep | 1:24714b45cd1b | 65 | * avoid overflows, as a total of numColsA additions are computed internally for each |
xorjoep | 1:24714b45cd1b | 66 | * output element. |
xorjoep | 1:24714b45cd1b | 67 | * |
xorjoep | 1:24714b45cd1b | 68 | * \par |
xorjoep | 1:24714b45cd1b | 69 | * See <code>arm_mat_mult_q15()</code> for a slower implementation of this function |
xorjoep | 1:24714b45cd1b | 70 | * which uses 64-bit accumulation to provide higher precision. |
xorjoep | 1:24714b45cd1b | 71 | */ |
xorjoep | 1:24714b45cd1b | 72 | |
xorjoep | 1:24714b45cd1b | 73 | arm_status arm_mat_mult_fast_q15( |
xorjoep | 1:24714b45cd1b | 74 | const arm_matrix_instance_q15 * pSrcA, |
xorjoep | 1:24714b45cd1b | 75 | const arm_matrix_instance_q15 * pSrcB, |
xorjoep | 1:24714b45cd1b | 76 | arm_matrix_instance_q15 * pDst, |
xorjoep | 1:24714b45cd1b | 77 | q15_t * pState) |
xorjoep | 1:24714b45cd1b | 78 | { |
xorjoep | 1:24714b45cd1b | 79 | q31_t sum; /* accumulator */ |
xorjoep | 1:24714b45cd1b | 80 | q15_t *pSrcBT = pState; /* input data matrix pointer for transpose */ |
xorjoep | 1:24714b45cd1b | 81 | q15_t *pInA = pSrcA->pData; /* input data matrix pointer A of Q15 type */ |
xorjoep | 1:24714b45cd1b | 82 | q15_t *pInB = pSrcB->pData; /* input data matrix pointer B of Q15 type */ |
xorjoep | 1:24714b45cd1b | 83 | q15_t *px; /* Temporary output data matrix pointer */ |
xorjoep | 1:24714b45cd1b | 84 | uint16_t numRowsA = pSrcA->numRows; /* number of rows of input matrix A */ |
xorjoep | 1:24714b45cd1b | 85 | uint16_t numColsB = pSrcB->numCols; /* number of columns of input matrix B */ |
xorjoep | 1:24714b45cd1b | 86 | uint16_t numColsA = pSrcA->numCols; /* number of columns of input matrix A */ |
xorjoep | 1:24714b45cd1b | 87 | uint16_t numRowsB = pSrcB->numRows; /* number of rows of input matrix A */ |
xorjoep | 1:24714b45cd1b | 88 | uint32_t col, i = 0U, row = numRowsB, colCnt; /* loop counters */ |
xorjoep | 1:24714b45cd1b | 89 | arm_status status; /* status of matrix multiplication */ |
xorjoep | 1:24714b45cd1b | 90 | |
xorjoep | 1:24714b45cd1b | 91 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 92 | |
xorjoep | 1:24714b45cd1b | 93 | q31_t in; /* Temporary variable to hold the input value */ |
xorjoep | 1:24714b45cd1b | 94 | q31_t inA1, inA2, inB1, inB2; |
xorjoep | 1:24714b45cd1b | 95 | q31_t sum2, sum3, sum4; |
xorjoep | 1:24714b45cd1b | 96 | q15_t *pInA2, *pInB2, *px2; |
xorjoep | 1:24714b45cd1b | 97 | uint32_t j = 0; |
xorjoep | 1:24714b45cd1b | 98 | |
xorjoep | 1:24714b45cd1b | 99 | #else |
xorjoep | 1:24714b45cd1b | 100 | |
xorjoep | 1:24714b45cd1b | 101 | q15_t in; /* Temporary variable to hold the input value */ |
xorjoep | 1:24714b45cd1b | 102 | q15_t inA1, inA2, inB1, inB2; |
xorjoep | 1:24714b45cd1b | 103 | |
xorjoep | 1:24714b45cd1b | 104 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
xorjoep | 1:24714b45cd1b | 105 | |
xorjoep | 1:24714b45cd1b | 106 | #ifdef ARM_MATH_MATRIX_CHECK |
xorjoep | 1:24714b45cd1b | 107 | /* Check for matrix mismatch condition */ |
xorjoep | 1:24714b45cd1b | 108 | if ((pSrcA->numCols != pSrcB->numRows) || |
xorjoep | 1:24714b45cd1b | 109 | (pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols)) |
xorjoep | 1:24714b45cd1b | 110 | { |
xorjoep | 1:24714b45cd1b | 111 | /* Set status as ARM_MATH_SIZE_MISMATCH */ |
xorjoep | 1:24714b45cd1b | 112 | status = ARM_MATH_SIZE_MISMATCH; |
xorjoep | 1:24714b45cd1b | 113 | } |
xorjoep | 1:24714b45cd1b | 114 | else |
xorjoep | 1:24714b45cd1b | 115 | #endif |
xorjoep | 1:24714b45cd1b | 116 | { |
xorjoep | 1:24714b45cd1b | 117 | /* Matrix transpose */ |
xorjoep | 1:24714b45cd1b | 118 | do |
xorjoep | 1:24714b45cd1b | 119 | { |
xorjoep | 1:24714b45cd1b | 120 | /* Apply loop unrolling and exchange the columns with row elements */ |
xorjoep | 1:24714b45cd1b | 121 | col = numColsB >> 2; |
xorjoep | 1:24714b45cd1b | 122 | |
xorjoep | 1:24714b45cd1b | 123 | /* The pointer px is set to starting address of the column being processed */ |
xorjoep | 1:24714b45cd1b | 124 | px = pSrcBT + i; |
xorjoep | 1:24714b45cd1b | 125 | |
xorjoep | 1:24714b45cd1b | 126 | /* First part of the processing with loop unrolling. Compute 4 outputs at a time. |
xorjoep | 1:24714b45cd1b | 127 | ** a second loop below computes the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 128 | while (col > 0U) |
xorjoep | 1:24714b45cd1b | 129 | { |
xorjoep | 1:24714b45cd1b | 130 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 131 | /* Read two elements from the row */ |
xorjoep | 1:24714b45cd1b | 132 | in = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 133 | |
xorjoep | 1:24714b45cd1b | 134 | /* Unpack and store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 135 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 136 | |
xorjoep | 1:24714b45cd1b | 137 | *px = (q15_t) in; |
xorjoep | 1:24714b45cd1b | 138 | |
xorjoep | 1:24714b45cd1b | 139 | #else |
xorjoep | 1:24714b45cd1b | 140 | |
xorjoep | 1:24714b45cd1b | 141 | *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
xorjoep | 1:24714b45cd1b | 142 | |
xorjoep | 1:24714b45cd1b | 143 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 144 | |
xorjoep | 1:24714b45cd1b | 145 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 146 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 147 | |
xorjoep | 1:24714b45cd1b | 148 | /* Unpack and store the second element in the destination */ |
xorjoep | 1:24714b45cd1b | 149 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 150 | |
xorjoep | 1:24714b45cd1b | 151 | *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
xorjoep | 1:24714b45cd1b | 152 | |
xorjoep | 1:24714b45cd1b | 153 | #else |
xorjoep | 1:24714b45cd1b | 154 | |
xorjoep | 1:24714b45cd1b | 155 | *px = (q15_t) in; |
xorjoep | 1:24714b45cd1b | 156 | |
xorjoep | 1:24714b45cd1b | 157 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 158 | |
xorjoep | 1:24714b45cd1b | 159 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 160 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 161 | |
xorjoep | 1:24714b45cd1b | 162 | /* Read two elements from the row */ |
xorjoep | 1:24714b45cd1b | 163 | in = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 164 | |
xorjoep | 1:24714b45cd1b | 165 | /* Unpack and store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 166 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 167 | |
xorjoep | 1:24714b45cd1b | 168 | *px = (q15_t) in; |
xorjoep | 1:24714b45cd1b | 169 | |
xorjoep | 1:24714b45cd1b | 170 | #else |
xorjoep | 1:24714b45cd1b | 171 | |
xorjoep | 1:24714b45cd1b | 172 | *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
xorjoep | 1:24714b45cd1b | 173 | |
xorjoep | 1:24714b45cd1b | 174 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 175 | |
xorjoep | 1:24714b45cd1b | 176 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 177 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 178 | |
xorjoep | 1:24714b45cd1b | 179 | /* Unpack and store the second element in the destination */ |
xorjoep | 1:24714b45cd1b | 180 | |
xorjoep | 1:24714b45cd1b | 181 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 182 | |
xorjoep | 1:24714b45cd1b | 183 | *px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
xorjoep | 1:24714b45cd1b | 184 | |
xorjoep | 1:24714b45cd1b | 185 | #else |
xorjoep | 1:24714b45cd1b | 186 | |
xorjoep | 1:24714b45cd1b | 187 | *px = (q15_t) in; |
xorjoep | 1:24714b45cd1b | 188 | |
xorjoep | 1:24714b45cd1b | 189 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 190 | |
xorjoep | 1:24714b45cd1b | 191 | #else |
xorjoep | 1:24714b45cd1b | 192 | |
xorjoep | 1:24714b45cd1b | 193 | /* Read one element from the row */ |
xorjoep | 1:24714b45cd1b | 194 | in = *pInB++; |
xorjoep | 1:24714b45cd1b | 195 | |
xorjoep | 1:24714b45cd1b | 196 | /* Store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 197 | *px = in; |
xorjoep | 1:24714b45cd1b | 198 | |
xorjoep | 1:24714b45cd1b | 199 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 200 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 201 | |
xorjoep | 1:24714b45cd1b | 202 | /* Read one element from the row */ |
xorjoep | 1:24714b45cd1b | 203 | in = *pInB++; |
xorjoep | 1:24714b45cd1b | 204 | |
xorjoep | 1:24714b45cd1b | 205 | /* Store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 206 | *px = in; |
xorjoep | 1:24714b45cd1b | 207 | |
xorjoep | 1:24714b45cd1b | 208 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 209 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 210 | |
xorjoep | 1:24714b45cd1b | 211 | /* Read one element from the row */ |
xorjoep | 1:24714b45cd1b | 212 | in = *pInB++; |
xorjoep | 1:24714b45cd1b | 213 | |
xorjoep | 1:24714b45cd1b | 214 | /* Store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 215 | *px = in; |
xorjoep | 1:24714b45cd1b | 216 | |
xorjoep | 1:24714b45cd1b | 217 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 218 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 219 | |
xorjoep | 1:24714b45cd1b | 220 | /* Read one element from the row */ |
xorjoep | 1:24714b45cd1b | 221 | in = *pInB++; |
xorjoep | 1:24714b45cd1b | 222 | |
xorjoep | 1:24714b45cd1b | 223 | /* Store one element in the destination */ |
xorjoep | 1:24714b45cd1b | 224 | *px = in; |
xorjoep | 1:24714b45cd1b | 225 | |
xorjoep | 1:24714b45cd1b | 226 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
xorjoep | 1:24714b45cd1b | 227 | |
xorjoep | 1:24714b45cd1b | 228 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 229 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 230 | |
xorjoep | 1:24714b45cd1b | 231 | /* Decrement the column loop counter */ |
xorjoep | 1:24714b45cd1b | 232 | col--; |
xorjoep | 1:24714b45cd1b | 233 | } |
xorjoep | 1:24714b45cd1b | 234 | |
xorjoep | 1:24714b45cd1b | 235 | /* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here. |
xorjoep | 1:24714b45cd1b | 236 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 237 | col = numColsB % 0x4U; |
xorjoep | 1:24714b45cd1b | 238 | |
xorjoep | 1:24714b45cd1b | 239 | while (col > 0U) |
xorjoep | 1:24714b45cd1b | 240 | { |
xorjoep | 1:24714b45cd1b | 241 | /* Read and store the input element in the destination */ |
xorjoep | 1:24714b45cd1b | 242 | *px = *pInB++; |
xorjoep | 1:24714b45cd1b | 243 | |
xorjoep | 1:24714b45cd1b | 244 | /* Update the pointer px to point to the next row of the transposed matrix */ |
xorjoep | 1:24714b45cd1b | 245 | px += numRowsB; |
xorjoep | 1:24714b45cd1b | 246 | |
xorjoep | 1:24714b45cd1b | 247 | /* Decrement the column loop counter */ |
xorjoep | 1:24714b45cd1b | 248 | col--; |
xorjoep | 1:24714b45cd1b | 249 | } |
xorjoep | 1:24714b45cd1b | 250 | |
xorjoep | 1:24714b45cd1b | 251 | i++; |
xorjoep | 1:24714b45cd1b | 252 | |
xorjoep | 1:24714b45cd1b | 253 | /* Decrement the row loop counter */ |
xorjoep | 1:24714b45cd1b | 254 | row--; |
xorjoep | 1:24714b45cd1b | 255 | |
xorjoep | 1:24714b45cd1b | 256 | } while (row > 0U); |
xorjoep | 1:24714b45cd1b | 257 | |
xorjoep | 1:24714b45cd1b | 258 | /* Reset the variables for the usage in the following multiplication process */ |
xorjoep | 1:24714b45cd1b | 259 | row = numRowsA; |
xorjoep | 1:24714b45cd1b | 260 | i = 0U; |
xorjoep | 1:24714b45cd1b | 261 | px = pDst->pData; |
xorjoep | 1:24714b45cd1b | 262 | |
xorjoep | 1:24714b45cd1b | 263 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 264 | /* Process two rows from matrix A at a time and output two rows at a time */ |
xorjoep | 1:24714b45cd1b | 265 | row = row >> 1; |
xorjoep | 1:24714b45cd1b | 266 | px2 = px + numColsB; |
xorjoep | 1:24714b45cd1b | 267 | #endif |
xorjoep | 1:24714b45cd1b | 268 | |
xorjoep | 1:24714b45cd1b | 269 | /* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */ |
xorjoep | 1:24714b45cd1b | 270 | /* row loop */ |
xorjoep | 1:24714b45cd1b | 271 | while (row > 0U) |
xorjoep | 1:24714b45cd1b | 272 | { |
xorjoep | 1:24714b45cd1b | 273 | /* For every row wise process, the column loop counter is to be initiated */ |
xorjoep | 1:24714b45cd1b | 274 | col = numColsB; |
xorjoep | 1:24714b45cd1b | 275 | |
xorjoep | 1:24714b45cd1b | 276 | /* For every row wise process, the pIn2 pointer is set |
xorjoep | 1:24714b45cd1b | 277 | ** to the starting address of the transposed pSrcB data */ |
xorjoep | 1:24714b45cd1b | 278 | pInB = pSrcBT; |
xorjoep | 1:24714b45cd1b | 279 | |
xorjoep | 1:24714b45cd1b | 280 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 281 | /* Process two (transposed) columns from matrix B at a time */ |
xorjoep | 1:24714b45cd1b | 282 | col = col >> 1; |
xorjoep | 1:24714b45cd1b | 283 | j = 0; |
xorjoep | 1:24714b45cd1b | 284 | #endif |
xorjoep | 1:24714b45cd1b | 285 | |
xorjoep | 1:24714b45cd1b | 286 | /* column loop */ |
xorjoep | 1:24714b45cd1b | 287 | while (col > 0U) |
xorjoep | 1:24714b45cd1b | 288 | { |
xorjoep | 1:24714b45cd1b | 289 | /* Set the variable sum, that acts as accumulator, to zero */ |
xorjoep | 1:24714b45cd1b | 290 | sum = 0; |
xorjoep | 1:24714b45cd1b | 291 | |
xorjoep | 1:24714b45cd1b | 292 | /* Initiate the pointer pInA to point to the starting address of the column being processed */ |
xorjoep | 1:24714b45cd1b | 293 | pInA = pSrcA->pData + i; |
xorjoep | 1:24714b45cd1b | 294 | |
xorjoep | 1:24714b45cd1b | 295 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 296 | sum2 = 0; |
xorjoep | 1:24714b45cd1b | 297 | sum3 = 0; |
xorjoep | 1:24714b45cd1b | 298 | sum4 = 0; |
xorjoep | 1:24714b45cd1b | 299 | pInB = pSrcBT + j; |
xorjoep | 1:24714b45cd1b | 300 | pInA2 = pInA + numColsA; |
xorjoep | 1:24714b45cd1b | 301 | pInB2 = pInB + numRowsB; |
xorjoep | 1:24714b45cd1b | 302 | |
xorjoep | 1:24714b45cd1b | 303 | /* Read in two elements at once - alows dual MAC instruction */ |
xorjoep | 1:24714b45cd1b | 304 | colCnt = numColsA >> 1; |
xorjoep | 1:24714b45cd1b | 305 | #else |
xorjoep | 1:24714b45cd1b | 306 | colCnt = numColsA >> 2; |
xorjoep | 1:24714b45cd1b | 307 | #endif |
xorjoep | 1:24714b45cd1b | 308 | |
xorjoep | 1:24714b45cd1b | 309 | /* matrix multiplication */ |
xorjoep | 1:24714b45cd1b | 310 | while (colCnt > 0U) |
xorjoep | 1:24714b45cd1b | 311 | { |
xorjoep | 1:24714b45cd1b | 312 | /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ |
xorjoep | 1:24714b45cd1b | 313 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 314 | |
xorjoep | 1:24714b45cd1b | 315 | inA1 = *__SIMD32(pInA)++; |
xorjoep | 1:24714b45cd1b | 316 | inB1 = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 317 | inA2 = *__SIMD32(pInA2)++; |
xorjoep | 1:24714b45cd1b | 318 | inB2 = *__SIMD32(pInB2)++; |
xorjoep | 1:24714b45cd1b | 319 | |
xorjoep | 1:24714b45cd1b | 320 | sum = __SMLAD(inA1, inB1, sum); |
xorjoep | 1:24714b45cd1b | 321 | sum2 = __SMLAD(inA1, inB2, sum2); |
xorjoep | 1:24714b45cd1b | 322 | sum3 = __SMLAD(inA2, inB1, sum3); |
xorjoep | 1:24714b45cd1b | 323 | sum4 = __SMLAD(inA2, inB2, sum4); |
xorjoep | 1:24714b45cd1b | 324 | |
xorjoep | 1:24714b45cd1b | 325 | #else |
xorjoep | 1:24714b45cd1b | 326 | |
xorjoep | 1:24714b45cd1b | 327 | inA1 = *pInA; |
xorjoep | 1:24714b45cd1b | 328 | inB1 = *pInB; |
xorjoep | 1:24714b45cd1b | 329 | sum += inA1 * inB1; |
xorjoep | 1:24714b45cd1b | 330 | |
xorjoep | 1:24714b45cd1b | 331 | inA2 = pInA[1]; |
xorjoep | 1:24714b45cd1b | 332 | inB2 = pInB[1]; |
xorjoep | 1:24714b45cd1b | 333 | sum += inA2 * inB2; |
xorjoep | 1:24714b45cd1b | 334 | |
xorjoep | 1:24714b45cd1b | 335 | inA1 = pInA[2]; |
xorjoep | 1:24714b45cd1b | 336 | inB1 = pInB[2]; |
xorjoep | 1:24714b45cd1b | 337 | sum += inA1 * inB1; |
xorjoep | 1:24714b45cd1b | 338 | |
xorjoep | 1:24714b45cd1b | 339 | inA2 = pInA[3]; |
xorjoep | 1:24714b45cd1b | 340 | inB2 = pInB[3]; |
xorjoep | 1:24714b45cd1b | 341 | sum += inA2 * inB2; |
xorjoep | 1:24714b45cd1b | 342 | |
xorjoep | 1:24714b45cd1b | 343 | pInA += 4; |
xorjoep | 1:24714b45cd1b | 344 | pInB += 4; |
xorjoep | 1:24714b45cd1b | 345 | |
xorjoep | 1:24714b45cd1b | 346 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
xorjoep | 1:24714b45cd1b | 347 | |
xorjoep | 1:24714b45cd1b | 348 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 349 | colCnt--; |
xorjoep | 1:24714b45cd1b | 350 | } |
xorjoep | 1:24714b45cd1b | 351 | |
xorjoep | 1:24714b45cd1b | 352 | /* process odd column samples */ |
xorjoep | 1:24714b45cd1b | 353 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 354 | if (numColsA & 1U) { |
xorjoep | 1:24714b45cd1b | 355 | inA1 = *pInA++; |
xorjoep | 1:24714b45cd1b | 356 | inB1 = *pInB++; |
xorjoep | 1:24714b45cd1b | 357 | inA2 = *pInA2++; |
xorjoep | 1:24714b45cd1b | 358 | inB2 = *pInB2++; |
xorjoep | 1:24714b45cd1b | 359 | sum += inA1 * inB1; |
xorjoep | 1:24714b45cd1b | 360 | sum2 += inA1 * inB2; |
xorjoep | 1:24714b45cd1b | 361 | sum3 += inA2 * inB1; |
xorjoep | 1:24714b45cd1b | 362 | sum4 += inA2 * inB2; |
xorjoep | 1:24714b45cd1b | 363 | } |
xorjoep | 1:24714b45cd1b | 364 | #else |
xorjoep | 1:24714b45cd1b | 365 | colCnt = numColsA % 0x4U; |
xorjoep | 1:24714b45cd1b | 366 | |
xorjoep | 1:24714b45cd1b | 367 | while (colCnt > 0U) |
xorjoep | 1:24714b45cd1b | 368 | { |
xorjoep | 1:24714b45cd1b | 369 | /* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ |
xorjoep | 1:24714b45cd1b | 370 | sum += (q31_t) (*pInA++) * (*pInB++); |
xorjoep | 1:24714b45cd1b | 371 | |
xorjoep | 1:24714b45cd1b | 372 | colCnt--; |
xorjoep | 1:24714b45cd1b | 373 | } |
xorjoep | 1:24714b45cd1b | 374 | #endif |
xorjoep | 1:24714b45cd1b | 375 | |
xorjoep | 1:24714b45cd1b | 376 | /* Saturate and store the result in the destination buffer */ |
xorjoep | 1:24714b45cd1b | 377 | *px++ = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 378 | |
xorjoep | 1:24714b45cd1b | 379 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 380 | *px++ = (q15_t) (sum2 >> 15); |
xorjoep | 1:24714b45cd1b | 381 | *px2++ = (q15_t) (sum3 >> 15); |
xorjoep | 1:24714b45cd1b | 382 | *px2++ = (q15_t) (sum4 >> 15); |
xorjoep | 1:24714b45cd1b | 383 | j += numRowsB * 2; |
xorjoep | 1:24714b45cd1b | 384 | #endif |
xorjoep | 1:24714b45cd1b | 385 | |
xorjoep | 1:24714b45cd1b | 386 | /* Decrement the column loop counter */ |
xorjoep | 1:24714b45cd1b | 387 | col--; |
xorjoep | 1:24714b45cd1b | 388 | |
xorjoep | 1:24714b45cd1b | 389 | } |
xorjoep | 1:24714b45cd1b | 390 | |
xorjoep | 1:24714b45cd1b | 391 | i = i + numColsA; |
xorjoep | 1:24714b45cd1b | 392 | |
xorjoep | 1:24714b45cd1b | 393 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 394 | i = i + numColsA; |
xorjoep | 1:24714b45cd1b | 395 | px = px2 + (numColsB & 1U); |
xorjoep | 1:24714b45cd1b | 396 | px2 = px + numColsB; |
xorjoep | 1:24714b45cd1b | 397 | #endif |
xorjoep | 1:24714b45cd1b | 398 | |
xorjoep | 1:24714b45cd1b | 399 | /* Decrement the row loop counter */ |
xorjoep | 1:24714b45cd1b | 400 | row--; |
xorjoep | 1:24714b45cd1b | 401 | |
xorjoep | 1:24714b45cd1b | 402 | } |
xorjoep | 1:24714b45cd1b | 403 | |
xorjoep | 1:24714b45cd1b | 404 | /* Compute any remaining odd row/column below */ |
xorjoep | 1:24714b45cd1b | 405 | |
xorjoep | 1:24714b45cd1b | 406 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 407 | |
xorjoep | 1:24714b45cd1b | 408 | /* Compute remaining output column */ |
xorjoep | 1:24714b45cd1b | 409 | if (numColsB & 1U) { |
xorjoep | 1:24714b45cd1b | 410 | |
xorjoep | 1:24714b45cd1b | 411 | /* Avoid redundant computation of last element */ |
xorjoep | 1:24714b45cd1b | 412 | row = numRowsA & (~0x1); |
xorjoep | 1:24714b45cd1b | 413 | |
xorjoep | 1:24714b45cd1b | 414 | /* Point to remaining unfilled column in output matrix */ |
xorjoep | 1:24714b45cd1b | 415 | px = pDst->pData+numColsB-1; |
xorjoep | 1:24714b45cd1b | 416 | pInA = pSrcA->pData; |
xorjoep | 1:24714b45cd1b | 417 | |
xorjoep | 1:24714b45cd1b | 418 | /* row loop */ |
xorjoep | 1:24714b45cd1b | 419 | while (row > 0) |
xorjoep | 1:24714b45cd1b | 420 | { |
xorjoep | 1:24714b45cd1b | 421 | |
xorjoep | 1:24714b45cd1b | 422 | /* point to last column in matrix B */ |
xorjoep | 1:24714b45cd1b | 423 | pInB = pSrcBT + numRowsB*(numColsB-1); |
xorjoep | 1:24714b45cd1b | 424 | |
xorjoep | 1:24714b45cd1b | 425 | /* Set the variable sum, that acts as accumulator, to zero */ |
xorjoep | 1:24714b45cd1b | 426 | sum = 0; |
xorjoep | 1:24714b45cd1b | 427 | |
xorjoep | 1:24714b45cd1b | 428 | /* Compute 4 columns at once */ |
xorjoep | 1:24714b45cd1b | 429 | colCnt = numColsA >> 2; |
xorjoep | 1:24714b45cd1b | 430 | |
xorjoep | 1:24714b45cd1b | 431 | /* matrix multiplication */ |
xorjoep | 1:24714b45cd1b | 432 | while (colCnt > 0U) |
xorjoep | 1:24714b45cd1b | 433 | { |
xorjoep | 1:24714b45cd1b | 434 | inA1 = *__SIMD32(pInA)++; |
xorjoep | 1:24714b45cd1b | 435 | inA2 = *__SIMD32(pInA)++; |
xorjoep | 1:24714b45cd1b | 436 | inB1 = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 437 | inB2 = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 438 | |
xorjoep | 1:24714b45cd1b | 439 | sum = __SMLAD(inA1, inB1, sum); |
xorjoep | 1:24714b45cd1b | 440 | sum = __SMLAD(inA2, inB2, sum); |
xorjoep | 1:24714b45cd1b | 441 | |
xorjoep | 1:24714b45cd1b | 442 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 443 | colCnt--; |
xorjoep | 1:24714b45cd1b | 444 | } |
xorjoep | 1:24714b45cd1b | 445 | |
xorjoep | 1:24714b45cd1b | 446 | colCnt = numColsA & 3U; |
xorjoep | 1:24714b45cd1b | 447 | while (colCnt > 0U) { |
xorjoep | 1:24714b45cd1b | 448 | sum += (q31_t) (*pInA++) * (*pInB++); |
xorjoep | 1:24714b45cd1b | 449 | colCnt--; |
xorjoep | 1:24714b45cd1b | 450 | } |
xorjoep | 1:24714b45cd1b | 451 | |
xorjoep | 1:24714b45cd1b | 452 | /* Store the result in the destination buffer */ |
xorjoep | 1:24714b45cd1b | 453 | *px = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 454 | px += numColsB; |
xorjoep | 1:24714b45cd1b | 455 | |
xorjoep | 1:24714b45cd1b | 456 | /* Decrement the row loop counter */ |
xorjoep | 1:24714b45cd1b | 457 | row--; |
xorjoep | 1:24714b45cd1b | 458 | } |
xorjoep | 1:24714b45cd1b | 459 | } |
xorjoep | 1:24714b45cd1b | 460 | |
xorjoep | 1:24714b45cd1b | 461 | /* Compute remaining output row */ |
xorjoep | 1:24714b45cd1b | 462 | if (numRowsA & 1U) { |
xorjoep | 1:24714b45cd1b | 463 | |
xorjoep | 1:24714b45cd1b | 464 | /* point to last row in output matrix */ |
xorjoep | 1:24714b45cd1b | 465 | px = pDst->pData+(numColsB)*(numRowsA-1); |
xorjoep | 1:24714b45cd1b | 466 | |
xorjoep | 1:24714b45cd1b | 467 | pInB = pSrcBT; |
xorjoep | 1:24714b45cd1b | 468 | col = numColsB; |
xorjoep | 1:24714b45cd1b | 469 | i = 0U; |
xorjoep | 1:24714b45cd1b | 470 | |
xorjoep | 1:24714b45cd1b | 471 | /* col loop */ |
xorjoep | 1:24714b45cd1b | 472 | while (col > 0) |
xorjoep | 1:24714b45cd1b | 473 | { |
xorjoep | 1:24714b45cd1b | 474 | |
xorjoep | 1:24714b45cd1b | 475 | /* point to last row in matrix A */ |
xorjoep | 1:24714b45cd1b | 476 | pInA = pSrcA->pData + (numRowsA-1)*numColsA; |
xorjoep | 1:24714b45cd1b | 477 | |
xorjoep | 1:24714b45cd1b | 478 | /* Set the variable sum, that acts as accumulator, to zero */ |
xorjoep | 1:24714b45cd1b | 479 | sum = 0; |
xorjoep | 1:24714b45cd1b | 480 | |
xorjoep | 1:24714b45cd1b | 481 | /* Compute 4 columns at once */ |
xorjoep | 1:24714b45cd1b | 482 | colCnt = numColsA >> 2; |
xorjoep | 1:24714b45cd1b | 483 | |
xorjoep | 1:24714b45cd1b | 484 | /* matrix multiplication */ |
xorjoep | 1:24714b45cd1b | 485 | while (colCnt > 0U) |
xorjoep | 1:24714b45cd1b | 486 | { |
xorjoep | 1:24714b45cd1b | 487 | inA1 = *__SIMD32(pInA)++; |
xorjoep | 1:24714b45cd1b | 488 | inA2 = *__SIMD32(pInA)++; |
xorjoep | 1:24714b45cd1b | 489 | inB1 = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 490 | inB2 = *__SIMD32(pInB)++; |
xorjoep | 1:24714b45cd1b | 491 | |
xorjoep | 1:24714b45cd1b | 492 | sum = __SMLAD(inA1, inB1, sum); |
xorjoep | 1:24714b45cd1b | 493 | sum = __SMLAD(inA2, inB2, sum); |
xorjoep | 1:24714b45cd1b | 494 | |
xorjoep | 1:24714b45cd1b | 495 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 496 | colCnt--; |
xorjoep | 1:24714b45cd1b | 497 | } |
xorjoep | 1:24714b45cd1b | 498 | |
xorjoep | 1:24714b45cd1b | 499 | colCnt = numColsA & 3U; |
xorjoep | 1:24714b45cd1b | 500 | while (colCnt > 0U) { |
xorjoep | 1:24714b45cd1b | 501 | sum += (q31_t) (*pInA++) * (*pInB++); |
xorjoep | 1:24714b45cd1b | 502 | colCnt--; |
xorjoep | 1:24714b45cd1b | 503 | } |
xorjoep | 1:24714b45cd1b | 504 | |
xorjoep | 1:24714b45cd1b | 505 | /* Store the result in the destination buffer */ |
xorjoep | 1:24714b45cd1b | 506 | *px++ = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 507 | |
xorjoep | 1:24714b45cd1b | 508 | /* Decrement the col loop counter */ |
xorjoep | 1:24714b45cd1b | 509 | col--; |
xorjoep | 1:24714b45cd1b | 510 | } |
xorjoep | 1:24714b45cd1b | 511 | } |
xorjoep | 1:24714b45cd1b | 512 | |
xorjoep | 1:24714b45cd1b | 513 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
xorjoep | 1:24714b45cd1b | 514 | |
xorjoep | 1:24714b45cd1b | 515 | /* set status as ARM_MATH_SUCCESS */ |
xorjoep | 1:24714b45cd1b | 516 | status = ARM_MATH_SUCCESS; |
xorjoep | 1:24714b45cd1b | 517 | } |
xorjoep | 1:24714b45cd1b | 518 | |
xorjoep | 1:24714b45cd1b | 519 | /* Return to application */ |
xorjoep | 1:24714b45cd1b | 520 | return (status); |
xorjoep | 1:24714b45cd1b | 521 | } |
xorjoep | 1:24714b45cd1b | 522 | |
xorjoep | 1:24714b45cd1b | 523 | /** |
xorjoep | 1:24714b45cd1b | 524 | * @} end of MatrixMult group |
xorjoep | 1:24714b45cd1b | 525 | */ |