CMSIS DSP library
Dependents: KL25Z_FFT_Demo Hat_Board_v5_1 KL25Z_FFT_Demo_tony KL25Z_FFT_Demo_tony ... more
Fork of mbed-dsp by
cmsis_dsp/FilteringFunctions/arm_conv_q15.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?
User | Revision | Line number | New 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_conv_q15.c |
emilmont | 1:fdd22bb7aa52 | 9 | * |
emilmont | 2:da51fb522205 | 10 | * Description: Convolution of Q15 sequences. |
emilmont | 1:fdd22bb7aa52 | 11 | * |
emilmont | 1:fdd22bb7aa52 | 12 | * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0 |
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. |
emilmont | 1:fdd22bb7aa52 | 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 groupFilters |
emilmont | 1:fdd22bb7aa52 | 45 | */ |
emilmont | 1:fdd22bb7aa52 | 46 | |
emilmont | 1:fdd22bb7aa52 | 47 | /** |
emilmont | 1:fdd22bb7aa52 | 48 | * @addtogroup Conv |
emilmont | 1:fdd22bb7aa52 | 49 | * @{ |
emilmont | 1:fdd22bb7aa52 | 50 | */ |
emilmont | 1:fdd22bb7aa52 | 51 | |
emilmont | 1:fdd22bb7aa52 | 52 | /** |
emilmont | 1:fdd22bb7aa52 | 53 | * @brief Convolution of Q15 sequences. |
emilmont | 1:fdd22bb7aa52 | 54 | * @param[in] *pSrcA points to the first input sequence. |
emilmont | 1:fdd22bb7aa52 | 55 | * @param[in] srcALen length of the first input sequence. |
emilmont | 1:fdd22bb7aa52 | 56 | * @param[in] *pSrcB points to the second input sequence. |
emilmont | 1:fdd22bb7aa52 | 57 | * @param[in] srcBLen length of the second input sequence. |
emilmont | 1:fdd22bb7aa52 | 58 | * @param[out] *pDst points to the location where the output result is written. Length srcALen+srcBLen-1. |
emilmont | 1:fdd22bb7aa52 | 59 | * @return none. |
emilmont | 1:fdd22bb7aa52 | 60 | * |
emilmont | 1:fdd22bb7aa52 | 61 | * @details |
emilmont | 1:fdd22bb7aa52 | 62 | * <b>Scaling and Overflow Behavior:</b> |
emilmont | 1:fdd22bb7aa52 | 63 | * |
emilmont | 1:fdd22bb7aa52 | 64 | * \par |
emilmont | 1:fdd22bb7aa52 | 65 | * The function is implemented using a 64-bit internal accumulator. |
emilmont | 1:fdd22bb7aa52 | 66 | * Both inputs are in 1.15 format and multiplications yield a 2.30 result. |
emilmont | 1:fdd22bb7aa52 | 67 | * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format. |
emilmont | 1:fdd22bb7aa52 | 68 | * This approach provides 33 guard bits and there is no risk of overflow. |
emilmont | 1:fdd22bb7aa52 | 69 | * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format. |
emilmont | 1:fdd22bb7aa52 | 70 | * |
emilmont | 1:fdd22bb7aa52 | 71 | * \par |
emilmont | 1:fdd22bb7aa52 | 72 | * Refer to <code>arm_conv_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4. |
emilmont | 1:fdd22bb7aa52 | 73 | * |
emilmont | 1:fdd22bb7aa52 | 74 | * \par |
emilmont | 1:fdd22bb7aa52 | 75 | * Refer the function <code>arm_conv_opt_q15()</code> for a faster implementation of this function using scratch buffers. |
emilmont | 1:fdd22bb7aa52 | 76 | * |
emilmont | 1:fdd22bb7aa52 | 77 | */ |
emilmont | 1:fdd22bb7aa52 | 78 | |
emilmont | 1:fdd22bb7aa52 | 79 | void arm_conv_q15( |
emilmont | 1:fdd22bb7aa52 | 80 | q15_t * pSrcA, |
emilmont | 1:fdd22bb7aa52 | 81 | uint32_t srcALen, |
emilmont | 1:fdd22bb7aa52 | 82 | q15_t * pSrcB, |
emilmont | 1:fdd22bb7aa52 | 83 | uint32_t srcBLen, |
emilmont | 1:fdd22bb7aa52 | 84 | q15_t * pDst) |
emilmont | 1:fdd22bb7aa52 | 85 | { |
emilmont | 1:fdd22bb7aa52 | 86 | |
emilmont | 1:fdd22bb7aa52 | 87 | #if (defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE) |
emilmont | 1:fdd22bb7aa52 | 88 | |
emilmont | 1:fdd22bb7aa52 | 89 | /* Run the below code for Cortex-M4 and Cortex-M3 */ |
emilmont | 1:fdd22bb7aa52 | 90 | |
emilmont | 1:fdd22bb7aa52 | 91 | q15_t *pIn1; /* inputA pointer */ |
emilmont | 1:fdd22bb7aa52 | 92 | q15_t *pIn2; /* inputB pointer */ |
emilmont | 1:fdd22bb7aa52 | 93 | q15_t *pOut = pDst; /* output pointer */ |
emilmont | 1:fdd22bb7aa52 | 94 | q63_t sum, acc0, acc1, acc2, acc3; /* Accumulator */ |
emilmont | 1:fdd22bb7aa52 | 95 | q15_t *px; /* Intermediate inputA pointer */ |
emilmont | 1:fdd22bb7aa52 | 96 | q15_t *py; /* Intermediate inputB pointer */ |
emilmont | 1:fdd22bb7aa52 | 97 | q15_t *pSrc1, *pSrc2; /* Intermediate pointers */ |
emilmont | 1:fdd22bb7aa52 | 98 | q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */ |
emilmont | 1:fdd22bb7aa52 | 99 | uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */ |
emilmont | 1:fdd22bb7aa52 | 100 | |
emilmont | 1:fdd22bb7aa52 | 101 | /* The algorithm implementation is based on the lengths of the inputs. */ |
emilmont | 1:fdd22bb7aa52 | 102 | /* srcB is always made to slide across srcA. */ |
emilmont | 1:fdd22bb7aa52 | 103 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
emilmont | 1:fdd22bb7aa52 | 104 | if(srcALen >= srcBLen) |
emilmont | 1:fdd22bb7aa52 | 105 | { |
emilmont | 1:fdd22bb7aa52 | 106 | /* Initialization of inputA pointer */ |
emilmont | 1:fdd22bb7aa52 | 107 | pIn1 = pSrcA; |
emilmont | 1:fdd22bb7aa52 | 108 | |
emilmont | 1:fdd22bb7aa52 | 109 | /* Initialization of inputB pointer */ |
emilmont | 1:fdd22bb7aa52 | 110 | pIn2 = pSrcB; |
emilmont | 1:fdd22bb7aa52 | 111 | } |
emilmont | 1:fdd22bb7aa52 | 112 | else |
emilmont | 1:fdd22bb7aa52 | 113 | { |
emilmont | 1:fdd22bb7aa52 | 114 | /* Initialization of inputA pointer */ |
emilmont | 1:fdd22bb7aa52 | 115 | pIn1 = pSrcB; |
emilmont | 1:fdd22bb7aa52 | 116 | |
emilmont | 1:fdd22bb7aa52 | 117 | /* Initialization of inputB pointer */ |
emilmont | 1:fdd22bb7aa52 | 118 | pIn2 = pSrcA; |
emilmont | 1:fdd22bb7aa52 | 119 | |
emilmont | 1:fdd22bb7aa52 | 120 | /* srcBLen is always considered as shorter or equal to srcALen */ |
emilmont | 1:fdd22bb7aa52 | 121 | j = srcBLen; |
emilmont | 1:fdd22bb7aa52 | 122 | srcBLen = srcALen; |
emilmont | 1:fdd22bb7aa52 | 123 | srcALen = j; |
emilmont | 1:fdd22bb7aa52 | 124 | } |
emilmont | 1:fdd22bb7aa52 | 125 | |
emilmont | 1:fdd22bb7aa52 | 126 | /* conv(x,y) at n = x[n] * y[0] + x[n-1] * y[1] + x[n-2] * y[2] + ...+ x[n-N+1] * y[N -1] */ |
emilmont | 1:fdd22bb7aa52 | 127 | /* The function is internally |
emilmont | 1:fdd22bb7aa52 | 128 | * divided into three stages according to the number of multiplications that has to be |
emilmont | 1:fdd22bb7aa52 | 129 | * taken place between inputA samples and inputB samples. In the first stage of the |
emilmont | 1:fdd22bb7aa52 | 130 | * algorithm, the multiplications increase by one for every iteration. |
emilmont | 1:fdd22bb7aa52 | 131 | * In the second stage of the algorithm, srcBLen number of multiplications are done. |
emilmont | 1:fdd22bb7aa52 | 132 | * In the third stage of the algorithm, the multiplications decrease by one |
emilmont | 1:fdd22bb7aa52 | 133 | * for every iteration. */ |
emilmont | 1:fdd22bb7aa52 | 134 | |
emilmont | 1:fdd22bb7aa52 | 135 | /* The algorithm is implemented in three stages. |
emilmont | 1:fdd22bb7aa52 | 136 | The loop counters of each stage is initiated here. */ |
emilmont | 1:fdd22bb7aa52 | 137 | blockSize1 = srcBLen - 1u; |
emilmont | 1:fdd22bb7aa52 | 138 | blockSize2 = srcALen - (srcBLen - 1u); |
emilmont | 1:fdd22bb7aa52 | 139 | |
emilmont | 1:fdd22bb7aa52 | 140 | /* -------------------------- |
emilmont | 1:fdd22bb7aa52 | 141 | * Initializations of stage1 |
emilmont | 1:fdd22bb7aa52 | 142 | * -------------------------*/ |
emilmont | 1:fdd22bb7aa52 | 143 | |
emilmont | 1:fdd22bb7aa52 | 144 | /* sum = x[0] * y[0] |
emilmont | 1:fdd22bb7aa52 | 145 | * sum = x[0] * y[1] + x[1] * y[0] |
emilmont | 1:fdd22bb7aa52 | 146 | * .... |
emilmont | 1:fdd22bb7aa52 | 147 | * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0] |
emilmont | 1:fdd22bb7aa52 | 148 | */ |
emilmont | 1:fdd22bb7aa52 | 149 | |
emilmont | 1:fdd22bb7aa52 | 150 | /* In this stage the MAC operations are increased by 1 for every iteration. |
emilmont | 1:fdd22bb7aa52 | 151 | The count variable holds the number of MAC operations performed */ |
emilmont | 1:fdd22bb7aa52 | 152 | count = 1u; |
emilmont | 1:fdd22bb7aa52 | 153 | |
emilmont | 1:fdd22bb7aa52 | 154 | /* Working pointer of inputA */ |
emilmont | 1:fdd22bb7aa52 | 155 | px = pIn1; |
emilmont | 1:fdd22bb7aa52 | 156 | |
emilmont | 1:fdd22bb7aa52 | 157 | /* Working pointer of inputB */ |
emilmont | 1:fdd22bb7aa52 | 158 | py = pIn2; |
emilmont | 1:fdd22bb7aa52 | 159 | |
emilmont | 1:fdd22bb7aa52 | 160 | |
emilmont | 1:fdd22bb7aa52 | 161 | /* ------------------------ |
emilmont | 1:fdd22bb7aa52 | 162 | * Stage1 process |
emilmont | 1:fdd22bb7aa52 | 163 | * ----------------------*/ |
emilmont | 1:fdd22bb7aa52 | 164 | |
emilmont | 1:fdd22bb7aa52 | 165 | /* For loop unrolling by 4, this stage is divided into two. */ |
emilmont | 1:fdd22bb7aa52 | 166 | /* First part of this stage computes the MAC operations less than 4 */ |
emilmont | 1:fdd22bb7aa52 | 167 | /* Second part of this stage computes the MAC operations greater than or equal to 4 */ |
emilmont | 1:fdd22bb7aa52 | 168 | |
emilmont | 1:fdd22bb7aa52 | 169 | /* The first part of the stage starts here */ |
emilmont | 1:fdd22bb7aa52 | 170 | while((count < 4u) && (blockSize1 > 0u)) |
emilmont | 1:fdd22bb7aa52 | 171 | { |
emilmont | 1:fdd22bb7aa52 | 172 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 173 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 174 | |
emilmont | 1:fdd22bb7aa52 | 175 | /* Loop over number of MAC operations between |
emilmont | 1:fdd22bb7aa52 | 176 | * inputA samples and inputB samples */ |
emilmont | 1:fdd22bb7aa52 | 177 | k = count; |
emilmont | 1:fdd22bb7aa52 | 178 | |
emilmont | 1:fdd22bb7aa52 | 179 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 180 | { |
emilmont | 1:fdd22bb7aa52 | 181 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 182 | sum = __SMLALD(*px++, *py--, sum); |
emilmont | 1:fdd22bb7aa52 | 183 | |
emilmont | 1:fdd22bb7aa52 | 184 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 185 | k--; |
emilmont | 1:fdd22bb7aa52 | 186 | } |
emilmont | 1:fdd22bb7aa52 | 187 | |
emilmont | 1:fdd22bb7aa52 | 188 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 189 | *pOut++ = (q15_t) (__SSAT((sum >> 15), 16)); |
emilmont | 1:fdd22bb7aa52 | 190 | |
emilmont | 1:fdd22bb7aa52 | 191 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 192 | py = pIn2 + count; |
emilmont | 1:fdd22bb7aa52 | 193 | px = pIn1; |
emilmont | 1:fdd22bb7aa52 | 194 | |
emilmont | 1:fdd22bb7aa52 | 195 | /* Increment the MAC count */ |
emilmont | 1:fdd22bb7aa52 | 196 | count++; |
emilmont | 1:fdd22bb7aa52 | 197 | |
emilmont | 1:fdd22bb7aa52 | 198 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 199 | blockSize1--; |
emilmont | 1:fdd22bb7aa52 | 200 | } |
emilmont | 1:fdd22bb7aa52 | 201 | |
emilmont | 1:fdd22bb7aa52 | 202 | /* The second part of the stage starts here */ |
emilmont | 1:fdd22bb7aa52 | 203 | /* The internal loop, over count, is unrolled by 4 */ |
emilmont | 1:fdd22bb7aa52 | 204 | /* To, read the last two inputB samples using SIMD: |
emilmont | 1:fdd22bb7aa52 | 205 | * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */ |
emilmont | 1:fdd22bb7aa52 | 206 | py = py - 1; |
emilmont | 1:fdd22bb7aa52 | 207 | |
emilmont | 1:fdd22bb7aa52 | 208 | while(blockSize1 > 0u) |
emilmont | 1:fdd22bb7aa52 | 209 | { |
emilmont | 1:fdd22bb7aa52 | 210 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 211 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 212 | |
emilmont | 1:fdd22bb7aa52 | 213 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emilmont | 1:fdd22bb7aa52 | 214 | k = count >> 2u; |
emilmont | 1:fdd22bb7aa52 | 215 | |
emilmont | 1:fdd22bb7aa52 | 216 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emilmont | 1:fdd22bb7aa52 | 217 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emilmont | 1:fdd22bb7aa52 | 218 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 219 | { |
emilmont | 1:fdd22bb7aa52 | 220 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 221 | /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */ |
emilmont | 1:fdd22bb7aa52 | 222 | sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum); |
emilmont | 1:fdd22bb7aa52 | 223 | /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */ |
emilmont | 1:fdd22bb7aa52 | 224 | sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum); |
emilmont | 1:fdd22bb7aa52 | 225 | |
emilmont | 1:fdd22bb7aa52 | 226 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 227 | k--; |
emilmont | 1:fdd22bb7aa52 | 228 | } |
emilmont | 1:fdd22bb7aa52 | 229 | |
emilmont | 1:fdd22bb7aa52 | 230 | /* For the next MAC operations, the pointer py is used without SIMD |
emilmont | 1:fdd22bb7aa52 | 231 | * So, py is incremented by 1 */ |
emilmont | 1:fdd22bb7aa52 | 232 | py = py + 1u; |
emilmont | 1:fdd22bb7aa52 | 233 | |
emilmont | 1:fdd22bb7aa52 | 234 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
emilmont | 1:fdd22bb7aa52 | 235 | ** No loop unrolling is used. */ |
emilmont | 1:fdd22bb7aa52 | 236 | k = count % 0x4u; |
emilmont | 1:fdd22bb7aa52 | 237 | |
emilmont | 1:fdd22bb7aa52 | 238 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 239 | { |
emilmont | 1:fdd22bb7aa52 | 240 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 241 | sum = __SMLALD(*px++, *py--, sum); |
emilmont | 1:fdd22bb7aa52 | 242 | |
emilmont | 1:fdd22bb7aa52 | 243 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 244 | k--; |
emilmont | 1:fdd22bb7aa52 | 245 | } |
emilmont | 1:fdd22bb7aa52 | 246 | |
emilmont | 1:fdd22bb7aa52 | 247 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 248 | *pOut++ = (q15_t) (__SSAT((sum >> 15), 16)); |
emilmont | 1:fdd22bb7aa52 | 249 | |
emilmont | 1:fdd22bb7aa52 | 250 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 251 | py = pIn2 + (count - 1u); |
emilmont | 1:fdd22bb7aa52 | 252 | px = pIn1; |
emilmont | 1:fdd22bb7aa52 | 253 | |
emilmont | 1:fdd22bb7aa52 | 254 | /* Increment the MAC count */ |
emilmont | 1:fdd22bb7aa52 | 255 | count++; |
emilmont | 1:fdd22bb7aa52 | 256 | |
emilmont | 1:fdd22bb7aa52 | 257 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 258 | blockSize1--; |
emilmont | 1:fdd22bb7aa52 | 259 | } |
emilmont | 1:fdd22bb7aa52 | 260 | |
emilmont | 1:fdd22bb7aa52 | 261 | /* -------------------------- |
emilmont | 1:fdd22bb7aa52 | 262 | * Initializations of stage2 |
emilmont | 1:fdd22bb7aa52 | 263 | * ------------------------*/ |
emilmont | 1:fdd22bb7aa52 | 264 | |
emilmont | 1:fdd22bb7aa52 | 265 | /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0] |
emilmont | 1:fdd22bb7aa52 | 266 | * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0] |
emilmont | 1:fdd22bb7aa52 | 267 | * .... |
emilmont | 1:fdd22bb7aa52 | 268 | * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0] |
emilmont | 1:fdd22bb7aa52 | 269 | */ |
emilmont | 1:fdd22bb7aa52 | 270 | |
emilmont | 1:fdd22bb7aa52 | 271 | /* Working pointer of inputA */ |
emilmont | 1:fdd22bb7aa52 | 272 | px = pIn1; |
emilmont | 1:fdd22bb7aa52 | 273 | |
emilmont | 1:fdd22bb7aa52 | 274 | /* Working pointer of inputB */ |
emilmont | 1:fdd22bb7aa52 | 275 | pSrc2 = pIn2 + (srcBLen - 1u); |
emilmont | 1:fdd22bb7aa52 | 276 | py = pSrc2; |
emilmont | 1:fdd22bb7aa52 | 277 | |
emilmont | 1:fdd22bb7aa52 | 278 | /* count is the index by which the pointer pIn1 to be incremented */ |
emilmont | 1:fdd22bb7aa52 | 279 | count = 0u; |
emilmont | 1:fdd22bb7aa52 | 280 | |
emilmont | 1:fdd22bb7aa52 | 281 | |
emilmont | 1:fdd22bb7aa52 | 282 | /* -------------------- |
emilmont | 1:fdd22bb7aa52 | 283 | * Stage2 process |
emilmont | 1:fdd22bb7aa52 | 284 | * -------------------*/ |
emilmont | 1:fdd22bb7aa52 | 285 | |
emilmont | 1:fdd22bb7aa52 | 286 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
emilmont | 1:fdd22bb7aa52 | 287 | * So, to loop unroll over blockSize2, |
emilmont | 1:fdd22bb7aa52 | 288 | * srcBLen should be greater than or equal to 4 */ |
emilmont | 1:fdd22bb7aa52 | 289 | if(srcBLen >= 4u) |
emilmont | 1:fdd22bb7aa52 | 290 | { |
emilmont | 1:fdd22bb7aa52 | 291 | /* Loop unroll over blockSize2, by 4 */ |
emilmont | 1:fdd22bb7aa52 | 292 | blkCnt = blockSize2 >> 2u; |
emilmont | 1:fdd22bb7aa52 | 293 | |
emilmont | 1:fdd22bb7aa52 | 294 | while(blkCnt > 0u) |
emilmont | 1:fdd22bb7aa52 | 295 | { |
emilmont | 1:fdd22bb7aa52 | 296 | py = py - 1u; |
emilmont | 1:fdd22bb7aa52 | 297 | |
emilmont | 1:fdd22bb7aa52 | 298 | /* Set all accumulators to zero */ |
emilmont | 1:fdd22bb7aa52 | 299 | acc0 = 0; |
emilmont | 1:fdd22bb7aa52 | 300 | acc1 = 0; |
emilmont | 1:fdd22bb7aa52 | 301 | acc2 = 0; |
emilmont | 1:fdd22bb7aa52 | 302 | acc3 = 0; |
emilmont | 1:fdd22bb7aa52 | 303 | |
emilmont | 1:fdd22bb7aa52 | 304 | |
emilmont | 1:fdd22bb7aa52 | 305 | /* read x[0], x[1] samples */ |
emilmont | 1:fdd22bb7aa52 | 306 | x0 = *__SIMD32(px); |
emilmont | 1:fdd22bb7aa52 | 307 | /* read x[1], x[2] samples */ |
emilmont | 1:fdd22bb7aa52 | 308 | x1 = _SIMD32_OFFSET(px+1); |
emilmont | 2:da51fb522205 | 309 | px+= 2u; |
emilmont | 1:fdd22bb7aa52 | 310 | |
emilmont | 1:fdd22bb7aa52 | 311 | |
emilmont | 1:fdd22bb7aa52 | 312 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emilmont | 1:fdd22bb7aa52 | 313 | k = srcBLen >> 2u; |
emilmont | 1:fdd22bb7aa52 | 314 | |
emilmont | 1:fdd22bb7aa52 | 315 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emilmont | 1:fdd22bb7aa52 | 316 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emilmont | 1:fdd22bb7aa52 | 317 | do |
emilmont | 1:fdd22bb7aa52 | 318 | { |
emilmont | 1:fdd22bb7aa52 | 319 | /* Read the last two inputB samples using SIMD: |
emilmont | 1:fdd22bb7aa52 | 320 | * y[srcBLen - 1] and y[srcBLen - 2] */ |
emilmont | 1:fdd22bb7aa52 | 321 | c0 = *__SIMD32(py)--; |
emilmont | 1:fdd22bb7aa52 | 322 | |
emilmont | 1:fdd22bb7aa52 | 323 | /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */ |
emilmont | 1:fdd22bb7aa52 | 324 | acc0 = __SMLALDX(x0, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 325 | |
emilmont | 1:fdd22bb7aa52 | 326 | /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */ |
emilmont | 1:fdd22bb7aa52 | 327 | acc1 = __SMLALDX(x1, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 328 | |
emilmont | 1:fdd22bb7aa52 | 329 | /* Read x[2], x[3] */ |
emilmont | 1:fdd22bb7aa52 | 330 | x2 = *__SIMD32(px); |
emilmont | 1:fdd22bb7aa52 | 331 | |
emilmont | 1:fdd22bb7aa52 | 332 | /* Read x[3], x[4] */ |
emilmont | 1:fdd22bb7aa52 | 333 | x3 = _SIMD32_OFFSET(px+1); |
emilmont | 1:fdd22bb7aa52 | 334 | |
emilmont | 1:fdd22bb7aa52 | 335 | /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */ |
emilmont | 1:fdd22bb7aa52 | 336 | acc2 = __SMLALDX(x2, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 337 | |
emilmont | 1:fdd22bb7aa52 | 338 | /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */ |
emilmont | 1:fdd22bb7aa52 | 339 | acc3 = __SMLALDX(x3, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 340 | |
emilmont | 1:fdd22bb7aa52 | 341 | /* Read y[srcBLen - 3] and y[srcBLen - 4] */ |
emilmont | 1:fdd22bb7aa52 | 342 | c0 = *__SIMD32(py)--; |
emilmont | 1:fdd22bb7aa52 | 343 | |
emilmont | 1:fdd22bb7aa52 | 344 | /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */ |
emilmont | 1:fdd22bb7aa52 | 345 | acc0 = __SMLALDX(x2, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 346 | |
emilmont | 1:fdd22bb7aa52 | 347 | /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */ |
emilmont | 1:fdd22bb7aa52 | 348 | acc1 = __SMLALDX(x3, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 349 | |
emilmont | 1:fdd22bb7aa52 | 350 | /* Read x[4], x[5] */ |
emilmont | 1:fdd22bb7aa52 | 351 | x0 = _SIMD32_OFFSET(px+2); |
emilmont | 1:fdd22bb7aa52 | 352 | |
emilmont | 1:fdd22bb7aa52 | 353 | /* Read x[5], x[6] */ |
emilmont | 1:fdd22bb7aa52 | 354 | x1 = _SIMD32_OFFSET(px+3); |
emilmont | 2:da51fb522205 | 355 | px += 4u; |
emilmont | 1:fdd22bb7aa52 | 356 | |
emilmont | 1:fdd22bb7aa52 | 357 | /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */ |
emilmont | 1:fdd22bb7aa52 | 358 | acc2 = __SMLALDX(x0, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 359 | |
emilmont | 1:fdd22bb7aa52 | 360 | /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */ |
emilmont | 1:fdd22bb7aa52 | 361 | acc3 = __SMLALDX(x1, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 362 | |
emilmont | 1:fdd22bb7aa52 | 363 | } while(--k); |
emilmont | 1:fdd22bb7aa52 | 364 | |
emilmont | 1:fdd22bb7aa52 | 365 | /* For the next MAC operations, SIMD is not used |
emilmont | 1:fdd22bb7aa52 | 366 | * So, the 16 bit pointer if inputB, py is updated */ |
emilmont | 1:fdd22bb7aa52 | 367 | |
emilmont | 1:fdd22bb7aa52 | 368 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emilmont | 1:fdd22bb7aa52 | 369 | ** No loop unrolling is used. */ |
emilmont | 1:fdd22bb7aa52 | 370 | k = srcBLen % 0x4u; |
emilmont | 1:fdd22bb7aa52 | 371 | |
emilmont | 1:fdd22bb7aa52 | 372 | if(k == 1u) |
emilmont | 1:fdd22bb7aa52 | 373 | { |
emilmont | 1:fdd22bb7aa52 | 374 | /* Read y[srcBLen - 5] */ |
emilmont | 1:fdd22bb7aa52 | 375 | c0 = *(py+1); |
emilmont | 1:fdd22bb7aa52 | 376 | |
emilmont | 1:fdd22bb7aa52 | 377 | #ifdef ARM_MATH_BIG_ENDIAN |
emilmont | 1:fdd22bb7aa52 | 378 | |
emilmont | 1:fdd22bb7aa52 | 379 | c0 = c0 << 16u; |
emilmont | 1:fdd22bb7aa52 | 380 | |
emilmont | 1:fdd22bb7aa52 | 381 | #else |
emilmont | 1:fdd22bb7aa52 | 382 | |
emilmont | 1:fdd22bb7aa52 | 383 | c0 = c0 & 0x0000FFFF; |
emilmont | 1:fdd22bb7aa52 | 384 | |
emilmont | 1:fdd22bb7aa52 | 385 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emilmont | 1:fdd22bb7aa52 | 386 | /* Read x[7] */ |
emilmont | 1:fdd22bb7aa52 | 387 | x3 = *__SIMD32(px); |
emilmont | 2:da51fb522205 | 388 | px++; |
emilmont | 1:fdd22bb7aa52 | 389 | |
emilmont | 1:fdd22bb7aa52 | 390 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 391 | acc0 = __SMLALD(x0, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 392 | acc1 = __SMLALD(x1, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 393 | acc2 = __SMLALDX(x1, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 394 | acc3 = __SMLALDX(x3, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 395 | } |
emilmont | 1:fdd22bb7aa52 | 396 | |
emilmont | 1:fdd22bb7aa52 | 397 | if(k == 2u) |
emilmont | 1:fdd22bb7aa52 | 398 | { |
emilmont | 1:fdd22bb7aa52 | 399 | /* Read y[srcBLen - 5], y[srcBLen - 6] */ |
emilmont | 1:fdd22bb7aa52 | 400 | c0 = _SIMD32_OFFSET(py); |
emilmont | 1:fdd22bb7aa52 | 401 | |
emilmont | 1:fdd22bb7aa52 | 402 | /* Read x[7], x[8] */ |
emilmont | 1:fdd22bb7aa52 | 403 | x3 = *__SIMD32(px); |
emilmont | 1:fdd22bb7aa52 | 404 | |
emilmont | 1:fdd22bb7aa52 | 405 | /* Read x[9] */ |
emilmont | 1:fdd22bb7aa52 | 406 | x2 = _SIMD32_OFFSET(px+1); |
emilmont | 2:da51fb522205 | 407 | px += 2u; |
emilmont | 1:fdd22bb7aa52 | 408 | |
emilmont | 1:fdd22bb7aa52 | 409 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 410 | acc0 = __SMLALDX(x0, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 411 | acc1 = __SMLALDX(x1, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 412 | acc2 = __SMLALDX(x3, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 413 | acc3 = __SMLALDX(x2, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 414 | } |
emilmont | 1:fdd22bb7aa52 | 415 | |
emilmont | 1:fdd22bb7aa52 | 416 | if(k == 3u) |
emilmont | 1:fdd22bb7aa52 | 417 | { |
emilmont | 1:fdd22bb7aa52 | 418 | /* Read y[srcBLen - 5], y[srcBLen - 6] */ |
emilmont | 1:fdd22bb7aa52 | 419 | c0 = _SIMD32_OFFSET(py); |
emilmont | 1:fdd22bb7aa52 | 420 | |
emilmont | 1:fdd22bb7aa52 | 421 | /* Read x[7], x[8] */ |
emilmont | 1:fdd22bb7aa52 | 422 | x3 = *__SIMD32(px); |
emilmont | 1:fdd22bb7aa52 | 423 | |
emilmont | 1:fdd22bb7aa52 | 424 | /* Read x[9] */ |
emilmont | 1:fdd22bb7aa52 | 425 | x2 = _SIMD32_OFFSET(px+1); |
emilmont | 1:fdd22bb7aa52 | 426 | |
emilmont | 1:fdd22bb7aa52 | 427 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 428 | acc0 = __SMLALDX(x0, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 429 | acc1 = __SMLALDX(x1, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 430 | acc2 = __SMLALDX(x3, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 431 | acc3 = __SMLALDX(x2, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 432 | |
emilmont | 2:da51fb522205 | 433 | c0 = *(py-1); |
emilmont | 1:fdd22bb7aa52 | 434 | |
emilmont | 1:fdd22bb7aa52 | 435 | #ifdef ARM_MATH_BIG_ENDIAN |
emilmont | 1:fdd22bb7aa52 | 436 | |
emilmont | 1:fdd22bb7aa52 | 437 | c0 = c0 << 16u; |
emilmont | 1:fdd22bb7aa52 | 438 | #else |
emilmont | 1:fdd22bb7aa52 | 439 | |
emilmont | 1:fdd22bb7aa52 | 440 | c0 = c0 & 0x0000FFFF; |
emilmont | 1:fdd22bb7aa52 | 441 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emilmont | 1:fdd22bb7aa52 | 442 | /* Read x[10] */ |
emilmont | 1:fdd22bb7aa52 | 443 | x3 = _SIMD32_OFFSET(px+2); |
emilmont | 2:da51fb522205 | 444 | px += 3u; |
emilmont | 1:fdd22bb7aa52 | 445 | |
emilmont | 1:fdd22bb7aa52 | 446 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 447 | acc0 = __SMLALDX(x1, c0, acc0); |
emilmont | 1:fdd22bb7aa52 | 448 | acc1 = __SMLALD(x2, c0, acc1); |
emilmont | 1:fdd22bb7aa52 | 449 | acc2 = __SMLALDX(x2, c0, acc2); |
emilmont | 1:fdd22bb7aa52 | 450 | acc3 = __SMLALDX(x3, c0, acc3); |
emilmont | 1:fdd22bb7aa52 | 451 | } |
emilmont | 1:fdd22bb7aa52 | 452 | |
emilmont | 1:fdd22bb7aa52 | 453 | |
emilmont | 1:fdd22bb7aa52 | 454 | /* Store the results in the accumulators in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 455 | |
emilmont | 1:fdd22bb7aa52 | 456 | #ifndef ARM_MATH_BIG_ENDIAN |
emilmont | 1:fdd22bb7aa52 | 457 | |
emilmont | 1:fdd22bb7aa52 | 458 | *__SIMD32(pOut)++ = |
emilmont | 1:fdd22bb7aa52 | 459 | __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16); |
emilmont | 1:fdd22bb7aa52 | 460 | *__SIMD32(pOut)++ = |
emilmont | 1:fdd22bb7aa52 | 461 | __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16); |
emilmont | 1:fdd22bb7aa52 | 462 | |
emilmont | 1:fdd22bb7aa52 | 463 | #else |
emilmont | 1:fdd22bb7aa52 | 464 | |
emilmont | 1:fdd22bb7aa52 | 465 | *__SIMD32(pOut)++ = |
emilmont | 1:fdd22bb7aa52 | 466 | __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16); |
emilmont | 1:fdd22bb7aa52 | 467 | *__SIMD32(pOut)++ = |
emilmont | 1:fdd22bb7aa52 | 468 | __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16); |
emilmont | 1:fdd22bb7aa52 | 469 | |
emilmont | 1:fdd22bb7aa52 | 470 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emilmont | 1:fdd22bb7aa52 | 471 | |
emilmont | 1:fdd22bb7aa52 | 472 | /* Increment the pointer pIn1 index, count by 4 */ |
emilmont | 1:fdd22bb7aa52 | 473 | count += 4u; |
emilmont | 1:fdd22bb7aa52 | 474 | |
emilmont | 1:fdd22bb7aa52 | 475 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 476 | px = pIn1 + count; |
emilmont | 1:fdd22bb7aa52 | 477 | py = pSrc2; |
emilmont | 1:fdd22bb7aa52 | 478 | |
emilmont | 1:fdd22bb7aa52 | 479 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 480 | blkCnt--; |
emilmont | 1:fdd22bb7aa52 | 481 | } |
emilmont | 1:fdd22bb7aa52 | 482 | |
emilmont | 1:fdd22bb7aa52 | 483 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
emilmont | 1:fdd22bb7aa52 | 484 | ** No loop unrolling is used. */ |
emilmont | 1:fdd22bb7aa52 | 485 | blkCnt = blockSize2 % 0x4u; |
emilmont | 1:fdd22bb7aa52 | 486 | |
emilmont | 1:fdd22bb7aa52 | 487 | while(blkCnt > 0u) |
emilmont | 1:fdd22bb7aa52 | 488 | { |
emilmont | 1:fdd22bb7aa52 | 489 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 490 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 491 | |
emilmont | 1:fdd22bb7aa52 | 492 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emilmont | 1:fdd22bb7aa52 | 493 | k = srcBLen >> 2u; |
emilmont | 1:fdd22bb7aa52 | 494 | |
emilmont | 1:fdd22bb7aa52 | 495 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emilmont | 1:fdd22bb7aa52 | 496 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emilmont | 1:fdd22bb7aa52 | 497 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 498 | { |
emilmont | 1:fdd22bb7aa52 | 499 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 500 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 501 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 502 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 503 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 504 | |
emilmont | 1:fdd22bb7aa52 | 505 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 506 | k--; |
emilmont | 1:fdd22bb7aa52 | 507 | } |
emilmont | 1:fdd22bb7aa52 | 508 | |
emilmont | 1:fdd22bb7aa52 | 509 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emilmont | 1:fdd22bb7aa52 | 510 | ** No loop unrolling is used. */ |
emilmont | 1:fdd22bb7aa52 | 511 | k = srcBLen % 0x4u; |
emilmont | 1:fdd22bb7aa52 | 512 | |
emilmont | 1:fdd22bb7aa52 | 513 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 514 | { |
emilmont | 1:fdd22bb7aa52 | 515 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 516 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 517 | |
emilmont | 1:fdd22bb7aa52 | 518 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 519 | k--; |
emilmont | 1:fdd22bb7aa52 | 520 | } |
emilmont | 1:fdd22bb7aa52 | 521 | |
emilmont | 1:fdd22bb7aa52 | 522 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 523 | *pOut++ = (q15_t) (__SSAT(sum >> 15, 16)); |
emilmont | 1:fdd22bb7aa52 | 524 | |
emilmont | 1:fdd22bb7aa52 | 525 | /* Increment the pointer pIn1 index, count by 1 */ |
emilmont | 1:fdd22bb7aa52 | 526 | count++; |
emilmont | 1:fdd22bb7aa52 | 527 | |
emilmont | 1:fdd22bb7aa52 | 528 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 529 | px = pIn1 + count; |
emilmont | 1:fdd22bb7aa52 | 530 | py = pSrc2; |
emilmont | 1:fdd22bb7aa52 | 531 | |
emilmont | 1:fdd22bb7aa52 | 532 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 533 | blkCnt--; |
emilmont | 1:fdd22bb7aa52 | 534 | } |
emilmont | 1:fdd22bb7aa52 | 535 | } |
emilmont | 1:fdd22bb7aa52 | 536 | else |
emilmont | 1:fdd22bb7aa52 | 537 | { |
emilmont | 1:fdd22bb7aa52 | 538 | /* If the srcBLen is not a multiple of 4, |
emilmont | 1:fdd22bb7aa52 | 539 | * the blockSize2 loop cannot be unrolled by 4 */ |
emilmont | 1:fdd22bb7aa52 | 540 | blkCnt = blockSize2; |
emilmont | 1:fdd22bb7aa52 | 541 | |
emilmont | 1:fdd22bb7aa52 | 542 | while(blkCnt > 0u) |
emilmont | 1:fdd22bb7aa52 | 543 | { |
emilmont | 1:fdd22bb7aa52 | 544 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 545 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 546 | |
emilmont | 1:fdd22bb7aa52 | 547 | /* srcBLen number of MACS should be performed */ |
emilmont | 1:fdd22bb7aa52 | 548 | k = srcBLen; |
emilmont | 1:fdd22bb7aa52 | 549 | |
emilmont | 1:fdd22bb7aa52 | 550 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 551 | { |
emilmont | 1:fdd22bb7aa52 | 552 | /* Perform the multiply-accumulate */ |
emilmont | 1:fdd22bb7aa52 | 553 | sum += (q63_t) ((q31_t) * px++ * *py--); |
emilmont | 1:fdd22bb7aa52 | 554 | |
emilmont | 1:fdd22bb7aa52 | 555 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 556 | k--; |
emilmont | 1:fdd22bb7aa52 | 557 | } |
emilmont | 1:fdd22bb7aa52 | 558 | |
emilmont | 1:fdd22bb7aa52 | 559 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 560 | *pOut++ = (q15_t) (__SSAT(sum >> 15, 16)); |
emilmont | 1:fdd22bb7aa52 | 561 | |
emilmont | 1:fdd22bb7aa52 | 562 | /* Increment the MAC count */ |
emilmont | 1:fdd22bb7aa52 | 563 | count++; |
emilmont | 1:fdd22bb7aa52 | 564 | |
emilmont | 1:fdd22bb7aa52 | 565 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 566 | px = pIn1 + count; |
emilmont | 1:fdd22bb7aa52 | 567 | py = pSrc2; |
emilmont | 1:fdd22bb7aa52 | 568 | |
emilmont | 1:fdd22bb7aa52 | 569 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 570 | blkCnt--; |
emilmont | 1:fdd22bb7aa52 | 571 | } |
emilmont | 1:fdd22bb7aa52 | 572 | } |
emilmont | 1:fdd22bb7aa52 | 573 | |
emilmont | 1:fdd22bb7aa52 | 574 | |
emilmont | 1:fdd22bb7aa52 | 575 | /* -------------------------- |
emilmont | 1:fdd22bb7aa52 | 576 | * Initializations of stage3 |
emilmont | 1:fdd22bb7aa52 | 577 | * -------------------------*/ |
emilmont | 1:fdd22bb7aa52 | 578 | |
emilmont | 1:fdd22bb7aa52 | 579 | /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1] |
emilmont | 1:fdd22bb7aa52 | 580 | * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2] |
emilmont | 1:fdd22bb7aa52 | 581 | * .... |
emilmont | 1:fdd22bb7aa52 | 582 | * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2] |
emilmont | 1:fdd22bb7aa52 | 583 | * sum += x[srcALen-1] * y[srcBLen-1] |
emilmont | 1:fdd22bb7aa52 | 584 | */ |
emilmont | 1:fdd22bb7aa52 | 585 | |
emilmont | 1:fdd22bb7aa52 | 586 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
emilmont | 1:fdd22bb7aa52 | 587 | The blockSize3 variable holds the number of MAC operations performed */ |
emilmont | 1:fdd22bb7aa52 | 588 | |
emilmont | 1:fdd22bb7aa52 | 589 | blockSize3 = srcBLen - 1u; |
emilmont | 1:fdd22bb7aa52 | 590 | |
emilmont | 1:fdd22bb7aa52 | 591 | /* Working pointer of inputA */ |
emilmont | 1:fdd22bb7aa52 | 592 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u); |
emilmont | 1:fdd22bb7aa52 | 593 | px = pSrc1; |
emilmont | 1:fdd22bb7aa52 | 594 | |
emilmont | 1:fdd22bb7aa52 | 595 | /* Working pointer of inputB */ |
emilmont | 1:fdd22bb7aa52 | 596 | pSrc2 = pIn2 + (srcBLen - 1u); |
emilmont | 1:fdd22bb7aa52 | 597 | pIn2 = pSrc2 - 1u; |
emilmont | 1:fdd22bb7aa52 | 598 | py = pIn2; |
emilmont | 1:fdd22bb7aa52 | 599 | |
emilmont | 1:fdd22bb7aa52 | 600 | /* ------------------- |
emilmont | 1:fdd22bb7aa52 | 601 | * Stage3 process |
emilmont | 1:fdd22bb7aa52 | 602 | * ------------------*/ |
emilmont | 1:fdd22bb7aa52 | 603 | |
emilmont | 1:fdd22bb7aa52 | 604 | /* For loop unrolling by 4, this stage is divided into two. */ |
emilmont | 1:fdd22bb7aa52 | 605 | /* First part of this stage computes the MAC operations greater than 4 */ |
emilmont | 1:fdd22bb7aa52 | 606 | /* Second part of this stage computes the MAC operations less than or equal to 4 */ |
emilmont | 1:fdd22bb7aa52 | 607 | |
emilmont | 1:fdd22bb7aa52 | 608 | /* The first part of the stage starts here */ |
emilmont | 1:fdd22bb7aa52 | 609 | j = blockSize3 >> 2u; |
emilmont | 1:fdd22bb7aa52 | 610 | |
emilmont | 1:fdd22bb7aa52 | 611 | while((j > 0u) && (blockSize3 > 0u)) |
emilmont | 1:fdd22bb7aa52 | 612 | { |
emilmont | 1:fdd22bb7aa52 | 613 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 614 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 615 | |
emilmont | 1:fdd22bb7aa52 | 616 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emilmont | 1:fdd22bb7aa52 | 617 | k = blockSize3 >> 2u; |
emilmont | 1:fdd22bb7aa52 | 618 | |
emilmont | 1:fdd22bb7aa52 | 619 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emilmont | 1:fdd22bb7aa52 | 620 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emilmont | 1:fdd22bb7aa52 | 621 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 622 | { |
emilmont | 1:fdd22bb7aa52 | 623 | /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied |
emilmont | 1:fdd22bb7aa52 | 624 | * with y[srcBLen - 1], y[srcBLen - 2] respectively */ |
emilmont | 1:fdd22bb7aa52 | 625 | sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum); |
emilmont | 1:fdd22bb7aa52 | 626 | /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied |
emilmont | 1:fdd22bb7aa52 | 627 | * with y[srcBLen - 3], y[srcBLen - 4] respectively */ |
emilmont | 1:fdd22bb7aa52 | 628 | sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum); |
emilmont | 1:fdd22bb7aa52 | 629 | |
emilmont | 1:fdd22bb7aa52 | 630 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 631 | k--; |
emilmont | 1:fdd22bb7aa52 | 632 | } |
emilmont | 1:fdd22bb7aa52 | 633 | |
emilmont | 1:fdd22bb7aa52 | 634 | /* For the next MAC operations, the pointer py is used without SIMD |
emilmont | 1:fdd22bb7aa52 | 635 | * So, py is incremented by 1 */ |
emilmont | 1:fdd22bb7aa52 | 636 | py = py + 1u; |
emilmont | 1:fdd22bb7aa52 | 637 | |
emilmont | 1:fdd22bb7aa52 | 638 | /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here. |
emilmont | 1:fdd22bb7aa52 | 639 | ** No loop unrolling is used. */ |
emilmont | 1:fdd22bb7aa52 | 640 | k = blockSize3 % 0x4u; |
emilmont | 1:fdd22bb7aa52 | 641 | |
emilmont | 1:fdd22bb7aa52 | 642 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 643 | { |
emilmont | 1:fdd22bb7aa52 | 644 | /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */ |
emilmont | 1:fdd22bb7aa52 | 645 | sum = __SMLALD(*px++, *py--, sum); |
emilmont | 1:fdd22bb7aa52 | 646 | |
emilmont | 1:fdd22bb7aa52 | 647 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 648 | k--; |
emilmont | 1:fdd22bb7aa52 | 649 | } |
emilmont | 1:fdd22bb7aa52 | 650 | |
emilmont | 1:fdd22bb7aa52 | 651 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 652 | *pOut++ = (q15_t) (__SSAT((sum >> 15), 16)); |
emilmont | 1:fdd22bb7aa52 | 653 | |
emilmont | 1:fdd22bb7aa52 | 654 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 655 | px = ++pSrc1; |
emilmont | 1:fdd22bb7aa52 | 656 | py = pIn2; |
emilmont | 1:fdd22bb7aa52 | 657 | |
emilmont | 1:fdd22bb7aa52 | 658 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 659 | blockSize3--; |
emilmont | 1:fdd22bb7aa52 | 660 | |
emilmont | 1:fdd22bb7aa52 | 661 | j--; |
emilmont | 1:fdd22bb7aa52 | 662 | } |
emilmont | 1:fdd22bb7aa52 | 663 | |
emilmont | 1:fdd22bb7aa52 | 664 | /* The second part of the stage starts here */ |
emilmont | 1:fdd22bb7aa52 | 665 | /* SIMD is not used for the next MAC operations, |
emilmont | 1:fdd22bb7aa52 | 666 | * so pointer py is updated to read only one sample at a time */ |
emilmont | 1:fdd22bb7aa52 | 667 | py = py + 1u; |
emilmont | 1:fdd22bb7aa52 | 668 | |
emilmont | 1:fdd22bb7aa52 | 669 | while(blockSize3 > 0u) |
emilmont | 1:fdd22bb7aa52 | 670 | { |
emilmont | 1:fdd22bb7aa52 | 671 | /* Accumulator is made zero for every iteration */ |
emilmont | 1:fdd22bb7aa52 | 672 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 673 | |
emilmont | 1:fdd22bb7aa52 | 674 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emilmont | 1:fdd22bb7aa52 | 675 | k = blockSize3; |
emilmont | 1:fdd22bb7aa52 | 676 | |
emilmont | 1:fdd22bb7aa52 | 677 | while(k > 0u) |
emilmont | 1:fdd22bb7aa52 | 678 | { |
emilmont | 1:fdd22bb7aa52 | 679 | /* Perform the multiply-accumulates */ |
emilmont | 1:fdd22bb7aa52 | 680 | /* sum += x[srcALen-1] * y[srcBLen-1] */ |
emilmont | 1:fdd22bb7aa52 | 681 | sum = __SMLALD(*px++, *py--, sum); |
emilmont | 1:fdd22bb7aa52 | 682 | |
emilmont | 1:fdd22bb7aa52 | 683 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 684 | k--; |
emilmont | 1:fdd22bb7aa52 | 685 | } |
emilmont | 1:fdd22bb7aa52 | 686 | |
emilmont | 1:fdd22bb7aa52 | 687 | /* Store the result in the accumulator in the destination buffer. */ |
emilmont | 1:fdd22bb7aa52 | 688 | *pOut++ = (q15_t) (__SSAT((sum >> 15), 16)); |
emilmont | 1:fdd22bb7aa52 | 689 | |
emilmont | 1:fdd22bb7aa52 | 690 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emilmont | 1:fdd22bb7aa52 | 691 | px = ++pSrc1; |
emilmont | 1:fdd22bb7aa52 | 692 | py = pSrc2; |
emilmont | 1:fdd22bb7aa52 | 693 | |
emilmont | 1:fdd22bb7aa52 | 694 | /* Decrement the loop counter */ |
emilmont | 1:fdd22bb7aa52 | 695 | blockSize3--; |
emilmont | 1:fdd22bb7aa52 | 696 | } |
emilmont | 1:fdd22bb7aa52 | 697 | |
emilmont | 1:fdd22bb7aa52 | 698 | #else |
emilmont | 1:fdd22bb7aa52 | 699 | |
emilmont | 1:fdd22bb7aa52 | 700 | /* Run the below code for Cortex-M0 */ |
emilmont | 1:fdd22bb7aa52 | 701 | |
emilmont | 1:fdd22bb7aa52 | 702 | q15_t *pIn1 = pSrcA; /* input pointer */ |
emilmont | 1:fdd22bb7aa52 | 703 | q15_t *pIn2 = pSrcB; /* coefficient pointer */ |
emilmont | 1:fdd22bb7aa52 | 704 | q63_t sum; /* Accumulator */ |
emilmont | 1:fdd22bb7aa52 | 705 | uint32_t i, j; /* loop counter */ |
emilmont | 1:fdd22bb7aa52 | 706 | |
emilmont | 1:fdd22bb7aa52 | 707 | /* Loop to calculate output of convolution for output length number of times */ |
emilmont | 1:fdd22bb7aa52 | 708 | for (i = 0; i < (srcALen + srcBLen - 1); i++) |
emilmont | 1:fdd22bb7aa52 | 709 | { |
emilmont | 1:fdd22bb7aa52 | 710 | /* Initialize sum with zero to carry on MAC operations */ |
emilmont | 1:fdd22bb7aa52 | 711 | sum = 0; |
emilmont | 1:fdd22bb7aa52 | 712 | |
emilmont | 1:fdd22bb7aa52 | 713 | /* Loop to perform MAC operations according to convolution equation */ |
emilmont | 1:fdd22bb7aa52 | 714 | for (j = 0; j <= i; j++) |
emilmont | 1:fdd22bb7aa52 | 715 | { |
emilmont | 1:fdd22bb7aa52 | 716 | /* Check the array limitations */ |
emilmont | 1:fdd22bb7aa52 | 717 | if(((i - j) < srcBLen) && (j < srcALen)) |
emilmont | 1:fdd22bb7aa52 | 718 | { |
emilmont | 1:fdd22bb7aa52 | 719 | /* z[i] += x[i-j] * y[j] */ |
emilmont | 1:fdd22bb7aa52 | 720 | sum += (q31_t) pIn1[j] * (pIn2[i - j]); |
emilmont | 1:fdd22bb7aa52 | 721 | } |
emilmont | 1:fdd22bb7aa52 | 722 | } |
emilmont | 1:fdd22bb7aa52 | 723 | |
emilmont | 1:fdd22bb7aa52 | 724 | /* Store the output in the destination buffer */ |
emilmont | 1:fdd22bb7aa52 | 725 | pDst[i] = (q15_t) __SSAT((sum >> 15u), 16u); |
emilmont | 1:fdd22bb7aa52 | 726 | } |
emilmont | 1:fdd22bb7aa52 | 727 | |
emilmont | 1:fdd22bb7aa52 | 728 | #endif /* #if (defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE)*/ |
emilmont | 1:fdd22bb7aa52 | 729 | |
emilmont | 1:fdd22bb7aa52 | 730 | } |
emilmont | 1:fdd22bb7aa52 | 731 | |
emilmont | 1:fdd22bb7aa52 | 732 | /** |
emilmont | 1:fdd22bb7aa52 | 733 | * @} end of Conv group |
emilmont | 1:fdd22bb7aa52 | 734 | */ |