Robert Lopez / CMSIS5

Aded CMSIS5 DSP and NN folder. Needs some work

NN/source/SoftmaxFunctions/arm_softmax_q15.c@0:eedb7d567a5d, 2018-04-12 (annotated)

Committer:
robert_lp
Date:
Thu Apr 12 01:31:58 2018 +0000
Revision:
0:eedb7d567a5d
CMSIS5 Library

Who changed what in which revision?

UserRevisionLine numberNew contents of line
robert_lp 0:eedb7d567a5d 1 /*
robert_lp 0:eedb7d567a5d 2 * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
robert_lp 0:eedb7d567a5d 3 *
robert_lp 0:eedb7d567a5d 4 * SPDX-License-Identifier: Apache-2.0
robert_lp 0:eedb7d567a5d 5 *
robert_lp 0:eedb7d567a5d 6 * Licensed under the Apache License, Version 2.0 (the License); you may
robert_lp 0:eedb7d567a5d 7 * not use this file except in compliance with the License.
robert_lp 0:eedb7d567a5d 8 * You may obtain a copy of the License at
robert_lp 0:eedb7d567a5d 9 *
robert_lp 0:eedb7d567a5d 10 * www.apache.org/licenses/LICENSE-2.0
robert_lp 0:eedb7d567a5d 11 *
robert_lp 0:eedb7d567a5d 12 * Unless required by applicable law or agreed to in writing, software
robert_lp 0:eedb7d567a5d 13 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
robert_lp 0:eedb7d567a5d 14 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
robert_lp 0:eedb7d567a5d 15 * See the License for the specific language governing permissions and
robert_lp 0:eedb7d567a5d 16 * limitations under the License.
robert_lp 0:eedb7d567a5d 17 */
robert_lp 0:eedb7d567a5d 18
robert_lp 0:eedb7d567a5d 19 /* ----------------------------------------------------------------------
robert_lp 0:eedb7d567a5d 20 * Project: CMSIS NN Library
robert_lp 0:eedb7d567a5d 21 * Title: arm_softmax_q15.c
robert_lp 0:eedb7d567a5d 22 * Description: Q15 softmax function
robert_lp 0:eedb7d567a5d 23 *
robert_lp 0:eedb7d567a5d 24 * $Date: 17. January 2018
robert_lp 0:eedb7d567a5d 25 * $Revision: V.1.0.0
robert_lp 0:eedb7d567a5d 26 *
robert_lp 0:eedb7d567a5d 27 * Target Processor: Cortex-M cores
robert_lp 0:eedb7d567a5d 28 *
robert_lp 0:eedb7d567a5d 29 * -------------------------------------------------------------------- */
robert_lp 0:eedb7d567a5d 30
robert_lp 0:eedb7d567a5d 31 #include "arm_math.h"
robert_lp 0:eedb7d567a5d 32 #include "arm_nnfunctions.h"
robert_lp 0:eedb7d567a5d 33
robert_lp 0:eedb7d567a5d 34 /**
robert_lp 0:eedb7d567a5d 35 * @ingroup groupNN
robert_lp 0:eedb7d567a5d 36 */
robert_lp 0:eedb7d567a5d 37
robert_lp 0:eedb7d567a5d 38 /**
robert_lp 0:eedb7d567a5d 39 * @addtogroup Softmax
robert_lp 0:eedb7d567a5d 40 * @{
robert_lp 0:eedb7d567a5d 41 */
robert_lp 0:eedb7d567a5d 42
robert_lp 0:eedb7d567a5d 43 /**
robert_lp 0:eedb7d567a5d 44 * @brief Q15 softmax function
robert_lp 0:eedb7d567a5d 45 * @param[in] vec_in pointer to input vector
robert_lp 0:eedb7d567a5d 46 * @param[in] dim_vec input vector dimention
robert_lp 0:eedb7d567a5d 47 * @param[out] p_out pointer to output vector
robert_lp 0:eedb7d567a5d 48 * @return none.
robert_lp 0:eedb7d567a5d 49 *
robert_lp 0:eedb7d567a5d 50 * @details
robert_lp 0:eedb7d567a5d 51 *
robert_lp 0:eedb7d567a5d 52 * Here, instead of typical e based softmax, we use
robert_lp 0:eedb7d567a5d 53 * 2-based softmax, i.e.,:
robert_lp 0:eedb7d567a5d 54 *
robert_lp 0:eedb7d567a5d 55 * y_i = 2^(x_i) / sum(2^x_j)
robert_lp 0:eedb7d567a5d 56 *
robert_lp 0:eedb7d567a5d 57 * The relative output will be different here.
robert_lp 0:eedb7d567a5d 58 * But mathematically, the gradient will be the same
robert_lp 0:eedb7d567a5d 59 * with a log(2) scaling factor.
robert_lp 0:eedb7d567a5d 60 *
robert_lp 0:eedb7d567a5d 61 */
robert_lp 0:eedb7d567a5d 62
robert_lp 0:eedb7d567a5d 63 void arm_softmax_q15(const q15_t * vec_in, const uint16_t dim_vec, q15_t * p_out)
robert_lp 0:eedb7d567a5d 64 {
robert_lp 0:eedb7d567a5d 65 q31_t sum;
robert_lp 0:eedb7d567a5d 66 int16_t i;
robert_lp 0:eedb7d567a5d 67 q31_t min, max;
robert_lp 0:eedb7d567a5d 68 max = -1 * 0x100000;
robert_lp 0:eedb7d567a5d 69 min = 0x100000;
robert_lp 0:eedb7d567a5d 70 for (i = 0; i < dim_vec; i++)
robert_lp 0:eedb7d567a5d 71 {
robert_lp 0:eedb7d567a5d 72 if (vec_in[i] > max)
robert_lp 0:eedb7d567a5d 73 {
robert_lp 0:eedb7d567a5d 74 max = vec_in[i];
robert_lp 0:eedb7d567a5d 75 }
robert_lp 0:eedb7d567a5d 76 if (vec_in[i] < min)
robert_lp 0:eedb7d567a5d 77 {
robert_lp 0:eedb7d567a5d 78 min = vec_in[i];
robert_lp 0:eedb7d567a5d 79 }
robert_lp 0:eedb7d567a5d 80 }
robert_lp 0:eedb7d567a5d 81
robert_lp 0:eedb7d567a5d 82 /* we ignore really small values
robert_lp 0:eedb7d567a5d 83 * anyway, they will be 0 after shrinking
robert_lp 0:eedb7d567a5d 84 * to q7_t
robert_lp 0:eedb7d567a5d 85 */
robert_lp 0:eedb7d567a5d 86 if (max - min > 16)
robert_lp 0:eedb7d567a5d 87 {
robert_lp 0:eedb7d567a5d 88 min = max - 16;
robert_lp 0:eedb7d567a5d 89 }
robert_lp 0:eedb7d567a5d 90
robert_lp 0:eedb7d567a5d 91 sum = 0;
robert_lp 0:eedb7d567a5d 92
robert_lp 0:eedb7d567a5d 93 for (i = 0; i < dim_vec; i++)
robert_lp 0:eedb7d567a5d 94 {
robert_lp 0:eedb7d567a5d 95 sum += 0x1 << (vec_in[i] - min);
robert_lp 0:eedb7d567a5d 96 }
robert_lp 0:eedb7d567a5d 97
robert_lp 0:eedb7d567a5d 98 for (i = 0; i < dim_vec; i++)
robert_lp 0:eedb7d567a5d 99 {
robert_lp 0:eedb7d567a5d 100 /* we leave 7-bit dynamic range, so that 128 -> 100% confidence */
robert_lp 0:eedb7d567a5d 101 p_out[i] = (q15_t) __SSAT(((0x1 << (vec_in[i] - min + 14)) / sum), 16);
robert_lp 0:eedb7d567a5d 102 }
robert_lp 0:eedb7d567a5d 103
robert_lp 0:eedb7d567a5d 104 }
robert_lp 0:eedb7d567a5d 105
robert_lp 0:eedb7d567a5d 106 /**
robert_lp 0:eedb7d567a5d 107 * @} end of Softmax group
robert_lp 0:eedb7d567a5d 108 */
robert_lp 0:eedb7d567a5d 109