• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2020 Huawei Technologies Co., Ltd
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #include "nnacl/int8/softmax_int8.h"
18 #include "nnacl/errorcode.h"
19 
SoftmaxInt8(const int8_t * input_ptr,int8_t * output_ptr,int count,int * exp_data,int * sum_data,const SoftmaxQuantArg * quant_param,const SoftmaxParameter * parameter)20 int SoftmaxInt8(const int8_t *input_ptr, int8_t *output_ptr, int count, int *exp_data, int *sum_data,
21                 const SoftmaxQuantArg *quant_param, const SoftmaxParameter *parameter) {
22   int32_t axis = parameter->axis_;
23   int n_dim = parameter->n_dim_;
24   const int *input_shape = parameter->input_shape_;
25   int axis_shape_size = input_shape[axis];
26 
27   int inner_size = 1;
28   if (n_dim > DIMENSION_5D) {
29     return NNACL_ERR;
30   }
31   for (int i = axis + 1; i < n_dim; i++) {
32     inner_size *= input_shape[i];
33   }
34 
35   for (int o = 0; o < count; o++) {
36     int outter_offset = o * axis_shape_size * inner_size;
37 
38     for (int c = 0; c < inner_size; c++) {
39       int8_t max_row = quant_param->output_activation_min_;
40       for (int i = 0; i < axis_shape_size; ++i) {
41         int axis_offset = outter_offset + c + i * inner_size;
42         max_row = MSMAX(max_row, input_ptr[axis_offset]);
43       }
44 
45       int32_t exp_sum = 0;
46       for (int i = 0; i < axis_shape_size; ++i) {
47         int axis_offset = outter_offset + c + i * inner_size;
48         const int32_t input_val = input_ptr[axis_offset] - max_row;
49         const int32_t input_scaled = SaturatingRoundingDoublingHighMul(
50           input_val * (1 << (unsigned int)quant_param->shift_left_), quant_param->output_multiplier_);
51         int exp_val = exp_on_negative_values(input_scaled, 5);
52         exp_data[axis_offset] = exp_val;
53         exp_sum = exp_sum + Rescale(exp_val, 0, 12);
54       }
55       sum_data[c] = exp_sum;
56     }
57     for (int i = 0; i < axis_shape_size; ++i) {
58       int axis_offset = outter_offset + i * inner_size;
59       for (int c = 0; c < inner_size; ++c) {
60         int num_bits_over_unit;
61         int shifted_scale = ComputerReciprocal(sum_data[c], 12, &num_bits_over_unit);
62         int unsat_output = RoundingDivideByPOT(
63           SaturatingRoundingDoublingHighMul(shifted_scale, exp_data[axis_offset + c]), num_bits_over_unit + 31 - 8);
64 
65         int raw_output = unsat_output + quant_param->output_activation_min_;
66         output_ptr[axis_offset + c] =
67           (int8_t)MSMAX(quant_param->output_activation_min_, MSMIN(raw_output, quant_param->output_activation_max_));
68       }
69     }
70   }
71   return 0;
72 }
73