1 /*
2 * SPDX-FileCopyrightText: Copyright 2010-2023 Arm Limited and/or its affiliates <open-source-office@arm.com>
3 *
4 * SPDX-License-Identifier: Apache-2.0
5 *
6 * Licensed under the Apache License, Version 2.0 (the License); you may
7 * not use this file except in compliance with the License.
8 * You may obtain a copy of the License at
9 *
10 * www.apache.org/licenses/LICENSE-2.0
11 *
12 * Unless required by applicable law or agreed to in writing, software
13 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
14 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 * See the License for the specific language governing permissions and
16 * limitations under the License.
17 */
18
19 /* ----------------------------------------------------------------------
20 * Project: CMSIS NN Library
21 * Title: arm_convolve_1_x_n_s8.c
22 * Description: s8 version of 1xN convolution using symmetric quantization.
23 *
24 * $Date: 8 March 2023
25 * $Revision: V.3.4.0
26 *
27 * Target : Arm(R) M-Profile Architecture
28 *
29 * -------------------------------------------------------------------- */
30
31 #include "arm_nnfunctions.h"
32 #include "arm_nnsupportfunctions.h"
33
34 /**
35 * @ingroup Public
36 */
37
38 /**
39 * @addtogroup NNConv
40 * @{
41 */
42
43 /*
44 * 1xN s8 convolution function.
45 *
46 * Refer header file for details.
47 *
48 */
49
arm_convolve_1_x_n_s8(const cmsis_nn_context * ctx,const cmsis_nn_conv_params * conv_params,const cmsis_nn_per_channel_quant_params * quant_params,const cmsis_nn_dims * input_dims,const int8_t * input_data,const cmsis_nn_dims * filter_dims,const int8_t * filter_data,const cmsis_nn_dims * bias_dims,const int32_t * bias_data,const cmsis_nn_dims * output_dims,int8_t * output_data)50 arm_cmsis_nn_status arm_convolve_1_x_n_s8(const cmsis_nn_context *ctx,
51 const cmsis_nn_conv_params *conv_params,
52 const cmsis_nn_per_channel_quant_params *quant_params,
53 const cmsis_nn_dims *input_dims,
54 const int8_t *input_data,
55 const cmsis_nn_dims *filter_dims,
56 const int8_t *filter_data,
57 const cmsis_nn_dims *bias_dims,
58 const int32_t *bias_data,
59 const cmsis_nn_dims *output_dims,
60 int8_t *output_data)
61 {
62 arm_cmsis_nn_status status = ARM_CMSIS_NN_SUCCESS;
63 int32_t buffer_size = arm_convolve_1_x_n_s8_get_buffer_size(input_dims, filter_dims);
64 /* The wrapper API is the ultimate reference for argument check */
65 if ((input_dims->h != 1) || conv_params->dilation.w != 1 || (buffer_size != 0 && ctx->buf == NULL) ||
66 conv_params->stride.w == 0 || (conv_params->stride.w * input_dims->c % 4 != 0))
67 {
68 status = ARM_CMSIS_NN_ARG_ERROR;
69 goto out;
70 }
71
72 #if defined(ARM_MATH_MVEI)
73 (void)bias_dims;
74 const uint16_t input_x = input_dims->w;
75 const uint16_t kernel_x = filter_dims->w;
76 const uint16_t output_x = output_dims->w;
77 const uint16_t output_ch = output_dims->c;
78 const uint16_t input_ch = input_dims->c;
79 const uint16_t pad_x = conv_params->padding.w;
80 const uint16_t stride_x = conv_params->stride.w;
81
82 // Total pad for dilation of 1
83 const int32_t total_pad = ((output_x - 1) * stride_x + kernel_x - input_x);
84 const int32_t asym_pad = total_pad % 2;
85
86 if (pad_x * 2 + asym_pad != total_pad)
87 {
88 return ARM_CMSIS_NN_FAILURE;
89 }
90
91 const int32_t right_pad_num = pad_x + asym_pad != 0 ? MAX(1, (pad_x + asym_pad + stride_x - 1) / stride_x) : 0;
92 const int32_t left_pad_num = pad_x != 0 ? MAX(1, (pad_x + stride_x - 1) / stride_x) : 0;
93 const int32_t no_pad_num = MAX(output_x - (right_pad_num + left_pad_num), 0);
94 if (right_pad_num + no_pad_num + left_pad_num != output_x)
95 {
96 return ARM_CMSIS_NN_FAILURE;
97 }
98
99 for (int i_batch = 0; i_batch < input_dims->n; i_batch++)
100 {
101 // Handle left padded sections
102 int32_t lhs_rows = left_pad_num;
103 const int32_t rhs_cols = kernel_x * input_dims->c;
104 const int32_t rhs_rows = output_dims->c;
105 const int32_t lhs_offset = input_ch * stride_x;
106
107 int32_t out_idx = 0;
108
109 for (int i = 0; i < lhs_rows; i++)
110 {
111 const int32_t est_input_x_idx = stride_x * i - pad_x;
112 const int32_t ker_begin_idx = -est_input_x_idx;
113
114 const int32_t actual_kernel_len = kernel_x - ker_begin_idx;
115
116 status = arm_nn_mat_mul_core_1x_s8(actual_kernel_len * input_ch,
117 ker_begin_idx * input_ch,
118 input_data,
119 filter_data + (ker_begin_idx * input_ch),
120 output_ch,
121 conv_params,
122 quant_params,
123 bias_data,
124 output_data);
125 output_data += output_ch;
126 }
127
128 out_idx += lhs_rows;
129 int32_t input_start = stride_x * lhs_rows - pad_x;
130
131 if (input_start < 0)
132 {
133 return ARM_CMSIS_NN_FAILURE;
134 }
135 /* Non padded elements */
136 input_start *= input_ch;
137 lhs_rows = no_pad_num;
138
139 arm_nn_mat_mult_nt_t_s8(input_data + input_start,
140 filter_data,
141 bias_data,
142 output_data,
143 quant_params->multiplier,
144 quant_params->shift,
145 lhs_rows,
146 rhs_rows,
147 rhs_cols,
148 conv_params->input_offset,
149 conv_params->output_offset,
150 conv_params->activation.min,
151 conv_params->activation.max,
152 lhs_offset);
153
154 output_data += lhs_rows * rhs_rows;
155
156 /* Right padded elements */
157 out_idx += lhs_rows;
158 lhs_rows = output_x - out_idx;
159
160 if (lhs_rows < 0)
161 {
162 return ARM_CMSIS_NN_FAILURE;
163 }
164
165 for (int i = out_idx; i < output_x; i++)
166 {
167 const int32_t est_input_x_idx = stride_x * i - pad_x;
168 const int32_t ker_end_idx = MIN(kernel_x, input_x - est_input_x_idx);
169
170 status = arm_nn_mat_mul_core_1x_s8(ker_end_idx * input_ch,
171 (kernel_x - ker_end_idx) * input_ch,
172 input_data + est_input_x_idx * input_ch,
173 filter_data,
174 output_ch,
175 conv_params,
176 quant_params,
177 bias_data,
178 output_data);
179 output_data += output_ch;
180 }
181 /* Advance to the next batch */
182 input_data += (input_x * input_ch);
183 }
184 #else
185 status = arm_convolve_s8(ctx,
186 conv_params,
187 quant_params,
188 input_dims,
189 input_data,
190 filter_dims,
191 filter_data,
192 bias_dims,
193 bias_data,
194 output_dims,
195 output_data);
196
197 #endif
198
199 out:
200 /* Return to application */
201 return status;
202 }
203
204 /**
205 * @} end of NNConv group
206 */
207
