1 /*
2 * SPDX-FileCopyrightText: Copyright 2024 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_transpose_conv_wrapper_s8.c
22 * Description: Wrapper API to select appropriate transpose conv API based
23 * on dimensions.
24 *
25 * $Date: 16 October 2024
26 * $Revision: V.1.0.0
27 *
28 * Target : Arm(R) M-Profile Architecture
29 *
30 * -------------------------------------------------------------------- */
31
32 #include "arm_nnfunctions.h"
33 #include "arm_nnsupportfunctions.h"
34
35 /**
36 * @ingroup Public
37 */
38
39 /**
40 * @addtogroup NNConv
41 * @{
42 */
43
44 /*
45 * s8 Transpose conv wrapper function
46 *
47 * Refer header file for details.
48 *
49 */
arm_transpose_conv_wrapper_s8(const cmsis_nn_context * ctx,const cmsis_nn_context * reverse_conv_ctx,const cmsis_nn_transpose_conv_params * transpose_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_transpose_conv_wrapper_s8(const cmsis_nn_context *ctx,
51 const cmsis_nn_context *reverse_conv_ctx,
52 const cmsis_nn_transpose_conv_params *transpose_conv_params,
53 const cmsis_nn_per_channel_quant_params *quant_params,
54 const cmsis_nn_dims *input_dims,
55 const int8_t *input_data,
56 const cmsis_nn_dims *filter_dims,
57 const int8_t *filter_data,
58 const cmsis_nn_dims *bias_dims,
59 const int32_t *bias_data,
60 const cmsis_nn_dims *output_dims,
61 int8_t *output_data)
62 {
63
64 if (ctx->buf == NULL)
65 {
66 return ARM_CMSIS_NN_ARG_ERROR;
67 }
68
69 const bool reverse_conv_possible =
70 ((transpose_conv_params->stride.w <= 2) && (transpose_conv_params->stride.h <= 2));
71 const bool reverse_conv_efficient = (input_dims->c > REVERSE_TCOL_EFFICIENT_THRESHOLD);
72
73 if (reverse_conv_possible && reverse_conv_efficient)
74 {
75
76 if (reverse_conv_ctx->buf == NULL)
77 {
78 return ARM_CMSIS_NN_ARG_ERROR;
79 }
80
81 const int32_t stride_w = transpose_conv_params->stride.w;
82 const int32_t stride_h = transpose_conv_params->stride.h;
83 const int32_t filter_h = filter_dims->h;
84 const int32_t filter_w = filter_dims->w;
85 const int32_t output_c = output_dims->c;
86 const int32_t input_n = input_dims->n;
87 const int32_t input_h = input_dims->h;
88 const int32_t input_w = input_dims->w;
89 const int32_t input_c = input_dims->c;
90 const int32_t padding_w = transpose_conv_params->padding.w;
91 const int32_t padding_h = transpose_conv_params->padding.h;
92
93 cmsis_nn_conv_params conv_params;
94 conv_params.padding.h = filter_h - 1 - padding_h;
95 conv_params.padding.w = filter_w - 1 - padding_w;
96 conv_params.input_offset = transpose_conv_params->input_offset;
97 conv_params.output_offset = transpose_conv_params->output_offset;
98 conv_params.stride.h = 1;
99 conv_params.stride.w = 1;
100 conv_params.dilation.h = 1;
101 conv_params.dilation.w = 1;
102 conv_params.activation = transpose_conv_params->activation;
103
104 const cmsis_nn_dims transposed_input_dims = {input_n, input_h * stride_h, input_w * stride_w, input_c};
105 const cmsis_nn_dims upscale_dims = {0, stride_h, stride_w, 0};
106
107 // Reverse filter in x and y-dimensions
108 int8_t *reversed_filter = reverse_conv_ctx->buf;
109 const int8_t *in_ptr = filter_data;
110 int8_t *out_ptr = reversed_filter;
111 const int32_t filter_size = filter_h * filter_w * input_c;
112
113 out_ptr += filter_size;
114 for (int32_t i = 0; i < output_c; i++)
115 {
116 for (int32_t y = 0; y < filter_h; y++)
117 {
118 for (int32_t x = 0; x < filter_w; x++)
119 {
120 out_ptr -= input_c;
121 arm_memcpy_s8(out_ptr, in_ptr, input_c * sizeof(int8_t));
122 in_ptr += input_c;
123 }
124 }
125 out_ptr += 2 * filter_size;
126 }
127
128 return arm_convolve_s8(ctx,
129 &conv_params,
130 quant_params,
131 &transposed_input_dims,
132 input_data,
133 filter_dims,
134 reversed_filter,
135 bias_dims,
136 bias_data,
137 &upscale_dims,
138 output_dims,
139 output_data);
140 }
141 else
142 {
143
144 return arm_transpose_conv_s8(ctx,
145 reverse_conv_ctx,
146 transpose_conv_params,
147 quant_params,
148 input_dims,
149 input_data,
150 filter_dims,
151 filter_data,
152 bias_dims,
153 bias_data,
154 output_dims,
155 output_data);
156 }
157 }
158
159 /**
160 * @} end of NNconv group
161 */
162
