1 /*
2 * Copyright (c) 2019 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "ConcatenateLayer.h"
25
26 #include "tests/validation/Helpers.h"
27 #include "tests/validation/reference/Permute.h"
28
29 namespace arm_compute
30 {
31 namespace test
32 {
33 namespace validation
34 {
35 namespace reference
36 {
37 namespace
38 {
39 template <typename T>
widthconcatenate_layer(const std::vector<SimpleTensor<T>> & srcs,SimpleTensor<T> & dst)40 SimpleTensor<T> widthconcatenate_layer(const std::vector<SimpleTensor<T>> &srcs, SimpleTensor<T> &dst)
41 {
42 // Create reference
43 std::vector<TensorShape> shapes;
44 shapes.reserve(srcs.size());
45 for(const auto &src : srcs)
46 {
47 shapes.emplace_back(src.shape());
48 }
49 // Compute reference
50 int width_offset = 0;
51 const int width_out = dst.shape().x();
52 // Set output tensor to 0
53 std::fill_n(dst.data(), dst.num_elements(), 0);
54 for(const auto &src : srcs)
55 {
56 ARM_COMPUTE_ERROR_ON(width_offset >= width_out);
57
58 const int width = src.shape().x();
59 const int height = src.shape().y();
60 const int depth = src.shape().z();
61 const int upper_dims = src.shape().total_size() / (width * height * depth);
62
63 const T *src_ptr = src.data();
64 T *dst_ptr = dst.data();
65
66 for(int u = 0; u < upper_dims; ++u)
67 {
68 for(int d = 0; d < depth; ++d)
69 {
70 for(int r = 0; r < height; ++r)
71 {
72 const int offset = u * height * depth + d * height + r;
73 if(is_data_type_quantized(src.data_type()) && src.quantization_info() != dst.quantization_info())
74 {
75 const UniformQuantizationInfo iq_info = src.quantization_info().uniform();
76 const UniformQuantizationInfo oq_info = dst.quantization_info().uniform();
77
78 if(src.data_type() == DataType::QASYMM8)
79 {
80 std::transform(src_ptr, src_ptr + width, dst_ptr + width_offset + offset * width_out, [&](T t)
81 {
82 const float dequantized_input = dequantize_qasymm8(t, iq_info);
83 return quantize_qasymm8(dequantized_input, oq_info);
84 });
85 }
86 else
87 {
88 std::transform(src_ptr, src_ptr + width, dst_ptr + width_offset + offset * width_out, [&](T t)
89 {
90 const float dequantized_input = dequantize_qasymm8_signed(t, iq_info);
91 return quantize_qasymm8_signed(dequantized_input, oq_info);
92 });
93 }
94 src_ptr += width;
95 }
96 else
97 {
98 std::copy(src_ptr, src_ptr + width, dst_ptr + width_offset + offset * width_out);
99 src_ptr += width;
100 }
101 }
102 }
103 }
104 width_offset += width;
105 }
106 return dst;
107 }
108
109 template SimpleTensor<float> widthconcatenate_layer(const std::vector<SimpleTensor<float>> &srcs, SimpleTensor<float> &dst);
110 template SimpleTensor<half> widthconcatenate_layer(const std::vector<SimpleTensor<half>> &srcs, SimpleTensor<half> &dst);
111 template SimpleTensor<uint8_t> widthconcatenate_layer(const std::vector<SimpleTensor<uint8_t>> &srcs, SimpleTensor<uint8_t> &dst);
112 template SimpleTensor<int8_t> widthconcatenate_layer(const std::vector<SimpleTensor<int8_t>> &srcs, SimpleTensor<int8_t> &dst);
113 } // namespace
114
115 template <typename T>
concatenate_layer(std::vector<SimpleTensor<T>> & srcs,SimpleTensor<T> & dst,unsigned int axis)116 SimpleTensor<T> concatenate_layer(std::vector<SimpleTensor<T>> &srcs, SimpleTensor<T> &dst, unsigned int axis)
117 {
118 switch(axis)
119 {
120 case Window::DimX:
121 {
122 return widthconcatenate_layer(srcs, dst);
123 }
124 case Window::DimY:
125 {
126 for(auto &t : srcs)
127 {
128 t = reference::permute<T>(t, PermutationVector(1U, 0U));
129 }
130 dst = reference::permute<T>(dst, PermutationVector(1U, 0U));
131 return reference::permute<T>(widthconcatenate_layer(srcs, dst), PermutationVector(1U, 0U));
132 }
133 case Window::DimZ:
134 {
135 for(auto &t : srcs)
136 {
137 t = reference::permute<T>(t, PermutationVector(2U, 1U, 0U));
138 }
139 dst = reference::permute<T>(dst, PermutationVector(2U, 1U, 0U));
140 return reference::permute<T>(widthconcatenate_layer(srcs, dst), PermutationVector(2U, 1U, 0U));
141 }
142 case 3:
143 {
144 for(auto &t : srcs)
145 {
146 t = reference::permute<T>(t, PermutationVector(3U, 2U, 1U, 0U));
147 }
148 dst = reference::permute<T>(dst, PermutationVector(3U, 2U, 1U, 0U));
149 auto ret = reference::permute<T>(widthconcatenate_layer(srcs, dst), PermutationVector(3U, 2U, 1U, 0U));
150 return ret;
151 }
152 default:
153 {
154 ARM_COMPUTE_ERROR("Not supported");
155 return dst;
156 }
157 }
158 }
159
160 template SimpleTensor<float> concatenate_layer(std::vector<SimpleTensor<float>> &srcs, SimpleTensor<float> &dst, unsigned int axis);
161 template SimpleTensor<half> concatenate_layer(std::vector<SimpleTensor<half>> &srcs, SimpleTensor<half> &dst, unsigned int axis);
162 template SimpleTensor<uint8_t> concatenate_layer(std::vector<SimpleTensor<uint8_t>> &srcs, SimpleTensor<uint8_t> &dst, unsigned int axis);
163 template SimpleTensor<int8_t> concatenate_layer(std::vector<SimpleTensor<int8_t>> &srcs, SimpleTensor<int8_t> &dst, unsigned int axis);
164 } // namespace reference
165 } // namespace validation
166 } // namespace test
167 } // namespace arm_compute
168