1 /*
2 * Copyright (c) 2018-2020 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 "src/core/NEON/kernels/NEConvertFullyConnectedWeightsKernel.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Types.h"
28 #include "src/core/helpers/AutoConfiguration.h"
29 #include "src/core/helpers/WindowHelpers.h"
30
31 namespace arm_compute
32 {
NEConvertFullyConnectedWeightsKernel()33 NEConvertFullyConnectedWeightsKernel::NEConvertFullyConnectedWeightsKernel()
34 : _input(nullptr), _output(nullptr), _factor1(0), _factor2(0)
35 {
36 }
37
configure(const ITensor * input,ITensor * output,const TensorShape & original_input_shape,DataLayout data_layout)38 void NEConvertFullyConnectedWeightsKernel::configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape,
39 DataLayout data_layout)
40 {
41 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
42
43 // Output tensor auto initialisation if not yet initialized
44 auto_init_if_empty(*output->info(), *input->info()->clone());
45
46 ARM_COMPUTE_ERROR_THROW_ON(NEConvertFullyConnectedWeightsKernel::validate(input->info(), output->info(), original_input_shape, data_layout));
47
48 _input = input;
49 _output = output;
50
51 const DataLayout input_data_layout = (data_layout == DataLayout::NCHW) ? DataLayout::NHWC : DataLayout::NCHW;
52
53 const int width_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::WIDTH);
54 const int height_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::HEIGHT);
55 const int channel_idx = get_data_layout_dimension_index(input_data_layout, DataLayoutDimension::CHANNEL);
56
57 const unsigned int num_elems_per_input_plane = original_input_shape[width_idx] * original_input_shape[height_idx];
58 const unsigned int num_channels = original_input_shape[channel_idx];
59
60 _factor1 = (data_layout == DataLayout::NCHW) ? num_elems_per_input_plane : num_channels;
61 _factor2 = (data_layout == DataLayout::NCHW) ? num_channels : num_elems_per_input_plane;
62
63 // Configure kernel window
64 Window win = calculate_max_window(*input->info(), Steps());
65 INEKernel::configure(win);
66 }
67
validate(const ITensorInfo * input,const ITensorInfo * output,const TensorShape & original_input_shape,DataLayout data_layout)68 Status NEConvertFullyConnectedWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape,
69 DataLayout data_layout)
70 {
71 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
72 //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
73 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
74 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
75 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != original_input_shape.total_size_lower(3));
76 ARM_COMPUTE_RETURN_ERROR_ON(data_layout == DataLayout::UNKNOWN);
77
78 // Checks performed when output is configured
79 if((output != nullptr) && (output->total_size() != 0))
80 {
81 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
83 }
84
85 return Status{};
86 }
87
88 template <typename T>
run_convert_fc_weights(const Window & window)89 void NEConvertFullyConnectedWeightsKernel::run_convert_fc_weights(const Window &window)
90 {
91 const unsigned int dst_stride_x = _output->info()->strides_in_bytes().x();
92 const unsigned int dst_stride_y = _output->info()->strides_in_bytes().y();
93
94 Iterator input(_input, window);
95 Iterator output(_output, window);
96
97 execute_window_loop(window, [&](const Coordinates & id)
98 {
99 *reinterpret_cast<T *>(output.ptr() + id.x() * dst_stride_x + (id.y() % _factor1 * _factor2 + id.y() / _factor1) * dst_stride_y) = *reinterpret_cast<T *>(input.ptr());
100 },
101 input);
102 }
103
run(const Window & window,const ThreadInfo & info)104 void NEConvertFullyConnectedWeightsKernel::run(const Window &window, const ThreadInfo &info)
105 {
106 ARM_COMPUTE_UNUSED(info);
107 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
108 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
109
110 switch(_input->info()->element_size())
111 {
112 case 1:
113 run_convert_fc_weights<uint8_t>(window);
114 break;
115 case 2:
116 run_convert_fc_weights<uint16_t>(window);
117 break;
118 case 4:
119 run_convert_fc_weights<uint32_t>(window);
120 break;
121 default:
122 ARM_COMPUTE_ERROR("Data type not supported.");
123 break;
124 }
125 }
126 } // namespace arm_compute
127