1 /*
2 * Copyright (c) 2017-2020 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/NEWeightsReshapeKernel.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Validate.h"
28 #include "src/core/helpers/AutoConfiguration.h"
29 #include "src/core/helpers/WindowHelpers.h"
30
31 namespace arm_compute
32 {
33 namespace
34 {
get_output_shape(const ITensorInfo * input,bool has_bias)35 TensorShape get_output_shape(const ITensorInfo *input, bool has_bias)
36 {
37 TensorShape output_shape{ input->tensor_shape() };
38
39 output_shape.collapse(3);
40 const size_t tmp_dim = output_shape[0];
41 output_shape.set(0, output_shape[1]);
42 output_shape.set(1, tmp_dim + (has_bias ? 1 : 0));
43
44 return output_shape;
45 }
46
validate_arguments(const ITensorInfo * input,const ITensorInfo * biases,const ITensorInfo * output)47 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output)
48 {
49 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
50 //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
51 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
52
53 if(biases != nullptr)
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()));
56 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
57 ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->num_dimensions() != 1));
58 ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->num_dimensions() != 2));
59 ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 4) && (biases->dimension(0) != input->tensor_shape()[3]));
60 ARM_COMPUTE_RETURN_ERROR_ON((input->num_dimensions() == 5) && (biases->dimension(0) != input->tensor_shape()[3] || biases->dimension(1) != input->tensor_shape()[4]));
61 }
62
63 // Checks performed when output is configured
64 if(output->total_size() != 0)
65 {
66 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, biases != nullptr));
67 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
68 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
69 }
70
71 return Status{};
72 }
73
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)74 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
75 {
76 Window window = calculate_max_window(*input, Steps());
77 window.set(Window::DimX, Window::Dimension(0, input->dimension(0), input->dimension(0)));
78 window.set(Window::DimY, Window::Dimension(0, input->dimension(1), input->dimension(1)));
79 window.set(Window::DimZ, Window::Dimension(0, input->dimension(2), input->dimension(2)));
80
81 // The NEConvolutionLayerWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
82 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
83
84 return std::make_pair(Status{}, window);
85 }
86 } // namespace
87
NEWeightsReshapeKernel()88 NEWeightsReshapeKernel::NEWeightsReshapeKernel()
89 : _input(nullptr), _bias(nullptr), _output(nullptr)
90 {
91 }
92
configure(const ITensor * input,const ITensor * bias,ITensor * output)93 void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output)
94 {
95 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
96
97 // Output tensor auto inizialitation if not yet initialized
98 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(get_output_shape(input->info(), (bias != nullptr))));
99
100 // Perform validation step
101 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(),
102 (bias != nullptr) ? bias->info() : nullptr,
103 output->info()));
104
105 _input = input;
106 _bias = bias;
107 _output = output;
108
109 // Configure kernel
110 auto win_config = validate_and_configure_window(input->info(), output->info());
111 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
112 INEKernel::configure(win_config.second);
113 }
114
validate(const ITensorInfo * input,const ITensorInfo * biases,const ITensorInfo * output)115 Status NEWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output)
116 {
117 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, biases, output));
118 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
119
120 return Status{};
121 }
122
run(const Window & window,const ThreadInfo & info)123 void NEWeightsReshapeKernel::run(const Window &window, const ThreadInfo &info)
124 {
125 ARM_COMPUTE_UNUSED(info);
126 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
127 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
128
129 const unsigned int kernel_size_x = _input->info()->dimension(0);
130 const unsigned int kernel_size_y = _input->info()->dimension(1);
131 const unsigned int kernel_depth = _input->info()->dimension(2);
132 const unsigned int input_stride_x = _input->info()->strides_in_bytes().x();
133 const unsigned int input_stride_y = _input->info()->strides_in_bytes().y();
134 const unsigned int input_stride_z = _input->info()->strides_in_bytes().z();
135 const unsigned int output_stride_y = _output->info()->strides_in_bytes().y();
136
137 // Create iterators
138 Iterator in(_input, window);
139 execute_window_loop(window, [&](const Coordinates & id)
140 {
141 // Get column index
142 const int kernel_idx = id[3];
143 const int kernel_idz = id[4];
144
145 // Setup pointers
146 const uint8_t *tmp_input_ptr = in.ptr();
147 uint8_t *tmp_output_ptr = _output->ptr_to_element(Coordinates(kernel_idx, 0, kernel_idz));
148 const uint8_t *curr_input_row_ptr = tmp_input_ptr;
149 const uint8_t *curr_input_depth_ptr = tmp_input_ptr;
150
151 // Linearize volume
152 for(unsigned int d = 0; d < kernel_depth; ++d)
153 {
154 for(unsigned int j = 0; j < kernel_size_y; ++j)
155 {
156 for(unsigned int i = 0; i < kernel_size_x; ++i)
157 {
158 std::memcpy(tmp_output_ptr, tmp_input_ptr, _input->info()->element_size());
159 tmp_input_ptr += input_stride_x;
160 tmp_output_ptr += output_stride_y;
161 }
162 curr_input_row_ptr += input_stride_y;
163 tmp_input_ptr = curr_input_row_ptr;
164 }
165 curr_input_depth_ptr += input_stride_z;
166 curr_input_row_ptr = curr_input_depth_ptr;
167 tmp_input_ptr = curr_input_depth_ptr;
168 }
169
170 // Add bias
171 if(_bias != nullptr)
172 {
173 std::memcpy(tmp_output_ptr, _bias->ptr_to_element(Coordinates(kernel_idx, kernel_idz)), _input->info()->element_size());
174 }
175 },
176 in);
177 }
178 } // namespace arm_compute
179