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/CL/kernels/CLMinMaxLayerKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "src/core/AccessWindowStatic.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 #include "support/StringSupport.h"
36
37 #include <climits>
38
39 using namespace arm_compute;
40 using namespace arm_compute::misc::shape_calculator;
41
42 namespace
43 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output)44 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
45 {
46 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
47 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
48 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() < 3);
49
50 if(output->tensor_shape().total_size() > 0)
51 {
52 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
53
54 TensorShape output_shape = compute_min_max_shape(input);
55
56 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
57 }
58
59 return Status{};
60 }
61
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)62 std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
63 {
64 TensorShape output_shape = compute_min_max_shape(input);
65
66 // Output auto initialization if not yet initialized
67 auto_init_if_empty(*output, output_shape, 1, input->data_type());
68
69 const unsigned int num_elems_processed_per_iteration = 1;
70
71 // Configure kernel window
72 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
73 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
74 AccessWindowStatic output_access(output, 0, 0, 2, output->dimension(1));
75
76 bool window_changed = update_window_and_padding(win, input_access, output_access);
77
78 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
79
80 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
81 return std::make_tuple(err, win);
82 }
83 } // namespace
84
CLMinMaxLayerKernel()85 CLMinMaxLayerKernel::CLMinMaxLayerKernel()
86 : _input(nullptr), _output(nullptr)
87 {
88 }
89
configure(const ICLTensor * input,ICLTensor * output)90 void CLMinMaxLayerKernel::configure(const ICLTensor *input, ICLTensor *output)
91 {
92 configure(CLKernelLibrary::get().get_compile_context(), input, output);
93 }
94
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output)95 void CLMinMaxLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output)
96 {
97 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
98 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
99
100 _input = input;
101 _output = output;
102
103 std::set<std::string> build_opts;
104 build_opts.emplace("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
105 build_opts.emplace("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
106 build_opts.emplace("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
107
108 // Create kernel
109 _kernel = create_kernel(compile_context, "minmax_layer", build_opts);
110
111 auto win_config = validate_and_configure_window(input->info(), output->info());
112
113 ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
114
115 ICLKernel::configure_internal(std::get<1>(win_config));
116 }
117
validate(const ITensorInfo * input,const ITensorInfo * output)118 Status CLMinMaxLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
119 {
120 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
121 ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get())));
122
123 return Status{};
124 }
125
reset(cl::CommandQueue & queue)126 void CLMinMaxLayerKernel::reset(cl::CommandQueue &queue)
127 {
128 _output->map(queue, true);
129
130 Window window_output;
131 window_output.use_tensor_dimensions(_output->info()->tensor_shape());
132 window_output.set(Window::DimX, Window::Dimension(0, 1, 1));
133
134 Iterator output(_output, window_output);
135
136 // Reset output
137 execute_window_loop(window_output, [&](const Coordinates &)
138 {
139 auto *ptr = reinterpret_cast<float *>(output.ptr());
140 ptr[0] = std::numeric_limits<float>::max();
141 ptr[1] = std::numeric_limits<float>::min();
142 },
143 output);
144
145 _output->unmap(queue);
146 }
147
run(const Window & window,cl::CommandQueue & queue)148 void CLMinMaxLayerKernel::run(const Window &window, cl::CommandQueue &queue)
149 {
150 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
151 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
152
153 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), 3);
154 Window slice = window_collapsed.first_slice_window_3D();
155 slice.set(Window::DimX, Window::Dimension(0, 1, 1));
156 slice.set(Window::DimY, Window::Dimension(0, 1, 1));
157 slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
158
159 do
160 {
161 Window output_slice = slice.shift_dimensions(2);
162
163 unsigned int idx = 0;
164 // Set inputs
165 add_3D_tensor_argument(idx, _input, slice);
166 add_1D_tensor_argument(idx, _output, output_slice);
167 enqueue(queue, *this, slice, lws_hint());
168 }
169 while(window_collapsed.slide_window_slice_3D(slice));
170 }
171