1 /*
2 * Copyright (c) 2016-2020 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/NEThresholdKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/Validate.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32
33 #include "src/core/NEON/wrapper/wrapper.h"
34
35 namespace arm_compute
36 {
37 namespace
38 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ThresholdKernelInfo & info)39 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ThresholdKernelInfo &info)
40 {
41 ARM_COMPUTE_UNUSED(info);
42 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
43 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
44
45 // Checks performed when output is configured
46 if((output != nullptr) && (output->total_size() != 0))
47 {
48 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
49 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
50 }
51
52 return Status{};
53 }
54
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)55 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
56 {
57 // Configure kernel window
58 Window win = calculate_max_window(*input, Steps());
59
60 // Output auto inizialitation if not yet initialized
61 auto_init_if_empty(*output, *input->clone());
62
63 // NEThresholdKernel doesn't need padding so update_window_and_padding() can be skipped
64 Coordinates coord;
65 coord.set_num_dimensions(output->num_dimensions());
66 output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
67
68 return std::make_pair(Status{}, win);
69 }
70 } // namespace
71
NEThresholdKernel()72 NEThresholdKernel::NEThresholdKernel()
73 : _func(nullptr), _input(nullptr), _output(nullptr), _info()
74 {
75 }
76
configure(const ITensor * input,ITensor * output,const ThresholdKernelInfo & info)77 void NEThresholdKernel::configure(const ITensor *input, ITensor *output, const ThresholdKernelInfo &info)
78 {
79 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
80 ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), info));
81
82 _input = input;
83 _output = output;
84 _info = info;
85
86 switch(_info.type)
87 {
88 case ThresholdType::BINARY:
89 _func = &NEThresholdKernel::run_binary;
90 break;
91 case ThresholdType::RANGE:
92 _func = &NEThresholdKernel::run_range;
93 break;
94 default:
95 ARM_COMPUTE_ERROR("Thresholding type not recognized");
96 break;
97 }
98
99 // Configure kernel window
100 auto win_config = validate_and_configure_window(input->info(), output->info());
101 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
102 ICPPKernel::configure(win_config.second);
103 }
104
validate(const ITensorInfo * input,const ITensorInfo * output,const ThresholdKernelInfo & info)105 Status NEThresholdKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ThresholdKernelInfo &info)
106 {
107 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, info));
108 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
109
110 return Status{};
111 }
112
run_binary(const Window & window)113 inline void NEThresholdKernel::run_binary(const Window &window)
114 {
115 /** NEON vector tag type. */
116 using Type = uint8_t;
117 using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<Type, wrapper::traits::BitWidth::W128>;
118
119 const int window_step_x = 16 / sizeof(Type);
120 const auto window_start_x = static_cast<int>(window.x().start());
121 const auto window_end_x = static_cast<int>(window.x().end());
122
123 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
124 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
125
126 const uint8_t threshold = _info.threshold;
127 const uint8_t true_value = _info.true_value;
128 const uint8_t false_value = _info.false_value;
129
130 const auto vthreshold = wrapper::vdup_n(threshold, ExactTagType{});
131 const auto vtrue_value = wrapper::vdup_n(true_value, ExactTagType{});
132 const auto vfalse_value = wrapper::vdup_n(false_value, ExactTagType{});
133
134 Iterator input(_input, win_collapsed);
135 Iterator output(_output, win_collapsed);
136
137 execute_window_loop(win_collapsed, [&](const Coordinates &)
138 {
139 const auto input_ptr = reinterpret_cast<const Type *>(input.ptr());
140 const auto output_ptr = reinterpret_cast<Type *>(output.ptr());
141
142 int x = window_start_x;
143 for(; x <= (window_end_x - window_step_x); x += window_step_x)
144 {
145 const auto vdata = wrapper::vloadq(input_ptr + x);
146 const auto vmask = wrapper::vcgt(vdata, vthreshold);
147 wrapper::vstore(output_ptr + x, wrapper::vbsl(vmask, vtrue_value, vfalse_value));
148 }
149
150 for(; x < window_end_x; ++x)
151 {
152 const Type data = *(reinterpret_cast<const Type *>(input_ptr + x));
153 *(output_ptr + x) = (data > threshold) ? true_value : false_value;
154 }
155 },
156 input, output);
157 }
158
run_range(const Window & window)159 inline void NEThresholdKernel::run_range(const Window &window)
160 {
161 /** NEON vector tag type. */
162 using Type = uint8_t;
163 using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<Type, wrapper::traits::BitWidth::W128>;
164
165 const int window_step_x = 16 / sizeof(Type);
166 const auto window_start_x = static_cast<int>(window.x().start());
167 const auto window_end_x = static_cast<int>(window.x().end());
168
169 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
170 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
171
172 const uint8_t lower_threshold = _info.threshold;
173 const uint8_t upper_threshold = _info.upper;
174 const uint8_t true_value = _info.true_value;
175 const uint8_t false_value = _info.false_value;
176
177 const auto vlower_threshold = wrapper::vdup_n(lower_threshold, ExactTagType{});
178 const auto vupper_threshold = wrapper::vdup_n(upper_threshold, ExactTagType{});
179 const auto vtrue_value = wrapper::vdup_n(true_value, ExactTagType{});
180 const auto vfalse_value = wrapper::vdup_n(false_value, ExactTagType{});
181
182 Iterator input(_input, win_collapsed);
183 Iterator output(_output, win_collapsed);
184
185 execute_window_loop(win_collapsed, [&](const Coordinates &)
186 {
187 const auto input_ptr = reinterpret_cast<const Type *>(input.ptr());
188 const auto output_ptr = reinterpret_cast<Type *>(output.ptr());
189
190 int x = window_start_x;
191 for(; x <= (window_end_x - window_step_x); x += window_step_x)
192 {
193 const auto vdata = wrapper::vloadq(input_ptr + x);
194 auto vmask = wrapper::vcle(vdata, vupper_threshold);
195 vmask = wrapper::vand(wrapper::vcge(vdata, vlower_threshold), vmask);
196 wrapper::vstore(output_ptr + x, wrapper::vbsl(vmask, vtrue_value, vfalse_value));
197 }
198
199 for(; x < window_end_x; ++x)
200 {
201 const Type data = *(reinterpret_cast<const Type *>(input_ptr + x));
202 *(output_ptr + x) = (data <= upper_threshold && data >= lower_threshold) ? true_value : false_value;
203 }
204 },
205 input, output);
206 }
207
run(const Window & window,const ThreadInfo & info)208 void NEThresholdKernel::run(const Window &window, const ThreadInfo &info)
209 {
210 ARM_COMPUTE_UNUSED(info);
211 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
212 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
213 ARM_COMPUTE_ERROR_ON(_func == nullptr);
214
215 (this->*_func)(window);
216 }
217 } // namespace arm_compute
218