1 /*
2 * Copyright (c) 2019-2020 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/CLArgMinMaxLayerKernel.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/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "src/core/AccessWindowStatic.h"
34 #include "src/core/CL/CLValidate.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 #include "support/StringSupport.h"
39
40 namespace arm_compute
41 {
42 namespace
43 {
44 constexpr unsigned int vector_size = 16;
45
validate_arguments(const ITensorInfo * input,const ITensorInfo * prev_output,const ITensorInfo * output,unsigned int axis,ReductionOperation op)46 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *prev_output, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
47 {
48 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
49 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
50 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::F16, DataType::F32);
51 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Only ARG_IDX_MAX and ARG_IDX_MIN are supported");
52 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
54
55 if(output->total_size() != 0)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32);
58 }
59 if(prev_output != nullptr && prev_output->total_size() != 0)
60 {
61 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(prev_output, 1, DataType::U32, DataType::S32);
62 if(output->total_size() != 0)
63 {
64 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(prev_output, output);
65 }
66 }
67
68 return Status{};
69 }
70
validate_and_configure_window(ITensorInfo * input,ITensorInfo * prev_output,ITensorInfo * output,unsigned int axis,ReductionOperation op)71 std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *prev_output, ITensorInfo *output, unsigned int axis, ReductionOperation op)
72 {
73 ARM_COMPUTE_UNUSED(op);
74 // Output tensor auto initialization if not yet initialized
75 TensorShape output_shape{ input->tensor_shape() };
76 output_shape.set(axis, 1);
77 DataType output_data_type = DataType::S32;
78 auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
79
80 Window win = calculate_max_window((prev_output != nullptr) ? (*prev_output) : (*input), Steps(vector_size));
81 bool window_changed = false;
82
83 switch(axis)
84 {
85 case 0:
86 {
87 ITensorInfo *input_tensor_access = prev_output != nullptr ? prev_output : input;
88 AccessWindowStatic input_access(input_tensor_access, 0, 0, static_cast<int>(input_tensor_access->dimension(0)), 1);
89 AccessWindowHorizontal output_access(output, 0, 1);
90 window_changed = update_window_and_padding(win, input_access, output_access);
91 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
92 }
93 break;
94 case 1:
95 case 2:
96 case 3:
97 {
98 AccessWindowHorizontal input_access(input, 0, vector_size);
99 AccessWindowHorizontal output_access(output, 0, vector_size);
100 window_changed = update_window_and_padding(win, input_access, output_access);
101 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
102 }
103 break;
104 default:
105 ARM_COMPUTE_ERROR("Not supported");
106 }
107
108 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
109 return std::make_tuple(err, win);
110 }
111 } // namespace
112
CLArgMinMaxLayerKernel()113 CLArgMinMaxLayerKernel::CLArgMinMaxLayerKernel()
114 : _input(nullptr), _prev_output(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::ARG_IDX_MAX)
115 {
116 }
117
configure(const ICLTensor * input,const ICLTensor * prev_output,ICLTensor * output,unsigned int axis,ReductionOperation op)118 void CLArgMinMaxLayerKernel::configure(const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op)
119 {
120 configure(CLKernelLibrary::get().get_compile_context(), input, prev_output, output, axis, op);
121 }
122
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * prev_output,ICLTensor * output,unsigned int axis,ReductionOperation op)123 void CLArgMinMaxLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *prev_output, ICLTensor *output, unsigned int axis, ReductionOperation op)
124 {
125 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
126 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr, output->info(), axis, op));
127 auto win_config = validate_and_configure_window(input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr, output->info(), axis, op);
128 ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
129
130 _input = input;
131 _prev_output = prev_output;
132 _output = output;
133 _reduction_axis = axis;
134 _op = op;
135
136 // Set build options
137 CLBuildOptions build_opts;
138
139 build_opts.add_option_if(_prev_output != nullptr, "-DPREV_OUTPUT");
140 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
141 build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DFLOAT_DATA_TYPE");
142 build_opts.add_option_if_else(op == ReductionOperation::ARG_IDX_MAX, "-DARG_MAX", "-DARG_MIN");
143 build_opts.add_option("-DDATA_TYPE_OUTPUT=" + get_cl_type_from_data_type(output->info()->data_type()));
144 build_opts.add_option("-DDATA_TYPE_SELECT=" + get_cl_signed_type_from_element_size(input->info()->element_size()));
145
146 // Create kernel
147 cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange();
148 std::string kernel_axis_name;
149 switch(axis)
150 {
151 case 0:
152 {
153 const ICLTensor *input_for_width = prev_output != nullptr ? _prev_output : _input;
154 build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input_for_width->info()->dimension(0)));
155
156 kernel_axis_name = "x";
157 lws_hint = create_lws_hint_parallel_implementations(input_for_width->info()->dimension(0), vector_size);
158 }
159 break;
160 case 1:
161 build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
162 kernel_axis_name = "y";
163 break;
164 case 2:
165 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
166 kernel_axis_name = "z";
167 break;
168 case 3:
169 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
170 build_opts.add_option("-DBATCH=" + support::cpp11::to_string(input->info()->dimension(3)));
171 kernel_axis_name = "w";
172 break;
173 default:
174 ARM_COMPUTE_ERROR("Not supported");
175 }
176 _kernel = create_kernel(compile_context, "arg_min_max_" + kernel_axis_name, build_opts.options());
177
178 // Configure kernel window
179 ICLKernel::configure_internal(std::get<1>(win_config), lws_hint);
180 }
181
validate(const ITensorInfo * input,const ITensorInfo * prev_output,const ITensorInfo * output,unsigned int axis,ReductionOperation op)182 Status CLArgMinMaxLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *prev_output, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
183 {
184 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, prev_output, output, axis, op));
185 ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), (prev_output != nullptr) ? prev_output->clone().get() : nullptr, output->clone().get(), axis, op)));
186 return Status{};
187 }
188
run(const Window & window,cl::CommandQueue & queue)189 void CLArgMinMaxLayerKernel::run(const Window &window, cl::CommandQueue &queue)
190 {
191 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
192 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
193
194 switch(_reduction_axis)
195 {
196 case 0:
197 {
198 // Set out window
199 Window out_window(window);
200 out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
201
202 // Get first input and output slices
203 Window in_slice = window.first_slice_window_2D();
204 Window out_slice = out_window.first_slice_window_2D();
205
206 // Reshape window
207 const unsigned int num_tensors = _prev_output != nullptr ? 3 : 2;
208
209 // Set local sums buffer
210 unsigned int local_res_size = lws_hint()[0] * _output->info()->element_size();
211 _kernel.setArg(num_arguments_per_2D_tensor() * num_tensors, local_res_size, nullptr);
212 do
213 {
214 unsigned int idx = 0;
215 add_2D_tensor_argument(idx, _input, in_slice);
216 if(_prev_output != nullptr)
217 {
218 add_2D_tensor_argument(idx, _prev_output, in_slice);
219 }
220 add_2D_tensor_argument(idx, _output, out_slice);
221 enqueue(queue, *this, in_slice, lws_hint());
222 }
223 while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
224 }
225 break;
226 case 1:
227 {
228 // Get first input and output slices
229 Window window_in{ window };
230 window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), _input->info()->dimension(1)));
231 Window in_slice = window_in.first_slice_window_2D();
232 Window out_slice = window.first_slice_window_2D();
233
234 do
235 {
236 unsigned int idx = 0;
237 add_2D_tensor_argument(idx, _input, in_slice);
238 add_2D_tensor_argument(idx, _output, out_slice);
239 enqueue(queue, *this, in_slice, lws_hint());
240 }
241 while(window_in.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
242 }
243 break;
244 case 2:
245 {
246 // Get first input and output slices
247 Window window_in{ window };
248 window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), _input->info()->dimension(2)));
249 Window in_slice = window_in.first_slice_window_3D();
250 Window out_slice = window.first_slice_window_3D();
251
252 do
253 {
254 unsigned int idx = 0;
255 add_3D_tensor_argument(idx, _input, in_slice);
256 add_3D_tensor_argument(idx, _output, out_slice);
257 enqueue(queue, *this, in_slice, lws_hint());
258 }
259 while(window_in.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(out_slice));
260 }
261 break;
262 case 3:
263 {
264 // Get first input and output slices
265 Window window_in{ window };
266 window_in.set(3, Window::Dimension(0, 1, 1));
267 Window in_slice = window_in.first_slice_window_4D();
268 Window out_slice = window.first_slice_window_4D();
269
270 do
271 {
272 unsigned int idx = 0;
273 add_4D_tensor_argument(idx, _input, in_slice);
274 add_4D_tensor_argument(idx, _output, out_slice);
275 enqueue(queue, *this, in_slice, lws_hint());
276 }
277 while(window_in.slide_window_slice_4D(in_slice) && window.slide_window_slice_4D(out_slice));
278 }
279 break;
280 default:
281 ARM_COMPUTE_ERROR("Not supported");
282 }
283 }
284 } // namespace arm_compute
285