1 /*
2 * Copyright (c) 2017-2021 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/CLReductionOperationKernel.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 "arm_compute/core/utils/misc/ShapeCalculator.h"
34 #include "src/core/AccessWindowStatic.h"
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38
39 #include "support/StringSupport.h"
40
41 namespace arm_compute
42 {
43 namespace
44 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,unsigned int axis,ReductionOperation op)45 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
46 {
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
48 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
49 if(input->num_channels() == 1)
50 {
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::F16, DataType::F32);
52 }
53 else
54 {
55 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 2, DataType::F16, DataType::F32);
56 ARM_COMPUTE_RETURN_ERROR_ON(axis == 0);
57 }
58 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op == ReductionOperation::SUM_SQUARE && input->data_type() == DataType::QASYMM8, "Not supported reduction operation for QASYMM8");
59 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
60 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
61 ARM_COMPUTE_RETURN_ERROR_ON((op == ReductionOperation::MEAN_SUM) && (axis == 0) && (input->dimension(0) == 0) && (input->data_type() != DataType::QASYMM8)
62 && (input->data_type() != DataType::QASYMM8_SIGNED));
63 ARM_COMPUTE_RETURN_ERROR_ON_MSG((op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN), "Not supported reduction operation, use CLArgMinMaxLayer");
64
65 if(output->total_size() != 0)
66 {
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 } // namespace
74
CLReductionOperationKernel()75 CLReductionOperationKernel::CLReductionOperationKernel()
76 : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::SUM_SQUARE)
77 {
78 _type = CLKernelType::ELEMENTWISE;
79 }
80
configure(const ICLTensor * input,ICLTensor * output,unsigned int axis,ReductionOperation op)81 void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op)
82 {
83 configure(CLKernelLibrary::get().get_compile_context(), input, output, axis, op);
84 }
85
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,unsigned int axis,ReductionOperation op)86 void CLReductionOperationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op)
87 {
88 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
89
90 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis, op));
91
92 auto padding_info = get_padding_info({ input, output });
93
94 _input = input;
95 _output = output;
96 _reduction_axis = axis;
97 _op = op;
98
99 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, true);
100 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).reset_padding().set_is_resizable(true));
101
102 // Set build options
103 CLBuildOptions build_opts;
104 DataType data_type = input->info()->data_type();
105 std::string data_type_promoted{};
106
107 if(is_data_type_quantized(data_type))
108 {
109 data_type_promoted = "int";
110 }
111 else
112 {
113 data_type_promoted = get_cl_type_from_data_type(data_type);
114 }
115
116 const unsigned int width = input->info()->dimension(0) * input->info()->num_channels();
117 unsigned int vec_size = (is_data_type_quantized(input->info()->data_type()) && (axis == 0)) ? 1 : 16;
118 vec_size = adjust_vec_size(vec_size, width);
119 const unsigned int vec_size_leftover = width % vec_size;
120
121 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
122 build_opts.add_option("-DDATA_TYPE_PROMOTED=" + data_type_promoted);
123 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size));
124 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover));
125 build_opts.add_option_if(is_data_type_float(data_type), "-DFLOAT_DATA_TYPE");
126 build_opts.add_option_if(op == ReductionOperation::SUM_SQUARE, "-DSUM_SQUARE");
127 build_opts.add_option_if(op == ReductionOperation::MEAN_SUM, "-DMEAN");
128 build_opts.add_option_if(op == ReductionOperation::SUM, "-DSUM");
129 build_opts.add_option_if(op == ReductionOperation::PROD, "-DPROD");
130 build_opts.add_option_if(op == ReductionOperation::MIN, "-DMIN");
131 build_opts.add_option_if(op == ReductionOperation::MAX, "-DMAX");
132 build_opts.add_option_if(is_data_type_quantized(data_type), "-DOFFSET=" + support::cpp11::to_string(input->info()->quantization_info().uniform().offset));
133 build_opts.add_option_if(is_data_type_quantized(data_type), "-DSCALE=" + float_to_string_with_full_precision(input->info()->quantization_info().uniform().scale));
134
135 switch(op)
136 {
137 case ReductionOperation::SUM_SQUARE:
138 build_opts.add_option(("-DOPERATION=square_sum"));
139 break;
140 case ReductionOperation::SUM:
141 case ReductionOperation::MEAN_SUM:
142 build_opts.add_option(("-DOPERATION=sum"));
143 break;
144 case ReductionOperation::MIN:
145 case ReductionOperation::MAX:
146 break;
147 case ReductionOperation::PROD:
148 build_opts.add_option(("-DOPERATION=product"));
149 break;
150 default:
151 ARM_COMPUTE_ERROR("Unsupported reduction operation");
152 }
153
154 // Create kernel
155 std::string kernel_axis_name;
156 const bool is_serial_op = needs_serialized_reduction(_op, _input->info()->data_type(), _reduction_axis);
157
158 switch(axis)
159 {
160 case 0:
161 {
162 build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(width));
163 kernel_axis_name = ((is_serial_op) ? "non_parallel_x" : "x");
164 }
165 break;
166 case 1:
167 build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
168 kernel_axis_name = "y";
169 break;
170 case 2:
171 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
172 kernel_axis_name = "z";
173 break;
174 case 3:
175 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
176 build_opts.add_option("-DBATCH=" + support::cpp11::to_string(input->info()->dimension(3)));
177 kernel_axis_name = "w";
178 break;
179 default:
180 ARM_COMPUTE_ERROR("Not supported");
181 }
182 _kernel = create_kernel(compile_context, "reduction_operation_" + kernel_axis_name, build_opts.options());
183
184 // Configure kernel window
185 Window win = calculate_max_window(*input->info(), Steps(vec_size));
186 win.set(Window::DimX, Window::Dimension(win.x().start(), win.x().end() * _input->info()->num_channels(), win.x().step()));
187 ICLKernel::configure_internal(win);
188
189 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
190 }
191
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int axis,ReductionOperation op)192 Status CLReductionOperationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op)
193 {
194 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op));
195 return Status{};
196 }
197
run(const Window & window,cl::CommandQueue & queue)198 void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &queue)
199 {
200 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
201 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
202
203 const bool is_serial_op = needs_serialized_reduction(_op, _input->info()->data_type(), _reduction_axis);
204 switch(_reduction_axis)
205 {
206 case 0:
207 {
208 // We use parallel reduction only in non quantized types
209 if(is_serial_op)
210 {
211 // Get first input and output slices
212 Window window_in{ window };
213 window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
214
215 Window out_window{ window };
216 out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
217
218 Window in_slice = window_in.first_slice_window_1D();
219 Window out_slice = out_window.first_slice_window_1D();
220
221 do
222 {
223 unsigned int idx = 0;
224 add_1D_tensor_argument(idx, _input, in_slice);
225 add_1D_tensor_argument(idx, _output, out_slice);
226 enqueue(queue, *this, in_slice);
227 }
228 while(window_in.slide_window_slice_1D(in_slice) && out_window.slide_window_slice_1D(out_slice));
229 }
230 else
231 {
232 // Set out window
233 bool has_collapsed = true;
234 Window window_in = window.collapse_if_possible(window, 2, &has_collapsed);
235 ARM_COMPUTE_ERROR_ON(!has_collapsed);
236
237 Window window_out = window_in;
238 window_out.set(0, Window::Dimension());
239
240 unsigned int idx = 0;
241 add_3D_tensor_argument(idx, _input, window_in);
242 add_3D_tensor_argument(idx, _output, window_out);
243 enqueue(queue, *this, window_in);
244 }
245 }
246 break;
247 case 1:
248 {
249 // Get first input and output slices
250 Window window_in{ window };
251 window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), _input->info()->dimension(1)));
252 Window in_slice = window_in.first_slice_window_2D();
253 Window out_slice = window.first_slice_window_2D();
254
255 do
256 {
257 unsigned int idx = 0;
258 add_2D_tensor_argument(idx, _input, in_slice);
259 add_2D_tensor_argument(idx, _output, out_slice);
260 enqueue(queue, *this, in_slice);
261 }
262 while(window_in.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
263 }
264 break;
265 case 2:
266 {
267 // Get first input and output slices
268 Window window_in{ window };
269 window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), _input->info()->dimension(2)));
270 Window in_slice = window_in.first_slice_window_3D();
271 Window out_slice = window.first_slice_window_3D();
272
273 do
274 {
275 unsigned int idx = 0;
276 add_3D_tensor_argument(idx, _input, in_slice);
277 add_3D_tensor_argument(idx, _output, out_slice);
278 enqueue(queue, *this, in_slice);
279 }
280 while(window_in.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(out_slice));
281 }
282 break;
283 case 3:
284 {
285 // Get first input and output slices
286 Window window_in{ window };
287 window_in.set(3, Window::Dimension(0, 1, 1));
288 Window in_slice = window_in.first_slice_window_4D();
289 Window out_slice = window.first_slice_window_4D();
290
291 do
292 {
293 unsigned int idx = 0;
294 add_4D_tensor_argument(idx, _input, in_slice);
295 add_4D_tensor_argument(idx, _output, out_slice);
296 enqueue(queue, *this, in_slice);
297 }
298 while(window_in.slide_window_slice_4D(in_slice) && window.slide_window_slice_4D(out_slice));
299 }
300 break;
301 default:
302 ARM_COMPUTE_ERROR("Not supported");
303 }
304 }
305 } // namespace arm_compute
306