blob: ef46325e4d07857e7069a5141c5d0ed09fc129cf [file] [log] [blame]
Michalis Spyrou04f089c2017-08-08 17:42:38 +01001/*
Michalis Spyrouf6402dd2018-01-26 15:06:19 +00002 * Copyright (c) 2017-2018 ARM Limited.
Michalis Spyrou04f089c2017-08-08 17:42:38 +01003 *
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 "arm_compute/core/CL/kernels/CLReductionOperationKernel.h"
25
26#include "arm_compute/core/AccessWindowStatic.h"
27#include "arm_compute/core/CL/CLHelpers.h"
28#include "arm_compute/core/CL/CLKernelLibrary.h"
29#include "arm_compute/core/CL/ICLTensor.h"
Michalis Spyrou04f089c2017-08-08 17:42:38 +010030#include "arm_compute/core/Helpers.h"
31#include "arm_compute/core/TensorInfo.h"
32#include "arm_compute/core/Utils.h"
33#include "arm_compute/core/Validate.h"
34#include "arm_compute/core/Window.h"
35
36#include "support/ToolchainSupport.h"
37
38using namespace arm_compute;
39
John Richardson62385bc2018-04-20 13:11:36 +010040namespace
41{
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010042// OpenCL kernel requires input width to be a power of 2 for x-axis.
Michalis Spyrou25747e22018-08-08 17:12:38 +010043constexpr unsigned int border_val = 64;
44
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010045Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, unsigned int width)
John Richardson62385bc2018-04-20 13:11:36 +010046{
John Richardson62385bc2018-04-20 13:11:36 +010047 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010048 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
Michalis Spyrou8aaf93e2018-10-11 17:33:32 +010049 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op == ReductionOperation::SUM_SQUARE && input->data_type() == DataType::QASYMM8, "Not supported reduction operation for QASYMM8");
John Richardson62385bc2018-04-20 13:11:36 +010050 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010051 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
52 ARM_COMPUTE_RETURN_ERROR_ON(op == ReductionOperation::MEAN_SUM && axis == 0 && width == 0 && input->data_type() != DataType::QASYMM8);
John Richardson62385bc2018-04-20 13:11:36 +010053
54 if(output->total_size() != 0)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010057 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
John Richardson62385bc2018-04-20 13:11:36 +010058 }
59
60 return Status{};
61}
62
63std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, unsigned int axis)
64{
65 // Output tensor auto initialization if not yet initialized
66 TensorShape output_shape{ input->tensor_shape() };
67 output_shape.set(axis, 1);
Vidhya Sudhan Loganathan7485d5a2018-07-04 09:34:00 +010068 auto_init_if_empty(*output, output_shape, 1, input->data_type());
John Richardson62385bc2018-04-20 13:11:36 +010069
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010070 const unsigned int num_elems_processed_per_iteration = (is_data_type_quantized(input->data_type()) && (axis == 0)) ? 1 : 16;
71 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
72 bool window_changed = false;
John Richardson62385bc2018-04-20 13:11:36 +010073
Michalis Spyrou7e9391b2018-10-05 14:49:28 +010074 switch(axis)
75 {
76 case 0:
77 {
78 if(is_data_type_quantized(input->data_type()))
79 {
80 AccessWindowHorizontal input_access(input, 0, input->dimension(0));
81 AccessWindowHorizontal output_access(output, 0, 1);
82 window_changed = update_window_and_padding(win, input_access, output_access);
83 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
84 }
85 else
86 {
87 const unsigned int border_width = ((input->dimension(0) % border_val) != 0) ? border_val - input->dimension(0) % border_val : 0;
88 AccessWindowStatic input_access(input, 0, 0, input->dimension(0) + border_width, 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 }
94 break;
95 case 1:
96 case 2:
97 case 3:
98 {
99 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
100 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
101 window_changed = update_window_and_padding(win, input_access, output_access);
102 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
103 }
104 break;
105 default:
106 ARM_COMPUTE_ERROR("Not supported");
107 }
John Richardson62385bc2018-04-20 13:11:36 +0100108
109 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
110
111 return std::make_tuple(err, win);
112}
113} // namespace
114
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100115CLReductionOperationKernel::CLReductionOperationKernel()
116 : _input(nullptr), _output(nullptr), _reduction_axis(0), _op(ReductionOperation::SUM_SQUARE), _border_size()
117{
118}
119
120BorderSize CLReductionOperationKernel::border_size() const
121{
122 return _border_size;
123}
124
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100125void CLReductionOperationKernel::configure(const ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, unsigned int width)
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100126{
John Richardson62385bc2018-04-20 13:11:36 +0100127 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100128
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100129 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis, op, width));
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100130
131 _input = input;
132 _output = output;
133 _reduction_axis = axis;
134 _op = op;
Michalis Spyrou343722b2018-06-05 13:04:40 +0100135
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100136 // Set build options
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100137 CLBuildOptions build_opts;
138 std::string data_type_promoted = get_cl_type_from_data_type(input->info()->data_type());
139 if(is_data_type_quantized(input->info()->data_type()) && axis != 0)
140 {
141 data_type_promoted = "uint";
142 }
143 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
144 build_opts.add_option("-DDATA_TYPE_PROMOTED=" + data_type_promoted);
Michalis Spyrou8aaf93e2018-10-11 17:33:32 +0100145 build_opts.add_option_if(op == ReductionOperation::SUM_SQUARE, "-DSUM_SQUARE=");
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100146 build_opts.add_option_if(op == ReductionOperation::MEAN_SUM, "-DMEAN");
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100147
148 switch(op)
149 {
150 case ReductionOperation::SUM_SQUARE:
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100151 build_opts.add_option(("-DOPERATION=square_sum"));
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100152 break;
153 case ReductionOperation::SUM:
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100154 case ReductionOperation::MEAN_SUM:
155 build_opts.add_option(("-DOPERATION=sum"));
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100156 break;
157 default:
158 ARM_COMPUTE_ERROR("Unsupported reduction operation");
159 }
160
161 // Create kernel
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100162 cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange();
163 std::string kernel_axis_name;
164 switch(axis)
165 {
166 case 0:
167 {
168 if(!is_data_type_quantized(input->info()->data_type()))
169 {
170 build_opts.add_option_if(op == ReductionOperation::MEAN_SUM, "-DWIDTH=" + support::cpp11::to_string(width));
171 const unsigned int width_leftover = input->info()->dimension(0) % border_val;
172 const unsigned int border_width = (width_leftover != 0) ? border_val - width_leftover : 0;
173 const unsigned int num_of_threads = ((input->info()->dimension(0) + border_width) / 16);
174 kernel_axis_name = "x";
175
176 // Set the number of WG based on the input size. If input width is < 128
177 // we can use fewer threads than 8.
178 lws_hint = cl::NDRange(std::min(8U, num_of_threads));
179 _border_size = BorderSize(0, border_width, 0, 0);
180 }
181 else
182 {
183 build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
184 kernel_axis_name = "quantized_x";
185 }
186 }
187 break;
188 case 1:
189 build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
190 kernel_axis_name = "y";
191 break;
192 case 2:
193 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
194 kernel_axis_name = "z";
195 break;
196 case 3:
197 build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
198 build_opts.add_option("-DBATCH=" + support::cpp11::to_string(input->info()->dimension(3)));
199 kernel_axis_name = "w";
200 break;
201 default:
202 ARM_COMPUTE_ERROR("Not supported");
203 }
204 _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reduction_operation_" + kernel_axis_name, build_opts.options()));
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100205
206 // Configure kernel window
John Richardson62385bc2018-04-20 13:11:36 +0100207 auto win_config = validate_and_configure_window(_input->info(), _output->info(), axis);
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100208
John Richardson62385bc2018-04-20 13:11:36 +0100209 ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100210
Anthony Barbierb6eb3532018-08-08 13:20:04 +0100211 ICLKernel::configure_internal(std::get<1>(win_config), lws_hint);
John Richardson62385bc2018-04-20 13:11:36 +0100212}
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100213
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100214Status CLReductionOperationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, unsigned int width)
John Richardson62385bc2018-04-20 13:11:36 +0100215{
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100216 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, op, width));
John Richardson62385bc2018-04-20 13:11:36 +0100217 ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), axis)));
218
219 return Status{};
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100220}
221
222void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &queue)
223{
224 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
225 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
226
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100227 switch(_reduction_axis)
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100228 {
Michalis Spyrou7e9391b2018-10-05 14:49:28 +0100229 case 0:
230 {
231 // We use parallel reduction only in non quantized types
232 if(!is_data_type_quantized(_input->info()->data_type()))
233 {
234 // Set out window
235 Window out_window(window);
236 out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
237
238 // Get first input and output slices
239 Window in_slice = window.first_slice_window_2D();
240 Window out_slice = out_window.first_slice_window_2D();
241
242 // Reshape window
243 const unsigned int border_width = ((in_slice.x().end() % border_val) != 0) ? border_val - in_slice.x().end() % border_val : 0;
244 in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start(), in_slice.x().end() + border_width, in_slice.x().step()));
245
246 // Set local sums buffer
247 unsigned int local_sum_size = lws_hint()[0] * _input->info()->element_size();
248 _kernel.setArg(num_arguments_per_2D_tensor() * 2, local_sum_size, nullptr);
249
250 do
251 {
252 unsigned int idx = 0;
253 add_2D_tensor_argument(idx, _input, in_slice);
254 add_2D_tensor_argument(idx, _output, out_slice);
255 enqueue(queue, *this, in_slice, lws_hint());
256 }
257 while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
258 }
259 else
260 {
261 // Get first input and output slices
262 Window window_in{ window };
263 window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));
264
265 Window in_slice = window.first_slice_window_1D();
266 Window out_slice = window.first_slice_window_1D();
267
268 do
269 {
270 unsigned int idx = 0;
271 add_1D_tensor_argument(idx, _input, in_slice);
272 add_1D_tensor_argument(idx, _output, out_slice);
273 enqueue(queue, *this, in_slice);
274 }
275 while(window_in.slide_window_slice_1D(in_slice) && window.slide_window_slice_1D(out_slice));
276 }
277 }
278 break;
279 case 1:
280 {
281 // Get first input and output slices
282 Window window_in{ window };
283 window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), _input->info()->dimension(1)));
284 Window in_slice = window_in.first_slice_window_2D();
285 Window out_slice = window.first_slice_window_2D();
286
287 do
288 {
289 unsigned int idx = 0;
290 add_2D_tensor_argument(idx, _input, in_slice);
291 add_2D_tensor_argument(idx, _output, out_slice);
292 enqueue(queue, *this, in_slice);
293 }
294 while(window_in.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
295 }
296 break;
297 case 2:
298 {
299 // Get first input and output slices
300 Window window_in{ window };
301 window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), _input->info()->dimension(2)));
302 Window in_slice = window_in.first_slice_window_3D();
303 Window out_slice = window.first_slice_window_3D();
304
305 do
306 {
307 unsigned int idx = 0;
308 add_3D_tensor_argument(idx, _input, in_slice);
309 add_3D_tensor_argument(idx, _output, out_slice);
310 enqueue(queue, *this, in_slice);
311 }
312 while(window_in.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(out_slice));
313 }
314 break;
315 case 3:
316 {
317 // Get first input and output slices
318 Window window_in{ window };
319 window_in.set(3, Window::Dimension(0, 1, 1));
320 Window in_slice = window_in.first_slice_window_4D();
321 Window out_slice = window.first_slice_window_4D();
322
323 do
324 {
325 unsigned int idx = 0;
326 add_4D_tensor_argument(idx, _input, in_slice);
327 add_4D_tensor_argument(idx, _output, out_slice);
328 enqueue(queue, *this, in_slice);
329 }
330 while(window_in.slide_window_slice_4D(in_slice) && window.slide_window_slice_4D(out_slice));
331 }
332 break;
333 default:
334 ARM_COMPUTE_ERROR("Not supported");
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100335 }
Michalis Spyrou04f089c2017-08-08 17:42:38 +0100336}