1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
|
/*
* Copyright (c) 2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
#include "arm_compute/core/AccessWindowStatic.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/IAccessWindow.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "support/ToolchainSupport.h"
#include <cmath>
using namespace arm_compute;
using namespace arm_compute::misc::shape_calculator;
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims, const Size2D &num_tiles)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_dims.width != 3 || kernel_dims.height != 3, "Only 3x3 kernels are supported");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast<unsigned int>(std::sqrt(input->dimension(2))) != 4, "Only 2x2 output tile is supported");
ARM_COMPUTE_UNUSED(kernel_dims);
if(bias != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
}
// Checks performed when output is configured
if(output->total_size() != 0)
{
const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input, output_convolved_dims, DataLayout::NCHW));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
}
return Status{};
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
constexpr unsigned int num_elems_processed_per_iteration = 1;
Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
bool window_changed = false;
AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration);
AccessWindowStatic output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2));
if(bias != nullptr)
{
AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
window_changed = update_window_and_padding(win, input_access, bias_access, output_access);
}
else
{
window_changed = update_window_and_padding(win, input_access, output_access);
}
output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
return std::make_pair(err, win);
}
} // namespace
CLWinogradOutputTransformKernel::CLWinogradOutputTransformKernel()
: _input(nullptr), _bias(nullptr), _output(nullptr)
{
}
void CLWinogradOutputTransformKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
const Size2D &num_tiles)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_UNUSED(kernel_dims);
// Output tensor auto initialization if not yet initialized
auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_output_transform_shape(*input->info(), output_convolved_dims, DataLayout::NCHW)));
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info(), kernel_dims, output_convolved_dims, num_tiles));
_input = input;
_bias = bias;
_output = output;
// Set build options
CLBuildOptions build_opts;
build_opts.add_option_if(_bias != nullptr, std::string("-DHAS_BIAS"));
build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(num_tiles.width));
// Create kernel
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("winograd_output_transform_2x2_3x3_nchw", build_opts.options()));
// Configure kernel window
auto win_config = validate_and_configure_window(input->info(), (bias != nullptr ? bias->info() : nullptr), output->info());
ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure(win_config.second);
// Set config_id for enabling LWS tuning
_config_id = "winograd_output_transform_2x2_3x3";
_config_id += lower_string(string_from_data_type(input->info()->data_type()));
_config_id += "_";
_config_id += support::cpp11::to_string(input->info()->dimension(0));
_config_id += "_";
_config_id += support::cpp11::to_string(input->info()->dimension(1));
_config_id += "_";
_config_id += support::cpp11::to_string(output->info()->dimension(0));
_config_id += "_";
_config_id += support::cpp11::to_string(output->info()->dimension(1));
}
Status CLWinogradOutputTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const Size2D &kernel_dims, const Size2D &output_convolved_dims,
const Size2D &num_tiles)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, (bias != nullptr ? bias->clone().get() : nullptr), output, kernel_dims, output_convolved_dims, num_tiles));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get()).first);
return Status{};
}
void CLWinogradOutputTransformKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
// Get initial windows
Window slice = window.first_slice_window_3D();
slice.set(Window::DimZ, Window::Dimension(0, 1, 1));
// Setup output slice
Window slice_out(slice);
slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
if(_bias != nullptr)
{
unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
Window slice_biases;
slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
add_1D_tensor_argument(idx1, _bias, slice_biases);
}
do
{
unsigned int idx = 0;
add_3D_tensor_argument(idx, _input, slice);
add_3D_tensor_argument(idx, _output, slice_out);
enqueue(queue, *this, slice, _lws_hint);
}
while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_out));
}
|
