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
|
/*
* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright (c) 2016-2018 ARM Limited.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "arm_compute/core/CL/kernels/CLPadLayerKernel.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibraryEx.h"
#include "arm_compute/core/CL/ICLTensor.h"
using namespace arm_compute;
namespace
{
Status validate_arguments(const ITensorInfo *input_info, const ITensorInfo *output_info,
const ITensorInfo *pad_size_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_info, 1, DataType::U8, DataType::QASYMM8,
DataType::S16, DataType::S32, DataType::F16,
DataType::F32);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_info, 1, DataType::U8, DataType::QASYMM8,
DataType::S16, DataType::S32, DataType::F16,
DataType::F32);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(pad_size_info, 1, DataType::S32);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_info->num_dimensions() > 0 &&
input_info->num_dimensions() <= 4,
"Pad kernel supports upto 4-D input tensor");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(
input_info->num_dimensions() == output_info->num_dimensions(),
"output tensor should have same number of dimensions as input tensor");
if (input_info->data_type() == DataType::QASYMM8)
{
ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_info->quantization_info() !=
output_info->quantization_info(),
"The input and output quantization info are different!");
}
return Status{};
}
} // namespace
CLPadLayerKernel::CLPadLayerKernel() : _input(nullptr), _output(nullptr), _pad_size(nullptr) {}
void CLPadLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ICLTensor *pad_size)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, pad_size);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pad_size->info()));
_input = input;
_output = output;
_pad_size = pad_size;
// Set kernel build options
std::set<std::string> build_opts;
build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
build_opts.emplace("-DDEPTH_OUT=" + support::cpp11::to_string(output->info()->dimension(2)));
build_opts.emplace("-DIB=" + support::cpp11::to_string(input->info()->dimension(3)));
build_opts.emplace("-DIW=" + support::cpp11::to_string(input->info()->dimension(0)));
build_opts.emplace("-DIH=" + support::cpp11::to_string(input->info()->dimension(1)));
build_opts.emplace("-DID=" + support::cpp11::to_string(input->info()->dimension(2)));
if (input->info()->data_type() == DataType::QASYMM8)
{
build_opts.emplace("-DZERO_VALUE=" +
support::cpp11::to_string(input->info()->quantization_info().offset));
}
else
{
build_opts.emplace("-DZERO_VALUE=" + support::cpp11::to_string(0));
}
// Create kernel
_kernel = static_cast<cl::Kernel>(CLKernelLibraryEx::get().create_kernel("pad", build_opts));
// Configure kernel window
Window win = calculate_max_window(*output->info(), Steps());
Coordinates coord;
coord.set_num_dimensions(output->info()->num_dimensions());
output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
ICLKernel::configure_internal(win);
}
void CLPadLayerKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
_pad_size->map(queue);
// Padding values only for up, top, left and front are required based on the rank of tensor
int rank = _pad_size->info()->dimension(1);
auto pad_batch_up =
(rank == 4) ? *reinterpret_cast<const int32_t *>(_pad_size->ptr_to_element({0, 0})) : 0;
auto pad_height_top =
(rank >= 2)
? *reinterpret_cast<const int32_t *>(_pad_size->ptr_to_element({0, (rank == 2) ? 0 : 1}))
: 0;
auto pad_width_left = (rank >= 1)
? *reinterpret_cast<const int32_t *>(
_pad_size->ptr_to_element({0, (rank == 4) ? 2 : rank - 1}))
: 0;
auto pad_depth_front =
(rank >= 3)
? *reinterpret_cast<const int32_t *>(_pad_size->ptr_to_element({0, (rank == 3) ? 0 : 3}))
: 0;
_pad_size->unmap(queue);
// Pad_values which needs to be passed
const cl_int4 paddingValues = {
{static_cast<cl_int>(pad_width_left), static_cast<cl_int>(pad_height_top),
static_cast<cl_int>(pad_depth_front), static_cast<cl_int>(pad_batch_up)}};
Window slice_out = window.first_slice_window_4D().collapse(ICLKernel::window(), 2, 4);
// Setup output slice
Window slice_in(slice_out);
slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
slice_in.set(3, Window::Dimension(0, 0, 0));
do
{
unsigned int idx = 0;
add_4D_tensor_argument(idx, _input, slice_in);
add_4D_tensor_argument(idx, _output, slice_out);
_kernel.setArg<cl_int4>(idx++, paddingValues);
enqueue(queue, *this, slice_out);
} while (window.slide_window_slice_4D(slice_out) && window.slide_window_slice_4D(slice_in));
}
|
