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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
/*
* 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/CLActivationLayerExKernel.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibraryEx.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/UtilsEx.h"
using namespace arm_compute;
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
const ActivationLayerInfoEx &act_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::QASYMM8,
DataType::F16, DataType::F32);
// Checks performed when output is configured
if ((output != nullptr) && (output->total_size() != 0))
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
}
return Status{};
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
{
if (output != nullptr)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
// Output auto inizialitation if not yet initialized
auto_init_if_empty(*output, *input);
}
const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
bool window_changed = false;
if (output != nullptr)
{
AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
window_changed = update_window_and_padding(win, input_access, output_access);
output_access.set_valid_region(win, input->valid_region());
}
else
{
window_changed = update_window_and_padding(
win, AccessWindowHorizontal(input, 0, num_elems_processed_per_iteration));
}
Status err = (window_changed)
? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!")
: Status{};
return std::make_pair(err, win);
}
} // namespace
CLActivationLayerExKernel::CLActivationLayerExKernel()
: _input(nullptr), _output(nullptr), _run_in_place(false)
{
}
void CLActivationLayerExKernel::configure(ICLTensor *input, ICLTensor *output,
ActivationLayerInfoEx act_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input);
_run_in_place = (output == nullptr) || (output == input);
if (output != nullptr)
{
// Output auto inizialitation if not yet initialized
auto_init_if_empty(*output->info(), *input->info()->clone());
}
ARM_COMPUTE_ERROR_THROW_ON(
validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr, act_info));
const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
const DataType dt = input->info()->data_type();
float a_const = act_info.a();
float b_const = act_info.b();
int a_const_int = 0;
int b_const_int = 0;
// Create quantized version of constants a, b if needed
if (is_data_type_quantized(dt))
{
a_const_int =
input->info()->quantization_info().quantize(a_const, RoundingPolicy::TO_NEAREST_UP);
b_const_int =
input->info()->quantization_info().quantize(b_const, RoundingPolicy::TO_NEAREST_UP);
}
// Set build options
std::set<std::string> build_opts;
build_opts.emplace(
("-DACT=" + lower_string(string_from_activation_func_ex(act_info.activation()))));
build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
build_opts.emplace(
("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
if (is_data_type_quantized(dt))
{
build_opts.emplace(("-DA_VAL=" + support::cpp11::to_string(a_const_int)));
build_opts.emplace(("-DB_VAL=" + support::cpp11::to_string(b_const_int)));
const int o1 = input->info()->quantization_info().offset;
// Quantized value of 0 corresponds to the offset o1
build_opts.emplace(("-DCONST_0=" + support::cpp11::to_string(o1)));
// Set scale and offset of the input and output if they have different quantization info
if (is_data_type_quantized_asymmetric(dt) && output != nullptr)
{
const float s1 = input->info()->quantization_info().scale;
const float s2 = output->info()->quantization_info().scale;
const int o2 = output->info()->quantization_info().offset;
if (o1 != o2 || s1 != s2)
{
build_opts.emplace(("-DS1_VAL=" + float_to_string_with_full_precision(s1)));
build_opts.emplace(("-DS2_VAL=" + float_to_string_with_full_precision(s2)));
build_opts.emplace(("-DO1_VAL=" + support::cpp11::to_string(o1)));
build_opts.emplace(("-DO2_VAL=" + support::cpp11::to_string(o2)));
}
}
}
else
{
build_opts.emplace(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
build_opts.emplace(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
}
build_opts.emplace((_run_in_place) ? "-DIN_PLACE" : "");
// Create kernel
std::string kernel_name = std::string("activation_layer_ex");
_kernel =
static_cast<cl::Kernel>(CLKernelLibraryEx::get().create_kernel(kernel_name, build_opts));
// Make sure _kernel is initialized before calling the parent's configure
_input = input;
_output = output;
// Configure kernel window
auto win_config =
validate_and_configure_window(input->info(), (_run_in_place) ? nullptr : output->info());
ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure_internal(win_config.second);
// Set config_id for enabling LWS tuning
_config_id = "activation_layer_ex_";
_config_id += lower_string(string_from_data_type(dt));
_config_id += "_";
_config_id += support::cpp11::to_string(input->info()->dimension(0));
_config_id += "_";
_config_id += support::cpp11::to_string(input->info()->dimension(1));
}
Status CLActivationLayerExKernel::validate(const ITensorInfo *input, const ITensorInfo *output,
const ActivationLayerInfoEx &act_info)
{
const bool run_in_place = (output == nullptr) || (output == input);
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, act_info));
ARM_COMPUTE_RETURN_ON_ERROR(
validate_and_configure_window(input->clone().get(),
(run_in_place) ? nullptr : output->clone().get())
.first);
return Status{};
}
void CLActivationLayerExKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
Window slice = collapsed.first_slice_window_3D();
do
{
unsigned int idx = 0;
add_3D_tensor_argument(idx, _input, slice);
if (!_run_in_place)
{
add_3D_tensor_argument(idx, _output, slice);
}
enqueue(queue, *this, slice, lws_hint());
} while (collapsed.slide_window_slice_3D(slice));
}
|
