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
|
/*
* Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
*
* 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 "KernelGenerator.h"
#include "ops/ConvolutionLayer.h"
#include "ops/FullyConnectedLayer.h"
#include <backend/Backend.h>
#include <backend/IConfig.h>
#include <memory>
#include <util/Utils.h>
#include <util/logging.h>
#include <exec/DynamicShapeInferer.h>
#include <stdexcept>
namespace onert
{
namespace backend
{
namespace ruy
{
std::unique_ptr<exec::FunctionSequence> KernelGenerator::generate(ir::OperationIndex ind)
{
auto ret = std::make_unique<exec::FunctionSequence>();
assert(_tensor_builder->dynamicTensorManager());
assert(_tensor_reg);
// Prepare to handle dynamic tensors later
auto dyn_ctx = std::make_shared<exec::FunctionSequence::DynamicTensorCtx>();
{
dyn_ctx->op = &_operations_ctx.at(ind);
dyn_ctx->dynamic_shape_inferer = std::make_shared<exec::DynamicShapeInferer>(_ctx, _tensor_reg);
}
ret->dynamic_tensor_ctx(dyn_ctx);
auto &op = _graph.operations().at(ind);
op.accept(*this);
assert(_return_fn); // _return_fn must have been generated
ret->append(std::move(_return_fn));
for (auto ind : (op.getInputs() | ir::Remove::UNDEFINED) + op.getOutputs())
{
auto portable_tensor = _tensor_reg->getPortableTensor(ind);
if (portable_tensor)
{
assert(portable_tensor->layout() == ir::Layout::NHWC);
}
auto tensor = _tensor_reg->getNativeTensor(ind);
if (tensor)
{
tensor->increase_ref();
}
}
return ret;
}
KernelGenerator::KernelGenerator(
const ir::Graph &graph, const std::shared_ptr<TensorBuilder> &tensor_builder,
const std::shared_ptr<basic::TensorRegistry> &tensor_reg,
const std::shared_ptr<backend::custom::IKernelBuilder> &kernel_builder,
const std::shared_ptr<ExternalContext> &external_context)
: basic::KernelGeneratorBase{graph},
_ctx(graph.operands()), _operations_ctx{graph.operations()}, _current_layout{graph.layout()},
_tensor_builder(tensor_builder), _tensor_reg{tensor_reg}, _kernel_builder(kernel_builder),
_external_context(external_context)
{
// DO NOTHING
}
void KernelGenerator::visit(const ir::operation::Conv2D &node)
{
using ir::operation::Conv2D;
const auto ofm_index{node.getOutputs().at(0)};
const auto ifm_index{node.getInputs().at(Conv2D::Input::INPUT)};
const auto ker_index{node.getInputs().at(Conv2D::Input::KERNEL)};
const auto bias_index{node.getInputs().at(Conv2D::Input::BIAS)};
auto ofm_tensor = _tensor_reg->getPortableTensor(ofm_index);
auto ifm_tensor = _tensor_reg->getPortableTensor(ifm_index);
auto ker_tensor = _tensor_reg->getPortableTensor(ker_index);
auto bias_tensor = _tensor_reg->getPortableTensor(bias_index);
const auto stride = node.param().stride;
const auto activation = node.param().activation;
const auto param_padding = node.param().padding;
const auto dilation = node.param().dilation;
auto fn = std::make_unique<ops::ConvolutionLayer>();
if (_ctx.at(ifm_index).info().isDynamic() || _ctx.at(ker_index).info().isDynamic())
{
fn->configure(ifm_tensor, ker_tensor, bias_tensor, param_padding.type, param_padding.param.left,
param_padding.param.right, param_padding.param.top, param_padding.param.bottom,
stride.horizontal, stride.vertical, dilation.width_factor, dilation.height_factor,
activation, ofm_tensor, _external_context);
_return_fn = std::move(fn);
return;
}
const auto ifm_shape = _ctx.at(ifm_index).shape().asFeature(_current_layout);
const auto ofm_shape = _ctx.at(ofm_index).shape().asFeature(_current_layout);
// Kernel format is [depth_out, kernel_height, kernel_width, depth_in].
const auto &ker_shape = _ctx.at(ker_index).shape();
const auto ker_height = ker_shape.dim(1);
const auto ker_width = ker_shape.dim(2);
const auto padding =
ir::calculatePadding(param_padding, ifm_shape, ofm_shape, stride, ker_width, ker_height,
dilation.width_factor, dilation.height_factor);
fn->configure(ifm_tensor, ker_tensor, bias_tensor, param_padding.type, padding.left,
padding.right, padding.top, padding.bottom, stride.horizontal, stride.vertical,
dilation.width_factor, dilation.height_factor, activation, ofm_tensor,
_external_context);
_return_fn = std::move(fn);
}
void KernelGenerator::visit(const ir::operation::FullyConnected &node)
{
using ir::operation::FullyConnected;
const auto output_index{node.getOutputs().at(0)};
const auto input_index{node.getInputs().at(FullyConnected::Input::INPUT)};
const auto weight_index{node.getInputs().at(FullyConnected::Input::WEIGHT)};
const auto bias_index{node.getInputs().at(FullyConnected::Input::BIAS)};
const auto activation = node.param().activation;
const auto weights_format = node.param().weights_format;
auto output_tensor = _tensor_reg->getPortableTensor(output_index);
auto input_tensor = _tensor_reg->getPortableTensor(input_index);
auto weight_tensor = _tensor_reg->getPortableTensor(weight_index);
auto bias_tensor = bias_index.undefined() ? nullptr : _tensor_reg->getPortableTensor(bias_index);
auto fn = std::make_unique<ops::FullyConnectedLayer>();
fn->configure(input_tensor, weight_tensor, bias_tensor, activation, weights_format, output_tensor,
_external_context);
_return_fn = std::move(fn);
}
} // namespace ruy
} // namespace backend
} // namespace onert
|
