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
|
/*
* Copyright (c) 2019 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 <cker/operation/FullyConnected.h>
#include "OperationUtil.h"
#include "exec/interp/Registration.h"
#include "model/operation/FullyConnectedNode.h"
#include "misc/polymorphic_downcast.h"
namespace neurun
{
namespace exec
{
namespace interp
{
namespace fc
{
void prepareFC(ExecEnv *env, const model::Operation &node)
{
const auto in_index = node.getInputs().at(model::operation::FullyConnectedNode::INPUT);
const auto kernel_index = node.getInputs().at(model::operation::FullyConnectedNode::WEIGHT);
const auto bias_index = node.getInputs().at(model::operation::FullyConnectedNode::BIAS);
const auto out_index = node.getOutputs().at(0);
const auto in_tensor = env->tensorAt(in_index);
const auto kernel_tensor = env->tensorAt(kernel_index);
const auto bias_tensor = env->tensorAt(bias_index);
UNUSED_RELEASE(in_tensor);
UNUSED_RELEASE(kernel_tensor);
UNUSED_RELEASE(bias_tensor);
assert(in_tensor->num_dimensions() >= 2);
assert(kernel_tensor->num_dimensions() == 2);
assert(bias_tensor->num_dimensions() == 1);
const auto input_size_with_batch = in_tensor->num_elements();
const auto num_units = kernel_tensor->dimension(0);
const auto input_size = kernel_tensor->dimension(1);
const auto batch_size = input_size_with_batch / input_size;
assert(input_size_with_batch % input_size == 0);
assert(num_units == bias_tensor->dimension(0));
// Make output tensor info
model::Shape output_shape(2);
output_shape.dim(0) = batch_size;
output_shape.dim(1) = num_units;
const model::OperandInfo out_info{output_shape, in_tensor->tensorInfo().typeInfo()};
env->allocateIfNeeded(out_index, out_info);
auto out_tensor = env->tensorAt(out_index);
UNUSED_RELEASE(out_tensor);
// Handle same ifm & ofm data type only
assert(in_tensor->data_type() == out_tensor->data_type());
assert(out_tensor->num_dimensions() == 2);
assert(out_tensor->dimension(0) == batch_size);
assert(out_tensor->dimension(1) == num_units);
}
void invoke(const ITensor *ifm_tensor, const ITensor *ker_tensor, const ITensor *bias_tensor,
const ITensor *ofm_tensor, const model::operation::FullyConnectedNode::Param ¶m)
{
const auto ifm_buffer = ifm_tensor->bufferRO();
const auto ker_buffer = ker_tensor->bufferRO();
const auto bias_buffer = bias_tensor->bufferRO();
auto ofm_buffer = ofm_tensor->buffer();
// Calculate
nnfw::cker::FullyConnectedParams cker_param;
calculateActivationRange(param.activation, &cker_param.float_activation_min,
&cker_param.float_activation_max);
const auto cker_ifm_shape = convertExtendShape(ifm_tensor->tensorInfo().shape());
const auto cker_ker_shape = convertExtendShape(ker_tensor->tensorInfo().shape());
const auto cker_bias_shape = convertExtendShape(bias_tensor->tensorInfo().shape());
const auto cker_ofm_shape = convertExtendShape(ofm_tensor->tensorInfo().shape());
const float *ifm_ptr = reinterpret_cast<const float *>(ifm_buffer);
const float *ker_ptr = reinterpret_cast<const float *>(ker_buffer);
const float *bias_ptr = reinterpret_cast<const float *>(bias_buffer);
float *ofm_ptr = reinterpret_cast<float *>(ofm_buffer);
nnfw::cker::FullyConnected(cker_param, cker_ifm_shape, ifm_ptr, cker_ker_shape, ker_ptr,
cker_bias_shape, bias_ptr, cker_ofm_shape, ofm_ptr);
}
void invokeFC(const ExecEnv *env, const model::Operation &node)
{
const auto &conv_node =
nnfw::misc::polymorphic_downcast<const model::operation::FullyConnectedNode &>(node);
const auto ifm_index = node.getInputs().at(model::operation::FullyConnectedNode::INPUT);
const auto ker_index = node.getInputs().at(model::operation::FullyConnectedNode::WEIGHT);
const auto bias_index = node.getInputs().at(model::operation::FullyConnectedNode::BIAS);
const auto ofm_index = node.getOutputs().at(0);
const auto ifm_tensor = env->tensorAt(ifm_index);
const auto ker_tensor = env->tensorAt(ker_index);
const auto bias_tensor = env->tensorAt(bias_index);
const auto ofm_tensor = env->tensorAt(ofm_index);
const auto data_type = ifm_tensor->data_type();
if (data_type == model::DataType::FLOAT32)
{
invoke(ifm_tensor, ker_tensor, bias_tensor, ofm_tensor, conv_node.param());
}
else
{
throw std::runtime_error{"NYI: Support float only"};
}
}
} // namespace fc
OpKernel *getFullyConnectedNode()
{
static OpKernel kernel = {fc::prepareFC, fc::invokeFC};
return &kernel;
}
} // namespace interp
} // namespace exec
} // namespace neurun
|
