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
|
/*
* 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 "util/Utils.h"
#include "ir/InternalType.h"
#include "ir/Shape.h"
#include "ir/operation/AvgPool2D.h"
#include "ir/operation/MaxPool2D.h"
#include "util/ShapeInference.h"
namespace neurun
{
namespace shape_inference
{
//
// Helper functions
//
namespace
{
template <typename T, typename U>
typename std::enable_if<std::is_integral<T>::value && std::is_integral<U>::value,
typename std::common_type<T, U>::type>::type
ceil_div(T dividend, U divisor)
{
assert(dividend > 0 && divisor > 0 && "this implementations is for positive numbers only");
return (dividend + divisor - 1) / divisor;
}
// Calculate the result of broadcast of two shapes
ir::Shape broadcastShapes(const ir::Shape &lhs_shape, const ir::Shape &rhs_shape)
{
ir::Shape out_shape;
auto max_rank = std::max(lhs_shape.rank(), rhs_shape.rank());
for (int idx = 0; idx < max_rank; ++idx)
{
// Go over operands dimensions from right to left
int lhs_idx = lhs_shape.rank() - idx - 1;
int rhs_idx = rhs_shape.rank() - idx - 1;
int32_t lhs_dim = lhs_idx >= 0 ? lhs_shape.dim(lhs_idx) : 1;
int32_t rhs_dim = rhs_idx >= 0 ? rhs_shape.dim(rhs_idx) : 1;
if (lhs_dim != 1 && rhs_dim != 1 && lhs_dim != rhs_dim)
throw std::runtime_error("Incompatible shapes for broadcast");
out_shape.prepend(std::max(lhs_dim, rhs_dim));
}
return out_shape;
}
// Calculate output height and width of convolution-like operation
std::pair<int, int> calcConvLikeHeightAndWidth(const int in_h, const int in_w, const int ker_h,
const int ker_w, const ir::Padding pad,
const ir::Stride stride)
{
int32_t out_h = 0, out_w = 0;
switch (pad.type)
{
case ir::PaddingType::SAME:
out_h = ceil_div(in_h, stride.vertical);
out_w = ceil_div(in_w, stride.horizontal);
break;
case ir::PaddingType::VALID:
out_h = ceil_div(in_h - ker_h + 1, stride.vertical);
out_w = ceil_div(in_w - ker_w + 1, stride.horizontal);
break;
case ir::PaddingType::EXPLICIT:
out_h = (in_h + pad.param.top + pad.param.bottom - ker_h) / stride.vertical + 1;
out_w = (in_w + pad.param.left + pad.param.right - ker_w) / stride.horizontal + 1;
break;
default:
assert(false);
}
return {out_h, out_w};
}
} // namespace
//
// Shape inference
//
Shapes inferEltwiseShape(const ir::Shape &lhs_shape, const ir::Shape &rhs_shape)
{
return {broadcastShapes(lhs_shape, rhs_shape)};
}
Shapes inferAvgPoolShape(const ir::Shape &in_shape, const ir::operation::AvgPool2D::Param ¶m,
const ir::Layout layout)
{
assert(layout == ir::Layout::NHWC);
auto ifm_shape = in_shape.asFeature(layout);
const auto out_h_w = calcConvLikeHeightAndWidth(ifm_shape.H, ifm_shape.W, param.kh, param.kw,
param.padding, param.stride);
// Pooling don't change number of channels and batch size
return {ir::Shape{ifm_shape.N, out_h_w.first, out_h_w.second, ifm_shape.C}};
}
Shapes inferConcatShape(const Shapes &in_shapes, const ir::operation::Concat::Param ¶m)
{
const int32_t concat_axis = param.axis;
const auto &first_in_shape = in_shapes[0];
// Check that all shapes are equal except for concat axis dimension
for (const auto &in_shape : in_shapes)
{
assert(in_shape.rank() == first_in_shape.rank());
for (int64_t dim_idx = 0; dim_idx < in_shape.rank(); ++dim_idx)
assert(dim_idx == concat_axis || in_shape.dim(dim_idx) == first_in_shape.dim(dim_idx));
}
// Calculate output shape
ir::Shape out_shape(first_in_shape);
out_shape.dim(concat_axis) = 0;
for (const auto &in_shape : in_shapes)
out_shape.dim(concat_axis) += in_shape.dim(concat_axis);
return {out_shape};
}
Shapes inferMaxPoolShape(const ir::Shape &in_shape, const ir::operation::MaxPool2D::Param ¶m,
const ir::Layout layout)
{
assert(layout == ir::Layout::NHWC);
auto ifm_shape = in_shape.asFeature(layout);
const auto out_h_w = calcConvLikeHeightAndWidth(ifm_shape.H, ifm_shape.W, param.kh, param.kw,
param.padding, param.stride);
// Pooling don't change number of channels and batch size
return {ir::Shape{ifm_shape.N, out_h_w.first, out_h_w.second, ifm_shape.C}};
}
Shapes inferConv2DShape(const ir::Shape &in_shape, const ir::Shape &ker_shape,
const ir::operation::Conv2D::Param ¶m, ir::Layout layout)
{
assert(layout == ir::Layout::NHWC);
auto ifm_shape = in_shape.asFeature(layout);
// Kernel format is [depth_out, kernel_height, kernel_width, depth_in]
auto kf_shape = ker_shape.asFeature(layout);
assert(ifm_shape.C == kf_shape.C);
const auto out_h_w = calcConvLikeHeightAndWidth(ifm_shape.H, ifm_shape.W, kf_shape.H, kf_shape.W,
param.padding, param.stride);
return {ir::Shape{ifm_shape.N, out_h_w.first, out_h_w.second, kf_shape.N}};
}
Shapes inferDepthwiseConv2DShape(const ir::Shape &in_shape, const ir::Shape &ker_shape,
const ir::operation::DepthwiseConv2D::Param ¶m,
ir::Layout layout)
{
assert(layout == ir::Layout::NHWC);
auto ifm_shape = in_shape.asFeature(layout);
// Kernel format is [1, kernel_height, kernel_width, depth_out]
auto kf_shape = ker_shape.asFeature(layout);
assert(kf_shape.C == static_cast<int32_t>(ifm_shape.C * param.multiplier));
assert(kf_shape.N == 1);
const auto out_h_w = calcConvLikeHeightAndWidth(ifm_shape.H, ifm_shape.W, kf_shape.H, kf_shape.W,
param.padding, param.stride);
return {ir::Shape{ifm_shape.N, out_h_w.first, out_h_w.second, kf_shape.C}};
}
Shapes inferFullyConnectedShape(const ir::Shape &in_shape, const ir::Shape &ker_shape)
{
assert(in_shape.rank() >= 2);
assert(ker_shape.rank() == 2);
const auto input_size_with_batch = in_shape.num_elements();
const auto num_units = ker_shape.dim(0);
const auto input_size = ker_shape.dim(1);
const auto batch_size = input_size_with_batch / input_size;
assert(input_size_with_batch % input_size == 0);
return {{ir::Shape({static_cast<int32_t>(batch_size), num_units})}};
}
} // namespace shape_inference
} // namespace neurun
|
