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
|
/*
* Copyright (c) 2018 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 "Convolution.h"
#include "ConvolutionSpec.h"
#include "Convert.h"
#include "IRBuilder.h"
#include <nncc/core/ADT/kernel/Overlay.h>
#include <nncc/core/ADT/kernel/NCHWLayout.h>
#include <morph/caffe.h>
#include <cassert>
using namespace nncc::core::ADT;
using namespace morph::caffe;
using tensor::num_elements;
namespace caffeimport
{
void ConvolutionBuilder::build(const ::caffe::LayerParameter &layer,
GraphBuilderContext *context) const
{
coco::Module *module = context->module();
coco::Data *data = context->data();
coco::Block *blk = context->block();
std::map<std::string, tensor::Shape> &shape_ctx = context->shape_ctx();
std::map<std::string, coco::Bag *> &bag_ctx = context->bag_ctx();
WeightContext &weight_ctx = context->weight_ctx();
assert(layer.bottom().size() == 1);
assert(layer.top().size() == 1);
assert(layer.has_convolution_param());
const auto ¶m = layer.convolution_param();
ConvolutionSpec spec{param};
{
const auto ifm_name = layer.bottom(0);
const auto ifm_shape = shape_ctx.at(ifm_name);
spec.ifm_shape(ifm_shape);
}
// NOTE The current implementation focuses on 2D convolution
// TODO Support general ND convolution
assert(spec.num_batch_axes() == 1);
assert(spec.num_spatial_axes() == 2);
// Create an object for an input feature map
const auto ifm_name = layer.bottom(0);
const auto ifm_shape = shape_ctx.at(ifm_name);
auto ifm_bag = bag_ctx.at(ifm_name);
auto ifm_obj = module->entity()->object()->create<coco::FeatureObject>();
ifm_obj->bag(ifm_bag);
ifm_obj->layout(coco::FeatureLayouts::BCHW::create(as_feature_shape(ifm_shape)));
// Create an object for an output feature map
const auto ofm_name = layer.top(0);
const auto ofm_shape = spec.ofm_shape();
auto ofm_bag = module->entity()->bag()->create(num_elements(ofm_shape));
auto ofm_obj = module->entity()->object()->create<coco::FeatureObject>();
ofm_obj->bag(ofm_bag);
ofm_obj->layout(coco::FeatureLayouts::BCHW::create(as_feature_shape(ofm_shape)));
// Create an object for kernel
using namespace coco::KernelLayouts;
const auto ker_shape = spec.ker_shape();
auto ker_bag = module->entity()->bag()->create(num_elements(ker_shape));
auto ker_obj = module->entity()->object()->create<coco::KernelObject>();
ker_obj->bag(ker_bag);
ker_obj->layout(NCHW::create(as_kernel_shape(ker_shape)));
// Create a kernel overlay for the kernel object
data->f32()->allocate(ker_bag);
// Initialize the kernel overlay
assert(weight_ctx.blob_count(layer.name()) >= 1);
auto ker_blob = weight_ctx.blob_get(layer.name(), 0);
assert(ker_shape == caffeimport::as_tensor_shape(ker_blob));
auto ker_dst = data->f32()->access(ker_obj);
auto ker_src = kernel::OverlayFactory<float, kernel::NCHWLayout>::make(
ker_obj->shape(), ker_blob->mutable_data()->begin());
for (uint32_t n = 0; n < ker_obj->shape().count(); ++n)
{
for (uint32_t ch = 0; ch < ker_obj->shape().depth(); ++ch)
{
for (uint32_t row = 0; row < ker_obj->shape().height(); ++row)
{
for (uint32_t col = 0; col < ker_obj->shape().width(); ++col)
{
ker_dst->at(n, ch, row, col) = ker_src.at(n, ch, row, col);
}
}
}
}
// Create a Load op
auto load = op_builder(module).load(ifm_obj).pop();
// Create a Conv2D op
auto op = module->entity()->op()->create<coco::Conv2D>();
op->group(spec.group());
op->ker(ker_obj);
op->stride()->vertical(spec.stride(0));
op->stride()->horizontal(spec.stride(1));
op->pad()->top(spec.pad(0));
op->pad()->bottom(spec.pad(0));
op->pad()->left(spec.pad(1));
op->pad()->right(spec.pad(1));
op->arg(load);
// Create an Eval instruction
auto ins = instr_builder(module).eval(ofm_obj, op);
// Append the instruction to the block
blk->instr()->append(ins);
//
// coco IR allows Conv2D fused with Add, but the current implementation of enco backend
// is unable to process such a tree.
//
// As a workaround, caffe frontend constructs a instruction for Conv2D and Add.
//
if (param.bias_term())
{
assert(weight_ctx.blob_count(layer.name()) >= 2);
// Create Bag & Object
auto bias_bag = module->entity()->bag()->create(ker_shape.dim(0));
auto bias_obj = module->entity()->object()->create<coco::FeatureObject>();
bias_obj->bag(bias_bag);
bias_obj->layout(coco::FeatureLayouts::BC::create(as_feature_shape(ofm_shape)));
auto added_bag = module->entity()->bag()->create(num_elements(ofm_shape));
auto added_obj = module->entity()->object()->create<coco::FeatureObject>();
added_obj->bag(added_bag);
added_obj->layout(coco::FeatureLayouts::BCHW::create(as_feature_shape(ofm_shape)));
// Create Op
auto bias_add = op_builder(module).load(bias_obj).load(ofm_obj).add().pop();
// Create Instr
auto bias_add_ins = instr_builder(module).eval(added_obj, bias_add);
// Append the instruction
blk->instr()->append(bias_add_ins);
// Fill bias data
data->f32()->allocate(bias_bag);
auto bias_span = data->f32()->weight(bias_bag);
auto bias_blob = weight_ctx.blob_get(layer.name(), 1);
for (uint32_t ch = 0; ch < ker_obj->shape().count(); ++ch)
{
bias_span[ch] = bias_blob->data(ch);
}
// Update output
ofm_bag = added_bag;
}
// Update bag and shape context
bag_ctx[ofm_name] = ofm_bag;
shape_ctx[ofm_name] = ofm_shape;
}
} // namespace caffeimport
|
