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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
|
/*
* 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 "DepthwiseConv2D.h"
#include "Convert.h"
#include "IRBuilder.h"
#include "GraphBuilder.h"
#include "Padding.h"
#include "Activation.h"
#include <morph/tflite.h>
#include <coco/IR/Module.h>
#include <coco/IR/KernelLayouts.h>
#include <coco/IR/FeatureLayouts.h>
#include <nncc/core/ADT/tensor/Shape.h>
#include <schema_generated.h>
#include <cassert>
using namespace nncc::core::ADT;
using namespace morph::tflite;
namespace tflimport
{
bool DepthwiseConv2DGraphBuilder::validate(const tflite::Operator *op) const
{
auto const options = op->builtin_options_as_DepthwiseConv2DOptions();
if ((options->stride_h() == 0) || (options->stride_w() == 0))
{
return false;
}
return true;
}
void DepthwiseConv2DGraphBuilder::build(const tflite::Operator *op,
GraphBuilderContext *context) const
{
assert(context != nullptr);
// preparation
coco::Module *m = context->m();
coco::Data *d = context->d();
coco::Block *blk = context->block();
TensorContext &tensor_context = context->tensor();
TensorBags &bags = context->bags();
TflBufferContext &buffer_context = context->buffer();
const tflite::SubGraph *graph = context->graph();
IndexVector opinputs = as_index_vector(op->inputs());
IndexVector opoutputs = as_index_vector(op->outputs());
// these are fixed in tflite
// input index 0 : input feature
// input index 1 : kernel
// input index 2 : bias (optional)
bool hasBias = (opinputs.size() == 3);
assert(opinputs.size() == 2 || hasBias);
assert(opoutputs.size() == 1);
int ifm_idx = opinputs.at(0);
int ker_idx = opinputs.at(1);
int ofm_idx = opoutputs.at(0);
const tensor::Shape &ifm_shape = tensor_context.shape(ifm_idx);
const tensor::Shape &ofm_shape = tensor_context.shape(ofm_idx);
tensor::Shape &ker_shape = const_cast<tensor::Shape &>(tensor_context.shape(ker_idx));
assert(ifm_shape.rank() == 4);
assert(ofm_shape.rank() == 4);
assert(ker_shape.rank() == 4);
assert(ker_shape.dim(0) == 1); // value > 1 was not tested. This value seems 1 in DepthwiseConv2D
assert(ifm_shape.dim(3) == ofm_shape.dim(3));
assert(ofm_shape.dim(3) == ker_shape.dim(3));
// Create an input feature map object
auto *ifm_obj = m->entity()->object()->create<coco::FeatureObject>();
auto *ifm_bag = bags.bag(ifm_idx);
ifm_obj->bag(ifm_bag);
ifm_obj->layout(coco::FeatureLayouts::BHWC::create(as_feature_shape(ifm_shape)));
// Create an an output feature map object
auto *ofm_obj = m->entity()->object()->create<coco::FeatureObject>();
auto *ofm_bag = bags.bag(ofm_idx);
ofm_obj->bag(ofm_bag);
ofm_obj->layout(coco::FeatureLayouts::BHWC::create(as_feature_shape(ofm_shape)));
// Create an kernel object
auto *ker_obj = m->entity()->object()->create<coco::KernelObject>();
auto *ker_bag = bags.bag(ker_idx);
ker_obj->bag(ker_bag);
// Adjust tflite kernel shape [1, h, w, channel_out] for coco::Kernel.
// coco::Kernel will have kernel.count = channel_out, kernel.depth = 1 ( == ker_shape.dim(0))
kernel::Shape new_shape{ker_shape.dim(3), 1, ker_shape.dim(1), ker_shape.dim(2)};
ker_obj->layout(coco::KernelLayouts::NHWC::create(new_shape));
// Create a kernel overlay for the kernel object
// TODO : support for other types
d->f32()->allocate(ker_bag);
TflBufferContext::TflBuffer<float> buffer = buffer_context.tensor_buffer<float>(graph, ker_idx);
auto ker_spn = d->f32()->weight(ker_bag);
// Copy data from tflBuffer of [1, h, w, channel_out] shape to coco::Data, which will be accessed
// by coco::KernelLayouts::NHWC
for (auto n = 0; n < new_shape.count(); n++)
{
auto tfl_c = n;
for (auto h = 0; h < new_shape.height(); h++)
{
for (auto w = 0; w < new_shape.width(); w++)
{
auto hw = new_shape.height() * new_shape.width();
for (auto c = 0; c < new_shape.depth(); c++)
{
auto tfl_n = c;
auto hwc = hw * new_shape.depth();
auto wc = new_shape.width() * new_shape.depth();
ker_spn[n * hwc + h * wc + w * new_shape.depth() + c] =
buffer.ptr[tfl_n * hw * new_shape.count() + /* new_shape.count() is old c */
h * new_shape.width() * new_shape.count() + w * new_shape.count() + tfl_c];
}
}
}
}
// Create a Load op
auto load = op_builder(m).load(ifm_obj).pop();
// Create a coco::Conv2D op for DepthwiseConv2D
auto coco_dconv2d = m->entity()->op()->create<coco::Conv2D>();
// populating objects and options such as stride and padding for DepthwiseConv2D
coco_dconv2d->ker(ker_obj);
// setting params passed from TFLITE DepthwiseConv2DOptions
auto dconv_params = op->builtin_options_as_DepthwiseConv2DOptions();
assert(dconv_params->depth_multiplier() == 1); // other depth_multiplier was not tested
coco_dconv2d->group(ifm_obj->asFeature()->shape().depth());
coco_dconv2d->stride()->vertical(dconv_params->stride_h());
coco_dconv2d->stride()->horizontal(dconv_params->stride_w());
coco::Padding2D padding = depthwiseConv2D_padding(dconv_params, ifm_shape, ker_shape);
coco_dconv2d->pad()->top(padding.top());
coco_dconv2d->pad()->bottom(padding.bottom());
coco_dconv2d->pad()->left(padding.left());
coco_dconv2d->pad()->right(padding.right());
// Link ops
coco_dconv2d->arg(load);
// Object to store output for DepthwiseConv2D
auto *dconv2d_obj = m->entity()->object()->create<coco::FeatureObject>();
auto *dconv2d_bag = m->entity()->bag()->create(num_elements(ofm_shape));
dconv2d_obj->bag(dconv2d_bag);
dconv2d_obj->layout(coco::FeatureLayouts::BHWC::create(as_feature_shape(ofm_shape)));
// Create an Eval instruction for DepthwiseConv2D
auto dconv2d_ins = instr_builder(m).eval(dconv2d_obj, coco_dconv2d);
// Append the instruction to the block
blk->instr()->append(dconv2d_ins);
// Last Object to make a copy to Output Object
coco::FeatureObject *last_obj = dconv2d_obj;
if (hasBias)
{
// When there is a bias, use btmp_obj as bias add output
// Bias is adding last_obj with bias weight values
auto *btmp_obj = m->entity()->object()->create<coco::FeatureObject>();
auto *btmp_bag = m->entity()->bag()->create(num_elements(ofm_shape));
btmp_obj->bag(btmp_bag);
btmp_obj->layout(coco::FeatureLayouts::BHWC::create(ofm_obj->shape()));
int bias_idx = opinputs.at(2);
// Create an object for bias
auto bias_obj = m->entity()->object()->create<coco::FeatureObject>();
coco::Bag *bias_bag = bags.bag(bias_idx);
bias_obj->bag(bias_bag);
bias_obj->layout(coco::FeatureLayouts::BC::create(ofm_obj->shape()));
// Create Op of conv2d output (last_obj) + bias values(bias_obj)
auto bias_add = op_builder(m).load(last_obj).load(bias_obj).add().pop();
// Create Instr as bias add result write to btmp_obj
auto bias_add_ins = instr_builder(m).eval(btmp_obj, bias_add);
// Append the instruction
blk->instr()->append(bias_add_ins);
// Update last_obj to btmp_obj
last_obj = btmp_obj;
}
// fused activation
coco::FeatureObject *act_output =
build_activation(dconv_params->fused_activation_function(), blk, last_obj);
// Create Copy Instr of last_obj to Output Object
auto copy_ins = instr_builder(m).copy(ofm_obj, act_output);
blk->instr()->append(copy_ins);
}
} // namespace tflimport
|
