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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
//
// Copyright (C) 2018 Intel Corporation
#include "precomp.hpp"
#include <iostream>
#include <ade/util/zip_range.hpp>
#include "opencv2/gapi/opencv_includes.hpp"
#include "executor/gexecutor.hpp"
#include "compiler/passes/passes.hpp"
cv::gimpl::GExecutor::GExecutor(std::unique_ptr<ade::Graph> &&g_model)
: m_orig_graph(std::move(g_model))
, m_island_graph(GModel::Graph(*m_orig_graph).metadata()
.get<IslandModel>().model)
, m_gm(*m_orig_graph)
, m_gim(*m_island_graph)
{
// NB: Right now GIslandModel is acyclic, so for a naive execution,
// simple unrolling to a list of triggers is enough
// Naive execution model is similar to current CPU (OpenCV) plugin
// execution model:
// 1. Allocate all internal resources first (NB - CPU plugin doesn't do it)
// 2. Put input/output GComputation arguments to the storage
// 3. For every Island, prepare vectors of input/output parameter descs
// 4. Iterate over a list of operations (sorted in the topological order)
// 5. For every operation, form a list of input/output data objects
// 6. Run GIslandExecutable
// 7. writeBack
auto sorted = m_gim.metadata().get<ade::passes::TopologicalSortData>();
for (auto nh : sorted.nodes())
{
switch (m_gim.metadata(nh).get<NodeKind>().k)
{
case NodeKind::ISLAND:
{
std::vector<RcDesc> input_rcs;
std::vector<RcDesc> output_rcs;
input_rcs.reserve(nh->inNodes().size());
output_rcs.reserve(nh->outNodes().size());
auto xtract = [&](ade::NodeHandle slot_nh, std::vector<RcDesc> &vec) {
const auto orig_data_nh
= m_gim.metadata(slot_nh).get<DataSlot>().original_data_node;
const auto &orig_data_info
= m_gm.metadata(orig_data_nh).get<Data>();
vec.emplace_back(RcDesc{ orig_data_info.rc
, orig_data_info.shape
, orig_data_info.ctor});
};
// (3)
for (auto in_slot_nh : nh->inNodes()) xtract(in_slot_nh, input_rcs);
for (auto out_slot_nh : nh->outNodes()) xtract(out_slot_nh, output_rcs);
m_ops.emplace_back(OpDesc{ std::move(input_rcs)
, std::move(output_rcs)
, m_gim.metadata(nh).get<IslandExec>().object});
}
break;
case NodeKind::SLOT:
{
const auto orig_data_nh
= m_gim.metadata(nh).get<DataSlot>().original_data_node;
// (1)
initResource(orig_data_nh);
m_slots.emplace_back(DataDesc{nh, orig_data_nh});
}
break;
default:
GAPI_Assert(false);
break;
} // switch(kind)
} // for(gim nodes)
}
void cv::gimpl::GExecutor::initResource(const ade::NodeHandle &orig_nh)
{
const Data &d = m_gm.metadata(orig_nh).get<Data>();
if ( d.storage != Data::Storage::INTERNAL
&& d.storage != Data::Storage::CONST)
return;
// INTERNALS+CONST only! no need to allocate/reset output objects
// to as it is bound externally (e.g. already in the m_res)
switch (d.shape)
{
case GShape::GMAT:
{
const auto desc = util::get<cv::GMatDesc>(d.meta);
const auto type = CV_MAKETYPE(desc.depth, desc.chan);
m_res.slot<cv::gapi::own::Mat>()[d.rc].create(desc.size, type);
}
break;
case GShape::GSCALAR:
if (d.storage == Data::Storage::CONST)
{
auto rc = RcDesc{d.rc, d.shape, d.ctor};
magazine::bindInArg(m_res, rc, m_gm.metadata(orig_nh).get<ConstValue>().arg);
}
break;
case GShape::GARRAY:
// Constructed on Reset, do nothing here
break;
default:
GAPI_Assert(false);
}
}
void cv::gimpl::GExecutor::run(cv::gimpl::GRuntimeArgs &&args)
{
// (2)
const auto proto = m_gm.metadata().get<Protocol>();
// Basic check if input/output arguments are correct
// FIXME: Move to GCompiled (do once for all GExecutors)
if (proto.inputs.size() != args.inObjs.size()) // TODO: Also check types
{
util::throw_error(std::logic_error
("Computation's input protocol doesn\'t "
"match actual arguments!"));
}
if (proto.outputs.size() != args.outObjs.size()) // TODO: Also check types
{
util::throw_error(std::logic_error
("Computation's output protocol doesn\'t "
"match actual arguments!"));
}
namespace util = ade::util;
//ensure that output Mat parameters are correctly allocated
for (auto index : util::iota(proto.out_nhs.size()) ) //FIXME: avoid copy of NodeHandle and GRunRsltComp ?
{
auto& nh = proto.out_nhs.at(index);
const Data &d = m_gm.metadata(nh).get<Data>();
if (d.shape == GShape::GMAT)
{
using cv::util::get;
const auto desc = get<cv::GMatDesc>(d.meta);
const auto type = CV_MAKETYPE(desc.depth, desc.chan);
#if !defined(GAPI_STANDALONE)
// Building as part of OpenCV - follow OpenCV behavior
// if output buffer is not enough to hold the result, reallocate it
auto& out_mat = *get<cv::Mat*>(args.outObjs.at(index));
out_mat.create(cv::gapi::own::to_ocv(desc.size), type);
#else
// Building standalone - output buffer should always exist,
// and _exact_ match our inferred metadata
auto& out_mat = *get<cv::gapi::own::Mat*>(args.outObjs.at(index));
GAPI_Assert( out_mat.type() == type
&& out_mat.data != nullptr
&& out_mat.rows == desc.size.height
&& out_mat.cols == desc.size.width)
#endif // !defined(GAPI_STANDALONE)
}
}
// Update storage with user-passed objects
for (auto it : ade::util::zip(ade::util::toRange(proto.inputs),
ade::util::toRange(args.inObjs)))
{
magazine::bindInArg(m_res, std::get<0>(it), std::get<1>(it));
}
for (auto it : ade::util::zip(ade::util::toRange(proto.outputs),
ade::util::toRange(args.outObjs)))
{
magazine::bindOutArg(m_res, std::get<0>(it), std::get<1>(it));
}
// Reset internal data
for (auto &sd : m_slots)
{
const auto& data = m_gm.metadata(sd.data_nh).get<Data>();
magazine::resetInternalData(m_res, data);
}
// Run the script
for (auto &op : m_ops)
{
// (5)
using InObj = GIslandExecutable::InObj;
using OutObj = GIslandExecutable::OutObj;
std::vector<InObj> in_objs;
std::vector<OutObj> out_objs;
in_objs.reserve (op.in_objects.size());
out_objs.reserve(op.out_objects.size());
for (const auto &rc : op.in_objects)
{
in_objs.emplace_back(InObj{rc, magazine::getArg(m_res, rc)});
}
for (const auto &rc : op.out_objects)
{
out_objs.emplace_back(OutObj{rc, magazine::getObjPtr(m_res, rc)});
}
// (6)
op.isl_exec->run(std::move(in_objs), std::move(out_objs));
}
// (7)
for (auto it : ade::util::zip(ade::util::toRange(proto.outputs),
ade::util::toRange(args.outObjs)))
{
magazine::writeBack(m_res, std::get<0>(it), std::get<1>(it));
}
}
const cv::gimpl::GModel::Graph& cv::gimpl::GExecutor::model() const
{
return m_gm;
}
bool cv::gimpl::GExecutor::canReshape() const
{
// FIXME: Introduce proper reshaping support on GExecutor level
// for all cases!
return (m_ops.size() == 1) && m_ops[0].isl_exec->canReshape();
}
void cv::gimpl::GExecutor::reshape(const GMetaArgs& inMetas, const GCompileArgs& args)
{
GAPI_Assert(canReshape());
auto& g = *m_orig_graph.get();
ade::passes::PassContext ctx{g};
passes::initMeta(ctx, inMetas);
passes::inferMeta(ctx, true);
m_ops[0].isl_exec->reshape(g, args);
}