megenginelite-sys 1.8.2

A safe megenginelite wrapper in Rust
Documentation 
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

/**
 * \file imperative/python/src/common.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */

#include "./common.h"

#include <pybind11/operators.h>

#include "./helper.h"
#include "./numpy_dtypes.h"
#include "megbrain/comp_node.h"
#include "megbrain/graph.h"
#include "megbrain/imperative/physical_tensor.h"

namespace py = pybind11;
using namespace mgb;
using namespace imperative;

namespace {

template <typename XTensorND>
auto def_TensorND(py::object parent, const char* name) {
    return py::class_<XTensorND>(parent, name)
            .def_property_readonly(
                    "shape", py::overload_cast<>(&XTensorND::shape, py::const_))
            .def_property_readonly(
                    "dtype", py::overload_cast<>(&XTensorND::dtype, py::const_))
            .def_property_readonly(
                    "comp_node", py::overload_cast<>(&XTensorND::comp_node, py::const_))
            .def("copy_from", &XTensorND::template copy_from<DeviceTensorStorage>)
            .def("copy_from", &XTensorND::template copy_from<HostTensorStorage>)
            .def("copy_from_fixlayout",
                 py::overload_cast<const DeviceTensorND&>(
                         &XTensorND::template copy_from_fixlayout<DeviceTensorStorage>))
            .def("copy_from_fixlayout",
                 py::overload_cast<const HostTensorND&>(
                         &XTensorND::template copy_from_fixlayout<HostTensorStorage>));
}

std::string default_device = "xpux";

}  // namespace

void set_default_device(const std::string& device) {
    default_device = device;
}

std::string get_default_device() {
    return default_device;
}

void init_common(py::module m) {
    auto PyCompNode =
            py::class_<CompNode>(m, "CompNode")
                    .def(py::init())
                    .def(py::init(
                            py::overload_cast<const std::string&>(&CompNode::load)))
                    .def_property_readonly(
                            "logical_name",
                            [](const CompNode& cn) { return cn.to_string_logical(); })
                    .def_property_readonly(
                            "physical_name",
                            [](const CompNode& cn) { return cn.to_string_physical(); })
                    .def_property_readonly(
                            "get_mem_status_bytes",
                            [](const CompNode& cn) {
                                return cn.get_mem_status_bytes();
                            })
                    .def_property_readonly(
                            "get_used_memory",
                            [](const CompNode& cn) { return cn.get_used_memory(); })
                    .def_property_readonly(
                            "get_max_used_memory",
                            [](const CompNode& cn) { return cn.get_max_used_memory(); })
                    .def_property_readonly(
                            "get_reserved_memory",
                            [](const CompNode& cn) { return cn.get_reserved_memory(); })
                    .def_property_readonly(
                            "get_max_reserved_memory",
                            [](const CompNode& cn) {
                                return cn.get_max_reserved_memory();
                            })
                    .def_static(
                            "reset_max_memory_stats",
                            [](const CompNode& cn) {
                                cn.reset_max_used_memory();
                                cn.reset_max_reserved_memory();
                            })
                    .def("create_event", &CompNode::create_event,
                         py::arg("flags") = 0ul)
                    .def_static("_set_default_device", &set_default_device)
                    .def_static("_get_default_device", &get_default_device)
                    .def("__str__", &CompNode::to_string_logical)
                    .def("__repr__",
                         [](const CompNode& cn) {
                             return mgb::ssprintf(
                                     "CompNode(\"%s\" from \"%s\")",
                                     cn.to_string_physical().c_str(),
                                     cn.to_string_logical().c_str());
                         })
                    .def("__hash__", [](CompNode cn) { return mgb::hash(cn); })
                    .def_static("_sync_all", &CompNode::sync_all)
                    .def(py::self == py::self)
                    .def_static(
                            "_get_device_count", &CompNode::get_device_count,
                            "Get total number of specific devices on this system")
                    .def(py::pickle(
                            [](const CompNode& cn) {
                                return py::str(cn.to_string_logical());
                            },
                            [](py::str cn) { return CompNode::load(cn); }));

    py::class_<CompNode::Event, std::shared_ptr<CompNode::Event>>(PyCompNode, "Event")
            .def("record", &CompNode::Event::record)
            .def("wait", &CompNode::Event::host_wait);

    py::implicitly_convertible<std::string, CompNode>();

    def_TensorND<DeviceTensorND>(m, "DeviceTensorND")
            .def("numpy", [](const DeviceTensorND& self) {
                HostTensorND hv;
                hv.copy_from(self).sync();
                return py::handle(
                        npy::ndarray_from_tensor(hv, npy::ShareType::TRY_SHARE));
            });

    def_TensorND<HostTensorND>(m, "HostTensorND")
            .def(py::init([](py::array data, CompNode cn, DType dtype) {
                if (!cn.valid()) {
                    throw py::type_error("device must not be None");
                }
                return npy::np2tensor(data.ptr(), npy::Meth::borrow(cn), dtype);
            }))
            .def("numpy", [](const HostTensorND& self) {
                return py::reinterpret_steal<py::object>(
                        npy::ndarray_from_tensor(self, npy::ShareType::TRY_SHARE));
            });

    py::class_<cg::OperatorNodeConfig>(m, "OperatorNodeConfig")
            .def(py::init())
            .def_property(
                    "name",
                    [](const OperatorNodeConfig& config) -> py::object {
                        auto name = config.name();
                        if (name.valid()) {
                            return py::str(name.val());
                        } else {
                            return py::none();
                        }
                    },
                    [](OperatorNodeConfig& config, std::string name) {
                        config.name(std::move(name));
                    })
            .def_property(
                    "dtype",
                    [](const OperatorNodeConfig& config) {
                        return config.output_dtype();
                    },
                    [](OperatorNodeConfig& config, DType dtype) {
                        config.output_dtype(dtype);
                    })
            .def_property(
                    "comp_node_arr",
                    [](const OperatorNodeConfig& config) -> py::tuple {
                        auto arr = config.comp_node();
                        std::vector<CompNode> tmp(arr.begin(), arr.end());
                        return py::cast(tmp);
                    },
                    [](OperatorNodeConfig& config, std::vector<CompNode> cns) {
                        config.comp_node_arr({cns.begin(), cns.end()});
                    })
            .def_property(
                    "comp_node",
                    [](const OperatorNodeConfig& config) {
                        auto arr = config.comp_node();
                        if (arr.size() != 1) {
                            throw py::value_error("invalid number of comp_node");
                        }
                        return arr[0];
                    },
                    [](OperatorNodeConfig& config, CompNode cn) {
                        OperatorNodeConfig::CompNodeArray arr{cn};
                        config.comp_node_arr(arr);
                    });

    py::class_<LogicalTensorDesc>(m, "TensorAttr")
            .def(py::init())
            .def(py::init([](const TensorShape& shape, const DType& dtype,
                             const CompNode& comp_node) {
                return LogicalTensorDesc{TensorLayout{shape, dtype}, comp_node};
            }))
            .def_property(
                    "shape",
                    [](const LogicalTensorDesc& desc) {
                        return static_cast<TensorShape>(desc.layout);
                    },
                    [](LogicalTensorDesc& desc, TensorShape shape) {})
            .def_property(
                    "dtype",
                    [](const LogicalTensorDesc& desc) { return desc.layout.dtype; },
                    [](LogicalTensorDesc& desc, DType dtype) {
                        desc.layout.dtype = dtype;
                    })
            .def_readwrite("comp_node", &LogicalTensorDesc::comp_node);

    py::enum_<CompNode::DeviceType>(m, "DeviceType")
            .value("UNSPEC", CompNode::DeviceType::UNSPEC)
            .value("CUDA", CompNode::DeviceType::CUDA)
            .value("ROCM", CompNode::DeviceType::ROCM)
            .value("CPU", CompNode::DeviceType::CPU)
            .value("CAMBRICON", CompNode::DeviceType::CAMBRICON)
            .value("ATLAS", CompNode::DeviceType::ATLAS)
            .value("MULTITHREAD", CompNode::DeviceType::MULTITHREAD)
            .value("MAX_DEVICE_ID", CompNode::DeviceType::MAX_DEVICE_ID);

    m.def("set_prealloc_config", &CompNode::set_prealloc_config,
          "specifies how to pre-allocate from raw dev allocator");

    m.def("get_cuda_compute_capability", &CompNode::get_compute_capability);

    m.def("what_is_xpu",
          [] { return CompNode::Locator::parse("xpux").to_physical().type; });

    init_npy_num_bfloat16(m);
    init_npy_num_intbx(m);
    init_dtypes(m);
}