graph_rt.cpp 32.0 KB
Newer Older
1 2
#include "./graph_rt.h"

M
Megvii Engine Team 已提交
3 4 5 6
#include "./common.h"
#include "./helper.h"
#include "./ops.h"
#include "megbrain/gopt/inference.h"
7
#include "megbrain/graph/cg.h"
M
Megvii Engine Team 已提交
8
#include "megbrain/imperative.h"
9
#include "megbrain/imperative/opr_utility.h"
M
Megvii Engine Team 已提交
10 11
#include "megbrain/imperative/profiler_plugin.h"
#include "megbrain/opr/basic_arith.h"
M
Megvii Engine Team 已提交
12
#include "megbrain/opr/io.h"
13
#include "megbrain/opr/utility.h"
14
#include "megbrain/plugin/profiler.h"
M
Megvii Engine Team 已提交
15
#include "megbrain/serialization/serializer.h"
16 17 18 19 20

namespace py = pybind11;

using namespace mgb;
using namespace imperative;
21
namespace ser = mgb::serialization;
22

23 24
using _OptimizeForInferenceOptions = mgb::gopt::OptimizeForInferenceOptions;
using _LayoutTransform = _OptimizeForInferenceOptions::LayoutTransform;
25
using _AlgoStrategy = opr::mixin::AlgoChooserHelper::ExecutionPolicy::Strategy;
26
using _SerializationMetadata = mgb::serialization::Metadata;
27
using _SerializationFormat = mgb::serialization::GraphDumpFormat;
28

29 30 31 32 33
namespace {
class _CompGraphProfilerImpl {
    std::shared_ptr<ComputingGraph> m_comp_graph;
    GraphProfiler m_profiler;

M
Megvii Engine Team 已提交
34 35 36 37 38 39 40 41
public:
    _CompGraphProfilerImpl(std::shared_ptr<ComputingGraph> cg)
            : m_comp_graph{cg}, m_profiler{m_comp_graph.get()} {}

    std::string _get_result() {
        auto json = m_profiler.to_json_full(m_comp_graph->current_comp_seq());
        return json->to_string();
    }
42
};
43

M
Megvii Engine Team 已提交
44 45
struct WeakRendezvousArray : public std::vector<std::weak_ptr<RendezvousBase>>,
                             public UserDataContainer::UserData {
46 47 48
    MGB_TYPEINFO_OBJ_DECL;
};
MGB_TYPEINFO_OBJ_IMPL(WeakRendezvousArray);
M
Megvii Engine Team 已提交
49
}  // namespace
50 51
#define DEF_READWRITE(name) .def_readwrite(#name, &CURRENT_CLASS::name)

M
Megvii Engine Team 已提交
52
template <typename T>
53 54
auto def_rendezvous(py::object m, const char* name) {
    return py::class_<Rendezvous<T>, std::shared_ptr<Rendezvous<T>>>(m, name)
M
Megvii Engine Team 已提交
55 56 57 58 59 60 61 62 63 64 65
            .def(py::init([]() { return Rendezvous<T>::make(); }))
            .def("set", [](Rendezvous<T>& r, T v) { r.set(std::move(v)); })
            .def(
                    "get", [](Rendezvous<T>& r) { return r.get(); },
                    py::call_guard<py::gil_scoped_release>())
            .def("drop", &Rendezvous<T>::drop)
            .def("reset", &Rendezvous<T>::reset)
            .def("set_exception", [](Rendezvous<T>& r, std::string&& message) {
                r.set_exception(std::make_exception_ptr(
                        std::runtime_error(std::move(message))));
            });
66 67 68
}

using TensorAttr = LogicalTensorDesc;
M
Megvii Engine Team 已提交
69
using HostNDWithEvent = std::pair<HostTensorND, std::shared_ptr<CompNode::Event>>;
70

M
Megvii Engine Team 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83
std::vector<mgb::cg::VarNode*> _replace_vars(
        const std::vector<mgb::cg::VarNode*>& repl_src,
        const std::vector<mgb::cg::VarNode*>& repl_dst,
        const std::vector<mgb::cg::VarNode*>& vars) {
    mgb::ThinHashMap<SymbolVar, SymbolVar> varmap;
    for (size_t i = 0; i < repl_src.size(); ++i) {
        varmap[SymbolVar(repl_src[i])] = SymbolVar(repl_dst[i]);
    }
    SymbolVarArray symvars(vars.begin(), vars.end());
    auto sym_result = mgb::cg::replace_vars(symvars, varmap);
    std::vector<mgb::cg::VarNode*> result;
    for (auto symvar : sym_result) {
        result.push_back(symvar.node());
84
    }
M
Megvii Engine Team 已提交
85 86
    return result;
}
87 88

typedef std::vector<mgb::cg::OperatorNodeBase*> OperatorArray;
M
Megvii Engine Team 已提交
89 90 91 92 93 94
std::vector<mgb::cg::VarNode*> _replace_oprs(
        const OperatorArray& repl_src, const OperatorArray& repl_dst,
        const std::vector<mgb::cg::VarNode*>& vars) {
    mgb::ThinHashMap<mgb::cg::OperatorNodeBase*, mgb::cg::OperatorNodeBase*> oprmap;
    for (size_t i = 0; i < repl_src.size(); ++i) {
        oprmap[repl_src[i]] = repl_dst[i];
95
    }
M
Megvii Engine Team 已提交
96 97 98 99 100 101 102 103
    const SymbolVarArray symvars(vars.begin(), vars.end());
    auto sym_result = mgb::cg::replace_oprs(symvars, oprmap);
    std::vector<mgb::cg::VarNode*> result;
    for (auto symvar : sym_result) {
        result.push_back(symvar.node());
    }
    return result;
}
104 105

void _set_priority_to_id(const std::vector<mgb::cg::VarNode*>& dest_vars) {
M
Megvii Engine Team 已提交
106 107 108
    auto on_opr = [](mgb::cg::OperatorNodeBase* opr) {
        if (opr->node_prop().attribute().priority == 0) {
            opr->node_prop().attribute().priority = opr->id();
109
        }
M
Megvii Engine Team 已提交
110 111 112 113 114
    };
    mgb::cg::DepOprIter dep_iter{on_opr};
    for (const auto& var : dest_vars) {
        dep_iter.add(SymbolVar(var));
    }
115 116
}

117 118
py::object Py_Varnode = py::none();

119
void init_graph_rt(py::module m) {
M
Megvii Engine Team 已提交
120 121
    static const std::unique_ptr<mgb::OprFootprint> _imperative_sm_opr_footprint_ptr{
            std::make_unique<mgb::OprFootprint>()};
122

123 124
    def_rendezvous<DeviceTensorND>(m, "DeviceTensorNDRendezvous");

M
Megvii Engine Team 已提交
125 126
    def_rendezvous<HostNDWithEvent>(m, "HostTensorNDRendezvous");

127 128
    def_rendezvous<TensorAttr>(m, "TensorAttrRendezvous");

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
    Py_Varnode =
            py::class_<cg::VarNode, GraphNodePtr<cg::VarNode>>(m, "VarNode")
                    .def_property_readonly(
                            "owner", [](cg::VarNode* v) { return v->owner_opr(); })
                    .def_property_readonly(
                            "graph", [](cg::VarNode* v) { return v->owner_graph(); })
                    .def_property(
                            "name", py::overload_cast<>(&VarNode::name, py::const_),
                            py::overload_cast<std::string>(&VarNode::name))
                    .def_property_readonly(
                            "dtype", [](cg::VarNode* v) { return v->dtype(); })
                    .def_property_readonly(
                            "comp_node", [](cg::VarNode* v) { return v->comp_node(); })
                    .def_property_readonly(
                            "shape",
                            [](cg::VarNode* v) -> const TensorShape* {
                                auto&& mgr = v->owner_graph()->static_infer_manager();
                                return mgr.infer_shape_fallible(v);
                            })
                    .def_property_readonly(
                            "value",
                            [](cg::VarNode* v) -> py::object {
                                auto&& mgr = v->owner_graph()->static_infer_manager();
                                auto&& type = mgr.get_infer_type(v);
                                using InferType = cg::static_infer::InferType;
                                if (!(type.value &
                                      (InferType::CONST | InferType::RT_STATIC))) {
                                    return py::none();
                                }
                                auto* val = mgr.infer_value_fallible(v);
                                if (!val) {
                                    return py::none();
                                }
                                return py::cast(*val).attr("numpy")();
                            })
                    .def_property_readonly(
                            "id", [](cg::VarNode* v) { return (v->id()); })
                    .def("__repr__", [](cg::VarNode* v) { return "Var:" + v->name(); });
M
Megvii Engine Team 已提交
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

    py::class_<cg::OperatorNodeBase, GraphNodePtr<cg::OperatorNodeBase>>(
            m, "OperatorNode")
            .def_property_readonly(
                    "graph",
                    [](cg::OperatorNodeBase* opr) { return opr->owner_graph(); })
            .def_property(
                    "name",
                    py::overload_cast<>(&cg::OperatorNodeBase::name, py::const_),
                    py::overload_cast<std::string>(&cg::OperatorNodeBase::name))
            .def_property_readonly(
                    "inputs",
                    [](cg::OperatorNodeBase* opr) { return to_tuple(opr->input()); })
            .def_property_readonly(
                    "outputs",
                    [](cg::OperatorNodeBase* opr) {
                        return to_tuple(opr->usable_output());
                    })
            .def_property_readonly(
                    "id", [](cg::OperatorNodeBase* opr) { return opr->id(); })
            .def_property_readonly(
                    "params",
                    [](cg::OperatorNodeBase* opr) {
                        return _imperative_sm_opr_footprint_ptr->calc_footprint(opr)
                                .param->to_string();
                    })
            .def_property_readonly(
                    "type",
                    [](cg::OperatorNodeBase* opr) { return opr->dyn_typeinfo()->name; })
            .def("__repr__",
                 [](cg::OperatorNodeBase* opr) { return "Opr:" + opr->name(); })
            .def_property(
                    "priority",
                    [](cg::OperatorNodeBase* opr) {
                        return opr->node_prop().attribute().priority;
                    },
                    [](cg::OperatorNodeBase* opr, int priority) {
                        opr->node_prop().attribute().priority = priority;
                    });
206

207
    py::class_<cg::AsyncExecutable>(m, "AsyncExecutable")
M
Megvii Engine Team 已提交
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
            .def("execute", &cg::AsyncExecutable::execute,
                 py::call_guard<py::gil_scoped_release>())
            .def("wait", &cg::AsyncExecutable::wait,
                 py::call_guard<py::gil_scoped_release>())
            .def("get_prev_exec_time", &cg::AsyncExecutable::get_prev_exec_time,
                 py::call_guard<py::gil_scoped_release>())
            .def("_to_json",
                 [](cg::AsyncExecutable* exec) {
                     py::call_guard<py::gil_scoped_release>();
                     // dump currently compiled computing graph for debugging
                     return exec->to_json()->to_string();
                 })
            // only used for exception handle
            .def_property_readonly(
                    "_all_rendezvous",
                    [](cg::AsyncExecutable* exec) {
                        auto ud =
                                exec->owner_graph()
                                        ->options()
                                        .user_data.get_user_data<WeakRendezvousArray>();
                        std::vector<std::shared_ptr<RendezvousBase>> ret;
                        if (ud.second) {
                            for (auto&& r : *ud.first[0]) {
                                if (auto p = r.lock()) {
                                    ret.emplace_back(std::move(p));
                                }
                            }
                        }
                        return ret;
                    })
            .def("get_static_memory_alloc_info",
                 &cg::AsyncExecutable::get_static_memory_alloc_info,
                 py::call_guard<py::gil_scoped_release>());

    auto PyComputingGraph =
            py::class_<cg::ComputingGraph, std::shared_ptr<cg::ComputingGraph>>(
                    m, "ComputingGraph")
                    .def(py::init(py::overload_cast<>(&cg::ComputingGraph::make)))
                    .def("compile",
                         [](cg::ComputingGraph& graph,
                            const std::vector<cg::VarNode*>& dest_vars) {
                             mgb_assert(!dest_vars.empty());
                             cg::ComputingGraph::OutputSpec spec;
                             for (auto v : dest_vars) {
                                 spec.emplace_back(v, nullptr);
                             }
                             return graph.compile(spec);
                         })
                    .def_property_readonly(
                            "options",
                            py::overload_cast<>(&cg::ComputingGraph::options));

    py::class_<_CompGraphProfilerImpl, std::shared_ptr<_CompGraphProfilerImpl>>(
            m, "GraphProfiler")
            .def(py::init([](std::shared_ptr<ComputingGraph> graph) {
263
                return std::make_shared<_CompGraphProfilerImpl>(graph);
M
Megvii Engine Team 已提交
264 265 266 267
            }))
            .def("get", [](_CompGraphProfilerImpl& profiler) {
                return profiler._get_result();
            });
268

269 270
    using interpreter::intl::ProfilerPlugin;
    py::class_<ProfilerPlugin, std::shared_ptr<ProfilerPlugin>>(m, "GraphProfiler2")
M
Megvii Engine Team 已提交
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
            .def(py::init<cg::ComputingGraph*>());

    auto GraphOptimizeOptions =
            py::class_<_OptimizeForInferenceOptions>(m, "GraphOptimizeOptions")
                    .def(py::init())
                    .def("serialize", &_OptimizeForInferenceOptions::serialize)
                    .def_static(
                            "deserialize", &_OptimizeForInferenceOptions::deserialize)
                    .def_readwrite(
                            "f16_io_f32_comp",
                            &_OptimizeForInferenceOptions::f16_io_f32_comp)
                    .def_readwrite(
                            "f16_io_comp", &_OptimizeForInferenceOptions::f16_io_comp)
                    .def_readwrite(
                            "fuse_conv_bias_nonlinearity",
                            &_OptimizeForInferenceOptions::fuse_conv_bias_nonlinearity)
                    .def_readwrite(
                            "fuse_conv_bias_with_z",
                            &_OptimizeForInferenceOptions::fuse_conv_bias_with_z)
                    .def_readwrite(
                            "fuse_preprocess",
                            &_OptimizeForInferenceOptions::fuse_preprocess)
                    .def_readwrite(
                            "layout_transform",
295 296 297
                            &_OptimizeForInferenceOptions::layout_transform)
                    .def_readwrite(
                            "fuse_grain", &_OptimizeForInferenceOptions::fuse_grain);
298 299

    py::enum_<_LayoutTransform>(GraphOptimizeOptions, "LayoutTransform")
M
Megvii Engine Team 已提交
300 301 302 303 304 305 306 307 308 309 310
            .value("DEFAULT", _LayoutTransform::DEFAULT)
            .value("NCHW4", _LayoutTransform::NCHW4)
            .value("NHWCD4", _LayoutTransform::NHWCD4)
            .value("NCHW88", _LayoutTransform::NCHW88)
            .value("NCHW44", _LayoutTransform::NCHW44)
            .value("NCHW44_DOT", _LayoutTransform::NCHW44_DOT)
            .value("NCHW32", _LayoutTransform::NCHW32)
            .value("CHWN4", _LayoutTransform::CHWN4)
            .value("NCHW64", _LayoutTransform::NCHW64)
            .export_values();

311 312
    py::enum_<_SerializationFormat>(m, "SerializationFormat")
            .value("FBS", _SerializationFormat::FLATBUFFERS)
313
            .value("FBS_V2", _SerializationFormat::FLATBUFFERS_V2)
314 315
            .export_values();

M
Megvii Engine Team 已提交
316 317 318 319 320 321 322 323 324 325 326 327 328 329
    m.def("optimize_for_inference",
          [](const VarNodeArray& dest_vars, const _OptimizeForInferenceOptions& opt) {
              SymbolVarArray symvars(dest_vars.begin(), dest_vars.end());
              auto res_symvars = mgb::gopt::optimize_for_inference(symvars, opt);
              VarNodeArray vars;
              for (auto& si : res_symvars)
                  vars.push_back(si.node());
              return vars;
          });

    m.def("modify_opr_algo_strategy_inplace",
          [](const VarNodeArray& dest_vars, const _AlgoStrategy& strategy) {
              mgb::gopt::modify_opr_algo_strategy_inplace(dest_vars, strategy);
          });
330

331 332
    m.def("get_info_for_strip", [](const std::vector<VarNode*>& dest_vars) {
        std::unordered_set<const char*> opr_types, dtype_names, elemwise_modes;
M
Megvii Engine Team 已提交
333
        auto on_opr = [&](cg::OperatorNodeBase* opr) {
334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
            if (ser::GraphDumper::should_remove_in_dump(opr))
                return;
            opr_types.insert(opr->dyn_typeinfo()->name);
            for (auto i : opr->output())
                dtype_names.insert(i->dtype().name());
            if (opr->same_type<opr::Elemwise>()) {
                auto mode = opr->cast_final<opr::Elemwise>().param().mode;
                elemwise_modes.insert(
                        megdnn::Elemwise::ModeTrait::from_mode(mode).name);
            }
        };
        cg::DepOprIter opr_iter{on_opr};
        for (auto i : dest_vars)
            opr_iter.add(i->owner_opr());

M
Megvii Engine Team 已提交
349
        auto to_json = [](const std::unordered_set<const char*>& v) {
350 351 352
            std::vector<std::string> vs(v.begin(), v.end());
            std::sort(vs.begin(), vs.end());
            auto ret = json::Array::make();
M
Megvii Engine Team 已提交
353
            for (auto&& i : vs)
354 355 356 357 358
                ret->add(json::String::make(i));
            return ret;
        };

        return json::Object::make({
M
Megvii Engine Team 已提交
359 360 361 362 363
                                          {"opr_types", to_json(opr_types)},
                                          {"dtypes", to_json(dtype_names)},
                                          {"elemwise_modes", to_json(elemwise_modes)},
                                  })
                ->to_string();
364 365
    });

366
    py::class_<_SerializationMetadata>(m, "SerializationMetadata")
M
Megvii Engine Team 已提交
367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385
            .def(py::init())
            .def_property(
                    "user_info",
                    [](const _SerializationMetadata& meta) {
                        return py::bytes(meta.get_user_info());
                    },
                    &_SerializationMetadata::set_user_info)
            .def_readonly(
                    "optimized_for_inference",
                    &_SerializationMetadata::optimized_for_inference)
            .def_property(
                    "optimize_options", &_SerializationMetadata::get_optimize_options,
                    &_SerializationMetadata::set_optimize_options)
            .def_readwrite("graph_modified", &_SerializationMetadata::graph_modified)
            .def_readwrite("is_valid", &_SerializationMetadata::is_valid);

    m.def("dump_graph",
          [](const std::vector<VarNode*>& dest_vars, int keep_var_name,
             bool keep_opr_name, bool keep_param_name, bool keep_opr_priority,
386 387 388 389
             bool no_change_graph, std::optional<_SerializationMetadata> metadata,
             std::optional<_SerializationFormat> dump_format,
             std::optional<int> model_version, py::list& stat, py::list& inputs,
             py::list& outputs, py::list& params) {
M
Megvii Engine Team 已提交
390
              std::vector<uint8_t> buf;
391
              ser::GraphDumpFormat format = ser::GraphDumpFormat::FLATBUFFERS_V2;
392
              int version = 2;
393 394 395
              if (dump_format.has_value()) {
                  format = dump_format.value();
              }
396 397 398
              if (model_version.has_value()) {
                  version = model_version.value();
              }
399
              auto dumper = ser::GraphDumper::make(
400
                      ser::OutputFile::make_vector_proxy(&buf), format, version);
M
Megvii Engine Team 已提交
401 402 403 404
              SymbolVarArray symvars(dest_vars.begin(), dest_vars.end());

              ser::GraphDumper::DumpConfig config{
                      keep_var_name, keep_param_name, keep_opr_priority, keep_opr_name};
405
              config.no_change_graph = no_change_graph;
M
Megvii Engine Team 已提交
406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466

              ser::GraphDumper::DumpResult rst;
              if (metadata)
                  rst = dumper->dump(symvars, config, *metadata);
              else
                  rst = dumper->dump(symvars, config);

              for (auto i : rst.inputs) {
                  inputs.append(py::cast(i));
              }
              for (auto i : rst.outputs) {
                  outputs.append(py::cast(i));
              }
              for (auto i : rst.params) {
                  params.append(py::cast(i));
              }
              auto rst_stat = std::vector{
                      rst.nr_opr, rst.tot_bytes, rst.tensor_value_bytes,
                      static_cast<size_t>(rst.content_hash)};
              for (auto i : rst_stat) {
                  stat.append(py::cast(i));
              }
              return py::bytes(reinterpret_cast<const char*>(&buf[0]), buf.size());
          });

    m.def("load_graph",
          [](std::string& buf, py::list& output_var_map, py::list& output_var_list) {
              auto file = ser::InputFile::make_mem_proxy(buf.c_str(), buf.length());
              auto format = ser::GraphLoader::identify_graph_dump_format(*file);
              auto loader = ser::GraphLoader::make(std::move(file), format.val());
              ser::GraphLoader::LoadConfig config;
              auto rst = loader->load(config);
              for (auto i : rst.output_var_map) {
                  output_var_map.append(py::make_tuple(i.first, i.second.node()));
              }
              for (auto i : rst.output_var_list) {
                  output_var_list.append(i.node());
              }
              std::unordered_map<HostTensorND*, const std::string*> tensor2name;
              for (const auto& pair : rst.tensor_map) {
                  tensor2name[pair.second.get()] = &pair.first;
              }
              auto cb = [&tensor2name, graph = rst.graph](cg::OperatorNodeBase* opr) {
                  if (!opr->same_type<opr::Host2DeviceCopy>())
                      return;
                  auto& h2d = opr->cast_final_safe<opr::Host2DeviceCopy>();
                  auto it = tensor2name.find(h2d.host_data().get());
                  mgb_throw_if(
                          it == tensor2name.end(), GraphError,
                          "unbound Host2DeviceCopy in loaded graph");
                  h2d.output(0)->name(*it->second);
              };
              cg::DepOprIter iter{cb};
              for (const auto& var : rst.output_var_list) {
                  iter.add(var);
              }
              auto ret = py::tuple(2);
              ret[0] = py::cast(rst.graph);
              ret[1] = py::cast(rst.metadata);
              return ret;
          });
467

468 469
#define CURRENT_CLASS cg::ComputingGraph::Options

470
    // clang-format off
M
Megvii Engine Team 已提交
471 472
    auto PyComputingGraphOptions =
            py::class_<cg::ComputingGraph::Options>(PyComputingGraph, "Options")
473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496
                // DEF_READWRITE(opr_attribute)
                DEF_READWRITE(seq_opt)
                DEF_READWRITE(graph_opt)
                DEF_READWRITE(graph_opt_level)
                DEF_READWRITE(log_level)
                DEF_READWRITE(async_exec_level)
                DEF_READWRITE(force_dynamic_alloc)
                DEF_READWRITE(var_sanity_check_first_run)
                DEF_READWRITE(allocate_static_mem_after_graph_compile)
                DEF_READWRITE(fake_next_exec)
                DEF_READWRITE(enable_sublinear_memory_opt)
                DEF_READWRITE(enable_dtr_memory_opt)
                DEF_READWRITE(no_profiling_on_shape_change)
                DEF_READWRITE(enable_var_mem_defragment)
                DEF_READWRITE(enable_grad_var_static_reshape)
                DEF_READWRITE(enable_memory_swap)
                DEF_READWRITE(comp_node_seq_record_level)
                DEF_READWRITE(no_force_inplace)
                DEF_READWRITE(sublinear_mem_config)
                DEF_READWRITE(dtr_config)
                // DEF_READWRITE(eager_evaluation)
                // DEF_READWRITE(imperative_proxy_graph)
                // DEF_READWRITE(extra_vardeps)
                // DEF_READWRITE(user_data)
M
Megvii Engine Team 已提交
497
            ;
498
    // clang-format on
499 500 501 502 503

#undef CURRENT_CLASS
#define CURRENT_CLASS cg::ComputingGraph::Options::SeqOpt

    py::class_<cg::ComputingGraph::Options::SeqOpt>(PyComputingGraphOptions, "SeqOpt")
M
Megvii Engine Team 已提交
504 505
            DEF_READWRITE(enable_mem_plan_opt) DEF_READWRITE(enable_mem_reuse_alloc)
                    DEF_READWRITE(enable_seq_comp_node_opt);
506 507 508 509

#undef CURRENT_CLASS
#define CURRENT_CLASS cg::ComputingGraph::Options::GraphOpt

510
    auto PyGraphOpt = py::class_<cg::ComputingGraph::Options::GraphOpt>(
M
Megvii Engine Team 已提交
511 512 513
            PyComputingGraphOptions, "GraphOpt") DEF_READWRITE(jit)
            DEF_READWRITE(jit_config)
            DEF_READWRITE(tensorrt);
514 515

#undef CURRENT_CLASS
516
#define CURRENT_CLASS cg::ComputingGraph::Options::GraphOpt::JITConfig
517

M
Megvii Engine Team 已提交
518 519 520
    py::class_<cg::ComputingGraph::Options::GraphOpt::JITConfig>(
            PyGraphOpt, "JITConfig") DEF_READWRITE(fuse_dimshuffle)
            DEF_READWRITE(fuse_reduce);
521 522

#undef CURRENT_CLASS
523 524
#define CURRENT_CLASS cg::ComputingGraph::Options::SublinearMemConfig

M
Megvii Engine Team 已提交
525 526 527 528
    py::class_<cg::ComputingGraph::Options::SublinearMemConfig>(
            PyComputingGraphOptions, "SublinearMemConfig") DEF_READWRITE(thresh_nr_try)
            DEF_READWRITE(genetic_nr_iter) DEF_READWRITE(genetic_pool_size)
                    DEF_READWRITE(lb_memory_mb) DEF_READWRITE(num_worker);
529

530 531 532 533
#undef CURRENT_CLASS

#define CURRENT_CLASS cg::ComputingGraph::Options::DTRConfig

M
Megvii Engine Team 已提交
534 535 536 537
    py::class_<cg::ComputingGraph::Options::DTRConfig>(
            PyComputingGraphOptions, "DTRConfig") DEF_READWRITE(eviction_threshold)
            DEF_READWRITE(evictee_minimum_size) DEF_READWRITE(recomp_memory_factor)
                    DEF_READWRITE(recomp_time_factor);
538

539
#undef CURRENT_CLASS
540 541
    auto common = rel_import("common", m, 1);

M
Megvii Engine Team 已提交
542 543 544 545 546 547 548 549
    common.def(
            "invoke_op",
            [](const OpDef& def, const std::vector<cg::VarNode*> inputs,
               cg::ComputingGraph* graph) {
                cg::VarNodeArray vinputs(inputs.begin(), inputs.end());
                return to_tuple(OpDef::apply_on_var_node(def, vinputs));
            },
            py::arg(), py::arg(), py::arg("graph") = py::none());
550

M
Megvii Engine Team 已提交
551 552
    auto input_callback = [](auto callback, const CompNode& comp_node,
                             const DType& dtype, const TensorShape& shape,
553
                             const std::vector<cg::VarNode*>& inputs,
M
Megvii Engine Team 已提交
554
                             cg::ComputingGraph* graph, bool use_static_shape) {
555 556 557 558 559 560 561 562
        if (!graph) {
            graph = inputs[0]->owner_graph();
        }
        SymbolVarArray sinputs;
        for (auto i : inputs) {
            sinputs.emplace_back(i);
        }
        static_assert(!std::is_reference<decltype(callback)>::value);
M
Megvii Engine Team 已提交
563 564 565
        auto soutputs = opr::InputCallback::make(
                *graph, std::move(callback), comp_node, dtype, shape, sinputs,
                use_static_shape);
566 567 568 569 570 571 572 573
        std::vector<VarNode*> outputs;
        outputs.reserve(soutputs.size());
        for (auto i : soutputs) {
            outputs.push_back(i.node());
        }
        return outputs;
    };

M
Megvii Engine Team 已提交
574
    m.def("make_shared", [](cg::ComputingGraph* graph, const DeviceTensorND& data) {
M
Megvii Engine Team 已提交
575 576 577 578
        return opr::SharedDeviceTensor::make(
                       *graph, std::make_shared<DeviceTensorND>(data))
                .node();
    });
M
Megvii Engine Team 已提交
579

M
Megvii Engine Team 已提交
580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654
    m.def(
            "make_const",
            [](cg::ComputingGraph* graph, py::array data, CompNode cn, DType dtype,
               std::optional<std::string> name) {
                if (!cn.valid()) {
                    cn = CompNode::load(get_default_device());
                }
                OperatorNodeConfig config(cn);
                if (name) {
                    config.name(*name);
                }
                auto hv = npy::np2tensor(data.ptr(), npy::Meth::borrow(cn), dtype);
                return opr::ImmutableTensor::make(*graph, hv, config).node();
            },
            py::arg(), py::arg(), py::arg(), py::arg(), py::arg() = py::none());

    m.def(
            "make_h2d",
            [](cg::ComputingGraph& graph, CompNode cn, DType dtype, TensorShape shape,
               std::optional<std::string> name) {
                if (!cn.valid()) {
                    throw py::type_error("device must be valid");
                }
                if (!dtype.valid()) {
                    throw py::type_error("dtype must be valid");
                }
                OperatorNodeConfig config;
                if (name) {
                    config.name(*name);
                }
                return opr::Host2DeviceCopy::make(
                               graph, std::make_shared<HostTensorND>(cn, shape, dtype),
                               config)
                        .node();
            },
            py::arg(), py::arg(), py::arg(), py::arg() = py::none(),
            py::arg() = py::none());

    m.def("_replace_vars", &_replace_vars, py::arg(), py::arg(), py::arg());
    m.def("_replace_oprs", &_replace_oprs, py::arg(), py::arg(), py::arg());
    m.def("_set_priority_to_id", &_set_priority_to_id, py::arg());

    m.def(
            "input_callback",
            [input_callback](
                    std::function<DeviceTensorND(void)> callback,
                    const CompNode& comp_node, const DType& dtype,
                    const TensorShape& shape, const std::vector<cg::VarNode*>& inputs,
                    cg::ComputingGraph* graph, bool use_static_shape) {
                return input_callback(
                        [f = std::move(callback)]() {
                            py::gil_scoped_acquire _;
                            return f();
                        },
                        comp_node, dtype, shape, inputs, graph, use_static_shape);
            },
            py::arg(), py::arg(), py::arg(), py::arg() = py::none(),
            py::arg() = py::tuple(), py::arg("graph") = py::none(),
            py::arg("use_static_shape") = false);

    m.def(
            "input_callback",
            [input_callback](
                    std::shared_ptr<Rendezvous<DeviceTensorND>> p,
                    const CompNode& comp_node, const DType& dtype,
                    const TensorShape& shape, const std::vector<cg::VarNode*>& inputs,
                    cg::ComputingGraph* graph, bool use_static_shape) {
                auto f = [p]() -> DeviceTensorND { return p->get(); };
                return input_callback(
                        std::move(f), comp_node, dtype, shape, inputs, graph,
                        use_static_shape);
            },
            py::arg(), py::arg(), py::arg(), py::arg() = py::none(),
            py::arg() = py::tuple(), py::arg("graph") = py::none(),
            py::arg("use_static_shape") = false);
655

656
    auto output_callback = [](auto callback, const std::vector<cg::VarNode*>& inputs,
M
Megvii Engine Team 已提交
657 658
                              std::shared_ptr<RendezvousBase> r = {},
                              bool borrow = false, bool prefer_host_value = false) {
659 660 661
        if (r) {
            mgb_assert(inputs.size());
            auto cg = inputs[0]->owner_graph();
M
Megvii Engine Team 已提交
662 663
            cg->options()
                    .user_data.get_user_data_or_create<WeakRendezvousArray>()
664 665
                    ->emplace_back(r);
        }
666 667 668 669 670
        SymbolVarArray sinputs;
        for (auto i : inputs) {
            sinputs.emplace_back(i);
        }
        static_assert(!std::is_reference<decltype(callback)>::value);
M
Megvii Engine Team 已提交
671 672
        opr::OutputCallback::Param param{
                std::move(callback), borrow, prefer_host_value};
673 674 675 676
        auto output = opr::OutputCallback::make(std::move(param), sinputs);
        return output.node();
    };

M
Megvii Engine Team 已提交
677 678 679 680 681
    m.def("output_callback", [output_callback](
                                     std::function<void(DeviceTensorND)> callback,
                                     std::vector<cg::VarNode*> inputs) {
        auto f = [f = std::move(callback)](DeviceTensorND dv) {
            auto task = [f = std::move(f), dv = std::move(dv)]() { f(dv); };
682 683 684 685 686
            py_task_q.add_task(std::move(task));
        };
        return output_callback(std::move(f), std::move(inputs));
    });

M
Megvii Engine Team 已提交
687 688 689 690
    m.def("output_callback", [output_callback](
                                     std::shared_ptr<Rendezvous<DeviceTensorND>> p,
                                     std::vector<cg::VarNode*> inputs) {
        auto f = [p](DeviceTensorND dv) { p->set(std::move(dv)); };
691
        return output_callback(std::move(f), std::move(inputs), p);
692 693
    });

M
Megvii Engine Team 已提交
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
    m.def("value_output_callback",
          [output_callback](
                  std::shared_ptr<Rendezvous<HostNDWithEvent>> p,
                  std::vector<cg::VarNode*> inputs) {
              auto f = [p](DeviceTensorND dv) {
                  HostNDWithEvent hv_with_event;
                  hv_with_event.first.copy_from(dv);
                  hv_with_event.second = dv.comp_node().create_event();
                  hv_with_event.second->record();
                  p->set(std::move(hv_with_event));
              };
              return output_callback(std::move(f), std::move(inputs), p, true, true);
          });

    m.def("attr_output_callback", [output_callback](
                                          std::shared_ptr<Rendezvous<TensorAttr>> p,
                                          std::vector<cg::VarNode*> inputs) {
711 712 713
        auto f = [p](DeviceTensorND dv) {
            p->set(TensorAttr{TensorLayout{dv.shape(), dv.dtype()}, dv.comp_node()});
        };
714
        return output_callback(std::move(f), std::move(inputs), p, true);
715
    });
716 717 718 719 720 721 722 723

    m.def("virtual_dep", [](std::vector<cg::VarNode*> inputs, std::string device) {
        auto&& graph = inputs[0]->owner_graph();
        VarNodeArray inps(inputs.begin(), inputs.end());
        cg::OperatorNodeConfig config;
        if (device.length() > 0) {
            config.comp_node(CompNode::load(device));
        }
M
Megvii Engine Team 已提交
724 725
        cg::OperatorNodeBase* opr =
                graph->insert_opr(std::make_unique<mgb::opr::VirtualDep>(inps, config));
726 727
        return opr;
    });
728
}