tensor.cpp 25.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/**
 * \file imperative/python/src/tensor.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2020 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.
 */

12 13
#include "megbrain/dtype.h"
#include "megbrain/common.h"
14
#include "megbrain/imperative/ops/utility.h"
15

16 17
#include "./tensor.h"
#include "./grad.h"
18
#include "./trace.h"
19 20
#include "./common.h"
#include "./numpy_dtypes.h"
21
#include "./graph_rt.h"
22
#include "./helper.h"
23 24 25

#include <pybind11/numpy.h>
#include <pybind11/operators.h>
26
#include <range/v3/all.hpp>
27 28 29

#include <unordered_map>

30
namespace py = pybind11;
31
namespace views = ranges::views;
32 33 34 35 36

namespace mgb::imperative::python {

std::unique_ptr<interpreter::Interpreter::Channel> interpreter_for_py;

37 38 39 40 41
py::object cpp_apply_with_tracing, cpp_apply_const_with_tracing,
           cpp_apply_compiled_mode, cpp_apply_const_compiled_mode;

py::object cpp_apply_backward_varnode;

42 43 44 45 46 47 48 49
void release_trace_apply_func(){
    cpp_apply_with_tracing.release();
    cpp_apply_const_with_tracing.release();
    cpp_apply_compiled_mode.release();
    cpp_apply_const_compiled_mode.release();
    cpp_apply_backward_varnode.release();
}

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
#define REGISTE_APPLY_FUNC(mode)                                    \
        void set_##mode(py::object pyf) {                           \
            mode = pybind11::reinterpret_steal<py::object>(pyf);    \
        }

REGISTE_APPLY_FUNC(cpp_apply_with_tracing)
REGISTE_APPLY_FUNC(cpp_apply_const_with_tracing)
REGISTE_APPLY_FUNC(cpp_apply_compiled_mode)
REGISTE_APPLY_FUNC(cpp_apply_const_compiled_mode)
REGISTE_APPLY_FUNC(cpp_apply_backward_varnode)

#undef REGISTE_APPLY_FUNC

bool is_tracing = false;
bool is_symbolic = false;
bool is_compiled = false;

#define SET_UNSET_PROP(mode)    \
    void set_##mode() {         \
        is_##mode = true;       \
    }                           \
    void unset_##mode() {       \
        is_##mode = false;      \
    }                           \

SET_UNSET_PROP(tracing)
SET_UNSET_PROP(symbolic)
SET_UNSET_PROP(compiled)

#undef SET_UNSET_PROP

bool skip_tracing = false;

83 84
Tensor::flags_t ApplyContext::global_disable = 0;

85 86 87 88
apply_result_t apply(ApplyContext& ctx) {
    // emulating scalar should be put to specific op's apply, e.g.,
    // elementwise, reduce, typecvt. Currently it's still handled at python
    // side. It could be move to C++ side if it has an impact on performance
89 90 91
    auto flags = ctx.flags & ~ApplyContext::global_disable;

    if (flags & Tensor::Flags::SCALAR) {
92 93 94
        // TODO: emulate scalar
    }

95
    if (flags & Tensor::Flags::GRAD) {
96 97 98
        return apply_grad(ctx);
    }

99
    if (flags & Tensor::Flags::TRACE) {
100
        return apply_trace(ctx);
101
    } else {
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121
        if (auto* op = ctx.op->try_cast_final<GenericPyOp>()) {
            py::tuple pyin(ctx.nargs);
            for (size_t i = 0; i < ctx.nargs; ++i) {
                pyin[i] = TensorWrapper::make(ctx.pytype, ctx.args[i]->shared_from_this());
            }
            auto f = py::getattr(op->obj, "_default_rule");
            auto pyout = py::reinterpret_steal<py::object>(PyObject_Call(f.ptr(), pyin.ptr(), nullptr));
            if (auto* tw = TensorWrapper::try_cast(pyout.ptr())) {
                return {tw->m_tensor};
            }
            apply_result_t ret;
            ret.reserve(py::len(pyout));
            for (auto&& i : pyout) {
                auto* tw = TensorWrapper::try_cast(i.ptr());
                mgb_assert(tw);
                ret.push_back(tw->m_tensor);
            }
            return ret;
        }

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
        SmallVector<interpreter::Interpreter::Handle> handles(ctx.nargs);
        for (size_t i = 0; i < ctx.nargs; ++i) {
            handles[i] = ctx.args[i]->m_handle.get();
        }

        auto output_handles = interpreter_for_py->apply_op(ctx.op, handles);

        apply_result_t outputs;
        outputs.reserve(output_handles.size());
        for (auto h : output_handles) {
            outputs.emplace_back(std::make_shared<Tensor>(h));
        }
        return outputs;
    }

    mgb_assert(0);
}

PyObject* py_apply(PyObject* self, PyObject*const* args, size_t nargs/* , PyObject* kwnames */) {
    try {
        // if (kwnames && PyTuple_GET_SIZE(kwnames)) {
        //     PyErr_SetString(PyExc_TypeError, "keyword argument not allowed");
        //     return nullptr;
        // }
        if (!nargs) {
            PyErr_SetString(PyExc_TypeError, "expect Op");
            return nullptr;
        }
150

151 152 153 154 155 156 157 158 159 160 161 162
        auto* op = args[0];

        PyTypeObject* pytype = args[1]->ob_type;
        ++args;
        --nargs;

        ApplyContext ctx;
        ctx.flags = 0;
        ctx.op = py::handle(op).cast<std::shared_ptr<OpDef>>();
        SmallVector<Tensor*, 64> tensors(nargs);
        ctx.args = &tensors[0];
        ctx.nargs = nargs;
163
        ctx.pytype = pytype;
164 165 166
        if (strstr(op->ob_type->tp_name, "BackwardGraph")) {
            ctx.backward = true;
        }
167 168

        for (size_t i = 0; i < nargs; ++i) {
169
            if (TensorWrapper* tw = TensorWrapper::try_cast(args[i])) {
170 171 172
                auto* t = tensors[i] = tw->m_tensor.get();
                ctx.flags |= t->m_flags;
            } else {
173 174 175 176 177
                PyErr_SetString(PyExc_TypeError, "expect Tensor");
                return nullptr;
            }
        }

178 179 180
        if (is_tracing) {
            ctx.flags |= Tensor::Flags::TRACE;
        }
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204

        auto outputs = apply(ctx);
        size_t nout = outputs.size();
        auto ret = py::tuple(nout);
        for (size_t i = 0; i < nout; ++i) {
            ret[i] = TensorWrapper::make(pytype, std::move(outputs[i]));
        }
        return ret.release().ptr();
    } catch (std::exception& e) {
        PyErr_SetString(PyExc_RuntimeError, e.what());
        return nullptr;
    }
}


TensorWrapper::TensorWrapper(PyObject* args, PyObject* kwargs) {
    if (kwargs && PyDict_Size(kwargs)) {
        throw py::type_error("keyword argument not allowed");
    }
    auto nargs = PyTuple_Size(args);
    auto tup = py::reinterpret_borrow<py::tuple>(args);
    if (nargs == 0) {
        throw py::type_error("too few arguments");
    }
205
    if (auto* t = try_cast(tup[0].ptr())) {
206 207 208 209 210
        if (nargs > 1) {
            throw py::type_error("expect 1 argument");
        }
        m_tensor = t->m_tensor;
    } else {
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
        if (nargs == 1) {
            auto arg0 = PyTuple_GetItem(args, 0);
            // for lazy_eval_tensor
            if (strstr(arg0->ob_type->tp_name, "VarNode")) {
                if (PyObject_HasAttrString(arg0, "_node")) {
                    arg0 = PyObject_GetAttrString(arg0, "_node");
                }
                m_tensor = std::make_shared<Tensor>(py::handle(arg0).cast<cg::VarNode *>());
            } else {
                // for DeviceTensorND
                if (strstr(arg0->ob_type->tp_name, "DeviceTensorND")) {
                    auto dv = py::handle(arg0).cast<DeviceTensorND>();
                    interpreter::Interpreter::Handle handle = interpreter_for_py->put(dv);
                    m_tensor = std::make_shared<Tensor>(handle);
                } else {
                    throw py::type_error("single argument is not tensor, varnode or devicetensor");
                }
            }
229
        } else {
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
            py::detail::loader_life_support life_sup; // required to cast DType
            auto data = tup[0].cast<py::array>();
            DType dtype = tup[1].cast<DType>();
            CompNode cn = tup[2].cast<CompNode>();
            bool is_const = tup[3].cast<bool>();
            if (nargs != 4) {
                throw py::type_error("expect 3 arguments");
            }

            // const op
            if (is_const && is_tracing) {
                py::object pyf;
                if (is_compiled) {
                    pyf = cpp_apply_const_compiled_mode;
                } else {
                    pyf = cpp_apply_const_with_tracing;
                }

                auto ret = pyf(*tup);
                auto py_ret = py::reinterpret_borrow<py::list>(ret);
250
                if (auto* t = try_cast(py_ret[0].ptr())) {
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
                    m_tensor = t->m_tensor;
                }
                return;
            }

            interpreter::Interpreter::Handle handle;
            constexpr auto size_threshhold = TensorShape::MAX_NDIM;
            if (data.size() > size_threshhold) {
                handle = interpreter_for_py->put(npy::np2tensor(data.ptr(), npy::Meth::borrow(cn), dtype));
            } else {
                HostTensorND ret(cn);
                handle = interpreter_for_py->put(npy::np2tensor(data.ptr(), npy::Meth::copy_into(&ret), dtype));
            }

            m_tensor = std::make_shared<Tensor>(handle);
266

267 268 269
            if (data.ndim() == 0) {
                m_tensor->m_flags |= Tensor::Flags::SCALAR;
            }
270 271 272 273 274
        }
    }
}


275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
#define REGISTE_TENSORWRAPPER_FUNC(type, member)                                    \
        PyObject* TensorWrapper::member() {                                         \
            return py::cast(m_tensor->m_trace_info.member).release().ptr();         \
        }                                                                           \
        void TensorWrapper::set_##member(PyObject* dest) {                          \
            auto py_dest = py::reinterpret_borrow<py::object>(dest);                \
            type real_dest = py_dest.cast<type>();                                  \
            m_tensor->m_trace_info.member = real_dest;                              \
        }

REGISTE_TENSORWRAPPER_FUNC(bool, data_read)
REGISTE_TENSORWRAPPER_FUNC(bool, value_read)
REGISTE_TENSORWRAPPER_FUNC(bool, shape_read)
REGISTE_TENSORWRAPPER_FUNC(int64_t, mixin_handle)

#undef REGISTE_TENSORWRAPPER_FUNC


PyObject* TensorWrapper::handle() {
    return py::cast(m_tensor->m_handle).release().ptr();
}


void TensorWrapper::set_handle(PyObject* dest) {
    auto py_dest = py::reinterpret_borrow<py::object>(dest);
    SharedHandle real_dest = py_dest.cast<SharedHandle>();
    auto&& t = std::move(m_tensor->m_handle);
    m_tensor->m_handle = std::move(real_dest);
}


306
PyObject* TensorWrapper::shape() {
307 308 309
    if (!skip_tracing) {
        set_shape_read(py::cast(true).  release().ptr());
    }
310 311 312
    if (m_tensor->m_flags & Tensor::Flags::SCALAR) {
        return PyTuple_New(0);
    }
313 314 315 316 317 318 319 320

    TensorShape shape;
    if (m_tensor->m_var) {
        shape = m_tensor->m_var->shape();
    } else {
        shape = m_tensor->shape();
    }

321 322 323 324 325 326 327 328 329 330 331 332
    if (!shape.ndim) {
        Py_RETURN_NONE;
    }
    py::tuple ret(shape.ndim);
    for (size_t i = 0; i < shape.ndim; ++i) {
        ret[i] = shape[i];
    }
    return ret.release().ptr();
}


PyObject* TensorWrapper::dtype() {
333 334 335
    if (m_tensor->m_var) {
        return py::cast(m_tensor->m_var->dtype()).release().ptr();
    }
336 337 338 339 340
    return py::cast(m_tensor->dtype()).release().ptr();
}


PyObject* TensorWrapper::device() {
341 342 343
    if (m_tensor->m_var) {
        return py::cast(m_tensor->m_var->comp_node()).release().ptr();
    }
344 345 346 347 348
    return py::cast(m_tensor->comp_node()).release().ptr();
}


PyObject* TensorWrapper::numpy() {
349 350 351 352 353 354 355 356
    if (!skip_tracing) {
        set_value_read(py::cast(true).release().ptr());
    }
    if (m_tensor->m_handle.get() == nullptr && m_tensor->m_var != nullptr) {
        auto&& mgr = m_tensor->m_var->owner_graph()->static_infer_manager();
        auto&& type = mgr.get_infer_type(m_tensor->m_var);
        using InferType = cg::static_infer::InferType;
        if (!(type.value & (InferType::CONST | InferType::RT_STATIC))) {
357
            PyErr_SetString(PyExc_ValueError, "tensor invalid");
358 359 360 361
            return nullptr;
        }
        auto* val = mgr.infer_value_fallible(m_tensor->m_var);
        if (!val) {
362
            PyErr_SetString(PyExc_ValueError, "tensor invalid");
363 364 365 366
            return nullptr;
        }
        return py::cast(*val).attr("numpy")().release().ptr();
    }
367 368
    auto&& hv = interpreter_for_py->get_value(m_tensor->m_handle.get());
    auto arr = py::reinterpret_steal<py::array>(npy::ndarray_from_tensor(hv, npy::ShareType::TRY_SHARE));
369 370 371 372
    if (!arr) {
        PyErr_SetString(PyExc_ValueError, "tensor invalid");
        return nullptr;
    }
373 374 375 376 377 378 379
    if (m_tensor->m_flags & Tensor::Flags::SCALAR) {
        mgb_assert(PyArray_Check(arr.ptr()));
        return PyArray_Squeeze(reinterpret_cast<PyArrayObject*>(arr.ptr()));
    }
    return arr.release().ptr();
}

380 381 382 383
PyObject* TensorWrapper::varnode() {
    if (m_tensor->m_var) {
        return py::cast(m_tensor->m_var).release().ptr();
    }
384
    return py::none().release().ptr();
385 386
}

387
void TensorWrapper::reset(PyObject* tensor) {
388
    TensorWrapper* t = TensorWrapper::try_cast(tensor);
389 390 391 392 393 394
    if (!t) {
        throw py::type_error("expect Tensor");
    }
    m_tensor = t->m_tensor;
}

395 396 397 398
void TensorWrapper::reset_varnode() {
    m_tensor->m_var = nullptr;
}

399 400 401
PyObject* TensorWrapper::detach() {
    PyObject* self = wrap_t::pycast(this);
    PyTypeObject* pytype = self->ob_type;
402 403 404 405 406 407 408

    std::shared_ptr<Tensor> new_tensor;
    if (m_tensor->m_handle.get()) {
        new_tensor = std::make_shared<Tensor>(m_tensor->m_handle);
    } else {
        new_tensor = std::make_shared<Tensor>(m_tensor->m_var);
    }
409
    new_tensor->m_trace_info = m_tensor->m_trace_info;
410 411 412 413 414
    auto ret = TensorWrapper::make(pytype, std::move(new_tensor));
    return ret.release().ptr();

}

415
PyObject* TensorWrapper::_dev_tensor(){
416 417 418
    if (!skip_tracing) {
        set_data_read(py::cast(true).release().ptr());
    }
419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435
    auto dev_tensor = interpreter_for_py->get_dev_tensor(m_tensor->m_handle.get());
    return py::cast(dev_tensor).release().ptr();
}

void TensorWrapper::_swap_out() {
    interpreter_for_py->swap_out(m_tensor->m_handle.get());
}

void TensorWrapper::_swap_in() {
    interpreter_for_py->swap_in(m_tensor->m_handle.get());
}

void TensorWrapper::_drop() {
    interpreter_for_py->drop(m_tensor->m_handle.get());
}


436 437 438 439 440 441 442 443
PyObject* TensorWrapper::isscalar() {
    if(m_tensor->m_flags & Tensor::Flags::SCALAR) {
        Py_RETURN_TRUE;
    } else {
        Py_RETURN_FALSE;
    }
}

444

445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
void TensorWrapper::setscalar() {
    m_tensor->m_flags |= Tensor::Flags::SCALAR;
}


struct TensorWeakRef {
    std::weak_ptr<Tensor> wptr;

    TensorWeakRef(const TensorWrapper& tw) : wptr(tw.m_tensor) {}

    py::object operator()() {
        if (auto p = wptr.lock()) {
            return TensorWrapper::make(p);
        }
        return py::none();
    }
};

463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
/* ============== convert inputs ============== */

// map numpy.dtype.kind to priority
inline uint8_t category_priority(char c) {
    switch (c) {
        case 'f': return 3; // floating-point
        case 'i': return 2; // signed integer
        case 'u': return 2; // unsigned integer
        case 'b': return 1; // boolean
        default: return 0;
    }
}

// Returns the maximum value of the priority of each type in the list `types`.
uint8_t max_priority(SmallVector<PyArray_Descr*> types) {
    if (types.size() == 0) {
        return 0;
    } else {
        uint8_t max_p = 0;
        for (auto&& desc: types) {
            max_p = std::max(max_p, category_priority(desc->kind));
        }
        return max_p;
    }
}

489
// Returns the data type with sufficient size to hold all types of
490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549
// category `cat` in the list `types`.
PyArray_Descr* promote_types(SmallVector<PyArray_Descr*> types, uint8_t cat) {
    // Return value: New reference
    SmallVector<PyArray_Descr*> used_types;
    for (auto&& desc: types) {
        auto&& v = category_priority(desc->kind);
        if (v == cat) {
            used_types.emplace_back(desc);
        }
    }
    mgb_assert(used_types.size() > 0, "size of used_types is 0");
    PyArray_Descr* res = used_types[0];
    Py_INCREF(res);

    for (size_t i = 1; i < used_types.size(); ++i) {
        PyArray_Descr* tmp = PyArray_PromoteTypes(used_types[i], res);
        Py_DECREF(res);
        res = tmp;
    }
    return res;
}

PyArray_Descr* scalar2dtype(PyObject* arg) {
    // Return value: New reference
    if (PyBool_Check(arg)) {
        auto&& descr = PyArray_DescrFromType(NPY_BOOL);
        return descr;
    }
    if (PyLong_CheckExact(arg)) {
        auto&& descr = PyArray_DescrFromType(NPY_INT32);
        return descr;
    }
    if (PyFloat_CheckExact(arg)) {
        auto&& descr = PyArray_DescrFromType(NPY_FLOAT32);
        return descr;
    }
    return nullptr;
}

PyArray_Descr* _dtype_promotion(PyObject*const* args, size_t nargs) {
    // Return value: New reference
    SmallVector<PyArray_Descr*> tensors;
    SmallVector<PyArray_Descr*> scalars;

    bool is_tuple = false;
    PyObject* tuple;
    if (nargs == 1 && (PyTuple_Check(args[0]) || PyList_Check(args[0]))) {
        if (PyList_Check(args[0])) {
            tuple = PyList_AsTuple(args[0]);
        } else {
            tuple = args[0];
            Py_INCREF(tuple);
        }
        nargs = PyTuple_Size(tuple);
        is_tuple = true;
    }

    for (size_t i = 0; i < nargs; ++i) {
        PyObject* handle = is_tuple ? PyTuple_GetItem(tuple, i): args[i];
        if (handle == Py_None) continue;
550
        TensorWrapper* tw = TensorWrapper::try_cast(handle);
551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604
        if (tw) {
            mgb::DType type = tw->m_tensor->dtype();
            auto&& descr = npy::dtype_mgb2np_descr(type);
            Py_INCREF(descr.get());
            tensors.emplace_back(descr.get());
        }else{
            if (PyArray_Check(handle) || PyArray_CheckScalar(handle)) {
                auto&& descr = PyArray_DescrFromObject(handle, nullptr);
                tensors.emplace_back(descr);
                continue;
            }
            PyArray_Descr* descr = scalar2dtype(handle);
            if (descr) {
                scalars.emplace_back(descr);
                continue;
            }
        }
    }

    auto max_pri_scalars = max_priority(scalars);
    auto max_pri_tensors = max_priority(tensors);

    if (max_pri_scalars <= 0 && max_pri_tensors <= 0) {
        throw py::value_error("invalid input, no dtype avaliable");
    }
    PyArray_Descr* res;
    if (max_pri_scalars > max_pri_tensors) {
        res = promote_types(scalars, max_pri_scalars);
    }else{
        res = promote_types(tensors, max_pri_tensors);
    }
    for (auto *p: tensors) { Py_DECREF(p); }
    for (auto *p: scalars) { Py_DECREF(p); }
    Py_DECREF(tuple);
    return res;
}

CompNode _get_device(PyObject*const* args, size_t nargs) {
    bool is_tuple = false;
    PyObject* tuple;
    if (nargs == 1 && (PyTuple_Check(args[0]) || PyList_Check(args[0]))) {
        if (PyList_Check(args[0])) {
            tuple = PyList_AsTuple(args[0]);
        } else {
            tuple = args[0];
            Py_INCREF(tuple);
        }
        nargs = PyTuple_Size(tuple);
        is_tuple = true;
    }
    bool valid = false;
    CompNode cn;
    for (size_t i = 0; i < nargs; ++i) {
        PyObject* handle = is_tuple ? PyTuple_GetItem(tuple, i): args[i];
605
        TensorWrapper* tw = TensorWrapper::try_cast(handle);
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
        if (tw) {
            if (!valid) {
                cn = tw->m_tensor->comp_node();
                valid = true;
            } else {
                CompNode cn1 = tw->m_tensor->comp_node();
                if (cn1 != cn) {
                    throw py::value_error(ssprintf("ambiguous device: %s vs %s",
                        cn.to_string().c_str(), cn1.to_string().c_str()));
                }
            }
        }
    }
    if (!valid) {
        mgb_assert(0, "expact at least 1 device");
    }
    Py_DECREF(tuple);
    return cn;
}

// Returns the dtype that would result from performing an arithmetic
// operation on the provided input tensors and scalars.
PyObject* dtype_promotion(PyObject* self, PyObject*const* args, size_t nargs) {
    if (!nargs) {
        PyErr_SetString(PyExc_TypeError, "empty input is not allowed");
        return nullptr;
    }
    try {
        PyArray_Descr* res = _dtype_promotion(args, nargs);
        return py::cast(npy::dtype_np2mgb_descr(res)).release().ptr();
    } catch (std::exception& e) {
        PyErr_SetString(PyExc_RuntimeError, e.what());
        return nullptr;
    }
}

PyObject* get_device(PyObject* self, PyObject*const* args, size_t nargs) {
    if (!nargs) {
        PyErr_SetString(PyExc_TypeError, "empty input is not allowed");
        return nullptr;
    }
    try {
        CompNode cn = _get_device(args, nargs);
        return py::cast(cn).release().ptr();
    } catch (std::exception& e) {
        PyErr_SetString(PyExc_RuntimeError, e.what());
        return nullptr;
    }
}
655

656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
#ifdef METH_FASTCALL
#define MGE_PY_INTERFACE(NAME, FUNC) \
    { #NAME, (PyCFunction)FUNC, METH_FASTCALL, nullptr }
#else
#define WRAP_FUNC_PY35(FUNC)                                \
    PyObject* py35_##FUNC(PyObject* self, PyObject* args) { \
        auto* arr = &PyTuple_GET_ITEM(args, 0);             \
        auto size = PyTuple_GET_SIZE(args);                 \
        return FUNC(self, arr, size);                       \
    }
WRAP_FUNC_PY35(py_apply);
WRAP_FUNC_PY35(dtype_promotion);
WRAP_FUNC_PY35(get_device);
#undef WRAP_FUNC_PY35
#define MGE_PY_INTERFACE(NAME, FUNC) \
    { #NAME, (PyCFunction)py35_##FUNC, METH_VARARGS, nullptr }
#endif

674 675 676 677
py::object make_empty_tensorwrapper() {
    return TensorWrapper::make(std::move(std::make_shared<Tensor>()));
}

678 679 680 681 682 683 684 685 686 687 688
void init_tensor(py::module m) {
    interpreter_for_py = interpreter::Interpreter::inst().create_channel();

    auto* tensor_type = TensorWrapper::wrap_t::type()
        .def<&TensorWrapper::numpy>("numpy")
        .def_getset<&TensorWrapper::shape>("shape")
        .def_getset<&TensorWrapper::dtype>("dtype")
        .def_getset<&TensorWrapper::device>("device")
        .def<&TensorWrapper::reset>("_reset")
        .def<&TensorWrapper::isscalar>("isscalar")
        .def<&TensorWrapper::setscalar>("setscalar")
689
        .def<&TensorWrapper::detach>("detach")
690 691 692 693
        .def<&TensorWrapper::_dev_tensor>("_dev_tensor")
        .def<&TensorWrapper::_swap_out>("_swap_out")
        .def<&TensorWrapper::_swap_in>("_swap_in")
        .def<&TensorWrapper::_drop>("_drop")
694
        .def<&TensorWrapper::reset_varnode>("_reset_varnode")
695 696 697 698 699 700
        .def_getset<&TensorWrapper::varnode>("_varnode")
        .def_getset<&TensorWrapper::data_read, &TensorWrapper::set_data_read>("data_read")
        .def_getset<&TensorWrapper::value_read, &TensorWrapper::set_value_read>("value_read")
        .def_getset<&TensorWrapper::shape_read, &TensorWrapper::set_shape_read>("shape_read")
        .def_getset<&TensorWrapper::mixin_handle, &TensorWrapper::set_mixin_handle>("mixin_handle")
        .def_getset<&TensorWrapper::handle, &TensorWrapper::set_handle>("_handle")
701 702 703 704 705 706 707 708
        .finalize();
    if (!tensor_type) throw py::error_already_set();
    py::setattr(m, "Tensor", tensor_type);

    py::class_<TensorWeakRef>(m, "TensorWeakRef")
        .def(py::init<const TensorWrapper&>())
        .def("__call__", &TensorWeakRef::operator());

709
    static PyMethodDef method_defs[] = {
710 711 712 713
            MGE_PY_INTERFACE(apply, py_apply),
            MGE_PY_INTERFACE(dtype_promotion, dtype_promotion),
            MGE_PY_INTERFACE(get_device, get_device),
            {nullptr, nullptr, 0, nullptr}};
714 715 716 717 718 719 720
    for (auto&& def: method_defs) {
        if (def.ml_meth != nullptr) {
            auto* func = PyCFunction_NewEx(&def, nullptr, nullptr);
            if (!func) throw py::error_already_set();
            py::setattr(m, def.ml_name, func);
        }
    }
721

722 723 724 725 726 727 728 729 730 731 732 733 734 735
    m.def("_set_swap_flag",
          [](bool flag) { interpreter_for_py->set_swap_flag(flag); });
    m.def("_set_drop_flag",
          [](bool flag) { interpreter_for_py->set_drop_flag(flag); });
    m.def("config_async_level",
          [](int level) { interpreter_for_py->config_async_level(level); });
    m.def("get_async_level",
          []() { return interpreter_for_py->get_async_level(); });
    m.def("sync",
          []() {
              interpreter_for_py->sync();
              py_task_q.wait_all_task_finish();
          },
          py::call_guard<py::gil_scoped_release>());
736

737
    m.def("release_trace_apply_func", &release_trace_apply_func);
738

739 740
    py::handle grad_key_type = GradKeyWrapper::wrap_t::type()
        .def<&GradKeyWrapper::attach>("attach")
741 742
        .def<&GradKeyWrapper::is_attached_to>("is_attached_to")
        .def_getset<&GradKeyWrapper::get_name, &GradKeyWrapper::set_name>("name")
743 744 745
        .finalize();
    if (!grad_key_type) throw py::error_already_set();
    py::setattr(m, "GradKey", grad_key_type);
746 747
    m.def("backward", &GradKeyWrapper::backward);

748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765
    m.def("set_cpp_apply_with_tracing", &set_cpp_apply_with_tracing);
    m.def("set_cpp_apply_const_with_tracing", &set_cpp_apply_const_with_tracing);
    m.def("set_cpp_apply_compiled_mode", &set_cpp_apply_compiled_mode);
    m.def("set_cpp_apply_const_compiled_mode", &set_cpp_apply_const_compiled_mode);
    m.def("set_cpp_apply_backward_varnode", &set_cpp_apply_backward_varnode);

    m.attr("skip_tracing") = &skip_tracing;

    py::class_<SharedHandle>(m, "SharedHandle")
        .def(py::init<const SharedHandle&>());

    m.def("set_tracing", &set_tracing);
    m.def("unset_tracing", &unset_tracing);
    m.def("set_symbolic", &set_symbolic);
    m.def("unset_symbolic", &unset_symbolic);
    m.def("set_compiled", &set_compiled);
    m.def("unset_compiled", &unset_compiled);

766
    m.def("__make_empty_tensor", &make_empty_tensorwrapper);
767 768
}

769 770
#undef MGE_PY_INTERFACE

771
} // namespace mgb::imperative::python