tensor_utils.cpp 59.4 KB
Newer Older
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
/**
 * \file imperative/python/src/tensor.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 "megbrain/common.h"
#include "megbrain/dtype.h"
#include "megbrain/imperative/ops/autogen.h"
#include "megbrain/imperative/ops/backward_graph.h"
#include "megbrain/imperative/ops/utility.h"
#include "megbrain/imperative/profiler.h"
#include "megbrain/imperative/transformations/eval.h"
#include "megbrain/imperative/transformations/lazy.h"
#include "megbrain/imperative/transformations/scalar.h"
#include "megbrain/imperative/transformations/symbol.h"
#include "megbrain/imperative/transformations/trace.h"
#include "megbrain/imperative/utils/map.h"
#include "megbrain/imperative/utils/stats.h"
#include "megbrain/opr/io.h"
#include "megbrain/plugin/profiler.h"

#include "./common.h"
#include "./grad.h"
#include "./graph_rt.h"
#include "./helper.h"
#include "./module_trace.h"
#include "./numpy_dtypes.h"
#include "./tensor.h"
#include "./tensor_utils.h"
#include "./transformation.h"

#include <object.h>
#include <pybind11/numpy.h>
#include <pybind11/operators.h>
#include <pybind11/pytypes.h>
#include <pyerrors.h>
#include <range/v3/all.hpp>
#include <string>

#include <unordered_map>

#include "../../src/impl/mgb_cg_impl.h"

namespace py = pybind11;
namespace views = ranges::views;

namespace mgb::imperative::python {

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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271
/* ============== 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;
    }
}

// Returns the data type with sufficient size to hold all types of
// 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 = nullptr;
    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;
        TensorWrapper* tw = TensorWrapper::try_cast(handle);
        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;
            }

            if (py::isinstance<PySymbolVar>(py::handle(handle))) {
                auto var = py::handle(handle).cast<PySymbolVar*>();
                mgb::DType type = var->m_node->dtype();
                auto&& descr = npy::dtype_mgb2np_descr(type);
                Py_INCREF(descr.get());
                tensors.emplace_back(descr.get());
                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_XDECREF(tuple);
    return res;
}

CompNode _get_device(PyObject* const* args, size_t nargs) {
    bool is_tuple = false;
    PyObject* tuple = nullptr;
    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];
        TensorWrapper* tw = TensorWrapper::try_cast(handle);

        bool is_symvar = py::isinstance<PySymbolVar>(py::handle(handle));
        if (tw || is_symvar) {
            if (!valid) {
                cn = tw ? tw->m_tensor->comp_node()
                        : py::handle(handle).cast<PySymbolVar*>()->m_node->comp_node();
                valid = true;
            } else {
                CompNode cn1 = tw ? tw->m_tensor->comp_node()
                                  : py::handle(handle)
                                               .cast<PySymbolVar*>()
                                               ->m_node->comp_node();
                if (cn1 != cn) {
                    throw py::value_error(ssprintf(
                            "ambiguous device: %s (from %s) vs %s (from %s)",
                            cn.to_string().c_str(), cn.to_string_logical().c_str(),
                            cn1.to_string().c_str(), cn1.to_string_logical().c_str()));
                }
            }
        }
    }
    if (!valid) {
        return CompNode::load(get_default_device());
    }
    Py_XDECREF(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();
    }
    PYEXT17_TRANSLATE_EXC_RET(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();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312
bool is_scalar(PyObject* tensor) {
    if (py::isinstance<PySymbolVar>(py::handle(tensor))) {
        auto var = py::handle(tensor).cast<PySymbolVar*>();
        return var->is_scalar;
    }
    auto* tw = TensorWrapper::try_cast(tensor);
    if (tw) {
        return tw->m_tensor->is_scalar();
    }
    return PyArray_CheckAnyScalar(tensor);
}

bool is_bool_list(PyObject* arg) {
    if (!PyList_Check(arg)) {
        return false;
    }
    size_t sz = PyList_Size(arg);
    if (!sz) {
        return false;
    }
    for (size_t i = 0; i < sz; ++i) {
        PyObject* handle = PyList_GetItem(arg, i);
        if (!PyBool_Check(handle)) {
            return false;
        }
    }
    return true;
}

bool is_bool_dtype(PyObject* args) {
    if (!PyObject_HasAttrString(args, "dtype"))
        return false;
    PyObject* dobj = PyObject_GetAttrString(args, "dtype");
    PyArray_Descr* dtype;
    PyArray_DescrConverter(dobj, &dtype);
    bool ret = (dtype->kind == 'b');
    Py_XDECREF(dtype);
    Py_XDECREF(dobj);
    return ret;
}

313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333
py::object device2obj(py::handle device, bool mapping = false) {
    if (device.ptr() == Py_None) {
        return py::cast(CompNode::load(get_default_device()));
    } else if (py::isinstance<py::str>(device)) {
        if (mapping) {
            py::object dmap = getattr(
                    py::reinterpret_borrow<py::object>((PyObject*)py_tensor_type),
                    "dmap_callback");
            if (dmap.ptr() != Py_None) {
                return device2obj(dmap(device), false);
            }
        }
        return py::cast(CompNode::load(device.cast<std::string>()));

    } else if (py::isinstance<CompNode>(device)) {
        return py::reinterpret_borrow<py::object>(device);
    } else {
        return getattr(device, "_cn");
    }
}

334
py::object _Const(
335
        py::handle value, py::handle dtype, py::handle device, py::handle ref_hdl) {
336 337 338 339 340 341 342 343 344 345 346 347
    py::object val = py::reinterpret_borrow<py::object>(value);
    if (PyArray_Check(value.ptr())) {
        py::tuple strides =
                py::reinterpret_borrow<py::tuple>(getattr(value, "strides"));
        bool need_squeeze = false;
        for (size_t i = 0; i < strides.size(); ++i) {
            if (strides[i].cast<ptrdiff_t>() == 0) {
                need_squeeze = true;
            }
        }
        if (need_squeeze) {
            val = py::reinterpret_borrow<py::array>(value);
348
            py::object orig_shp = val.attr("shape");
349
            val = val.attr("squeeze")();
350
            val = val.attr("reshape")(orig_shp);
351 352
        }
    }
353 354 355 356 357 358 359 360 361 362 363
    py::object ref;
    if (py::isinstance<py::tuple>(ref_hdl)) {
        py::tuple tup = py::reinterpret_borrow<py::tuple>(ref_hdl);
        if (tup.size()) {
            ref = tup[0];
        } else {
            ref = py::none();
        }
    } else {
        ref = py::reinterpret_borrow<py::object>(ref_hdl);
    }
364 365 366
    if (py::isinstance<PySymbolVar>(ref)) {
        auto ref_var = ref.cast<PySymbolVar*>();
        auto* graph = ref_var->m_node->owner_graph();
367 368 369 370
        CompNode cn;
        if (device.ptr() == Py_None) {
            cn = ref_var->m_node->comp_node();
        } else {
371
            cn = device2obj(device).cast<CompNode>();
372
        }
373 374 375 376 377 378
        OperatorNodeConfig config(cn);
        auto hv = npy::np2tensor(
                val.ptr(), npy::Meth::borrow(cn), dtype.cast<mgb::DType>());
        auto typeobj = ref.get_type();
        return typeobj(opr::ImmutableTensor::make(*graph, hv, config).node());
    }
379
    py::object device_obj = device2obj(device, true);
380
    py::tuple tup = py::make_tuple(val, dtype, device_obj, true, false, py::none());
381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
    return TensorWrapper::make(py_tensor_type, tup.ptr(), nullptr);
}

py::tuple _make_shape_tuple(py::handle shape) {
    py::list orig;
    py::list ret(0);
    auto solve_one = [&](py::handle val) {
        if (TensorWrapper::try_cast(val.ptr()) || py::isinstance<PySymbolVar>(val)) {
            py::object np = getattr(val, "numpy")();
            PyArrayObject* arr = (PyArrayObject*)np.ptr();
            PyObject* maybe_list = PyArray_ToList(arr);
            if (PyList_Check(maybe_list)) {
                py::list may = py::reinterpret_steal<py::list>(maybe_list);
                for (size_t i = 0; i < may.size(); ++i) {
                    ret.append(may[i]);
                }
            } else {
                mgb_assert(PyLong_Check(maybe_list));
                ret.append(PyLong_AsLong(maybe_list));
                Py_XDECREF(maybe_list);
            }
        } else if (PyArray_Check(val.ptr())) {
            ret.append(PyArray_PyIntAsInt(val.ptr()));
        } else {
            ret.append(PyLong_AsLong(val.ptr()));
        }
    };
    if (PyArray_Check(shape.ptr()) && !PyArray_CheckAnyScalar(shape.ptr())) {
        orig = py::reinterpret_steal<py::list>(
                PyArray_ToList((PyArrayObject*)shape.ptr()));
        for (size_t i = 0; i < orig.size(); ++i) {
            solve_one(orig[i]);
        }
    } else if (PyList_Check(shape.ptr())) {
        orig = py::reinterpret_borrow<py::list>(shape);
        for (size_t i = 0; i < orig.size(); ++i) {
            solve_one(orig[i]);
        }
    } else if (PyTuple_Check(shape.ptr())) {
        py::tuple tup = py::reinterpret_borrow<py::tuple>(shape);
        for (size_t i = 0; i < tup.size(); ++i) {
            solve_one(tup[i]);
        }
    } else {
        solve_one(shape);
    }
    return py::reinterpret_steal<py::tuple>(PyList_AsTuple(ret.ptr()));
}

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
bool is_tensor_or_symbolvar(py::handle arg) {
    return bool(TensorWrapper::try_cast(arg.ptr())) || py::isinstance<PySymbolVar>(arg);
}

bool is_py_sequence(py::handle arg) {
    if (PyArray_Check(arg.ptr()) || TensorWrapper::try_cast(arg.ptr()) ||
        py::isinstance<PySymbolVar>(arg)) {
        return false;
    }
    return PySequence_Check(arg.ptr());
}

mgb::DType _get_dtype(py::handle tensor) {
    if (auto tw = TensorWrapper::try_cast(tensor.ptr())) {
        return tw->m_tensor->dtype();
    } else {
        auto var = tensor.cast<PySymbolVar*>();
        return var->m_node->dtype();
    }
}

py::object _astype_cpp(py::handle tensor, py::handle dtype_hdl) {
    PyArray_Descr* descr;
    if (!PyArray_DescrConverter(dtype_hdl.ptr(), &descr)) {
        throw py::value_error(ssprintf(
                "can not convert to numpy.dtype from %s",
                dtype_hdl.ptr()->ob_type->tp_name));
    }
    PyArray_Descr* cur = npy::dtype_mgb2np_descr(_get_dtype(tensor)).get();
    if (!dtype_equal(cur, descr)) {
        std::shared_ptr<OpDef> op = TypeCvt::make(npy::dtype_np2mgb_descr(descr));
        py::object Op = py::cast(op);
462 463
        PyObject* p[2] = {Op.ptr(), tensor.ptr()};
        py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512
        return ret[0];
    } else {
        return py::reinterpret_borrow<py::object>(tensor);
    }
}

py::object _convert_single_value_cpp(
        py::handle value, py::handle dtype, py::handle device) {
    if (is_tensor_or_symbolvar(value)) {
        if (_get_dtype(value).category() != DTypeCategory::QUANTIZED) {
            return _astype_cpp(value, dtype);
        }
    } else {
        return _Const(value, dtype, device, py::none());
    }
    return py::reinterpret_borrow<py::object>(value);
}

py::object _convert_inputs_cpp(
        PyObject* const* args, size_t nargs, py::object dtype, py::object device) {
    ComputingGraph* graph = nullptr;
    py::handle typeobj;
    py::list lis;
    for (size_t i = 0; i < nargs; ++i) {
        py::handle h = py::handle(args[i]);
        lis.append(h);
        if (py::isinstance<PySymbolVar>(h)) {
            auto var = h.cast<PySymbolVar*>();
            auto g = var->m_node->owner_graph();
            if (!graph) {
                graph = g;
                typeobj = h.get_type();
            } else {
                mgb_assert(graph == g);
            }
        }
    }
    if (graph) {
        CompNode cn = device2obj(device).cast<CompNode>();
        for (size_t i = 0; i < nargs; ++i) {
            OperatorNodeConfig config(cn);
            auto hv = npy::np2tensor(
                    lis[i].ptr(), npy::Meth::borrow(cn), dtype.cast<mgb::DType>());
            if (!py::isinstance<PySymbolVar>(lis[i])) {
                lis[i] = typeobj(opr::ImmutableTensor::make(*graph, hv, config).node());
            }
        }
    }
    auto convert = [&](py::object value) {
513
        if (value.is_none()) {
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
            return value;
        }
        return _convert_single_value_cpp(value, dtype, device);
    };
    for (size_t i = 0; i < lis.size(); ++i) {
        lis[i] = convert(lis[i]);
    }
    return py::reinterpret_steal<py::tuple>(PyList_AsTuple(lis.ptr()));
}

py::object _astensor1d_cpp(
        py::handle value, py::handle dtype, py::handle device, py::handle ref) {
    py::object ret;
    py::object device_obj = py::none();
    py::object ndim_obj = py::none();
    if (device.ptr() != Py_None) {
        device_obj = device2obj(device);
    }
    if (py::isinstance<PySymbolVar>(value)) {
        try {
            getattr(value, "ndim");
        } catch (py::error_already_set& err) {
            if (dtype.ptr() != Py_None) {
                ret = _astype_cpp(value, dtype);
            } else {
                ret = py::reinterpret_borrow<py::object>(value);
            }
            if (device.ptr() != Py_None) {
                std::shared_ptr<OpDef> op = Copy::make(device_obj.cast<CompNode>());
                py::object Op = py::cast(op);
544 545 546
                PyObject* p[2] = {Op.ptr(), ret.ptr()};
                py::tuple copy_ret =
                        py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585
                return copy_ret[0];
            }
            return ret;
        }
    }
    size_t ndim = 999;
    if (hasattr(value, "ndim")) {
        ndim = getattr(value, "ndim").cast<size_t>();
        if (ndim != 0 && ndim != 1) {
            throw py::value_error("ndim != 1 or 0, get : " + std::to_string(ndim));
        }
        if (!is_tensor_or_symbolvar(value)) {
            return _Const(value, dtype, device, ref);
        } else {
            return py::reinterpret_borrow<py::object>(value);
        }
    }
    if (!is_py_sequence(value)) {
        throw py::type_error();
    }
    py::list lis = py::reinterpret_steal<py::list>(PySequence_List(value.ptr()));
    bool need_concat = false;
    for (size_t i = 0; i < lis.size(); ++i) {
        if (is_tensor_or_symbolvar(lis[i])) {
            need_concat = true;
            break;
        }
    }
    if (!need_concat) {
        return _Const(value, dtype, device, ref);
    }
    if (lis.size() > 1) {
        std::vector<PyObject*> c_args(lis.size() + 1);
        for (size_t i = 0; i < lis.size(); ++i) {
            c_args[i] = lis[i].ptr();
        }
        c_args[lis.size()] = Py_None;
        py::tuple inp_tup = py::reinterpret_steal<py::tuple>(
                convert_inputs_cpp(NULL, c_args.data(), c_args.size()));
586
        if (device_obj.is_none()) {
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
            std::vector<PyObject*> inp(inp_tup.size());
            for (size_t i = 0; i < inp_tup.size(); ++i) {
                inp[i] = inp_tup[i].ptr();
            }
            device_obj = py::cast(_get_device(inp.data(), inp.size()));
        }
        std::shared_ptr<OpDef> op = Concat::make(0, device_obj.cast<CompNode>());
        py::object Op = py::cast(op);
        std::vector<PyObject*> p;
        p.resize(inp_tup.size() + 1);
        p[0] = Op.ptr();
        for (size_t i = 0; i < inp_tup.size(); ++i) {
            p[i + 1] = inp_tup[i].ptr();
        }
        py::tuple concat_ret =
                py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
        ret = concat_ret[0];
    } else {
        ret = lis[0];
    }
    if (dtype.ptr() != Py_None) {
        return _astype_cpp(ret, dtype);
    } else {
        return ret;
    }
}

614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
py::object _get_index(py::object tensor, py::object src) {
    if (!TensorWrapper::try_cast(tensor.ptr()) &&
        !py::isinstance<PySymbolVar>(tensor)) {
        auto get_const = [&](mgb::DType dtype) -> py::object {
            return _Const(tensor, py::cast(dtype), src.attr("device"), src);
        };
        if (is_bool_list(tensor.ptr()) || is_bool_dtype(tensor.ptr())) {
            tensor = get_const(dtype::Bool());
        } else {
            tensor = get_const(dtype::Int32());
        }
        if (!is_bool_dtype(tensor.ptr())) {
            return tensor;
        }
    } else {
        if (!is_bool_dtype(tensor.ptr())) {
            return tensor;
        }
    }
633
    std::shared_ptr<OpDef> op = CondTake::make();
634
    py::object Op = py::cast(op);
635 636
    PyObject* p[3] = {Op.ptr(), tensor.ptr(), tensor.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 3));
637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656
    return ret[1];
}

py::tuple _try_cond_take(py::handle tensor, py::handle index) {
    if (!hasattr(index, "dtype") || !hasattr(index, "shape")) {
        return py::tuple();
    }
    if (!is_bool_dtype(index.ptr()) ||
        _make_shape_tuple(getattr(index, "shape"))
                .not_equal(_make_shape_tuple(getattr(tensor, "shape")))) {
        return py::tuple();
    }
    py::object iobj;
    if (PyArray_Check(index.ptr())) {
        iobj =
                _Const(index, py::cast((mgb::DType)dtype::Bool()),
                       getattr(tensor, "device"), tensor);
    } else {
        iobj = py::reinterpret_borrow<py::object>(index);
    }
657
    std::shared_ptr<OpDef> op = CondTake::make();
658
    py::object Op = py::cast(op);
659 660
    PyObject* p[3] = {Op.ptr(), tensor.ptr(), iobj.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 3));
661 662 663 664 665 666 667 668 669 670
    return ret;
}

py::tuple _remove_ellipsis(py::object tensor, py::tuple tuple_val) {
    size_t tuple_size = tuple_val.size();
    size_t ndim_sum = 0, cur_sum = 0;
    int pos = -1;
    bool has_unknown_ndim_bool_index = false;
    for (size_t i = 0; i < tuple_size; ++i) {
        py::object handle = tuple_val[i];
671 672 673
        if (handle.is_none()) {
            continue;
        } else if (handle.ptr() == Py_Ellipsis) {
674 675 676 677 678 679 680 681 682 683 684
            pos = static_cast<int>(i);
            for (size_t j = 0; j < i; ++j) {
                py::object t = tuple_val[j];
                if (t.ptr() == Py_Ellipsis) {
                    throw py::index_error("only one ellipsis is allowed.");
                }
            }
        } else {
            size_t ndim_incr = 1;
            if (hasattr(handle, "dtype") && is_bool_dtype(handle.ptr()) &&
                hasattr(handle, "ndim")) {
685 686 687 688 689 690
                py::object ndim;
                try {
                    ndim = getattr(handle, "ndim");
                } catch (py::error_already_set& err) {
                    has_unknown_ndim_bool_index = true;
                }
691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728
                if (PyLong_Check(ndim.ptr())) {
                    ndim_incr = PyLong_AsLong(ndim.ptr());
                } else {
                    has_unknown_ndim_bool_index = true;
                }
            }
            cur_sum += ndim_incr;
        }
    }
    if (pos == -1) {
        return tuple_val;
    } else {
        if (has_unknown_ndim_bool_index) {
            throw py::index_error(
                    "does not support bool index with unknown shape when using "
                    "Ellipsis.");
        }
        try {
            ndim_sum = getattr(tensor, "ndim").cast<size_t>();
        } catch (py::error_already_set& err) {
            throw py::index_error(
                    "does not support Ellipsis when tensor's ndim is unknown.");
        }
        py::tuple ret(ndim_sum - cur_sum + tuple_size - 1);
        size_t idx = 0;
        for (size_t i = 0; i < tuple_size; ++i) {
            if (i == pos) {
                for (size_t j = cur_sum; j < ndim_sum; ++j) {
                    ret[idx++] = PySlice_New(NULL, NULL, NULL);
                }
            } else {
                ret[idx++] = tuple_val[i];
            }
        }
        return ret;
    }
}

729 730
py::object _reshape_cpp(py::handle inp_hdl, py::handle args);

731 732 733 734 735 736
py::tuple _expand_bool_dim(py::object tensor, py::tuple tuple_val) {
    py::tuple cur_shape = _make_shape_tuple(py::handle(getattr(tensor, "shape")));
    py::list new_tuple_val(0);

    size_t offset = 0;
    size_t tdim = 0;
737
    size_t nonedim = 0;
738 739
    for (size_t i = 0; i < tuple_val.size(); ++i) {
        py::handle k = tuple_val[i];
740 741 742 743 744
        if (k.ptr() == Py_None) {
            nonedim++;
            new_tuple_val.append(k);
            continue;
        }
745 746 747 748 749 750 751 752
        if (is_bool_dtype(k.ptr())) {
            size_t ndim = getattr(k, "ndim").cast<size_t>();
            if (ndim > 1) {
                py::tuple ishape = _make_shape_tuple(py::handle(getattr(k, "shape")));
                for (size_t j = 0; j < ndim; ++j) {
                    if (cur_shape[tdim + j - offset].cast<size_t>() !=
                        ishape[j].cast<size_t>()) {
                        std::string msg =
753 754
                                "boolean index did not match tensor along "
                                "dimension " +
755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770
                                std::to_string(tdim + j) + "; dimension is " +
                                std::to_string(
                                        cur_shape[tdim + j - offset].cast<size_t>()) +
                                " but corresponding boolean dimension is " +
                                std::to_string(ishape[j].cast<size_t>());
                        throw py::index_error(msg.c_str());
                    }
                }
                py::object new_k = getattr(k, "reshape")(-1);
                py::object kshape = getattr(new_k, "shape");
                py::list new_shape(0);
                PyObject* sym = PyObject_CallObject(cpp_use_symbolic_shape, nullptr);
                bool is_sym = (sym == Py_True);
                Py_XDECREF(sym);
                if (is_sym) {
                    py::object tshape = getattr(tensor, "shape");
771
                    for (size_t j = 0; j < i - nonedim; ++j) {
772 773 774 775 776 777
                        new_shape.append(tshape[py::int_(j)]);
                    }
                    new_shape.append(kshape[py::int_(0)]);
                    for (size_t j = tdim + ndim - offset; j < cur_shape.size(); ++j) {
                        new_shape.append(cur_shape[j]);
                    }
778 779 780 781
                    py::object shape_tensor = _astensor1d_cpp(
                            new_shape, py::none(), py::none(), py::none());
                    tensor = _reshape_cpp(tensor, shape_tensor);
                    cur_shape = _make_shape_tuple(shape_tensor);
782
                } else {
783
                    for (size_t j = 0; j < i - nonedim; ++j) {
784 785 786 787 788 789 790
                        new_shape.append(cur_shape[j]);
                    }
                    new_shape.append(py::reinterpret_borrow<py::tuple>(kshape)[0]);
                    for (size_t j = tdim + ndim - offset; j < cur_shape.size(); ++j) {
                        new_shape.append(cur_shape[j]);
                    }
                    cur_shape = new_shape;
791
                    tensor = _reshape_cpp(tensor, cur_shape);
792 793 794 795 796 797 798 799 800 801 802 803 804
                }
                offset++;
                tdim += ndim;
            }
            new_tuple_val.append(k);
        } else {
            new_tuple_val.append(k);
            tdim++;
        }
    }
    return py::make_tuple(tensor, py::reinterpret_borrow<py::tuple>(new_tuple_val));
}

805 806 807 808 809 810 811 812 813 814 815 816
std::pair<size_t, bool> get_ndim_safe(py::handle tensor) {
    if (auto p = TensorWrapper::try_cast(tensor.ptr())) {
        return {p->m_tensor->shape()->ndim, true};
    }

    try {
        return {getattr(tensor, "ndim").cast<size_t>(), true};
    } catch (py::error_already_set& err) {
        return {0, false};
    }
}

817 818 819 820 821 822 823 824 825 826 827 828 829 830 831
py::tuple _unpack_indexes(py::handle inp_hdl, py::handle idx_hdl) {
    py::object inp = py::reinterpret_borrow<py::object>(inp_hdl);
    py::tuple tuple_val;
    if (py::isinstance<py::tuple>(idx_hdl)) {
        tuple_val = py::reinterpret_borrow<py::tuple>(idx_hdl);
    } else {
        tuple_val = py::make_tuple(idx_hdl);
    }

    bool use_subtensor = true;
    bool need_remove_ellipsis = false;
    bool need_expand_bool_dim = false;
    size_t idx_ndim = 0;
    for (size_t i = 0; i < tuple_val.size(); ++i) {
        py::object k = tuple_val[i];
832 833
        if (k.is_none()) {
            continue;
834 835 836 837
        } else if (k.ptr() == Py_Ellipsis) {
            need_remove_ellipsis = true;
        } else {
            if (is_bool_dtype(k.ptr()) && hasattr(k, "ndim")) {
838
                size_t ndim = get_ndim_safe(k).first;
839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871
                idx_ndim += ndim;
                if (ndim > 1) {
                    need_expand_bool_dim = true;
                }
            } else {
                idx_ndim++;
            }
        }
    }
    try {
        size_t inp_ndim = getattr(inp, "ndim").cast<size_t>();
        if (idx_ndim > inp_ndim) {
            std::string msg = "too many indices for tensor: tensor is " +
                              std::to_string(inp_ndim) + "-dimensional, but " +
                              std::to_string(idx_ndim) + " were indexed";
            throw py::index_error(msg.c_str());
        }
    } catch (py::error_already_set& err) {
        ;  // ignore
    }
    if (need_remove_ellipsis) {
        tuple_val = _remove_ellipsis(inp, tuple_val);
    }

    if (need_expand_bool_dim) {
        py::object shape = getattr(inp, "shape");
        if (shape.ptr() != Py_None) {
            py::tuple ret = _expand_bool_dim(inp, tuple_val);
            inp = ret[0];
            tuple_val = ret[1];
        }
    }

872 873 874 875 876 877 878 879 880 881 882 883 884 885
    std::vector<int32_t> axis;
    for (size_t i = 0; i < tuple_val.size(); ++i) {
        if (tuple_val[i].is_none()) {
            axis.push_back(i);
        }
    }
    if (axis.size()) {
        std::shared_ptr<OpDef> op = AddAxis::make(axis);
        py::object Op = py::cast(op);
        PyObject* p[2] = {Op.ptr(), inp.ptr()};
        py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
        inp = ret[0];
    }

886 887 888 889 890 891 892
    py::list items;
    py::list tensors;
    int cur_axis = -1;

    for (size_t i = 0; i < tuple_val.size(); ++i) {
        py::object handle = tuple_val[i];
        cur_axis++;
893 894 895
        if (handle.is_none()) {
            continue;
        }
896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929
        if (!is_scalar(handle.ptr()) && !PySlice_Check(handle.ptr())) {
            use_subtensor = false;
        }
        py::list item;
        item.append(cur_axis);
        auto push = [&](PyObject* v) {
            if (v == Py_None) {
                item.append(false);
            } else {
                item.append(true);
                tensors.append(_get_index(py::reinterpret_borrow<py::object>(v), inp));
            }
        };

        if (PySlice_Check(handle.ptr())) {
            PySliceObject* s = (PySliceObject*)handle.ptr();
            if (s->start == Py_None && s->stop == Py_None && s->step == Py_None) {
                continue;
            }
            push(s->start);
            push(s->stop);
            push(s->step);
            item.append(false);
        } else {
            for (size_t j = 0; j < 3; j++)
                item.append(false);
            push(handle.ptr());
        }
        items.append(item);
    }

    return py::make_tuple(inp, tensors, items, use_subtensor, need_expand_bool_dim);
}

930 931 932
py::object _expand_args(py::handle args) {
    if (!PyTuple_Check(args.ptr())) {
        return py::reinterpret_borrow<py::object>(args);
933
    }
934 935 936 937
    py::tuple args_tup = py::reinterpret_borrow<py::tuple>(args.ptr());
    if (args_tup.size() == 1 && (PySequence_Check(args_tup[0].ptr()) ||
                                 is_tensor_or_symbolvar(args_tup[0].ptr()))) {
        return py::reinterpret_borrow<py::object>(args_tup[0]);
938
    } else {
939
        return py::reinterpret_steal<py::list>(PySequence_List(args_tup.ptr()));
940 941 942
    }
}

943 944 945 946
std::tuple<std::vector<int32_t>, bool> tuple2vector(py::object shape) {
    std::vector<int32_t> shp;
    if (!PyTuple_Check(shape.ptr())) {
        return {shp, false};
947
    }
948 949 950
    py::tuple tup = py::reinterpret_borrow<py::tuple>(shape);
    for (size_t i = 0; i < tup.size(); ++i) {
        if (!PyLong_Check(tup[i].ptr())) {
951
            shp.clear();
952 953 954 955
            return {shp, false};
        } else {
            shp.push_back(tup[i].cast<int32_t>());
        }
956
    }
957 958 959 960
    return {shp, true};
}

bool enable_fastpath(py::handle inp) {
961
    // FIXME: the way to judge whether it is in traced module is inaccurate
962 963 964 965 966 967 968 969
    if (!TensorWrapper::try_cast(inp.ptr()) ||
        TransformationManager::get_instance()
                        .segments[TransformationManager::Segment::Trace]
                        .size() > 0 ||
        TransformationManager::get_instance()
                        .segments[TransformationManager::Segment::ModuleTrace]
                        .size() > 0) {
        return false;
970
    }
971 972
    return true;
}
973

974 975 976 977 978 979 980 981
py::object _broadcast_cpp(py::handle inp_hdl, py::handle args) {
    py::object shape_hdl = _expand_args(args);
    bool auto_infer = false;
    py::list lis;
    py::list new_shape;
    if (PyList_Check(shape_hdl.ptr()) || PyTuple_Check(shape_hdl.ptr())) {
        lis = py::reinterpret_steal<py::list>(PySequence_List(shape_hdl.ptr()));
        for (size_t i = 0; i < lis.size(); ++i) {
982
            if (lis[i].is_none()) {
983 984 985
                auto_infer = true;
                size_t right = lis.size() - i;
                py::object tshp = getattr(inp_hdl, "_tuple_shape");
986
                if (tshp.is_none()) {
987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111
                    throw py::index_error("does not support `None` with unknown shape");
                }
                py::tuple inp_shape = py::reinterpret_borrow<py::tuple>(tshp);
                if (inp_shape.size() >= right) {
                    if (enable_fastpath(inp_hdl)) {
                        lis[i] = inp_shape[inp_shape.size() - right];
                    }
                    new_shape.append(inp_shape[inp_shape.size() - right]);
                } else {
                    throw py::value_error("invalid broadcast shape");
                }
            } else {
                new_shape.append(lis[i]);
                if (PyLong_Check(lis[i].ptr())) {
                    int32_t s = lis[i].cast<int32_t>();
                    if (s < 0) {
                        throw py::value_error(
                                "expect shape[" + std::to_string(i) +
                                "] >= 0 or use `None` to auto infer, got " +
                                std::to_string(s));
                    }
                }
            }
        }
    }
    if (auto_infer) {
        if (enable_fastpath(inp_hdl)) {
            shape_hdl = py::reinterpret_borrow<py::tuple>(lis);
        } else {
            shape_hdl = _astensor1d_cpp(
                    new_shape, py::cast((mgb::DType)dtype::Int32()),
                    getattr(inp_hdl, "device"), inp_hdl);
        }
    }
    py::object shape_tuple;
    try {
        shape_tuple = _make_shape_tuple(shape_hdl);
    } catch (py::error_already_set& err) {
        shape_tuple = py::reinterpret_borrow<py::object>(shape_hdl);
    }
    auto [shape, fastpath] = tuple2vector(shape_tuple);
    fastpath &= enable_fastpath(inp_hdl);
    std::shared_ptr<OpDef> op;
    std::vector<PyObject*> p;
    py::object shape_tensor;
    if (fastpath) {
        op = Broadcast::make(shape);
        p.resize(2);
    } else {
        op = Broadcast::make();
        shape_tensor = _astensor1d_cpp(
                shape_hdl, py::cast((mgb::DType)dtype::Int32()),
                getattr(inp_hdl, "device"), inp_hdl);
        p.resize(3);
        p[2] = shape_tensor.ptr();
    }
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = inp_hdl.ptr();
    py::tuple ret =
            py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
    return ret[0];
}

py::object _reshape_cpp(py::handle inp_hdl, py::handle args) {
    py::object shape_hdl = _expand_args(args);
    py::object shape_tuple;
    try {
        shape_tuple = _make_shape_tuple(shape_hdl);
    } catch (py::error_already_set& err) {
        shape_tuple = py::reinterpret_borrow<py::object>(shape_hdl);
    }
    int32_t unspec_axis = -1;
    if (PyTuple_Check(shape_tuple.ptr())) {
        py::tuple tup = py::reinterpret_borrow<py::tuple>(shape_tuple);
        for (size_t i = 0; i < tup.size(); ++i) {
            py::object obj = py::reinterpret_borrow<py::object>(tup[i]);
            if (obj < py::int_(0)) {
                if (obj.not_equal(py::int_(-1))) {
                    throw py::value_error(
                            "expect shape [" + std::to_string(i) + "] >= -1, got " +
                            repr(obj).cast<std::string>());
                }
                if (unspec_axis >= 0) {
                    throw py::value_error(
                            "multiple -1 in shape: " + std::to_string(unspec_axis) +
                            " & " + std::to_string(i));
                }
                unspec_axis = i;
            }
        }
    }
    auto [shape, fastpath] = tuple2vector(shape_tuple);
    fastpath &= enable_fastpath(inp_hdl);
    std::shared_ptr<OpDef> op;
    std::vector<PyObject*> p;
    py::object shape_tensor;
    if (fastpath) {
        if (unspec_axis >= 0) {
            op = Reshape::make(unspec_axis, shape);
        } else {
            op = Reshape::make(::megdnn::param::OptionalAxisV1::INVALID_AXIS, shape);
        }
        p.resize(2);
    } else {
        shape.clear();
        if (unspec_axis >= 0) {
            op = Reshape::make(unspec_axis, shape);
        } else {
            op = Reshape::make();
        }
        shape_tensor = _astensor1d_cpp(
                shape_hdl, py::cast((mgb::DType)dtype::Int32()),
                getattr(inp_hdl, "device"), inp_hdl);
        p.resize(3);
        p[2] = shape_tensor.ptr();
    }
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = inp_hdl.ptr();
    py::tuple ret =
            py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
    return ret[0];
}

1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157
py::object _adaptive_pool2d_cpp(
        py::handle inp_hdl, py::handle shape_val_hdl, py::handle pool_mode_hdl) {
    py::object shape_hdl = py::reinterpret_borrow<py::object>(shape_val_hdl);
    py::list shps(0);
    if (!PyTuple_Check(shape_val_hdl.ptr())) {
        shps.append(PyLong_AsLong(shape_val_hdl.ptr()));
        shps.append(PyLong_AsLong(shape_val_hdl.ptr()));

        shape_hdl = py::reinterpret_borrow<py::object>(shps);
    }
    py::object shape_tuple;
    try {
        shape_tuple = _make_shape_tuple(shape_hdl);
    } catch (py::error_already_set& err) {
        shape_tuple = py::reinterpret_borrow<py::object>(shape_hdl);
    }
    auto mode_string = pool_mode_hdl.cast<std::string>();
    ::megdnn::param::AdaptivePooling::Mode pool_mode =
            ::megdnn::param::AdaptivePooling::Mode::MAX;
    if (mode_string.compare(std::string("AVERAGE")) == 0) {
        pool_mode = ::megdnn::param::AdaptivePooling::Mode::AVERAGE;
    }
    auto [shape, fastpath] = tuple2vector(shape_tuple);
    fastpath &= enable_fastpath(inp_hdl);
    std::shared_ptr<OpDef> op;
    std::vector<PyObject*> p;
    py::object shape_tensor;
    op = AdaptivePooling::make(
            pool_mode, ::megdnn::param::AdaptivePooling::Format::NCHW, shape);
    if (fastpath) {
        p.resize(2);
    } else {
        p.resize(3);
        shape_tensor = _astensor1d_cpp(
                shape_hdl, py::cast((mgb::DType)dtype::Int32()),
                getattr(inp_hdl, "device"), inp_hdl);
        p[2] = shape_tensor.ptr();
    }
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = inp_hdl.ptr();
    py::tuple ret =
            py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
    return ret[0];
}

1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
py::object _getitem_cpp(py::handle inp_hdl, py::handle idx_hdl) {
    py::tuple try_res = _try_cond_take(inp_hdl, idx_hdl);
    if (try_res.size() == 2) {
        return try_res[0];
    }
    py::tuple up = _unpack_indexes(inp_hdl, idx_hdl);
    py::object tensor = py::reinterpret_borrow<py::object>(up[0]);
    py::list tensors = py::reinterpret_borrow<py::list>(up[1]);
    py::list py_items = py::reinterpret_borrow<py::list>(up[2]);
    std::vector<std::tuple<int8_t, bool, bool, bool, bool>> cpp_items;
    for (size_t i = 0; i < py_items.size(); ++i) {
        py::list item = py::reinterpret_borrow<py::list>(py_items[i]);
        cpp_items.push_back(
                {item[0].cast<int8_t>(), item[1].cast<bool>(), item[2].cast<bool>(),
                 item[3].cast<bool>(), item[4].cast<bool>()});
    }
1174
    std::shared_ptr<OpDef> op;
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212
    if (up[3].cast<bool>()) {
        op = Subtensor::make(cpp_items);
    } else {
        op = IndexingMultiAxisVec::make(cpp_items);
    }
    std::vector<PyObject*> p;
    p.resize(tensors.size() + 2);
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = tensor.ptr();
    for (size_t i = 0; i < tensors.size(); ++i) {
        p[i + 2] = tensors[i].ptr();
    }
    py::tuple ret =
            py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
    return ret[0];
}

py::object _setitem_cpp(py::handle inp_hdl, py::handle idx_hdl, py::handle val_hdl) {
    py::object org_shape = getattr(inp_hdl, "shape");
    py::object val = py::reinterpret_borrow<py::object>(val_hdl);
    if (!TensorWrapper::try_cast(val.ptr()) && !py::isinstance<PySymbolVar>(val)) {
        val =
                _Const(val_hdl, getattr(inp_hdl, "dtype"), getattr(inp_hdl, "device"),
                       inp_hdl);
    }

    py::tuple up = _unpack_indexes(inp_hdl, idx_hdl);
    py::object tensor = py::reinterpret_borrow<py::object>(up[0]);
    py::list tensors = py::reinterpret_borrow<py::list>(up[1]);
    py::list py_items = py::reinterpret_borrow<py::list>(up[2]);
    std::vector<std::tuple<int8_t, bool, bool, bool, bool>> cpp_items;
    for (size_t i = 0; i < py_items.size(); ++i) {
        py::list item = py::reinterpret_borrow<py::list>(py_items[i]);
        cpp_items.push_back(
                {item[0].cast<int8_t>(), item[1].cast<bool>(), item[2].cast<bool>(),
                 item[3].cast<bool>(), item[4].cast<bool>()});
    }
1213
    std::shared_ptr<OpDef> op, set_op;
1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235
    if (up[3].cast<bool>()) {
        op = Subtensor::make(cpp_items);
    } else {
        op = IndexingMultiAxisVec::make(cpp_items);
    }
    std::vector<PyObject*> p;
    p.resize(tensors.size() + 2);
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = tensor.ptr();
    for (size_t i = 0; i < tensors.size(); ++i) {
        p[i + 2] = tensors[i].ptr();
    }
    py::tuple ret =
            py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
    py::object tmp_result = ret[0];

    try {
        py::tuple value_shape =
                py::reinterpret_borrow<py::tuple>(val.attr("_tuple_shape"));
        py::tuple tmp_result_shape =
                py::reinterpret_borrow<py::tuple>(tmp_result.attr("_tuple_shape"));
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259
        for (size_t i = 0; i < value_shape.size() && i < tmp_result_shape.size(); ++i) {
            size_t vs = value_shape[value_shape.size() - i - 1].cast<size_t>();
            size_t ts =
                    tmp_result_shape[tmp_result_shape.size() - i - 1].cast<size_t>();
            if (vs != 1 && vs != ts) {
                std::string lhs = "", rhs = "";
                for (size_t j = 0; j < tmp_result_shape.size(); ++j) {
                    lhs += std::to_string(tmp_result_shape[j].cast<size_t>());
                    if (j)
                        lhs += ",";
                }
                for (size_t j = 0; j < value_shape.size(); ++j) {
                    rhs += std::to_string(value_shape[j].cast<size_t>());
                    if (j)
                        rhs += ",";
                }
                throw py::value_error(
                        "cannot copy tensor with shape (" + rhs +
                        ") to subtensor with shape (" + lhs + ")");
            }
        }
    } catch (py::error_already_set& err) {
        ;
    }
1260
    val = _broadcast_cpp(val, getattr(tmp_result, "shape"));
1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280
    if (up[3].cast<bool>()) {
        set_op = SetSubtensor::make(cpp_items);
    } else {
        set_op = IndexingSetMultiAxisVec::make(cpp_items);
    }

    std::vector<PyObject*> q;
    q.resize(tensors.size() + 3);
    py::object Set_Op = py::cast(set_op);
    q[0] = Set_Op.ptr();
    q[1] = tensor.ptr();
    q[2] = val.ptr();
    for (size_t i = 0; i < tensors.size(); ++i) {
        q[i + 3] = tensors[i].ptr();
    }
    py::tuple result =
            py::reinterpret_steal<py::object>(py_apply(NULL, q.data(), q.size()));
    py::object res = result[0];

    if (up[4].cast<bool>()) {
1281
        res = _reshape_cpp(res, org_shape);
1282 1283 1284 1285 1286
    }

    return res;
}

1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
py::object _split_cpp(
        py::handle inp_hdl, py::handle nsplits_or_sections_hdl, py::handle axis_hdl) {
    py::object shape_obj = getattr(inp_hdl, "shape");
    py::object n_total = shape_obj[axis_hdl];
    int ndim = shape_obj.attr("__len__")().cast<int>();
    int axis = axis_hdl.cast<int>();
    if (axis >= ndim) {
        throw py::value_error("Invalid axis " + std::to_string(axis));
    }
    int n_sections;
    bool is_array;
    if (is_py_sequence(nsplits_or_sections_hdl)) {
        n_sections = PySequence_Length(nsplits_or_sections_hdl.ptr()) + 1;
        is_array = true;
    } else {
        n_sections = getattr(nsplits_or_sections_hdl, "__int__")().cast<int>();
        is_array = false;
    }
    py::list partitions;
    std::shared_ptr<OpDef> op;
    std::vector<PyObject*> p;
    if (is_array) {
        py::list div_points;
        py::list sections = py::reinterpret_borrow<py::object>(nsplits_or_sections_hdl);
        div_points.append(0);
        for (size_t i = 0; i < sections.size(); ++i) {
            div_points.append(sections[i]);
        }
        div_points.append(n_total);
        for (size_t i = 1; i < div_points.size(); ++i) {
            if (div_points[i - 1] > div_points[i]) {
                throw py::value_error(
                        "Invalid nsplits_or_secions: " +
                        repr(nsplits_or_sections_hdl).cast<std::string>());
            }
            py::object pos = div_points[i] - div_points[i - 1];
            if (is_tensor_or_symbolvar(pos)) {
                partitions.append(pos);
            } else {
                partitions.append(
                        _Const(pos, py::cast((mgb::DType)dtype::Int32()),
                               getattr(inp_hdl, "device"), inp_hdl));
            }
        }
        op = Split::make(axis, 0);
        p.resize(partitions.size() + 2);
        for (size_t i = 0; i < partitions.size(); ++i) {
            p[i + 2] = partitions[i].ptr();
        }
    } else {
        if (n_sections <= 0) {
            throw py::value_error("Number sections must be larger than 0");
        }
        if (py::int_(n_sections) > n_total) {
            throw py::value_error(
                    "The size " + repr(n_total).cast<std::string>() + " at dim " +
                    std::to_string(axis) + " cannot be split into " +
                    std::to_string(n_sections) + " sections");
        }
        op = Split::make(axis, n_sections);
        p.resize(2);
    }
    py::object Op = py::cast(op);
    p[0] = Op.ptr();
    p[1] = inp_hdl.ptr();
    return py::reinterpret_steal<py::object>(py_apply(NULL, p.data(), p.size()));
}

1355
std::vector<int32_t> list2vector(py::handle li) {
1356
    std::vector<int32_t> axis;
1357
    if (is_py_sequence(li)) {
1358
        py::list tmp_list = py::reinterpret_steal<py::list>(PySequence_List(li.ptr()));
1359 1360 1361 1362
        for (size_t i = 0; i < tmp_list.size(); ++i) {
            axis.push_back(tmp_list[i].attr("__int__")().cast<int32_t>());
        }
    } else {
1363
        axis.push_back(getattr(li, "__int__")().cast<int32_t>());
1364
    }
1365 1366 1367 1368 1369
    return axis;
}

py::object _expand_dims_cpp(py::handle inp_hdl, py::handle axis_hdl) {
    std::vector<int32_t> axis = list2vector(axis_hdl);
1370 1371 1372 1373 1374 1375 1376 1377 1378
    bool unknown_ndim = true;
    size_t ndim = axis.size();
    if (auto p = TensorWrapper::try_cast(inp_hdl.ptr())) {
        auto&& shape = p->m_tensor->shape();
        if (shape) {
            unknown_ndim = false;
            ndim += shape->ndim;
        }
    } else {
1379 1380
        auto&& inp_ndim = get_ndim_safe(inp_hdl);
        ndim += inp_ndim.first;
1381
        unknown_ndim &= !inp_ndim.second;
1382 1383 1384 1385 1386 1387 1388 1389
    }
    for (size_t i = 0; i < axis.size(); ++i) {
        if (axis[i] < 0) {
            if (unknown_ndim) {
                throw py::index_error(
                        "Does not support negative index when tensor's ndim is "
                        "unknown");
            }
1390
            axis[i] += static_cast<int32_t>(ndim);
1391 1392 1393 1394 1395 1396 1397 1398
        }
    }
    if (!axis.size()) {
        throw py::index_error("axis could not be empty");
    }
    std::sort(axis.begin(), axis.end());
    std::shared_ptr<OpDef> op = AddAxis::make(axis = axis);
    py::object Op = py::cast(op);
1399 1400
    PyObject* p[2] = {Op.ptr(), inp_hdl.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
1401 1402 1403
    return ret[0];
}

1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422
py::object _squeeze_cpp(py::handle inp_hdl, py::handle axis_hdl) {
    std::vector<int32_t> axis;
    size_t ndim;
    if (axis_hdl.ptr() != Py_None) {
        axis = list2vector(axis_hdl);
    }
    if (auto p = TensorWrapper::try_cast(inp_hdl.ptr())) {
        auto&& shape = p->m_tensor->shape();
        if (shape) {
            ndim = shape->ndim;
            if (axis_hdl.ptr() == Py_None) {
                for (size_t i = 0; i < shape->ndim; ++i) {
                    if (shape->shape[i] == 1) {
                        axis.push_back(i);
                    }
                }
            }
        }
    } else {
1423 1424 1425 1426 1427 1428 1429
        py::tuple shape =
                py::reinterpret_borrow<py::tuple>(getattr(inp_hdl, "_tuple_shape"));
        ndim = shape.size();
        if (axis_hdl.ptr() == Py_None) {
            for (size_t i = 0; i < shape.size(); ++i) {
                if (shape[i].cast<size_t>() == 1) {
                    axis.push_back(i);
1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444
                }
            }
        }
    }
    for (size_t i = 0; i < axis.size(); ++i) {
        if (axis[i] < 0) {
            axis[i] += static_cast<int32_t>(ndim);
        }
    }
    std::sort(axis.begin(), axis.end());
    for (size_t i = 0; i < axis.size(); ++i) {
        axis[i] -= static_cast<int32_t>(i);
    }
    std::shared_ptr<OpDef> op = RemoveAxis::make(axis = axis);
    py::object Op = py::cast(op);
1445 1446
    PyObject* p[2] = {Op.ptr(), inp_hdl.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
1447 1448
    return ret[0];
}
1449

1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462
py::object _transpose_cpp(py::handle inp_hdl, py::handle args) {
    py::object obj = _expand_args(args);
    py::list lis;
    if (!is_tensor_or_symbolvar(obj.ptr()) && PySequence_Check(obj.ptr())) {
        lis = py::reinterpret_steal<py::list>(PySequence_List(obj.ptr()));
    } else {
        py::object np = getattr(obj, "numpy")();
        PyArrayObject* arr = (PyArrayObject*)np.ptr();
        PyObject* maybe_list = PyArray_ToList(arr);
        if (PyList_Check(maybe_list)) {
            lis = py::reinterpret_steal<py::list>(maybe_list);
        }
    }
1463
    if (get_ndim_safe(inp_hdl).first == 0) {
1464
        if (lis.size() != 0) {
1465 1466 1467 1468 1469 1470
            throw py::index_error(
                    "transpose for scalar does not accept additional args");
        }
        return getattr(inp_hdl, "to")(getattr(inp_hdl, "device"));
    }
    std::vector<int32_t> pattern;
1471
    if (!lis.size()) {
1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488
        size_t ndim = getattr(inp_hdl, "ndim").cast<size_t>();
        for (size_t i = 0; i < ndim; ++i) {
            pattern.push_back(ndim - i - 1);
        }
    } else {
        for (size_t i = 0; i < lis.size(); ++i) {
            if (PyLong_Check(lis[i].ptr())) {
                pattern.push_back(lis[i].cast<int32_t>());
            } else {
                if (lis[i].cast<std::string>() == "x") {
                    pattern.push_back(-1);
                }
            }
        }
    }
    std::shared_ptr<OpDef> op = Dimshuffle::make(pattern);
    py::object Op = py::cast(op);
1489 1490
    PyObject* p[2] = {Op.ptr(), inp_hdl.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
1491 1492 1493
    return ret[0];
}

1494 1495 1496
py::object _matmul_cpp(
        py::handle inp1, py::handle inp2, py::handle dim1, py::handle dim2,
        py::handle transpose_a, py::handle transpose_b, py::handle compute_mode,
1497 1498 1499 1500 1501 1502 1503 1504 1505 1506
        py::handle profile, py::handle determistic) {
    ::megdnn::param::MatrixMul::ComputeMode mode =
            ::megdnn::param::MatrixMul::ComputeMode::DEFAULT;
    if (compute_mode.cast<std::string>().compare(std::string("float32")) == 0) {
        mode = ::megdnn::param::MatrixMul::ComputeMode::FLOAT32;
    }
    ::megdnn::param::ExecutionPolicy::Strategy cstrategy =
            static_cast<::megdnn::param::ExecutionPolicy::Strategy>(0);
    if (profile.cast<bool>()) {
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::PROFILE;
1507
    } else {
1508
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::HEURISTIC;
1509
    }
1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521
    if (determistic.cast<bool>()) {
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::REPRODUCIBLE;
    }
    std::shared_ptr<OpDef> op = MatrixMul::make(
            transpose_a.cast<bool>(), transpose_b.cast<bool>(), mode,
            ::megdnn::param::MatrixMul::Format::DEFAULT, cstrategy, UINT64_MAX,
            dim1.cast<uint32_t>(), dim2.cast<uint32_t>());

    py::object Op = py::cast(op);
    PyObject* p[3] = {Op.ptr(), inp1.ptr(), inp2.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 3));
    return ret[0];
1522 1523 1524 1525 1526
}

py::object _batched_matmul_cpp(
        py::handle inp1, py::handle inp2, py::handle dim1, py::handle dim2,
        py::handle transpose_a, py::handle transpose_b, py::handle compute_mode,
1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
        py::handle profile, py::handle determistic) {
    ::megdnn::param::MatrixMul::ComputeMode mode =
            ::megdnn::param::MatrixMul::ComputeMode::DEFAULT;
    if (compute_mode.cast<std::string>().compare(std::string("float32")) == 0) {
        mode = ::megdnn::param::MatrixMul::ComputeMode::FLOAT32;
    }
    ::megdnn::param::ExecutionPolicy::Strategy cstrategy =
            static_cast<::megdnn::param::ExecutionPolicy::Strategy>(0);
    if (profile.cast<bool>()) {
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::PROFILE;
1537
    } else {
1538
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::HEURISTIC;
1539
    }
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551
    if (determistic.cast<bool>()) {
        cstrategy |= ::megdnn::param::ExecutionPolicy::Strategy::REPRODUCIBLE;
    }
    std::shared_ptr<OpDef> op = BatchedMatrixMul::make(
            transpose_a.cast<bool>(), transpose_b.cast<bool>(), mode,
            ::megdnn::param::MatrixMul::Format::DEFAULT, cstrategy, UINT64_MAX,
            dim1.cast<uint32_t>(), dim2.cast<uint32_t>());

    py::object Op = py::cast(op);
    PyObject* p[3] = {Op.ptr(), inp1.ptr(), inp2.ptr()};
    py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 3));
    return ret[0];
1552 1553
}

1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566
py::object _pixel_shuffle_cpp(py::handle inp, py::handle val, py::handle func) {
    if (enable_fastpath(inp) && PyLong_Check(val.ptr())) {
        std::shared_ptr<OpDef> op = PixelShuffle::make(val.cast<int32_t>());
        py::object Op = py::cast(op);
        PyObject* p[2] = {Op.ptr(), inp.ptr()};
        py::tuple ret = py::reinterpret_steal<py::object>(py_apply(NULL, p, 2));
        return ret[0];
    } else {
        // fallback to traceable subgraph implement
        return func(inp, val);
    }
}

1567 1568
PyObject* make_shape_tuple(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1569
        return _make_shape_tuple(args[0]).release().ptr();
1570 1571 1572 1573 1574 1575
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

PyObject* getitem_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1576
        return _getitem_cpp(args[0], args[1]).release().ptr();
1577 1578 1579 1580 1581 1582
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

PyObject* setitem_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1583
        return _setitem_cpp(args[0], args[1], args[2]).release().ptr();
1584 1585 1586 1587
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1588 1589
PyObject* split_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1590
        return _split_cpp(args[0], args[1], args[2]).release().ptr();
1591 1592 1593 1594
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1595 1596
PyObject* expand_dims_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1597
        return _expand_dims_cpp(args[0], args[1]).release().ptr();
1598 1599 1600 1601
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1602 1603
PyObject* squeeze_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1604
        return _squeeze_cpp(args[0], args[1]).release().ptr();
1605 1606 1607 1608
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1609 1610
PyObject* transpose_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1611
        return _transpose_cpp(args[0], args[1]).release().ptr();
1612 1613 1614 1615
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1616 1617
PyObject* broadcast_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1618
        return _broadcast_cpp(args[0], args[1]).release().ptr();
1619 1620 1621 1622 1623 1624
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

PyObject* reshape_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1625
        return _reshape_cpp(args[0], args[1]).release().ptr();
1626 1627 1628 1629
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1630 1631 1632 1633 1634 1635 1636
PyObject* adaptive_pool2d_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        return _adaptive_pool2d_cpp(args[0], args[1], args[2]).release().ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1637 1638 1639 1640 1641 1642 1643
PyObject* pixel_shuffle_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        return _pixel_shuffle_cpp(args[0], args[1], args[2]).release().ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1644 1645
PyObject* Const(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1646
        return _Const(args[0], args[1], args[2], args[3]).release().ptr();
1647 1648 1649 1650
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1651 1652
PyObject* astype_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1653
        return _astype_cpp(args[0], args[1]).release().ptr();
1654 1655 1656 1657
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1658 1659 1660 1661
PyObject* matmul_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        return _matmul_cpp(
                       args[0], args[1], args[2], args[3], args[4], args[5], args[6],
1662
                       args[7], args[8])
1663 1664 1665 1666 1667 1668 1669 1670 1671 1672
                .release()
                .ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

PyObject* batched_matmul_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        return _batched_matmul_cpp(
                       args[0], args[1], args[2], args[3], args[4], args[5], args[6],
1673
                       args[7], args[8])
1674 1675 1676 1677 1678 1679
                .release()
                .ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1680 1681 1682
PyObject* convert_single_value_cpp(
        PyObject* self, PyObject* const* args, size_t nargs) {
    try {
1683
        return _convert_single_value_cpp(args[0], args[1], args[2]).release().ptr();
1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

PyObject* convert_inputs_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        py::object dtype = py::reinterpret_steal<py::object>(
                dtype_promotion(self, args, nargs - 1));
        py::object device;
        if (args[nargs - 1] == Py_None) {
            device = py::reinterpret_steal<py::object>(
                    get_device(self, args, nargs - 1));
        } else {
            device = py::reinterpret_borrow<py::object>(args[nargs - 1]);
        }
        return _convert_inputs_cpp(args, nargs - 1, dtype, device).release().ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1704 1705 1706 1707 1708 1709 1710
PyObject* astensor1d_cpp(PyObject* self, PyObject* const* args, size_t nargs) {
    try {
        return _astensor1d_cpp(args[0], args[1], args[2], args[3]).release().ptr();
    }
    PYEXT17_TRANSLATE_EXC_RET(nullptr)
}

1711
}  // namespace mgb::imperative::python