perf(mge/imperative): move convert_inputs from python to C++

GitOrigin-RevId: baef3d348c

imperative/python/megengine/core/ops/special.py

@@ -8,7 +8,7 @@
 # "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 import numpy as np
 
-from .._imperative_rt.core2 import Tensor
+# from .._imperative_rt.core2 import Tensor
 from ..tensor.core import OpBase, TensorBase, apply
 
 
@@ -19,5 +19,10 @@ class Const:
         self.device = device
 
     def __call__(self, *reference):
-        Wrapper = type(reference[0])
-        return (Wrapper(self.value, self.dtype, self.device, True),)
+        from ...tensor import Tensor
+
+        device = self.device
+        if device is None:
+            device = reference[0].device
+
+        return (Tensor(self.value, self.dtype, self.device, True),)

imperative/python/megengine/core/tensor/dtype.py

@@ -13,6 +13,12 @@ import numpy as np
 
 # normal dtype related
 from .._imperative_rt import bfloat16, intb1, intb2, intb4
+from .._imperative_rt.common import (
+    get_scale,
+    get_zero_point,
+    is_dtype_equal,
+    is_quantize,
+)
 
 
 def is_lowbit(dtype):
@@ -42,41 +48,6 @@ _metadata_dict = {
 }
 
 
-def is_quantize(dtype):
-    return (
-        hasattr(dtype, "metadata")
-        and dtype.metadata is not None
-        and "mgb_dtype" in dtype.metadata
-    )
-
-
-def get_scale(dtype):
-    assert is_quantize(dtype)
-    return dtype.metadata["mgb_dtype"]["scale"]
-
-
-def get_zero_point(dtype):
-    assert is_quantize(dtype)
-    metadata = dtype.metadata["mgb_dtype"]
-    assert metadata["name"] in ("Quantized8Asymm", "Quantized4Asymm")
-    return metadata["zero_point"]
-
-
-def is_equal(dt0, dt1):
-    def _get_zero_point(dtype):
-        assert is_quantize(dtype)
-        metadata = dtype.metadata["mgb_dtype"]
-        return metadata.get("zero_point")
-
-    if is_quantize(dt0) and is_quantize(dt1):
-        return get_scale(dt0) == get_scale(dt1) and _get_zero_point(
-            dt0
-        ) == _get_zero_point(dt1)
-    if not (is_quantize(dt0) or is_quantize(dt1)):
-        return dt0 == dt1
-    return False
-
-
 def _check_zero_point(zp: int, dtype_str: str):
     qmin = _metadata_dict[dtype_str].qmin
     qmax = _metadata_dict[dtype_str].qmax

imperative/python/megengine/core/tensor/indexing.py

@@ -151,9 +151,9 @@ def unpack_getitem(inp, tuple_val, *, allow_newaxis=True):
         def get_index(i):
             if not isinstance(i, (Tensor)):
                 if is_bool_list(i) or isinstance(i, np.ndarray) and i.dtype == np.bool_:
-                    (i,) = Const(i, dtype=np.bool_, device=inp.device)(inp)
+                    (i,) = Const(i, dtype=np.bool_, device=inp.device)()
                 else:
-                    (i,) = Const(i, dtype=np.int32, device=inp.device)(inp)
+                    (i,) = Const(i, dtype=np.int32, device=inp.device)()
                     return i
             assert isinstance(i, Tensor)
             if i.dtype != np.bool_:
@@ -197,7 +197,7 @@ def try_condtake(tensor, index):
     ):
         return []
     if isinstance(index, np.ndarray):
-        (index,) = Const(index, dtype=np.bool_, device=tensor.device)(tensor)
+        (index,) = Const(index, dtype=np.bool_, device=tensor.device)()
     assert isinstance(index, Tensor)
     if not isinstance(tensor, Tensor):
         raise TypeError("input must be a tensor")
@@ -217,9 +217,7 @@ def getitem(tensor, index):
         if isinstance(v.shape, v.__class__):
             break
         if len(v.shape) > 0 and v.shape[0] == 0:
-            (empty_tensor,) = Const([], dtype=tensor.dtype, device=tensor.device)(
-                tensor
-            )
+            (empty_tensor,) = Const([], dtype=tensor.dtype, device=tensor.device)()
             return empty_tensor
     if use_subtensor:
         op = builtin.Subtensor(items=items)
@@ -240,8 +238,7 @@ def setitem(tensor, index, value):
             return tensor
         tensor = tensor.reshape(-1)
     if not isinstance(value, Tensor):
-        op = Const(value, dtype=tensor.dtype, device=tensor.device)
-        (value,) = op(tensor)
+        (value,) = Const(value, dtype=tensor.dtype, device=tensor.device)()
     tensor, tensors, items, use_subtensor, _ = unpack_getitem(tensor, index)
     for v in tensors:
         if len(v.shape) > 0 and v.shape[0] == 0:

imperative/python/megengine/core/tensor/utils.py

@@ -11,10 +11,10 @@ from typing import Iterable, Union
 
 import numpy as np
 
-from .._imperative_rt.core2 import Tensor, apply
+from .._imperative_rt.core2 import Tensor, apply, dtype_promotion, get_device
 from ..ops import builtin
 from ..ops.special import Const
-from .dtype import is_equal, is_quantize
+from .dtype import is_dtype_equal, is_quantize
 from .megbrain_graph import VarNode
 
 _enable_convert_inputs = True
@@ -37,94 +37,12 @@ def set_convert_inputs(flag):
     return backup
 
 
-def dtype_promotion(inputs):
-    """
-    Returns the dtype that would result from performing an arithmetic
-    operation on the provided input tensors and scalars.
-    """
-    # map numpy.dtype.kind to priority
-    category_priority = {
-        "f": 3,  # floating-point
-        "i": 2,  # signed integer
-        "u": 2,  # unsigned integer
-        "b": 1,  # boolean
-    }
-
-    def scalar2dtype(x):
-        """
-        For scalar `x`, returns its corresponding type. A floating point scalar
-        has dtype 'float32'. An integral non-boolean scalar has dtype 'int32'.
-        A boolean scalar has dtype 'bool'.
-        """
-        if isinstance(x, bool):
-            return np.bool_
-        if isinstance(x, int):
-            return np.int32
-        if isinstance(x, float):
-            return np.float32
-
-    def promote_types(types, cat):
-        """
-        Returns the data type with sufficient size to hold all types of
-        category `cat` in the list `types`.
-        """
-        used_types = [
-            i for i in types if category_priority.get(np.dtype(i).kind, 0) == cat
-        ]
-        assert len(used_types) > 0
-        res = used_types[0]
-        for i in used_types:
-            res = np.promote_types(res, i)
-        return res
-
-    def max_priority(types):
-        """
-        Returns the maximum value of the priority of each type in the list
-        `types`.
-        """
-        if not types:
-            return 0
-        else:
-            return max([category_priority.get(np.dtype(i).kind, 0) for i in types])
-
-    scalars = []
-    tensors = []
-
-    for data in inputs:
-        if hasattr(data, "dtype"):
-            tensors.append(data.dtype)
-        elif isinstance(data, (float, int, bool)):
-            scalars.append(scalar2dtype(data))
-
-    max_pri_scalars = max_priority(scalars)
-    max_pri_tensors = max_priority(tensors)
-
-    assert max_pri_scalars > 0 or max_pri_tensors > 0
-
-    if max_pri_scalars > max_pri_tensors:
-        return promote_types(scalars, max_pri_scalars)
-    else:
-        return promote_types(tensors, max_pri_tensors)
-
-
-def get_device(inputs):
-    device = None
-    for i in inputs:
-        if isinstance(i, (Tensor, VarNode)):
-            if device is None:
-                device = i.device
-            elif device != i.device:
-                raise ValueError("ambiguous device: {} vs {}".format(device, i.device))
-    assert device is not None
-    return device
-
-
 def concatenate(inputs, axis=0, *, device=None):
     dtype = dtype_promotion(inputs)
     device = get_device(inputs)
 
     def convert(x):
-        return convert_single_value(x, inputs, dtype=dtype)
+        return convert_single_value(x, dtype=dtype, device=device)
 
     inputs = tuple(map(convert, inputs))
     (result,) = apply(builtin.Concat(axis=axis, comp_node=device), *inputs)
@@ -133,7 +51,7 @@ def concatenate(inputs, axis=0, *, device=None):
 
 def astype(x, dtype):
     dtype = np.dtype(dtype)
-    if not is_equal(x.dtype, dtype):
+    if not is_dtype_equal(x.dtype, dtype):
         isscalar = x.isscalar()
         (x,) = apply(builtin.TypeCvt(dtype=dtype), x)
         if isscalar:
@@ -141,13 +59,12 @@ def astype(x, dtype):
     return x
 
 
-def convert_single_value(v, inputs, *, dtype=None, device=None):
-    tensors = [i for i in inputs if isinstance(i, (Tensor, VarNode))]
-    assert len(tensors) > 0
+def convert_single_value(v, *, dtype=None, device=None):
     if isinstance(v, (Tensor, VarNode)):
-        v = astype(v, v.dtype if is_quantize(v.dtype) else dtype)
+        if not is_quantize(v.dtype):
+            v = astype(v, dtype)
     else:
-        (v,) = Const(v, dtype=dtype, device=device)(*tensors)
+        (v,) = Const(v, dtype=dtype, device=device)()
     return v
 
 
@@ -161,7 +78,7 @@ def convert_inputs(*args: Tensor):
     def convert(value):
         if value is None:
             return value
-        return convert_single_value(value, args, dtype=dtype, device=device)
+        return convert_single_value(value, dtype=dtype, device=device)
 
     return tuple(map(convert, args))
 

imperative/python/megengine/functional/math.py

@@ -703,7 +703,7 @@ def topk(
     op = builtin.TopK(mode=mode)
 
     if not isinstance(k, Tensor):
-        (k,) = Const(k, dtype="int32", device=inp.device)(inp)
+        (k,) = Const(k, dtype="int32", device=inp.device)()
 
     if len(inp.shape) == 1:
         inp = inp.reshape(1, -1)

imperative/python/megengine/functional/nn.py

@@ -658,7 +658,7 @@ def batch_norm(
 
     def make_full_if_none(x, value):
         if x is None:
-            (x,) = Const(value, dtype=inp.dtype, device=inp.device)(inp)
+            (x,) = Const(value, dtype=inp.dtype, device=inp.device)()
             shape = utils.astensor1d(
                 (1, C, 1, 1), inp, dtype="int32", device=inp.device
             )
@@ -1567,7 +1567,7 @@ def indexing_one_hot(
     """
     assert isinstance(src, Tensor), "src must be of Tensor type"
     op = builtin.IndexingOneHot(axis=axis)
-    index = utils.convert_single_value(index, (src,), dtype="int32", device=src.device)
+    index = utils.convert_single_value(index, dtype="int32", device=src.device)
     (result,) = apply(op, src, index)
     if not keepdims:
         result = squeeze(result, axis)

imperative/python/megengine/functional/tensor.py

@@ -107,9 +107,7 @@ def full(shape, value, dtype="float32", device=None):
         shape = (shape,)
     if device is None:
         device = get_default_device()
-    (x,) = Const(value, dtype=dtype, device=device)(
-        Tensor(value, dtype=dtype, device=device)
-    )
+    (x,) = Const(value, dtype=dtype, device=device)()
     if len(shape) == 0:  # scalar
         return x
     return broadcast_to(x, shape)
@@ -265,7 +263,7 @@ def concat(inps: Iterable[Tensor], axis: int = 0, device=None) -> Tensor:
     device = as_device(device)
 
     def convert(x):
-        return convert_single_value(x, inps, dtype=dtype)
+        return convert_single_value(x, dtype=dtype, device=device)
 
     inps = tuple(map(convert, inps))
     (result,) = apply(builtin.Concat(axis=axis, comp_node=device.to_c()), *inps)

imperative/python/megengine/tensor.py

@@ -37,8 +37,10 @@ class Tensor(_Tensor, ArrayMethodMixin):
             else:
                 cn = CompNode(device)
         else:
-            assert isinstance(device, CompNode)
-            cn = device
+            if isinstance(device, CompNode):
+                cn = device
+            else:
+                cn = device._cn
 
         # import pdb; pdb.set_trace()
         if isinstance(data, _Tensor):

imperative/python/src/common.cpp

@@ -179,4 +179,5 @@ void init_common(py::module m) {
 
     init_npy_num_bfloat16(m);
     init_npy_num_intbx(m);
+    init_dtypes(m);
 }

imperative/python/src/helper.cpp

@@ -158,7 +158,7 @@ void PyExceptionForward::throw_() {
 
 /* ============== namespace npy ============== */
 
-namespace {
+namespace npy {
 
 int to_mgb_supported_dtype_raw(int dtype) {
     if (dtype == NPY_INT64)
@@ -199,12 +199,6 @@ int dtype_mgb2np_raw(DType dtype) {
                 "can not convert dtype %s to numpy dtype", dtype.name()));
 }
 
-struct PyArrayDescrDeleter {
-    void operator()(PyArray_Descr* obj) {
-        Py_XDECREF(obj);
-    }
-};
-
 //! Convert MegBrain DType to NumPy DType descriptor, the caller receives a new
 //! reference to the descriptor.
 std::unique_ptr<PyArray_Descr, PyArrayDescrDeleter> dtype_mgb2np_descr(
@@ -585,9 +579,7 @@ void ndarray_shared_from_tensor_py_capsule_dtor(PyObject *cap) {
     HostTensorNDRefHolder::free(static_cast<HostTensorNDRefHolder*>(ptr));
 }
 
-} // anonymous namespace
-
-PyObject* npy::ndarray_from_tensor(
+PyObject* ndarray_from_tensor(
         const HostTensorND &val, ShareType share_type) {
     if (!val.layout().is_contiguous() && !val.shape().is_empty()) {
         mgb_assert(share_type != ShareType::MUST_SHARE);
@@ -634,7 +626,7 @@ PyObject* npy::ndarray_from_tensor(
     return ret;
 }
 
-HostTensorND npy::np2tensor(PyObject* obj, const Meth& meth, DType dtype) {
+HostTensorND np2tensor(PyObject* obj, const Meth& meth, DType dtype) {
     auto ret_full = np2tensor_try_borrow(obj, meth, dtype);
     if (meth.must_borrow_) {
         mgb_assert(ret_full.second,
@@ -645,7 +637,7 @@ HostTensorND npy::np2tensor(PyObject* obj, const Meth& meth, DType dtype) {
     return ret_full.first;
 }
 
-PyObject* npy::dtype_mgb2np(mgb::DType dtype) {
+PyObject* dtype_mgb2np(mgb::DType dtype) {
     PYTHON_GIL;
     // According to
     // https://docs.scipy.org/doc/numpy/reference/c-api.array.html#c.PyArray_TypeObjectFromType
@@ -668,7 +660,7 @@ PyObject* npy::dtype_mgb2np(mgb::DType dtype) {
     return typeobj;
 }
 
-mgb::DType npy::dtype_np2mgb(PyObject *obj) {
+mgb::DType dtype_np2mgb(PyObject *obj) {
     mgb_assert(obj && obj != Py_None,
                "can not convert null PyObject to numpy dtype");
     // see
@@ -686,7 +678,7 @@ mgb::DType npy::dtype_np2mgb(PyObject *obj) {
     return result;
 }
 
-PyObject* npy::to_mgb_supported_dtype(PyObject* dtype) {
+PyObject* to_mgb_supported_dtype(PyObject* dtype) {
     PYTHON_GIL;
 
     PyArray_Descr* descr;
@@ -702,7 +694,7 @@ PyObject* npy::to_mgb_supported_dtype(PyObject* dtype) {
     return PyArray_TypeObjectFromType(type_num);
 }
 
-TensorShape npy::vec2shape(const std::vector<size_t> &vec) {
+TensorShape vec2shape(const std::vector<size_t> &vec) {
     TensorShape shape;
     mgb_assert(vec.size() <= TensorShape::MAX_NDIM,
             "dim too large: %zd (max %zd)",
@@ -718,3 +710,5 @@ TensorShape npy::vec2shape(const std::vector<size_t> &vec) {
     mgb_assert(shape.ndim, "shape should not be empty");
     return shape;
 }
+
+} // namespace npy

imperative/python/src/helper.h

@@ -11,7 +11,7 @@
 
 #pragma once
 
-#include "megbrain/graph.h"
+#include "megbrain/common.h"
 #include "megbrain/utils/persistent_cache.h"
 #include "megbrain/imperative/op_def.h"
 
@@ -26,6 +26,8 @@
 #include <pybind11/numpy.h>
 #include <pybind11/functional.h>
 
+#include "./numpy_dtypes.h"
+
 pybind11::module submodule(pybind11::module parent, const char* name, const char* doc = nullptr);
 
 pybind11::module rel_import(pybind11::str name, pybind11::module m, int level);
@@ -182,6 +184,18 @@ namespace npy {
     //! convert raw vector to tensor shape
     mgb::TensorShape vec2shape(const std::vector<size_t> &vec);
 
+    struct PyArrayDescrDeleter {
+        void operator()(PyArray_Descr* obj) {
+            Py_XDECREF(obj);
+        }
+    };
+
+    //! Convert MegBrain DType to NumPy DType descriptor, the caller receives a new
+    //! reference to the descriptor.
+    std::unique_ptr<PyArray_Descr, PyArrayDescrDeleter> dtype_mgb2np_descr(mgb::DType dtype);
+
+    mgb::DType dtype_np2mgb_descr(PyArray_Descr* descr);
+
     //! convert megbrain dtype to numpy dtype object; return new reference
     PyObject* dtype_mgb2np(mgb::DType dtype);
 

imperative/python/src/numpy_dtypes.cpp

@@ -0,0 +1,179 @@
+/**
+ * \file imperative/python/src/numpy_dtypes.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.
+ */
+
+#include "./numpy_dtypes.h"
+#include "./helper.h"
+#include "./pyext17.h"
+
+#include "pybind11/pybind11.h"
+
+#include <cstring>
+
+namespace py = pybind11;
+
+namespace mgb {
+namespace {
+
+inline bool _is_quantize(PyArray_Descr* dtype) {
+    static PyObject* PY_MGB_DTYPE_KEY = PyUnicode_FromString("mgb_dtype");
+    return dtype->metadata &&
+           PyDict_CheckExact(dtype->metadata) &&
+           PyDict_Contains(dtype->metadata, PY_MGB_DTYPE_KEY) == 1;
+}
+
+PyObject* _get_mgb_dtype(PyArray_Descr* dtype) {
+    // Return value: New reference.
+    if (!_is_quantize(dtype)) {
+        throw py::type_error("expact quantize dtype");
+    }
+    PyObject* ob = PyDict_GetItemString(dtype->metadata, "mgb_dtype");
+    if (!PyDict_CheckExact(ob)) {
+        throw py::type_error("mgb_dtype is not dict");
+    }
+    Py_INCREF(ob);
+    return ob;
+}
+
+double _get_scale(PyArray_Descr* dtype) {
+    PyObject* ob = _get_mgb_dtype(dtype);
+    PyObject* scale = PyDict_GetItemString(ob, "scale");
+    if (!scale) {
+        Py_DECREF(ob);
+        throw py::key_error("scale");
+    }
+    if (!PyFloat_Check(scale)) {
+        Py_DECREF(ob);
+        throw py::type_error("scale is not float");
+    }
+    double ret = PyFloat_AsDouble(scale);
+    Py_DECREF(ob);
+    return ret;
+}
+
+long _get_zero_point(PyArray_Descr* dtype) {
+    PyObject* ob = _get_mgb_dtype(dtype);
+    PyObject* name = PyDict_GetItemString(ob, "name");
+    if (!name) {
+        Py_DECREF(ob);
+        throw py::key_error("name");
+    }
+    const char* s = PyUnicode_AsUTF8(name);
+    if (strcmp(s, "Quantized8Asymm") != 0 && strcmp(s, "Quantized4Asymm") != 0) {
+        Py_DECREF(ob);
+        throw py::value_error(ssprintf("expect name to be \"Quantized8Asymm\" or \"Quantized4Asymm\", got %s", s));
+    }
+    PyObject* zp = PyDict_GetItemString(ob, "zero_point");
+    if (!zp) {
+        Py_DECREF(ob);
+        throw py::key_error("zero_point");
+    }
+    long ret = PyLong_AsLong(zp);
+    Py_DECREF(ob);
+    return ret;
+}
+
+bool _is_dtype_equal(PyArray_Descr* dt1, PyArray_Descr* dt2) {
+    bool q1 = _is_quantize(dt1),
+         q2 = _is_quantize(dt2);
+    if (q1 && q2) {
+        if (_get_scale(dt1) != _get_scale(dt2)) {
+            return false;
+        }
+        PyObject* zp1 = PyDict_GetItemString(
+            PyDict_GetItemString(dt1->metadata, "mgb_dtype"), "zero_point");
+        PyObject* zp2 = PyDict_GetItemString(
+            PyDict_GetItemString(dt2->metadata, "mgb_dtype"), "zero_point");
+        if (!zp1 || !zp2) {
+            throw py::key_error("zero_point");
+        }
+        return PyLong_AsLong(zp1) == PyLong_AsLong(zp2);
+    }
+    if (!q1 && !q2) {
+        return dt1->type_num == dt2->type_num;
+    }
+    return false;
+}
+
+template<auto f>
+struct _wrap {
+    static constexpr size_t n_args = []() {
+        using F = decltype(f);
+        using T = PyArray_Descr*;
+        static_assert(std::is_pointer<F>::value);
+        if constexpr (std::is_invocable<F, T>::value) {
+            return 1;
+        } else if constexpr (std::is_invocable<F, T, T>::value) {
+            return 2;
+        } else {
+            static_assert(!std::is_same_v<F, F>, "unreachable");
+        }
+    }();
+
+    static PyObject* impl(PyObject* self, PyObject*const* args, size_t nargs) {
+        if (nargs != n_args) {
+            PyErr_Format(PyExc_ValueError, "expected %lu arguments", n_args);
+            return nullptr;
+        }
+        for (size_t i=0; i<nargs; ++i) {
+            if (args[i] == Py_None) {
+                PyErr_SetString(PyExc_ValueError, "can not convert null PyObject to numpy dtype");
+                return nullptr;
+            }
+        }
+        try {
+            PyArray_Descr *dt1;
+            if(!PyArray_DescrConverter(args[0], &dt1)) {
+                throw ConversionError(ssprintf("can not convert to numpy.dtype from %s",
+                            args[0]->ob_type->tp_name));
+            }
+            if constexpr (n_args == 1) {
+                auto res = (*f)(dt1);
+                Py_DECREF(dt1);
+                return py::cast(res).release().ptr();
+            } else {
+                PyArray_Descr *dt2;            
+                if(!PyArray_DescrConverter(args[1], &dt2)) {
+                    Py_DECREF(dt1);
+                    throw ConversionError(ssprintf("can not convert to numpy.dtype from %s",
+                                args[1]->ob_type->tp_name));
+                }
+                auto&& res = (*f)(dt1, dt2);
+                Py_DECREF(dt1);
+                Py_DECREF(dt2);
+                return py::cast(res).release().ptr();
+            }
+        } catch (std::exception& e) {
+            PyErr_SetString(PyExc_RuntimeError, e.what());
+            return nullptr;
+        }
+    }
+};
+
+} // anonymous namespace
+
+void init_dtypes(py::module m) {
+    static PyMethodDef method_defs[] = {
+        {"is_quantize",     (PyCFunction)_wrap<&_is_quantize>::impl,    METH_FASTCALL, nullptr},
+        {"get_scale",       (PyCFunction)_wrap<&_get_scale>::impl,      METH_FASTCALL, nullptr},
+        {"get_zero_point",  (PyCFunction)_wrap<&_get_zero_point>::impl, METH_FASTCALL, nullptr},
+        {"is_dtype_equal",  (PyCFunction)_wrap<&_is_dtype_equal>::impl, METH_FASTCALL, nullptr},
+        {nullptr, nullptr, 0, nullptr}
+    };
+    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);    
+        }
+    }
+}
+
+} // namespace mgb

imperative/python/src/numpy_dtypes.h

@@ -36,6 +36,7 @@ namespace mgb {
     int npy_num_intb##n();
 FOREACH_MGB_LOW_BIT(DEFINE_NPY_INTBX)
 #undef DEFINE_NPY_INTBX
+    void init_dtypes(pybind11::module m);
     void init_npy_num_intbx(pybind11::module m);
 
     //! numpy type num for bfloat16 type

imperative/python/src/tensor.cpp

@@ -9,16 +9,22 @@
  * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  */
 
+#include "megbrain/dtype.h"
+#include "megbrain/common.h"
+
 #include "./tensor.h"
 #include "./grad.h"
 #include "./trace.h"
 #include "./common.h"
 #include "./numpy_dtypes.h"
 #include "./graph_rt.h"
+#include "./helper.h"
 
 #include <pybind11/numpy.h>
 #include <pybind11/operators.h>
-#include "./helper.h"
+
+#include <unordered_map>
+
 namespace py = pybind11;
 
 namespace mgb::imperative::python {
@@ -413,6 +419,198 @@ struct TensorWeakRef {
     }
 };
 
+/* ============== 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;
+    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::cast_safe(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;
+            }
+            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];
+        TensorWrapper* tw = TensorWrapper::cast_safe(handle);
+        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;
+    }
+}
 
 void init_tensor(py::module m) {
     interpreter_for_py = interpreter::Interpreter::inst().create_channel();
@@ -444,10 +642,19 @@ void init_tensor(py::module m) {
         .def(py::init<const TensorWrapper&>())
         .def("__call__", &TensorWeakRef::operator());
 
-    static PyMethodDef apply_def{"apply", (PyCFunction)py_apply, METH_FASTCALL, nullptr};
-    auto* apply_func = PyCFunction_NewEx(&apply_def, nullptr, nullptr);
-    if (!apply_func) throw py::error_already_set();
-    py::setattr(m, "apply", apply_func);
+    static PyMethodDef method_defs[] = {
+        {"apply", (PyCFunction)py_apply, METH_FASTCALL, nullptr},
+        {"dtype_promotion", (PyCFunction)dtype_promotion, METH_FASTCALL, nullptr},
+        {"get_device", (PyCFunction)get_device, METH_FASTCALL, nullptr},
+        {nullptr, nullptr, 0, nullptr}
+    };
+    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);
+        }
+    }
 
     m.def("_set_swap_flag",
           [](bool flag) { interpreter_for_py->set_swap_flag(flag); });

imperative/python/test/unit/core/test_dtype_quant.py

@@ -113,7 +113,7 @@ def test_quint8_typecvt():
     data = np.random.random(shape).astype(np.float32) * 5 - 1
 
     def typecvt(x, dt=None):
-        (y,) = apply(ops.TypeCvt(dtype=dt), x)
+        (y,) = G.apply_normal_op(ops.TypeCvt(dtype=dt), x)
         return y
 
     # convert to quint8
@@ -194,7 +194,7 @@ def test_quint4_typecvt():
     data = np.random.random(shape).astype(np.float32) * 5 - 1
 
     def typecvt(x, dt=None):
-        (y,) = apply(ops.TypeCvt(dtype=dt), x)
+        (y,) = G.apply_normal_op(ops.TypeCvt(dtype=dt), x)
         return y
 
     # convert to quint4