gstaichi 2.1.1__cp311-cp311-macosx_11_0_arm64.whl

gstaichi/lang/impl.py

@@ -0,0 +1,1245 @@
+import numbers
+from types import FunctionType, MethodType
+from typing import Any, Iterable, Sequence
+
+import numpy as np
+
+from gstaichi._lib import core as _ti_core
+from gstaichi._lib.core.gstaichi_python import (
+    DataTypeCxx,
+    Function,
+    KernelCxx,
+    Program,
+)
+from gstaichi._snode.fields_builder import FieldsBuilder
+from gstaichi.lang._ndarray import ScalarNdarray
+from gstaichi.lang._ndrange import GroupedNDRange, _Ndrange
+from gstaichi.lang._texture import RWTextureAccessor
+from gstaichi.lang.any_array import AnyArray
+from gstaichi.lang.exception import (
+    GsTaichiCompilationError,
+    GsTaichiRuntimeError,
+    GsTaichiSyntaxError,
+    GsTaichiTypeError,
+)
+from gstaichi.lang.expr import Expr, make_expr_group
+from gstaichi.lang.field import Field, ScalarField
+from gstaichi.lang.kernel_arguments import SparseMatrixProxy
+from gstaichi.lang.kernel_impl import BoundGsTaichiCallable, GsTaichiCallable, Kernel
+from gstaichi.lang.matrix import (
+    Matrix,
+    MatrixField,
+    MatrixNdarray,
+    MatrixType,
+    Vector,
+    VectorNdarray,
+    make_matrix,
+)
+from gstaichi.lang.mesh import (
+    ConvType,
+    MeshElementFieldProxy,
+    MeshInstance,
+    MeshRelationAccessProxy,
+    MeshReorderedMatrixFieldProxy,
+    MeshReorderedScalarFieldProxy,
+    element_type_name,
+)
+from gstaichi.lang.simt.block import SharedArray
+from gstaichi.lang.snode import SNode
+from gstaichi.lang.struct import Struct, StructField, _IntermediateStruct
+from gstaichi.lang.util import (
+    cook_dtype,
+    get_traceback,
+    gstaichi_scope,
+    is_gstaichi_class,
+    python_scope,
+    warning,
+)
+from gstaichi.types.enums import SNodeGradType
+from gstaichi.types.primitive_types import (
+    all_types,
+    f16,
+    f32,
+    f64,
+    i32,
+    i64,
+    u8,
+    u32,
+    u64,
+)
+
+
+@gstaichi_scope
+def expr_init_shared_array(shape, element_type):
+    ast_builder = get_runtime().compiling_callable.ast_builder()
+    debug_info = _ti_core.DebugInfo(get_runtime().get_current_src_info())
+    return ast_builder.expr_alloca_shared_array(shape, element_type, debug_info)
+
+
+@gstaichi_scope
+def expr_init(rhs):
+    compiling_callable = get_runtime().compiling_callable
+    if rhs is None:
+        return Expr(
+            compiling_callable.ast_builder().expr_alloca(_ti_core.DebugInfo(get_runtime().get_current_src_info()))
+        )
+    if isinstance(rhs, Matrix) and (hasattr(rhs, "_DIM")):
+        return Matrix(*rhs.to_list(), ndim=rhs.ndim)  # type: ignore
+    if isinstance(rhs, Matrix):
+        return make_matrix(rhs.to_list())
+    if isinstance(rhs, SharedArray):
+        return rhs
+    if isinstance(rhs, Struct):
+        return Struct(rhs.to_dict(include_methods=True, include_ndim=True))
+    if isinstance(rhs, list):
+        return [expr_init(e) for e in rhs]
+    if isinstance(rhs, tuple):
+        return tuple(expr_init(e) for e in rhs)
+    if isinstance(rhs, dict):
+        return dict((key, expr_init(val)) for key, val in rhs.items())
+    if isinstance(rhs, _ti_core.DataTypeCxx):
+        return rhs
+    if isinstance(rhs, _ti_core.Arch):
+        return rhs
+    if isinstance(rhs, _Ndrange):
+        return rhs
+    if isinstance(rhs, MeshElementFieldProxy):
+        return rhs
+    if isinstance(rhs, MeshRelationAccessProxy):
+        return rhs
+    if hasattr(rhs, "_data_oriented"):
+        return rhs
+    return Expr(
+        compiling_callable.ast_builder().expr_var(
+            Expr(rhs).ptr, _ti_core.DebugInfo(get_runtime().get_current_src_info())
+        )
+    )
+
+
+@gstaichi_scope
+def expr_init_func(rhs):  # temporary solution to allow passing in fields as arguments
+    if isinstance(rhs, Field):
+        return rhs
+    return expr_init(rhs)
+
+
+def begin_frontend_struct_for(ast_builder, group, loop_range):
+    if not isinstance(loop_range, (AnyArray, Field, SNode, RWTextureAccessor, _Root)):
+        raise TypeError(
+            f"Cannot loop over the object {type(loop_range)} in GsTaichi scope. Only GsTaichi fields (via template) or dense arrays (via types.ndarray) are supported."
+        )
+    if group.size() != len(loop_range.shape):
+        raise IndexError(
+            "Number of struct-for indices does not match loop variable dimensionality "
+            f"({group.size()} != {len(loop_range.shape)}). Maybe you wanted to "
+            'use "for I in ti.grouped(x)" to group all indices into a single vector I?'
+        )
+    dbg_info = _ti_core.DebugInfo(get_runtime().get_current_src_info())
+    if isinstance(loop_range, (AnyArray, RWTextureAccessor)):
+        ast_builder.begin_frontend_struct_for_on_external_tensor(group, loop_range._loop_range(), dbg_info)
+    else:
+        ast_builder.begin_frontend_struct_for_on_snode(group, loop_range._loop_range(), dbg_info)
+
+
+def begin_frontend_if(ast_builder, cond, stmt_dbg_info):
+    assert ast_builder is not None
+    if is_gstaichi_class(cond):
+        raise ValueError(
+            "The truth value of vectors/matrices is ambiguous.\n"
+            "Consider using `any` or `all` when comparing vectors/matrices:\n"
+            "    if all(x == y):\n"
+            "or\n"
+            "    if any(x != y):\n"
+        )
+    ast_builder.begin_frontend_if(Expr(cond).ptr, stmt_dbg_info)
+
+
+@gstaichi_scope
+def _calc_slice(index, default_stop):
+    start, stop, step = index.start or 0, index.stop or default_stop, index.step or 1
+
+    def check_validity(x):
+        #  TODO(mzmzm): support variable in slice
+        if isinstance(x, Expr):
+            raise GsTaichiCompilationError(
+                "GsTaichi does not support variables in slice now, please use constant instead of it."
+            )
+
+    _ = check_validity(start), check_validity(stop), check_validity(step)
+    return [_ for _ in range(start, stop, step)]
+
+
+def validate_subscript_index(value, index):
+    if isinstance(value, Field):
+        # field supports negative indices
+        return
+
+    if isinstance(index, Expr):
+        return
+
+    if isinstance(index, Iterable):
+        for ind in index:
+            validate_subscript_index(value, ind)
+
+    if isinstance(index, slice):
+        validate_subscript_index(value, index.start)
+        validate_subscript_index(value, index.stop)
+
+    if isinstance(index, int) and index < 0:
+        raise GsTaichiSyntaxError("Negative indices are not supported in GsTaichi kernels.")
+
+
+@gstaichi_scope
+def subscript(ast_builder, value, *_indices, skip_reordered=False):
+    dbg_info = _ti_core.DebugInfo(get_runtime().get_current_src_info())
+    ast_builder = get_runtime().compiling_callable.ast_builder()
+    # Directly evaluate in Python for non-GsTaichi types
+    if not isinstance(
+        value,
+        (
+            Expr,
+            Field,
+            AnyArray,
+            SparseMatrixProxy,
+            MeshElementFieldProxy,
+            MeshRelationAccessProxy,
+            SharedArray,
+        ),
+    ):
+        if len(_indices) == 1:
+            _indices = _indices[0]
+        return value.__getitem__(_indices)
+
+    has_slice = False
+
+    flattened_indices = []
+    for _index in _indices:
+        if isinstance(_index, Matrix):
+            ind = _index.to_list()
+        elif isinstance(_index, slice):
+            ind = [_index]
+            has_slice = True
+        else:
+            ind = [_index]
+        flattened_indices += ind
+    indices = tuple(flattened_indices)
+    validate_subscript_index(value, indices)
+
+    if len(indices) == 1 and indices[0] is None:
+        indices = ()
+
+    indices_expr_group = None
+    if has_slice:
+        if not (isinstance(value, Expr) and value.is_tensor()):
+            raise GsTaichiSyntaxError(f"The type {type(value)} do not support index of slice type")
+    else:
+        indices_expr_group = make_expr_group(*indices)
+
+    if isinstance(value, SharedArray):
+        return value.subscript(*indices)
+    if isinstance(value, MeshElementFieldProxy):
+        return value.subscript(*indices)  # type: ignore
+    if isinstance(value, MeshRelationAccessProxy):
+        return value.subscript(*indices)
+    if isinstance(value, (MeshReorderedScalarFieldProxy, MeshReorderedMatrixFieldProxy)) and not skip_reordered:
+        assert len(indices) > 0
+        reordered_index = tuple(
+            [
+                Expr(
+                    ast_builder.mesh_index_conversion(
+                        value.mesh_ptr, value.element_type, Expr(indices[0]).ptr, ConvType.g2r, dbg_info
+                    )
+                )
+            ]
+        )
+        return subscript(ast_builder, value, *reordered_index, skip_reordered=True)
+    if isinstance(value, SparseMatrixProxy):
+        return value.subscript(*indices)
+    if isinstance(value, Field):
+        _var = value._get_field_members()[0].ptr
+        snode = _var.snode()
+        if snode is None:
+            if _var.is_primal():
+                raise RuntimeError(f"{_var.get_expr_name()} has not been placed.")
+            else:
+                raise RuntimeError(
+                    f"Gradient {_var.get_expr_name()} has not been placed, check whether `needs_grad=True`"
+                )
+
+        assert indices_expr_group is not None
+        if isinstance(value, MatrixField):
+            return Expr(ast_builder.expr_subscript(value.ptr, indices_expr_group, dbg_info))
+        if isinstance(value, StructField):
+            entries = {k: subscript(ast_builder, v, *indices) for k, v in value._items}
+            entries["__struct_methods"] = value.struct_methods
+            return _IntermediateStruct(entries)
+        return Expr(ast_builder.expr_subscript(_var, indices_expr_group, dbg_info))
+    if isinstance(value, AnyArray):
+        assert indices_expr_group is not None
+        return Expr(ast_builder.expr_subscript(value.ptr, indices_expr_group, dbg_info))
+    assert isinstance(value, Expr)
+    # Index into TensorType
+    # value: IndexExpression with ret_type = TensorType
+    assert value.is_tensor()
+
+    if has_slice:
+        shape = value.get_shape()
+        dim = len(shape)
+        assert dim == len(indices)
+        indices = [
+            _calc_slice(index, shape[i]) if isinstance(index, slice) else index for i, index in enumerate(indices)
+        ]
+        if dim == 1:
+            assert isinstance(indices[0], list)
+            multiple_indices = [make_expr_group(i) for i in indices[0]]
+            return_shape = (len(indices[0]),)
+        else:
+            assert dim == 2
+            if isinstance(indices[0], list) and isinstance(indices[1], list):
+                multiple_indices = [make_expr_group(i, j) for i in indices[0] for j in indices[1]]
+                return_shape = (len(indices[0]), len(indices[1]))
+            elif isinstance(indices[0], list):  # indices[1] is not list
+                multiple_indices = [make_expr_group(i, indices[1]) for i in indices[0]]
+                return_shape = (len(indices[0]),)
+            else:  # indices[0] is not list while indices[1] is list
+                multiple_indices = [make_expr_group(indices[0], j) for j in indices[1]]
+                return_shape = (len(indices[1]),)
+        return Expr(
+            _ti_core.subscript_with_multiple_indices(
+                value.ptr,
+                multiple_indices,
+                return_shape,
+                dbg_info,
+            )
+        )
+    return Expr(ast_builder.expr_subscript(value.ptr, indices_expr_group, dbg_info))
+
+
+class SrcInfoGuard:
+    def __init__(self, info_stack, info):
+        self.info_stack = info_stack
+        self.info = info
+
+    def __enter__(self):
+        self.info_stack.append(self.info)
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.info_stack.pop()
+
+
+class PyGsTaichi:
+    def __init__(self, kernels=None):
+        self.materialized = False
+        self._prog: Program | None = None
+        self.src_info_stack = []
+        self.inside_kernel: bool = False
+        self._compiling_callable: KernelCxx | Kernel | Function | None = None
+        self._current_kernel: "Kernel | None" = None
+        self.global_vars = []
+        self.grad_vars = []
+        self.dual_vars = []
+        self.matrix_fields = []
+        self.default_fp = f32
+        self.default_ip = i32
+        self.default_up = u32
+        self.print_full_traceback: bool = False
+        self.target_tape = None
+        self.fwd_mode_manager = None
+        self.grad_replaced = False
+        self.kernels: list[Kernel] = kernels or []
+        self._signal_handler_registry = None
+        self.unfinalized_fields_builder = {}
+        self.short_circuit_operators: bool = False
+        self.unrolling_limit: int = 0
+        self.src_ll_cache: bool = True
+
+    @property
+    def compiling_callable(self) -> KernelCxx | Kernel | Function:
+        if self._compiling_callable is None:
+            raise GsTaichiRuntimeError(
+                "_compiling_callable attribute not initialized. Maybe you forgot to call `ti.init()` first?"
+            )
+        return self._compiling_callable
+
+    @property
+    def prog(self) -> Program:
+        if self._prog is None:
+            raise GsTaichiRuntimeError("_prog attribute not initialized. Maybe you forgot to call `ti.init()` first?")
+        return self._prog
+
+    @property
+    def current_kernel(self) -> Kernel:
+        if self._current_kernel is None:
+            raise GsTaichiRuntimeError(
+                "_current_kernel attribute not initialized. Maybe you forgot to call `ti.init()` first?"
+            )
+        return self._current_kernel
+
+    def initialize_fields_builder(self, builder):
+        self.unfinalized_fields_builder[builder] = get_traceback(2)
+
+    def clear_compiled_functions(self):
+        for k in self.kernels:
+            k.materialized_kernels.clear()
+
+    def finalize_fields_builder(self, builder):
+        self.unfinalized_fields_builder.pop(builder)
+
+    def validate_fields_builder(self):
+        for builder, tb in self.unfinalized_fields_builder.items():
+            if builder == _root_fb:
+                continue
+
+            raise GsTaichiRuntimeError(
+                f"Field builder {builder} is not finalized. " f"Please call finalize() on it. Traceback:\n{tb}"
+            )
+
+    def get_num_compiled_functions(self):
+        count = 0
+        for k in self.kernels:
+            count += len(k.materialized_kernels)
+        return count
+
+    def src_info_guard(self, info):
+        return SrcInfoGuard(self.src_info_stack, info)
+
+    def get_current_src_info(self):
+        return self.src_info_stack[-1]
+
+    def set_default_fp(self, fp):
+        assert fp in [f16, f32, f64]
+        self.default_fp = fp
+        default_cfg().default_fp = self.default_fp
+
+    def set_default_ip(self, ip):
+        assert ip in [i32, i64]
+        self.default_ip = ip
+        self.default_up = u32 if ip == i32 else u64
+        default_cfg().default_ip = self.default_ip
+        default_cfg().default_up = self.default_up
+
+    def create_program(self):
+        if self._prog is None:
+            self._prog = _ti_core.Program()
+
+    @staticmethod
+    def materialize_root_fb(is_first_call):
+        if root.finalized:
+            return
+        if not is_first_call and root.empty:
+            # We have to forcefully finalize when `is_first_call` is True (even
+            # if the root itself is empty), so that there is a valid struct
+            # llvm::Module, if no field has been declared before the first kernel
+            # invocation. Example case:
+            # https://github.com/taichi-dev/gstaichi/blob/27bb1dc3227d9273a79fcb318fdb06fd053068f5/tests/python/test_ad_basics.py#L260-L266
+            return
+
+        if get_runtime().prog.config().debug:
+            if not root.finalized:
+                root._allocate_adjoint_checkbit()
+
+        root.finalize(raise_warning=not is_first_call)
+        global _root_fb
+        _root_fb = FieldsBuilder()
+
+    @staticmethod
+    def _get_tb(_var):
+        return getattr(_var, "declaration_tb", str(_var.ptr))
+
+    def _check_field_not_placed(self):
+        not_placed = []
+        for _var in self.global_vars:
+            if _var.ptr.snode() is None:
+                not_placed.append(self._get_tb(_var))
+
+        if len(not_placed):
+            bar = "=" * 44 + "\n"
+            raise RuntimeError(
+                f"These field(s) are not placed:\n{bar}"
+                + f"{bar}".join(not_placed)
+                + f"{bar}Please consider specifying a shape for them. E.g.,"
+                + "\n\n  x = ti.field(float, shape=(2, 3))"
+            )
+
+    def _check_gradient_field_not_placed(self, gradient_type):
+        not_placed = set()
+        gradient_vars = []
+        if gradient_type == "grad":
+            gradient_vars = self.grad_vars
+        elif gradient_type == "dual":
+            gradient_vars = self.dual_vars
+        for _var in gradient_vars:
+            if _var.ptr.snode() is None:
+                not_placed.add(self._get_tb(_var))
+
+        if len(not_placed):
+            bar = "=" * 44 + "\n"
+            raise RuntimeError(
+                f"These field(s) requrie `needs_{gradient_type}=True`, however their {gradient_type} field(s) are not placed:\n{bar}"
+                + f"{bar}".join(not_placed)
+                + f"{bar}Please consider place the {gradient_type} field(s). E.g.,"
+                + "\n\n  ti.root.dense(ti.i, 1).place(x.{gradient_type})"
+                + "\n\n Or specify a shape for the field(s). E.g.,"
+                + "\n\n  x = ti.field(float, shape=(2, 3), needs_{gradient_type}=True)"
+            )
+
+    def _check_matrix_field_member_shape(self):
+        for _field in self.matrix_fields:
+            shapes = [_field.get_scalar_field(i, j).shape for i in range(_field.n) for j in range(_field.m)]
+            if any(shape != shapes[0] for shape in shapes):
+                raise RuntimeError(
+                    "Members of the following field have different shapes "
+                    + f"{shapes}:\n{self._get_tb(_field._get_field_members()[0])}"
+                )
+
+    def _calc_matrix_field_dynamic_index_stride(self):
+        for _field in self.matrix_fields:
+            _field._calc_dynamic_index_stride()
+
+    def materialize(self):
+        self.materialize_root_fb(not self.materialized)
+        self.materialized = True
+
+        self.validate_fields_builder()
+
+        self._check_field_not_placed()
+        self._check_gradient_field_not_placed("grad")
+        self._check_gradient_field_not_placed("dual")
+        self._check_matrix_field_member_shape()
+        self._calc_matrix_field_dynamic_index_stride()
+        self.global_vars = []
+        self.grad_vars = []
+        self.dual_vars = []
+        self.matrix_fields = []
+
+    def _register_signal_handlers(self):
+        if self._signal_handler_registry is None:
+            self._signal_handler_registry = _ti_core.HackedSignalRegister()
+
+    def clear(self):
+        if self._prog:
+            self._prog.finalize()
+            self._prog = None
+        self._signal_handler_registry = None
+        self.materialized = False
+
+    def sync(self):
+        self.materialize()
+        assert self._prog is not None
+        self._prog.synchronize()
+
+
+pygstaichi = PyGsTaichi()
+
+
+def get_runtime() -> PyGsTaichi:
+    return pygstaichi
+
+
+def reset():
+    global pygstaichi
+    old_kernels = pygstaichi.kernels
+    pygstaichi.clear()
+    pygstaichi = PyGsTaichi(old_kernels)
+    for k in old_kernels:
+        k.reset()
+    _ti_core.reset_default_compile_config()
+
+
+@gstaichi_scope
+def static_print(*args, __p=print, **kwargs):
+    """The print function in GsTaichi scope.
+
+    This function is called at compile time and has no runtime overhead.
+    """
+    __p(*args, **kwargs)
+
+
+# we don't add @gstaichi_scope decorator for @ti.pyfunc to work
+def static_assert(cond, msg=None):
+    """Throw AssertionError when `cond` is False.
+
+    This function is called at compile time and has no runtime overhead.
+    The bool value in `cond` must can be determined at compile time.
+
+    Args:
+        cond (bool): an expression with a bool value.
+        msg (str): assertion message.
+
+    Example::
+
+        >>> year = 2001
+        >>> @ti.kernel
+        >>> def test():
+        >>>     ti.static_assert(year % 4 == 0, "the year must be a lunar year")
+        AssertionError: the year must be a lunar year
+    """
+    if isinstance(cond, Expr):
+        raise GsTaichiTypeError("Static assert with non-static condition")
+    if msg is not None:
+        assert cond, msg
+    else:
+        assert cond
+
+
+def inside_kernel():
+    return pygstaichi.inside_kernel
+
+
+def index_nd(dim):
+    return axes(*range(dim))
+
+
+class _UninitializedRootFieldsBuilder:
+    def __getattr__(self, item):
+        if item == "__qualname__":
+            # For sphinx docstring extraction.
+            return "_UninitializedRootFieldsBuilder"
+        raise GsTaichiRuntimeError("Please call init() first")
+
+
+# `root` initialization must be delayed until after the program is
+# created. Unfortunately, `root` exists in both gstaichi.lang.impl module and
+# the top-level gstaichi module at this point; so if `root` itself is written, we
+# would have to make sure that `root` in all the modules get updated to the same
+# instance. This is an error-prone process.
+#
+# To avoid this situation, we create `root` once during the import time, and
+# never write to it. The core part, `_root_fb`, is the one whose initialization
+# gets delayed. `_root_fb` will only exist in the gstaichi.lang.impl module, so
+# writing to it is would result in less for maintenance cost.
+#
+# `_root_fb` will be overridden inside :func:`gstaichi.lang.init`.
+_root_fb = _UninitializedRootFieldsBuilder()
+
+
+def deactivate_all_snodes():
+    """Recursively deactivate all SNodes."""
+    for root_fb in FieldsBuilder._finalized_roots():
+        root_fb.deactivate_all()
+
+
+class _Root:
+    """Wrapper around the default root FieldsBuilder instance."""
+
+    @staticmethod
+    def parent(n=1):
+        """Same as :func:`gstaichi.SNode.parent`"""
+        assert isinstance(_root_fb, FieldsBuilder)
+        return _root_fb.root.parent(n)
+
+    @staticmethod
+    def _loop_range():
+        """Same as :func:`gstaichi.SNode.loop_range`"""
+        assert isinstance(_root_fb, FieldsBuilder)
+        return _root_fb.root._loop_range()
+
+    @staticmethod
+    def _get_children():
+        """Same as :func:`gstaichi.SNode.get_children`"""
+        assert isinstance(_root_fb, FieldsBuilder)
+        return _root_fb.root._get_children()
+
+    # TODO: Record all of the SNodeTrees that finalized under 'ti.root'
+    @staticmethod
+    def deactivate_all():
+        warning("""'ti.root.deactivate_all()' would deactivate all finalized snodes.""")
+        deactivate_all_snodes()
+
+    @property
+    def shape(self):
+        """Same as :func:`gstaichi.SNode.shape`"""
+        assert isinstance(_root_fb, FieldsBuilder)
+        return _root_fb.root.shape
+
+    @property
+    def _id(self):
+        assert isinstance(_root_fb, FieldsBuilder)
+        return _root_fb.root._id
+
+    def __getattr__(self, item):
+        return getattr(_root_fb, item)
+
+    def __repr__(self):
+        return "ti.root"
+
+
+root = _Root()
+"""Root of the declared GsTaichi :func:`~gstaichi.lang.impl.field`s.
+
+See also https://docs.taichi-lang.org/docs/layout
+
+Example::
+
+    >>> x = ti.field(ti.f32)
+    >>> ti.root.pointer(ti.ij, 4).dense(ti.ij, 8).place(x)
+"""
+
+
+def _create_snode(axis_seq: Sequence[int], shape_seq: Sequence[numbers.Number], same_level: bool):
+    dim = len(axis_seq)
+    assert dim == len(shape_seq)
+    snode = root
+    if same_level:
+        snode = snode.dense(axes(*axis_seq), shape_seq)
+    else:
+        for i in range(dim):
+            snode = snode.dense(axes(axis_seq[i]), (shape_seq[i],))
+    return snode
+
+
+@python_scope
+def create_field_member(dtype, name, needs_grad, needs_dual):
+    dtype = cook_dtype(dtype)
+
+    # primal
+    prog = get_runtime().prog
+
+    x = Expr(prog.make_id_expr(""))
+    x.declaration_tb = get_traceback(stacklevel=4)
+    x.ptr = _ti_core.expr_field(x.ptr, dtype)
+    x.ptr.set_name(name)
+    x.ptr.set_grad_type(SNodeGradType.PRIMAL)
+    pygstaichi.global_vars.append(x)
+
+    x_grad = None
+    x_dual = None
+    # The x_grad_checkbit is used for global data access rule checker
+    x_grad_checkbit = None
+    if _ti_core.is_real(dtype):
+        # adjoint
+        x_grad = Expr(prog.make_id_expr(""))
+        x_grad.declaration_tb = get_traceback(stacklevel=4)
+        x_grad.ptr = _ti_core.expr_field(x_grad.ptr, dtype)
+        x_grad.ptr.set_name(name + ".grad")
+        x_grad.ptr.set_grad_type(SNodeGradType.ADJOINT)
+        x.ptr.set_adjoint(x_grad.ptr)
+        if needs_grad:
+            pygstaichi.grad_vars.append(x_grad)
+
+        if prog.config().debug:
+            # adjoint checkbit
+            x_grad_checkbit = Expr(prog.make_id_expr(""))
+            dtype = u8
+            if prog.config().arch == _ti_core.vulkan:
+                dtype = i32
+            x_grad_checkbit.ptr = _ti_core.expr_field(x_grad_checkbit.ptr, cook_dtype(dtype))
+            x_grad_checkbit.ptr.set_name(name + ".grad_checkbit")
+            x_grad_checkbit.ptr.set_grad_type(SNodeGradType.ADJOINT_CHECKBIT)
+            x.ptr.set_adjoint_checkbit(x_grad_checkbit.ptr)
+
+        # dual
+        x_dual = Expr(prog.make_id_expr(""))
+        x_dual.ptr = _ti_core.expr_field(x_dual.ptr, dtype)
+        x_dual.ptr.set_name(name + ".dual")
+        x_dual.ptr.set_grad_type(SNodeGradType.DUAL)
+        x.ptr.set_dual(x_dual.ptr)
+        if needs_dual:
+            pygstaichi.dual_vars.append(x_dual)
+    elif needs_grad or needs_dual:
+        raise GsTaichiRuntimeError(f"{dtype} is not supported for field with `needs_grad=True` or `needs_dual=True`.")
+
+    return x, x_grad, x_dual
+
+
+@python_scope
+def _field(
+    dtype,
+    shape=None,
+    order=None,
+    name="",
+    offset=None,
+    needs_grad=False,
+    needs_dual=False,
+):
+    x, x_grad, x_dual = create_field_member(dtype, name, needs_grad, needs_dual)
+    x = ScalarField(x)
+    if x_grad:
+        x_grad = ScalarField(x_grad)
+        x._set_grad(x_grad)
+    if x_dual:
+        x_dual = ScalarField(x_dual)
+        x._set_dual(x_dual)
+
+    if shape is None:
+        if offset is not None:
+            raise GsTaichiSyntaxError("shape cannot be None when offset is set")
+        if order is not None:
+            raise GsTaichiSyntaxError("shape cannot be None when order is set")
+    else:
+        if isinstance(shape, numbers.Number):
+            shape = (shape,)
+        if isinstance(offset, numbers.Number):
+            offset = (offset,)
+        dim = len(shape)
+        if offset is not None and dim != len(offset):
+            raise GsTaichiSyntaxError(
+                f"The dimensionality of shape and offset must be the same ({dim} != {len(offset)})"
+            )
+        axis_seq = []
+        shape_seq = []
+        if order is not None:
+            if dim != len(order):
+                raise GsTaichiSyntaxError(
+                    f"The dimensionality of shape and order must be the same ({dim} != {len(order)})"
+                )
+            if dim != len(set(order)):
+                raise GsTaichiSyntaxError("The axes in order must be different")
+            for ch in order:
+                axis = ord(ch) - ord("i")
+                if axis < 0 or axis >= dim:
+                    raise GsTaichiSyntaxError(f"Invalid axis {ch}")
+                axis_seq.append(axis)
+                shape_seq.append(shape[axis])
+        else:
+            axis_seq = list(range(dim))
+            shape_seq = list(shape)
+        same_level = order is None
+        _create_snode(axis_seq, shape_seq, same_level).place(x, offset=offset)
+        if needs_grad:
+            _create_snode(axis_seq, shape_seq, same_level).place(x_grad, offset=offset)
+        if needs_dual:
+            _create_snode(axis_seq, shape_seq, same_level).place(x_dual, offset=offset)
+    return x
+
+
+@python_scope
+def field(dtype, *args, **kwargs):
+    """Defines a GsTaichi field.
+
+    A GsTaichi field can be viewed as an abstract N-dimensional array, hiding away
+    the complexity of how its underlying :class:`~gstaichi.lang.snode.SNode` are
+    actually defined. The data in a GsTaichi field can be directly accessed by
+    a GsTaichi :func:`~gstaichi.lang.kernel_impl.kernel`.
+
+    See also https://docs.taichi-lang.org/docs/field
+
+    Args:
+        dtype (DataType): data type of the field. Note it can be vector or matrix types as well.
+        shape (Union[int, tuple[int]], optional): shape of the field.
+        order (str, optional): order of the shape laid out in memory.
+        name (str, optional): name of the field.
+        offset (Union[int, tuple[int]], optional): offset of the field domain.
+        needs_grad (bool, optional): whether this field participates in autodiff (reverse mode)
+            and thus needs an adjoint field to store the gradients.
+        needs_dual (bool, optional): whether this field participates in autodiff (forward mode)
+            and thus needs an dual field to store the gradients.
+
+    Example::
+
+        The code below shows how a GsTaichi field can be declared and defined::
+
+            >>> x1 = ti.field(ti.f32, shape=(16, 8))
+            >>> # Equivalently
+            >>> x2 = ti.field(ti.f32)
+            >>> ti.root.dense(ti.ij, shape=(16, 8)).place(x2)
+            >>>
+            >>> x3 = ti.field(ti.f32, shape=(16, 8), order='ji')
+            >>> # Equivalently
+            >>> x4 = ti.field(ti.f32)
+            >>> ti.root.dense(ti.j, shape=8).dense(ti.i, shape=16).place(x4)
+            >>>
+            >>> x5 = ti.field(ti.math.vec3, shape=(16, 8))
+
+    """
+    if isinstance(dtype, MatrixType):
+        if dtype.ndim == 1:
+            return Vector.field(dtype.n, dtype.dtype, *args, **kwargs)
+        return Matrix.field(dtype.n, dtype.m, dtype.dtype, *args, **kwargs)
+    return _field(dtype, *args, **kwargs)
+
+
+@python_scope
+def ndarray(dtype, shape, needs_grad=False):
+    """Defines a GsTaichi ndarray with scalar elements.
+
+    Args:
+        dtype (Union[DataType, MatrixType]): Data type of each element. This can be either a scalar type like ti.f32 or a compound type like ti.types.vector(3, ti.i32).
+        shape (Union[int, tuple[int]]): Shape of the ndarray.
+
+    Example:
+        The code below shows how a GsTaichi ndarray with scalar elements can be declared and defined::
+
+            >>> x = ti.ndarray(ti.f32, shape=(16, 8))  # ndarray of shape (16, 8), each element is ti.f32 scalar.
+            >>> vec3 = ti.types.vector(3, ti.i32)
+            >>> y = ti.ndarray(vec3, shape=(10, 2))  # ndarray of shape (10, 2), each element is a vector of 3 ti.i32 scalars.
+            >>> matrix_ty = ti.types.matrix(3, 4, float)
+            >>> z = ti.ndarray(matrix_ty, shape=(4, 5))  # ndarray of shape (4, 5), each element is a matrix of (3, 4) ti.float scalars.
+    """
+    # primal
+    if isinstance(shape, numbers.Number):
+        shape = (shape,)
+    if not all((isinstance(x, int) or isinstance(x, np.integer)) and x > 0 and x <= 2**31 - 1 for x in shape):
+        raise GsTaichiRuntimeError(f"{shape} is not a valid shape for ndarray")
+    if dtype in all_types:
+        dt = cook_dtype(dtype)
+        x = ScalarNdarray(dt, shape)
+    elif isinstance(dtype, MatrixType):
+        if dtype.ndim == 1:
+            x = VectorNdarray(dtype.n, dtype.dtype, shape)
+        else:
+            x = MatrixNdarray(dtype.n, dtype.m, dtype.dtype, shape)
+        dt = dtype.dtype
+    else:
+        raise GsTaichiRuntimeError(f"{dtype} is not supported as ndarray element type")
+    if needs_grad:
+        assert isinstance(dt, DataTypeCxx)
+        if not _ti_core.is_real(dt):
+            raise GsTaichiRuntimeError(
+                f"{dt} is not supported for ndarray with `needs_grad=True` or `needs_dual=True`."
+            )
+        x_grad = ndarray(dtype, shape, needs_grad=False)
+        x._set_grad(x_grad)
+    return x
+
+
+@gstaichi_scope
+def ti_format_list_to_content_entries(raw):
+    # return a pair of [content, format]
+    def entry2content(_var):
+        if isinstance(_var, str):
+            return [_var, None]
+        if isinstance(_var, list):
+            assert len(_var) == 2 and (isinstance(_var[1], str) or _var[1] is None)
+            _var[0] = Expr(_var[0]).ptr
+            return _var
+        return [Expr(_var).ptr, None]
+
+    def list_ti_repr(_var):
+        yield "["  # distinguishing tuple & list will increase maintenance cost
+        for i, v in enumerate(_var):
+            if i:
+                yield ", "
+            yield v
+        yield "]"
+
+    def vars2entries(_vars):
+        for _var in _vars:
+            # If the first element is '__ti_fmt_value__', this list is an Expr and its format.
+            if isinstance(_var, list) and len(_var) == 3 and isinstance(_var[0], str) and _var[0] == "__ti_fmt_value__":
+                # yield [Expr, format] as a whole and don't pass it to vars2entries() again
+                yield _var[1:]
+                continue
+            elif hasattr(_var, "__ti_repr__"):
+                res = _var.__ti_repr__()  # type: ignore
+            elif isinstance(_var, (list, tuple)):
+                # If the first element is '__ti_format__', this list is the result of ti_format.
+                if len(_var) > 0 and isinstance(_var[0], str) and _var[0] == "__ti_format__":
+                    res = _var[1:]
+                else:
+                    res = list_ti_repr(_var)
+            else:
+                yield _var
+                continue
+
+            for v in vars2entries(res):
+                yield v
+
+    def fused_string(entries):
+        accumated = ""
+        for entry in entries:
+            if isinstance(entry, str):
+                accumated += entry
+            else:
+                if accumated:
+                    yield accumated
+                    accumated = ""
+                yield entry
+        if accumated:
+            yield accumated
+
+    def extract_formats(entries):
+        contents, formats = zip(*entries)
+        return list(contents), list(formats)
+
+    entries = vars2entries(raw)
+    entries = fused_string(entries)
+    entries = [entry2content(entry) for entry in entries]
+    return extract_formats(entries)
+
+
+@gstaichi_scope
+def ti_print(*_vars, sep=" ", end="\n"):
+    def add_separators(_vars):
+        for i, _var in enumerate(_vars):
+            if i:
+                yield sep
+            yield _var
+        yield end
+
+    _vars = add_separators(_vars)
+    contents, formats = ti_format_list_to_content_entries(_vars)
+    ast_builder = get_runtime().compiling_callable.ast_builder()
+    debug_info = _ti_core.DebugInfo(get_runtime().get_current_src_info())
+    ast_builder.create_print(contents, formats, debug_info)
+
+
+@gstaichi_scope
+def ti_format(*args):
+    content = args[0]
+    mixed = args[1:]
+    new_mixed = []
+    args = []
+    for x in mixed:
+        # x is a (formatted) Expr
+        if isinstance(x, Expr) or (isinstance(x, list) and len(x) == 3 and x[0] == "__ti_fmt_value__"):
+            new_mixed.append("{}")
+            args.append(x)
+        else:
+            new_mixed.append(x)
+    content = content.format(*new_mixed)
+    res = content.split("{}")
+    assert len(res) == len(args) + 1, "Number of args is different from number of positions provided in string"
+
+    for i, arg in enumerate(args):
+        res.insert(i * 2 + 1, arg)
+    res.insert(0, "__ti_format__")
+    return res
+
+
+@gstaichi_scope
+def ti_assert(cond, msg, extra_args, dbg_info):
+    # Mostly a wrapper to help us convert from Expr (defined in Python) to
+    # _ti_core.Expr (defined in C++)
+    ast_builder = get_runtime().compiling_callable.ast_builder()
+    ast_builder.create_assert_stmt(Expr(cond).ptr, msg, extra_args, dbg_info)
+
+
+@gstaichi_scope
+def ti_int(_var):
+    if hasattr(_var, "__ti_int__"):
+        return _var.__ti_int__()
+    return int(_var)
+
+
+@gstaichi_scope
+def ti_bool(_var):
+    if hasattr(_var, "__ti_bool__"):
+        return _var.__ti_bool__()
+    return bool(_var)
+
+
+@gstaichi_scope
+def ti_float(_var):
+    if hasattr(_var, "__ti_float__"):
+        return _var.__ti_float__()
+    return float(_var)
+
+
+@gstaichi_scope
+def zero(x):
+    # TODO: get dtype from Expr and Matrix:
+    """Returns an array of zeros with the same shape and type as the input. It's also a scalar
+    if the input is a scalar.
+
+    Args:
+        x (Union[:mod:`~gstaichi.types.primitive_types`, :class:`~gstaichi.Matrix`]): The input.
+
+    Returns:
+        A new copy of the input but filled with zeros.
+
+    Example::
+
+        >>> x = ti.Vector([1, 1])
+        >>> @ti.kernel
+        >>> def test():
+        >>>     y = ti.zero(x)
+        >>>     print(y)
+        [0, 0]
+    """
+    return x * 0
+
+
+@gstaichi_scope
+def one(x):
+    """Returns an array of ones with the same shape and type as the input. It's also a scalar
+    if the input is a scalar.
+
+    Args:
+        x (Union[:mod:`~gstaichi.types.primitive_types`, :class:`~gstaichi.Matrix`]): The input.
+
+    Returns:
+        A new copy of the input but filled with ones.
+
+    Example::
+
+        >>> x = ti.Vector([0, 0])
+        >>> @ti.kernel
+        >>> def test():
+        >>>     y = ti.one(x)
+        >>>     print(y)
+        [1, 1]
+    """
+    return zero(x) + 1
+
+
+def axes(*x: int):
+    """Defines a list of axes to be used by a field.
+
+    Args:
+        *x: A list of axes to be activated
+
+    Note that GsTaichi has already provided a set of commonly used axes. For example,
+    `ti.ij` is just `axes(0, 1)` under the hood.
+    """
+    return [_ti_core.Axis(i) for i in x]
+
+
+Axis = _ti_core.Axis
+
+
+def static(x, *xs) -> Any:
+    """Evaluates a GsTaichi-scope expression at compile time.
+
+    `static()` is what enables the so-called metaprogramming in GsTaichi. It is
+    in many ways similar to ``constexpr`` in C++.
+
+    See also https://docs.taichi-lang.org/docs/meta.
+
+    Args:
+        x (Any): an expression to be evaluated
+        *xs (Any): for Python-ish swapping assignment
+
+    Example:
+        The most common usage of `static()` is for compile-time evaluation::
+
+            >>> cond = False
+            >>>
+            >>> @ti.kernel
+            >>> def run():
+            >>>     if ti.static(cond):
+            >>>         do_a()
+            >>>     else:
+            >>>         do_b()
+
+        Depending on the value of ``cond``, ``run()`` will be directly compiled
+        into either ``do_a()`` or ``do_b()``. Thus there won't be a runtime
+        condition check.
+
+        Another common usage is for compile-time loop unrolling::
+
+            >>> @ti.kernel
+            >>> def run():
+            >>>     for i in ti.static(range(3)):
+            >>>         print(i)
+            >>>
+            >>> # The above will be unrolled to:
+            >>> @ti.kernel
+            >>> def run():
+            >>>     print(0)
+            >>>     print(1)
+            >>>     print(2)
+    """
+    if len(xs):  # for python-ish pointer assign: x, y = ti.static(y, x)
+        return [static(x)] + [static(x) for x in xs]
+
+    if (
+        isinstance(
+            x,
+            (
+                bool,
+                int,
+                float,
+                range,
+                list,
+                tuple,
+                enumerate,
+                GroupedNDRange,
+                _Ndrange,
+                zip,
+                filter,
+                map,
+            ),
+        )
+        or x is None
+    ):
+        return x
+    if isinstance(x, (np.bool_, np.integer, np.floating)):
+        return x
+
+    if isinstance(x, AnyArray):
+        return x
+    if isinstance(x, Field):
+        return x
+    if isinstance(x, (FunctionType, MethodType, BoundGsTaichiCallable, GsTaichiCallable)):
+        return x
+    raise ValueError(f"Input to ti.static must be compile-time constants or global pointers, instead of {type(x)}")
+
+
+@gstaichi_scope
+def grouped(x):
+    """Groups the indices in the iterator returned by `ndrange()` into a 1-D vector.
+
+    This is often used when you want to iterate over all indices returned by `ndrange()`
+    in one `for` loop and a single index.
+
+    Args:
+        x (:func:`~gstaichi.ndrange`): an iterator object returned by `ti.ndrange`.
+
+    Example::
+        >>> # without ti.grouped
+        >>> for I in ti.ndrange(2, 3):
+        >>>     print(I)
+        prints 0, 1, 2, 3, 4, 5
+
+        >>> # with ti.grouped
+        >>> for I in ti.grouped(ti.ndrange(2, 3)):
+        >>>     print(I)
+        prints [0, 0], [0, 1], [0, 2], [1, 0], [1, 1], [1, 2]
+    """
+    if isinstance(x, _Ndrange):
+        return x.grouped()
+    return x
+
+
+def stop_grad(x):
+    """Stops computing gradients during back propagation.
+
+    Args:
+        x (:class:`~gstaichi.Field`): A field.
+    """
+    compiling_callable = get_runtime().compiling_callable
+    assert compiling_callable is not None
+    compiling_callable.ast_builder().stop_grad(x.snode.ptr)
+
+
+def current_cfg():
+    return get_runtime().prog.config()
+
+
+def default_cfg():
+    return _ti_core.default_compile_config()
+
+
+def call_internal(name, *args, with_runtime_context=True):
+    return expr_init(_ti_core.insert_internal_func_call(getattr(_ti_core.InternalOp, name), make_expr_group(args)))
+
+
+def get_cuda_compute_capability():
+    return _ti_core.query_int64("cuda_compute_capability")
+
+
+@gstaichi_scope
+def mesh_relation_access(mesh, from_index, to_element_type):
+    # to support ti.mesh_local and access mesh attribute as field
+    if isinstance(from_index, MeshInstance):
+        return getattr(from_index, element_type_name(to_element_type))
+    if isinstance(mesh, MeshInstance):
+        return MeshRelationAccessProxy(mesh, from_index, to_element_type)
+    raise RuntimeError("Relation access should be with a mesh instance!")
+
+
+__all__ = [
+    "axes",
+    "deactivate_all_snodes",
+    "field",
+    "grouped",
+    "ndarray",
+    "one",
+    "root",
+    "static",
+    "static_assert",
+    "static_print",
+    "stop_grad",
+    "zero",
+]