vllm/lib/python3.10/site-packages/cupyx/jit/_compile.py

import ast
import collections
import inspect
import linecache
import numbers
import re
import sys
from typing import Any, Dict, List, Optional, Sequence, Tuple, TypeVar, Union
import warnings
import types

import numpy

from cupy.exceptions import ComplexWarning
from cupy_backends.cuda.api import runtime
from cupy._core._codeblock import CodeBlock, _CodeType
from cupy._core import _kernel
from cupy._core._dtype import _raise_if_invalid_cast
from cupyx import jit
from cupyx.jit import _cuda_types
from cupyx.jit import _cuda_typerules
from cupyx.jit import _internal_types
from cupyx.jit._internal_types import Data
from cupyx.jit._internal_types import Constant
from cupyx.jit import _builtin_funcs
from cupyx.jit import _interface


_is_debug_mode = False

_typeclasses = (bool, numpy.bool_, numbers.Number)

if (3, 8) <= sys.version_info:
    from typing import Literal
    _CastingType = Optional[
        Literal['no', 'equiv', 'safe', 'same_kind', 'unsafe']]
else:
    _CastingType = str

Result = collections.namedtuple(
    'Result',
    [
        'func_name',
        'code',
        'return_type',
        'enable_cooperative_groups',
        'backend',
        'options',
        'jitify',
    ])


class _JitCompileError(Exception):

    def __init__(self, e, node):
        self.error_type = type(e)
        self.mes = str(e)
        self.node = node

    def reraise(self, pycode):
        start = self.node.lineno
        end = getattr(self.node, 'end_lineno', start)
        pycode = '\n'.join([
            (f'> {line}' if start <= i + 1 <= end else f'  {line}').rstrip()
            for i, line in enumerate(pycode.split('\n'))])
        raise self.error_type(self.mes + '\n\n' + pycode)


def transpile_function_wrapper(func):
    def new_func(node, *args, **kwargs):
        try:
            return func(node, *args, **kwargs)
        except _JitCompileError:
            raise
        except Exception as e:
            raise _JitCompileError(e, node)

    return new_func


def _parse_function_object(func):
    """Returns the tuple of ``ast.FunctionDef`` object and the source string
    for the given callable ``func``.

    ``func`` can be a ``def`` function or a ``lambda`` expression.

    The source is returned only for informational purposes (i.e., rendering
    an exception message in case of an error).
    """
    if not callable(func):
        raise ValueError('`func` must be a callable object.')

    try:
        # ``filename`` can be any of:
        # - A "real" file path on the filesystem
        # - "<stdin>" (within Python interpreter)
        # - "<ipython-input-XXXXXXXX>" (within IPython interpreter)
        filename = inspect.getsourcefile(func)
    except TypeError:
        # Built-in function or method, or inside Doctest
        filename = None

    if filename == '<stdin>':
        raise RuntimeError(
            f'JIT needs access to the Python source code for {func}'
            ' but it cannot be retrieved within the Python interactive'
            ' interpreter. Consider using IPython instead.')

    if func.__name__ != '<lambda>':
        lines, _ = inspect.getsourcelines(func)
        num_indent = len(lines[0]) - len(lines[0].lstrip())
        source = ''.join([
            line.replace(' ' * num_indent, '', 1) for line in lines])
        tree = ast.parse(source)
        assert isinstance(tree, ast.Module)
        assert len(tree.body) == 1
        return tree.body[0], source

    if filename is None:
        # filename is needed for lambdas.
        raise ValueError(
            f'JIT needs access to Python source code for {func}'
            ' but could not be located.\n'
            '(hint: it is likely you passed a built-in function or method)')

    # Extract the AST of the lambda from the AST of the whole source file
    # that defines that lambda.
    # This is needed because ``inspect.getsourcelines(lambda_expr)`` may
    # return unparsable code snippet.

    # Use ``linecache.getlines`` instead of directly opening a file to
    # support notebook environments.
    full_source = ''.join(linecache.getlines(filename))
    source, start_line = inspect.getsourcelines(func)
    end_line = start_line + len(source)
    source = ''.join(source)

    tree = ast.parse(full_source)

    nodes = [node for node in ast.walk(tree)
             if isinstance(node, ast.Lambda)
             and start_line <= node.lineno < end_line]
    if len(nodes) > 1:
        # TODO(kmaehashi): can be improved by heuristics (e.g. number of args)
        raise ValueError('Multiple callables are found near the'
                         f' definition of {func}, and JIT could not'
                         ' identify the source code for it.')
    node = nodes[0]
    return ast.FunctionDef(
        name='_lambda_kernel', args=node.args,
        body=[ast.Return(node.body)],
        decorator_list=[], returns=None, type_comment=None,
    ), source


class Generated:

    def __init__(self) -> None:
        # list of str
        self.codes: List[str] = []
        # (function, in_types) => Optional(function_name, return_type)
        self.device_function: \
            Dict[Tuple[Any, Tuple[_cuda_types.TypeBase, ...]],
                 Tuple[str, _cuda_types.TypeBase]] = {}
        # whether to use cooperative launch
        self.enable_cg = False
        # whether to include cooperative_groups.h
        self.include_cg = False
        # whether to include cooperative_groups/memcpy_async.h
        self.include_cg_memcpy_async = False
        # whether to include cuda/barrier
        self.include_cuda_barrier = False
        # compiler options
        self.options = ('-DCUPY_JIT_MODE', '--std=c++17',
                        # WAR: for compiling any CCCL header
                        '-DCUB_DISABLE_BF16_SUPPORT',)
        # workaround for hipRTC: as of ROCm 4.1.0 hipRTC still does not
        # recognize "-D", so we have to compile using hipcc...
        self.backend = 'nvcc' if runtime.is_hip else 'nvrtc'
        # workaround for CUB/libcudacxx headers: they can be compiled by NVRTC
        # but they need Jitify; Thrust headers can only be compiled by NVCC
        # for now. We keep Jitify off by default to reduce overhead.
        self.jitify = False

    def add_code(self, code: str) -> None:
        if code not in self.codes:
            self.codes.append(code)
            if len(self.codes) > jit._n_functions_upperlimit:
                raise ValueError("Number of functions exceeds upper limit.")


def transpile(func, attributes, mode, in_types, ret_type):
    """Transpiles the target function.

    Args:
        func (function): Target function.
        attributes (list of str): Attributes of the generated CUDA function.
        mode ('numpy' or 'cuda'): The rule for typecast.
        in_types (list of _cuda_types.TypeBase): Types of the arguments.
        ret_type (_cuda_types.TypeBase or None): Type of the return value.
    """
    generated = Generated()
    in_types = tuple(in_types)
    name, return_type = _transpile_func_obj(
        func, attributes, mode, in_types, ret_type, generated)
    func_name, _ = generated.device_function[(func, in_types)]
    code = '\n'.join(generated.codes)
    backend = generated.backend
    options = generated.options
    jitify = generated.jitify
    enable_cg = generated.enable_cg

    if _is_debug_mode:
        print(code)

    return Result(
        func_name=func_name, code=code, return_type=return_type, jitify=jitify,
        enable_cooperative_groups=enable_cg, backend=backend, options=options)


def _transpile_func_obj(func, attributes, mode, in_types, ret_type, generated):
    if (func, in_types) in generated.device_function:
        result = generated.device_function[(func, in_types)]
        if result is None:
            raise ValueError("Recursive function is not supported.")
        return result

    # Do sanity check first.
    tree, source = _parse_function_object(func)

    cvars = inspect.getclosurevars(func)
    consts = dict(**cvars.globals, **cvars.nonlocals, **cvars.builtins)
    attributes = ' '.join(attributes)
    name = tree.name
    if len(generated.device_function) > 0:
        name += '_' + str(len(generated.device_function))
    generated.device_function[(func, in_types)] = None

    cuda_code, env = _transpile_function(
        tree, name, attributes, mode, consts,
        in_types, ret_type, generated, source=source)

    generated.device_function[(func, in_types)] = (name, env.ret_type)
    generated.add_code(cuda_code)
    return name, env.ret_type


def _indent(lines: List[str], spaces: str = '  ') -> List[str]:
    return [spaces + line for line in lines]


def is_constants(*values: _internal_types.Expr) -> bool:
    assert all(isinstance(x, _internal_types.Expr) for x in values)
    return all(isinstance(x, Constant) for x in values)


class Environment:
    """Environment of the scope

    Attributes:
        mode ('numpy' or 'cuda'): The rule for typecast.
        consts (dict): The dictionary with keys as the variable names and
            the values as the data that is determined at compile-time.
        params (dict): The dictionary of function arguments with keys as
            the variable names and the values as the Data.
        locals (dict): The dictionary with keys as the variable names and the
            values as the Data stored at the local scope of the function.
        ret_type (_cuda_types.TypeBase):
            The type of return value of the function.
            If it is initialized to be ``None``, the return type must be
            inferred until the end of transpilation of the function.
        generated (Generated): Generated CUDA functions.
    """

    def __init__(
            self,
            mode: str,
            consts: Dict[str, Constant],
            params: Dict[str, Data],
            ret_type: _cuda_types.TypeBase,
            generated: Generated,
    ):
        self.mode = mode
        self.consts = consts
        self.params = params
        self.locals: Dict[str, Data] = {}
        self.decls: Dict[str, Data] = {}
        self.ret_type = ret_type
        self.generated = generated
        self.count = 0

    def __getitem__(self, key: str) -> Optional[Union[Constant, Data]]:
        if key in self.locals:
            return self.locals[key]
        if key in self.params:
            return self.params[key]
        if key in self.consts:
            return self.consts[key]
        return None

    def get_fresh_variable_name(
            self, prefix: str = '', suffix: str = '') -> str:
        self.count += 1
        name = f'{prefix}{self.count}{suffix}'
        if self[name] is None:
            return name
        else:
            return self.get_fresh_variable_name(prefix, suffix)


def _transpile_function(
        func, name, attributes, mode, consts,
        in_types, ret_type, generated, *, source):
    """Transpile the function
    Args:
        func (ast.FunctionDef): Target function.
        name (str): Function name.
        attributes (str): The attributes of target function.
        mode ('numpy' or 'cuda'): The rule for typecast.
        consts (dict): The dictionary with keys as variable names and
            values as concrete data object.
        in_types (list of _cuda_types.TypeBase): The types of arguments.
        ret_type (_cuda_types.TypeBase): The type of return value.

    Returns:
        code (str): The generated CUDA code.
        env (Environment): More details of analysis result of the function,
            which includes preambles, estimated return type and more.
    """
    try:
        return _transpile_function_internal(
            func, name, attributes, mode, consts,
            in_types, ret_type, generated)
    except _JitCompileError as e:
        exc = e
        if _is_debug_mode:
            exc.reraise(source)

    # Raises the error out of `except` block to clean stack trace.
    exc.reraise(source)
    assert False


def _transpile_function_internal(
        func, name, attributes, mode, consts, in_types, ret_type, generated):
    consts = dict([(k, Constant(v)) for k, v, in consts.items()])

    if not isinstance(func, ast.FunctionDef):
        # TODO(asi1024): Support for `ast.ClassDef`.
        raise NotImplementedError('Not supported: {}'.format(type(func)))
    if len(func.decorator_list) > 0:
        if sys.version_info >= (3, 9):
            # Code path for Python versions that support `ast.unparse`.
            for deco in func.decorator_list:
                deco_code = ast.unparse(deco)
                if not any(word in deco_code
                           for word in ['rawkernel', 'vectorize']):
                    warnings.warn(
                        f'Decorator {deco_code} may not supported in JIT.',
                        RuntimeWarning)
    arguments = func.args
    if arguments.vararg is not None:
        raise NotImplementedError('`*args` is not supported currently.')
    if len(arguments.kwonlyargs) > 0:  # same length with `kw_defaults`.
        raise NotImplementedError(
            'keyword only arguments are not supported currently .')
    if arguments.kwarg is not None:
        raise NotImplementedError('`**kwargs` is not supported currently.')
    if len(arguments.defaults) > 0:
        raise NotImplementedError(
            'Default values are not supported currently.')

    args = [arg.arg for arg in arguments.args]
    if len(args) != len(in_types):
        raise TypeError(
            f'{name}() takes {len(args)} positional arguments '
            f'but {len(in_types)} were given.')
    params = dict([(x, Data(x, t)) for x, t in zip(args, in_types)])
    env = Environment(mode, consts, params, ret_type, generated)
    body = _transpile_stmts(func.body, True, env)
    params_s = ', '.join([t.declvar(x, None) for x, t in zip(args, in_types)])
    local_vars = [v.ctype.declvar(n, None) + ';' for n, v in env.decls.items()]

    if env.ret_type is None:
        env.ret_type = _cuda_types.void

    head = f'{attributes} {env.ret_type} {name}({params_s})'
    code = CodeBlock(head, local_vars + body)
    return str(code), env


def _eval_operand(
        op: ast.AST,
        args: Sequence[Union[Constant, Data]],
        env: Environment,
) -> Union[Constant, Data]:
    if is_constants(*args):
        pyfunc = _cuda_typerules.get_pyfunc(type(op))
        return Constant(pyfunc(*[x.obj for x in args]))

    if isinstance(op, ast.Add):
        x, y = args
        x = Data.init(x, env)
        y = Data.init(y, env)
        if hasattr(x.ctype, '_add'):
            out = x.ctype._add(env, x, y)
            if out is not NotImplemented:
                return out
        if hasattr(y.ctype, '_radd'):
            out = y.ctype._radd(env, x, y)
            if out is not NotImplemented:
                return out

    if isinstance(op, ast.Sub):
        x, y = args
        x = Data.init(x, env)
        y = Data.init(y, env)
        if hasattr(x.ctype, '_sub'):
            out = x.ctype._sub(env, x, y)
            if out is not NotImplemented:
                return out
        if hasattr(y.ctype, '_rsub'):
            out = y.ctype._rsub(env, x, y)
            if out is not NotImplemented:
                return out

    ufunc = _cuda_typerules.get_ufunc(env.mode, type(op))
    return _call_ufunc(ufunc, args, None, env)


def _call_ufunc(
        ufunc: _kernel.ufunc,
        args: Sequence[Union[Constant, Data]],
        dtype: Optional[numpy.dtype],
        env: Environment,
) -> Data:
    if len(args) != ufunc.nin:
        raise ValueError('invalid number of arguments')

    in_types = []
    for x in args:
        if isinstance(x, Constant):
            t = _cuda_typerules.get_ctype_from_scalar(env.mode, x.obj).dtype
        elif isinstance(x.ctype, _cuda_types.Scalar):
            t = x.ctype.dtype
        else:
            raise TypeError(f'cupy.ufunc: {x.ctype} is unsupported')
        in_types.append(t)

    op = _cuda_typerules.guess_routine(ufunc, tuple(in_types), dtype, env.mode)

    if op is None:
        raise TypeError(
            f'"{ufunc.name}" does not support for the input types: {in_types}')

    if op.error_func is not None:
        op.error_func()

    if ufunc.nout == 1 and op.routine.startswith('out0 = '):
        out_type = _cuda_types.Scalar(op.out_types[0])
        expr = op.routine.replace('out0 = ', '')

        in_params = []
        for x, t in zip(args, op.in_types):
            x = _astype_scalar(x, _cuda_types.Scalar(t), 'same_kind', env)
            x = Data.init(x, env)
            in_params.append(x)

        can_use_inline_expansion = True
        for i in range(ufunc.nin):
            if len(list(re.finditer(r'in{}'.format(i), op.routine))) > 1:
                can_use_inline_expansion = False
            if f'in{i}_type' in op.routine:
                can_use_inline_expansion = False

        env.generated.add_code(ufunc._preamble)

        if can_use_inline_expansion:
            # Code pass for readable generated code
            for i, x in enumerate(in_params):
                expr = expr.replace(f'in{i}', x.code)
            expr = '(' + expr.replace('out0_type', str(out_type)) + ')'
        else:
            template_typenames = ', '.join([
                f'typename in{i}_type' for i in range(ufunc.nin)])
            ufunc_name = f'{ufunc.name}_{str(numpy.dtype(op.out_types[0]))}'
            params = ', '.join([f'in{i}_type in{i}' for i in range(ufunc.nin)])
            ufunc_code = f"""template <{template_typenames}>
__device__ {out_type} {ufunc_name}({params}) {{
    typedef {out_type} out0_type;
    return {expr};
}}
"""
            env.generated.add_code(ufunc_code)
            in_params_code = ', '.join([a.code for a in in_params])
            expr = f'{ufunc_name}({in_params_code})'
        return Data(expr, out_type)

    raise NotImplementedError(f'ufunc `{ufunc.name}` is not supported.')


def _transpile_stmts(
        stmts: List[ast.stmt],
        is_toplevel: bool,
        env: Environment,
) -> _CodeType:
    codeblocks = []
    for stmt in stmts:
        codeblocks.extend(_transpile_stmt(stmt, is_toplevel, env))
    return codeblocks


@transpile_function_wrapper
def _transpile_stmt(
        stmt: ast.stmt,
        is_toplevel: bool,
        env: Environment,
) -> _CodeType:
    """Transpile the statement.

    Returns (list of [CodeBlock or str]): The generated CUDA code.
    """

    if isinstance(stmt, ast.ClassDef):
        raise NotImplementedError('class is not supported currently.')
    if isinstance(stmt, (ast.FunctionDef, ast.AsyncFunctionDef)):
        raise NotImplementedError(
            'Nested functions are not supported currently.')
    if isinstance(stmt, ast.Return):
        value = _transpile_expr(stmt.value, env)
        value = Data.init(value, env)
        t = value.ctype
        if env.ret_type is None:
            env.ret_type = t
        elif env.ret_type != t:
            raise ValueError(
                f'Failed to infer the return type: {env.ret_type} or {t}')
        return [f'return {value.code};']
    if isinstance(stmt, ast.Delete):
        raise NotImplementedError('`del` is not supported currently.')

    if isinstance(stmt, ast.Assign):
        if len(stmt.targets) != 1:
            raise NotImplementedError('Not implemented.')

        value = _transpile_expr(stmt.value, env)
        var = stmt.targets[0]

        if is_constants(value) and isinstance(var, ast.Name):
            name = var.id
            if not isinstance(value.obj, _typeclasses):
                if is_toplevel:
                    if isinstance(env[name], Data):
                        raise TypeError(f'Type mismatch of variable: `{name}`')
                    env.consts[name] = value
                    return []
                else:
                    raise TypeError(
                        'Cannot assign constant value not at top-level.')

        value = Data.init(value, env)
        return _transpile_assign_stmt(var, env, value, is_toplevel)

    if isinstance(stmt, ast.AugAssign):
        value = _transpile_expr(stmt.value, env)
        target = _transpile_expr(stmt.target, env)
        if not isinstance(target, Data):
            raise TypeError(f'Cannot augassign to {target.code}')
        value = Data.init(value, env)
        tmp = Data(env.get_fresh_variable_name('_tmp_'), target.ctype)
        result = _eval_operand(stmt.op, (tmp, value), env)
        assert isinstance(target, Data)
        assert isinstance(result, Data)
        assert isinstance(target.ctype, _cuda_types.Scalar)
        assert isinstance(result.ctype, _cuda_types.Scalar)
        _raise_if_invalid_cast(
            result.ctype.dtype, target.ctype.dtype, 'same_kind')

        return ['{ ' + target.ctype.declvar('&' + tmp.code, target) + '; ' +
                target.ctype.assign(tmp, result) + '; }']

    if isinstance(stmt, ast.For):
        if len(stmt.orelse) > 0:
            raise NotImplementedError('while-else is not supported.')
        assert isinstance(stmt.target, ast.Name)
        name = stmt.target.id
        iters = _transpile_expr(stmt.iter, env)
        loop_var = env[name]

        if loop_var is None:
            target = Data(stmt.target.id, iters.ctype)
            env.locals[name] = target
            env.decls[name] = target
        elif isinstance(loop_var, Constant):
            raise TypeError('loop counter must not be constant value')
        elif loop_var.ctype != iters.ctype:
            raise TypeError(
                f'Data type mismatch of variable: `{name}`: '
                f'{loop_var.ctype} != {iters.ctype}')

        if not isinstance(iters, _internal_types.Range):
            raise NotImplementedError(
                'for-loop is supported only for range iterator.')

        body = _transpile_stmts(stmt.body, False, env)

        init_code = (f'{iters.ctype} '
                     f'__it = {iters.start.code}, '
                     f'__stop = {iters.stop.code}, '
                     f'__step = {iters.step.code}')
        cond = '__step >= 0 ? __it < __stop : __it > __stop'
        if iters.step_is_positive is True:
            cond = '__it < __stop'
        elif iters.step_is_positive is False:
            cond = '__it > __stop'

        head = f'for ({init_code}; {cond}; __it += __step)'
        code: _CodeType = [CodeBlock(head, [f'{name} = __it;'] + body)]

        unroll = iters.unroll
        if unroll is True:
            code = ['#pragma unroll'] + code
        elif unroll is not None:
            code = [f'#pragma unroll({unroll})'] + code
        return code

    if isinstance(stmt, ast.AsyncFor):
        raise ValueError('`async for` is not allowed.')
    if isinstance(stmt, ast.While):
        if len(stmt.orelse) > 0:
            raise NotImplementedError('while-else is not supported.')
        condition = _transpile_expr(stmt.test, env)
        condition = _astype_scalar(condition, _cuda_types.bool_, 'unsafe', env)
        condition = Data.init(condition, env)
        body = _transpile_stmts(stmt.body, False, env)
        head = f'while ({condition.code})'
        return [CodeBlock(head, body)]
    if isinstance(stmt, ast.If):
        condition = _transpile_expr(stmt.test, env)
        if is_constants(condition):
            stmts = stmt.body if condition.obj else stmt.orelse
            return _transpile_stmts(stmts, is_toplevel, env)
        head = f'if ({condition.code})'
        then_body = _transpile_stmts(stmt.body, False, env)
        else_body = _transpile_stmts(stmt.orelse, False, env)
        return [CodeBlock(head, then_body), CodeBlock('else', else_body)]
    if isinstance(stmt, (ast.With, ast.AsyncWith)):
        raise ValueError('Switching contexts are not allowed.')
    if isinstance(stmt, (ast.Raise, ast.Try)):
        raise ValueError('throw/catch are not allowed.')
    if isinstance(stmt, ast.Assert):
        value = _transpile_expr(stmt.test, env)
        if is_constants(value):
            assert value.obj
            return [';']
        else:
            return ['assert(' + value + ');']
    if isinstance(stmt, (ast.Import, ast.ImportFrom)):
        raise ValueError('Cannot import modules from the target functions.')
    if isinstance(stmt, (ast.Global, ast.Nonlocal)):
        raise ValueError('Cannot use global/nonlocal in the target functions.')
    if isinstance(stmt, ast.Expr):
        value = _transpile_expr(stmt.value, env)
        return [';'] if is_constants(value) else [value.code + ';']
    if isinstance(stmt, ast.Pass):
        return [';']
    if isinstance(stmt, ast.Break):
        raise NotImplementedError('Not implemented.')
    if isinstance(stmt, ast.Continue):
        raise NotImplementedError('Not implemented.')
    assert False


@transpile_function_wrapper
def _transpile_expr(expr: ast.expr, env: Environment) -> _internal_types.Expr:
    """Transpile the statement.

    Returns (Data): The CUDA code and its type of the expression.
    """
    res = _transpile_expr_internal(expr, env)

    if isinstance(res, Constant) and isinstance(res.obj, _internal_types.Expr):
        return res.obj
    else:
        return res


def _transpile_expr_internal(
        expr: ast.expr,
        env: Environment,
) -> _internal_types.Expr:
    if isinstance(expr, ast.BoolOp):
        values = [_transpile_expr(e, env) for e in expr.values]
        value = values[0]
        for rhs in values[1:]:
            value = _eval_operand(expr.op, (value, rhs), env)
        return value
    if isinstance(expr, ast.BinOp):
        left = _transpile_expr(expr.left, env)
        right = _transpile_expr(expr.right, env)
        return _eval_operand(expr.op, (left, right), env)
    if isinstance(expr, ast.UnaryOp):
        value = _transpile_expr(expr.operand, env)
        return _eval_operand(expr.op, (value,), env)
    if isinstance(expr, ast.Lambda):
        raise NotImplementedError('Not implemented.')
    if isinstance(expr, ast.Compare):
        values = [expr.left] + expr.comparators
        if len(values) != 2:
            raise NotImplementedError(
                'Comparison of 3 or more values is not implemented.')
        values = [_transpile_expr(e, env) for e in values]
        return _eval_operand(expr.ops[0], values, env)
    if isinstance(expr, ast.IfExp):
        cond = _transpile_expr(expr.test, env)
        x = _transpile_expr(expr.body, env)
        y = _transpile_expr(expr.orelse, env)

        if isinstance(expr, Constant):
            return x if expr.obj else y
        if cond.ctype.dtype.kind == 'c':
            raise TypeError("Complex type value cannot be boolean condition.")
        x, y = _infer_type(x, y, env), _infer_type(y, x, env)
        if x.ctype.dtype != y.ctype.dtype:
            raise TypeError(
                'Type mismatch in conditional expression.: '
                f'{x.ctype.dtype} != {y.ctype.dtype}')
        cond = _astype_scalar(cond, _cuda_types.bool_, 'unsafe', env)
        return Data(f'({cond.code} ? {x.code} : {y.code})', x.ctype)

    if isinstance(expr, ast.Call):
        func = _transpile_expr(expr.func, env)
        args = [_transpile_expr(x, env) for x in expr.args]
        kwargs: Dict[str, Union[Constant, Data]] = {}
        for kw in expr.keywords:
            assert kw.arg is not None
            kwargs[kw.arg] = _transpile_expr(kw.value, env)

        builtin_funcs = _builtin_funcs.builtin_functions_dict
        if is_constants(func) and (func.obj in builtin_funcs):
            func = builtin_funcs[func.obj]

        if isinstance(func, _internal_types.BuiltinFunc):
            return func.call(env, *args, **kwargs)

        if not isinstance(func, Constant):
            raise TypeError(f"'{func}' is not callable.")

        func = func.obj

        if isinstance(func, _interface._cuda_types.TypeBase):
            return func._instantiate(env, *args, **kwargs)

        if isinstance(func, _interface._JitRawKernel):
            if not func._device:
                raise TypeError(
                    f'Calling __global__ function {func._func.__name__} '
                    'from __global__ function is not allowed.')
            args = [Data.init(x, env) for x in args]
            in_types = tuple([x.ctype for x in args])
            fname, return_type = _transpile_func_obj(
                func._func, ['__device__'], env.mode,
                in_types, None, env.generated)
            in_params = ', '.join([x.code for x in args])
            return Data(f'{fname}({in_params})', return_type)

        if isinstance(func, _kernel.ufunc):
            # ufunc call
            dtype = kwargs.pop('dtype', Constant(None)).obj
            if len(kwargs) > 0:
                name = next(iter(kwargs))
                raise TypeError(
                    f"'{name}' is an invalid keyword to ufunc {func.name}")
            return _call_ufunc(func, args, dtype, env)

        if is_constants(*args, *kwargs.values()):
            # compile-time function call
            args = [x.obj for x in args]
            kwargs = dict([(k, v.obj) for k, v in kwargs.items()])
            return Constant(func(*args, **kwargs))

        if inspect.isclass(func) and issubclass(func, _typeclasses):
            # explicit typecast
            if len(args) != 1:
                raise TypeError(
                    f'function takes {func} invalid number of argument')
            ctype = _cuda_types.Scalar(func)
            return _astype_scalar(args[0], ctype, 'unsafe', env)

        raise TypeError(f"Invalid function call '{func.__name__}'.")

    if isinstance(expr, ast.Constant):
        return Constant(expr.value)
    if isinstance(expr, ast.Subscript):
        array = _transpile_expr(expr.value, env)
        index = _transpile_expr(expr.slice, env)
        return _indexing(array, index, env)
    if isinstance(expr, ast.Name):
        value = env[expr.id]
        if value is None:
            raise NameError(f'Unbound name: {expr.id}')
        return value
    if isinstance(expr, ast.Attribute):
        value = _transpile_expr(expr.value, env)
        if isinstance(value, Constant):
            return Constant(getattr(value.obj, expr.attr))
        if isinstance(value, Data) and hasattr(value.ctype, expr.attr):
            attr = getattr(value.ctype, expr.attr, None)
            if isinstance(attr, types.MethodType):
                return attr(value)
            if isinstance(attr, Data):
                return attr
        raise AttributeError(f'Unknown attribute: {expr.attr}')

    if isinstance(expr, ast.Tuple):
        elts = [_transpile_expr(x, env) for x in expr.elts]

        if all([isinstance(x, Constant) for x in elts]):
            return Constant(tuple([x.obj for x in elts]))

        elts = [Data.init(x, env) for x in elts]
        elts_code = ', '.join([x.code for x in elts])
        # STD is defined in carray.cuh
        if len(elts) == 2:
            return Data(
                f'STD::make_pair({elts_code})',
                _cuda_types.Tuple([x.ctype for x in elts]))
        else:
            return Data(
                f'STD::make_tuple({elts_code})',
                _cuda_types.Tuple([x.ctype for x in elts]))

    if isinstance(expr, ast.Index):
        # Deprecated in Python 3.9
        return _transpile_expr(expr.value, env)  # type: ignore

    raise ValueError('Not supported: type {}'.format(type(expr)))


def _emit_assign_stmt(
        lvalue: Union[Constant, Data],
        rvalue: Data,
        env: Environment,
) -> _CodeType:
    if isinstance(lvalue, Constant):
        raise TypeError('lvalue of assignment must not be constant value')

    if (isinstance(lvalue.ctype, _cuda_types.Scalar)
            and isinstance(rvalue.ctype, _cuda_types.Scalar)):
        rvalue = _astype_scalar(rvalue, lvalue.ctype, 'same_kind', env)
    elif lvalue.ctype != rvalue.ctype:
        raise TypeError(
            f'Data type mismatch of variable: `{lvalue.code}`: '
            f'{lvalue.ctype} != {rvalue.ctype}')

    return [lvalue.ctype.assign(lvalue, rvalue) + ';']


def _transpile_assign_stmt(
        target: ast.expr,
        env: Environment,
        value: Data,
        is_toplevel: bool,
        depth: int = 0,
) -> _CodeType:
    if isinstance(target, ast.Name):
        name = target.id
        lvalue = env[name]
        if lvalue is None:
            lvalue = Data(name, value.ctype)
            env.locals[name] = lvalue
            if is_toplevel and depth == 0:
                return [value.ctype.declvar(name, value) + ';']
            env.decls[name] = lvalue
        return _emit_assign_stmt(lvalue, value, env)

    if isinstance(target, ast.Subscript):
        lvalue = Data.init(_transpile_expr(target, env), env)
        return _emit_assign_stmt(lvalue, value, env)

    if isinstance(target, ast.Tuple):
        if not isinstance(value.ctype, _cuda_types.Tuple):
            raise ValueError(f'{value.ctype} cannot be unpack')
        size = len(target.elts)
        if len(value.ctype.types) > size:
            raise ValueError(f'too many values to unpack (expected {size})')
        if len(value.ctype.types) < size:
            raise ValueError(f'not enough values to unpack (expected {size})')
        codes = [value.ctype.declvar(f'_temp{depth}', value) + ';']
        for i in range(size):
            # STD is defined in carray.cuh
            code = f'STD::get<{i}>(_temp{depth})'
            ctype = value.ctype.types[i]
            stmt = _transpile_assign_stmt(
                target.elts[i], env, Data(code, ctype), is_toplevel, depth + 1)
            codes.extend(stmt)
        return [CodeBlock('', codes)]

    assert False


def _indexing(
        array: _internal_types.Expr,
        index: _internal_types.Expr,
        env: Environment,
) -> Union[Data, Constant]:
    if isinstance(array, Constant):
        if isinstance(index, Constant):
            return Constant(array.obj[index.obj])
        raise TypeError(
            f'{type(array.obj)} is not subscriptable with non-constants.')

    array = Data.init(array, env)

    if isinstance(array.ctype, _cuda_types.Tuple):
        if isinstance(index, Constant):
            i = index.obj
            t = array.ctype.types[i]
            # STD is defined in carray.cuh
            return Data(f'STD::get<{i}>({array.code})', t)
        raise TypeError('Tuple is not subscriptable with non-constants.')

    if isinstance(array.ctype, _cuda_types.ArrayBase):
        index = Data.init(index, env)
        ndim = array.ctype._ndim
        if isinstance(index.ctype, _cuda_types.Scalar):
            index_dtype = index.ctype.dtype
            if ndim == 0:
                raise TypeError(
                    'Scalar indexing is not supported for 0-dim array.')
            if ndim > 1:
                new_carray = _cuda_types.CArray(
                    array.ctype.dtype,
                    array.ctype._ndim - 1,
                    array.ctype._c_contiguous,
                    array.ctype._index_32_bits)
                return Data(
                    f'{array.code}._slicing({index.code})', new_carray)
            if index_dtype.kind not in 'ui':
                raise TypeError('Array indices must be integers.')
            return Data(
                f'{array.code}[{index.code}]', array.ctype.child_type)
        if isinstance(index.ctype, _cuda_types.Tuple):
            if ndim < len(index.ctype.types):
                raise IndexError(
                    f'The number of indices is beyond array dim: {ndim}')
            for t in index.ctype.types:
                if not isinstance(t, _cuda_types.Scalar):
                    raise TypeError('Array indices must be scalar.')
                if t.dtype.kind not in 'iu':
                    raise TypeError('Array indices must be integer.')
            if ndim > len(index.ctype.types):
                new_carray = _cuda_types.CArray(
                    array.ctype.dtype,
                    array.ctype._ndim - len(index.ctype.types),
                    array.ctype._c_contiguous,
                    array.ctype._index_32_bits)
                params = f'{index.code}, Dim<{len(index.ctype.types)}>()'
                return Data(f'{array.code}._slicing({params})', new_carray)
            if ndim == 0:
                return Data(
                    f'{array.code}[0]', array.ctype.child_type)
            if ndim == 1:
                # STD is defined in carray.cuh
                return Data(
                    f'{array.code}[STD::get<0>({index.code})]',
                    array.ctype.child_type)
            return Data(
                f'{array.code}._indexing({index.code})',
                array.ctype.child_type)
        if isinstance(index.ctype, _cuda_types.CArray):
            raise TypeError('Advanced indexing is not supported.')
        assert False  # Never reach.

    raise TypeError(f'{array.code} is not subscriptable.')


_T = TypeVar('_T', Constant, Data, Union[Constant, Data])


def _astype_scalar(
        x: _T,
        ctype: _cuda_types.Scalar,
        casting: _CastingType,
        env: Environment,
) -> _T:
    if isinstance(x, Constant):
        assert not isinstance(x, Data)
        return Constant(ctype.dtype.type(x.obj))
    # # TODO
    # if not isinstance(x, Data):
    #     raise TypeError(f'{x} is not scalar type.')
    if not isinstance(x.ctype, _cuda_types.Scalar):
        raise TypeError(f'{x.code} is not scalar type.')
    from_t = x.ctype.dtype
    to_t = ctype.dtype
    if from_t == to_t:
        return x
    # Uses casting rules for scalar values.
    # TODO(seberg): NEP 50/NumPy 2.0 hopefully makes `from_t.type(0)`
    #               unnecessary allowing to use `_raise_if_invalid_cast()`.
    if not numpy.can_cast(from_t.type(0), to_t, casting):
        raise TypeError(
            f"Cannot cast from '{from_t}' to {to_t} "
            f"with casting rule {casting}.")
    if from_t.kind == 'c' and to_t.kind != 'c':
        if to_t.kind != 'b':
            warnings.warn(
                'Casting complex values to real discards the imaginary part',
                ComplexWarning)
        return Data(f'({ctype})({x.code}.real())', ctype)
    return Data(f'({ctype})({x.code})', ctype)


def _infer_type(
        x: Union[Constant, Data],
        hint: Union[Constant, Data],
        env: Environment,
) -> Data:
    if not isinstance(x, Constant) or isinstance(x.obj, numpy.generic):
        return Data.init(x, env)
    hint = Data.init(hint, env)
    assert isinstance(hint.ctype, _cuda_types.Scalar)
    cast_x = _astype_scalar(x, hint.ctype, 'same_kind', env)
    return Data.init(cast_x, env)