vllm/lib/python3.10/site-packages/cupy/cuda/compiler.py

import copy
import hashlib
import math
import os
import platform
import re
import shutil
import subprocess
import sys
import tempfile
from typing import Optional
import warnings

from cupy.cuda import device
from cupy.cuda import function
from cupy.cuda import get_rocm_path
from cupy_backends.cuda.api import driver
from cupy_backends.cuda.api import runtime
from cupy_backends.cuda.libs import nvrtc
from cupy import _environment
from cupy import _util

_cuda_hip_version = driver.get_build_version()


_nvrtc_version = None
_win32 = sys.platform.startswith('win32')
_rdc_flags = ('--device-c', '-dc', '-rdc=true',
              '--relocatable-device-code=true')
_cudadevrt = None


class NVCCException(Exception):
    pass


class HIPCCException(Exception):
    pass


class JitifyException(Exception):
    pass


def _run_cc(cmd, cwd, backend, log_stream=None):
    # backend in ('nvcc', 'hipcc')
    try:
        # Inherit the environment variable as NVCC refers to PATH, TMPDIR/TMP,
        # NVCC_PREPEND_FLAGS, NVCC_APPEND_FLAGS.
        env = os.environ
        if _win32:
            # Adds the extra PATH for NVCC invocation.
            # When running NVCC, a host compiler must be available in PATH,
            # but this is not true in general Windows environment unless
            # running inside the SDK Tools command prompt.
            # To mitigate the situation CuPy automatically adds a path to
            # the VC++ compiler (cl.exe) found via setuptools, if it is not
            # on the PATH.
            extra_path = _get_extra_path_for_msvc()
            if extra_path is not None:
                path = extra_path + os.pathsep + os.environ.get('PATH', '')
                env = copy.deepcopy(env)
                env['PATH'] = path
        log = subprocess.check_output(
            cmd, cwd=cwd, env=env,
            stderr=subprocess.STDOUT,
            universal_newlines=True,
            creationflags=(subprocess.CREATE_NO_WINDOW if _win32 else 0))
        if log_stream is not None:
            log_stream.write(log)
        return log
    except subprocess.CalledProcessError as e:
        msg = ('`{0}` command returns non-zero exit status. \n'
               'command: {1}\n'
               'return-code: {2}\n'
               'stdout/stderr: \n'
               '{3}'.format(backend,
                            e.cmd,
                            e.returncode,
                            e.output))
        if backend == 'nvcc':
            raise NVCCException(msg)
        elif backend == 'hipcc':
            raise HIPCCException(msg)
        else:
            raise RuntimeError(msg)
    except OSError as e:
        msg = 'Failed to run `{0}` command. ' \
              'Check PATH environment variable: ' \
              + str(e)
        raise OSError(msg.format(backend))


@_util.memoize()
def _get_extra_path_for_msvc():
    cl_exe = shutil.which('cl.exe')
    if cl_exe:
        # The compiler is already on PATH, no extra path needed.
        return None

    cl_exe_dir = _get_cl_exe_dir()
    if cl_exe_dir:
        return cl_exe_dir

    cl_exe_dir = _get_cl_exe_dir_fallback()
    if cl_exe_dir:
        return cl_exe_dir

    return None


def _get_cl_exe_dir() -> Optional[str]:
    try:
        try:
            # setuptools.msvc is missing in setuptools v74.0.0.
            # setuptools.msvc requires explicit import in setuptools v74.1.0+.
            import setuptools.msvc
        except Exception:
            return None
        vctools = setuptools.msvc.EnvironmentInfo(platform.machine()).VCTools
        for path in vctools:
            cl_exe = os.path.join(path, 'cl.exe')
            if os.path.exists(cl_exe):
                return path
        warnings.warn(f'cl.exe could not be found in {vctools}')
    except Exception as e:
        warnings.warn(
            f'Failed to find cl.exe with setuptools.msvc: {type(e)}: {e}')
    return None


def _get_cl_exe_dir_fallback() -> Optional[str]:
    # Discover cl.exe without relying on undocumented setuptools.msvc API.
    # As of now this code path exists only for setuptools 74.0.0 (see #8583).
    # N.B. This takes few seconds as this incurs cmd.exe (vcvarsall.bat)
    # invocation.
    try:
        from setuptools import Distribution
        from setuptools.command.build_ext import build_ext
        ext = build_ext(Distribution({'name': 'cupy_cl_exe_discover'}))
        ext.setup_shlib_compiler()
        ext.shlib_compiler.initialize()  # MSVCCompiler only
        return os.path.dirname(ext.shlib_compiler.cc)
    except Exception as e:
        warnings.warn(
            f'Failed to find cl.exe with setuptools: {type(e)}: {e}')
    return None


def _get_nvrtc_version():
    global _nvrtc_version
    if _nvrtc_version is None:
        _nvrtc_version = nvrtc.getVersion()

    return _nvrtc_version


@_util.memoize()
def _get_cupy_cache_key():
    from cupy._core import core
    return core.CUPY_CACHE_KEY


# Known archs for Tegra/Jetson/Xavier/etc
_tegra_archs = ('32', '53', '62', '72', '87')


@_util.memoize()
def _get_max_compute_capability():
    major, minor = _get_nvrtc_version()
    if major < 11:
        # CUDA 10.2
        nvrtc_max_compute_capability = '75'
    elif major == 11 and minor == 0:
        # CUDA 11.0
        nvrtc_max_compute_capability = '80'
    elif major == 11 and minor < 8:
        # CUDA 11.1 - 11.7
        # Note: 87 is for Jetson Orin
        nvrtc_max_compute_capability = '86'
    elif (major == 11 and minor == 8) or (major == 12 and minor < 8):
        # CUDA 11.8, 12.0 - 12.7
        nvrtc_max_compute_capability = '90'
    else:
        # CUDA 12.8+
        nvrtc_max_compute_capability = '120'

    return nvrtc_max_compute_capability


@_util.memoize()
def _get_extra_include_dir_opts():
    major, minor = _get_nvrtc_version()
    return tuple(
        f'-I{d}'
        for d in _environment._get_include_dir_from_conda_or_wheel(
            major, minor
        )
    )


@_util.memoize(for_each_device=True)
def _get_arch():
    # See Supported Compile Options section of NVRTC User Guide for
    # the maximum value allowed for `--gpu-architecture`.
    nvrtc_max_compute_capability = _get_max_compute_capability()

    arch = device.Device().compute_capability
    if arch in _tegra_archs:
        return arch
    else:
        return min(arch, nvrtc_max_compute_capability, key=int)


@_util.memoize(for_each_device=True)
def _get_arch_for_options_for_nvrtc(arch=None):
    # NVRTC in CUDA 11.3+ generates PTX that cannot be run an earlier driver
    # version than the one included in the used CUDA version, as
    # documented in:
    # https://docs.nvidia.com/cuda/archive/11.3.0/nvrtc/index.html#versioning
    # Here we use `-arch=sm_*` instead of `-arch=compute_*` to directly
    # generate cubin (SASS) instead of PTX. See #5097 for details.
    if arch is None:
        arch = _get_arch()
    if (
        not _use_ptx
        and int(arch) <= int(_get_max_compute_capability())
    ):
        return f'-arch=sm_{arch}', 'cubin'
    return f'-arch=compute_{arch}', 'ptx'


def _is_cudadevrt_needed(options):
    return any(o for o in options if o in _rdc_flags)


def _get_cudadevrt_path():
    global _cudadevrt
    if _cudadevrt is not None:
        return _cudadevrt

    # defer import to here to avoid circular dependency
    from cupy.cuda import get_cuda_path
    global _win32

    cudadevrt = get_cuda_path()
    if cudadevrt is None:
        raise RuntimeError('CUDA is not found.')

    if _win32:
        # rely on os.altsep
        cudadevrt += '/lib/x64/cudadevrt.lib'
    else:  # linux & osx: search twice as in cupy/install/build.py
        cudadevrt64 = cudadevrt + '/lib64/libcudadevrt.a'
        if not os.path.isfile(cudadevrt64):
            cudadevrt += '/lib/libcudadevrt.a'
        else:
            cudadevrt = cudadevrt64
    if not os.path.isfile(cudadevrt):
        raise RuntimeError(
            'Relocatable PTX code is requested, but cudadevrt '
            'is not found.')
    return cudadevrt


def _remove_rdc_option(options):
    return tuple(o for o in options if o not in _rdc_flags)


def _get_bool_env_variable(name, default):
    val = os.environ.get(name)
    if val is None or len(val) == 0:
        return default
    try:
        return int(val) == 1
    except ValueError:
        return False


_use_ptx = _get_bool_env_variable('CUPY_COMPILE_WITH_PTX', False)
_jitify_header_source_map_populated = False


def _jitify_prep(source, options, cu_path):
    from cupy.cuda import jitify

    # TODO(leofang): refactor this?
    global _jitify_header_source_map_populated
    if not _jitify_header_source_map_populated:
        from cupy._core import core
        jitify._init_module()
        jitify._add_sources(core._get_header_source_map())
        _jitify_header_source_map_populated = True

    # jitify requires the 1st line to be the program name
    old_source = source
    source = cu_path + '\n' + source

    # Upon failure, in addition to throw an error Jitify also prints the log
    # to stdout. In principle we could intercept that by hijacking stdout's
    # file descriptor (tested locally), but the problem is pytest also does
    # the same thing internally, causing strange errors when running the tests.
    # As a result, we currently maintain Jitify's default behavior for easy
    # debugging, and wait for the upstream to address this issue
    # (NVIDIA/jitify#79).

    try:
        name, options, headers, include_names = jitify.jitify(source, options)
    except Exception as e:  # C++ could throw all kinds of errors
        cex = CompileException(str(e), old_source, cu_path, options, 'jitify')
        dump = _get_bool_env_variable(
            'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
        if dump:
            cex.dump(sys.stderr)
        raise JitifyException(str(cex)) from e
    assert name == cu_path

    return options, headers, include_names


_has_usedforsecurity = (sys.version_info >= (3, 9))


def _hash_hexdigest(value):
    if _has_usedforsecurity:
        hashobj = hashlib.sha1(value, usedforsecurity=False)
    else:
        hashobj = hashlib.sha1(value)
    return hashobj.hexdigest()


_hash_length = len(_hash_hexdigest(b''))  # 40 for SHA1


def compile_using_nvrtc(source, options=(), arch=None, filename='kern.cu',
                        name_expressions=None, log_stream=None,
                        cache_in_memory=False, jitify=False):
    def _compile(
            source, options, cu_path, name_expressions, log_stream, jitify):

        if not runtime.is_hip:
            arch_opt, method = _get_arch_for_options_for_nvrtc(arch)
            options += (arch_opt,)
        else:
            method = 'ptx'

        if jitify:
            options, headers, include_names = _jitify_prep(
                source, options, cu_path)
        else:
            headers = include_names = ()
            major_version, minor_version = _get_nvrtc_version()
            if major_version >= 12:
                # Starting with CUDA 12.0, even without using jitify, some
                # tests cause an error if the following option is not included.
                options += ('--device-as-default-execution-space',)

        prog = _NVRTCProgram(source, cu_path, headers, include_names,
                             name_expressions=name_expressions, method=method)
        try:
            compiled_obj, mapping = prog.compile(options, log_stream)
        except CompileException as e:
            dump = _get_bool_env_variable(
                'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
            if dump:
                e.dump(sys.stderr)
            raise
        return compiled_obj, mapping

    if not cache_in_memory:
        with tempfile.TemporaryDirectory() as root_dir:
            cu_path = os.path.join(root_dir, filename)

            with open(cu_path, 'w') as cu_file:
                cu_file.write(source)

            return _compile(source, options, cu_path,
                            name_expressions, log_stream, jitify)
    else:
        cu_path = '' if not jitify else filename
        return _compile(source, options, cu_path, name_expressions,
                        log_stream, jitify)


def compile_using_nvcc(source, options=(), arch=None,
                       filename='kern.cu', code_type='cubin',
                       separate_compilation=False, log_stream=None):
    # defer import to here to avoid circular dependency
    from cupy.cuda import get_nvcc_path

    if not arch:
        arch = _get_arch()

    if code_type not in ('cubin', 'ptx'):
        raise ValueError('Invalid code_type %s. Should be cubin or ptx')
    if code_type == 'ptx':
        assert not separate_compilation

    arch_str = '-gencode=arch=compute_{cc},code=sm_{cc}'.format(cc=arch)
    _nvcc = get_nvcc_path()
    # split() is needed because _nvcc could come from the env var NVCC
    cmd = _nvcc.split()
    cmd.append(arch_str)

    with tempfile.TemporaryDirectory() as root_dir:
        first_part = filename.split('.')[0]

        path = os.path.join(root_dir, first_part)
        cu_path = '%s.cu' % path
        result_path = '%s.%s' % (path, code_type)

        with open(cu_path, 'w') as cu_file:
            cu_file.write(source)

        if not separate_compilation:  # majority cases
            cmd.append('--%s' % code_type)
            cmd += list(options)
            cmd.append(cu_path)

            try:
                _run_cc(cmd, root_dir, 'nvcc', log_stream)
            except NVCCException as e:
                cex = CompileException(str(e), source, cu_path, options,
                                       'nvcc')

                dump = _get_bool_env_variable(
                    'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
                if dump:
                    cex.dump(sys.stderr)

                raise cex
        else:  # two steps: compile to object and device-link
            cmd_partial = cmd.copy()
            cmd_partial.append('--cubin')

            obj = path + '.o'
            cmd += list(options + ('-o', obj))
            cmd.append(cu_path)

            try:
                _run_cc(cmd, root_dir, 'nvcc', log_stream)
            except NVCCException as e:
                cex = CompileException(str(e), source, cu_path, options,
                                       'nvcc')

                dump = _get_bool_env_variable(
                    'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
                if dump:
                    cex.dump(sys.stderr)

                raise cex

            options = _remove_rdc_option(options)
            options += ('--device-link', obj, '-o', path + '.cubin')
            cmd = cmd_partial + list(options)

            try:
                _run_cc(cmd, root_dir, 'nvcc', log_stream)
            except NVCCException as e:
                cex = CompileException(str(e), '', '', options, 'nvcc')
                raise cex

        if code_type == 'ptx':
            with open(result_path, 'rb') as ptx_file:
                return ptx_file.read()
        elif code_type == 'cubin':
            with open(result_path, 'rb') as bin_file:
                return bin_file.read()
        else:
            assert False, code_type


def _preprocess(source, options, arch, backend):
    if backend == 'nvrtc':
        # For the preprocess it is enough to use PTX method
        # we don't need to explicitly obtain a CUBIN file.
        options += ('-arch=compute_{}'.format(arch),)
        prog = _NVRTCProgram(source)
        try:
            result, _ = prog.compile(options)
        except CompileException as e:
            dump = _get_bool_env_variable(
                'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
            if dump:
                e.dump(sys.stderr)
            raise
    elif backend == 'nvcc':
        try:
            result = compile_using_nvcc(source, options, arch, 'preprocess.cu',
                                        code_type='ptx')
        except CompileException as e:
            dump = _get_bool_env_variable(
                'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
            if dump:
                e.dump(sys.stderr)
            raise
    else:
        raise ValueError('Invalid backend %s' % backend)

    assert isinstance(result, bytes)

    # Extract the part containing version information.
    return '\n'.join(
        x for x in result.decode().splitlines() if x.startswith('//'))


_default_cache_dir = os.path.expanduser('~/.cupy/kernel_cache')


def get_cache_dir():
    return os.environ.get('CUPY_CACHE_DIR', _default_cache_dir)


_empty_file_preprocess_cache: dict = {}


def _compile_module_with_cache(
        source, options=(), arch=None, cache_dir=None, extra_source=None,
        backend='nvrtc', *, enable_cooperative_groups=False,
        name_expressions=None, log_stream=None, jitify=False):

    if enable_cooperative_groups:
        if runtime.is_hip:
            raise ValueError(
                'Cooperative groups is not supported in HIP.')

    if name_expressions is not None and backend != 'nvrtc':
        raise NotImplementedError

    # We silently ignore CUPY_CACHE_IN_MEMORY if nvcc/hipcc are in use, because
    # they must dump files to disk.
    cache_in_memory = (
        _get_bool_env_variable('CUPY_CACHE_IN_MEMORY', False)
        and backend == 'nvrtc')

    if runtime.is_hip:
        backend = 'hiprtc' if backend == 'nvrtc' else 'hipcc'
        return _compile_with_cache_hip(
            source, options, arch, cache_dir, extra_source, backend,
            name_expressions, log_stream, cache_in_memory)
    else:
        return _compile_with_cache_cuda(
            source, options, arch, cache_dir, extra_source, backend,
            enable_cooperative_groups, name_expressions, log_stream,
            cache_in_memory, jitify)


def _compile_with_cache_cuda(
        source, options, arch, cache_dir, extra_source=None, backend='nvrtc',
        enable_cooperative_groups=False, name_expressions=None,
        log_stream=None, cache_in_memory=False, jitify=False):
    # NVRTC does not use extra_source. extra_source is used for cache key.
    global _empty_file_preprocess_cache
    if cache_dir is None:
        cache_dir = get_cache_dir()
    if arch is None:
        arch = _get_arch()

    options += ('-ftz=true',)

    if enable_cooperative_groups:
        # `cooperative_groups` requires relocatable device code.
        options += ('--device-c',)

    if _get_bool_env_variable('CUPY_CUDA_COMPILE_WITH_DEBUG', False):
        options += ('--device-debug', '--generate-line-info')

    is_jitify_requested = ('-DCUPY_USE_JITIFY' in options)
    if jitify and not is_jitify_requested:
        # jitify is set in RawKernel/RawModule, translate it to an option
        # that is useless to the compiler, but can be used as part of the
        # hash key
        options += ('-DCUPY_USE_JITIFY',)
    elif is_jitify_requested and not jitify:
        # jitify is requested internally, just set the flag
        jitify = True
    if jitify and backend != 'nvrtc':
        raise ValueError('jitify only works with NVRTC')

    options += _get_extra_include_dir_opts()
    env = ((arch, options, _get_nvrtc_version(), backend)
           + _get_arch_for_options_for_nvrtc(arch))
    base = _empty_file_preprocess_cache.get(env, None)
    if base is None:
        # This is for checking NVRTC/NVCC compiler internal version
        base = _preprocess('', options, arch, backend)
        _empty_file_preprocess_cache[env] = base

    key_src = '%s %s %s %s %s' % (
        env, base, source, extra_source, _get_cupy_cache_key())
    key_src = key_src.encode('utf-8')
    name = _hash_hexdigest(key_src) + '.cubin'

    mod = function.Module()

    if not cache_in_memory:
        # Read from disk cache
        if not os.path.isdir(cache_dir):
            os.makedirs(cache_dir, exist_ok=True)

        # To handle conflicts in concurrent situation, we adopt lock-free
        # method to avoid performance degradation.
        # We force recompiling to retrieve C++ mangled names if so desired.
        path = os.path.join(cache_dir, name)
        if os.path.exists(path) and not name_expressions:
            with open(path, 'rb') as file:
                data = file.read()
            if len(data) >= _hash_length:
                hash = data[:_hash_length]
                cubin = data[_hash_length:]
                cubin_hash = _hash_hexdigest(cubin).encode('ascii')
                if hash == cubin_hash:
                    mod.load(cubin)
                    return mod
    else:
        # Enforce compiling -- the resulting kernel will be cached elsewhere,
        # so we do nothing
        pass

    if backend == 'nvrtc':
        cu_name = '' if cache_in_memory else name + '.cu'
        ptx, mapping = compile_using_nvrtc(
            source, options, arch, cu_name, name_expressions,
            log_stream, cache_in_memory, jitify)
        if _is_cudadevrt_needed(options):
            # for separate compilation
            ls = function.LinkState()
            ls.add_ptr_data(ptx, 'cupy.ptx')
            _cudadevrt = _get_cudadevrt_path()
            ls.add_ptr_file(_cudadevrt)
            cubin = ls.complete()
        else:
            cubin = ptx
        mod._set_mapping(mapping)
    elif backend == 'nvcc':
        rdc = _is_cudadevrt_needed(options)
        cubin = compile_using_nvcc(source, options, arch,
                                   name + '.cu', code_type='cubin',
                                   separate_compilation=rdc,
                                   log_stream=log_stream)
    else:
        raise ValueError('Invalid backend %s' % backend)

    if not cache_in_memory:
        # Write to disk cache
        cubin_hash = _hash_hexdigest(cubin).encode('ascii')

        # shutil.move is not atomic operation, so it could result in a
        # corrupted file. We detect it by appending a hash at the beginning
        # of each cache file. If the file is corrupted, it will be ignored
        # next time it is read.
        with tempfile.NamedTemporaryFile(dir=cache_dir, delete=False) as tf:
            tf.write(cubin_hash)
            tf.write(cubin)
            temp_path = tf.name
        shutil.move(temp_path, path)

        # Save .cu source file along with .cubin
        if _get_bool_env_variable('CUPY_CACHE_SAVE_CUDA_SOURCE', False):
            with open(path + '.cu', 'w') as f:
                f.write(source)
    else:
        # we don't do any disk I/O
        pass

    mod.load(cubin)
    return mod


class CompileException(Exception):

    def __init__(self, msg, source, name, options, backend='nvrtc'):
        self._msg = msg
        self.source = source
        self.name = name
        self.options = options
        self.backend = backend
        super(CompileException, self).__init__()

    def __reduce__(self):
        return (type(self), (self._msg, self.source, self.name,
                             self.options, self.backend))

    def __repr__(self):
        return str(self)

    def __str__(self):
        return self.get_message()

    def get_message(self):
        return self._msg

    def dump(self, f):
        lines = self.source.split('\n')
        digits = int(math.floor(math.log10(len(lines)))) + 1
        linum_fmt = '{{:0{}d}} '.format(digits)
        f.write('{} '.format(self.backend.upper()))
        f.write('compilation error: {}\n'.format(self))
        f.write('-----\n')
        f.write('Name: {}\n'.format(self.name))
        f.write('Options: {}\n'.format(' '.join(self.options)))
        f.write('CUDA source:\n')
        for i, line in enumerate(lines):
            f.write(linum_fmt.format(i + 1) + line.rstrip() + '\n')
        f.write('-----\n')
        f.flush()


class _NVRTCProgram(object):

    def __init__(self, src, name='default_program', headers=(),
                 include_names=(), name_expressions=None, method='ptx'):
        self.ptr = None

        if isinstance(src, bytes):
            src = src.decode('UTF-8')
        if isinstance(name, bytes):
            name = name.decode('UTF-8')

        self.src = src
        self.name = name
        self.ptr = nvrtc.createProgram(src, name, headers, include_names)
        self.name_expressions = name_expressions
        self.method = method

    def __del__(self, is_shutting_down=_util.is_shutting_down):
        if is_shutting_down():
            return
        if self.ptr:
            nvrtc.destroyProgram(self.ptr)

    def compile(self, options=(), log_stream=None):
        try:
            if self.name_expressions:
                for ker in self.name_expressions:
                    nvrtc.addNameExpression(self.ptr, ker)
            nvrtc.compileProgram(self.ptr, options)
            mapping = None
            if self.name_expressions:
                mapping = {}
                for ker in self.name_expressions:
                    mapping[ker] = nvrtc.getLoweredName(self.ptr, ker)
            if log_stream is not None:
                log_stream.write(nvrtc.getProgramLog(self.ptr))
            # This is to ensure backwards compatibility with nvrtc
            if self.method == 'cubin':
                return nvrtc.getCUBIN(self.ptr), mapping
            elif self.method == 'ptx':
                return nvrtc.getPTX(self.ptr), mapping
            # TODO(leofang): support JIT LTO using nvrtc.getNVVM()?
            # need -dlto and -arch=compute_XX
            else:
                raise RuntimeError('Unknown NVRTC compile method')
        except nvrtc.NVRTCError:
            log = nvrtc.getProgramLog(self.ptr)
            raise CompileException(log, self.src, self.name, options,
                                   'nvrtc' if not runtime.is_hip else 'hiprtc')


def is_valid_kernel_name(name):
    return re.match('^[a-zA-Z_][a-zA-Z_0-9]*$', name) is not None


def compile_using_hipcc(source, options, arch, log_stream=None):
    # As of ROCm 3.5.0 hiprtc/hipcc can automatically pick up the
    # right arch without setting HCC_AMDGPU_TARGET, so we don't need
    # to set arch here
    cmd = ['hipcc', '--genco'] + list(options)

    with tempfile.TemporaryDirectory() as root_dir:
        path = os.path.join(root_dir, 'kern')
        in_path = path + '.cpp'
        out_path = path + '.hsaco'

        with open(in_path, 'w') as f:
            f.write(source)

        cmd += [in_path, '-o', out_path]

        try:
            output = _run_cc(cmd, root_dir, 'hipcc', log_stream)
        except HIPCCException as e:
            cex = CompileException(str(e), source, in_path, options,
                                   'hipcc')

            dump = _get_bool_env_variable(
                'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
            if dump:
                cex.dump(sys.stderr)

            raise cex
        if not os.path.isfile(out_path):
            raise HIPCCException(
                '`hipcc` command does not generate output file. \n'
                'command: {0}\n'
                'stdout/stderr: \n'
                '{1}'.format(cmd, output))
        with open(out_path, 'rb') as f:
            return f.read()


# TODO(leofang): consider merge _preprocess_hipcc with _preprocess_hiprtc,
# perhaps also with _preprocess?
def _preprocess_hipcc(source, options):
    cmd = ['hipcc', '--preprocess'] + list(options)
    with tempfile.TemporaryDirectory() as root_dir:
        path = os.path.join(root_dir, 'kern')
        cu_path = '%s.cpp' % path

        with open(cu_path, 'w') as cu_file:
            cu_file.write(source)

        cmd.append(cu_path)
        pp_src = _run_cc(cmd, root_dir, 'hipcc')
        assert isinstance(pp_src, str)
        return re.sub('(?m)^#.*$', '', pp_src)


def _preprocess_hiprtc(source, options):
    # source is ignored
    if _cuda_hip_version >= 40400000:
        # HIP runtime headers can be no longer explicitly included on ROCm 4.5+
        code = '''
        // hiprtc segfaults if the input code is empty
        __global__ void _cupy_preprocess_dummy_kernel_() { }
        '''
    else:
        code = '''
        // hiprtc segfaults if the input code is empty
        #include <hip/hip_runtime.h>
        __global__ void _cupy_preprocess_dummy_kernel_() { }
        '''

    prog = _NVRTCProgram(code)
    try:
        result, _ = prog.compile(options)
    except CompileException as e:
        dump = _get_bool_env_variable(
            'CUPY_DUMP_CUDA_SOURCE_ON_ERROR', False)
        if dump:
            e.dump(sys.stderr)
        raise
    assert isinstance(result, bytes)
    return result


_hip_extra_source = None


def _convert_to_hip_source(source, extra_source, is_hiprtc):
    if not is_hiprtc:
        return '#include <hip/hip_runtime.h>\n' + source
    if _cuda_hip_version >= 40400000:
        # HIP runtime headers can be no longer explicitly included on ROCm 4.5+
        return source
    if _cuda_hip_version >= 402:
        # "-I" is fixed on ROCm 4.2.0+
        return '#include <hip/hip_runtime.h>\n' + source

    # Workaround for hiprtc: it does not follow the -I option to search
    # headers (as of ROCm 3.5.0), so we must prepend all CuPy's headers
    global _hip_extra_source
    if _hip_extra_source is None:
        if extra_source is not None:
            extra_source = extra_source.split('\n')
            extra_source = [line for line in extra_source if (
                not line.startswith('#include')
                and not line.startswith('#pragma once'))]
            _hip_extra_source = extra_source = '\n'.join(extra_source)

    source = source.split('\n')
    source = [line for line in source if not line.startswith('#include')]
    source = ('#include <hip/hip_runtime.h>\n#include <hip/hip_fp16.h>\n'
              + _hip_extra_source + '\n'.join(source))

    return source


# TODO(leofang): evaluate if this can be merged with _compile_with_cache_cuda()
def _compile_with_cache_hip(source, options, arch, cache_dir, extra_source,
                            backend='hiprtc', name_expressions=None,
                            log_stream=None, cache_in_memory=False,
                            use_converter=True):
    global _empty_file_preprocess_cache

    # TODO(leofang): this might be possible but is currently undocumented
    if _is_cudadevrt_needed(options):
        raise ValueError('separate compilation is not supported in HIP')

    # HIP's equivalent of -ftz=true, see ROCm-Developer-Tools/HIP#2252
    # Notes:
    # - For hipcc, this should just work, as invalid options would cause errors
    #   See https://clang.llvm.org/docs/ClangCommandLineReference.html.
    # - For hiprtc, this is a no-op until the compiler options like -D and -I
    #   are accepted, see ROCm-Developer-Tools/HIP#2182 and
    #   ROCm-Developer-Tools/HIP#2248
    options += ('-fcuda-flush-denormals-to-zero',)

    # Workaround ROCm 4.3 LLVM_PATH issue in hipRTC #5689
    rocm_build_version = driver.get_build_version()
    if rocm_build_version >= 40300000 and rocm_build_version < 40500000:
        options += (
            '-I' + get_rocm_path() + '/llvm/lib/clang/13.0.0/include/',)

    if cache_dir is None:
        cache_dir = get_cache_dir()
    # As of ROCm 3.5.0 hiprtc/hipcc can automatically pick up the
    # right arch without setting HCC_AMDGPU_TARGET, so we don't need
    # to tell the compiler which arch we are targeting. But, we still
    # need to know arch as part of the cache key:
    if arch is None:
        # On HIP, gcnArch is computed from "compute capability":
        # https://github.com/ROCm-Developer-Tools/HIP/blob/rocm-4.0.0/rocclr/hip_device.cpp#L202
        arch = device.Device().compute_capability
    if use_converter:
        source = _convert_to_hip_source(source, extra_source,
                                        is_hiprtc=(backend == 'hiprtc'))

    env = (arch, options, _get_nvrtc_version(), backend)
    base = _empty_file_preprocess_cache.get(env, None)
    if base is None:
        # This is for checking HIPRTC/HIPCC compiler internal version
        if backend == 'hiprtc':
            base = _preprocess_hiprtc('', options)
        else:
            base = _preprocess_hipcc('', options)
        _empty_file_preprocess_cache[env] = base

    key_src = '%s %s %s %s' % (env, base, source, extra_source)
    key_src = key_src.encode('utf-8')
    name = _hash_hexdigest(key_src) + '.hsaco'

    mod = function.Module()

    if not cache_in_memory:
        # Read from disk cache
        if not os.path.isdir(cache_dir):
            os.makedirs(cache_dir, exist_ok=True)

        # To handle conflicts in concurrent situation, we adopt lock-free
        # method to avoid performance degradation.
        # We force recompiling to retrieve C++ mangled names if so desired.
        path = os.path.join(cache_dir, name)
        if os.path.exists(path) and not name_expressions:
            with open(path, 'rb') as f:
                data = f.read()
            if len(data) >= _hash_length:
                hash_value = data[:_hash_length]
                binary = data[_hash_length:]
                binary_hash = _hash_hexdigest(binary).encode('ascii')
                if hash_value == binary_hash:
                    mod.load(binary)
                    return mod
    else:
        # Enforce compiling -- the resulting kernel will be cached elsewhere,
        # so we do nothing
        pass

    if backend == 'hiprtc':
        # compile_using_nvrtc calls hiprtc for hip builds
        binary, mapping = compile_using_nvrtc(
            source, options, arch, name + '.cu', name_expressions,
            log_stream, cache_in_memory)
        mod._set_mapping(mapping)
    else:
        binary = compile_using_hipcc(source, options, arch, log_stream)

    if not cache_in_memory:
        # Write to disk cache
        binary_hash = _hash_hexdigest(binary).encode('ascii')

        # shutil.move is not atomic operation, so it could result in a
        # corrupted file. We detect it by appending a hash at the beginning
        # of each cache file. If the file is corrupted, it will be ignored
        # next time it is read.
        with tempfile.NamedTemporaryFile(dir=cache_dir, delete=False) as tf:
            tf.write(binary_hash)
            tf.write(binary)
            temp_path = tf.name
        shutil.move(temp_path, path)

        # Save .cu source file along with .hsaco
        if _get_bool_env_variable('CUPY_CACHE_SAVE_CUDA_SOURCE', False):
            with open(path + '.cpp', 'w') as f:
                f.write(source)
    else:
        # we don't do any disk I/O
        pass

    mod.load(binary)
    return mod