import subprocess
import os
import os.path
import errno
import re
import shutil
import sys
import fnmatch
import errno
import platform
import glob

from contextlib import contextmanager
from setuptools import Command
from setuptools import setup, Distribution, Extension
from setuptools.command.install import install as InstallCommandBase
from setuptools.command.egg_info import egg_info


class BinaryDistribution(Distribution):
    def has_ext_modules(foo):
        return True

RC      = 0

ext_name = '.dll' if os.name == 'nt' else ('.dylib' if sys.platform == 'darwin' else '.so')

def git_commit():
    try:
        cmd = ['git', 'rev-parse', 'HEAD']
        git_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE,
            cwd="@PADDLE_SOURCE_DIR@").communicate()[0].strip()
    except:
        git_commit = 'Unknown'
    git_commit = git_commit.decode()
    return str(git_commit)

def _get_version_detail(idx):
    assert idx < 3, "vesion info consists of %(major)d.%(minor)d.%(patch)d, \
        so detail index must less than 3"

    if re.match('@TAG_VERSION_REGEX@', '@PADDLE_VERSION@'):
        version_details = '@PADDLE_VERSION@'.split('.')

        if len(version_details) >= 3:
            return version_details[idx]

    return 0

def get_major():
    return int(_get_version_detail(0))

def get_minor():
    return int(_get_version_detail(1))

def get_patch():
    return str(_get_version_detail(2))

def get_cuda_version():
    if '@WITH_GPU@' == 'ON':
        return '@CUDA_VERSION@'
    else:
        return 'False'

def get_cudnn_version():
    if '@WITH_GPU@' == 'ON':
        temp_cudnn_version = ''
        if '@CUDNN_MAJOR_VERSION@':
            temp_cudnn_version += '@CUDNN_MAJOR_VERSION@'
            if '@CUDNN_MINOR_VERSION@':
                temp_cudnn_version += '.@CUDNN_MINOR_VERSION@'
                if '@CUDNN_PATCHLEVEL_VERSION@':
                    temp_cudnn_version += '.@CUDNN_PATCHLEVEL_VERSION@'
        return temp_cudnn_version
    else:
        return 'False'

def get_xpu_version():
    if '@WITH_XPU@' == 'ON':
        return '@XPU_BASE_DATE@'
    else:
        return 'False'

def get_xpu_xccl_version():
    if '@WITH_XPU_BKCL@' == 'ON':
        return '@XPU_XCCL_BASE_VERSION@'
    else:
        return 'False'

def is_taged():
    try:
        cmd = ['git', 'describe', '--exact-match', '--tags', 'HEAD', '2>/dev/null']
        git_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE, cwd="@PADDLE_SOURCE_DIR@").communicate()[0].strip()
        git_tag = git_tag.decode()
    except:
        return False

    if str(git_tag).replace('v', '') == '@PADDLE_VERSION@':
        return True
    else:
        return False

def get_cinn_version():
    if '@WITH_CINN@' != 'ON':
        return "False"

    cinn_git_version = 'Unknown'
    try:
        cmd = ['git', 'describe', '--exact-match', '--tags', 'HEAD', '2>/dev/null']
        cinn_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE, cwd='@CINN_SOURCE_DIR@').communicate()[0].strip()
        if len(cinn_tag) > 0:
            cinn_git_version = cinn_tag
    except:
        pass

    if cinn_git_version == 'Unknown':
        try:
            cmd = ['git', 'rev-parse', 'HEAD']
            cinn_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE,
                cwd='@CINN_SOURCE_DIR@').communicate()[0].strip()
            if len(cinn_commit) > 0:
                cinn_git_version = cinn_commit
        except:
            pass

    cinn_git_version = cinn_git_version.decode('utf-8')
    return str(cinn_git_version)

def write_version_py(filename='paddle/version/__init__.py'):
    cnt = '''# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY
#
full_version     = '%(major)d.%(minor)d.%(patch)s'
major            = '%(major)d'
minor            = '%(minor)d'
patch            = '%(patch)s'
rc               = '%(rc)d'
cuda_version     = '%(cuda)s'
cudnn_version    = '%(cudnn)s'
xpu_version      = '%(xpu)s'
xpu_xccl_version = '%(xpu_xccl)s'
istaged          = %(istaged)s
commit           = '%(commit)s'
with_mkl         = '%(with_mkl)s'
cinn_version      = '%(cinn)s'

__all__ = ['cuda', 'cudnn', 'show', 'xpu', 'xpu_xccl']

def show():
    """Get the version of paddle if `paddle` package if tagged. Otherwise, output the corresponding commit id.

    Returns:
        If paddle package is not tagged, the commit-id of paddle will be output.
        Otherwise, the following information will be output.

        full_version: version of paddle

        major: the major version of paddle

        minor: the minor version of paddle

        patch: the patch level version of paddle

        rc: whether it's rc version

        cuda: the cuda version of package. It will return `False` if CPU version paddle package is installed

        cudnn: the cudnn version of package. It will return `False` if CPU version paddle package is installed

        xpu: the xpu version of package. It will return `False` if non-XPU version paddle package is installed

        xpu_xccl: the xpu xccl version of package. It will return `False` if non-XPU version paddle package is installed

        cinn: the cinn version of package. It will return `False` if paddle package is not compiled with CINN

    Examples:
        .. code-block:: python

            import paddle

            # Case 1: paddle is tagged with 2.2.0
            paddle.version.show()
            # full_version: 2.2.0
            # major: 2
            # minor: 2
            # patch: 0
            # rc: 0
            # cuda: '10.2'
            # cudnn: '7.6.5'
            # xpu: '20230114'
            # xpu_xccl: '1.0.7'
            # cinn: False

            # Case 2: paddle is not tagged
            paddle.version.show()
            # commit: cfa357e984bfd2ffa16820e354020529df434f7d
            # cuda: '10.2'
            # cudnn: '7.6.5'
            # xpu: '20230114'
            # xpu_xccl: '1.0.7'
            # cinn: False
    """
    if istaged:
        print('full_version:', full_version)
        print('major:', major)
        print('minor:', minor)
        print('patch:', patch)
        print('rc:', rc)
    else:
        print('commit:', commit)
    print('cuda:', cuda_version)
    print('cudnn:', cudnn_version)
    print('xpu:', xpu_version)
    print('xpu_xccl:', xpu_xccl_version)
    print('cinn:', cinn_version)

def mkl():
    return with_mkl

def cuda():
    """Get cuda version of paddle package.

    Returns:
        string: Return the version information of cuda. If paddle package is CPU version, it will return False.

    Examples:
        .. code-block:: python

            import paddle

            paddle.version.cuda()
            # '10.2'

    """
    return cuda_version

def cudnn():
    """Get cudnn version of paddle package.

    Returns:
        string: Return the version information of cudnn. If paddle package is CPU version, it will return False.

    Examples:
        .. code-block:: python

            import paddle

            paddle.version.cudnn()
            # '7.6.5'

    """
    return cudnn_version

def xpu():
    """Get xpu version of paddle package.

    Returns:
        string: Return the version information of xpu. If paddle package is non-XPU version, it will return False.

    Examples:
        .. code-block:: python

            import paddle

            paddle.version.xpu()
            # '20230114'

    """
    return xpu_version

def xpu_xccl():
    """Get xpu xccl version of paddle package.

    Returns:
        string: Return the version information of xpu xccl. If paddle package is non-XPU version, it will return False.

    Examples:
        .. code-block:: python

            import paddle

            paddle.version.xpu_xccl()
            # '1.0.7'

    """
    return xpu_xccl_version

def cinn():
    """Get CINN version of paddle package.

    Returns:
        string: Return the version information of CINN. If paddle package is not compiled with CINN, it will return False.

    Examples:
        .. code-block:: python

            import paddle

            paddle.version.cinn()
            # False

    """
    return cinn_version
'''
    commit = git_commit()

    dirname = os.path.dirname(filename)

    try:
        os.makedirs(dirname)
    except OSError as e:
        if e.errno != errno.EEXIST:
            raise

    with open(filename, 'w') as f:
        f.write(cnt % {
            'major': get_major(),
            'minor': get_minor(),
            'patch': get_patch(),
            'rc': RC,
            'version': '${PADDLE_VERSION}',
            'cuda': get_cuda_version(),
            'cudnn': get_cudnn_version(),
            'xpu': get_xpu_version(),
            'xpu_xccl': get_xpu_xccl_version(),
            'commit': commit,
            'istaged': is_taged(),
            'with_mkl': '@WITH_MKL@',
            'cinn': get_cinn_version()})

write_version_py(filename='@PADDLE_BINARY_DIR@/python/paddle/version/__init__.py')

def write_cuda_env_config_py(filename='paddle/cuda_env.py'):
    cnt = ""
    if '${JIT_RELEASE_WHL}' == 'ON':
        cnt = '''# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY
#
import os
os.environ['CUDA_CACHE_MAXSIZE'] = '805306368'
'''

    with open(filename, 'w') as f:
        f.write(cnt)

write_cuda_env_config_py(filename='@PADDLE_BINARY_DIR@/python/paddle/cuda_env.py')

def write_distributed_training_mode_py(filename='paddle/incubate/distributed/fleet/parameter_server/version.py'):
    cnt = '''

# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY

from paddle.incubate.distributed.fleet.base import Mode

BUILD_MODE=Mode.%(mode)s

def is_transpiler():
    return Mode.TRANSPILER == BUILD_MODE

'''

    dirname = os.path.dirname(filename)

    try:
        os.makedirs(dirname)
    except OSError as e:
        if e.errno != errno.EEXIST:
            raise

    with open(filename, 'w') as f:
        f.write(cnt % {
            'mode': 'PSLIB' if '${WITH_PSLIB}' == 'ON' else 'TRANSPILER'
        })

write_distributed_training_mode_py(filename='@PADDLE_BINARY_DIR@/python/paddle/incubate/distributed/fleet/parameter_server/version.py')


packages=['paddle',
          'paddle.libs',
          'paddle.utils',
          'paddle.utils.gast',
          'paddle.utils.cpp_extension',
          'paddle.dataset',
          'paddle.reader',
          'paddle.distributed',
          'paddle.distributed.communication',
          'paddle.distributed.communication.stream',
          'paddle.distributed.metric',
          'paddle.distributed.ps',
          'paddle.distributed.ps.utils',
          'paddle.incubate',
          'paddle.incubate.autograd',
          'paddle.incubate.optimizer',
          'paddle.incubate.checkpoint',
          'paddle.incubate.operators',
          'paddle.incubate.tensor',
          'paddle.incubate.multiprocessing',
          'paddle.incubate.nn',
          'paddle.incubate.asp',
          'paddle.incubate.passes',
          'paddle.distribution',
          'paddle.distributed.utils',
          'paddle.distributed.sharding',
          'paddle.distributed.fleet',
          'paddle.distributed.launch',
          'paddle.distributed.launch.context',
          'paddle.distributed.launch.controllers',
          'paddle.distributed.launch.job',
          'paddle.distributed.launch.plugins',
          'paddle.distributed.launch.utils',
          'paddle.distributed.fleet.base',
          'paddle.distributed.fleet.recompute',
          'paddle.distributed.fleet.elastic',
          'paddle.distributed.fleet.meta_optimizers',
          'paddle.distributed.fleet.meta_optimizers.sharding',
          'paddle.distributed.fleet.meta_optimizers.dygraph_optimizer',
          'paddle.distributed.fleet.runtime',
          'paddle.distributed.rpc',
          'paddle.distributed.fleet.dataset',
          'paddle.distributed.fleet.data_generator',
          'paddle.distributed.fleet.metrics',
          'paddle.distributed.fleet.proto',
          'paddle.distributed.fleet.utils',
          'paddle.distributed.fleet.layers',
          'paddle.distributed.fleet.layers.mpu',
          'paddle.distributed.fleet.meta_parallel',
          'paddle.distributed.fleet.meta_parallel.pp_utils',
          'paddle.distributed.fleet.meta_parallel.sharding',
          'paddle.distributed.fleet.meta_parallel.parallel_layers',
          'paddle.distributed.auto_parallel',
          'paddle.distributed.auto_parallel.dygraph',
          'paddle.distributed.auto_parallel.static',
          'paddle.distributed.auto_parallel.static.operators',
          'paddle.distributed.auto_parallel.static.tuner',
          'paddle.distributed.auto_parallel.static.cost',
          'paddle.distributed.passes',
          'paddle.distributed.models',
          'paddle.distributed.models.moe',
          'paddle.distributed.transpiler',
          'paddle.distributed.transpiler.details',
          'paddle.framework',
          'paddle.jit',
          'paddle.jit.dy2static',
          'paddle.inference',
          'paddle.inference.contrib',
          'paddle.inference.contrib.utils',
          'paddle.fluid',
          'paddle.fluid.dygraph',
          'paddle.fluid.proto',
          'paddle.fluid.proto.profiler',
          'paddle.fluid.layers',
          'paddle.fluid.contrib',
          'paddle.fluid.contrib.extend_optimizer',
          'paddle.fluid.incubate',
          'paddle.incubate.distributed.fleet',
          'paddle.fluid.incubate.checkpoint',
          'paddle.amp',
          'paddle.cost_model',
          'paddle.hapi',
          'paddle.vision',
          'paddle.vision.models',
          'paddle.vision.transforms',
          'paddle.vision.datasets',
          'paddle.audio',
	      'paddle.audio.functional',
	      'paddle.audio.features',
	      'paddle.audio.datasets',
	      'paddle.audio.backends',
          'paddle.text',
          'paddle.text.datasets',
          'paddle.incubate',
          'paddle.incubate.nn',
          'paddle.incubate.nn.functional',
          'paddle.incubate.nn.layer',
          'paddle.incubate.optimizer.functional',
          'paddle.incubate.autograd',
          'paddle.incubate.distributed',
          'paddle.incubate.distributed.utils',
          'paddle.incubate.distributed.utils.io',
          'paddle.incubate.distributed.fleet',
          'paddle.incubate.distributed.models',
          'paddle.incubate.distributed.models.moe',
          'paddle.incubate.distributed.models.moe.gate',
          'paddle.incubate.distributed.fleet.parameter_server',
          'paddle.incubate.distributed.fleet.parameter_server.distribute_transpiler',
          'paddle.incubate.distributed.fleet.parameter_server.pslib',
          'paddle.incubate.distributed.fleet.parameter_server.ir',
          'paddle.incubate.layers',
          'paddle.quantization',
          'paddle.quantization.quanters',
          'paddle.quantization.observers',
          'paddle.sparse',
          'paddle.sparse.nn',
          'paddle.sparse.nn.layer',
          'paddle.sparse.nn.functional',
          'paddle.incubate.xpu',
          'paddle.io',
          'paddle.io.dataloader',
          'paddle.optimizer',
          'paddle.nn',
          'paddle.nn.functional',
          'paddle.nn.layer',
          'paddle.nn.quant',
          'paddle.nn.quant.qat',
          'paddle.nn.initializer',
          'paddle.nn.utils',
          'paddle.metric',
          'paddle.static',
          'paddle.static.nn',
          'paddle.static.amp',
          'paddle.static.amp.bf16',
          'paddle.static.quantization',
          'paddle.quantization',
          'paddle.quantization.imperative',
          'paddle.tensor',
          'paddle.onnx',
          'paddle.autograd',
          'paddle.device',
          'paddle.device.cuda',
          'paddle.device.xpu',
          'paddle.version',
          'paddle.profiler',
          'paddle.geometric',
          'paddle.geometric.message_passing',
          'paddle.geometric.sampling',
          ]

with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f:
    setup_requires = f.read().splitlines()

if sys.version_info < (3,7):
    raise RuntimeError("Paddle only support Python version>=3.7 now")

if sys.version_info >= (3,7):
    setup_requires_tmp = []
    for setup_requires_i in setup_requires:
        if "<\"3.6\"" in setup_requires_i or "<=\"3.6\"" in setup_requires_i or "<\"3.5\"" in setup_requires_i or "<=\"3.5\"" in setup_requires_i or "<\"3.7\"" in setup_requires_i:
            continue
        setup_requires_tmp+=[setup_requires_i]
    setup_requires = setup_requires_tmp

# the prefix is sys.prefix which should always be usr
paddle_bins = ''

if not '${WIN32}':
    paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle']

if os.name != 'nt':
    package_data={'paddle.fluid': ['${FLUID_CORE_NAME}' + '.so']}
else:
    package_data={'paddle.fluid': ['${FLUID_CORE_NAME}' + '.pyd', '${FLUID_CORE_NAME}' + '.lib']}

package_data['paddle.fluid'] += ['${PADDLE_BINARY_DIR}/python/paddle/cost_model/static_op_benchmark.json']

package_dir={
    '': '${PADDLE_BINARY_DIR}/python',
    # The paddle.fluid.proto will be generated while compiling.
    # So that package points to other directory.
    'paddle.fluid.proto.profiler': '${PADDLE_BINARY_DIR}/paddle/fluid/platform',
    'paddle.fluid.proto': '${PADDLE_BINARY_DIR}/paddle/fluid/framework',
    'paddle.fluid': '${PADDLE_BINARY_DIR}/python/paddle/fluid',
}

# put all thirdparty libraries in paddle.libs
libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'

package_data['paddle.libs']= []
if('${WITH_SHARED_PHI}' == 'ON'):
    package_data['paddle.libs'] += [('libphi' if os.name != 'nt' else 'phi') + ext_name]
    shutil.copy('${PHI_LIB}', libs_path)

if('${WITH_SHARED_IR}' == 'ON'):
    package_data['paddle.libs'] += [('libir' if os.name != 'nt' else 'ir') + ext_name]
    shutil.copy('${IR_LIB}', libs_path)

package_data['paddle.libs']+=[
    ('libwarpctc' if os.name != 'nt' else 'warpctc') + ext_name,
    ('libwarprnnt' if os.name != 'nt' else 'warprnnt') + ext_name,
]
shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
shutil.copy('${WARPRNNT_LIBRARIES}', libs_path)

package_data['paddle.libs']+=[
    os.path.basename('${LAPACK_LIB}'),
    os.path.basename('${BLAS_LIB}'),
    os.path.basename('${GFORTRAN_LIB}'),
    os.path.basename('${GNU_RT_LIB_1}')]
shutil.copy('${BLAS_LIB}', libs_path)
shutil.copy('${LAPACK_LIB}', libs_path)
shutil.copy('${GFORTRAN_LIB}', libs_path)
shutil.copy('${GNU_RT_LIB_1}', libs_path)
if '${WITH_CUDNN_DSO}' == 'ON' and os.path.exists('${CUDNN_LIBRARY}'):
    package_data['paddle.libs']+=[os.path.basename('${CUDNN_LIBRARY}')]
    shutil.copy('${CUDNN_LIBRARY}', libs_path)
    if sys.platform.startswith("linux") and '${CUDNN_MAJOR_VERSION}' == '8':
        # libcudnn.so includes libcudnn_ops_infer.so, libcudnn_ops_train.so,
        # libcudnn_cnn_infer.so, libcudnn_cnn_train.so, libcudnn_adv_infer.so,
        # libcudnn_adv_train.so
        cudnn_lib_files = glob.glob(os.path.dirname('${CUDNN_LIBRARY}') + '/libcudnn_*so.8')
        for cudnn_lib in cudnn_lib_files:
            if os.path.exists(cudnn_lib):
                package_data['paddle.libs']+=[os.path.basename(cudnn_lib)]
                shutil.copy(cudnn_lib, libs_path)

if not sys.platform.startswith("linux"):
    package_data['paddle.libs']+=[os.path.basename('${GNU_RT_LIB_2}')]
    shutil.copy('${GNU_RT_LIB_2}', libs_path)

if len('${FLASHATTN_LIBRARIES}') > 1:
    package_data['paddle.libs']+=[os.path.basename('${FLASHATTN_LIBRARIES}')]
    shutil.copy('${FLASHATTN_LIBRARIES}', libs_path)

if '${WITH_MKL}' == 'ON':
    shutil.copy('${MKLML_SHARED_LIB}', libs_path)
    shutil.copy('${MKLML_SHARED_IOMP_LIB}', libs_path)
    package_data['paddle.libs']+=[('libmklml_intel' if os.name != 'nt' else 'mklml') + ext_name, ('libiomp5' if os.name != 'nt' else 'libiomp5md') + ext_name]
else:
    if os.name == 'nt':
        # copy the openblas.dll
        shutil.copy('${OPENBLAS_SHARED_LIB}', libs_path)
        package_data['paddle.libs'] += ['openblas' + ext_name]
    elif os.name == 'posix' and platform.machine() == 'aarch64' and '${OPENBLAS_LIB}'.endswith('so'):
        # copy the libopenblas.so on linux+aarch64
        # special: libpaddle.so without avx depends on 'libopenblas.so.0', not 'libopenblas.so'
        if os.path.exists('${OPENBLAS_LIB}' + '.0'):
            shutil.copy('${OPENBLAS_LIB}' + '.0', libs_path)
            package_data['paddle.libs'] += ['libopenblas.so.0']

if '${WITH_LITE}' == 'ON':
    shutil.copy('${LITE_SHARED_LIB}', libs_path)
    package_data['paddle.libs']+=['libpaddle_full_api_shared' + ext_name]
    if '${LITE_WITH_NNADAPTER}' == 'ON':
        shutil.copy('${LITE_NNADAPTER_LIB}', libs_path)
        package_data['paddle.libs']+=['libnnadapter' + ext_name]
        if '${NNADAPTER_WITH_HUAWEI_ASCEND_NPU}' == 'ON':
            shutil.copy('${LITE_NNADAPTER_NPU_LIB}', libs_path)
            package_data['paddle.libs']+=['libnnadapter_driver_huawei_ascend_npu' + ext_name]

if '${WITH_CINN}' == 'ON':
    shutil.copy('${CINN_LIB_LOCATION}/${CINN_LIB_NAME}', libs_path)
    shutil.copy('${CINN_INCLUDE_DIR}/cinn/runtime/cuda/cinn_cuda_runtime_source.cuh', libs_path)
    package_data['paddle.libs']+=['libcinnapi.so']
    package_data['paddle.libs']+=['cinn_cuda_runtime_source.cuh']

    cinn_fp16_file = '${CINN_INCLUDE_DIR}/cinn/runtime/cuda/float16.h'
    if os.path.exists(cinn_fp16_file):
        shutil.copy(cinn_fp16_file, libs_path)
        package_data['paddle.libs']+=['float16.h']

    cinn_bf16_file = '${CINN_INCLUDE_DIR}/cinn/runtime/cuda/bfloat16.h'
    if os.path.exists(cinn_bf16_file):
        shutil.copy(cinn_bf16_file, libs_path)
        package_data['paddle.libs']+=['bfloat16.h']

    if '${CMAKE_BUILD_TYPE}' == 'Release' and os.name != 'nt':
        command = "patchelf --set-rpath '$ORIGIN/' %s/${CINN_LIB_NAME}" % libs_path
        if os.system(command) != 0:
            raise Exception("patch %s/${CINN_LIB_NAME} failed, command: %s" % (libs_path, command))

if '${WITH_PSLIB}' == 'ON':
    shutil.copy('${PSLIB_LIB}', libs_path)
    shutil.copy('${JVM_LIB}', libs_path)
    if os.path.exists('${PSLIB_VERSION_PY}'):
        shutil.copy('${PSLIB_VERSION_PY}', '${PADDLE_BINARY_DIR}/python/paddle/incubate/distributed/fleet/parameter_server/pslib/')
    package_data['paddle.libs'] += ['libps' + ext_name]
    package_data['paddle.libs'] += ['libjvm' + ext_name]

if '${WITH_MKLDNN}' == 'ON':
    if '${CMAKE_BUILD_TYPE}' == 'Release' and os.name != 'nt':
        # only change rpath in Release mode.
        # TODO(typhoonzero): use install_name_tool to patch mkl libs once
        # we can support mkl on mac.
        #
        # change rpath of libdnnl.so.1, add $ORIGIN/ to it.
        # The reason is that all thirdparty libraries in the same directory,
        # thus, libdnnl.so.1 will find libmklml_intel.so and libiomp5.so.
        command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}"
        if os.system(command) != 0:
            raise Exception("patch libdnnl.so failed, command: %s" % command)
    shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
    if os.name != 'nt':
        shutil.copy('${MKLDNN_SHARED_LIB_1}', libs_path)
        shutil.copy('${MKLDNN_SHARED_LIB_2}', libs_path)
        package_data['paddle.libs']+=['libmkldnn.so.0', 'libdnnl.so.1', 'libdnnl.so.2']
    else:
        package_data['paddle.libs']+=['mkldnn.dll']

if '${WITH_ONNXRUNTIME}' == 'ON':
    shutil.copy('${ONNXRUNTIME_SHARED_LIB}', libs_path)
    shutil.copy('${PADDLE2ONNX_LIB}', libs_path)
    if os.name == 'nt':
        package_data['paddle.libs']+=['paddle2onnx.dll', 'onnxruntime.dll']
    else:
        package_data['paddle.libs']+=['${PADDLE2ONNX_LIB_NAME}', '${ONNXRUNTIME_LIB_NAME}']

if '${WITH_XPU}' == 'ON':
    # only change rpath in Release mode,
    if '${CMAKE_BUILD_TYPE}' == 'Release':
        if os.name != 'nt':
            if "@APPLE@" == "1":
                command = "install_name_tool -id \"@loader_path/\" ${XPU_API_LIB}"
            else:
                command = "patchelf --set-rpath '$ORIGIN/' ${XPU_API_LIB}"
            if os.system(command) != 0:
                raise Exception("patch ${XPU_API_LIB} failed, command: %s" % command)
    shutil.copy('${XPU_API_LIB}', libs_path)
    package_data['paddle.libs']+=['${XPU_API_LIB_NAME}']
    xpu_rt_lib_list = glob.glob('${XPU_RT_LIB}*')
    for xpu_rt_lib_file in xpu_rt_lib_list:
        shutil.copy(xpu_rt_lib_file, libs_path)
        package_data['paddle.libs']+=[os.path.basename(xpu_rt_lib_file)]

if '${WITH_XPU_BKCL}' == 'ON':
    shutil.copy('${XPU_BKCL_LIB}', libs_path)
    package_data['paddle.libs']+=['${XPU_BKCL_LIB_NAME}']

if '${WITH_XPU_XFT}' == 'ON':
    shutil.copy('${XPU_XFT_LIB}', libs_path)
    package_data['paddle.libs']+=['${XPU_XFT_LIB_NAME}']

if '${WITH_XPTI}' == 'ON':
    shutil.copy('${XPU_XPTI_LIB}', libs_path)
    package_data['paddle.libs']+=['${XPU_XPTI_LIB_NAME}']

# remove unused paddle/libs/__init__.py
if os.path.isfile(libs_path+'/__init__.py'):
    os.remove(libs_path+'/__init__.py')
package_dir['paddle.libs']=libs_path


# change rpath of ${FLUID_CORE_NAME}.ext, add $ORIGIN/../libs/ to it.
# The reason is that libwarpctc.ext, libiomp5.ext etc are in paddle.libs, and
# ${FLUID_CORE_NAME}.ext is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries.
# This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213
if '${CMAKE_BUILD_TYPE}' == 'Release':
    if os.name != 'nt':
        # only change rpath in Release mode, since in Debug mode, ${FLUID_CORE_NAME}.xx is too large to be changed.
        if "@APPLE@" == "1":
            commands = ["install_name_tool -id '@loader_path/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so']
            commands.append("install_name_tool -add_rpath '@loader_path/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so')
            if('${WITH_SHARED_PHI}' == 'ON'):
                # change rpath of phi.ext for loading 3rd party libb
                commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_NAME}")
            if('${WITH_SHARED_IR}' == 'ON'):
                # change rpath of ir.ext for loading 3rd party libb
                commands.append("install_name_tool -add_rpath '@loader_path' ${PADDLE_BINARY_DIR}/python/paddle/libs/${IR_NAME}")
        else:
            commands = ["patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/${FLUID_CORE_NAME}" + '.so']
            if('${WITH_SHARED_PHI}' == 'ON'):
                # change rpath of phi.ext for loading 3rd party lib
                commands.append("patchelf --set-rpath '$ORIGIN' ${PADDLE_BINARY_DIR}/python/paddle/libs/${PHI_NAME}")
            if('${WITH_SHARED_IR}' == 'ON'):
                # change rpath of ir.ext for loading 3rd party lib
                commands.append("patchelf --set-rpath '$ORIGIN' ${PADDLE_BINARY_DIR}/python/paddle/libs/${IR_NAME}")
        # The sw_64 not suppot patchelf, so we just disable that.
        if platform.machine() != 'sw_64' and platform.machine() != 'mips64':
            for command in commands:
                if os.system(command) != 0:
                    raise Exception("patch ${FLUID_CORE_NAME}.%s failed, command: %s" % (ext_name, command))

ext_modules = [Extension('_foo', ['stub.cc'])]
if os.name == 'nt':
    # fix the path separator under windows
    fix_package_dir = {}
    for k, v in package_dir.items():
        fix_package_dir[k] = v.replace('/', '\\')
    package_dir = fix_package_dir
    ext_modules = []
elif sys.platform == 'darwin':
    ext_modules = []

def find_files(pattern, root, recursive=False):
    for dirpath, _, files in os.walk(root):
        for filename in fnmatch.filter(files, pattern):
            yield os.path.join(dirpath, filename)
        if not recursive:
            break

headers = (
    # paddle level api headers (high level api, for both training and inference)
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle')) +
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/api')) +  # phi unify api header
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/api/ext')) +  # custom op api
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/api/include')) +  # phi api
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/common')) +  # phi common headers
    # phi level api headers (low level api, for training only)
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi')) +  # phi extension header
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/include', recursive=True)) +  # phi include headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/backends', recursive=True)) +  # phi backends headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/core', recursive=True)) +  # phi core headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/infermeta', recursive=True)) +  # phi infermeta headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/kernels', recursive=True)) +  # phi kernels headers
    # capi headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/capi', recursive=True)) +  # phi capi headers
    # phi profiler headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/phi/api/profiler')) +
    # utils api headers
    list(find_files('*.h', '@PADDLE_SOURCE_DIR@/paddle/utils', recursive=True)) + # paddle utils headers
    # init headers
    list(find_files('init_phi.h', '@PADDLE_SOURCE_DIR@/paddle/fluid/platform')))  # phi init headers

jit_layer_headers = ['layer.h', 'serializer.h', 'serializer_utils.h', 'all.h', 'function.h']
for f in jit_layer_headers:
    headers += list(find_files(f, '@PADDLE_SOURCE_DIR@/paddle/fluid/jit', recursive=True))

if '${WITH_MKLDNN}' == 'ON':
    headers += list(find_files('*', '${MKLDNN_INSTALL_DIR}/include')) # mkldnn

if '${WITH_GPU}' == 'ON' or '${WITH_ROCM}' == 'ON':
    # externalErrorMsg.pb for External Error message
    headers += list(find_files('*.pb', '${externalError_INCLUDE_DIR}'))

if '${WITH_XPU}' == 'ON':
    headers += list(find_files('*.h', '@PADDLE_BINARY_DIR@/third_party/xpu/src/extern_xpu/xpu', recursive=True)) # xdnn api headers

headers += list(find_files('*.h', '${PYBIND_INCLUDE_DIR}', True)) # pybind headers

def get_header_install_dir(header):
    if 'pb.h' in header:
        install_dir = re.sub('${PADDLE_BINARY_DIR}/', '', header)
    elif 'third_party' not in header:
        # paddle headers
        install_dir = re.sub('@PADDLE_SOURCE_DIR@/', '', header)
        print('install_dir: ', install_dir)
        if 'fluid/jit' in install_dir:
            install_dir = re.sub('fluid/jit', 'jit', install_dir)
            print('fluid/jit install_dir: ', install_dir)
    else:
        # third_party
        install_dir = re.sub('${THIRD_PARTY_PATH}', 'third_party', header)
        patterns = ['install/mkldnn/include', 'pybind/src/extern_pybind/include', 'third_party/xpu/src/extern_xpu/xpu/include/']
        for pattern in patterns:
            install_dir = re.sub(pattern, '', install_dir)
    return install_dir

class InstallCommand(InstallCommandBase):
    def finalize_options(self):
        ret = InstallCommandBase.finalize_options(self)
        self.install_lib = self.install_platlib
        self.install_headers = os.path.join(self.install_platlib, 'paddle', 'include')
        return ret

class InstallHeaders(Command):
    """Override how headers are copied.
    """
    description = 'install C/C++ header files'

    user_options = [('install-dir=', 'd',
                     'directory to install header files to'),
                    ('force', 'f',
                     'force installation (overwrite existing files)'),
                   ]

    boolean_options = ['force']

    def initialize_options(self):
        self.install_dir = None
        self.force = 0
        self.outfiles = []

    def finalize_options(self):
        self.set_undefined_options('install',
                                   ('install_headers', 'install_dir'),
                                   ('force', 'force'))

    def mkdir_and_copy_file(self, header):
        install_dir = get_header_install_dir(header)
        install_dir = os.path.join(self.install_dir, os.path.dirname(install_dir))
        if not os.path.exists(install_dir):
            self.mkpath(install_dir)
        return self.copy_file(header, install_dir)

    def run(self):
        hdrs = self.distribution.headers
        if not hdrs:
            return
        self.mkpath(self.install_dir)
        for header in hdrs:
            (out, _) = self.mkdir_and_copy_file(header)
            self.outfiles.append(out)

    def get_inputs(self):
        return self.distribution.headers or []

    def get_outputs(self):
        return self.outfiles

class EggInfo(egg_info):
    """Copy license file into `.dist-info` folder."""

    def run(self):
        # don't duplicate license into `.dist-info` when building a distribution
        if not self.distribution.have_run.get('install', True):
            self.mkpath(self.egg_info)
            self.copy_file("@PADDLE_SOURCE_DIR@/LICENSE", self.egg_info)

        egg_info.run(self)

# we redirect setuptools log for non-windows
if sys.platform != 'win32':
    @contextmanager
    def redirect_stdout():
        f_log = open('${SETUP_LOG_FILE}', 'w')
        origin_stdout = sys.stdout
        sys.stdout = f_log
        yield
        f_log = sys.stdout
        sys.stdout = origin_stdout
        f_log.close()
else:
    @contextmanager
    def redirect_stdout():
        yield

# Log for PYPI
with open("@PADDLE_BINARY_DIR@/python/paddle/README.md", "r", encoding='UTF-8') as f:
    long_description = f.read()

# strip *.so to reduce package size
if '${WITH_STRIP}' == 'ON':
    command = 'find ${PADDLE_BINARY_DIR}/python/paddle -name "*.so" | xargs -i strip {}'
    if os.system(command) != 0:
        raise Exception("strip *.so failed, command: %s" % command)

def install_cpp_dist_and_build_test(paddle_install_dir, paddle_lib_test_dir):
    """install cpp distribution and build test target

    TODO(huangjiyi): 
    1. This function will be moved when seperating C++ distribution
    installation from python package installation.
    2. Reduce the header and library files to be installed.
    """
    if '${CMAKE_BUILD_TYPE}' != 'Release':
        return
    os.makedirs(paddle_install_dir, exist_ok=True)
    # install C++ header files
    for header in headers: 
        install_dir = get_header_install_dir(header)
        install_dir = os.path.join(
            paddle_install_dir, 'include', os.path.dirname(install_dir)
        )
        os.makedirs(install_dir, exist_ok=True)
        shutil.copy(header, install_dir)

    # install C++ shared libraries
    lib_install_dir = os.path.join(paddle_install_dir, 'lib')
    os.makedirs(lib_install_dir, exist_ok=True)
    # install libpaddle.ext
    paddle_libs = glob.glob('${PADDLE_BINARY_DIR}/paddle/fluid/pybind/${FLUID_CORE_NAME}.*')
    for lib in paddle_libs:
        shutil.copy(lib, lib_install_dir)
    # install dependent libraries
    libs_path = package_dir['paddle.libs']
    for lib in package_data['paddle.libs']:
        lib_path = os.path.join(libs_path, lib)
        shutil.copy(lib_path, lib_install_dir)
    
    # build test target
    cmake_args = ["cmake", paddle_lib_test_dir, "-B", paddle_lib_test_dir]
    if os.getenv("GENERATOR") == "Ninja":
        cmake_args.append("-GNinja")
    subprocess.check_call(cmake_args)
    subprocess.check_call(["cmake", "--build", paddle_lib_test_dir])


# install cpp distribution
if '${WITH_CPP_DIST}' == 'ON':
    paddle_install_dir = '${PADDLE_INSTALL_DIR}'
    paddle_lib_test_dir = '${PADDLE_LIB_TEST_DIR}'
    install_cpp_dist_and_build_test(paddle_install_dir, paddle_lib_test_dir)
    

with redirect_stdout():
    setup(name='${PACKAGE_NAME}',
        version='${PADDLE_VERSION}',
        description='Parallel Distributed Deep Learning',
        long_description=long_description,
        long_description_content_type="text/markdown",
        author_email="Paddle-better@baidu.com",
        maintainer="PaddlePaddle",
        maintainer_email="Paddle-better@baidu.com",
        project_urls = {
            'Homepage': 'https://www.paddlepaddle.org.cn/',
            'Downloads': 'https://github.com/paddlepaddle/paddle'
        },
        license='Apache Software License',
        packages=packages,
        install_requires=setup_requires,
        ext_modules=ext_modules,
        package_data=package_data,
        package_dir=package_dir,
        scripts=paddle_bins,
        distclass=BinaryDistribution,
        headers=headers,
        cmdclass={
            'install_headers': InstallHeaders,
            'install': InstallCommand,
            'egg_info': EggInfo,
        },
        entry_points={
            'console_scripts': [
                'fleetrun = paddle.distributed.launch.main:launch'
            ]
        },
        classifiers=[
            'Development Status :: 5 - Production/Stable',
            'Operating System :: OS Independent',
            'Intended Audience :: Developers',
            'Intended Audience :: Education',
            'Intended Audience :: Science/Research',
            'License :: OSI Approved :: Apache Software License',
            'Programming Language :: C++',
            'Programming Language :: Python :: 3.7',
            'Programming Language :: Python :: 3.8',
            'Programming Language :: Python :: 3.9',
            'Programming Language :: Python :: 3.10',
        ],
    )

# As there are a lot of files in purelib which causes many logs,
# we don't print them on the screen, and you can open `setup.py.log`
# for the full logs.
if os.path.exists('${SETUP_LOG_FILE}'):
    os.system('grep -v "purelib" ${SETUP_LOG_FILE}')