test_cpp_extensions_jit.py

# Owner(s): ["module: cpp-extensions"]

import os
import shutil
import sys
import unittest
import warnings
import re
import tempfile
import subprocess
import glob

import textwrap
from multiprocessing import Process

import torch.testing._internal.common_utils as common
import torch
import torch.backends.cudnn
import torch.utils.cpp_extension
from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME
from torch.testing._internal.common_utils import gradcheck, TEST_WITH_ASAN, has_breakpad


TEST_CUDA = torch.cuda.is_available() and CUDA_HOME is not None
TEST_CUDNN = False
TEST_ROCM = torch.cuda.is_available() and torch.version.hip is not None and ROCM_HOME is not None
if TEST_CUDA and torch.version.cuda is not None:  # the skip CUDNN test for ROCm
    CUDNN_HEADER_EXISTS = os.path.isfile(os.path.join(CUDA_HOME, "include/cudnn.h"))
    TEST_CUDNN = (
        TEST_CUDA and CUDNN_HEADER_EXISTS and torch.backends.cudnn.is_available()
    )
IS_WINDOWS = sys.platform == "win32"


def remove_build_path():
    if sys.platform == "win32":
        print("Not wiping extensions build folder because Windows")
        return
    default_build_root = torch.utils.cpp_extension.get_default_build_root()
    if os.path.exists(default_build_root):
        shutil.rmtree(default_build_root)


class TestCppExtensionJIT(common.TestCase):
    """Tests just-in-time cpp extensions.
    Don't confuse this with the PyTorch JIT (aka TorchScript).
    """

    def setUp(self):
        super().setUp()
        # cpp extensions use relative paths. Those paths are relative to
        # this file, so we'll change the working directory temporarily
        self.old_working_dir = os.getcwd()
        os.chdir(os.path.dirname(os.path.abspath(__file__)))

    def tearDown(self):
        super().tearDown()
        # return the working directory (see setUp)
        os.chdir(self.old_working_dir)

    @classmethod
    def setUpClass(cls):
        remove_build_path()

    @classmethod
    def tearDownClass(cls):
        remove_build_path()

    def test_jit_compile_extension(self):
        module = torch.utils.cpp_extension.load(
            name="jit_extension",
            sources=[
                "cpp_extensions/jit_extension.cpp",
                "cpp_extensions/jit_extension2.cpp",
            ],
            extra_include_paths=["cpp_extensions"],
            extra_cflags=["-g"],
            verbose=True,
        )
        x = torch.randn(4, 4)
        y = torch.randn(4, 4)

        z = module.tanh_add(x, y)
        self.assertEqual(z, x.tanh() + y.tanh())

        # Checking we can call a method defined not in the main C++ file.
        z = module.exp_add(x, y)
        self.assertEqual(z, x.exp() + y.exp())

        # Checking we can use this JIT-compiled class.
        doubler = module.Doubler(2, 2)
        self.assertIsNone(doubler.get().grad)
        self.assertEqual(doubler.get().sum(), 4)
        self.assertEqual(doubler.forward().sum(), 8)

    @unittest.skipIf(not (TEST_CUDA or TEST_ROCM), "CUDA not found")
    def test_jit_cuda_extension(self):
        # NOTE: The name of the extension must equal the name of the module.
        module = torch.utils.cpp_extension.load(
            name="torch_test_cuda_extension",
            sources=[
                "cpp_extensions/cuda_extension.cpp",
                "cpp_extensions/cuda_extension.cu",
            ],
            extra_cuda_cflags=["-O2"],
            verbose=True,
            keep_intermediates=False,
        )

        x = torch.zeros(100, device="cuda", dtype=torch.float32)
        y = torch.zeros(100, device="cuda", dtype=torch.float32)

        z = module.sigmoid_add(x, y).cpu()

        # 2 * sigmoid(0) = 2 * 0.5 = 1
        self.assertEqual(z, torch.ones_like(z))

    def _run_jit_cuda_archflags(self, flags, expected):
        # Compile an extension with given `flags`
        def _check_cuobjdump_output(expected_values, is_ptx=False):
            elf_or_ptx = '--list-ptx' if is_ptx else '--list-elf'
            lib_ext = '.pyd' if IS_WINDOWS else '.so'
            # Note, .extension name may include _v1, _v2, so first find exact name
            ext_filename = glob.glob(os.path.join(temp_dir,
                                                  'cudaext_archflag*' + lib_ext))[0]
            command = ['cuobjdump', elf_or_ptx, ext_filename]
            p = subprocess.Popen(command,
                                 stdout=subprocess.PIPE,
                                 stderr=subprocess.PIPE)
            output, err = p.communicate()
            output = output.decode("ascii")
            err = err.decode("ascii")

            if not p.returncode == 0 or not err == '':
                raise AssertionError("Flags: {}\nReturncode: {}\nStderr: {}\n"
                                     "Output: {} ".format(flags, p.returncode,
                                                          err, output))

            actual_arches = sorted(re.findall(r'sm_\d\d', output))
            expected_arches = sorted(['sm_' + xx for xx in expected_values])
            self.assertEqual(actual_arches, expected_arches,
                             msg="Flags: {},  Actual: {},  Expected: {}\n"
                                 "Stderr: {}\nOutput: {}".format(
                                     flags, actual_arches, expected_arches,
                                     err, output))

        temp_dir = tempfile.mkdtemp()
        old_envvar = os.environ.get('TORCH_CUDA_ARCH_LIST', None)
        try:
            os.environ['TORCH_CUDA_ARCH_LIST'] = flags
            torch.utils.cpp_extension.load(
                name="cudaext_archflags",
                sources=[
                    "cpp_extensions/cuda_extension.cpp",
                    "cpp_extensions/cuda_extension.cu",
                ],
                extra_cuda_cflags=["-O2"],
                verbose=True,
                build_directory=temp_dir,
            )

            # Expected output for --list-elf:
            #   ELF file    1: cudaext_archflags.1.sm_61.cubin
            #   ELF file    2: cudaext_archflags.2.sm_52.cubin
            _check_cuobjdump_output(expected[0])
            if expected[1] is not None:
                # Expected output for --list-ptx:
                #   PTX file    1: cudaext_archflags.1.sm_61.ptx
                _check_cuobjdump_output(expected[1], is_ptx=True)
        finally:
            if IS_WINDOWS:
                print("Not wiping extensions build folder because Windows")
            else:
                shutil.rmtree(temp_dir)

            if old_envvar is None:
                os.environ.pop('TORCH_CUDA_ARCH_LIST')
            else:
                os.environ['TORCH_CUDA_ARCH_LIST'] = old_envvar

    @unittest.skipIf(not TEST_CUDA, "CUDA not found")
    @unittest.skipIf(TEST_ROCM, "disabled on rocm")
    def test_jit_cuda_archflags(self):
        # Test a number of combinations:
        #   - the default for the machine we're testing on
        #   - Separators, can be ';' (most common) or ' '
        #   - Architecture names
        #   - With/without '+PTX'

        n = torch.cuda.device_count()
        capabilities = {torch.cuda.get_device_capability(i) for i in range(n)}
        # expected values is length-2 tuple: (list of ELF, list of PTX)
        # note: there should not be more than one PTX value
        archflags = {
            '': (['{}{}'.format(capability[0], capability[1]) for capability in capabilities], None),
            "Maxwell+Tegra;6.1": (['53', '61'], None),
            "Pascal 3.5": (['35', '60', '61'], None),
            "Volta": (['70'], ['70']),
        }
        if int(torch.version.cuda.split('.')[0]) >= 10:
            # CUDA 9 only supports compute capability <= 7.2
            archflags["7.5+PTX"] = (['75'], ['75'])
            archflags["5.0;6.0+PTX;7.0;7.5"] = (['50', '60', '70', '75'], ['60'])

        for flags, expected in archflags.items():
            self._run_jit_cuda_archflags(flags, expected)

    @unittest.skipIf(not TEST_CUDNN, "CuDNN not found")
    def test_jit_cudnn_extension(self):
        # implementation of CuDNN ReLU
        if IS_WINDOWS:
            extra_ldflags = ["cudnn.lib"]
        else:
            extra_ldflags = ["-lcudnn"]
        module = torch.utils.cpp_extension.load(
            name="torch_test_cudnn_extension",
            sources=["cpp_extensions/cudnn_extension.cpp"],
            extra_ldflags=extra_ldflags,
            verbose=True,
            with_cuda=True,
        )

        x = torch.randn(100, device="cuda", dtype=torch.float32)
        y = torch.zeros(100, device="cuda", dtype=torch.float32)
        module.cudnn_relu(x, y)  # y=relu(x)
        self.assertEqual(torch.nn.functional.relu(x), y)
        with self.assertRaisesRegex(RuntimeError, "same size"):
            y_incorrect = torch.zeros(20, device="cuda", dtype=torch.float32)
            module.cudnn_relu(x, y_incorrect)

    def test_inline_jit_compile_extension_with_functions_as_list(self):
        cpp_source = """
        torch::Tensor tanh_add(torch::Tensor x, torch::Tensor y) {
          return x.tanh() + y.tanh();
        }
        """

        module = torch.utils.cpp_extension.load_inline(
            name="inline_jit_extension_with_functions_list",
            cpp_sources=cpp_source,
            functions="tanh_add",
            verbose=True,
        )

        self.assertEqual(module.tanh_add.__doc__.split("\n")[2], "tanh_add")

        x = torch.randn(4, 4)
        y = torch.randn(4, 4)

        z = module.tanh_add(x, y)
        self.assertEqual(z, x.tanh() + y.tanh())

    def test_inline_jit_compile_extension_with_functions_as_dict(self):
        cpp_source = """
        torch::Tensor tanh_add(torch::Tensor x, torch::Tensor y) {
          return x.tanh() + y.tanh();
        }
        """

        module = torch.utils.cpp_extension.load_inline(
            name="inline_jit_extension_with_functions_dict",
            cpp_sources=cpp_source,
            functions={"tanh_add": "Tanh and then sum :D"},
            verbose=True,
        )

        self.assertEqual(module.tanh_add.__doc__.split("\n")[2], "Tanh and then sum :D")

    def test_inline_jit_compile_extension_multiple_sources_and_no_functions(self):
        cpp_source1 = """
        torch::Tensor sin_add(torch::Tensor x, torch::Tensor y) {
          return x.sin() + y.sin();
        }
        """

        cpp_source2 = """
        #include <torch/extension.h>
        torch::Tensor sin_add(torch::Tensor x, torch::Tensor y);
        PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
          m.def("sin_add", &sin_add, "sin(x) + sin(y)");
        }
        """

        module = torch.utils.cpp_extension.load_inline(
            name="inline_jit_extension",
            cpp_sources=[cpp_source1, cpp_source2],
            verbose=True,
        )

        x = torch.randn(4, 4)
        y = torch.randn(4, 4)

        z = module.sin_add(x, y)
        self.assertEqual(z, x.sin() + y.sin())

    @unittest.skip("Temporarily disabled")
    @unittest.skipIf(not (TEST_CUDA or TEST_ROCM), "CUDA not found")
    def test_inline_jit_compile_extension_cuda(self):
        cuda_source = """
        __global__ void cos_add_kernel(
            const float* __restrict__ x,
            const float* __restrict__ y,
            float* __restrict__ output,
            const int size) {
          const auto index = blockIdx.x * blockDim.x + threadIdx.x;
          if (index < size) {
            output[index] = __cosf(x[index]) + __cosf(y[index]);
          }
        }

        torch::Tensor cos_add(torch::Tensor x, torch::Tensor y) {
          auto output = torch::zeros_like(x);
          const int threads = 1024;
          const int blocks = (output.numel() + threads - 1) / threads;
          cos_add_kernel<<<blocks, threads>>>(x.data<float>(), y.data<float>(), output.data<float>(), output.numel());
          return output;
        }
        """

        # Here, the C++ source need only declare the function signature.
        cpp_source = "torch::Tensor cos_add(torch::Tensor x, torch::Tensor y);"

        module = torch.utils.cpp_extension.load_inline(
            name="inline_jit_extension_cuda",
            cpp_sources=cpp_source,
            cuda_sources=cuda_source,
            functions=["cos_add"],
            verbose=True,
        )

        self.assertEqual(module.cos_add.__doc__.split("\n")[2], "cos_add")

        x = torch.randn(4, 4, device="cuda", dtype=torch.float32)
        y = torch.randn(4, 4, device="cuda", dtype=torch.float32)

        z = module.cos_add(x, y)
        self.assertEqual(z, x.cos() + y.cos())

    @unittest.skip("Temporarily disabled")
    @unittest.skipIf(not (TEST_CUDA or TEST_ROCM), "CUDA not found")
    def test_inline_jit_compile_custom_op_cuda(self):
        cuda_source = """
        __global__ void cos_add_kernel(
            const float* __restrict__ x,
            const float* __restrict__ y,
            float* __restrict__ output,
            const int size) {
          const auto index = blockIdx.x * blockDim.x + threadIdx.x;
          if (index < size) {
            output[index] = __cosf(x[index]) + __cosf(y[index]);
          }
        }

        torch::Tensor cos_add(torch::Tensor x, torch::Tensor y) {
          auto output = torch::zeros_like(x);
          const int threads = 1024;
          const int blocks = (output.numel() + threads - 1) / threads;
          cos_add_kernel<<<blocks, threads>>>(x.data_ptr<float>(), y.data_ptr<float>(), output.data_ptr<float>(), output.numel());
          return output;
        }
        """

        # Here, the C++ source need only declare the function signature.
        cpp_source = """
           #include <torch/library.h>
           torch::Tensor cos_add(torch::Tensor x, torch::Tensor y);

           TORCH_LIBRARY(inline_jit_extension_custom_op_cuda, m) {
             m.def("cos_add", cos_add);
           }
        """

        torch.utils.cpp_extension.load_inline(
            name="inline_jit_extension_custom_op_cuda",
            cpp_sources=cpp_source,
            cuda_sources=cuda_source,
            verbose=True,
            is_python_module=False,
        )

        x = torch.randn(4, 4, device="cuda", dtype=torch.float32)
        y = torch.randn(4, 4, device="cuda", dtype=torch.float32)

        z = torch.ops.inline_jit_extension_custom_op_cuda.cos_add(x, y)
        self.assertEqual(z, x.cos() + y.cos())

    def test_inline_jit_compile_extension_throws_when_functions_is_bad(self):
        with self.assertRaises(ValueError):
            torch.utils.cpp_extension.load_inline(
                name="invalid_jit_extension", cpp_sources="", functions=5
            )

    def test_lenient_flag_handling_in_jit_extensions(self):
        cpp_source = """
        torch::Tensor tanh_add(torch::Tensor x, torch::Tensor y) {
          return x.tanh() + y.tanh();
        }
        """

        module = torch.utils.cpp_extension.load_inline(
            name="lenient_flag_handling_extension",
            cpp_sources=cpp_source,
            functions="tanh_add",
            extra_cflags=["-g\n\n", "-O0 -Wall"],
            extra_include_paths=["       cpp_extensions\n"],
            verbose=True,
        )

        x = torch.zeros(100, dtype=torch.float32)
        y = torch.zeros(100, dtype=torch.float32)
        z = module.tanh_add(x, y).cpu()
        self.assertEqual(z, x.tanh() + y.tanh())

    @unittest.skip("Temporarily disabled")
    @unittest.skipIf(not (TEST_CUDA or TEST_ROCM), "CUDA not found")
    def test_half_support(self):
        """
        Checks for an issue with operator< ambiguity for half when certain
        THC headers are included.

        See https://github.com/pytorch/pytorch/pull/10301#issuecomment-416773333
        for the corresponding issue.
        """
        cuda_source = """
        template<typename T, typename U>
        __global__ void half_test_kernel(const T* input, U* output) {
            if (input[0] < input[1] || input[0] >= input[1]) {
                output[0] = 123;
            }
        }

        torch::Tensor half_test(torch::Tensor input) {
            auto output = torch::empty(1, input.options().dtype(torch::kFloat));
            AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "half_test", [&] {
                half_test_kernel<scalar_t><<<1, 1>>>(
                    input.data<scalar_t>(),
                    output.data<float>());
            });
            return output;
        }
        """

        module = torch.utils.cpp_extension.load_inline(
            name="half_test_extension",
            cpp_sources="torch::Tensor half_test(torch::Tensor input);",
            cuda_sources=cuda_source,
            functions=["half_test"],
            verbose=True,
        )

        x = torch.randn(3, device="cuda", dtype=torch.half)
        result = module.half_test(x)
        self.assertEqual(result[0], 123)

    def test_reload_jit_extension(self):
        def compile(code):
            return torch.utils.cpp_extension.load_inline(
                name="reloaded_jit_extension",
                cpp_sources=code,
                functions="f",
                verbose=True,
            )

        module = compile("int f() { return 123; }")
        self.assertEqual(module.f(), 123)

        module = compile("int f() { return 456; }")
        self.assertEqual(module.f(), 456)
        module = compile("int f() { return 456; }")
        self.assertEqual(module.f(), 456)

        module = compile("int f() { return 789; }")
        self.assertEqual(module.f(), 789)

    def test_cpp_frontend_module_has_same_output_as_python(self, dtype=torch.double):
        extension = torch.utils.cpp_extension.load(
            name="cpp_frontend_extension",
            sources="cpp_extensions/cpp_frontend_extension.cpp",
            verbose=True,
        )

        input = torch.randn(2, 5, dtype=dtype)
        cpp_linear = extension.Net(5, 2)
        cpp_linear.to(dtype)
        python_linear = torch.nn.Linear(5, 2).to(dtype)

        # First make sure they have the same parameters
        cpp_parameters = dict(cpp_linear.named_parameters())
        with torch.no_grad():
            python_linear.weight.copy_(cpp_parameters["fc.weight"])
            python_linear.bias.copy_(cpp_parameters["fc.bias"])

        cpp_output = cpp_linear.forward(input)
        python_output = python_linear(input)
        self.assertEqual(cpp_output, python_output)

        cpp_output.sum().backward()
        python_output.sum().backward()

        for p in cpp_linear.parameters():
            self.assertFalse(p.grad is None)

        self.assertEqual(cpp_parameters["fc.weight"].grad, python_linear.weight.grad)
        self.assertEqual(cpp_parameters["fc.bias"].grad, python_linear.bias.grad)

    def test_cpp_frontend_module_python_inter_op(self):
        extension = torch.utils.cpp_extension.load(
            name="cpp_frontend_extension",
            sources="cpp_extensions/cpp_frontend_extension.cpp",
            verbose=True,
        )

        # Create a torch.nn.Module which uses the C++ module as a submodule.
        class M(torch.nn.Module):
            def __init__(self):
                super(M, self).__init__()
                self.x = torch.nn.Parameter(torch.tensor(1.0))
                self.net = extension.Net(3, 5)

            def forward(self, input):
                return self.net.forward(input) + self.x

        net = extension.Net(5, 2)
        net.double()
        net.to(torch.get_default_dtype())
        self.assertEqual(str(net), "Net")

        # Further embed the torch.nn.Module into a Sequential, and also add the
        # C++ module as an element of the Sequential.
        sequential = torch.nn.Sequential(M(), torch.nn.Tanh(), net, torch.nn.Sigmoid())

        input = torch.randn(2, 3)
        # Try calling the module!
        output = sequential.forward(input)
        # The call operator is bound to forward too.
        self.assertEqual(output, sequential(input))
        self.assertEqual(list(output.shape), [2, 2])

        # Do changes on the module hierarchy.
        old_dtype = torch.get_default_dtype()
        sequential.to(torch.float64)
        sequential.to(torch.float32)
        sequential.to(old_dtype)
        self.assertEqual(sequential[2].parameters()[0].dtype, old_dtype)

        # Make sure we can access these methods recursively.
        self.assertEqual(len(list(sequential.parameters())), len(net.parameters()) * 2 + 1)
        self.assertEqual(len(list(sequential.named_parameters())), len(net.named_parameters()) * 2 + 1)
        self.assertEqual(len(list(sequential.buffers())), len(net.buffers()) * 2)
        self.assertEqual(len(list(sequential.modules())), 8)

        # Test clone()
        net2 = net.clone()
        self.assertEqual(len(net.parameters()), len(net2.parameters()))
        self.assertEqual(len(net.buffers()), len(net2.buffers()))
        self.assertEqual(len(net.modules()), len(net2.modules()))

        # Try differentiating through the whole module.
        for parameter in net.parameters():
            self.assertIsNone(parameter.grad)
        output.sum().backward()
        for parameter in net.parameters():
            self.assertFalse(parameter.grad is None)
            self.assertGreater(parameter.grad.sum(), 0)

        # Try calling zero_grad()
        net.zero_grad()
        for p in net.parameters():
            self.assertEqual(p.grad, torch.zeros_like(p))

        # Test train(), eval(), training (a property)
        self.assertTrue(net.training)
        net.eval()
        self.assertFalse(net.training)
        net.train()
        self.assertTrue(net.training)
        net.eval()

        # Try calling the additional methods we registered.
        biased_input = torch.randn(4, 5)
        output_before = net.forward(biased_input)
        bias = net.get_bias().clone()
        self.assertEqual(list(bias.shape), [2])
        net.set_bias(bias + 1)
        self.assertEqual(net.get_bias(), bias + 1)
        output_after = net.forward(biased_input)

        self.assertNotEqual(output_before, output_after)

        # Try accessing parameters
        self.assertEqual(len(net.parameters()), 2)
        np = net.named_parameters()
        self.assertEqual(len(np), 2)
        self.assertIn("fc.weight", np)
        self.assertIn("fc.bias", np)

        self.assertEqual(len(net.buffers()), 1)
        nb = net.named_buffers()
        self.assertEqual(len(nb), 1)
        self.assertIn("buf", nb)
        self.assertEqual(nb[0][1], torch.eye(5))

    def test_cpp_frontend_module_has_up_to_date_attributes(self):
        extension = torch.utils.cpp_extension.load(
            name="cpp_frontend_extension",
            sources="cpp_extensions/cpp_frontend_extension.cpp",
            verbose=True,
        )

        net = extension.Net(5, 2)

        self.assertEqual(len(net._parameters), 0)
        net.add_new_parameter("foo", torch.eye(5))
        self.assertEqual(len(net._parameters), 1)

        self.assertEqual(len(net._buffers), 1)
        net.add_new_buffer("bar", torch.eye(5))
        self.assertEqual(len(net._buffers), 2)

        self.assertEqual(len(net._modules), 1)
        net.add_new_submodule("fc2")
        self.assertEqual(len(net._modules), 2)

    @unittest.skipIf(not (TEST_CUDA or TEST_ROCM), "CUDA not found")
    def test_cpp_frontend_module_python_inter_op_with_cuda(self):
        extension = torch.utils.cpp_extension.load(
            name="cpp_frontend_extension",
            sources="cpp_extensions/cpp_frontend_extension.cpp",
            verbose=True,
        )

        net = extension.Net(5, 2)
        for p in net.parameters():
            self.assertTrue(p.device.type == "cpu")
        cpu_parameters = [p.clone() for p in net.parameters()]

        device = torch.device("cuda", 0)
        net.to(device)

        for i, p in enumerate(net.parameters()):
            self.assertTrue(p.device.type == "cuda")
            self.assertTrue(p.device.index == 0)
            self.assertEqual(cpu_parameters[i], p)

        net.cpu()
        net.add_new_parameter("a", torch.eye(5))
        net.add_new_parameter("b", torch.eye(5))
        net.add_new_buffer("c", torch.eye(5))
        net.add_new_buffer("d", torch.eye(5))
        net.add_new_submodule("fc2")
        net.add_new_submodule("fc3")

        for p in net.parameters():
            self.assertTrue(p.device.type == "cpu")

        net.cuda()

        for p in net.parameters():
            self.assertTrue(p.device.type == "cuda")

    def test_returns_shared_library_path_when_is_python_module_is_true(self):
        source = """
        #include <torch/script.h>
        torch::Tensor func(torch::Tensor x) { return x; }
        static torch::RegisterOperators r("test::func", &func);
        """
        torch.utils.cpp_extension.load_inline(
            name="is_python_module",
            cpp_sources=source,
            functions="func",
            verbose=True,
            is_python_module=False,
        )
        self.assertEqual(torch.ops.test.func(torch.eye(5)), torch.eye(5))

    def test_set_default_type_also_changes_aten_default_type(self):
        module = torch.utils.cpp_extension.load_inline(
            name="test_set_default_type",
            cpp_sources="torch::Tensor get() { return torch::empty({}); }",
            functions="get",
            verbose=True,
        )

        initial_default = torch.get_default_dtype()
        try:
            self.assertEqual(module.get().dtype, initial_default)
            torch.set_default_dtype(torch.float64)
            self.assertEqual(module.get().dtype, torch.float64)
            torch.set_default_dtype(torch.float32)
            self.assertEqual(module.get().dtype, torch.float32)
            torch.set_default_dtype(torch.float16)
            self.assertEqual(module.get().dtype, torch.float16)
        finally:
            torch.set_default_dtype(initial_default)

    def test_compilation_error_formatting(self):
        # Test that the missing-semicolon error message has linebreaks in it.
        # This'll fail if the message has been munged into a single line.
        # It's hard to write anything more specific as every compiler has it's own
        # error formatting.
        with self.assertRaises(RuntimeError) as e:
            torch.utils.cpp_extension.load_inline(
                name="test_compilation_error_formatting",
                cpp_sources="int main() { return 0 }")
        pattern = r'.*(\\n|\\r).*'
        self.assertNotRegex(str(e), pattern)

    def test_warning(self):
        # Note: the module created from this source will include the py::key_error
        # symbol. But because of visibility and the fact that it lives in a
        # different compilation unit than pybind, this trips up ubsan even though
        # it is fine. "ubsan.supp" thus needs to contain "vptr:warn_mod.so".
        source = '''
        // error_type:
        // 0: no error
        // 1: torch::TypeError
        // 2: python_error()
        // 3: py::error_already_set
        at::Tensor foo(at::Tensor x, int error_type) {
            std::ostringstream err_stream;
            err_stream << "Error with "  << x.type();

            TORCH_WARN(err_stream.str());
            if(error_type == 1) {
                throw torch::TypeError(err_stream.str().c_str());
            }
            if(error_type == 2) {
                PyObject* obj = PyTuple_New(-1);
                TORCH_CHECK(!obj);
                // Pretend it was caught in a different thread and restored here
                auto e = python_error();
                e.persist();
                e.restore();
                throw e;
            }
            if(error_type == 3) {
                throw py::key_error(err_stream.str());
            }
            return x.cos();
        }
        '''

        # Ensure double type for hard-coded c name below
        t = torch.rand(2).double()
        cpp_tensor_name = r"CPUDoubleType"

        # Without error handling, the warnings cannot be catched
        warn_mod = torch.utils.cpp_extension.load_inline(name='warn_mod',
                                                         cpp_sources=[source],
                                                         functions=['foo'],
                                                         with_pytorch_error_handling=False)

        with warnings.catch_warnings(record=True) as w:
            warn_mod.foo(t, 0)
            self.assertEqual(len(w), 0)

            with self.assertRaisesRegex(TypeError, t.type()):
                warn_mod.foo(t, 1)
            self.assertEqual(len(w), 0)

            with self.assertRaisesRegex(SystemError, "bad argument to internal function"):
                warn_mod.foo(t, 2)
            self.assertEqual(len(w), 0)

            with self.assertRaisesRegex(KeyError, cpp_tensor_name):
                warn_mod.foo(t, 3)
            self.assertEqual(len(w), 0)


        warn_mod = torch.utils.cpp_extension.load_inline(name='warn_mod',
                                                         cpp_sources=[source],
                                                         functions=['foo'],
                                                         with_pytorch_error_handling=True)


        with warnings.catch_warnings(record=True) as w:
            # Catched with no error should be detected
            warn_mod.foo(t, 0)
            self.assertEqual(len(w), 1)

            # Catched with cpp error should also be detected
            with self.assertRaisesRegex(TypeError, t.type()):
                warn_mod.foo(t, 1)
            self.assertEqual(len(w), 2)

            # Catched with python error should also be detected
            with self.assertRaisesRegex(SystemError, "bad argument to internal function"):
                warn_mod.foo(t, 2)
            self.assertEqual(len(w), 3)

            # Catched with pybind error should also be detected
            # Note that there is no type name translation for pybind errors
            with self.assertRaisesRegex(KeyError, cpp_tensor_name):
                warn_mod.foo(t, 3)
            self.assertEqual(len(w), 4)

        # Make sure raising warnings are handled properly
        with warnings.catch_warnings(record=True) as w:
            warnings.simplefilter("error")

            # No error, the warning should raise
            with self.assertRaisesRegex(UserWarning, t.type()):
                warn_mod.foo(t, 0)
            self.assertEqual(len(w), 0)

            # Another error happened, the warning is ignored
            with self.assertRaisesRegex(TypeError, t.type()):
                warn_mod.foo(t, 1)
            self.assertEqual(len(w), 0)

    def test_autograd_from_cpp(self):
        source = '''
        void run_back(at::Tensor x) {
            x.backward({});
        }

        void run_back_no_gil(at::Tensor x) {
            pybind11::gil_scoped_release no_gil;
            x.backward({});
        }
        '''

        class MyFn(torch.autograd.Function):
            @staticmethod
            def forward(ctx, x):
                return x.clone()

            @staticmethod
            def backward(ctx, gx):
                return gx

        test_backward_deadlock = torch.utils.cpp_extension.load_inline(name='test_backward_deadlock',
                                                                       cpp_sources=[source],
                                                                       functions=['run_back', 'run_back_no_gil'],)

        # This used to deadlock
        inp = torch.rand(20, requires_grad=True)
        loss = MyFn.apply(inp).sum()
        with self.assertRaisesRegex(RuntimeError, "The autograd engine was called while holding the GIL."):
            test_backward_deadlock.run_back(loss)

        inp = torch.rand(20, requires_grad=True)
        loss = MyFn.apply(inp).sum()
        test_backward_deadlock.run_back_no_gil(loss)

    def test_custom_compound_op_autograd(self):
        # Test that a custom compound op (i.e. a custom op that just calls other aten ops)
        # correctly returns gradients of those other ops

        source = """
        #include <torch/library.h>
        torch::Tensor my_add(torch::Tensor x, torch::Tensor y) {
          return x + y;
        }
        TORCH_LIBRARY(my, m) {
            m.def("add", &my_add);
        }
        """

        torch.utils.cpp_extension.load_inline(
            name="is_python_module",
            cpp_sources=source,
            verbose=True,
            is_python_module=False,
        )

        a = torch.randn(5, 5, requires_grad=True)
        b = torch.randn(5, 5, requires_grad=True)

        gradcheck(torch.ops.my.add, [a, b], eps=1e-2)

    @staticmethod
    def _crash_handler_test_process(stderr_file, destination):
        # Code to enable dumps and trigger a segfault
        if sys.platform == "win32":
            destination = destination.replace("\\", "\\\\")
            csrc = textwrap.dedent(f"""
            #include <torch/torch.h>
            #include <locale>
            #include <iostream>
            #include <codecvt>
            #include <string>

            int fail() {{
                std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
                std::string narrow("{destination}");
                std::wstring wide = converter.from_bytes(narrow);
                torch::crash_handler::enable_minidumps(wide.c_str());

                volatile int* bad = nullptr;
                return *bad;
            }}
            """)
        else:
            csrc = textwrap.dedent(f"""
            #include <torch/torch.h>

            int fail() {{
                torch::crash_handler::enable_minidumps("{destination}");

                volatile int* bad = nullptr;
                return *bad;
            }}
            """)

        # Some special stuff to overwrite stderr for a C++ extension
        # Copied from: https://stackoverflow.com/questions/8804893/redirect-stdout-from-python-for-c-calls
        sys.stdout.flush()
        newstdout = os.dup(2)
        devnull = os.open(stderr_file, os.O_WRONLY)
        os.dup2(devnull, 2)
        os.close(devnull)
        sys.stdout = os.fdopen(newstdout, 'w')

        module = torch.utils.cpp_extension.load_inline(
            name="segfault",
            cpp_sources=csrc,
            functions=["fail"],
        )
        module.fail()

    @unittest.skipIf(TEST_WITH_ASAN, "ASAN disables the crash handler's signal handler")
    @unittest.skipIf(not has_breakpad(), "Built without breakpad")
    @unittest.skipIf(os.environ.get("TEST_CONFIG") == "force_on_cpu", "fails on force_on_cpu config, tracked w/ #65253")
    def test_crash_handler(self):
        with tempfile.TemporaryDirectory() as temp_dir, tempfile.NamedTemporaryFile(delete=not sys.platform == "win32") as stderr:
            # Use multiprocessing to spin up a separate process to make catching
            # the segfault easier
            p = Process(target=self._crash_handler_test_process, args=(stderr.name, temp_dir))
            p.start()
            p.join()

            with open(stderr.name) as f:
                result = f.read().strip()

            # Check that the signal handler was called
            self.assertTrue(result.startswith(f"Wrote minidump to {temp_dir}"))

            with open(result.replace("Wrote minidump to ", ""), "rb") as dump_file:
                dump_bytes = dump_file.read()

                # Check that the file has the correct magic number
                self.assertEqual(b"MDMP", dump_bytes[0:4])



if __name__ == "__main__":
    common.run_tests()