# Owner(s): ["module: tests"]

import sys

import torch
from torch.testing._internal.common_utils import NoTest, run_tests, TestCase


if not torch.accelerator.is_available():
    print("No available accelerator detected, skipping tests", file=sys.stderr)
    TestCase = NoTest  # noqa: F811
    sys.exit()


class TestAccelerator(TestCase):
    def test_current_accelerator(self):
        self.assertTrue(torch.accelerator.is_available())
        accelerators = ["cuda", "xpu", "mps", "npu"]
        for accelerator in accelerators:
            if torch.get_device_module(accelerator).is_available():
                self.assertEqual(
                    torch.accelerator.current_accelerator().type, accelerator
                )
                self.assertIsNone(torch.accelerator.current_accelerator().index)
                with self.assertRaisesRegex(
                    ValueError, "doesn't match the current accelerator"
                ):
                    torch.accelerator.set_device_index("cpu")

    def test_generic_stream_behavior(self):
        s1 = torch.Stream()
        s2 = torch.Stream()
        torch.accelerator.set_stream(s1)
        self.assertEqual(torch.accelerator.current_stream(), s1)
        event = torch.Event()
        a = torch.randn(1000)
        b = torch.randn(1000)
        c = a + b
        torch.accelerator.set_stream(s2)
        self.assertEqual(torch.accelerator.current_stream(), s2)
        a_acc = a.to(torch.accelerator.current_accelerator(), non_blocking=True)
        b_acc = b.to(torch.accelerator.current_accelerator(), non_blocking=True)
        torch.accelerator.set_stream(s1)
        self.assertEqual(torch.accelerator.current_stream(), s1)
        event.record(s2)
        event.synchronize()
        c_acc = a_acc + b_acc
        event.record(s2)
        torch.accelerator.synchronize()
        self.assertTrue(event.query())
        self.assertEqual(c_acc.cpu(), c)

    def test_current_stream_query(self):
        s = torch.accelerator.current_stream()
        self.assertEqual(torch.accelerator.current_stream(s.device), s)
        self.assertEqual(torch.accelerator.current_stream(s.device.index), s)
        self.assertEqual(torch.accelerator.current_stream(str(s.device)), s)
        other_device = torch.device("cpu")
        with self.assertRaisesRegex(
            ValueError, "doesn't match the current accelerator"
        ):
            torch.accelerator.current_stream(other_device)

    def test_stream_context_manager(self):
        prev_stream = torch.accelerator.current_stream()
        with torch.Stream() as s:
            self.assertEqual(torch.accelerator.current_stream(), s)
        self.assertEqual(torch.accelerator.current_stream(), prev_stream)

    def test_pin_memory_on_non_blocking_copy(self):
        t_acc = torch.randn(100).to(torch.accelerator.current_accelerator())
        t_host = t_acc.to("cpu", non_blocking=True)
        torch.accelerator.synchronize()
        self.assertTrue(t_host.is_pinned())
        self.assertEqual(t_acc.cpu(), t_host)

    def test_event_elapsed_time(self):
        start_event = torch.Event(enable_timing=True)
        end_event = torch.Event(enable_timing=True)
        start_event.record()
        x = torch.randn(1000, 1000, device='npu')
        y = torch.randn(1000, 1000, device='npu')
        z = torch.matmul(x, y)
        end_event.record()
        torch.npu.synchronize()
        ms = start_event.elapsed_time(end_event)
        self.assertGreater(ms, 0)


if __name__ == "__main__":
    run_tests()