Creates spectral ops test suite (pytorch#42157)

Summary:
In preparation for creating the new torch.fft namespace and NumPy-like fft functions, as well as supporting our goal of refactoring and reducing the size of test_torch.py, this PR creates a test suite for our spectral ops.

The existing spectral op tests from test_torch.py and test_cuda.py are moved to test_spectral_ops.py and updated to run under the device generic test framework.

Pull Request resolved: pytorch#42157

Reviewed By: albanD

Differential Revision: D22811096

Pulled By: mruberry

fbshipit-source-id: e5c50f0016ea6bb8b093cd6df2dbcef6db9bb6b6

test/run_test.py

@@ -54,6 +54,7 @@
     'test_vulkan',
     'test_quantization',
     'test_sparse',
+    'test_spectral_ops',
     'test_serialization',
     'test_show_pickle',
     'test_torch',

test/test_cuda.py

@@ -6,7 +6,6 @@
 from itertools import repeat, chain
 import os
 import gc
-from contextlib import contextmanager
 import threading
 import queue
 import pickle
@@ -1366,77 +1365,6 @@ def test_prod_large(self):
         x = torch.ones(240000, device='cuda', dtype=torch.float32)
         self.assertEqual(x.prod(), 1)
 
-    @skipIfRocm
-    def test_fft_ifft_rfft_irfft(self):
-        AbstractTestCases._TestTorchMixin._test_fft_ifft_rfft_irfft(self, device=torch.device('cuda'))
-
-        @contextmanager
-        def plan_cache_max_size(n, device=None):
-            if device is None:
-                plan_cache = torch.backends.cuda.cufft_plan_cache
-            else:
-                plan_cache = torch.backends.cuda.cufft_plan_cache[device]
-            original = plan_cache.max_size
-            plan_cache.max_size = n
-            yield
-            plan_cache.max_size = original
-
-        with plan_cache_max_size(max(1, torch.backends.cuda.cufft_plan_cache.size - 10)):
-            AbstractTestCases._TestTorchMixin._test_fft_ifft_rfft_irfft(self, device=torch.device('cuda'))
-
-        with plan_cache_max_size(0):
-            AbstractTestCases._TestTorchMixin._test_fft_ifft_rfft_irfft(self, device=torch.device('cuda'))
-
-        torch.backends.cuda.cufft_plan_cache.clear()
-
-        # check that stll works after clearing cache
-        with plan_cache_max_size(10):
-            AbstractTestCases._TestTorchMixin._test_fft_ifft_rfft_irfft(self, device=torch.device('cuda'))
-
-        with self.assertRaisesRegex(RuntimeError, r"must be non-negative"):
-            torch.backends.cuda.cufft_plan_cache.max_size = -1
-
-        with self.assertRaisesRegex(RuntimeError, r"read-only property"):
-            torch.backends.cuda.cufft_plan_cache.size = -1
-
-        with self.assertRaisesRegex(RuntimeError, r"but got device with index"):
-            torch.backends.cuda.cufft_plan_cache[torch.cuda.device_count() + 10]
-
-        if TEST_MULTIGPU:
-            # Test that different GPU has different cache
-            x0 = torch.randn(2, 3, 3, device='cuda:0')
-            x1 = x0.cuda(1)
-            self.assertEqual(x0.rfft(2), x1.rfft(2))
-            # If a plan is used across different devices, the following line (or
-            # the assert above) would trigger illegal memory access. Other ways
-            # to trigger the error include
-            #   (1) setting CUDA_LAUNCH_BLOCKING=1 (pytorch/pytorch#19224) and
-            #   (2) printing a device 1 tensor.
-            x0.copy_(x1)
-
-            # Test that un-indexed `torch.backends.cuda.cufft_plan_cache` uses current device
-            with plan_cache_max_size(10, device='cuda:0'):
-                with plan_cache_max_size(11, device='cuda:1'):
-                    self.assertEqual(torch.backends.cuda.cufft_plan_cache[0].max_size, 10)
-                    self.assertEqual(torch.backends.cuda.cufft_plan_cache[1].max_size, 11)
-
-                    self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 10)  # default is cuda:0
-                    with torch.cuda.device(1):
-                        self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 11)  # default is cuda:1
-                        with torch.cuda.device(0):
-                            self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 10)  # default is cuda:0
-
-                self.assertEqual(torch.backends.cuda.cufft_plan_cache[0].max_size, 10)
-                with torch.cuda.device(1):
-                    with plan_cache_max_size(11):  # default is cuda:1
-                        self.assertEqual(torch.backends.cuda.cufft_plan_cache[0].max_size, 10)
-                        self.assertEqual(torch.backends.cuda.cufft_plan_cache[1].max_size, 11)
-
-                        self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 11)  # default is cuda:1
-                        with torch.cuda.device(0):
-                            self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 10)  # default is cuda:0
-                        self.assertEqual(torch.backends.cuda.cufft_plan_cache.max_size, 11)  # default is cuda:1
-
     def test_multinomial_ext(self):
         # Test two corner cases from older PyTorch (Issue #4858)
         freqs = torch.cuda.FloatTensor([