class TestForeach(TestCase): @property def is_cuda(self): return self.device_type == 'cuda' # note(mkozuki): It might be the case that the expected number of `cudaLaunchKernel`s # is greater than 1 once foreach functions internally separate their input `TensorList`s by # devices & dtypes into vectors of tensors. def _get_funcs(self, op, n_expected_cudaLaunchKernels): return ( ForeachFuncWrapper(op.method_variant, n_expected_cudaLaunchKernels), RegularFuncWrapper(op.ref), ForeachFuncWrapper(op.inplace_variant, n_expected_cudaLaunchKernels), RegularFuncWrapper(op.ref_inplace), ) def _binary_test(self, dtype, op, ref, inputs, is_fastpath, is_inplace, *, alpha=None): ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1] ] if is_inplace else inputs try: actual = op(inputs, self.is_cuda, is_fastpath) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): ref(ref_inputs) else: expected = ref(ref_inputs) self.assertEqual(actual, expected) if alpha is not None: kwargs = {'alpha': alpha} ref_inputs = inputs try: actual = op(inputs, self.is_cuda, is_fastpath, **kwargs) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): ref(ref_inputs, **kwargs) else: expected = ref(ref_inputs, **kwargs) if dtype in (torch.float16, torch.bfloat16) and TEST_WITH_ROCM: self.assertEqual(expected, actual, atol=1.e-3, rtol=self.dtype_precisions[dtype][0]) else: self.assertEqual(expected, actual) def _test_binary_op_tensorlists(self, device, dtype, opinfo, N, is_fastpath, disable_fastpath): n_expected_cudaLaunchKernels = N if disable_fastpath else 1 op, ref, inplace_op, inplace_ref = self._get_funcs( opinfo, n_expected_cudaLaunchKernels) inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), ] self._binary_test(dtype, op, ref, inputs, is_fastpath, is_inplace=False) self._binary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath, is_inplace=True) if opinfo.supports_alpha_param: alpha = None if dtype in get_all_int_dtypes(): alpha = 3 elif dtype.is_complex: alpha = complex(3, 3) else: alpha = 3.14 self._binary_test(dtype, op, ref, inputs, is_fastpath, is_inplace=False, alpha=alpha) self._binary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath, is_inplace=True, alpha=alpha) # Tests of implicit broadcasting # When sizes of tensors don't match, foreach functions are supposed to choose slow path # even if this methods's argument `is_fastpath` is True. # `cudaLaunchKernel` will be equal to `N`. For assert in `ForeachFuncWrapper` to pass, # we pass `is_fastpath and disable_fastpath` to `_binary_test`'s argument of is_fastpath. # as n_expected_cudaLaunchKernels is N if disable_fastpath. inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), [ make_tensor((N - i, 1), device=device, dtype=dtype, noncontiguous=not is_fastpath) for i in range(N) ], ] self._binary_test(dtype, op, ref, inputs, is_fastpath and disable_fastpath, is_inplace=False) self._binary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath and disable_fastpath, is_inplace=True) # note(mkozuki): Why ROCm? # ROCm is supposed to compile slow path as in # https://github.com/pytorch/pytorch/blob/7e032f18cf1405804c4f787b05ea2de5e08a091e/aten/src/ATen/native/ForeachUtils.h#L148-L164, # noqa: E501 # Therefore `[torch.add(*args, alpha=alpha) for args in zip(tensors1, tensors2)]` and # `torch._foreach_add(tensors1, tensors2, alpha=alpha)` # are expected to return the same outputs, however, the outputs look unstable for torch.bfloat16 and torch.half. # log: https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-linux-bionic-rocm4.2-py3.6-test1/2741/console @skipCUDAIfRocm @skipMeta @ops(foreach_binary_op_db) def test_binary_op_tensorlists_fastpath(self, device, dtype, op): for N in N_values: disable_fastpath = op.ref == torch.div and dtype in get_all_int_dtypes( ) + [torch.bool] if op.ref == torch.add and dtype == torch.bool: disable_fastpath = True self._test_binary_op_tensorlists(device, dtype, op, N, True, disable_fastpath) @ops(foreach_binary_op_db) def test_binary_op_tensorlists_slowpath(self, device, dtype, op): for N in N_values: self._test_binary_op_tensorlists(device, dtype, op, N, False, False) def _test_binary_op_scalar(self, device, dtype, opinfo, N, scalar, is_fastpath, disable_fastpath): n_expected_cudaLaunchKernels = N if disable_fastpath else 1 op, ref, inplace_op, inplace_ref = self._get_funcs( opinfo, n_expected_cudaLaunchKernels) inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), scalar ] self._binary_test(dtype, op, ref, inputs, is_fastpath, is_inplace=False) self._binary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath, is_inplace=True) @skipCUDAIfRocm @skipMeta @ops(foreach_binary_op_db) def test_binary_op_scalar_fastpath(self, device, dtype, op): for N, scalar in itertools.product(N_values, Scalars): disable_fastpath = op.ref == torch.div and dtype in get_all_int_dtypes( ) + [torch.bool] if isinstance(scalar, int): disable_fastpath |= dtype == torch.bool if isinstance(scalar, float): disable_fastpath |= dtype in get_all_int_dtypes() + [ torch.bool ] if isinstance(scalar, bool): disable_fastpath |= dtype == torch.bool if op.ref in (torch.add, torch.mul): disable_fastpath = False if isinstance(scalar, complex): disable_fastpath |= dtype not in get_all_complex_dtypes() self._test_binary_op_scalar(device, dtype, op, N, scalar, True, disable_fastpath) @ops(foreach_binary_op_db) def test_binary_op_scalar_slowpath(self, device, dtype, op): for N, scalar in itertools.product(N_values, Scalars): self._test_binary_op_scalar(device, dtype, op, N, scalar, False, False) def _test_binary_op_scalarlist(self, device, dtype, opinfo, N, scalarlist, is_fastpath, disable_fastpath): n_expected_cudaLaunchKernels = N if disable_fastpath else 1 op, ref, inplace_op, inplace_ref = self._get_funcs( opinfo, n_expected_cudaLaunchKernels) inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), scalarlist ] self._binary_test(dtype, op, ref, inputs, is_fastpath, is_inplace=False) self._binary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath, is_inplace=True) # note(mkozuki): Why two functions depending on with/without bool? # `foreach_sub` & `foreach_sub_` do `sub_check(tensors[i], scalars[i])` from i=1...N. # So, if scalarlist has one or more bool values, `foreach_sub` and `foreach_sub_` # raise bool subtraction error before doing any math. # While regular `sub` and `sub_` do some math until they encounter bool. # So, foreach sub's throw bool sub error first. However, regular sub's throw different # errors depending on the order of scalarlist. To keep actual unit test impl simple, # separating mixed scalarlist tests. By setting the first element of scalarlist to bool, # they are expected to throw bool sub error even in inplace test. @skipCUDAIfRocm @skipMeta @ops(foreach_binary_op_db) def test_binary_op_scalarlist_fastpath(self, device, dtype, op): for N in N_values: for type_str, scalarlist in getScalarLists(N): bool_int_div = op.ref == torch.div and dtype in get_all_int_dtypes( ) + [torch.bool] disable_fastpath = bool_int_div if type_str == "int": disable_fastpath |= dtype == torch.bool if type_str == "float": disable_fastpath |= dtype in get_all_int_dtypes() + [ torch.bool ] if type_str == "complex": disable_fastpath |= dtype not in get_all_complex_dtypes() if type_str == "mixed": disable_fastpath |= True and dtype not in get_all_complex_dtypes( ) self._test_binary_op_scalarlist(device, dtype, op, N, scalarlist, True, disable_fastpath) @ops(foreach_binary_op_db) def test_binary_op_scalarlist_slowpath(self, device, dtype, op): for N in N_values: for _, scalarlist in getScalarLists(N): self._test_binary_op_scalarlist(device, dtype, op, N, scalarlist, False, False) def _pointwise_test(self, dtype, op, ref, inputs, is_fastpath, is_inplace, *, values=None): ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1], inputs[2]] if is_inplace else inputs try: actual = op(inputs, self.is_cuda, is_fastpath) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): ref(ref_inputs) else: expected = ref(ref_inputs) self.assertEqual(expected, actual) if values is not None: try: actual = op(inputs + [values], self.is_cuda, is_fastpath) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): ref(ref_inputs, values=values) else: expected = ref(ref_inputs, values=values) self.assertEqual(expected, actual) def _test_pointwise_op(self, device, dtype, opinfo, N, is_fastpath, disable_fastpath, *, values=None): n_expected_cudaLaunchKernels = N if disable_fastpath else 1 op, ref, inplace_op, inplace_ref = self._get_funcs( opinfo, n_expected_cudaLaunchKernels) inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), ] self._pointwise_test(dtype, op, ref, inputs, is_fastpath, is_inplace=False, values=values) self._pointwise_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath, is_inplace=True, values=values) # Tests of implicit broadcasting inputs = [ opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath, same_size=True), [ make_tensor((N - i, 1), device=device, dtype=dtype, noncontiguous=not is_fastpath) for i in range(N) ], [ make_tensor((1, N - i), device=device, dtype=dtype, noncontiguous=not is_fastpath) for i in range(N) ], ] self._pointwise_test(dtype, op, ref, inputs, is_fastpath and disable_fastpath, is_inplace=False, values=values) self._pointwise_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath and disable_fastpath, is_inplace=True, values=values) @skipMeta @ops(foreach_pointwise_op_db) def test_pointwise_op_fastpath(self, device, dtype, op): disable_fastpath = dtype in get_all_int_dtypes() + [torch.bool] # for N, scalar in itertools.product(N_values, Scalars): for N in N_values: self._test_pointwise_op(device, dtype, op, N, True, disable_fastpath) for scalar in Scalars: self._test_pointwise_op(device, dtype, op, N, True, disable_fastpath, values=scalar) for _, scalarlist in getScalarLists(N): self._test_pointwise_op(device, dtype, op, N, True, disable_fastpath, values=scalarlist) @ops(foreach_pointwise_op_db) def test_pointwise_op_slowpath(self, device, dtype, op): # for N, scalar in itertools.product(N_values, Scalars): for N in N_values: self._test_pointwise_op(device, dtype, op, N, False, False) for scalar in Scalars: self._test_pointwise_op(device, dtype, op, N, False, False, values=scalar) for _, scalarlist in getScalarLists(N): self._test_pointwise_op(device, dtype, op, N, False, False, values=scalarlist) # note(mkozuki): fastpath test uses dtypes which fastpath implementation supports. # To confirm the dtypes of `OpInfo` cover the dtypes that the function support, # this test does not use `try-except` for fastpath. def _regular_unary_test(self, dtype, op, ref, inputs, is_fastpath): if is_fastpath: self.assertEqual(ref(inputs), op(inputs, self.is_cuda, is_fastpath)) return try: actual = op(inputs, self.is_cuda, is_fastpath) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): ref(inputs) else: expected = ref(inputs) self.assertEqual(actual, expected) # note(mkozuki): why `try-except` for both fastpath? # - inputs for fastpath can be integer tensors. # - this is becase opinfo dtypes are configured for outpulace implementation # - for integer inputs, trigonometric functions and exponential function returns float outputs, # which causes "result type Float can't be case to the desired type" error. # Thus, `try-except` is used even if `is_fastpath` is `True`. def _inplace_unary_test(self, dtype, inplace, inplace_ref, inputs, is_fastpath): copied_inputs = [[t.clone().detach() for t in tensors] for tensors in inputs] try: inplace(inputs, self.is_cuda, is_fastpath) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): inplace_ref(copied_inputs) else: inplace_ref(copied_inputs), self.assertEqual(copied_inputs, inputs) def _test_unary(self, device, dtype, opinfo, N, is_fastpath): op, ref, inplace_op, inplace_ref = self._get_funcs(opinfo, 1) inputs = opinfo.sample_inputs(device, dtype, N, noncontiguous=not is_fastpath), # note(mkozuki): Complex inputs for `_foreach_abs` go through slowpath. if opinfo.name == "_foreach_abs" and dtype in get_all_complex_dtypes(): is_fastpath = False self._regular_unary_test(dtype, op, ref, inputs, is_fastpath) self._inplace_unary_test(dtype, inplace_op, inplace_ref, inputs, is_fastpath) @skipMeta @ops(foreach_unary_op_db) def test_unary_fastpath(self, device, dtype, op): for N in N_values: self._test_unary(device, dtype, op, N, is_fastpath=True) @ops(foreach_unary_op_db, dtypes=get_all_dtypes()) def test_unary_slowpath(self, device, dtype, op): for N in N_values: self._test_unary(device, dtype, op, N, is_fastpath=False) def _minmax_test(self, opinfo, inputs, is_fastpath, n_expected_cudaLaunchKernels): op, ref, _, _ = self._get_funcs(opinfo, n_expected_cudaLaunchKernels) self.assertEqual(ref(inputs), op(inputs, self.is_cuda, is_fastpath)) # note(mkozuki): in-place of foreach_minimum and foreach_maximum aren't implemented. @ops(foreach_minmax_op_db) def test_minmax_fastpath(self, device, dtype, op): for N in N_values: inputs = tuple( op.sample_inputs(device, dtype, N) for _ in range(2)) self._minmax_test(op, inputs, True, N if dtype == torch.bool else 1) @ops(foreach_minmax_op_db, dtypes=get_all_dtypes(include_half=True, include_bfloat16=True, include_complex=False)) def test_minmax_slowpath(self, device, dtype, op): for N in N_values: inputs = tuple( op.sample_inputs(device, dtype, N, noncontiguous=True) for _ in range(2)) self._minmax_test(op, inputs, False, 1) # note(mkozuki): ForeachFuncInfo's of both `_foreach_maximum` and `_foreach_minimum` include integer types. # so, manually limit dtypes to fp types for inf&nan tests. @ops(foreach_minmax_op_db, dtypes=get_all_fp_dtypes(include_bfloat16=True, include_half=True)) def test_minmax_float_inf_nan(self, device, dtype, op): inputs = ( [ torch.tensor([float('inf')], device=device, dtype=dtype), torch.tensor([-float('inf')], device=device, dtype=dtype), torch.tensor([float('nan')], device=device, dtype=dtype), torch.tensor([float('nan')], device=device, dtype=dtype) ], [ torch.tensor([-float('inf')], device=device, dtype=dtype), torch.tensor([float('inf')], device=device, dtype=dtype), torch.tensor([float('inf')], device=device, dtype=dtype), torch.tensor([float('nan')], device=device, dtype=dtype) ], ) self._minmax_test(op, inputs, True, 1) def _reduce_test(self, opinfo, inputs, ord, is_fastpath, n_expected_cudaLaunchKernels): op, ref, _, _ = self._get_funcs(opinfo, n_expected_cudaLaunchKernels) self.assertEqual(ref(inputs, ord=ord), op(inputs, self.is_cuda, is_fastpath, ord=ord)) @ops(foreach_reduce_op_db) def test_reduce_fastpath(self, device, dtype, op): for N, ord in itertools.product(N_values, (0, 1, 2, -1, -2)): if ord in (1, 2) and dtype in torch.testing.get_all_fp_dtypes(): n_expected_cudaLaunchKernels = 3 else: n_expected_cudaLaunchKernels = N inputs = op.sample_inputs(device, dtype, N, noncontiguous=False), self._reduce_test(op, inputs, ord, True, n_expected_cudaLaunchKernels) @ops(foreach_reduce_op_db) def test_reduce_slowpath(self, device, dtype, op): for N, ord in itertools.product(N_values, (0, 1, 2, -1, -2)): inputs = op.sample_inputs(device, dtype, N, noncontiguous=True), self._reduce_test(op, inputs, ord, False, 1) @dtypes(*get_all_dtypes()) def test_add_scalar_with_empty_list_and_empty_tensor(self, device, dtype): # TODO: enable empty list case for tensors in [[torch.randn([0])]]: res = torch._foreach_add(tensors, 1) self.assertEqual(res, tensors) torch._foreach_add_(tensors, 1) self.assertEqual(res, tensors) @ops(foreach_binary_op_db, dtypes=get_all_dtypes()) def test_binary_op_scalar_with_overlapping_tensors(self, device, dtype, op): foreach_op, ref = op.method_variant, op.ref tensors = [ torch.ones(1, 1, device=device, dtype=dtype).expand(2, 1, 3) ] if ref == torch.sub and dtype == torch.bool: with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)): [ref(t, 1) for t in tensors] with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)): foreach_op(tensors, 1) return expected = [ref(t, 1) for t in tensors] res = foreach_op(tensors, 1) self.assertEqual(res, expected) # note(mkozuki): this test case fails with Meta at least in my local environment. # The message was # `AssertionError: NotImplementedError("Could not run 'aten::_foreach_add.Scalar' with arguments from the 'Meta' backend.` @skipMeta @ops(foreach_binary_op_db, allowed_dtypes=[torch.float]) def test_binary_op_scalar_with_different_tensor_dtypes( self, device, dtype, op): foreach_op = op.method_variant tensors = [ torch.tensor([1.1], dtype=torch.float, device=device), torch.tensor([1], dtype=torch.long, device=device) ] runtime_error = None try: foreach_op(tensors, 1) except RuntimeError as e: runtime_error = e self.assertIsNone(runtime_error) @ops(foreach_binary_op_db, dtypes=get_all_dtypes()) def test_binary_op_list_error_cases(self, device, dtype, op): foreach_op, foreach_op_, ref, ref_ = op.method_variant, op.inplace_variant, op.ref, op.ref_inplace tensors1 = [] tensors2 = [] # Empty lists with self.assertRaisesRegex( RuntimeError, "There were no tensor arguments to this function"): foreach_op(tensors1, tensors2) with self.assertRaisesRegex( RuntimeError, "There were no tensor arguments to this function"): foreach_op_(tensors1, tensors2) # One empty list tensors1.append(torch.tensor([1], device=device, dtype=dtype)) with self.assertRaisesRegex( RuntimeError, "Tensor list must have same number of elements as scalar list." ): foreach_op(tensors1, tensors2) with self.assertRaisesRegex( RuntimeError, "Tensor list must have same number of elements as scalar list." ): foreach_op_(tensors1, tensors2) # Lists have different amount of tensors tensors2.append(torch.tensor([1], device=device)) tensors2.append(torch.tensor([1], device=device)) with self.assertRaisesRegex( RuntimeError, "Tensor lists must have the same number of tensors, got 1 and 2" ): foreach_op(tensors1, tensors2) with self.assertRaisesRegex( RuntimeError, "Tensor lists must have the same number of tensors, got 1 and 2" ): foreach_op_(tensors1, tensors2) # Corresponding tensors with different sizes that aren't compatible with broadcast # If sizes are different then foreach chooses slow path, thus error messages are expected # to be the same as torch regular function. tensors1 = [ torch.zeros(10, 10, device=device, dtype=dtype) for _ in range(10) ] tensors2 = [ torch.ones(11, 11, device=device, dtype=dtype) for _ in range(10) ] try: foreach_op(tensors1, tensors2) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): [ref(t1, t2) for t1, t2 in zip(tensors1, tensors2)] try: foreach_op_(tensors1, tensors2) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): [ref_(t1, t2) for t1, t2 in zip(tensors1, tensors2)] # different devices if self.device_type == "cuda" and torch.cuda.device_count() > 1: tensor1 = torch.zeros(10, 10, device="cuda:0", dtype=dtype) tensor2 = torch.ones(10, 10, device="cuda:1", dtype=dtype) if dtype == torch.bool and foreach_op == torch._foreach_sub: with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)): foreach_op([tensor1], [tensor2]) with self.assertRaisesRegex(RuntimeError, re.escape(_BOOL_SUB_ERR_MSG)): foreach_op_([tensor1], [tensor2]) return with self.assertRaisesRegex( RuntimeError, "Expected all tensors to be on the same device"): foreach_op([tensor1], [tensor2]) if dtype in get_all_int_dtypes() + [ torch.bool ] and foreach_op == torch._foreach_div: with self.assertRaisesRegex(RuntimeError, "result type"): foreach_op_([tensor1], [tensor2]) else: with self.assertRaisesRegex( RuntimeError, "Expected all tensors to be on the same device"): foreach_op_([tensor1], [tensor2]) @skipMeta @unittest.skipIf(not torch.cuda.is_available(), "CUDA not found") @ops(foreach_binary_op_db, dtypes=get_all_dtypes()) def test_binary_op_list_slow_path(self, device, dtype, op): # note(mkozuki): why `n_expected_cudaLaunchKernels=0`? # In this test, foreach functions don't go through fast path, # but as there is only one tensor in each list of tensors, # `cudaLaunchKernel` is 1 so ForeachFuncWrapper internal assert fails. foreach_op, native_op, foreach_op_, native_op_ = self._get_funcs( op, n_expected_cudaLaunchKernels=0) # 0-strides tensor1 = make_tensor((10, 10), dtype=dtype, device=device) tensor2 = make_tensor((1, ), device=device, dtype=dtype).expand_as(tensor1) inputs = ([tensor1], [tensor2]) self._binary_test(dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False) self._binary_test(dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True) # different strides tensor1 = torch.zeros(10, 10, device=device, dtype=dtype) tensor2 = torch.ones(10, 10, device=device, dtype=dtype) inputs = ([tensor1], [tensor2.t()]) self._binary_test(dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False) self._binary_test(dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True) # non contiguous tensor1 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype, noncontiguous=True) tensor2 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype, noncontiguous=True) self.assertFalse(tensor1.is_contiguous()) self.assertFalse(tensor2.is_contiguous()) inputs = ([tensor1], [tensor2]) self._binary_test(dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False) self._binary_test(dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True) # sliced tensor tensor1 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype) tensor2 = make_tensor((5, 2, 1, 3 * 7), device=device, dtype=dtype)[:, :, :, ::7] inputs = ([tensor1], [tensor2]) self._binary_test(dtype, foreach_op, native_op, inputs, is_fastpath=False, is_inplace=False) self._binary_test(dtype, foreach_op_, native_op_, inputs, is_fastpath=False, is_inplace=True) # note: Below three tests (postfixed with `_tensors_on_different_devices`) # checks whether foreach works with lists of tensors on different devices # but tensors of the same index are on the same device, e.g., ['cuda', 'cpu]. @onlyCUDA @ops(foreach_unary_op_db) def test_unary_op_tensors_on_different_devices(self, device, dtype, op): method, ref, inplace_method, ref_inplace = self._get_funcs(op, 1) # tensors: ['cuda', 'cpu] tensors = op.sample_inputs(device, dtype, 2) tensors[1] = tensors[1].to('cpu') try: actual = method((tensors, ), False, False) except RuntimeError as e: with self.assertRaisesRegex(type(e), str(e)): ref((tensors, )) else: expected = ref((tensors, )) self.assertEqual(expected, actual) try: inplace_method((tensors, ), False, False) except RuntimeError as e: with self.assertRaisesRegex(type(e), str(e)): ref_inplace((tensors, )) else: self.assertEqual(expected, tensors) @onlyCUDA @ops(foreach_binary_op_db) def test_binary_op_tensors_on_different_devices(self, device, dtype, op): # `tensors1`: ['cuda', 'cpu'] # `tensors2`: ['cuda', 'cpu'] _cuda_tensors = op.sample_inputs(device, dtype, 2, same_size=True) _cpu_tensors = op.sample_inputs('cpu', dtype, 2, same_size=True) tensors1, tensors2 = list( tensors for tensors in zip(_cuda_tensors, _cpu_tensors)) foreach_op, foreach_op_ = op.method_variant, op.inplace_variant native_op, native_op_ = op.ref, op.ref_inplace try: actual = foreach_op(tensors1, tensors2) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): [native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2)] else: expected = [ native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2) ] self.assertEqual(expected, actual) try: foreach_op_(tensors1, tensors2) except RuntimeError as e: with self.assertRaisesRegex(type(e), re.escape(str(e))): [native_op_(t1, t2) for t1, t2 in zip(tensors1, tensors2)] else: self.assertEqual(actual, tensors1) @onlyCUDA @ops(foreach_pointwise_op_db, allowed_dtypes=get_all_fp_dtypes(include_half=False, include_bfloat16=False)) def test_pointwise_op_tensors_on_different_devices(self, device, dtype, op): # tensors1: ['cuda', 'cpu] # tensors2: ['cuda', 'cpu] # tensors3: ['cuda', 'cpu] _cuda_tensors = op.sample_inputs(device, dtype, 3, same_size=True) _cpu_tensors = op.sample_inputs('cpu', dtype, 3, same_size=True) tensors1, tensors2, tensors3 = list( tensors for tensors in zip(_cuda_tensors, _cpu_tensors)) foreach_op, foreach_op_, native_op = op.method_variant, op.inplace_variant, op.ref actual = foreach_op(tensors1, tensors2, tensors3) expected = [native_op(*_cuda_tensors), native_op(*_cpu_tensors)] self.assertEqual(expected, actual) # note(mkozuki): Limiting dtypes to FP32&FP64, we can safely run inplace ops. foreach_op_(tensors1, tensors2, tensors3) self.assertEqual(expected, tensors1)
class TestSortAndSelect(TestCase): def assertIsOrdered(self, order, x, mxx, ixx, task): SIZE = x.size(1) if order == 'descending': def check_order(a, b): # `a != a` because we put NaNs # at the end of ascending sorted lists, # and the beginning of descending ones. return ((a != a) | (a >= b)).all().item() elif order == 'ascending': def check_order(a, b): # see above return ((b != b) | (a <= b)).all().item() else: error('unknown order "{}", must be "ascending" or "descending"'.format(order)) are_ordered = True for k in range(1, SIZE): self.assertTrue(check_order(mxx[:, k - 1], mxx[:, k]), 'torch.sort ({}) values unordered for {}'.format(order, task)) seen = set() indicesCorrect = True size0 = x.size(0) size = x.size(x.dim() - 1) x = x.tolist() mxx = mxx.tolist() ixx = ixx.tolist() for k in range(size0): seen.clear() for j in range(size): self.assertEqual(x[k][ixx[k][j]], mxx[k][j], msg='torch.sort ({}) indices wrong for {}'.format(order, task)) seen.add(ixx[k][j]) self.assertEqual(len(seen), size) def test_sort(self, device): # on CUDA 2048 vs >2048 have different code path for the dim being sorted for SIZE in (4, 2049): x = torch.rand(4, SIZE, device=device) res1val, res1ind = torch.sort(x) # Test inplace y = x.clone() y_inds = torch.tensor((), dtype=torch.int64, device=device) torch.sort(y, out=(y, y_inds)) x_vals, x_inds = torch.sort(x) self.assertEqual(x_vals, y) self.assertEqual(x_inds, y_inds) # Test use of result tensor res2val = torch.tensor((), device=device) res2ind = torch.tensor((), device=device, dtype=torch.long) torch.sort(x, out=(res2val, res2ind)) self.assertEqual(res1val, res2val, atol=0, rtol=0) self.assertEqual(res1ind, res2ind, atol=0, rtol=0) self.assertEqual(torch.argsort(x), res1ind) self.assertEqual(x.argsort(), res1ind) # Test sorting of random numbers self.assertIsOrdered('ascending', x, res2val, res2ind, 'random') # Test simple sort self.assertEqual( torch.sort(torch.tensor((50, 40, 30, 20, 10), device=device))[0], torch.tensor((10, 20, 30, 40, 50), device=device), atol=0, rtol=0 ) # Test that we still have proper sorting with duplicate keys x = torch.floor(torch.rand(4, SIZE, device=device) * 10) torch.sort(x, out=(res2val, res2ind)) self.assertIsOrdered('ascending', x, res2val, res2ind, 'random with duplicate keys') # DESCENDING SORT x = torch.rand(4, SIZE, device=device) res1val, res1ind = torch.sort(x, x.dim() - 1, True) # Test use of result tensor res2val = torch.tensor((), device=device) res2ind = torch.tensor((), device=device, dtype=torch.long) torch.sort(x, x.dim() - 1, True, out=(res2val, res2ind)) self.assertEqual(res1val, res2val, atol=0, rtol=0) self.assertEqual(res1ind, res2ind, atol=0, rtol=0) self.assertEqual(torch.argsort(x, x.dim() - 1, True), res1ind) self.assertEqual(x.argsort(x.dim() - 1, True), res1ind) # Test sorting of random numbers self.assertIsOrdered('descending', x, res2val, res2ind, 'random') # Test simple sort task self.assertEqual( torch.sort(torch.tensor((10, 20, 30, 40, 50), device=device), 0, True)[0], torch.tensor((50, 40, 30, 20, 10), device=device), atol=0, rtol=0 ) # Test that we still have proper sorting with duplicate keys self.assertIsOrdered('descending', x, res2val, res2ind, 'random with duplicate keys') # Test sorting with NaNs x = torch.rand(4, SIZE, device=device) x[1][2] = float('NaN') x[3][0] = float('NaN') torch.sort(x, out=(res2val, res2ind)) self.assertIsOrdered('ascending', x, res2val, res2ind, 'random with NaNs') torch.sort(x, out=(res2val, res2ind), descending=True) self.assertIsOrdered('descending', x, res2val, res2ind, 'random with NaNs') # FIXME: remove torch.bool from unsupported types once support is added for cub sort @dtypes(*set(get_all_dtypes()) - {torch.bool, torch.complex64, torch.complex128}) def test_stable_sort(self, device, dtype): if TEST_WITH_ROCM and dtype == torch.bfloat16: return sizes = (100, 1000, 10000) for ncopies in sizes: x = torch.tensor([0, 1] * ncopies, dtype=dtype, device=device) _, idx = x.sort(stable=True) self.assertEqual( idx[:ncopies], torch.arange(start=0, end=2 * ncopies, step=2, device=device) ) self.assertEqual( idx[ncopies:], torch.arange(start=1, end=2 * ncopies, step=2, device=device) ) @onlyCUDA @dtypes(torch.uint8) @largeTensorTest('200GB') # Unfortunately 80GB A100 is not large enough def test_sort_large(self, device, dtype): t0 = torch.randperm(8192, device=device).to(dtype) t = t0.view(1, 8192).expand(2 ** 18 + 1, -1).contiguous() v, i = t.sort() del t iv, im = i.var_mean(dim=0) del i vv, vm = v.var_mean(dim=0) del v self.assertEqual(vv, torch.zeros_like(vv)) self.assertEqual(iv, torch.zeros_like(iv)) self.assertEqual(vm, torch.arange(255, dtype=dtype, device=device)) self.assertEqual(im, t0.sort().indices) def _test_sort_discontiguous(self, device, dtype): # on CUDA 2048 vs >2048 have different code path for the dim being sorted sizes = (5, 7, 2049) for shape in permutations(sizes): for perm in permutations((0, 1, 2)): for dim in range(3): t = torch.randn(shape, device=device, dtype=dtype).permute(perm) r1 = t.sort(dim=dim) r2 = t.contiguous().sort(dim=dim) self.assertEqual(r1, r2) n = t.size(dim) # assert ordered self.assertTrue((r1.values.narrow(dim, 1, n - 1) >= r1.values.narrow(dim, 0, n - 1)).all()) # assert that different segments does not mix, which can easily happen # if the stride is not handled correctly self.assertTrue((t.unsqueeze(-1).transpose(dim, -1) == r1.values.unsqueeze(-1)).any(dim=dim).any(dim=-1).all()) # assert stride is preserved if self.device_type == 'cuda': # FIXME: this behavior should be true for all cases, not # just the one specified in if condition self.assertEqual(r1.values.stride(), t.stride()) self.assertEqual(r1.indices.stride(), t.stride()) @onlyCUDA @dtypes(torch.float32) def test_sort_discontiguous(self, device, dtype): self._test_sort_discontiguous(device, dtype) @slowTest # this test is slow on CPU, but not on CUDA @onlyCPU @dtypes(torch.float32) def test_sort_discontiguous_slow(self, device, dtype): self._test_sort_discontiguous(device, dtype) @dtypes(torch.float32) def test_sort_1d_output_discontiguous(self, device, dtype): tensor = torch.randn(12, device=device, dtype=dtype)[:6] values = torch.empty_like(tensor)[::2] indices = torch.empty(18, device=device, dtype=torch.long)[::3] torch.sort(tensor, out=(values, indices)) values_cont, indices_cont = tensor.sort() self.assertEqual(indices, indices_cont) self.assertEqual(values, values_cont) @dtypes(torch.float32) def test_topk_1d_output_discontiguous(self, device, dtype): tensor = torch.randn(12, device=device, dtype=dtype) values = torch.empty_like(tensor)[::2] indices = torch.empty(18, device=device, dtype=torch.long)[::3] for sorted in (True, False): # outputs of `sorted=False` test are not guaranteed to be the same, # but with current implementation they are torch.topk(tensor, 6, sorted=sorted, out=(values, indices)) values_cont, indices_cont = tensor.topk(6, sorted=sorted) self.assertEqual(indices, indices_cont) self.assertEqual(values, values_cont) # FIXME: remove torch.bool from unsupported types once support is added for cub sort @dtypes(*set(get_all_dtypes()) - {torch.bool, torch.complex64, torch.complex128}) def test_stable_sort_against_numpy(self, device, dtype): if TEST_WITH_ROCM and dtype == torch.bfloat16: return if dtype in floating_types_and(torch.float16, torch.bfloat16): inf = float('inf') neg_inf = -float('inf') nan = float('nan') else: if dtype != torch.bool: # no torch.iinfo support for torch.bool inf = torch.iinfo(dtype).max neg_inf = torch.iinfo(dtype).min else: inf = True neg_inf = ~inf # no nan for integral types, we use inf instead for simplicity nan = inf def generate_samples(): from itertools import chain, combinations for sizes in [(1025,), (10000,)]: size = sizes[0] # binary strings yield (torch.tensor([0, 1] * size, dtype=dtype, device=device), 0) if self.device_type == 'cuda': return yield (torch.tensor([0, 1] * 100, dtype=dtype, device=device), 0) def repeated_index_fill(t, dim, idxs, vals): res = t for idx, val in zip(idxs, vals): res = res.index_fill(dim, idx, val) return res for sizes in [(1, 10), (10, 1), (10, 10), (10, 10, 10)]: size = min(*sizes) x = (torch.randn(*sizes, device=device) * size).to(dtype) yield (x, 0) # Generate tensors which are being filled at random locations # with values from the non-empty subsets of the set (inf, neg_inf, nan) # for each dimension. n_fill_vals = 3 # cardinality of (inf, neg_inf, nan) for dim in range(len(sizes)): idxs = (torch.randint(high=size, size=(size // 10,)) for i in range(n_fill_vals)) vals = (inf, neg_inf, nan) subsets = chain.from_iterable(combinations(list(zip(idxs, vals)), r) for r in range(1, n_fill_vals + 1)) for subset in subsets: idxs_subset, vals_subset = zip(*subset) yield (repeated_index_fill(x, dim, idxs_subset, vals_subset), dim) for sample, dim in generate_samples(): _, idx_torch = sample.sort(dim=dim, stable=True) if dtype is torch.bfloat16: sample_numpy = sample.float().cpu().numpy() else: sample_numpy = sample.cpu().numpy() idx_numpy = np.argsort(sample_numpy, axis=dim, kind='stable') self.assertEqual(idx_torch, idx_numpy) @dtypes(*(get_all_int_dtypes() + get_all_fp_dtypes())) def test_msort(self, device, dtype): if TEST_WITH_ROCM and dtype == torch.bfloat16: return def test(shape): tensor = make_tensor(shape, device, dtype, low=-9, high=9) if tensor.size() != torch.Size([]): if dtype is torch.bfloat16: expected = torch.from_numpy(np.msort(tensor.float().cpu().numpy())).bfloat16() else: expected = torch.from_numpy(np.msort(tensor.cpu().numpy())) else: expected = tensor # numpy.msort() does not support empty shapes tensor result = torch.msort(tensor) self.assertEqual(result, expected) out = torch.empty_like(result) torch.msort(tensor, out=out) self.assertEqual(out, expected) shapes = ( [], [0, ], [20, ], [1, 20], [30, 30], [10, 20, 30] ) for shape in shapes: test(shape) def test_topk(self, device): def topKViaSort(t, k, dim, dir): sorted, indices = t.sort(dim, dir) return sorted.narrow(dim, 0, k), indices.narrow(dim, 0, k) def compareTensors(t, res1, ind1, res2, ind2, dim): # Values should be exactly equivalent self.assertEqual(res1, res2, atol=0, rtol=0) # Indices might differ based on the implementation, since there is # no guarantee of the relative order of selection if not ind1.eq(ind2).all(): # To verify that the indices represent equivalent elements, # gather from the input using the topk indices and compare against # the sort indices vals = t.gather(dim, ind2) self.assertEqual(res1, vals, atol=0, rtol=0) def compare(t, k, dim, dir): topKVal, topKInd = t.topk(k, dim, dir, True) sortKVal, sortKInd = topKViaSort(t, k, dim, dir) compareTensors(t, sortKVal, sortKInd, topKVal, topKInd, dim) t = torch.rand(random.randint(1, SIZE), random.randint(1, SIZE), random.randint(1, SIZE), device=device) for _kTries in range(3): for _dimTries in range(3): for transpose in (True, False): for dir in (True, False): testTensor = t if transpose: dim1 = random.randrange(t.ndimension()) dim2 = dim1 while dim1 == dim2: dim2 = random.randrange(t.ndimension()) testTensor = t.transpose(dim1, dim2) dim = random.randrange(testTensor.ndimension()) k = random.randint(1, testTensor.size(dim)) compare(testTensor, k, dim, dir) # This tests the code path where on CUDA, topk is implemented with sort. t = torch.randn((2, 100000), device=device) compare(t, 2000, 1, True) compare(t, 2000, 1, False) def test_topk_arguments(self, device): q = torch.randn(10, 2, 10, device=device) # Make sure True isn't mistakenly taken as the 2nd dimension (interpreted as 1) self.assertRaises(TypeError, lambda: q.topk(4, True)) @skipCUDAIfRocm def test_unique_dim(self, device): self.assertFalse(hasattr(torch, 'unique_dim')) def run_test(device, dtype): x = torch.tensor([[[1., 1.], [0., 1.], [2., 1.], [0., 1.]], [[1., 1.], [0., 1.], [2., 1.], [0., 1.]]], dtype=dtype, device=device) x_empty = torch.empty(5, 0, dtype=dtype, device=device) x_ill_formed_empty = torch.empty(5, 0, 0, dtype=dtype, device=device) x_ill_formed_empty_another = torch.empty(5, 0, 5, dtype=dtype, device=device) if dtype in floating_types_and(torch.float16, torch.bfloat16): x_nan = torch.tensor([float("nan"), 0, 0, float("nan"), float("nan"), 1], dtype=dtype, device=device) expected_unique_dim0 = torch.tensor([[[1., 1.], [0., 1.], [2., 1.], [0., 1.]]], dtype=dtype, device=device) expected_inverse_dim0 = torch.tensor([0, 0]) expected_counts_dim0 = torch.tensor([2]) expected_unique_dim1 = torch.tensor([[[0., 1.], [1., 1.], [2., 1.]], [[0., 1.], [1., 1.], [2., 1.]]], dtype=dtype, device=device) expected_unique_dim1_bool = torch.tensor([[[False, True], [True, True]], [[False, True], [True, True]]], dtype=torch.bool, device=device) expected_inverse_dim1 = torch.tensor([1, 0, 2, 0]) expected_inverse_dim1_bool = torch.tensor([1, 0, 1, 0]) expected_counts_dim1 = torch.tensor([2, 1, 1]) expected_counts_dim1_bool = torch.tensor([2, 2]) expected_unique_dim2 = torch.tensor([[[1., 1.], [0., 1.], [2., 1.], [0., 1.]], [[1., 1.], [0., 1.], [2., 1.], [0., 1.]]], dtype=dtype, device=device) expected_inverse_dim2 = torch.tensor([0, 1]) expected_counts_dim2 = torch.tensor([1, 1]) expected_unique_empty = torch.tensor([], dtype=dtype, device=device) expected_inverse_empty = torch.tensor([], dtype=torch.long, device=device) expected_counts_empty = torch.tensor([], dtype=torch.long, device=device) if dtype in floating_types_and(torch.float16, torch.bfloat16): expected_unique_nan = torch.tensor([float("nan"), 0, float("nan"), float("nan"), 1], dtype=dtype, device=device) expected_inverse_nan = torch.tensor([0, 1, 1, 2, 3, 4], dtype=torch.long, device=device) expected_counts_nan = torch.tensor([1, 2, 1, 1, 1], dtype=torch.long, device=device) # dim0 x_unique = torch.unique(x, dim=0) self.assertEqual(expected_unique_dim0, x_unique) x_unique, x_inverse = torch.unique( x, return_inverse=True, dim=0) self.assertEqual(expected_unique_dim0, x_unique) self.assertEqual(expected_inverse_dim0, x_inverse) x_unique, x_counts = torch.unique( x, return_inverse=False, return_counts=True, dim=0) self.assertEqual(expected_unique_dim0, x_unique) self.assertEqual(expected_counts_dim0, x_counts) x_unique, x_inverse, x_counts = torch.unique( x, return_inverse=True, return_counts=True, dim=0) self.assertEqual(expected_unique_dim0, x_unique) self.assertEqual(expected_inverse_dim0, x_inverse) self.assertEqual(expected_counts_dim0, x_counts) # dim1 x_unique = torch.unique(x, dim=1) if x.dtype == torch.bool: self.assertEqual(expected_unique_dim1_bool, x_unique) else: self.assertEqual(expected_unique_dim1, x_unique) x_unique, x_inverse = torch.unique( x, return_inverse=True, dim=1) if x.dtype == torch.bool: self.assertEqual(expected_unique_dim1_bool, x_unique) self.assertEqual(expected_inverse_dim1_bool, x_inverse) else: self.assertEqual(expected_unique_dim1, x_unique) self.assertEqual(expected_inverse_dim1, x_inverse) x_unique, x_counts = torch.unique( x, return_inverse=False, return_counts=True, dim=1) if x.dtype == torch.bool: self.assertEqual(expected_unique_dim1_bool, x_unique) self.assertEqual(expected_counts_dim1_bool, x_counts) else: self.assertEqual(expected_unique_dim1, x_unique) self.assertEqual(expected_counts_dim1, x_counts) x_unique, x_inverse, x_counts = torch.unique( x, return_inverse=True, return_counts=True, dim=1) if x.dtype == torch.bool: self.assertEqual(expected_unique_dim1_bool, x_unique) self.assertEqual(expected_inverse_dim1_bool, x_inverse) self.assertEqual(expected_counts_dim1_bool, x_counts) else: self.assertEqual(expected_unique_dim1, x_unique) self.assertEqual(expected_inverse_dim1, x_inverse) self.assertEqual(expected_counts_dim1, x_counts) # dim2 x_unique = torch.unique(x, dim=2) self.assertEqual(expected_unique_dim2, x_unique) x_unique, x_inverse = torch.unique( x, return_inverse=True, dim=2) self.assertEqual(expected_unique_dim2, x_unique) self.assertEqual(expected_inverse_dim2, x_inverse) x_unique, x_counts = torch.unique( x, return_inverse=False, return_counts=True, dim=2) self.assertEqual(expected_unique_dim2, x_unique) self.assertEqual(expected_counts_dim2, x_counts) x_unique, x_inverse, x_counts = torch.unique( x, return_inverse=True, return_counts=True, dim=2) self.assertEqual(expected_unique_dim2, x_unique) self.assertEqual(expected_inverse_dim2, x_inverse) self.assertEqual(expected_counts_dim2, x_counts) # test empty tensor x_unique, x_inverse, x_counts = torch.unique( x_empty, return_inverse=True, return_counts=True, dim=1) self.assertEqual(expected_unique_empty, x_unique) self.assertEqual(expected_inverse_empty, x_inverse) self.assertEqual(expected_counts_empty, x_counts) # test tensor with nan if dtype in floating_types_and(torch.float16, torch.bfloat16): x_unique, x_inverse, x_counts = torch.unique( x_nan, return_inverse=True, return_counts=True, dim=0) self.assertEqual(expected_unique_nan, x_unique) self.assertEqual(expected_inverse_nan, x_inverse) self.assertEqual(expected_counts_nan, x_counts) # test not a well formed tensor # Checking for runtime error, as this is the expected behaviour with self.assertRaises(RuntimeError): torch.unique( x_ill_formed_empty, return_inverse=True, return_counts=True, dim=1) # test along dim2 with self.assertRaises(RuntimeError): torch.unique( x_ill_formed_empty_another, return_inverse=True, return_counts=True, dim=2) # test consecutive version y = torch.tensor( [[0, 1], [0, 1], [0, 1], [1, 2], [1, 2], [3, 4], [0, 1], [0, 1], [3, 4], [1, 2]], dtype=dtype, device=device ) # test tensor with nan if dtype in floating_types_and(torch.float16, torch.bfloat16): y_nan = torch.tensor([float("nan"), 0, 0, float("nan"), float("nan"), 1], dtype=dtype, device=device) expected_y_unique = torch.tensor( [[0, 1], [1, 2], [3, 4], [0, 1], [3, 4], [1, 2]], dtype=dtype, device=device ) expected_y_inverse = torch.tensor([0, 0, 0, 1, 1, 2, 3, 3, 4, 5], dtype=torch.int64, device=device) expected_y_counts = torch.tensor([3, 2, 1, 2, 1, 1], dtype=torch.int64, device=device) expected_y_inverse_bool = torch.tensor([0, 0, 0, 1, 1, 1, 2, 2, 3, 3], dtype=torch.int64, device=device) expected_y_counts_bool = torch.tensor([3, 3, 2, 2], dtype=torch.int64, device=device) if dtype in floating_types_and(torch.float16, torch.bfloat16): expected_y_unique_nan = torch.tensor([float("nan"), 0, float("nan"), float("nan"), 1], dtype=dtype, device=device) expected_y_inverse_nan = torch.tensor([0, 1, 1, 2, 3, 4], dtype=torch.long, device=device) expected_y_counts_nan = torch.tensor([1, 2, 1, 1, 1], dtype=torch.long, device=device) y_unique, y_inverse, y_counts = torch.unique_consecutive(y, return_inverse=True, return_counts=True, dim=0) if x.dtype == torch.bool: self.assertEqual(expected_y_inverse_bool, y_inverse) self.assertEqual(expected_y_counts_bool, y_counts) else: self.assertEqual(expected_y_inverse, y_inverse) self.assertEqual(expected_y_counts, y_counts) # test tensor with nan if dtype in floating_types_and(torch.float16, torch.bfloat16): y_unique, y_inverse, y_counts = torch.unique_consecutive( y_nan, return_inverse=True, return_counts=True, dim=0) self.assertEqual(expected_y_unique_nan, y_unique) self.assertEqual(expected_y_inverse_nan, y_inverse) self.assertEqual(expected_y_counts_nan, y_counts) run_test(device, torch.float) run_test(device, torch.double) run_test(device, torch.long) run_test(device, torch.uint8) run_test(device, torch.bool) @onlyCUDA def test_topk_noncontiguous_gpu(self, device): t = torch.randn(20, device=device)[::2] top1, idx1 = t.topk(5) top2, idx2 = t.contiguous().topk(5) self.assertEqual(top1, top2) self.assertEqual(idx1, idx2) def _test_topk_dtype(self, device, dtype, integral, size): if integral: a = torch.randint(torch.iinfo(dtype).min, torch.iinfo(dtype).max, size=(size,), dtype=dtype, device=device) else: a = torch.randn(size=(size,), dtype=dtype, device=device) sort_topk = a.sort()[0][-(size // 2):].flip(0) topk = a.topk(size // 2) self.assertEqual(sort_topk, topk[0]) # check values self.assertEqual(sort_topk, a[topk[1]]) # check indices @dtypes(torch.int8, torch.uint8, torch.int16, torch.int32, torch.int64) def test_topk_integral(self, device, dtype): small = 10 large = 4096 for curr_size in (small, large): self._test_topk_dtype(device, dtype, True, curr_size) @onlyCUDA @dtypes(torch.bfloat16) @skipCUDAIfRocm def test_topk_bfloat16(self, device, dtype): small = 10 large = 8192 for curr_size in (small, large): self._test_topk_dtype(device, dtype, False, curr_size) @dtypesIfCUDA(*get_all_fp_dtypes()) @dtypes(torch.float, torch.double, torch.bfloat16) def test_topk_nonfinite(self, device, dtype): if TEST_WITH_ROCM and dtype == torch.bfloat16: return x = torch.tensor([float('nan'), float('inf'), 1e4, 0, -1e4, -float('inf')], device=device, dtype=dtype) val, idx = x.topk(4) expect = torch.tensor([float('nan'), float('inf'), 1e4, 0], device=device, dtype=dtype) self.assertEqual(val, expect) self.assertEqual(idx, [0, 1, 2, 3]) val, idx = x.topk(4, largest=False) expect = torch.tensor([-float('inf'), -1e4, 0, 1e4], device=device, dtype=dtype) self.assertEqual(val, expect) self.assertEqual(idx, [5, 4, 3, 2]) def test_topk_4d(self, device): x = torch.ones(2, 3072, 2, 2, device=device) x[:, 1, :, :] *= 2. x[:, 10, :, :] *= 1.5 val, ind = torch.topk(x, k=2, dim=1) expected_ind = torch.ones(2, 2, 2, 2, dtype=torch.long, device=device) expected_ind[:, 1, :, :] = 10 expected_val = torch.ones(2, 2, 2, 2, device=device) expected_val[:, 0, :, :] *= 2. expected_val[:, 1, :, :] *= 1.5 self.assertEqual(val, expected_val, atol=0, rtol=0) self.assertEqual(ind, expected_ind, atol=0, rtol=0) @onlyNativeDeviceTypes @dtypesIfCUDA(*(get_all_dtypes(include_complex=False, include_bool=False, include_half=False, include_bfloat16=True))) @dtypes(*(get_all_dtypes(include_complex=False, include_bool=False, include_half=False, include_bfloat16=False))) def test_topk_zero(self, device, dtype): if TEST_WITH_ROCM and dtype == torch.bfloat16: return # https://github.com/pytorch/pytorch/issues/49205 t = torch.rand(2, 2, device=device).to(dtype=dtype) val, idx = torch.topk(t, k=0, largest=False) self.assertEqual(val.size(), torch.Size([2, 0])) self.assertEqual(idx.size(), torch.Size([2, 0])) def _test_unique_scalar_empty(self, dtype, device, f): # test scalar x = torch.tensor(0, dtype=dtype, device=device) unique, inverse, counts = f(x, return_inverse=True, return_counts=True) expected_unique = torch.tensor([0], dtype=dtype, device=device) expected_inverse = torch.tensor(0, device=device) expected_counts = torch.tensor([1], device=device) self.assertEqual(unique, expected_unique) self.assertEqual(inverse, expected_inverse) self.assertEqual(counts, expected_counts) # test zero sized tensor x = torch.zeros((0, 0, 3), dtype=dtype, device=device) unique, inverse, counts = f(x, return_inverse=True, return_counts=True) expected_unique = torch.tensor([], dtype=dtype, device=device) expected_inverse = torch.empty((0, 0, 3), dtype=torch.long, device=device) expected_counts = torch.tensor([], dtype=torch.long, device=device) self.assertEqual(unique, expected_unique) self.assertEqual(inverse, expected_inverse) self.assertEqual(counts, expected_counts) def _test_unique_with_expects(self, device, dtype, f, x, expected_unique, expected_inverse, expected_counts, additional_shape): def ensure_tuple(x): if isinstance(x, torch.Tensor): return (x,) return x for return_inverse in [True, False]: for return_counts in [True, False]: # test with expected ret = ensure_tuple(f(x, return_inverse=return_inverse, return_counts=return_counts)) self.assertEqual(len(ret), 1 + int(return_inverse) + int(return_counts)) self.assertEqual(expected_unique, ret[0]) if return_inverse: self.assertEqual(expected_inverse, ret[1]) if return_counts: count_index = 1 + int(return_inverse) self.assertEqual(expected_counts, ret[count_index]) # tests per-element unique on a higher rank tensor. y = x.view(additional_shape) y_unique, y_inverse, y_counts = f(y, return_inverse=True, return_counts=True) self.assertEqual(expected_unique, y_unique) self.assertEqual(expected_inverse.view(additional_shape), y_inverse) self.assertEqual(expected_counts, y_counts) @dtypesIfCPU(*set(get_all_dtypes()) - {torch.complex64, torch.complex128}) @dtypes(*set(get_all_dtypes()) - {torch.bfloat16, torch.complex64, torch.complex128}) def test_unique(self, device, dtype): if dtype is torch.half and self.device_type == 'cpu': return # CPU does not have half support def ensure_tuple(x): if isinstance(x, torch.Tensor): return (x,) return x if dtype is torch.bool: x = torch.tensor([True, False, False, False, True, False, True, False], dtype=torch.bool, device=device) expected_unique = torch.tensor([False, True], dtype=torch.bool, device=device) expected_inverse = torch.tensor([1, 0, 0, 0, 1, 0, 1, 0], dtype=torch.long, device=device) expected_counts = torch.tensor([5, 3], dtype=torch.long, device=device) else: x = torch.tensor([1, 2, 3, 2, 8, 5, 2, 3], dtype=dtype, device=device) expected_unique = torch.tensor([1, 2, 3, 5, 8], dtype=dtype, device=device) expected_inverse = torch.tensor([0, 1, 2, 1, 4, 3, 1, 2], device=device) expected_counts = torch.tensor([1, 3, 2, 1, 1], device=device) # test sorted unique fs = ( lambda x, **kwargs: torch.unique(x, sorted=True, **kwargs), lambda x, **kwargs: x.unique(sorted=True, **kwargs), ) x_sliced = torch.empty(x.size(0) * 2, dtype=dtype, device=device)[::2].copy_(x) xs = (x, x_sliced) for f, x in product(fs, xs): self._test_unique_with_expects(device, dtype, f, x, expected_unique, expected_inverse, expected_counts, (2, 2, 2)) self._test_unique_scalar_empty(dtype, device, f) # test unsorted unique fs = ( lambda x, **kwargs: torch.unique(x, sorted=False, **kwargs), lambda x, **kwargs: x.unique(sorted=False, **kwargs) ) for f, x in product(fs, xs): self._test_unique_scalar_empty(dtype, device, f) for return_inverse, return_counts in product((True, False), repeat=2): ret = ensure_tuple(f(x, return_inverse=return_inverse, return_counts=return_counts)) self.assertEqual(len(ret), 1 + int(return_inverse) + int(return_counts)) x_list = x.tolist() x_unique_list = ret[0].tolist() self.assertEqual(expected_unique.tolist(), sorted(x_unique_list)) if return_inverse: x_inverse_list = ret[1].tolist() for i, j in enumerate(x_inverse_list): self.assertEqual(x_list[i], x_unique_list[j]) if return_counts: count_index = 1 + int(return_inverse) x_counts_list = ret[count_index].tolist() for i, j in zip(x_unique_list, x_counts_list): count = 0 for k in x_list: if k == i: count += 1 self.assertEqual(j, count) @dtypesIfCPU(*set(get_all_dtypes()) - {torch.complex64, torch.complex128}) @dtypes(*set(get_all_dtypes()) - {torch.bfloat16, torch.complex64, torch.complex128}) def test_unique_consecutive(self, device, dtype): if dtype is torch.half and self.device_type == 'cpu': return # CPU does not have half support if dtype is torch.bool: x = torch.tensor([True, False, False, False, True, True, False, False, False], dtype=torch.bool, device=device) expected_unique = torch.tensor([True, False, True, False], dtype=torch.bool, device=device) expected_inverse = torch.tensor([0, 1, 1, 1, 2, 2, 3, 3, 3], dtype=torch.long, device=device) expected_counts = torch.tensor([1, 3, 2, 3], dtype=torch.long, device=device) else: x = torch.tensor([1, 2, 2, 2, 5, 5, 2, 2, 3], dtype=dtype, device=device) expected_unique = torch.tensor([1, 2, 5, 2, 3], dtype=dtype, device=device) expected_inverse = torch.tensor([0, 1, 1, 1, 2, 2, 3, 3, 4], device=device) expected_counts = torch.tensor([1, 3, 2, 2, 1], device=device) for f in [torch.unique_consecutive, lambda x, **kwargs: x.unique_consecutive(**kwargs)]: self._test_unique_with_expects(device, dtype, f, x, expected_unique, expected_inverse, expected_counts, (3, 3)) self._test_unique_scalar_empty(dtype, device, f) @dtypes(torch.double) def test_kthvalue(self, device, dtype): SIZE = 50 x = torch.rand(SIZE, SIZE, SIZE, dtype=dtype, device=device) x0 = x.clone() k = random.randint(1, SIZE) res1val, res1ind = torch.kthvalue(x, k, keepdim=False) res2val, res2ind = torch.sort(x) self.assertEqual(res1val[:, :], res2val[:, :, k - 1], atol=0, rtol=0) self.assertEqual(res1ind[:, :], res2ind[:, :, k - 1], atol=0, rtol=0) # test use of result tensors k = random.randint(1, SIZE) res1val = torch.tensor([], dtype=dtype, device=device) res1ind = torch.tensor([], dtype=torch.long, device=device) torch.kthvalue(x, k, keepdim=False, out=(res1val, res1ind)) res2val, res2ind = torch.sort(x) self.assertEqual(res1val[:, :], res2val[:, :, k - 1], atol=0, rtol=0) self.assertEqual(res1ind[:, :], res2ind[:, :, k - 1], atol=0, rtol=0) # test non-default dim k = random.randint(1, SIZE) res1val, res1ind = torch.kthvalue(x, k, 0, keepdim=False) res2val, res2ind = torch.sort(x, 0) self.assertEqual(res1val, res2val[k - 1], atol=0, rtol=0) self.assertEqual(res1ind, res2ind[k - 1], atol=0, rtol=0) # non-contiguous y = x.narrow(1, 0, 1) y0 = y.contiguous() k = random.randint(1, SIZE) res1val, res1ind = torch.kthvalue(y, k) res2val, res2ind = torch.kthvalue(y0, k) self.assertEqual(res1val, res2val, atol=0, rtol=0) self.assertEqual(res1ind, res2ind, atol=0, rtol=0) # non-contiguous [Reference: https://github.com/pytorch/pytorch/issues/45721] non_contig_t = torch.tensor([0, -1, 1, -2, 2], dtype=dtype, device=device)[::2] expected_val, expected_ind = non_contig_t.contiguous().kthvalue(2) non_contig_cpu_t = non_contig_t.cpu() expected_val_cpu, expected_ind_cpu = non_contig_cpu_t.kthvalue(2) out_val, out_ind = non_contig_t.kthvalue(2) self.assertEqual(expected_val, out_val, atol=0, rtol=0) self.assertEqual(expected_ind, out_ind, atol=0, rtol=0) self.assertEqual(expected_val_cpu, out_val, atol=0, rtol=0) self.assertEqual(expected_ind_cpu, out_ind, atol=0, rtol=0) # check that the input wasn't modified self.assertEqual(x, x0, atol=0, rtol=0) # simple test case (with repetitions) y = torch.tensor((3., 5, 4, 1, 1, 5), dtype=dtype, device=device) self.assertEqual(torch.kthvalue(y, 3)[0], 3, atol=0, rtol=0) self.assertEqual(torch.kthvalue(y, 2)[0], 1, atol=0, rtol=0) # simple test case (with NaN) SIZE = 50 x = torch.rand(SIZE, SIZE, SIZE, dtype=dtype, device=device) x[torch.arange(SIZE), :, torch.randint(50, (50,))] = nan ks = [random.randint(1, SIZE), 1, SIZE, SIZE - 1] res2val, res2ind = torch.sort(x) for k in ks: res1val, res1ind = torch.kthvalue(x, k, keepdim=False) self.assertEqual(res1val[:, :], res2val[:, :, k - 1], atol=0, rtol=0) self.assertEqual(res1ind[:, :], res2ind[:, :, k - 1], atol=0, rtol=0) @dtypes(torch.float) @onlyNativeDeviceTypes # Fails on XLA def test_kthvalue_scalar(self, device, dtype): # Test scalar input (test case from https://github.com/pytorch/pytorch/issues/30818) # Tests that passing a scalar tensor or 1D tensor with 1 element work either way res = torch.tensor(2, device=device, dtype=dtype).kthvalue(1) ref = torch.tensor([2], device=device, dtype=dtype).kthvalue(1) self.assertEqual(res[0], ref[0].squeeze()) self.assertEqual(res[1], ref[1].squeeze()) @dtypes(*all_types()) @dtypesIfCUDA(*all_types_and(torch.half)) def test_isin(self, device, dtype): def assert_isin_equal(a, b): # Compare to the numpy reference implementation. x = torch.isin(a, b) a = a.cpu().numpy() if torch.is_tensor(a) else np.array(a) b = b.cpu().numpy() if torch.is_tensor(b) else np.array(b) y = np.isin(a, b) self.assertEqual(x, y) # multi-dim tensor, multi-dim tensor a = torch.arange(24, device=device, dtype=dtype).reshape([2, 3, 4]) b = torch.tensor([[10, 20, 30], [0, 1, 3], [11, 22, 33]], device=device, dtype=dtype) assert_isin_equal(a, b) # zero-dim tensor zero_d = torch.tensor(3, device=device, dtype=dtype) assert_isin_equal(zero_d, b) assert_isin_equal(a, zero_d) assert_isin_equal(zero_d, zero_d) # empty tensor empty = torch.tensor([], device=device, dtype=dtype) assert_isin_equal(empty, b) assert_isin_equal(a, empty) assert_isin_equal(empty, empty) # scalar assert_isin_equal(a, 6) assert_isin_equal(5, b) def define_expected(lst, invert=False): expected = torch.tensor(lst, device=device) if invert: expected = expected.logical_not() return expected # Adapted from numpy's in1d tests for mult in [1, 10]: for invert in [False, True]: a = torch.tensor([5, 7, 1, 2], device=device, dtype=dtype) b = torch.tensor([2, 4, 3, 1, 5] * mult, device=device, dtype=dtype) ec = define_expected([True, False, True, True], invert=invert) c = torch.isin(a, b, assume_unique=True, invert=invert) self.assertEqual(c, ec) a[0] = 8 ec = define_expected([False, False, True, True], invert=invert) c = torch.isin(a, b, assume_unique=True, invert=invert) self.assertEqual(c, ec) a[0], a[3] = 4, 8 ec = define_expected([True, False, True, False], invert=invert) c = torch.isin(a, b, assume_unique=True, invert=invert) self.assertEqual(c, ec) a = torch.tensor([5, 4, 5, 3, 4, 4, 3, 4, 3, 5, 2, 1, 5, 5], device=device, dtype=dtype) b = torch.tensor([2, 3, 4] * mult, device=device, dtype=dtype) ec = define_expected([False, True, False, True, True, True, True, True, True, False, True, False, False, False], invert=invert) c = torch.isin(a, b, invert=invert) self.assertEqual(c, ec) b = torch.tensor([2, 3, 4] * mult + [5, 5, 4] * mult, device=device, dtype=dtype) ec = define_expected([True, True, True, True, True, True, True, True, True, True, True, False, True, True], invert=invert) c = torch.isin(a, b, invert=invert) self.assertEqual(c, ec) a = torch.tensor([5, 7, 1, 2], device=device, dtype=dtype) b = torch.tensor([2, 4, 3, 1, 5] * mult, device=device, dtype=dtype) ec = define_expected([True, False, True, True], invert=invert) c = torch.isin(a, b, invert=invert) self.assertEqual(c, ec) a = torch.tensor([5, 7, 1, 1, 2], device=device, dtype=dtype) b = torch.tensor([2, 4, 3, 3, 1, 5] * mult, device=device, dtype=dtype) ec = define_expected([True, False, True, True, True], invert=invert) c = torch.isin(a, b, invert=invert) self.assertEqual(c, ec) a = torch.tensor([5, 5], device=device, dtype=dtype) b = torch.tensor([2, 2] * mult, device=device, dtype=dtype) ec = define_expected([False, False], invert=invert) c = torch.isin(a, b, invert=invert) self.assertEqual(c, ec) # multi-dimensional input case using sort-based algo for assume_unique in [False, True]: a = torch.arange(6, device=device, dtype=dtype).reshape([2, 3]) b = torch.arange(3, 30, device=device, dtype=dtype) ec = define_expected([[False, False, False], [True, True, True]], invert=invert) c = torch.isin(a, b, invert=invert, assume_unique=assume_unique) self.assertEqual(c, ec) def test_isin_different_dtypes(self, device): supported_types = all_types() if device == 'cpu' else all_types_and(torch.half) for mult in [1, 10]: for assume_unique in [False, True]: for dtype1, dtype2 in product(supported_types, supported_types): a = torch.tensor([1, 2, 3], device=device, dtype=dtype1) b = torch.tensor([3, 4, 5] * mult, device=device, dtype=dtype2) ec = torch.tensor([False, False, True], device=device) c = torch.isin(a, b, assume_unique=assume_unique) self.assertEqual(c, ec) @onlyCUDA @dtypes(*all_types()) def test_isin_different_devices(self, device, dtype): a = torch.arange(6, device=device, dtype=dtype).reshape([2, 3]) b = torch.arange(3, 30, device='cpu', dtype=dtype) with self.assertRaises(RuntimeError): torch.isin(a, b) c = torch.arange(6, device='cpu', dtype=dtype).reshape([2, 3]) d = torch.arange(3, 30, device=device, dtype=dtype) with self.assertRaises(RuntimeError): torch.isin(c, d)
class TestScatterGather(TestCase): # Fills an index tensor with valid indices def _fill_indices(self, idx, dim, dim_size, elems_per_row, m, n, o, unique_indices=True): for i in range(1 if dim == 0 else m): for j in range(1 if dim == 1 else n): for k in range(1 if dim == 2 else o): ii = [i, j, k] ii[dim] = slice(0, idx.size(dim) + 1) if unique_indices: idx[tuple(ii)] = torch.randperm(dim_size)[0:elems_per_row] else: idx[tuple(ii)] = torch.randint(dim_size, (elems_per_row,)) @dtypes(torch.float32, torch.complex64) def test_gather(self, device, dtype): m, n, o = random.randint(10, 20), random.randint(10, 20), random.randint(10, 20) elems_per_row = random.randint(1, 10) dim = random.randrange(3) src = make_tensor((m, n, o), device=device, dtype=dtype) idx_size = [m, n, o] idx_size[dim] = elems_per_row idx = make_tensor(idx_size, device=device, dtype=torch.long) self._fill_indices(idx, dim, src.size(dim), elems_per_row, m, n, o) actual = torch.gather(src, dim, idx) expected = torch.zeros(idx_size, device=device, dtype=dtype) for i in range(idx_size[0]): for j in range(idx_size[1]): for k in range(idx_size[2]): ii = [i, j, k] ii[dim] = idx[i, j, k] expected[i, j, k] = src[tuple(ii)] self.assertEqual(actual, expected, atol=0, rtol=0) # Guarded because torch.max isn't defined for complex types if not dtype.is_complex: src = make_tensor((3, 4, 5), device=device, dtype=dtype) expected, idx = src.max(2, True) actual = torch.gather(src, 2, idx) self.assertEqual(actual, expected, atol=0, rtol=0) @dtypes(torch.bool) def test_gather_bool(self, device, dtype): src = torch.tensor(((False, True), (True, True)), device=device, dtype=dtype) idx = torch.tensor(((0, 0), (1, 0)), device=device, dtype=torch.long) actual = torch.gather(src, 1, idx) expected = torch.tensor(((False, False), (True, True)), device=device, dtype=dtype) self.assertEqual(actual, expected, atol=0, rtol=0) def _test_scatter_base(self, fn, *, device, dtype, is_scalar, reduction, unique_indices=True, include_self=True): m, n, o = random.randint(10, 20), random.randint(10, 20), random.randint(10, 20) elems_per_row = random.randint(1, 10) dim = random.randrange(3) idx_size = [m, n, o] idx_size[dim] = elems_per_row idx = torch.empty(tuple(idx_size), device=device, dtype=torch.long) self._fill_indices(idx, dim, ([m, n, o])[dim], elems_per_row, m, n, o, unique_indices) if is_scalar: src = random.random() else: src_size = [random.randint(1, 5) + s for s in idx_size] src = make_tensor(tuple(src_size), device=device, dtype=dtype) base = make_tensor((m, n, o), device=device, dtype=dtype) if reduction is not None: if fn is torch.Tensor.scatter_reduce_: actual = fn(base.clone(), dim, idx, src, reduce=reduction, include_self=include_self) else: actual = fn(base.clone(), dim, idx, src, reduce=reduction) else: actual = fn(base.clone(), dim, idx, src) expected = base.clone() counts = torch.zeros(base.shape, dtype=torch.long, device=device) + include_self for i in range(idx_size[0]): for j in range(idx_size[1]): for k in range(idx_size[2]): ii = [i, j, k] ii[dim] = idx[i, j, k] if fn is torch.Tensor.scatter_add_: expected[tuple(ii)] += src[i, j, k] else: # method may be 'scatter_', 'scatter', 'scatter_reduce' # or 'scatter_reduce_', the former two might have a reduction argument # while the latter two always do value = src if is_scalar else src[i, j, k] if ((not include_self) and counts[tuple(ii)] == 0): expected[tuple(ii)] = value else: if reduction == "add" or reduction == "sum": expected[tuple(ii)] += value elif reduction == "multiply" or reduction == "prod": expected[tuple(ii)] *= value elif reduction == "amax": expected[tuple(ii)] = max(expected[tuple(ii)], value) elif reduction == "amin": expected[tuple(ii)] = min(expected[tuple(ii)], value) elif reduction == "mean": expected[tuple(ii)] += value else: expected[tuple(ii)] = value counts[tuple(ii)] += 1 if (reduction == "mean"): counts.masked_fill_(counts == 0, 1) if (dtype.is_floating_point or dtype.is_complex): expected /= counts else: expected.div_(counts, rounding_mode="floor") self.assertEqual(actual, expected, atol=0, rtol=0) # Tests empty index dst = make_tensor((2, 2), device=device, dtype=dtype) idx = torch.tensor((), device=device, dtype=torch.long) src = make_tensor((2, 2), device=device, dtype=dtype) if reduction is not None: actual = fn(dst, 0, idx, src, reduce=reduction) else: actual = fn(dst, 0, idx, src) self.assertEqual(actual, dst, atol=0, rtol=0) @dtypes(torch.float16, torch.float32, torch.complex64) def test_scatter_(self, device, dtype): self._test_scatter_base(torch.Tensor.scatter_, device=device, dtype=dtype, is_scalar=False, reduction=None) @dtypes(torch.float16, torch.float32, torch.complex64) def test_scatter__scalar(self, device, dtype): self._test_scatter_base(torch.Tensor.scatter_, device=device, dtype=dtype, is_scalar=True, reduction=None) # FIXME: RuntimeError: "cuda_scatter_gather_base_kernel_reduce_multiply" not implemented for 'ComplexFloat' @toleranceOverride({torch.float16: tol(atol=1e-2, rtol=0)}) @dtypesIfCUDA(torch.float16, torch.float32) @dtypes(torch.float16, torch.float32, torch.complex64) def test_scatter__reductions(self, device, dtype): for reduction in ("add", "multiply"): self._test_scatter_base(torch.Tensor.scatter_, device=device, dtype=dtype, is_scalar=False, reduction=reduction) self._test_scatter_base(torch.Tensor.scatter_, device=device, dtype=dtype, is_scalar=True, reduction=reduction) @dtypes(torch.float16, torch.float32, torch.complex64) def test_scatter_add_(self, device, dtype): self._test_scatter_base(torch.Tensor.scatter_add_, device=device, dtype=dtype, is_scalar=False, reduction=None) @dtypes(torch.float32) def test_scatter_add_mult_index_base(self, device, dtype): m, n = 30, 40 idx = torch.zeros(m, n, device=device, dtype=torch.long) src = torch.ones(m, n, device=device, dtype=dtype) res0 = torch.zeros(m, n, device=device, dtype=dtype).scatter_add_(0, idx, src) res1 = torch.zeros(m, n, device=device, dtype=dtype).scatter_add_(1, idx, src) self.assertEqual(res0[0, :], m * torch.ones(n, device=device, dtype=dtype), atol=0, rtol=0) self.assertEqual(res1[:, 0], n * torch.ones(m, device=device, dtype=dtype), atol=0, rtol=0) # FIXME: discrepancy between bool ReduceAdd on CUDA and CPU (a + b on CPU and buggy a && b on CUDA) @dtypes(*get_all_dtypes(include_half=True, include_bfloat16=True, include_bool=False)) def test_scatter_reduce_sum(self, device, dtype): for include_self in (True, False): self._test_scatter_base(torch.Tensor.scatter_reduce_, device=device, dtype=dtype, is_scalar=False, reduction='sum', unique_indices=False, include_self=include_self) @dtypes(*get_all_dtypes(include_half=True, include_bfloat16=True)) @dtypesIfCUDA(*get_all_fp_dtypes(include_half=True, include_bfloat16=True)) def test_scatter_reduce_prod(self, device, dtype): for include_self in (True, False): self._test_scatter_base(torch.Tensor.scatter_reduce_, device=device, dtype=dtype, is_scalar=False, reduction='prod', unique_indices=False, include_self=include_self) @dtypes(*get_all_dtypes(include_half=True, include_bfloat16=True, include_bool=False)) @dtypesIfCUDA(*get_all_fp_dtypes(include_half=True, include_bfloat16=True)) def test_scatter_reduce_mean(self, device, dtype): for include_self in (True, False): self._test_scatter_base(torch.Tensor.scatter_reduce_, device=device, dtype=dtype, is_scalar=False, reduction='mean', unique_indices=False, include_self=include_self) @dtypes(*get_all_dtypes(include_half=True, include_bfloat16=True, include_complex=False)) @dtypesIfCUDA(*get_all_fp_dtypes(include_half=True, include_bfloat16=True)) def test_scatter_reduce_amax(self, device, dtype): for include_self in (True, False): self._test_scatter_base(torch.Tensor.scatter_reduce_, device=device, dtype=dtype, is_scalar=False, reduction='amax', unique_indices=False, include_self=include_self) # simple test for nan/inf propagation if (dtype.is_floating_point): input = torch.zeros(3, device=device, dtype=dtype) src = torch.tensor([1, float('nan'), -float('inf'), -float('inf'), 2, float('inf')], device=device, dtype=dtype) idx = torch.tensor([0, 0, 1, 1, 2, 2], device=device) input.scatter_reduce_(0, idx, src, 'amax', include_self=include_self) expected_result = torch.tensor([float('nan'), -float('inf'), float('inf')], device=device, dtype=dtype) if (include_self): expected_result[1] = 0 self.assertEqual(input, expected_result) @dtypes(*get_all_dtypes(include_half=True, include_bfloat16=True, include_complex=False)) @dtypesIfCUDA(*get_all_fp_dtypes(include_half=True, include_bfloat16=True)) def test_scatter_reduce_amin(self, device, dtype): for include_self in (True, False): self._test_scatter_base(torch.Tensor.scatter_reduce_, device=device, dtype=dtype, is_scalar=False, reduction='amin', unique_indices=False, include_self=include_self) # simple test for nan/inf propagation if (dtype.is_floating_point): input = torch.zeros(3, device=device, dtype=dtype) src = torch.tensor([1, float('nan'), -2, -float('inf'), float('inf'), float('inf')], device=device, dtype=dtype) idx = torch.tensor([0, 0, 1, 1, 2, 2], device=device) input.scatter_reduce_(0, idx, src, 'amin', include_self=include_self) expected_result = torch.tensor([float('nan'), -float('inf'), float('inf')], device=device, dtype=dtype) if (include_self): expected_result[2] = 0 self.assertEqual(input, expected_result)