[Navi] [Inductor] Unskip Navi inductor UTs (#1514)

Relates to https://ontrack-internal.amd.com/browse/SWDEV-461590

Author: jataylo

test/inductor/test_cuda_repro.py

@@ -333,7 +333,6 @@ def foo(x):
         out_ref.add_(2)
         # self.assertEqual(out_ref, out)
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_accuracy_issue1(self):
         class Repro(torch.nn.Module):
             def __init__(self) -> None:
@@ -370,7 +369,6 @@ def forward(self, start_positions: torch.Tensor, x: torch.Tensor):
             assert same_two_models(mod, opt_mod, args), "Dynamo failed"
 
     @config.patch(allow_buffer_reuse=False)
-    @skipIfRocmArch(NAVI_ARCH)
     def test_issue103461(self):
         def forward(add_1):
             var_mean = torch.ops.aten.var_mean.correction(
@@ -869,7 +867,6 @@ def forward(self, x):
             res2 = jit_func(x)
             self.assertEqual(res1, res2)
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_issue103481(self):
         def fn(x, y):
             # NOTE: 6 dimensions is important! does not fail for 5 dimensions

test/inductor/test_torchinductor.py

@@ -1804,7 +1804,6 @@ def fn(x):
             # make sure things also work if they aren't unrolled
             self.common(fn, (torch.randn(8, 3),))
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_multilayer_sum_low_prec(self):
         # fp16 nyi for cpu
         if self.device == "cpu":
@@ -1815,7 +1814,6 @@ def fn(a):
 
         self.common(fn, ((torch.rand((10, 3, 352, 352), dtype=torch.float16),)))
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_multilayer_prime_size(self):
         def fn(a):
             return torch.max(a), torch.sum(a)
@@ -1827,7 +1825,6 @@ def fn(a):
 
     @skip_if_gpu_halide
     @skipCPUIf(IS_MACOS, "fails on macos")
-    @skipIfRocmArch(NAVI_ARCH)
     def test_multilayer_var(self):
         def fn(a):
             return torch.var(a)
@@ -2940,7 +2937,6 @@ def fn(a, b):
         self.common(fn, (torch.randn(8, 8), torch.randn(8, 8)))
 
     @skip_if_halide  # only 32-bit indexing
-    @skipIfRocmArch(NAVI_ARCH)
     def test_large_tensor_reduction(self):
         if not _has_sufficient_memory(self.device, 4.5 * 1024**3):  # 4.5 GiB
             raise unittest.SkipTest("insufficient memory")
@@ -2962,7 +2958,6 @@ def fn(a):
         self.assertEqual(actual, expect)
 
     @skip_if_gpu_halide  # only 32-bit indexing
-    @skipIfRocmArch(NAVI_ARCH)
     def test_large_broadcast_reduction(self):
         if self.device == "cpu":
             raise unittest.SkipTest("Fails on CPU")
@@ -4113,7 +4108,6 @@ def test_conv2d_channels_last(self):
             check_lowp=False,
         )
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_conv2d_backward_channels_last(self):
         def fn(grad_output, inp, weight):
             convolution_backward_8 = torch.ops.aten.convolution_backward.default(
@@ -4899,7 +4893,6 @@ def fn(x, y):
         self.assertEqual(c.stride()[2], 1)
 
     @skip_if_gpu_halide
-    @skipIfRocmArch(NAVI_ARCH)
     def test_std(self):
         def fn(x):
             return (
@@ -4942,7 +4935,6 @@ def test_batch_norm_2d(self):
 
     # From yolov3
     @with_tf32_off
-    @skipIfRocmArch(NAVI_ARCH)
     def test_batch_norm_2d_2(self):
         if self.device == "cpu":
             raise unittest.SkipTest(f"requires {GPU_TYPE}")
@@ -5090,7 +5082,6 @@ def fn(dist, angle):
         self.common(fn, (*inp,))
 
     @skip_if_gpu_halide  # incorrect result on CUDA
-    @skipIfRocmArch(NAVI_ARCH)
     def test_cauchy(self):
         def fn(x, y):
             return torch.sum(1 / (torch.unsqueeze(x, -1) - y))
@@ -6491,7 +6482,6 @@ def fn(a):
         y = fn_compiled(x)
         self.assertTrue(y is not x)
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_l1_loss(self):
         def fn(a, b):
             return torch.nn.functional.l1_loss(a, b), torch.nn.functional.mse_loss(a, b)
@@ -6899,7 +6889,6 @@ def fn(x):
             fn, (torch.tensor([1, float("inf"), 2, float("-inf"), float("nan")]),)
         )
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_any(self):
         def fn(x):
             return (
@@ -7652,7 +7641,6 @@ def fn(a, dim, index, b, reduce):
 
     @skip_if_gpu_halide
     # issue #1150
-    @skipIfRocmArch(NAVI_ARCH)
     def test_dense_mask_index(self):
         r"""
         There will be a little difference for reduce order between aten and inductor
@@ -8662,7 +8650,6 @@ def fn(a, b):
         b = torch.rand(2, 2, 1, 4, 1).int()
         self.common(fn, (a, b))
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_argmax_argmin1(self):
         def fn(x):
             return (aten.argmax(x), aten.argmin(x))
@@ -8674,7 +8661,6 @@ def fn(x):
             ],
         )
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_argmax_argmin2(self):
         def fn(x):
             return (
@@ -8686,7 +8672,6 @@ def fn(x):
 
         self.common(fn, (torch.randn([144, 144]),))
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_argmax_argmin_with_duplicates(self):
         def fn(x):
             return (
@@ -8709,7 +8694,6 @@ def fn(x):
         self.common(fn, (t1,))
 
     @skip_if_halide  # nan behavior
-    @skipIfRocmArch(NAVI_ARCH)
     def test_argmax_argmin_with_nan(self):
         def fn(x):
             return (
@@ -8833,7 +8817,6 @@ def fn(x):
             ],
         )
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_tmp_not_defined_issue1(self):
         def forward(
             primals_3,
@@ -9234,7 +9217,6 @@ def __torch_dispatch__(self, func, types, args=(), kwargs=None):
                 else:
                     self.assertEqual(len(inps), 0)
 
-    @skipIfRocmArch(NAVI_ARCH)
     def test_dtype_mismatch_issue(self):
         def fn(x):
             attn = torch.nn.functional.pad(x, [0, 1])
@@ -12235,7 +12217,6 @@ def test_rnn_compile_safe(self):
 
     class NanCheckerTest(TestCase):
         @config.patch("nan_asserts", True)
-        @skipIfRocmArch(NAVI_ARCH)
         def test_nan_checker_pass(self):
             def f(x):
                 return torch.softmax(x, dim=-1)

test/inductor/test_torchinductor_opinfo.py

@@ -32,7 +32,6 @@
     dtype_abbrs,
     IS_MACOS,
     IS_X86,
-    is_navi_arch,
     skipCUDAMemoryLeakCheckIf,
     skipIfCrossRef,
     skipIfTorchDynamo,
@@ -204,19 +203,6 @@ def format_op(op):
     # Tensors are not alike
     inductor_skips["cuda"]["logcumsumexp"] = {f32}
     inductor_skips["cuda"]["special.modified_bessel_i1"] = {f64}
-    if is_navi_arch():
-        inductor_skips["cuda"]["aminmax"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["dist"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["kron"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["masked.std"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["masked.var"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"][("max", "reduction_no_dim")] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"][("min", "reduction_no_dim")] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["nn.functional.conv_transpose3d"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["std"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["std_mean"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["var"] = {b8, f16, f32, f64, i32, i64}
-        inductor_skips["cuda"]["var_mean"] = {b8, f16, f32, f64, i32, i64}
 
 inductor_expected_failures_single_sample = defaultdict(dict)