Revert "Fix conv2d strided prologue (#150697)"

pytorchmergebot · pytorchmergebot · commit caf8d9bc1744 · 2025-04-06T04:50:15.000Z
This reverts commit 2e4ae2a. Reverted #150697 on behalf of https://github.com/ngimel due to breaks rocm build ([comment](#150697 (comment)))
diff --git a/test/inductor/test_max_autotune.py b/test/inductor/test_max_autotune.py
@@ -1380,44 +1380,6 @@ def check_code(self, code_str, num_kernels, num_allocs, num_deallocs):
                 "del", num_deallocs, exactly=True
             ).run(code_str)
 
-    @parametrize("prologue", (False, True))
-    def test_conv1x1_cast(self, prologue):
-        with torch._inductor.config.patch(prologue_fusion=prologue):
-            conv1x1 = (
-                torch.nn.Conv2d(in_channels=3, out_channels=16, kernel_size=1)
-                .to(memory_format=torch.channels_last)
-                .to(GPU_TYPE)
-                .to(dtype=torch.float16)
-            )
-            input_tensor = (
-                torch.randn(4, 3, 32, 32)
-                .contiguous(memory_format=torch.channels_last)
-                .to(GPU_TYPE)
-            )
-
-            def foo(mod, input):
-                return torch.nn.functional.conv2d(
-                    input,
-                    mod.weight.to(input.dtype),
-                    None,
-                    mod.stride,
-                    mod.padding,
-                    mod.dilation,
-                    mod.groups,
-                )
-
-            with torch.no_grad():
-                out_eager = foo(conv1x1, input_tensor)
-                foo_c = torch.compile(foo)
-                out, code = run_and_get_code(foo_c, conv1x1, input_tensor)
-
-                FileCheck().check_not("extern_kernels.convolution").run(code[0])
-                if prologue:
-                    self.check_code(
-                        code[0], num_kernels=1, num_allocs=1, num_deallocs=2
-                    )
-                self.assertEqual(out_eager, out, atol=1e-2, rtol=0)
-
     @parametrize("sizes", ((64, 128, 256), (128, 128, 128), (63, 120, 250)))
     def test_upcast(self, sizes):
         M, K, N = sizes
diff --git a/torch/_inductor/select_algorithm.py b/torch/_inductor/select_algorithm.py
@@ -746,10 +746,11 @@ def load_input(
                 indices, self.range_trees[0].construct_entries(lengths)
             ):
                 range_tree_entry.set_name(name)
-
-            strided_index = sympy_dot(input_node.get_stride(), index_symbols)
-            strided_index = self.rename_indexing(strided_index)
-            self.body.writeline("xindex = " + texpr(strided_index))
+            contiguous_index = sympy_dot(
+                ir.FlexibleLayout.contiguous_strides(lengths), index_symbols
+            )
+            contiguous_index = self.rename_indexing(contiguous_index)
+            self.body.writeline("xindex = " + texpr(contiguous_index))
 
             xindex_range_root = self.range_trees[0].lookup(
                 sympy.Integer(1), sympy_product(lengths)
@@ -822,7 +823,7 @@ def store(
 
             output_index = self.rename_indexing(output_index)
 
-            if output_index == strided_index:
+            if output_index == contiguous_index:
                 output_index_str = "xindex"
             else:
                 out_indexing = self.indexing(
