Fix Windows cpu_profiling_allocator_test same pointer check flakiness (#92264)
This is a small follow-up from https://github.com/pytorch/pytorch/pull/91727 to fix the flaky same pointer check on Windows https://hud.pytorch.org/failure/%5B%20%20FAILED%20%20%5D%20CPUAllocationPlanTest.with_profiling_alloc. AFAICT, keeping the same memory pointer is not a guarantee in non-mobile memory allocator (or may be this is Windows-specific behavior).
The test might be flaky when the tensor is copied to a different memory location with the default allocator. This's ok as long as the values remain equal.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92264
Approved by: https://github.com/ZainRizvi
diff --git a/aten/src/ATen/test/cpu_profiling_allocator_test.cpp b/aten/src/ATen/test/cpu_profiling_allocator_test.cpp
index 1d69ca4..73f69d7 100644
--- a/aten/src/ATen/test/cpu_profiling_allocator_test.cpp
+++ b/aten/src/ATen/test/cpu_profiling_allocator_test.cpp
@@ -164,6 +164,12 @@
// NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
ASSERT_NO_THROW(validate_allocation_plan(false, false, false));
ASSERT_TRUE(ref_output.equal(output));
+
+ // Returning the same pointers is not a guarantee when the default
+ // allocator is used. It looks like the underlying memory pointer
+ // can change as long as output and ref_output remain equal. This
+ // has already been confirmed in the previous two tests
+ #ifdef C10_MOBILE
// Furthermore profiling allocator should return the same pointers
// back for the intermediate tensors
// NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
@@ -173,7 +179,6 @@
ASSERT_NO_THROW(validate_allocation_plan(false, false, true));
ASSERT_TRUE(ref_output.equal(output));
- #ifdef C10_MOBILE
// When control flow conditions are different between profiling and evaluation
// profiling allocator should throw.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-goto,hicpp-avoid-goto)
