`TestRotatedBoxIou::test_iou[xyxyxyxy-float32]` fails on arm64 Linux due to float32 precision in format round-trip

🐛 Describe the bug

Description

test_iou[xyxyxyxy-dtype0-cpu] fails on arm64 (aarch64) Linux with float32 dtype. The test passes on x86_64 and ppc64el.

The existing tolerance handling in the test accounts for macOS (atol=0.5) and CUDA (xfail), but the else fallback uses atol=1e-4 which is too tight for arm64 Linux where the observed absolute difference is 0.5058.

log.gz

Error

FAILED test_ops.py::TestRotatedBoxIou::test_iou[xyxyxyxy-dtype0-cpu]

Mismatched elements: 2 / 64 (3.1%)
Greatest absolute difference: 0.5057777166366577 at index (3, 0) (up to 0.0001 allowed)
Greatest relative difference: 0.5057777166366577 at index (3, 0) (up to 0.0001 allowed)

The cause

Guessed cause is that torch.cos and torch.sin on tensors use architecture-specific SIMD implementation. The Sleef NEON backend (arm64) produces slightly different float32 rounding than the AVX backend (x86_64), similar to the already handled macOS arm64 case which uses Apple's vDSP implementation.

Versions

PyTorch version: 2.12.0+debian
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux forky/sid (aarch64)
GCC version: (Debian 15.2.0-17) 15.2.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.42

Python version: 3.13.12 (main, Feb 4 2026, 15:06:39) [GCC 15.2.0] (64-bit runtime)
Python platform: Linux-6.12.88+deb13-arm64-aarch64-with-glibc2.42
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture: aarch64
CPU op-mode(s): 64-bit
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
...
Flags: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma dcpop asimddp asimdfhm ssbs
L1d cache: 4 MiB (64 instances)
L1i cache: 4 MiB (64 instances)
L2 cache: 32 MiB (64 instances)
L3 cache: 64 MiB (2 instances)

Versions of relevant libraries:
[pip3] numpy==2.4.4
[pip3] torch==2.12.0+debian
[pip3] torchvision==0.27.0
[conda] Could not collect

Start with test_ops.py and the TestRotatedBoxIou::test_iou[xyxyxyxy-dtype0-cpu] case, then review its existing macOS, CUDA, and fallback tolerance handling. Reproduce or inspect the arm64 failure and update the test expectations so the float32 case is valid on arm64 Linux while existing platforms retain their behavior; the test should pass on the reported configuration.