softshrink is expected to return nan when the input is nan on ARC

[daisyden]

🐛 Describe the bug

FAILED test_ops_xpu.py::TestCommonXPU::test_compare_cpu_nn_functional_softshrink_xpu_bfloat16 - AssertionError: Tensor-likes are not close!
FAILED test_ops_xpu.py::TestCommonXPU::test_compare_cpu_nn_functional_softshrink_xpu_float16 - AssertionError: Tensor-likes are not close!
FAILED test_ops_xpu.py::TestCommonXPU::test_compare_cpu_nn_functional_softshrink_xpu_float32 - AssertionError: Tensor-likes are not close!

When input is nan, PVC and CPU return nan while ARC returns 0. However I write a small case and it also returns nan on ARC, looks not a compiler issue. Need more investigations.

Traceback (most recent call last):
File "/root/actions-runner/_work/torch-xpu-ops/pytorch/third_party/torch-xpu-ops/test/xpu/extended/test_ops_xpu.py", line 97, in test_compare_cpu
test_common_test_fn(self.proxy, device, dtype, op)
File "/root/actions-runner/_work/torch-xpu-ops/pytorch/third_party/torch-xpu-ops/test/xpu/../../../../test/test_ops.py", line 414, in test_compare_cpu
self.assertEqual(cuda_results, cpu_results, atol=1e-3, rtol=1e-3)
File "/daisyenv/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4007, in assertEqual
raise error_metas.pop()[0].to_error(
AssertionError: Tensor-likes are not close!

Mismatched elements: 3 / 9 (33.3%)
Greatest absolute difference: nan at index (6,) (up to 0.001 allowed)
Greatest relative difference: nan at index (6,) (up to 0.001 allowed)

#include "sycl/sycl.hpp"
#include <cmath>
#include <iostream>

void tanh_kernel(sycl::queue &q) {
  const size_t local_size = 1;
  const size_t global_size = 1;

  float* nan = (float*)sycl::malloc_shared(sizeof(float), q);
  float* output = (float*)sycl::malloc_shared(sizeof(float), q);
  //*nan = std::numeric_limits<float>::quiet_NaN();
  nan = std::numeric_limits<float>::signaling_NaN();
  output[0] = 555;

  auto e = q.submit([&](sycl::handler& h) {
  sycl::ext::oneapi::experimental::printf(
             "on host: result: %.16f\n",
             *nan);

  h.parallel_for(
      sycl::nd_range<1>{global_size, local_size},
      [=](sycl::nd_item<1> it) {
      output[0] = std::isnan(nan[0]) ? nan[0] : 3;
  });
      });
  e.wait();
  std::cout << "device output " << nan[0] <<  " " << output[0]  << "\n";
}

int main(int argc, char * argv[]) {
  auto devs = sycl::device::get_devices(sycl::info::device_type::gpu);
  auto first = devs[1];

  sycl::device dg1(first);
  sycl::queue q(dg1, sycl::property::queue::enable_profiling());

  tanh_kernel(q);

  return 0;
}

on host: result: nan
device output nan nan

Intel(R) oneAPI DPC++/C++ Compiler 2025.0.1 (2025.0.1.20241113)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/intel/oneapi/compiler/2025.0/bin/compiler
Configuration file: /opt/intel/oneapi/compiler/2025.0/bin/compiler/../icpx.cfg
Found candidate GCC installation: /usr/lib/gcc/x86_64-linux-gnu/13
Selected GCC installation: /usr/lib/gcc/x86_64-linux-gnu/13
Candidate multilib: .;@m64
Selected multilib: .;@m64

Versions

pip install --pre torch==2.6.0.dev20241202+xpu torchvision==0.20.0.dev20241202+xpu torchaudio==2.5.0.dev20241202+xpu --index-url https://download.pytorch.org/whl/nightly/xpu