Add a new API torch.xpu.is_tf32_supported for Intel GPU (pytorch#163141)

# Motivation
Aligned with other backends, this PR introduces a new API `torch.xpu.is_tf32_supported`, which should be used before `torch.backends.mkldnn.allow_tf32=True` or provide hardware capability information to the Triton

# Additional Context
On Intel Xe architecture and newer, TF32 operations can be accelerated through DPAS (Dot Product Accumulate Systolic) instructions. Therefore, TF32 support can be determined by checking whether the device supports subgroup matrix multiply-accumulate operations.
Pull Request resolved: pytorch#163141
Approved by: https://github.com/EikanWang

Author: guangyey

docs/source/xpu.md

@@ -28,6 +28,7 @@
     is_available
     is_bf16_supported
     is_initialized
+    is_tf32_supported
     set_device
     set_stream
     stream

test/test_xpu.py

@@ -776,6 +776,10 @@ def test_is_bf16_supported(self):
             torch.xpu.is_available(),
         )
 
+    def test_is_tf32_supported(self):
+        if not torch.xpu.is_available():
+            self.assertFalse(torch.xpu.is_tf32_supported())
+
     def test_get_arch_list(self):
         if not torch.xpu._is_compiled():
             self.assertEqual(len(torch.xpu.get_arch_list()), 0)

torch/xpu/__init__.py

@@ -78,6 +78,17 @@ def is_bf16_supported(including_emulation: bool = True) -> bool:
     )
 
 
+def is_tf32_supported() -> bool:
+    r"""Return a bool indicating if the current XPU device supports dtype tf32."""
+    if not is_available():
+        return False
+    # On Intel Xe architecture and newer, TF32 operations can be accelerated
+    # through DPAS (Dot Product Accumulate Systolic) instructions. Therefore,
+    # TF32 support can be determined by checking whether the device supports
+    # subgroup matrix multiply-accumulate operations.
+    return torch.xpu.get_device_properties().has_subgroup_matrix_multiply_accumulate
+
+
 def is_initialized():
     r"""Return whether PyTorch's XPU state has been initialized."""
     return _initialized and not _is_in_bad_fork()
@@ -559,6 +570,7 @@ def _get_rng_state_offset(device: Union[int, str, torch.device] = "xpu") -> int:
     "is_available",
     "is_bf16_supported",
     "is_initialized",
+    "is_tf32_supported",
     "manual_seed",
     "manual_seed_all",
     "max_memory_allocated",