Comment why we ignore some spurious matmul RuntimeWarnings on macOS arm64

Author: FrancescAlted

bench/ndarray/matmul_path_compare.py

@@ -69,6 +69,7 @@ def run_case(
 ):
     a, b, a_np, b_np = build_arrays(shape_a, shape_b, dtype, chunks_a, chunks_b, blocks_a, blocks_b)
     with warnings.catch_warnings():
+        # NumPy + Accelerate can emit spurious matmul RuntimeWarnings on macOS arm64.
         warnings.simplefilter("ignore", RuntimeWarning)
         expected = np.matmul(a_np, b_np)
     original_flag = set_path_mode(mode)
@@ -87,6 +88,7 @@ def wrapped_set_pref_matmul(self, inputs, fp_accuracy):
             before = len(selected_paths)
             t0 = time.perf_counter()
             with warnings.catch_warnings():
+                # NumPy + Accelerate can emit spurious matmul RuntimeWarnings on macOS arm64.
                 warnings.simplefilter("ignore", RuntimeWarning)
                 result = blosc2.matmul(a, b, chunks=chunks_out, blocks=blocks_out)
             times.append(time.perf_counter() - t0)

tests/ndarray/test_linalg.py

@@ -108,6 +108,7 @@ def test_matmul_uses_fast_path_for_supported_2d(monkeypatch, dtype):
         b = blosc2.full(shape=(400, 400), fill_value=2, dtype=dtype, chunks=(200, 200), blocks=(100, 100))
 
         with warnings.catch_warnings():
+            # NumPy + Accelerate can emit spurious matmul RuntimeWarnings on macOS arm64.
             warnings.simplefilter("ignore", RuntimeWarning)
             c = blosc2.matmul(a, b, chunks=(200, 200), blocks=(100, 100))
             expected = np.matmul(a[:], b[:])