Address code review comments

Signed-off-by: Tiotto, Ettore <[email protected]>

Author: etiotto

python/test/unit/language/test_block_pointer.py

@@ -7,53 +7,51 @@
 
 
 @triton.jit
-def block_copy_kernel(a_ptr, b_ptr, N, BLOCK_SIZE: tl.constexpr):
+def block_copy_kernel(a_ptr, b_ptr, N, BLOCK_SIZE: tl.constexpr, padding_option: tl.constexpr):
     pid = tl.program_id(0)
     # We only copy half of the data to see if the padding works
     a_block_ptr = tl.make_block_ptr(base=a_ptr, shape=(N // 2, ), strides=(1, ), offsets=(pid * BLOCK_SIZE, ),
                                     block_shape=(BLOCK_SIZE, ), order=(0, ))
     b_block_ptr = tl.make_block_ptr(base=b_ptr, shape=(N, ), strides=(1, ), offsets=(pid * BLOCK_SIZE, ),
                                     block_shape=(BLOCK_SIZE, ), order=(0, ))
-    #    if padding_option is None:
-    a = tl.load(a_block_ptr, boundary_check=(0, ))
-    #    else:
-    #        a = tl.load(a_block_ptr, boundary_check=(0, ), padding_option=padding_option)
+    if padding_option is None:
+        a = tl.load(a_block_ptr, boundary_check=(0, ))
+    else:
+        a = tl.load(a_block_ptr, boundary_check=(0, ), padding_option=padding_option)
     tl.store(b_block_ptr, a, boundary_check=(0, ))
 
 
 @pytest.mark.interpreter
-@pytest.mark.parametrize("dtypes_str, n", [  #
-    (dtypes_str, n)
-    #    for dtypes_str in (("bool", "bool"), ("int16", "int16"), ("int32", "int32"), ("float16", "float16"),
-    #                       ("float32", "float32"), ("bfloat16", "bfloat16"))
-    for dtypes_str in [("float16", "float16")]
-    for n in [64]
+@pytest.mark.parametrize("dtypes_str, n, padding_option", [  #
+    (dtypes_str, n, padding)
+    for dtypes_str in (("bool", "bool"), ("int16", "int16"), ("int32", "int32"), ("float16", "float16"),
+                       ("float32", "float32"), ("bfloat16", "bfloat16"))
+    for n in (64, 128, 256, 512, 1024)
+    for padding in (None, "zero", "nan")  #
 ])
-def test_block_copy(dtypes_str, n, device):
+def test_block_copy(dtypes_str, n, padding_option, device):
     src_dtype_str = dtypes_str[0]
     dst_dtype_str = dtypes_str[1]
     src_dtype = getattr(torch, src_dtype_str)
     dst_dtype = getattr(torch, dst_dtype_str)
     check_type_supported(src_dtype, device)
     check_type_supported(dst_dtype, device)
     if src_dtype_str in ("bool", "int16", "int32"):
-        #        if padding_option == "nan":
-        #            pytest.xfail("Padding with NaN is not supported for integer types")
+        if padding_option == "nan":
+            pytest.xfail("Padding with NaN is not supported for integer types")
         a = torch.randint(0, 2, (n, ), device=device, dtype=src_dtype)
     else:
         a = torch.randn((n, ), device=device, dtype=src_dtype)
     b = torch.zeros((n, ), device=device, dtype=dst_dtype)
 
     grid = lambda meta: (triton.cdiv(n, meta["BLOCK_SIZE"]), )
-    block_copy_kernel[grid](a_ptr=a, b_ptr=b, N=n, BLOCK_SIZE=64)
+    block_copy_kernel[grid](a_ptr=a, b_ptr=b, N=n, BLOCK_SIZE=64, padding_option=padding_option)
     a.to(dst_dtype)
     assert torch.all(a[0:n // 2] == b[0:n // 2])
-
-
-#    if padding_option == "zero":
-#        assert torch.all(b[n // 2:n] == 0)
-#    elif padding_option == "nan":
-#        assert torch.all(torch.isnan(b[n // 2:n]))
+    if padding_option == "zero":
+        assert torch.all(b[n // 2:n] == 0)
+    elif padding_option == "nan":
+        assert torch.all(torch.isnan(b[n // 2:n]))
 
 
 @triton.jit