test(op): The zeros operator adds an ub overflow constraint and fix matmul of Floor Division

Author: wangzhanpeng5

ascend/examples/generalization_cases/test_matmul.py

@@ -61,7 +61,7 @@ def test_matmul(shape, dtype):
     # bisheng not support yet
     if M % 16 != 0 or N % 16 != 0 or get_dtype_size(dtype) * K % 32 != 0:
         return
-    kalign = 32 / get_dtype_size(dtype)  # 32byte/Dtype_bytes
+    kalign = 32 // get_dtype_size(dtype)  # 32byte/Dtype_bytes
     BLOCK_M, BLOCK_N, BLOCK_K = min(max(M, 16), 32), min(max(N, 16), 32), min(max(K, kalign), 32)
     a = test_common.generate_tensor((M, K), dtype)
     b = test_common.generate_tensor((K, N), dtype)

ascend/examples/generalization_cases/test_zeros_op.py

@@ -519,7 +519,8 @@ def fn_npu_multi_d(output_ptr, XB: tl.constexpr, YB: tl.constexpr, ZB: tl.conste
                          )
 def test_case_4d_5d(param_list):
     dtype, shape = param_list
-
+    if check_ub_mem_overflow(sigtype, shape):
+        pytest.skip(f"dtype:{sigtype} shape:{shape} mem overflow")
     y_ref = torch.full(shape, 0, dtype=eval('torch.' + dtype)).npu()
     print(f"y_ref = {torch.flatten(y_ref)[0:4]}")
 

ascend/examples/generalization_cases/test_zeroslike.py

@@ -136,6 +136,8 @@ def fn_npu_multi_d(output_ptr, x_ptr, XB: tl.constexpr, YB: tl.constexpr, ZB: tl
                          )
 def test_case_4d_5d(param_list):
     dtype, shape = param_list
+    if check_ub_mem_overflow(dtype, shape):
+        return
     x0 = test_common.generate_tensor(shape, dtype)
     y_ref = torch.zeros_like(x0, dtype=eval('torch.' + dtype)).npu()
     print(f"y_ref = {torch.flatten(y_ref)[0:4]}")