fix the v100 cuda11.2 matmul_v2 and elementwise_div bug (#42479)

Author: wawltor

paddle/fluid/pybind/pybind.cc

@@ -2972,6 +2972,10 @@ All parameter, weight, gradient are variables in Paddle.
     // Only GPUs with Compute Capability >= 53 support float16
     return platform::GetGPUComputeCapability(place.device) >= 53;
   });
+  m.def("is_bfloat16_supported", [](const platform::CUDAPlace &place) -> bool {
+    // Only GPUs with Compute Capability >= 80 support bfloat16
+    return platform::GetGPUComputeCapability(place.device) >= 80;
+  });
 #endif
 
   m.def("set_feed_variable",

python/paddle/fluid/tests/unittests/test_elementwise_div_op.py

@@ -60,9 +60,9 @@ def init_dtype(self):
         pass
 
 
-@unittest.skipIf(
-    not core.is_compiled_with_cuda() or core.cudnn_version() < 8100,
-    "core is not compiled with CUDA and cudnn version need larger than 8.1.0")
+@unittest.skipIf(not core.is_compiled_with_cuda() or
+                 not core.is_bfloat16_supported(core.CUDAPlace(0)),
+                 "core is not compiled with CUDA and not support the bfloat16")
 class TestElementwiseDivOpBF16(OpTest):
     def setUp(self):
         self.op_type = "elementwise_div"

python/paddle/fluid/tests/unittests/test_matmul_v2_op.py

@@ -385,9 +385,9 @@ def test_check_grad(self):
 
 def create_test_bf16_class(parent, atol=0.01):
     @unittest.skipIf(
-        not core.is_compiled_with_cuda() or core.cudnn_version() < 8100,
-        "core is not compiled with CUDA and cudnn version need larger than 8.1.0"
-    )
+        not core.is_compiled_with_cuda() or
+        not core.is_bfloat16_supported(core.CUDAPlace(0)),
+        "core is not compiled with CUDA and not support the bfloat16")
     class TestMatMulOpBf16Case(parent):
         def get_numeric_grad(self, place, check_name):
             scope = core.Scope()