gpu maximum & minimum kernel with fp16 input

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/broadcast_impl.cu

@@ -202,7 +202,8 @@ void ElewiseArith(const int &nums, enum BroadcastOpType op, const T *x0, const T
 template <>
 void ElewiseArith(const int &nums, enum BroadcastOpType op, const half *x0, const half *x1, half *y,
                   cudaStream_t stream) {
-  if (nums % 2 == 0) {
+  // `>` return true iff both half result are true. fallback to half
+  if (nums % 2 == 0 && op != BROADCAST_TYPE_MINIMUM && op != BROADCAST_TYPE_MAXIMUM && op != BROADCAST_TYPE_ABSGRAD) {
     ElewiseArithKernel<half2>(nums / 2, op, reinterpret_cast<const half2 *>(x0), reinterpret_cast<const half2 *>(x1),
                               reinterpret_cast<half2 *>(y), stream);
   } else {

tests/st/ops/gpu/test_broadcast_op.py

@@ -68,6 +68,48 @@ def test_nobroadcast():
     assert np.allclose(output_ms.asnumpy(), output_np)
 
 
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_nobroadcast_fp16():
+    context.set_context(mode=context.GRAPH_MODE, device_target='GPU')
+
+    x1_np = np.random.rand(10, 20).astype(np.float16)
+    x2_np = np.random.rand(10, 20).astype(np.float16)
+
+    output_ms = P.Minimum()(Tensor(x1_np), Tensor(x2_np))
+    output_np = np.minimum(x1_np, x2_np)
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Maximum()(Tensor(x1_np), Tensor(x2_np))
+    output_np = np.maximum(x1_np, x2_np)
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Greater()(Tensor(x1_np), Tensor(x2_np))
+    output_np = x1_np > x2_np
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Less()(Tensor(x1_np), Tensor(x2_np))
+    output_np = x1_np < x2_np
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Pow()(Tensor(x1_np), Tensor(x2_np))
+    output_np = np.power(x1_np, x2_np)
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.RealDiv()(Tensor(x1_np), Tensor(x2_np))
+    output_np = x1_np / x2_np
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Mul()(Tensor(x1_np), Tensor(x2_np))
+    output_np = x1_np * x2_np
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+    output_ms = P.Sub()(Tensor(x1_np), Tensor(x2_np))
+    output_np = x1_np - x2_np
+    assert np.allclose(output_ms.asnumpy(), output_np)
+
+
 @pytest.mark.level0
 @pytest.mark.platform_x86_gpu_training
 @pytest.mark.env_onecard