fix eval branch of prim vjp of batch_norm in amp mode (#53598)

Author: cyber-pioneer

paddle/fluid/prim/api/composite_backward/composite_backward_api.h

@@ -1699,6 +1699,9 @@ void batch_norm_grad(const Tensor& x,
         if (use_global_stats) {
           auto nhwc_x_grad = scale * rsqrt_var * nhwc_out_grad;
           auto nchw_x_grad = transpose<T>(nhwc_x_grad, nhwc_to_nchw_dim);
+          if (x.dtype() == phi::DataType::FLOAT16) {
+            nchw_x_grad = cast<T>(nchw_x_grad, x.dtype());
+          }
           set_output<T>(nchw_x_grad, x_grad);
         } else {
           auto part1 = scale * rsqrt_var;
@@ -1732,6 +1735,9 @@ void batch_norm_grad(const Tensor& x,
             sum<T>(out_grad_data * x_sub_mean, reduce_axis, dtype, false);
         if (use_global_stats) {
           auto x_grad_data = scale * rsqrt_var * out_grad_data;
+          if (x.dtype() == phi::DataType::FLOAT16) {
+            x_grad_data = cast<T>(x_grad_data, x.dtype());
+          }
           set_output<T>(x_grad_data, x_grad);
         } else {
           auto part1 = scale * rsqrt_var;

test/prim/composite_ops/test_composite_batch_norm.py

@@ -386,10 +386,12 @@ def apply_to_static(net, use_cinn):
 
 
 class PrimeNet(paddle.nn.Layer):
-    def __init__(self, data_layout='NCHW'):
+    def __init__(self, data_layout='NCHW', is_test=False):
         super().__init__()
         self.conv = nn.Conv2D(2, 4, (3, 3), bias_attr=False)
-        self.bn = BatchNorm(4, act="relu", data_layout=data_layout)
+        self.bn = BatchNorm(
+            4, act="relu", data_layout=data_layout, is_test=is_test
+        )
 
     def forward(self, x):
         y = self.conv(x)
@@ -408,10 +410,10 @@ def setUp(self):
         self.x = paddle.randn([4, 2, 6, 6], dtype="float32")
         self.x.stop_gradient = False
 
-    def train(self, use_prim, data_layout="NCHW"):
+    def train(self, use_prim, data_layout="NCHW", is_test=False):
         core._set_prim_all_enabled(use_prim)
         paddle.seed(2022)
-        net = PrimeNet(data_layout)
+        net = PrimeNet(data_layout=data_layout, is_test=is_test)
         sgd = paddle.optimizer.SGD(
             learning_rate=0.1, parameters=net.parameters()
         )
@@ -429,8 +431,19 @@ def train(self, use_prim, data_layout="NCHW"):
 
     def test_amp_nchw(self):
         if not isinstance(framework._current_expected_place(), core.CPUPlace):
-            expected = self.train(False)
-            actual = self.train(True)
+            expected = self.train(use_prim=False)
+            actual = self.train(use_prim=True)
+            np.testing.assert_allclose(
+                expected,
+                actual,
+                rtol=1e-3,
+                atol=1e-3,
+            )
+
+    def test_amp_nchw_eval(self):
+        if not isinstance(framework._current_expected_place(), core.CPUPlace):
+            expected = self.train(use_prim=False, is_test=True)
+            actual = self.train(use_prim=True, is_test=True)
             np.testing.assert_allclose(
                 expected,
                 actual,
@@ -449,6 +462,19 @@ def test_amp_nhwc(self):
                 atol=1e-3,
             )
 
+    def test_amp_nhwc_eval(self):
+        if not isinstance(framework._current_expected_place(), core.CPUPlace):
+            expected = self.train(
+                use_prim=False, data_layout="NHWC", is_test=True
+            )
+            actual = self.train(use_prim=True, data_layout="NHWC", is_test=True)
+            np.testing.assert_allclose(
+                expected,
+                actual,
+                rtol=1e-3,
+                atol=1e-3,
+            )
+
 
 class TestPrimEvalBranch(unittest.TestCase):
     """