support some prim ops bf16 dtype (#54263)

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

@@ -542,7 +542,15 @@ void log_grad(const Tensor& x, const Tensor& out_grad, Tensor* x_grad) {
 template <typename T>
 void exp_grad(const Tensor& out, const Tensor& out_grad, Tensor* x_grad) {
   if (x_grad) {
-    set_output<T>(out_grad * out, x_grad);
+    if (out.dtype() == phi::DataType::FLOAT16 ||
+        out.dtype() == phi::DataType::BFLOAT16) {
+      Tensor out_promote = cast<T>(out, phi::DataType::FLOAT32);
+      Tensor out_grad_promote = cast<T>(out_grad, phi::DataType::FLOAT32);
+      set_output<T>(cast<T>(out_promote * out_grad_promote, out.dtype()),
+                    x_grad);
+    } else {
+      set_output<T>(out_grad * out, x_grad);
+    }
   }
 }
 

test/legacy_test/test_activation_op.py

@@ -113,6 +113,7 @@ class TestExpFp32_Prim(OpTest):
         self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
         self.outputs = {'Out': out}
         self.if_enable_cinn()
+        self.convert_input_output()
 
     def test_check_output(self):
         self.check_output()
@@ -129,6 +130,9 @@ class TestExpFp32_Prim(OpTest):
     def if_enable_cinn(self):
         pass
 
+    def convert_input_output(self):
+        pass
+
 
 class TestExpFp64_Prim(TestExpFp32_Prim):
     def init_dtype(self):
@@ -4003,6 +4007,7 @@ def create_test_act_fp16_class(
 
 
 create_test_act_fp16_class(TestActivation)
+create_test_act_fp16_class(TestExpFp32_Prim, check_prim=True, enable_cinn=True)
 create_test_act_fp16_class(TestExpm1)
 create_test_act_fp16_class(TestSigmoid, check_prim=True, enable_cinn=True)
 create_test_act_fp16_class(TestSilu, check_prim=True, enable_cinn=True)
@@ -4133,6 +4138,7 @@ def create_test_act_bf16_class(
 
 
 create_test_act_bf16_class(TestActivation)
+create_test_act_bf16_class(TestExpFp32_Prim, check_prim=True)
 create_test_act_bf16_class(TestExpm1)
 create_test_act_bf16_class(TestSigmoid, check_prim=True)
 create_test_act_bf16_class(TestSilu, check_prim=True)

test/legacy_test/test_cast_op.py

@@ -108,6 +108,9 @@ class TestCastOpBf16ToFp32(OpTest):
         self.prim_op_type = "prim"
         self.python_api = cast_wrapper
         self.public_python_api = cast_wrapper
+        self.if_enable_cinn()
+
+    def if_enable_cinn(self):
         self.enable_cinn = False
 
     def test_check_output(self):
@@ -130,6 +133,9 @@ class TestCastOpFp32ToBf16(OpTest):
         self.prim_op_type = "prim"
         self.python_api = cast_wrapper
         self.public_python_api = cast_wrapper
+        self.if_enable_cinn()
+
+    def if_enable_cinn(self):
         self.enable_cinn = False
 
     def test_check_output(self):

test/legacy_test/test_elementwise_div_op.py

@@ -222,9 +222,8 @@ class TestElementwiseDivOpBF16(ElementwiseDivOp):
                 check_args.insert(0, self.place)
                 self.check_grad_with_place(*check_args, **check_kwargs)
 
-    # elementwise_pow does't support bfloat16
     def if_check_prim(self):
-        self.check_prim = False
+        self.check_prim = True
 
     def if_enable_cinn(self):
         self.enable_cinn = False

test/legacy_test/test_mean_op.py

@@ -211,7 +211,7 @@ class TestReduceMeanBF16Op(OpTest):
         self.axis = [0]
         self.keepdim = False
         self.set_attrs()
-        self.enable_cinn = False
+        self.if_enable_cinn()
 
         np.random.seed(10)
         x_np = np.random.uniform(-1, 1, self.shape).astype(np.float32)
@@ -227,6 +227,9 @@ class TestReduceMeanBF16Op(OpTest):
             'reduce_all': self.reduce_all,
         }
 
+    def if_enable_cinn(self):
+        self.enable_cinn = False
+
     def set_attrs(self):
         pass
 

test/legacy_test/test_softmax_op.py

@@ -398,7 +398,9 @@ class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp):
 class TestSoftmaxBF16Op(OpTest):
     def setUp(self):
         self.op_type = "softmax"
+        self.prim_op_type = "comp"
         self.python_api = F.softmax
+        self.public_python_api = F.softmax
         self.use_cudnn = self.init_cudnn()
         self.use_mkldnn = False
         self.dtype = np.uint16
@@ -424,7 +426,9 @@ class TestSoftmaxBF16Op(OpTest):
 
     def test_check_output(self):
         place = core.CUDAPlace(0)
-        self.check_output_with_place(place, check_dygraph=(not self.use_mkldnn))
+        self.check_output_with_place(
+            place, check_dygraph=(not self.use_mkldnn), check_prim=True
+        )
 
     def test_check_grad(self):
         place = core.CUDAPlace(0)
@@ -434,6 +438,7 @@ class TestSoftmaxBF16Op(OpTest):
             "Out",
             numeric_grad_delta=0.05,
             check_dygraph=(not self.use_mkldnn),
+            check_prim=True,
         )