【Hackathon No.55】add erf FP16 test and BF16 test (#52136)

* add erf FP16 test

paddle/phi/kernels/funcs/eigen/erf.cc

@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+#include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 #include "paddle/phi/kernels/funcs/eigen/extensions.h"
@@ -46,10 +47,11 @@ struct EigenErfGrad<Eigen::DefaultDevice, T> {
   }
 };
 
-#define INSTANTIATION(FUNCTOR)                           \
-  template struct FUNCTOR<Eigen::DefaultDevice, float>;  \
-  template struct FUNCTOR<Eigen::DefaultDevice, double>; \
-  template struct FUNCTOR<Eigen::DefaultDevice, dtype::float16>
+#define INSTANTIATION(FUNCTOR)                                   \
+  template struct FUNCTOR<Eigen::DefaultDevice, float>;          \
+  template struct FUNCTOR<Eigen::DefaultDevice, double>;         \
+  template struct FUNCTOR<Eigen::DefaultDevice, dtype::float16>; \
+  template struct FUNCTOR<Eigen::DefaultDevice, dtype::bfloat16>
 INSTANTIATION(EigenErf);
 INSTANTIATION(EigenErfGrad);
 #undef INSTANTIATION

paddle/phi/kernels/funcs/eigen/erf.cu

@@ -14,6 +14,7 @@ limitations under the License. */
 #ifndef _USE_MATH_DEFINES
 #define _USE_MATH_DEFINES
 #endif
+#include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 #include "paddle/phi/kernels/funcs/eigen/extensions.h"
@@ -47,10 +48,11 @@ struct EigenErfGrad<Eigen::GpuDevice, T> {
   }
 };
 
-#define INSTANTIATION(FUNCTOR)                       \
-  template struct FUNCTOR<Eigen::GpuDevice, float>;  \
-  template struct FUNCTOR<Eigen::GpuDevice, double>; \
-  template struct FUNCTOR<Eigen::GpuDevice, dtype::float16>
+#define INSTANTIATION(FUNCTOR)                               \
+  template struct FUNCTOR<Eigen::GpuDevice, float>;          \
+  template struct FUNCTOR<Eigen::GpuDevice, double>;         \
+  template struct FUNCTOR<Eigen::GpuDevice, dtype::float16>; \
+  template struct FUNCTOR<Eigen::GpuDevice, dtype::bfloat16>
 INSTANTIATION(EigenErf);
 INSTANTIATION(EigenErfGrad);
 #undef INSTANTIATION

paddle/phi/kernels/gpu/erf_grad_kernel.cu

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/phi/kernels/erf_grad_kernel.h"
 
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/erf_grad_kernel_impl.h"
@@ -25,4 +26,5 @@ PD_REGISTER_KERNEL(erf_grad,
                    phi::ErfGradKernel,
                    float,
                    double,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/erf_kernel.cu

@@ -15,9 +15,16 @@ limitations under the License. */
 #include "paddle/phi/kernels/erf_kernel.h"
 
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/erf_kernel_impl.h"
 
-PD_REGISTER_KERNEL(
-    erf, GPU, ALL_LAYOUT, phi::ErfKernel, float, double, phi::dtype::float16) {}
+PD_REGISTER_KERNEL(erf,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::ErfKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

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

@@ -15,7 +15,7 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 from scipy.special import erf
 
 import paddle
@@ -68,5 +68,54 @@ class TestErfLayer(unittest.TestCase):
             self.assertTrue('erf' in y.name)
 
 
+class TestErfFP16OP(OpTest):
+    def setUp(self):
+        self.op_type = "erf"
+        self.prim_op_type = "prim"
+        self.public_python_api = paddle.erf
+        self.python_api = paddle.erf
+        self.dtype = np.float16
+        self.x_shape = [11, 17]
+        x = np.random.uniform(-1, 1, size=self.x_shape).astype(self.dtype)
+        y_ref = erf(x).astype(self.dtype)
+        self.inputs = {'X': x}
+        self.outputs = {'Out': y_ref}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out', check_prim=True)
+
+
+@unittest.skipIf(
+    not paddle.fluid.core.is_compiled_with_cuda()
+    or not paddle.fluid.core.is_bfloat16_supported(
+        paddle.fluid.core.CUDAPlace(0)
+    ),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestErfBF16OP(OpTest):
+    def setUp(self):
+        self.op_type = "erf"
+        self.prim_op_type = "prim"
+        self.public_python_api = paddle.erf
+        self.python_api = paddle.erf
+        self.dtype = np.uint16
+        self.x_shape = [11, 17]
+        x = np.random.uniform(-1, 1, size=self.x_shape).astype(np.float32)
+        y_ref = erf(x).astype(np.float32)
+        self.inputs = {'X': convert_float_to_uint16(x)}
+        self.outputs = {'Out': convert_float_to_uint16(y_ref)}
+
+    def test_check_output(self):
+        place = paddle.fluid.core.CUDAPlace(0)
+        self.check_output_with_place(place)
+
+    def test_check_grad(self):
+        place = paddle.fluid.core.CUDAPlace(0)
+        self.check_grad_with_place(place, ['X'], 'Out', check_prim=True)
+
+
 if __name__ == '__main__':
     unittest.main()