move sign kernel impl (#38363)

paddle/fluid/operators/sign_op.h

@@ -19,9 +19,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/operators/eigen/eigen_function.h"
 
-// only can include the headers in paddle/pten/api dirs
 #include "paddle/pten/include/core.h"
-#include "paddle/pten/include/math.h"
+#include "paddle/pten/kernels/sign_kernel.h"
 
 namespace paddle {
 namespace operators {

paddle/pten/api/lib/kernel_declare.h

@@ -22,13 +22,13 @@ limitations under the License. */
 
 PT_DECLARE_KERNEL(matmul, CPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(cast, CPU, ALL_LAYOUT);
-PT_DECLARE_KERNEL(sign, CPU, ALL_LAYOUT);
+PT_DECLARE_KERNEL(mean, CPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(conj, CPU, ALL_LAYOUT);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 PT_DECLARE_KERNEL(matmul, GPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(cast, GPU, ALL_LAYOUT);
-PT_DECLARE_KERNEL(sign, GPU, ALL_LAYOUT);
+PT_DECLARE_KERNEL(mean, GPU, ALL_LAYOUT);
 PT_DECLARE_KERNEL(conj, GPU, ALL_LAYOUT);
 #endif
 

paddle/pten/kernels/cpu/math.cc

@@ -17,7 +17,6 @@
 #include "paddle/pten/api/ext/dispatch.h"
 #include "paddle/pten/kernels/hybird/cpu/elementwise.h"
 #include "paddle/pten/kernels/hybird/eigen/reduce.h"
-#include "paddle/pten/kernels/hybird/eigen/sign.h"
 #include "paddle/pten/kernels/hybird/general/elementwise_functor.h"
 #include "paddle/pten/kernels/hybird/general/reduce_impl.h"
 
@@ -28,11 +27,6 @@
 
 namespace pten {
 
-template <typename T>
-void Sign(const CPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out) {
-  eigen::Sign<CPUContext, T>(dev_ctx, x, out);
-}
-
 template <typename T>
 void Mean(const CPUContext& dev_ctx,
           const DenseTensor& x,
@@ -97,7 +91,6 @@ using complex128 = ::paddle::platform::complex<double>;
 
 // NOTE(chenweihang): using bfloat16 will cause redefine with xpu bfloat16
 // using bfloat16 = ::paddle::platform::bfloat16;
-PT_REGISTER_KERNEL(sign, CPU, ALL_LAYOUT, pten::Sign, float, double) {}
 PT_REGISTER_KERNEL(mean, CPU, ALL_LAYOUT, pten::Mean, float, double, bool) {}
 PT_REGISTER_KERNEL(add,
                    CPU,

paddle/pten/kernels/cpu/math.h

@@ -21,9 +21,6 @@ limitations under the License. */
 
 namespace pten {
 
-template <typename T>
-void Sign(const CPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out);
-
 template <typename T>
 void Mean(const CPUContext& dev_ctx,
           const DenseTensor& x,

paddle/pten/kernels/cpu/sign_kernel.cc

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+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/pten/kernels/sign_kernel.h"
+#include "paddle/pten/kernels/impl/sign_kernel_impl.h"
+
+#include "paddle/pten/backends/cpu/cpu_context.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/fluid/platform/bfloat16.h"
+
+PT_REGISTER_CTX_KERNEL(sign, CPU, ALL_LAYOUT, pten::Sign, float, double) {}

paddle/pten/kernels/gpu/math.cu

@@ -17,7 +17,6 @@ limitations under the License. */
 #include "paddle/fluid/operators/reduce_ops/reduce_functor_op.h"
 #include "paddle/pten/kernels/hybird/cuda/elementwise/elementwise.h"
 #include "paddle/pten/kernels/hybird/cuda/reduce/reduce.h"
-#include "paddle/pten/kernels/hybird/eigen/sign.h"
 #include "paddle/pten/kernels/hybird/general/elementwise_functor.h"
 #include "paddle/pten/kernels/hybird/general/reduce_impl.h"
 
@@ -57,11 +56,6 @@ struct DivideFunctor {
  * Kernels
  */
 
-template <typename T>
-void Sign(const GPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out) {
-  eigen::Sign<GPUContext, T>(dev_ctx, x, out);
-}
-
 template <typename T>
 void Mean(const GPUContext& dev_ctx,
           const DenseTensor& x,
@@ -101,7 +95,6 @@ using float16 = paddle::platform::float16;
 using complex64 = ::paddle::platform::complex<float>;
 using complex128 = ::paddle::platform::complex<double>;
 
-PT_REGISTER_KERNEL(sign, GPU, ALL_LAYOUT, pten::Sign, float, double, float16) {}
 PT_REGISTER_KERNEL(mean, GPU, ALL_LAYOUT, pten::Mean, float, double, bool) {}
 PT_REGISTER_KERNEL(add,
                    GPU,

paddle/pten/kernels/gpu/math.h

@@ -23,9 +23,6 @@ limitations under the License. */
 
 namespace pten {
 
-template <typename T>
-void Sign(const GPUContext& dev_ctx, const DenseTensor& x, DenseTensor* out);
-
 template <typename T>
 void Mean(const GPUContext& dev_ctx,
           const DenseTensor& x,

paddle/pten/kernels/gpu/sign_kernel.cu

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+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/pten/kernels/impl/sign_kernel_impl.h"
+#include "paddle/pten/kernels/sign_kernel.h"
+
+#include "paddle/pten/backends/gpu/gpu_context.h"
+#include "paddle/pten/core/kernel_registry.h"
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/fluid/platform/float16.h"
+
+using float16 = paddle::platform::float16;
+
+PT_REGISTER_CTX_KERNEL(
+    sign, GPU, ALL_LAYOUT, pten::Sign, float, double, float16) {}

paddle/pten/kernels/impl/sign_kernel_impl.h

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+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. */
+
+#pragma once
+
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/kernels/hybird/eigen/common.h"
+
+// See Note [ Why still include the fluid headers? ]
+#include "paddle/fluid/operators/eigen/eigen_function.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+void Sign(const ContextT& dev_ctx, const DenseTensor& x, DenseTensor* out) {
+  out->mutable_data<T>();
+  auto eigen_out = pten::EigenVector<T>::Flatten(*out);
+  auto eigen_x = pten::EigenVector<T>::Flatten(x);
+
+  auto& dev = *dev_ctx.eigen_device();
+  paddle::operators::EigenSign<std::decay_t<decltype(dev)>, T>::Eval(
+      dev, eigen_out, eigen_x);
+}
+
+}  // namespace pten

paddle/pten/kernels/sign_kernel.h

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+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. */
+
+#pragma once
+
+#include "paddle/pten/core/dense_tensor.h"
+
+namespace pten {
+
+template <typename T, typename ContextT>
+void Sign(const ContextT& dev_ctx, const DenseTensor& x, DenseTensor* out);
+
+}  // namespace pten