[pten]rm reduce_sum and reduce_mean raw kernel (#39484)

* rm reduce_sum raw kernel

* remove reduce_mean kernel

* remove reduce_mean kernel

* reduce support int and int64_t

* mean support int and int64_t type

paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc

@@ -176,4 +176,4 @@ TEST(Benchmark, EagerIntermediateMLPCPU) {
 USE_OP_ITSELF(scale);
 USE_OP_ITSELF(elementwise_add);
 USE_OP_ITSELF(matmul_v2);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_sum);

paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc

@@ -185,7 +185,7 @@ TEST(Benchmark, EagerIntermediateMLPCUDA) {
 
 USE_OP_ITSELF(scale);
 USE_OP_ITSELF(matmul_v2);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_sum);
 USE_OP(reduce_sum_grad);
 USE_OP_ITSELF(elementwise_add);
 

paddle/fluid/eager/tests/performance_tests/benchmark_fluid_cpu.cc

@@ -214,4 +214,4 @@ TEST(Benchmark, FluidMLPCPU) {
 USE_OP_ITSELF(scale);
 USE_OP_ITSELF(elementwise_add);
 USE_OP_ITSELF(matmul_v2);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_sum);

paddle/fluid/eager/tests/performance_tests/benchmark_fluid_cuda.cc

@@ -247,7 +247,7 @@ TEST(Benchmark, FluidMLPCUDA) {
 
 USE_OP_ITSELF(scale);
 USE_OP_ITSELF(matmul_v2);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_sum);
 USE_OP(reduce_sum_grad);
 USE_OP_ITSELF(elementwise_add);
 

paddle/fluid/framework/new_executor/standalone_executor_test.cc

@@ -35,8 +35,8 @@ USE_OP(sigmoid);
 USE_OP(tanh);
 USE_OP(elementwise_mul);
 USE_OP(softmax_with_cross_entropy);
-USE_OP(reduce_mean);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_mean);
+USE_OP_ITSELF(reduce_sum);
 USE_OP(reduce_sum_grad);
 USE_OP(reduce_mean_grad);
 USE_OP_ITSELF(reshape2_grad);

paddle/fluid/imperative/tests/test_tracer.cc

@@ -590,6 +590,6 @@ TEST(test_tracer, eager_tracer) {
 
 USE_OP(mul);
 USE_OP(mul_grad);
-USE_OP(reduce_sum);
+USE_OP_ITSELF(reduce_sum);
 USE_OP(reduce_sum_grad);
 USE_OP_ITSELF(elementwise_add);

paddle/fluid/operators/reduce_ops/reduce_mean_op.cc

@@ -99,13 +99,6 @@ REGISTER_OPERATOR(reduce_mean_grad, ops::ReduceGradOp,
                   ops::ReduceMeanDoubleGradDescMaker,
                   ops::ReduceMeanDoubleGradOpBaseMaker,
                   ops::ReduceMeanGradNoNeedBufferVarInferer);
-REGISTER_OP_CPU_KERNEL(reduce_mean,
-                       ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                                         bool, ops::MeanFunctor>,
-                       ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                                         float, ops::MeanFunctor>,
-                       ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                                         double, ops::MeanFunctor>);
 
 template <typename T>
 using CPUReduceMeanGradKernel =

paddle/fluid/operators/reduce_ops/reduce_mean_op.cu

-// Copyright (c) 2018 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 <vector>
-#include "paddle/fluid/operators/reduce_ops/reduce_mean_op.h"
-#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
-
-REGISTER_OP_CUDA_KERNEL(
-    reduce_mean,
-    ops::ReduceCudaKernel<bool, kps::AddFunctor, kps::DivideFunctor>,
-    ops::ReduceCudaKernel<paddle::platform::float16, kps::AddFunctor,
-                          kps::DivideFunctor>,
-    ops::ReduceCudaKernel<float, kps::AddFunctor, kps::DivideFunctor>,
-    ops::ReduceCudaKernel<double, kps::AddFunctor, kps::DivideFunctor>,
-    ops::ReduceCudaKernel<int, kps::AddFunctor, kps::DivideFunctor>,
-    ops::ReduceCudaKernel<int64_t, kps::AddFunctor, kps::DivideFunctor>);

paddle/fluid/operators/reduce_ops/reduce_sum_op.cc

@@ -107,27 +107,6 @@ REGISTER_OPERATOR(reduce_sum_grad, ops::ReduceGradOp,
                   ops::ReduceSumDoubleOpGradMaker<paddle::imperative::OpBase>,
                   ops::ReduceSumGradNoNeedBufferVarInferer);
 
-REGISTER_OP_CPU_KERNEL(
-    reduce_sum, ops::ReduceKernel<paddle::platform::CPUDeviceContext, bool,
-                                  ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext, float,
-                      ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext, double,
-                      ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                      paddle::platform::float16, ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext, int16_t,
-                      ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext, int, ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext, int64_t,
-                      ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                      paddle::platform::complex<float>, ops::SumFunctor>,
-    ops::ReduceKernel<paddle::platform::CPUDeviceContext,
-                      paddle::platform::complex<double>,
-
-                      ops::SumFunctor>);
-
 template <typename T>
 using CPUReduceSumGradKernel =
     ops::ReduceSumGradKernel<paddle::platform::CPUDeviceContext, T,

paddle/fluid/operators/reduce_ops/reduce_sum_op.cu

-// Copyright (c) 2018 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/fluid/operators/reduce_ops/reduce_op.h"
-#include "paddle/fluid/operators/reduce_ops/reduce_sum_op.h"
-REGISTER_OP_CUDA_KERNEL(
-    reduce_sum,
-    ops::ReduceCudaKernel<bool, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<float, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<double, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<paddle::platform::float16, kps::AddFunctor,
-                          kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<int16_t, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<int, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<int64_t, kps::AddFunctor, kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<paddle::platform::complex<float>, kps::AddFunctor,
-                          kps::IdentityFunctor>,
-    ops::ReduceCudaKernel<paddle::platform::complex<double>, kps::AddFunctor,
-                          kps::IdentityFunctor>);

paddle/fluid/operators/reduce_ops/unity_build_rule.cmake

@@ -21,5 +21,4 @@ register_unity_group(cu
 register_unity_group(cu frobenius_norm_op.cu)
 register_unity_group(cu logsumexp_op.cu)
 register_unity_group(cu reduce_max_op.cu)
-register_unity_group(cu reduce_mean_op.cu)
 register_unity_group(cu reduce_min_op.cu)

paddle/phi/kernels/gpu/math_kernel.cu

@@ -165,4 +165,6 @@ PT_REGISTER_KERNEL(mean_raw,
                    float,
                    double,
                    bool,
-                   float16) {}
+                   float16,
+                   int,
+                   int64_t) {}

paddle/phi/kernels/math_kernel.cc

@@ -152,6 +152,8 @@ PT_REGISTER_KERNEL(mean,
                    float,
                    double,
                    bool,
+                   int,
+                   int64_t,
                    phi::dtype::float16) {}
 PT_REGISTER_KERNEL(sum,
                    GPU,