Revert "【Hackathon 6th Fundable Projects 3 No.242】 [fluid_ops] mp_all…

…reduce_s…" (#67832)

This reverts commit 29eb8d7.

paddle/fluid/operators/collective/mp_allreduce_sum_op.cc

@@ -82,3 +82,13 @@ REGISTER_OPERATOR(mp_allreduce_sum,
                   ops::MpAllReduceSumOpGradMaker<paddle::imperative::OpBase>,
                   ops::MpAllReduceSumOpMaker,
                   ops::MpAllReduceSumInplaceInferer);
+
+PD_REGISTER_STRUCT_KERNEL(mp_allreduce_sum,
+                          CPU,
+                          ALL_LAYOUT,
+                          ops::MpAllReduceSumCPUKernel,
+                          float,
+                          double,
+                          int,
+                          int64_t,
+                          phi::dtype::float16) {}

paddle/fluid/operators/collective/mp_allreduce_sum_op.cu.cc

@@ -0,0 +1,39 @@
+// Copyright (c) 2022 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/framework/op_registry.h"
+#include "paddle/fluid/operators/collective/c_allreduce_op.h"
+
+namespace paddle {
+namespace operators {
+DEFINE_C_ALLREDUCE_CUDA_KERNEL(MpAllReduceSum, kRedSum)
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+PD_REGISTER_STRUCT_KERNEL(mp_allreduce_sum,
+                          GPU,
+                          ALL_LAYOUT,
+                          ops::MpAllReduceSumCUDAKernel,
+                          float,
+                          double,
+                          int,
+                          int64_t,
+#if (NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000) || \
+    defined(PADDLE_WITH_HIP)
+                          phi::dtype::bfloat16,
+#endif
+                          phi::dtype::float16) {
+}

paddle/fluid/operators/collective/mp_allreduce_sum_op.kps

@@ -0,0 +1,29 @@
+#ifdef PADDLE_WITH_XPU_KP
+
+// Please do not modify the following code
+#if defined(__CUDA_ARCH__)
+#undef __CUDA_ARCH__
+#endif
+
+#if defined(__CUDACC__)
+#undef __CUDACC__
+#endif
+
+#if defined(__CUDA__)
+#undef __CUDA__
+#endif
+
+#if defined(__NVCC__)
+#undef __NVCC__
+#endif
+
+#include "paddle/fluid/operators/collective/c_allreduce_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_KERNEL(mp_allreduce_sum,
+                   KP,
+                   phi::XPUPlace,
+                   ops::CAllReduceOpXPUKernel<ops::kRedSum, float>);
+
+#endif

paddle/fluid/operators/collective/mp_allreduce_sum_op_xpu.cc

@@ -0,0 +1,31 @@
+// Copyright (c) 2022 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/collective/c_allreduce_op.h"
+
+namespace paddle {
+namespace operators {
+DEFINE_C_ALLREDUCE_XPU_KERNEL(CAllReduceSum, kRedSum)
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+PD_REGISTER_STRUCT_KERNEL(mp_allreduce_sum,
+                          XPU,
+                          ALL_LAYOUT,
+                          ops::CAllReduceSumXPUKernel,
+                          float,
+                          int,
+                          phi::dtype::float16) {}

paddle/fluid/pir/dialect/op_generator/ops_api_gen.py

@@ -202,8 +202,6 @@
     'nce',
     'lrn',
     'max_pool2d_v2',
-    'mp_allreduce_sum',
-    'mp_allreduce_sum_',
     'partial_sum',
     'pull_gpups_sparse',
     'pull_gpups_sparse_',

paddle/phi/ops/yaml/inconsistent/static_backward.yaml

@@ -313,12 +313,6 @@
     func : minimum_grad
   composite : minimum_grad(x, y, out_grad, x_grad, y_grad)
 
-- backward_op : mp_allreduce_sum_grad
-  forward : mp_allreduce_sum(Tensor x, int ring_id = 0, bool use_calc_stream = false) -> Tensor(out)
-  args : (Tensor out_grad, int ring_id = 0, bool use_calc_stream = false)
-  output : Tensor(x_grad)
-  invoke : c_identity(out_grad, ring_id, false, false)
-
 - backward_op : multiply_double_grad
   forward : multiply_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
   args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)

paddle/phi/ops/yaml/inconsistent/static_ops.yaml

@@ -576,18 +576,6 @@
   backward : minimum_grad
   interfaces : paddle::dialect::InferSymbolicShapeInterface
 
-- op : mp_allreduce_sum
-  args : (Tensor x, int ring_id = 0, bool use_calc_stream = false)
-  output : Tensor(out)
-  infer_meta :
-    func : AllReduceInferMeta
-    param: [x]
-  kernel :
-    func : mp_allreduce_sum
-    param: [x, ring_id, use_calc_stream]
-  backward: mp_allreduce_sum_grad
-  inplace: (x -> out)
-
 - op : multiply
   args : (Tensor x, Tensor y)
   output : Tensor

paddle/phi/ops/yaml/op_compat.yaml

@@ -4431,12 +4431,6 @@
   outputs:
     out: Out
 
-- op: mp_allreduce_sum
-  inputs :
-    x : X
-  outputs :
-    out: Out
-
 - op: nce
   backward: nce_grad
   inputs: