PaddlePaddle
diff --git a/‎paddle/fluid/operators/custom_device_common_op_registry.cc
Copy file name to clipboard
-216Lines changed: 0 additions & 216 deletions b/‎paddle/fluid/operators/custom_device_common_op_registry.cc
Copy file name to clipboard
-216Lines changed: 0 additions & 216 deletions
diff --git a/‎paddle/phi/kernels/custom/barrier_kernel.cc
Copy file name to clipboard
+63Lines changed: 63 additions & 0 deletions b/‎paddle/phi/kernels/custom/barrier_kernel.cc
Copy file name to clipboard
+63Lines changed: 63 additions & 0 deletions
@@ -1317,22 +1317,6 @@ void FeedDenseTensorKernel(const Context& dev_ctx,
 void RegisterCustomDeviceCommonKernel(const std::string& dev_type) {
 #ifdef PADDLE_WITH_CUSTOM_DEVICE
   auto device_type = dev_type.c_str();
-  /* see [Why use single type kernel] */
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      save_combine,
-      device_type,
-      paddle::operators ::SaveCombineOpKernel<phi::CustomContext, float>,
-      paddle::operators ::SaveCombineOpKernel<phi::CustomContext, double>,
-      paddle::operators ::SaveCombineOpKernel<phi::CustomContext, int>,
-      paddle::operators ::SaveCombineOpKernel<phi::CustomContext, int64_t>);
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      load_combine,
-      device_type,
-      paddle::operators::LoadCombineOpKernel<float, phi::CustomContext>,
-      paddle::operators::LoadCombineOpKernel<double, phi::CustomContext>,
-      paddle::operators::LoadCombineOpKernel<int, phi::CustomContext>,
-      paddle::operators::LoadCombineOpKernel<int8_t, phi::CustomContext>,
-      paddle::operators::LoadCombineOpKernel<int64_t, phi::CustomContext>);
   REGISTER_OP_CUSTOM_DEVICE_KERNEL(
       c_concat,
       device_type,
@@ -1341,31 +1325,6 @@ void RegisterCustomDeviceCommonKernel(const std::string& dev_type) {
                                                      phi::dtype::float16>,
       paddle::operators::CConcatOpCustomDeviceKernel<phi::CustomContext,
                                                      phi::dtype::bfloat16>);
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_split,
-      device_type,
-      paddle::operators::CSplitOpCustomDeviceKernel<phi::CustomContext, float>,
-      paddle::operators::CSplitOpCustomDeviceKernel<phi::CustomContext, int>,
-      paddle::operators::CSplitOpCustomDeviceKernel<phi::CustomContext,
-                                                    phi::dtype::float16>,
-      paddle::operators::CSplitOpCustomDeviceKernel<phi::CustomContext,
-                                                    phi::dtype::bfloat16>);
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_embedding,
-      device_type,
-      paddle::operators::CEmbeddingOpCustomDeviceKernel<phi::CustomContext,
-                                                        float>,
-      paddle::operators::CEmbeddingOpCustomDeviceKernel<phi::CustomContext,
-                                                        phi::dtype::float16>);
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_embedding_grad,
-      device_type,
-      paddle::operators::CEmbeddingGradOpCustomDeviceKernel<phi::CustomContext,
-                                                            float>,
-      paddle::operators::CEmbeddingGradOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16>);
-
   REGISTER_OP_CUSTOM_DEVICE_KERNEL(
       c_softmax_with_cross_entropy,
       device_type,
@@ -1392,181 +1351,6 @@ void RegisterCustomDeviceCommonKernel(const std::string& dev_type) {
           phi::CustomContext,
           phi::dtype::float16>) {}
 
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_identity,
-      device_type,
-      paddle::operators::CIdentityOpCustomDeviceKernel<phi::CustomContext,
-                                                       float>,
-      paddle::operators::CIdentityOpCustomDeviceKernel<phi::CustomContext,
-                                                       double>,
-      paddle::operators::CIdentityOpCustomDeviceKernel<phi::CustomContext, int>,
-      paddle::operators::CIdentityOpCustomDeviceKernel<phi::CustomContext,
-                                                       int64_t>,
-      paddle::operators::CIdentityOpCustomDeviceKernel<phi::CustomContext,
-                                                       phi::dtype::float16>) {}
-
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      sync_calc_stream,
-      device_type,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<phi::CustomContext,
-                                                           int16_t>,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<phi::CustomContext,
-                                                           int32_t>,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<phi::CustomContext,
-                                                           int64_t>,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<phi::CustomContext,
-                                                           float>,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<phi::CustomContext,
-                                                           double>,
-      paddle::operators::CSyncCalcStreamCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_allreduce_sum,
-      device_type,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          float,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          double,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int32_t,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int64_t,
-          phi::ccl::CCLReduceOp::SUM>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      mp_allreduce_sum,
-      device_type,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          float,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          double,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int32_t,
-          phi::ccl::CCLReduceOp::SUM>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int64_t,
-          phi::ccl::CCLReduceOp::SUM>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_allreduce_min,
-      device_type,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          float,
-          phi::ccl::CCLReduceOp::MIN>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          double,
-          phi::ccl::CCLReduceOp::MIN>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16,
-          phi::ccl::CCLReduceOp::MIN>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int32_t,
-          phi::ccl::CCLReduceOp::MIN>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int64_t,
-          phi::ccl::CCLReduceOp::MIN>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_allreduce_max,
-      device_type,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          float,
-          phi::ccl::CCLReduceOp::MAX>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          double,
-          phi::ccl::CCLReduceOp::MAX>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16,
-          phi::ccl::CCLReduceOp::MAX>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int32_t,
-          phi::ccl::CCLReduceOp::MAX>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int64_t,
-          phi::ccl::CCLReduceOp::MAX>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_allreduce_prod,
-      device_type,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          float,
-          phi::ccl::CCLReduceOp::PRODUCT>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          double,
-          phi::ccl::CCLReduceOp::PRODUCT>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          phi::dtype::float16,
-          phi::ccl::CCLReduceOp::PRODUCT>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int32_t,
-          phi::ccl::CCLReduceOp::PRODUCT>,
-      paddle::operators::CAllReduceOpCustomDeviceKernel<
-          phi::CustomContext,
-          int64_t,
-          phi::ccl::CCLReduceOp::PRODUCT>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      c_broadcast,
-      device_type,
-      paddle::operators::CBroadcastOpCustomDeviceKernel<int32_t>,
-      paddle::operators::CBroadcastOpCustomDeviceKernel<int64_t>,
-      paddle::operators::CBroadcastOpCustomDeviceKernel<float>,
-      paddle::operators::CBroadcastOpCustomDeviceKernel<double>,
-      paddle::operators::CBroadcastOpCustomDeviceKernel<phi::dtype::float16>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      barrier,
-      device_type,
-      paddle::operators::BarrierOpCustomDeviceKernel<int>) {}
-
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      global_scatter,
-      device_type,
-      paddle::operators::GlobalScatterOpCustomDeviceKernel<float>,
-      paddle::operators::GlobalScatterOpCustomDeviceKernel<double>,
-      paddle::operators::GlobalScatterOpCustomDeviceKernel<int32_t>,
-      paddle::operators::GlobalScatterOpCustomDeviceKernel<int64_t>,
-      paddle::operators::GlobalScatterOpCustomDeviceKernel<
-          phi::dtype::float16>) {}
-  REGISTER_OP_CUSTOM_DEVICE_KERNEL(
-      global_gather,
-      device_type,
-      paddle::operators::GlobalGatherOpCustomDeviceKernel<float>,
-      paddle::operators::GlobalGatherOpCustomDeviceKernel<double>,
-      paddle::operators::GlobalGatherOpCustomDeviceKernel<int32_t>,
-      paddle::operators::GlobalGatherOpCustomDeviceKernel<int64_t>,
-      paddle::operators::GlobalGatherOpCustomDeviceKernel<
-          phi::dtype::float16>) {}
 #endif
 }
 
 
@@ -0,0 +1,63 @@
+// Copyright (c) 2025 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
+//
+//     https://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/phi/api/backward/backward_api.h"
+#include "paddle/phi/api/include/api.h"
+#include "paddle/phi/backends/all_context.h"
+#include "paddle/phi/backends/device_manager.h"
+#include "paddle/phi/core/distributed/collective/process_group.h"
+#include "paddle/phi/core/distributed/comm_context_manager.h"
+#include "paddle/phi/core/distributed/xccl_comm_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/core/tensor_utils.h"
+#ifdef PADDLE_WITH_CUSTOM_DEVICE
+namespace phi {
+
+template <typename T, typename Context>
+void BarrierKernel(const Context& dev_ctx,
+                   const DenseTensor& x_in,
+                   int ring_id,
+                   bool use_calc_stream,
+                   DenseTensor* out) {
+  auto in = &x_in;
+
+  auto place = dev_ctx.GetPlace();
+  int64_t numel = in->numel();
+  const void* sendbuff = in->data();
+  void* recvbuff = dev_ctx.template Alloc<T>(out);
+  int rid = ring_id;
+
+  auto comm = reinterpret_cast<phi::distributed::XCCLCommContext*>(
+      phi::distributed::CommContextManager::GetInstance().Get(
+          std::to_string(rid)));
+
+  std::shared_ptr<phi::stream::Stream> stream;
+  if (use_calc_stream) {
+    stream = dev_ctx.GetStream();
+  } else {
+    stream = comm->GetStream();
+  }
+  phi::DeviceManager::CCLAllReduce(place.GetDeviceType(),
+                                   const_cast<void*>(sendbuff),
+                                   recvbuff,
+                                   numel,
+                                   in->dtype(),
+                                   phi::ccl::CCLReduceOp::SUM,
+                                   comm->GetXcclComm(),
+                                   *stream);
+}
+}  // namespace phi
+
+PD_REGISTER_KERNEL(barrier, Custom, ALL_LAYOUT, phi::BarrierKernel, int) {}
+#endif