-
Notifications
You must be signed in to change notification settings - Fork 5.6k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
1 parent
9923be5
commit d885304
Showing
5 changed files
with
405 additions
and
9 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
145 changes: 145 additions & 0 deletions
145
paddle/fluid/framework/details/reduce_blockop_handle.cc
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,145 @@ | ||
| // 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/framework/details/reduce_blockop_handle.h" | ||
| #include "paddle/fluid/framework/details/container_cast.h" | ||
| #include "paddle/fluid/framework/details/reduce_and_gather.h" | ||
| #include "paddle/fluid/framework/details/variable_visitor.h" | ||
|
|
||
| namespace paddle { | ||
| namespace framework { | ||
| namespace details { | ||
|
|
||
| void ReduceBlockOpHandle::RunImpl() { | ||
| if (places_.size() == 1) return; | ||
|
|
||
| WaitInputVarGenerated(); | ||
|
|
||
| auto in_var_handles = DynamicCast<VarHandle>(inputs_); | ||
| auto out_var_handles = DynamicCast<VarHandle>(outputs_); | ||
|
|
||
| PADDLE_ENFORCE_EQ(in_var_handles.size(), | ||
| places_.size() * out_var_handles.size(), | ||
| "The number of input and output is not consistent."); | ||
|
|
||
| #ifdef PADDLE_WITH_CUDA | ||
| std::vector<const LoDTensor *> lod_tensors = GetInputValues(); | ||
| auto type = lod_tensors[0]->type(); | ||
| auto numel = lod_tensors[0]->numel(); | ||
| std::vector<std::function<void()>> nccl_reduce_calls; | ||
| auto root_dev_id = | ||
| boost::get<platform::CUDAPlace>(places_[dst_scope_id_]).device; | ||
| for (size_t i = 0; i < local_scopes_.size(); ++i) { | ||
| int dev_id = | ||
| boost::get<platform::CUDAPlace>(lod_tensors[i]->place()).device; | ||
| auto &nccl_ctx = nccl_ctxs_->at(dev_id); | ||
|
|
||
| void *buffer = const_cast<void *>(lod_tensors[i]->data<void>()); | ||
| void *recvbuffer = nullptr; | ||
| if (i == dst_scope_id_) { | ||
| auto reduce_var = local_scopes_[dst_scope_id_]->FindVar(var_name_); | ||
| recvbuffer = reduce_var->GetMutable<framework::LoDTensor>()->mutable_data( | ||
| lod_tensors[dst_scope_id_]->place()); | ||
| } | ||
|
|
||
| int nccl_type = platform::ToNCCLDataType(type); | ||
| nccl_reduce_calls.emplace_back( | ||
| [buffer, recvbuffer, nccl_type, numel, root_dev_id, &nccl_ctx] { | ||
| PADDLE_ENFORCE(platform::dynload::ncclReduce( | ||
| buffer, recvbuffer, numel, static_cast<ncclDataType_t>(nccl_type), | ||
| ncclSum, root_dev_id, nccl_ctx.comm_, nccl_ctx.stream())); | ||
| }); | ||
| } | ||
| #endif | ||
|
|
||
| std::vector<const Scope *> var_scopes; | ||
| for (auto *s : local_scopes_) { | ||
| auto var = s->FindVar(kLocalExecScopeName); | ||
| PADDLE_ENFORCE_NOT_NULL(var); | ||
| var_scopes.emplace_back(var->Get<Scope *>()); | ||
| } | ||
|
|
||
| // Reduce the variable which is in CPU side. | ||
| std::vector<std::vector<const LoDTensor *>> cpu_lod_tensors; | ||
| std::vector<Variable *> cpu_out_vars; | ||
| for (auto out_var_h : out_var_handles) { | ||
| auto out_var = var_scopes[dst_scope_id_]->FindVar(out_var_h->name_); | ||
| PADDLE_ENFORCE_NOT_NULL(out_var); | ||
|
|
||
| if (platform::is_cpu_place(out_var->Get<LoDTensor>().place())) { | ||
| std::vector<const LoDTensor *> lod_tensors; | ||
| for (size_t i = 0; i < places_.size(); ++i) { | ||
| lod_tensors.emplace_back( | ||
| &(var_scopes[i]->FindVar(out_var_h->name_)->Get<LoDTensor>())); | ||
| PADDLE_ENFORCE(platform::is_cpu_place(lod_tensors.back()->place())); | ||
| } | ||
|
|
||
| cpu_lod_tensors.push_back(lod_tensors); | ||
| cpu_out_vars.push_back(out_var); | ||
| } | ||
| } | ||
| #ifdef PADDLE_WITH_CUDA | ||
| this->RunAndRecordEvent([&] { | ||
| { | ||
| platform::NCCLGroupGuard guard; | ||
| for (auto &call : nccl_reduce_calls) { | ||
| call(); | ||
| } | ||
| } | ||
| for (size_t i = 0; i < cpu_lod_tensors.size(); ++i) { | ||
| ReduceLoDTensor func(cpu_lod_tensors[i], | ||
| cpu_out_vars[i]->GetMutable<framework::LoDTensor>()); | ||
| VisitDataType(ToDataType(cpu_lod_tensors[i][0]->type()), func); | ||
| } | ||
| }); | ||
| #endif | ||
| } | ||
|
|
||
| std::vector<const LoDTensor *> ReduceBlockOpHandle::GetInputValues() { | ||
| std::vector<const LoDTensor *> lod_tensors; | ||
|
|
||
| auto &value_0 = GetInputValue(0); | ||
| lod_tensors.emplace_back(&value_0); | ||
|
|
||
| std::unordered_set<int> dev_id_set; | ||
| auto dev_id = boost::get<platform::CUDAPlace>(value_0.place()).device; | ||
| dev_id_set.insert(dev_id); | ||
|
|
||
| for (size_t i = 1; i < local_scopes_.size(); ++i) { | ||
| auto &value = GetInputValue(i); | ||
| PADDLE_ENFORCE_EQ(value_0.type(), value.type()); | ||
| PADDLE_ENFORCE_EQ(value_0.numel(), value.numel()); | ||
| lod_tensors.emplace_back(&value); | ||
|
|
||
| auto dev_id = boost::get<platform::CUDAPlace>(value.place()).device; | ||
| if (dev_id_set.count(dev_id)) { | ||
| PADDLE_THROW("dev_%d has been in dev_id_set.", dev_id); | ||
| } | ||
| dev_id_set.insert(dev_id); | ||
| } | ||
| return lod_tensors; | ||
| } | ||
|
|
||
| const LoDTensor &ReduceBlockOpHandle::GetInputValue(size_t idx) { | ||
| auto reduce_var = local_scopes_.at(idx)->FindVar(var_name_); | ||
| PADDLE_ENFORCE_NOT_NULL(reduce_var, "%s is not found.", var_name_); | ||
| auto &lod_tensor = reduce_var->Get<LoDTensor>(); | ||
| PADDLE_ENFORCE(platform::is_gpu_place(lod_tensor.place())); | ||
| return lod_tensor; | ||
| } | ||
|
|
||
| std::string ReduceBlockOpHandle::Name() const { return "reduce_block"; } | ||
| } // namespace details | ||
| } // namespace framework | ||
| } // namespace paddle |
Oops, something went wrong.