Cpp parallel executor #9080
Cpp parallel executor #9080
Conversation
doc/design/parallel_executor.md Outdated
| opt = fluid.optimizer.SGDOptimizer() | ||
| opt.minimize(avg_cost) | ||
| | ||
| # change Executor -> ParallelExecutor |
There was a problem hiding this comment.
Maybe we can let the user still use the Executor interface, and add an optional argument "gpu_list", and underlying if there are multiple GPUs available (either len(gpu_list) > 0, or gpu_list == None and multiple GPUs initialized), create and return the parallel executor instance.
doc/design/parallel_executor.md Outdated
| // e.g. sgd should wait for allreduce to be finished | ||
| CallBack->BeforeOp(op); | ||
| | ||
| op->Run(*local_scope, place_); |
There was a problem hiding this comment.
From my understanding, the reason we need callback is ParallelExecutor will call Executor::Run, but need to be notified before and after each OP::Run. Do we even need the Executor implementation anymore? Maybe we can consolidate them into a single executor, so that we don't need the callback anymore.
And it will be easier for the Python side, Python always create the same executor.
| std::vector<OpHandle *> to_run; | ||
| for (auto *var : to_remove) { | ||
| for (auto *op : var->pending_ops_) { | ||
| if (var->name_ == "mean_0.tmp_0@GRAD") { |
There was a problem hiding this comment.
what is the purpose of this special case?
There was a problem hiding this comment.
Just debug code... Sorry
| struct OpHandle { | ||
| std::vector<VarHandle *> inputs_; | ||
| std::vector<VarHandle *> outputs_; | ||
| platform::DeviceContext *dev_ctx_; |
| } | ||
| | ||
| std::vector<LoDTensor> ParallelExecutor::Run( | ||
| const std::vector<std::string> &fetch_tensors) { |
There was a problem hiding this comment.
Instantiate Variables here?
Paddle/paddle/fluid/framework/executor.cc
Lines 276 to 305 in 41894da
There was a problem hiding this comment.
Yes. We need to instantiate variables here. We might extract this routine to a global function.
| } | ||
| }; | ||
| | ||
| member_->pool_.Run(op_run); |
There was a problem hiding this comment.
Actually, we should add a callback after we push operator run job to memory pool. In this callback, we change pending_var state.
There was a problem hiding this comment.
The pending_var state has been changed during L404-L406
| member_->local_scopes_.size() != 1) { // Is CUDA | ||
| BuildNCCLCommunicator(); | ||
| BCastParamsToGPUs(startup_program); | ||
| } |
There was a problem hiding this comment.
Why do not initialize parameters on the respective devices？
There was a problem hiding this comment.
Since randomize result might not same when seed = 0
… cpp_parallel_executor
| void RunOp(std::unordered_map<VarHandleBase*, bool>& pending_vars, | ||
| OpHandle* op) const; | ||
| | ||
| void PolishGraphToSupportDataHarzaeds() const; |
reyoung force-pushed the cpp_parallel_executor branch from 17f0491 to c0c2e15 Compare reyoung force-pushed the cpp_parallel_executor branch from 9441175 to 201f79d Compare 
There was a problem hiding this comment.
Yes. Since the operators in a sub-block will be executed by a control flow operator, e.g., While. The behaviour between control flow operators and computational operators should be same.
… cpp_parallel_executor
…o cpp_parallel_executor
There was a problem hiding this comment.
I have finished my review. @chengduoZH Let's verify the correctness and speed of transfomer and resnext. If they are OK, let's merge it soon so that everyone can start improving it
| namespace details { | ||
| | ||
| struct FetchOpHandle : public OpHandleBase { | ||
| FeedFetchList *data_; |
There was a problem hiding this comment.
Let's add a "private:" then?
There was a problem hiding this comment.
I see. Can you add your reply as comments in the code?
| // if size is 0. We just make sure it does. | ||
| if (size <= 0) return nullptr; | ||
| void* p; | ||
| int prev_id; |
There was a problem hiding this comment.
I would suggest this to be a PADDLE_ENFORCE. if the current behavior works. Otherwise, reader will think the Allocator currently works on multiple GPUs.
| PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error"); | ||
| } | ||
| | ||
| void FetchOpHandle::WaitAndMergeCPUTensors() const { |
There was a problem hiding this comment.
Does this mean really wait on anything? Maybe just MergeCPUTensors?
| // FIXME: Currently ScaleLossGradOp only use device_count as scale | ||
| // factor. So it does not depend on any other operators. | ||
| // VarHandle *loss = GetVarHandle(loss_var_name, place); | ||
| // loss->pending_ops_.emplace_back(op_handle); |
There was a problem hiding this comment.
if this op doesn't not depend on anything, when will it be scheduled?
There was a problem hiding this comment.
It will be run at the first time.
| * | ||
| * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR) | ||
| */ | ||
| static void PolishGraphToSupportDataHazards(SSAGraph *graph); |
There was a problem hiding this comment.
Add more comments to describe when does data hazard happens?
| std::unordered_set<OpHandleBase *> ready_ops; | ||
| | ||
| auto InsertPendingVar = [&pending_vars, &ready_vars](VarHandleBase &var) { | ||
| pending_vars.insert(&var); |
There was a problem hiding this comment.
should this be skipped if generated_op_ is nullptr?
| PADDLE_ENFORCE_EQ(actual.size(), expect.size()); | ||
| for (int j = 0; j < actual.size(); ++j) { | ||
| PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); | ||
| // PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); |
| // Create local scopes. | ||
| for (auto &scope : local_scopes_) { | ||
| auto &local_scope = scope->NewScope(); | ||
| *scope->Var("@TMP_SCOPE@")->GetMutable<Scope *>() = &local_scope; |
There was a problem hiding this comment.
Why are all the scopes using the same var name?
There was a problem hiding this comment.
When Op::Run, there are some temporary variables will be created in local scopes. So, here just use Variable @TMP_SCOPE@ to holds these temporary variables. They will be destroied after a period.
| auto *dep_var = new DummyVarHandle(); | ||
| read_op->AddOutput(dep_var); | ||
| write_op->AddInput(dep_var); | ||
| graph->dep_vars_.emplace(dep_var); |
There was a problem hiding this comment.
should this be called data_hazard_vars_?
There was a problem hiding this comment.
Comments have been added.
… cpp_parallel_executor
6 participants
I just add a dependency engine to parse the dependencies of operators. There are still a lot of jobs need to be done.
Complete Broadcast parameters.Use thread pool to invoke operator parallelly.Complete NCCL AllReduce OpHandle in this implementation.I just use
VarHandleandOpHandleto parseProgramas a SSA form graph. A variable is assigned by only one OpHandle. When all inputs ofOpHandleis ready, theOpHandlecan be run.The speed of ResNeXt152 is