p04896132
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Cancel NoOp optimizer in GPU until memory reuse ready

This commit is contained in:
wilfChen 2020-05-23 15:57:55 +08:00
parent 23b4b4d106
commit 56d751b5d2
2 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
mindspore/ccsrc/pre_activate/common/helper.cc
View File

@ -328,7 +328,7 @@ tensor::TensorPtr CreateTupleTensor(const ValueTuplePtr &value_tuple) {
bool IsNopNode(const AnfNodePtr &node) {
auto context_ptr = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(context_ptr);
if (context_ptr->device_target() != kAscendDevice && context_ptr->device_target() != kGPUDevice) {
if (context_ptr->device_target() != kAscendDevice) {
return false;
}
static std::unordered_set<std::string> nop_nodes = {prim::kPrimReshape->name(), kExpandDimsOpName,

6
mindspore/ccsrc/session/gpu_session.cc
View File

@ -70,7 +70,7 @@ void GPUSession::AllocateMemory(KernelGraph *kernel_graph) const {
MS_EXCEPTION_IF_NULL(kernel_graph);
auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
MS_EXCEPTION_IF_NULL(runtime_instance);
opt::RemoveNopNode(kernel_graph);
// opt::RemoveNopNode(kernel_graph);
runtime_instance->AssignMemory(kernel_graph);
}
@ -79,7 +79,7 @@ void GPUSession::RunOpAllocateMemory(const std::vector<tensor::TensorPtr> &input
MS_EXCEPTION_IF_NULL(kernel_graph);
auto runtime_instance = device::KernelRuntimeManager::Instance().GetSingleKernelRuntime(kGPUDevice, device_id_);
MS_EXCEPTION_IF_NULL(runtime_instance);
opt::RemoveNopNode(kernel_graph);
// opt::RemoveNopNode(kernel_graph);
runtime_instance->RunOpAssignMemory(input_tensors, kernel_graph);
}
@ -106,7 +106,7 @@ GraphId GPUSession::CompileGraph(const AnfNodePtrList &lst, const AnfNodePtrList
// Assign CUDA streams
AssignStream(graph);
// Remove NoOp from execution graph
opt::HideNopNode(graph.get());
// opt::HideNopNode(graph.get());
// Build kernel if node is cnode
BuildKernel(graph);
// Set graph execution order before memory alloc, ensure that memory alloc is according to the reorder graph