m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Support GetNext Parallel

Browse Source
pull/390/head
gukecai 6 years ago
parent 27a88a6bc3
commit f8208c7c52
10 changed files with 304 additions and 223 deletions
Show Stats Download Patch File Download Diff File

9
mindspore/ccsrc/device/ascend/ascend_kernel_runtime.cc
Unescape View File

@ -283,18 +283,19 @@ bool AscendKernelRuntime::GenTask(const session::KernelGraph *graph) {
AscendStreamAssign &assign_instance = AscendStreamAssign::GetInstance();
// the streams' flag not HEAD_STREAM
std::vector<uint32_t> wait_active_stream_list = assign_instance.GetWaitStreams();
std::vector<uint32_t> force_copy_stream_list = assign_instance.GetHcomStreams();
std::vector<uint32_t> wait_active_stream_list;
assign_instance.GetWaitStreams(&wait_active_stream_list);
auto force_copy_stream_list = assign_instance.hcom_streams();
MS_LOG(INFO) << "call DavinciModel total stream num:" << assign_instance.GetTotalStreamNum()
<< ", total event num:" << assign_instance.GetTotalEventNum()
<< ", total event num:" << assign_instance.total_event_num()
<< ", wait_active_stream_list size:" << wait_active_stream_list.size()
<< ", force_copy_stream_list size:" << force_copy_stream_list.size();
std::vector<std::shared_ptr<ge::model_runner::OpInfo>> empty_list;
std::shared_ptr<ge::model_runner::DavinciModel> model = std::make_shared<ge::model_runner::DavinciModel>(
task_info_list, empty_list, empty_list, empty_list, empty_list, wait_active_stream_list, force_copy_stream_list, 0,
0, 0, 0, 0, 0, assign_instance.GetTotalStreamNum(), 1, assign_instance.GetTotalEventNum(), 0);
0, 0, 0, 0, 0, assign_instance.GetTotalStreamNum(), 1, assign_instance.total_event_num(), 0);
auto ret = graph_model_map_.insert(std::make_pair(graph->graph_id(), model));
if (!ret.second) {

154
mindspore/ccsrc/device/ascend/ascend_stream_assign.cc
View File

File diff suppressed because it is too large Load Diff

33
mindspore/ccsrc/device/ascend/ascend_stream_assign.h
Unescape View File

@ -49,37 +49,35 @@ class AscendStreamAssign {
uint32_t GetTotalStreamNum() const;
// new stream policy
uint32_t GetTotalCommonStreamNum() const { return total_common_stream_num_; }
uint32_t GetTotalIndependStreamNum() const { return total_independ_stream_num_; }
uint32_t GetTotalEventNum() const { return total_event_num_; }
const uint32_t GetFisrtPhysicId() const { return first_physic_id_; }
const uint32_t GetFirstLogicId() const { return first_logic_id_; }
uint32_t total_common_stream_num() const { return total_common_stream_num_; }
uint32_t total_independ_stream_num() const { return total_independ_stream_num_; }
uint32_t total_event_num() const { return total_event_num_; }
void InsertActiveNew(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void AssignAllNodesStream(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void ResetNew();
void AssignStreamNew(const std::shared_ptr<session::KernelGraph>& graph_ptr);
bool IsIndependentNode(const CNodePtr& node_ptr);
const std::unordered_map<uint32_t, uint32_t> GetIndependentMap() { return logic_to_independent_map_; }
const std::unordered_map<uint32_t, uint32_t> GetPhysicMap() { return logic_to_physic_map_; }
std::vector<uint32_t> GetWaitStreams();
std::vector<uint32_t> GetHcomStreams();
private:
AscendStreamAssign() = default;
~AscendStreamAssign() = default;
const std::unordered_map<uint32_t, uint32_t>& logic_to_independent_map() { return logic_to_independent_map_; }
const std::unordered_map<uint32_t, uint32_t>& logic_to_physic_map() { return logic_to_physic_map_; }
const std::vector<std::vector<uint32_t>>& inner_parallel_streams() { return inner_parallel_streams_; }
void GetWaitStreams(vector<uint32_t>* wait_active_stream_list);
const std::vector<uint32_t>& hcom_streams() { return hcom_stream_list_; }
CNodePtr CreateSendApplyKernel(const std::shared_ptr<session::KernelGraph>& graph_ptr, uint32_t event_id,
uint32_t stream_id);
CNodePtr CreateRecvApplyKernel(const std::shared_ptr<session::KernelGraph>& graph_ptr, uint32_t event_id,
uint32_t stream_id);
private:
AscendStreamAssign() = default;
~AscendStreamAssign() = default;
vector<CNodePtr>::iterator FindTargetOp(vector<CNodePtr>::iterator begin, vector<CNodePtr>::iterator end,
const CNodePtr& node);
bool IsHcom(const CNodePtr& apply_kernel);
bool IsProcessed(uint32_t logic_id);
vector<uint32_t> TransLogicToPhysic(const vector<uint32_t>& logic_ids);
void TransLogicToPhysic(const vector<uint32_t>& logic_ids, vector<uint32_t>* physic_ids);
void AssignCommonStreamId(const CNodePtr& cur_cnode_ptr, CNodePtr* pre_cnode_ptr, uint32_t* cur_index,
uint32_t* cur_stream_id);
void RecordIdMap(uint32_t logic_id, uint32_t physic_id);
@ -88,15 +86,17 @@ class AscendStreamAssign {
bool IsTaskSink();
void AssignIndependentStreamId(const CNodePtr& cur_cnode_ptr, uint32_t deal_logic_id);
void UpdateStreamId(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void UpdateEventId(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void PrintGraphExeOrders(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void RecordFirstCommonOp(const CNodePtr& cur_cnode_ptr, uint32_t cur_node_logic_id, uint32_t cur_stream_id);
uint32_t GetLogicId(const CNodePtr& cur_cnode_ptr);
void SetCommonStreamNum(uint32_t cur_stream_id);
void FindAllReduceParallel(const std::shared_ptr<session::KernelGraph>& graph_ptr);
bool IsProcessedParallelStream(uint32_t stream_id);
vector<uint32_t> GetParallelStream(uint32_t cur_stream_id, uint32_t stream_acitve_id);
void GetParallelStream(uint32_t cur_stream_id, uint32_t stream_acitve_id, std::vector<uint32_t>* parallel_streams);
void InsertSendRecvForIndependent(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void InsertSendRecvForHcomParallel(const std::shared_ptr<session::KernelGraph>& graph_ptr);
void GetNeedActiveStreams(const std::shared_ptr<session::KernelGraph>& graph_ptr);
uint32_t total_common_stream_num_{0};
uint32_t total_independ_stream_num_{0};
@ -112,6 +112,7 @@ class AscendStreamAssign {
std::vector<std::vector<uint32_t>> inner_parallel_streams_{};
std::vector<uint32_t> processed_parallel_streams_{};
std::vector<uint32_t> hcom_stream_list_{};
std::vector<uint32_t> need_first_active_streams_{};
// new policy end
};
} // namespace ascend

286
mindspore/ccsrc/device/kernel_adjust.cc
View File

File diff suppressed because it is too large Load Diff

25
mindspore/ccsrc/device/kernel_adjust.h
Unescape View File

@ -28,10 +28,22 @@
#include "session/session_context.h"
#include "ir/meta_tensor.h"
#include "device/ascend/profiling/profiling_utils.h"
#include "device/kernel_info.h"
using mindspore::device::ascend::ProfilingTraceInfo;
using mindspore::device::ascend::ProfilingUtils;
namespace mindspore {
constexpr auto kLoopCountParamName = "loop_count";
constexpr auto kIterLoopParamName = "iter_loop";
constexpr auto kZeroParamName = "zero";
constexpr auto kOneParamName = "one";
constexpr auto kStreamNeedActivedFirst = "stream_need_active_first";
const uint32_t kFirstStreamSwitchLabel = kInvalidDistincLabel - 1;
const uint32_t kGetNextLabel = kInvalidDistincLabel - 2;
const uint32_t kSecondStreamSwitchLabel = kInvalidDistincLabel - 3;
const uint32_t kInvalidEventId = UINT32_MAX;
const uint32_t kFirstEventId = kInvalidEventId / 2;
namespace device {
class KernelAdjust {
public:
@ -41,26 +53,23 @@ class KernelAdjust {
}
void Reorder(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
void InsertSwitchLoop(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
void SetStreamActiveOPs(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const std::unordered_set<uint32_t> &ctrl_stream_list,
const std::unordered_set<uint32_t> &comm_stream_list,
const std::unordered_set<uint32_t> &momentum_stream_list);
void SetStreamSwitchOps(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
bool StepLoadCtrlInputs(const std::shared_ptr<session::Context> &context,
const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
void Profiling(NotNull<session::KernelGraph *> kernel_graph_ptr);
static bool NeedInsertSwitch();
CNodePtr CreateSteamActiveOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
CNodePtr CreateStreamActiveOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
private:
KernelAdjust() = default;
~KernelAdjust() = default;
CNodePtr CreateRecvApplyKernel(const std::shared_ptr<session::KernelGraph> &graph_ptr, uint32_t event_id);
CNodePtr CreateSendApplyKernel(const std::shared_ptr<session::KernelGraph> &graph_ptr, uint32_t event_id);
uint32_t FindFirstStreamSwitchLabel(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
void CreateSwitchOpParameters(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
std::map<std::string, mindspore::ParameterPtr> *switch_loop_input);
CNodePtr CreateStreamSwitchOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const std::map<std::string, mindspore::ParameterPtr> &switch_loop_input);
CNodePtr CreateStreamActiveSwitchOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
CNodePtr CreateStreamActiveOtherOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr);
CNodePtr CreateStreamAssignAddnOP(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const std::map<std::string, mindspore::ParameterPtr> &switch_loop_input);
kernel::KernelBuildInfo::KernelBuildInfoBuilder CreateMngKernelBuilder(const std::vector<std::string> &formats,

7
mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
Unescape View File

@ -62,6 +62,7 @@
#include "pre_activate/ascend/format_type/insert_transdata_for_runop.h"
#include "pre_activate/ascend/enhancer/getnext_memcpy_elimination.h"
#include "pre_activate/ascend/ir_fission/addn_fission.h"
#include "pre_activate/ascend/enhancer/insert_memcpy_async_for_getnext.h"
#include "utils/context/ms_context.h"
#include "utils/config_manager.h"
#include "debug/anf_ir_dump.h"
@ -187,6 +188,12 @@ void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGrap
ir_fusion_pm->AddPass(std::make_shared<GetitemTuple>());
ir_fusion_pm->AddPass(std::make_shared<TransposeTransDataFusion>());
}
if (context_ptr->enable_task_sink() && context_ptr->loop_sink_flag() && ConfigManager::GetInstance().iter_num() > 1) {
ir_fusion_pm->AddPass(std::make_shared<InsertMemcpyAsyncForGetNext>());
ir_fusion_pm->AddPass(std::make_shared<GetitemTuple>());
ir_fusion_pm->AddPass(std::make_shared<EraseVisitAttr>());
}
optimizer->AddPassManager(ir_fusion_pm);
(void)optimizer->Optimize(kernel_graph);
kernel_graph->SetExecOrderByDefault();

4
mindspore/ccsrc/pre_activate/mem_reuse/stream_reuse.cc
Unescape View File

@ -20,8 +20,8 @@ namespace mindspore {
namespace memreuse {
void StreamReuse::SetStreamReuseResource() {
#ifdef ENABLE_D
auto logic_physic_map = device::ascend::AscendStreamAssign::GetInstance().GetPhysicMap();
auto logic_independent_map = device::ascend::AscendStreamAssign::GetInstance().GetIndependentMap();
auto logic_physic_map = device::ascend::AscendStreamAssign::GetInstance().logic_to_physic_map();
auto logic_independent_map = device::ascend::AscendStreamAssign::GetInstance().logic_to_independent_map();
MS_LOG(INFO) << "stream mem reuse for Davici";
if (!logic_independent_map.empty() && !logic_physic_map.empty()) {
set_logic_physic_map(logic_physic_map);

2
mindspore/ccsrc/session/ascend_session.cc
Unescape View File

@ -610,7 +610,7 @@ void AscendSession::CopyOutputOfIf(GraphId false_graph_id) {
if (context_ptr->enable_task_sink() && context_ptr->loop_sink_flag() &&
ConfigManager::GetInstance().iter_num() > 1) {
// insert active in true graph, another active will be inserted in kernel adjust
InsertStreamActiveToGraph(true_last_id, kInvalidDistincLabel - 1);
InsertStreamActiveToGraph(true_last_id, kSecondStreamSwitchLabel);
}
break;
}

3
mindspore/ccsrc/utils/utils.h
Unescape View File

@ -114,6 +114,9 @@ constexpr auto kFusedMulAddNOpName = "FusedMulAddN";
constexpr auto kFusedMulApplyMomentumOpName = "FusedMulApplyMomentum";
constexpr auto kBiasAddOpName = "BiasAdd";
constexpr auto kConfusionMulGradOpName = "ConfusionMulGrad";
constexpr auto kStreamSwitchOpName = "StreamSwitch";
constexpr auto kStreamActiveOpName = "StreamActive";
constexpr auto kAssignAddOpName = "AssignAdd";
constexpr auto kSendOpName = "Send";
constexpr auto kRecvOpName = "Recv";
constexpr auto kReluV2OpName = "ReluV2";

4
tests/ut/cpp/stub/tasksink/ascend_stream_assign_stub.cc
Unescape View File

@ -24,9 +24,7 @@ void AscendStreamAssign::AssignStreamNew(const KernelGraphPtr &graph) { return;
uint32_t AscendStreamAssign::GetTotalStreamNum() const { return 1; }
std::vector<uint32_t> AscendStreamAssign::GetWaitStreams() { return vector<uint32_t>(); }
std::vector<uint32_t> AscendStreamAssign::GetHcomStreams() { return vector<uint32_t>(); }
void AscendStreamAssign::GetWaitStreams(vector<uint32_t> *wait_active_stream_list) { return; }
namespace tasksink {
bool TaskGenerator::GenTasks(const std::vector<CNodePtr> &anf_node_list, std::vector<TaskInfoPtr> *const task_info_list,

Write Preview
Loading…
Cancel
Save