diff --git a/ge/graph/load/new_model_manager/davinci_model.cc b/ge/graph/load/new_model_manager/davinci_model.cc index 93f4624e..f2bf0aea 100755 --- a/ge/graph/load/new_model_manager/davinci_model.cc +++ b/ge/graph/load/new_model_manager/davinci_model.cc @@ -139,7 +139,6 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr kernel_list; - std::vector arg_list; - std::vector dump_flag_list; - std::vector op_desc_list; - std::vector dump_args_list; - uint32_t last_dump_flag; - int64_t last_group_key; - uintptr_t last_dump_args; - OpDescPtr last_op; -}; - struct TaskMemInfo { int64_t input_size{0}; int64_t output_size{0}; @@ -223,14 +204,13 @@ class DavinciModel { // get total mem size size_t TotalMemSize() const { return runtime_param_.mem_size; } - const std::map &P2PMemInfos() const { return runtime_param_.memory_infos; } + const std::map &P2PMemInfos() const {return runtime_param_.memory_infos;} // model name string Name() const { return name_; } // om_name string OmName() const { return om_name_; } - // version uint32_t Version() const { return version_; } @@ -275,16 +255,12 @@ class DavinciModel { } return nullptr; } - // get task info for profiling const std::vector &GetTaskDescInfo() const { return task_desc_info_; } // get updated task info list std::vector GetTaskList() { return task_list_; } - // Modified from KernelTaskInfo. - SuperKernelTaskInfo &GetSuperKernelTaskInfo() { return skt_info_; } - /// /// @ingroup ge /// @brief get model input and output format @@ -634,7 +610,7 @@ class DavinciModel { uint8_t *MallocWeightsMem(size_t weights_size); - uint8_t *MallocP2PMem(size_t p2p_data_size); + uint8_t* MallocP2PMem(size_t p2p_data_size); void FreeFeatureMapMem(); @@ -1020,9 +996,6 @@ class DavinciModel { std::multimap op_id_map_; std::vector profile_list_; - - // For super kernel. - SuperKernelTaskInfo skt_info_; }; } // namespace ge #endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_DAVINCI_MODEL_H_ diff --git a/ge/graph/load/new_model_manager/task_info/hccl_task_info.cc b/ge/graph/load/new_model_manager/task_info/hccl_task_info.cc index 2d476662..4fb64aab 100644 --- a/ge/graph/load/new_model_manager/task_info/hccl_task_info.cc +++ b/ge/graph/load/new_model_manager/task_info/hccl_task_info.cc @@ -59,40 +59,40 @@ Status HcclTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_m GELOGI("HcclTaskInfo Init, op_index is: %u", op_index); // Get HCCL op - const auto op_desc = davinci_model_->GetOpByIndex(op_index); - GE_CHECK_NOTNULL(op_desc); + op_desc_ = davinci_model->GetOpByIndex(op_index); + GE_CHECK_NOTNULL(op_desc_); // Create the kernel hccl infos - CreateKernelHcclInfo(op_desc); + CreateKernelHcclInfo(op_desc_); // Initialize the hccl_type of all kernel hccl info HcomOmeUtil::GetHcclType(task_def, kernel_hccl_infos_); // Only in Horovod scenario should get the inputName and GeShape - ret = HcomOmeUtil::GetHorovodInputs(op_desc, kernel_hccl_infos_); + ret = HcomOmeUtil::GetHorovodInputs(op_desc_, kernel_hccl_infos_); if (ret != SUCCESS) { GELOGE(ret, "davinci_model: GetHorovodInputs fail! domi error: %u", ret); return ret; } - Status dmrt = HcomOmeUtil::GetHcclDataType(op_desc, kernel_hccl_infos_); + Status dmrt = HcomOmeUtil::GetHcclDataType(op_desc_, kernel_hccl_infos_); if (dmrt != SUCCESS) { GELOGE(dmrt, "davinci_model: GetHcomDataType fail! domi error: %u", dmrt); return dmrt; } - dmrt = HcomOmeUtil::GetHcclCount(op_desc, kernel_hccl_infos_); + dmrt = HcomOmeUtil::GetHcclCount(op_desc_, kernel_hccl_infos_); if (dmrt != SUCCESS) { GELOGE(dmrt, "davinci_model: GetHcomCount fail! domi error: %u", dmrt); return dmrt; } // Only HCOMBROADCAST and HVDCALLBACKBROADCAST need to get the rootId - dmrt = HcomOmeUtil::GetAllRootId(op_desc, kernel_hccl_infos_); + dmrt = HcomOmeUtil::GetAllRootId(op_desc_, kernel_hccl_infos_); if (dmrt != SUCCESS) { GELOGE(dmrt, "davinci_model: Get rootId fail! domi error: %u", dmrt); return dmrt; } // GE's new process: hccl declares the number of streams required, creates a stream by GE, and sends it to hccl - ret = SetFollowStream(op_desc, davinci_model); + ret = SetFollowStream(op_desc_, davinci_model); if (ret != SUCCESS) { GELOGE(ret, "SetStream Fail."); return ret; @@ -100,28 +100,21 @@ Status HcclTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_m if (davinci_model_->IsKnownNode()) { args_ = davinci_model_->GetCurrentArgsAddr(args_offset_); - GELOGI("Known node %s args addr %p, offset %u.", op_desc->GetName().c_str(), args_, args_offset_); + GELOGI("Known node %s args addr %p, offset %u.", op_desc_->GetName().c_str(), args_, args_offset_); } - ret = SetAddrs(op_desc, kernel_hccl_infos_); + ret = SetAddrs(op_desc_, kernel_hccl_infos_); if (ret != SUCCESS) { GELOGE(ret, "Setaddrs Fail."); return ret; } // GE's new process: hccl declares the need for Workspace size, and GE allocates Workspace - ret = SetWorkspace(op_desc, kernel_hccl_infos_); + ret = SetWorkspace(op_desc_, kernel_hccl_infos_); if (ret != SUCCESS) { GELOGE(ret, "SetWorkspace Fail."); return ret; } - const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); - const auto input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); - const auto output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); - const auto workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc); - io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); - io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); - io_addrs_.insert(io_addrs_.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); GELOGI("HcclTaskInfo Init Success"); return SUCCESS; } @@ -238,7 +231,18 @@ Status HcclTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel * Status HcclTaskInfo::UpdateArgs() { GELOGI("HcclTaskInfo::UpdateArgs in."); - davinci_model_->SetTotalIOAddrs(io_addrs_); + const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); + input_data_addrs_ = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); + output_data_addrs_ = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); + workspace_data_addrs_ = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc_); + + vector io_addrs; + io_addrs.insert(io_addrs.end(), input_data_addrs_.begin(), input_data_addrs_.end()); + io_addrs.insert(io_addrs.end(), output_data_addrs_.begin(), output_data_addrs_.end()); + io_addrs.insert(io_addrs.end(), workspace_data_addrs_.begin(), workspace_data_addrs_.end()); + + davinci_model_->SetTotalIOAddrs(io_addrs); + GELOGI("HcclTaskInfo::UpdateArgs success."); return SUCCESS; } @@ -257,11 +261,9 @@ Status HcclTaskInfo::SetAddrs(const std::shared_ptr &op_desc, HcclReduceOp op_type = HCCL_REDUCE_SUM; GE_CHECK_NOTNULL(davinci_model_); GELOGI("Calc opType[%s] input address before. Node name[%s]", op_desc->GetType().c_str(), op_desc->GetName().c_str()); - vector input_data_addrs; - vector output_data_addrs; if (!davinci_model_->IsKnownNode()) { - input_data_addrs = ModelUtils::GetInputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); - output_data_addrs = ModelUtils::GetOutputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); + input_data_addrs_ = ModelUtils::GetInputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); + output_data_addrs_ = ModelUtils::GetOutputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); } void *input_data_addr = nullptr; void *output_data_addr = nullptr; @@ -273,8 +275,8 @@ Status HcclTaskInfo::SetAddrs(const std::shared_ptr &op_desc, output_data_addr = reinterpret_cast(reinterpret_cast(args_) + op_desc->GetInputsSize() + i); GELOGI("Hccl task info known input addr %p, output addr %p.", input_data_addr, output_data_addr); } else { - input_data_addr = input_data_addrs.empty() ? nullptr : input_data_addrs[i]; - output_data_addr = output_data_addrs.empty() ? nullptr : output_data_addrs[i]; + input_data_addr = input_data_addrs_.empty() ? nullptr : input_data_addrs_[i]; + output_data_addr = output_data_addrs_.empty() ? nullptr : output_data_addrs_[i]; } kernel_hccl_infos[i].inputDataAddr = input_data_addr; if (hccl_type == HCOMALLGATHER || hccl_type == HCOMRECEIVE || hccl_type == HVDCALLBACKALLGATHER) { @@ -364,8 +366,8 @@ Status HcclTaskInfo::SetWorkspace(const std::shared_ptr &op_desc, workspace_addr = reinterpret_cast(reinterpret_cast(args_) + op_desc->GetInputsSize() + op_desc->GetOutputsSize()); } else { - const auto workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); - workspace_addr = workspace_data_addrs.empty() ? nullptr : workspace_data_addrs[0]; + workspace_data_addrs_ = ModelUtils::GetWorkspaceDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); + workspace_addr = workspace_data_addrs_.empty() ? nullptr : workspace_data_addrs_[0]; } } } diff --git a/ge/graph/load/new_model_manager/task_info/hccl_task_info.h b/ge/graph/load/new_model_manager/task_info/hccl_task_info.h index 5498cf6c..f7ce3468 100644 --- a/ge/graph/load/new_model_manager/task_info/hccl_task_info.h +++ b/ge/graph/load/new_model_manager/task_info/hccl_task_info.h @@ -35,6 +35,7 @@ class HcclTaskInfo : public TaskInfo { ops_kernel_store_(nullptr), private_def_(nullptr), private_def_len_(0), + op_desc_(nullptr), args_(nullptr), args_offset_(0) {} @@ -75,7 +76,10 @@ class HcclTaskInfo : public TaskInfo { uint32_t private_def_len_; static std::mutex hccl_follow_stream_mutex_; vector kernel_hccl_infos_; - vector io_addrs_; + vector input_data_addrs_; + vector output_data_addrs_; + vector workspace_data_addrs_; + OpDescPtr op_desc_; void *args_; uint32_t args_offset_; }; diff --git a/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.cc b/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.cc index 944423ab..e5574e47 100644 --- a/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.cc +++ b/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.cc @@ -30,7 +30,11 @@ namespace ge { Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) { GELOGI("KernelExTaskInfo Init Start."); - GE_CHECK_NOTNULL(davinci_model); + if (davinci_model == nullptr) { + GELOGE(PARAM_INVALID, "davinci_model is null!"); + return PARAM_INVALID; + } + davinci_model_ = davinci_model; Status ret = SetStream(task_def.stream_id(), davinci_model_->GetStreamList()); if (ret != SUCCESS) { @@ -47,6 +51,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin GELOGE(INTERNAL_ERROR, "Init aicpu task info error, index is out of range!"); return INTERNAL_ERROR; } + op_desc_ = op_desc; // 2. Reconstruct kernelExDef.args to STR_FWK_OP_KERNEL STR_FWK_OP_KERNEL fwk_op_kernel = {0}; @@ -74,8 +79,8 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin return RT_ERROR_TO_GE_STATUS(rt_ret);) } - GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, ext_info_addr_=%p", op_desc->GetName().c_str(), - op_desc->GetType().c_str(), ext_info.size(), ext_info_addr_); + GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, ext_info_addr_=%p", op_desc_->GetName().c_str(), + op_desc_->GetType().c_str(), ext_info.size(), ext_info_addr_); // 2.1 get loop cond variable for tensor array write uint64_t step_id_addr = 0; @@ -128,7 +133,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin return RT_ERROR_TO_GE_STATUS(rt_ret);) GELOGI("KernelExTaskInfo knonw node Init Success."); - return SetIoAddrs(op_desc); + return SUCCESS; } // 3. Set workspaceaddr, inputOutputDataAddr @@ -192,7 +197,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin davinci_model_->SetZeroCopyAddr(op_desc, io_addrs, io_addrs.data(), input_output_addr_, addrs_size, 0); GELOGI("KernelExTaskInfo Init Success. session id: %lu", session_id); - return SetIoAddrs(op_desc); + return SUCCESS; } Status KernelExTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) { @@ -231,40 +236,36 @@ Status KernelExTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciMod return SUCCESS; } -Status KernelExTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) { +Status KernelExTaskInfo::UpdateArgs() { + GELOGI("KernelExTaskInfo::UpdateArgs in."); const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); - vector input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); - vector output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); - if (!op_desc->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { - io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); - io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); + vector input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); + vector output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); + vector io_addrs; + if (!op_desc_->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { + io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); + io_addrs.insert(io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end()); } else { string peer_input_name; - if (AttrUtils::GetStr(op_desc, ATTR_DYNAMIC_SHAPE_FIXED_ADDR, peer_input_name)) { + if (AttrUtils::GetStr(op_desc_, ATTR_DYNAMIC_SHAPE_FIXED_ADDR, peer_input_name)) { uint32_t output_index = davinci_model_->GetFixedAddrOutputIndex(peer_input_name); if (output_index > output_data_addrs.size()) { GELOGE(FAILED, "The output data addr size[%zu] and output index[%u] are inconsistent.", output_data_addrs.size(), output_index); return FAILED; } - io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); + io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); for (size_t i = 0; i < output_data_addrs.size(); ++i) { if (i == output_index) { void *fixed_addr = davinci_model_->GetCurrentFixedAddr(fixed_addr_offset_); - io_addrs_.emplace_back(fixed_addr); + io_addrs.emplace_back(fixed_addr); continue; } - io_addrs_.emplace_back(output_data_addrs[i]); + io_addrs.emplace_back(output_data_addrs[i]); } } } - - return SUCCESS; -} - -Status KernelExTaskInfo::UpdateArgs() { - GELOGI("KernelExTaskInfo::UpdateArgs in."); - davinci_model_->SetTotalIOAddrs(io_addrs_); + davinci_model_->SetTotalIOAddrs(io_addrs); GELOGI("KernelExTaskInfo::UpdateArgs success."); return SUCCESS; } diff --git a/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.h b/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.h index cb7647dc..e4d3e6fd 100644 --- a/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.h +++ b/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.h @@ -59,7 +59,6 @@ class KernelExTaskInfo : public TaskInfo { }; private: Status CopyTaskInfo(const domi::KernelExDef &kernel_def, const RuntimeParam &rts_param, const OpDescPtr &op_desc); - Status SetIoAddrs(const OpDescPtr &op_desc); uint32_t task_id_; uint32_t stream_id_; @@ -70,7 +69,7 @@ class KernelExTaskInfo : public TaskInfo { void *input_output_addr_; void *ext_info_addr_; void *dump_args_; - vector io_addrs_; + OpDescPtr op_desc_ = nullptr; uint32_t args_offset_ = 0; int64_t fixed_addr_offset_ = 0; }; diff --git a/ge/graph/load/new_model_manager/task_info/kernel_task_info.cc b/ge/graph/load/new_model_manager/task_info/kernel_task_info.cc index 1906a0c9..ce31ef30 100755 --- a/ge/graph/load/new_model_manager/task_info/kernel_task_info.cc +++ b/ge/graph/load/new_model_manager/task_info/kernel_task_info.cc @@ -53,6 +53,9 @@ const int kArgsAttrHandle = 4; } // namespace namespace ge { +KernelTaskInfo::SuperKernelTaskInfo KernelTaskInfo::skt_info_ = { + 0, 0, 0, 0, nullptr, nullptr, {}, {}, {}, {}, {}, RT_KERNEL_DEFAULT, kInvalidGroupKey, 0, nullptr}; + Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) { GE_CHECK_NOTNULL(davinci_model); davinci_model_ = davinci_model; @@ -134,7 +137,6 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci ret = InitCceTask(kernel_def); } - SetIoAddrs(op_desc_); GELOGD("KernelTaskInfo init finish, result=%u.", ret); return ret; } @@ -146,10 +148,9 @@ Status KernelTaskInfo::SaveSKTDumpInfo() { return SUCCESS; } // all op in super kernel share one taskid and streamid - const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - for (size_t i = 0; i < skt_info.op_desc_list.size(); i++) { - davinci_model_->SaveDumpTask(skt_info.last_task_id, skt_info.last_stream_id, skt_info.op_desc_list[i], - skt_info.dump_args_list[i]); + for (size_t i = 0; i < skt_info_.op_desc_list.size(); i++) { + davinci_model_->SaveDumpTask(skt_info_.last_task_id, skt_info_.last_stream_id, skt_info_.op_desc_list[i], + skt_info_.dump_args_list[i]); } return SUCCESS; } @@ -163,10 +164,9 @@ void KernelTaskInfo::UpdateSKTTaskId() { GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret); return; } - SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - skt_info.last_task_id = task_id; - skt_info.last_stream_id = stream_id; - skt_id_ = skt_info.last_task_id; + skt_info_.last_task_id = task_id; + skt_info_.last_stream_id = stream_id; + skt_id_ = skt_info_.last_task_id; GELOGI("UpdateTaskId:UpdateSKTTaskId [%u],stream id [%u]", task_id, stream_id); } @@ -191,25 +191,23 @@ Status KernelTaskInfo::SKTFinalize() { UpdateSKTTaskId(); GE_CHK_STATUS_RET(SaveSKTDumpInfo(), "skt save dump info failed"); GELOGI("SuperKernel Distribute [skt_id:%u]", skt_id_); - SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - skt_info.kernel_list.clear(); - skt_info.arg_list.clear(); - skt_info.dump_flag_list.clear(); - skt_info.op_desc_list.clear(); - skt_info.dump_args_list.clear(); - skt_info.last_stream = nullptr; - skt_info.last_block_dim = 0; - skt_info.last_sm_desc = sm_desc_; - skt_info.last_group_key = kInvalidGroupKey; - skt_info.last_dump_flag = RT_KERNEL_DEFAULT; - skt_info.last_dump_args = 0; - skt_info.last_op = nullptr; + skt_info_.kernel_list.clear(); + skt_info_.arg_list.clear(); + skt_info_.dump_flag_list.clear(); + skt_info_.op_desc_list.clear(); + skt_info_.dump_args_list.clear(); + skt_info_.last_stream = nullptr; + skt_info_.last_block_dim = 0; + skt_info_.last_sm_desc = sm_desc_; + skt_info_.last_group_key = kInvalidGroupKey; + skt_info_.last_dump_flag = RT_KERNEL_DEFAULT; + skt_info_.last_dump_args = 0; + skt_info_.last_op = nullptr; return SUCCESS; } uint32_t KernelTaskInfo::GetDumpFlag() { - const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - for (auto flag : skt_info.dump_flag_list) { + for (auto flag : skt_info_.dump_flag_list) { if (flag == RT_KERNEL_DUMPFLAG) { return RT_KERNEL_DUMPFLAG; } @@ -218,20 +216,19 @@ uint32_t KernelTaskInfo::GetDumpFlag() { } Status KernelTaskInfo::SuperKernelLaunch() { - SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - if (skt_info.kernel_list.empty()) { + if (skt_info_.kernel_list.empty()) { GELOGI("SuperKernelLaunch: Skt_kernel_list has no task, just return"); return SUCCESS; } rtError_t rt_ret; - auto &skt_kernel_list = skt_info.kernel_list; - auto &skt_arg_list = skt_info.arg_list; + auto &skt_kernel_list = skt_info_.kernel_list; + auto &skt_arg_list = skt_info_.arg_list; GELOGI("SuperKernelLaunch: Skt_kernel_list size[%zu] skt_arg_list[%zu]", skt_kernel_list.size(), skt_arg_list.size()); if (skt_kernel_list.size() == kSKTSingleSize && skt_arg_list.size() == kSKTSingleSize) { - rt_ret = rtKernelLaunchWithFlag(skt_info.kernel_list[0], static_cast(skt_info.last_block_dim), - skt_info.arg_list[0], skt_info.last_args_size, - static_cast(skt_info.last_sm_desc), skt_info.last_stream, - skt_info.last_dump_flag); + rt_ret = rtKernelLaunchWithFlag(skt_info_.kernel_list[0], static_cast(skt_info_.last_block_dim), + skt_info_.arg_list[0], skt_info_.last_args_size, + static_cast(skt_info_.last_sm_desc), skt_info_.last_stream, + skt_info_.last_dump_flag); if (rt_ret != RT_ERROR_NONE) { GELOGE(RT_FAILED, "SuperKernelLaunch: Call rt api failed, ret: 0x%X", rt_ret); return RT_ERROR_TO_GE_STATUS(rt_ret); @@ -250,14 +247,14 @@ Status KernelTaskInfo::SuperKernelLaunch() { } // Call the fuse API std::unique_ptr superKernel = nullptr; - ge_ret = factory->FuseKernels(skt_kernel_list, skt_arg_list, skt_info.last_block_dim, superKernel); + ge_ret = factory->FuseKernels(skt_kernel_list, skt_arg_list, skt_info_.last_block_dim, superKernel); if (ge_ret != SUCCESS) { GELOGE(ge_ret, "SuperKernelLaunch: fuse call failed"); return ge_ret; } // Launch a super kernel skt_dump_flag_ = GetDumpFlag(); - ge_ret = superKernel->Launch(skt_info.last_stream, skt_dump_flag_); + ge_ret = superKernel->Launch(skt_info_.last_stream, skt_dump_flag_); if (ge_ret != SUCCESS) { GELOGE(ge_ret, "SuperKernelLaunch: launch failed"); return ge_ret; @@ -272,26 +269,23 @@ Status KernelTaskInfo::SuperKernelLaunch() { } Status KernelTaskInfo::SaveSuperKernelInfo() { - SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - skt_info.kernel_list.push_back(stub_func_); - skt_info.arg_list.push_back(args_); - skt_info.last_stream = stream_; - skt_info.last_block_dim = block_dim_; - skt_info.last_args_size = args_size_; - skt_info.last_sm_desc = sm_desc_; - skt_info.last_dump_flag = dump_flag_; - skt_info.dump_flag_list.push_back(dump_flag_); - skt_info.op_desc_list.push_back(op_desc_); - skt_info.dump_args_list.push_back(reinterpret_cast(skt_dump_args_)); - skt_info.last_group_key = group_key_; - skt_info.last_dump_args = reinterpret_cast(skt_dump_args_); - skt_info.last_op = op_desc_; + skt_info_.kernel_list.push_back(stub_func_); + skt_info_.arg_list.push_back(args_); + skt_info_.last_stream = stream_; + skt_info_.last_block_dim = block_dim_; + skt_info_.last_args_size = args_size_; + skt_info_.last_sm_desc = sm_desc_; + skt_info_.last_dump_flag = dump_flag_; + skt_info_.dump_flag_list.push_back(dump_flag_); + skt_info_.op_desc_list.push_back(op_desc_); + skt_info_.dump_args_list.push_back(reinterpret_cast(skt_dump_args_)); + skt_info_.last_group_key = group_key_; + skt_info_.last_dump_args = reinterpret_cast(skt_dump_args_); + skt_info_.last_op = op_desc_; // last node in a stream, just launch if (IsMarkedLastNode()) { return SuperKernelLaunch(); } - - GELOGI("Save Current task [block_dim:%u, size:%zu].", block_dim_, skt_info.kernel_list.size()); return SUCCESS; } @@ -328,9 +322,8 @@ bool KernelTaskInfo::IsMarkedFirstNode() { // then may be saved to skt task list; else // call skt launch those saved tasks before bool KernelTaskInfo::FirstCallSKTLaunchCheck() { - const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); - return ((block_dim_ != skt_info.last_block_dim) || (stream_ != skt_info.last_stream) || - (has_group_key_ && (group_key_ != skt_info.last_group_key))); + return ((block_dim_ != skt_info_.last_block_dim) || (stream_ != skt_info_.last_stream) || + (has_group_key_ && (group_key_ != skt_info_.last_group_key))); } // current task has group_id or has n ATTR_N_BATCH_SPLIT then save it to skt task list; else @@ -369,6 +362,7 @@ Status KernelTaskInfo::SuperKernelDistribute() { GELOGE(ret, "Call SuperKernelLaunch failed!"); return ret; } + GELOGI("Save Current task [block_dim:%u, size:%zu].", block_dim_, skt_info_.kernel_list.size()); } return SUCCESS; } @@ -397,11 +391,10 @@ Status KernelTaskInfo::Distribute() { call_save_dump_ = true; } else { /* default: not skt launch */ - const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); GELOGD( "KernelTaskInfo Distribute Start, sktenable:%d taskid:%u sktid:%u last_sktid:%u stubfunc_name:%s " "stubfunc:%p blockdim:%u stream:%p", - call_skt, task_id_, skt_id_, skt_info.last_task_id, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); + call_skt, task_id_, skt_id_, skt_info_.last_task_id, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); // l1 fusion enable and env flag open (kCloseSkt for skt debug) bool open_dump = false; auto all_dump_model = davinci_model_->GetDumpProperties().GetAllDumpModel(); @@ -429,30 +422,23 @@ Status KernelTaskInfo::Distribute() { "KernelTaskInfo Distribute Success. sktenable:%d taskid:%d sktid:%d stubfunc_name:%s stubfunc:%p " "blockdim:%d stream:%p", call_skt, task_id_, skt_id_, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); - op_desc_.reset(); // Not hold OpDesc after distribute. return SUCCESS; } -void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) { - const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); - vector input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); - vector output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); - - io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); - io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); - if (kernel_type_ == ccKernelType::TE) { - vector workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc); - io_addrs_.insert(io_addrs_.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); - } -} - Status KernelTaskInfo::UpdateArgs() { GELOGI("KernelTaskInfo::UpdateArgs in."); + const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); + vector input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); + vector output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); + vector io_addrs; + io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); + io_addrs.insert(io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end()); if (kernel_type_ == ccKernelType::TE) { - davinci_model_->SetTotalIOAddrs(io_addrs_); + vector workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc_); + io_addrs.insert(io_addrs.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); + davinci_model_->SetTotalIOAddrs(io_addrs); } else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) { - vector io_addrs = io_addrs_; davinci_model_->UpdateKnownZeroCopyAddr(io_addrs); uintptr_t io_addr = reinterpret_cast(args_addr.get()) + sizeof(aicpu::AicpuParamHead); auto addrs_size = sizeof(uint64_t) * io_addrs.size(); @@ -803,6 +789,7 @@ Status KernelTaskInfo::InitCceTask(const domi::KernelDef &kernel_def) { GELOGE(FAILED, "flowtable is null."); return FAILED; } + flowtable_size_ = flowtable.size(); } // get smDesc stored in model @@ -867,14 +854,14 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k GELOGI("Do InitAicpuTask"); so_name_ = kernel_def.so_name(); kernel_name_ = kernel_def.kernel_name(); + GELOGI("node[%s] test so name %s, kernel name %s", op_desc_->GetName().c_str(), so_name_.c_str(), + kernel_name_.c_str()); OpDescPtr op_desc = davinci_model_->GetOpByIndex(op_index); if (op_desc == nullptr) { GELOGE(INTERNAL_ERROR, "index is out of range, index: %u", op_index); return INTERNAL_ERROR; } - GELOGI("node[%s] test so name %s, kernel name %s", op_desc->GetName().c_str(), so_name_.c_str(), - kernel_name_.c_str()); if (kernel_type_ == ccKernelType::CUST_AI_CPU) { bool loaded = false; @@ -898,8 +885,8 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k GELOGE(init_ret, "Init aicpu task ext info failed, ext_info size=%zu", ext_info.size()); return init_ret; } - GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, aicpu_ext_info_addr_=%p", op_desc->GetName().c_str(), - op_desc->GetType().c_str(), ext_info.size(), aicpu_ext_info_addr_); + GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, aicpu_ext_info_addr_=%p", op_desc_->GetName().c_str(), + op_desc_->GetType().c_str(), ext_info.size(), aicpu_ext_info_addr_); aicpu_param_head->extInfoAddr = reinterpret_cast(aicpu_ext_info_addr_); aicpu_param_head->extInfoLength = static_cast(ext_info.size()); diff --git a/ge/graph/load/new_model_manager/task_info/kernel_task_info.h b/ge/graph/load/new_model_manager/task_info/kernel_task_info.h index cb96797f..7717edd3 100644 --- a/ge/graph/load/new_model_manager/task_info/kernel_task_info.h +++ b/ge/graph/load/new_model_manager/task_info/kernel_task_info.h @@ -38,6 +38,7 @@ class KernelTaskInfo : public TaskInfo { flowtable_(nullptr), block_dim_(0), args_size_(0), + flowtable_size_(0), task_id_(0), stream_id_(0), so_name_(""), @@ -45,6 +46,7 @@ class KernelTaskInfo : public TaskInfo { kernel_type_(ccKernelType::CCE_AI_CORE), dump_flag_(RT_KERNEL_DEFAULT), dump_args_(nullptr), + op_desc_(nullptr), davinci_model_(nullptr), skt_id_(0), stub_func_name_(""), @@ -126,7 +128,6 @@ class KernelTaskInfo : public TaskInfo { Status SuperKernelDistribute(); bool IsL1FusionOp(const OpDescPtr &op_desc); - void SetIoAddrs(const OpDescPtr &op_desc); // For super kernel Status SaveSKTDumpInfo(); @@ -147,6 +148,7 @@ class KernelTaskInfo : public TaskInfo { void *flowtable_; uint32_t block_dim_; uint32_t args_size_; + uint32_t flowtable_size_; uint32_t task_id_; uint32_t stream_id_; std::string so_name_; @@ -154,8 +156,7 @@ class KernelTaskInfo : public TaskInfo { ccKernelType kernel_type_; uint32_t dump_flag_; void *dump_args_; - OpDescPtr op_desc_; // Clear after distribute. - vector io_addrs_; + OpDescPtr op_desc_; DavinciModel *davinci_model_; uint32_t args_offset_ = 0; uint32_t hybrid_args_offset_ = 0; @@ -185,6 +186,25 @@ class KernelTaskInfo : public TaskInfo { void *output_addrs = nullptr; void *attr_handle = nullptr; } custom_info_; + + // For super kernel + static struct SuperKernelTaskInfo { + uint32_t last_block_dim; + uint32_t last_args_size; + uint32_t last_task_id; + uint32_t last_stream_id; + void *last_stream; + void *last_sm_desc; + std::vector kernel_list; + std::vector arg_list; + std::vector dump_flag_list; + std::vector op_desc_list; + std::vector dump_args_list; + uint32_t last_dump_flag; + int64_t last_group_key; + uintptr_t last_dump_args; + OpDescPtr last_op; + } skt_info_; }; } // namespace ge #endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_TASK_INFO_H_ diff --git a/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc b/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc index 6851b1bc..aa37dd07 100755 --- a/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc +++ b/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc @@ -30,13 +30,14 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da return ret; } - const domi::MemcpyAsyncDef &memcpy_async = task_def.memcpy_async(); - count_ = memcpy_async.count(); - kind_ = memcpy_async.kind(); - dst_max_ = memcpy_async.dst_max(); - OpDescPtr op_desc = davinci_model_->GetOpByIndex(memcpy_async.op_index()); + memcpy_async_ = task_def.memcpy_async(); + count_ = memcpy_async_.count(); + kind_ = memcpy_async_.kind(); + dst_max_ = memcpy_async_.dst_max(); + OpDescPtr op_desc = davinci_model_->GetOpByIndex(memcpy_async_.op_index()); + op_desc_ = op_desc; if (op_desc == nullptr) { - GELOGE(INTERNAL_ERROR, "Task op index:%u out of range", memcpy_async.op_index()); + GELOGE(INTERNAL_ERROR, "Task op index:%u out of range", memcpy_async_.op_index()); return INTERNAL_ERROR; } @@ -51,7 +52,7 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da } const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); - ret = ModelUtils::GetRtAddress(rts_param, memcpy_async.src(), src_); + ret = ModelUtils::GetRtAddress(rts_param, memcpy_async_.src(), src_); if (ret != SUCCESS) { return ret; } @@ -60,32 +61,23 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da vector memory_type_list; (void)AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memory_type_list); if (!memory_type_list.empty() && memory_type_list[0] == RT_MEMORY_TS_4G) { // TS Feature, Just one. - uint64_t mem_offset = memcpy_async.dst() - rts_param.logic_mem_base; - dst_ = static_cast(rts_param.ts_mem_mall->Acquire(mem_offset, memcpy_async.dst_max())); + uint64_t mem_offset = memcpy_async_.dst() - rts_param.logic_mem_base; + dst_ = static_cast(rts_param.ts_mem_mall->Acquire(mem_offset, memcpy_async_.dst_max())); if (dst_ == nullptr) { return FAILED; } } else { - ret = ModelUtils::GetRtAddress(rts_param, memcpy_async.dst(), dst_); + ret = ModelUtils::GetRtAddress(rts_param, memcpy_async_.dst(), dst_); if (ret != SUCCESS) { return ret; } } GELOGI("MemcpyAsyncTaskInfo Init Success, logic[0x%lx, 0x%lx], src:%p, dst:%p, max:%lu, count:%lu", - memcpy_async.src(), memcpy_async.dst(), src_, dst_, dst_max_, count_); + memcpy_async_.src(), memcpy_async_.dst(), src_, dst_, dst_max_, count_); davinci_model_->DisableZeroCopy(src_); davinci_model_->DisableZeroCopy(dst_); - - io_addrs_.emplace_back(reinterpret_cast(src_)); - if (op_desc->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { - void *fixed_addr = davinci_model_->GetCurrentFixedAddr(fixed_addr_offset_); - io_addrs_.emplace_back(fixed_addr); - } else { - io_addrs_.emplace_back(reinterpret_cast(dst_)); - } - return SUCCESS; } @@ -126,7 +118,25 @@ Status MemcpyAsyncTaskInfo::CalculateArgs(const domi::TaskDef &task_def, Davinci Status MemcpyAsyncTaskInfo::UpdateArgs() { GELOGI("MemcpyAsyncTaskInfo::UpdateArgs in."); GE_CHECK_NOTNULL(davinci_model_); - davinci_model_->SetTotalIOAddrs(io_addrs_); + Status ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async_.src(), src_); + if (ret != SUCCESS) { + return ret; + } + + ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async_.dst(), dst_); + if (ret != SUCCESS) { + return ret; + } + + vector io_addrs; + io_addrs.emplace_back(reinterpret_cast(src_)); + if (op_desc_->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { + void *fixed_addr = davinci_model_->GetCurrentFixedAddr(fixed_addr_offset_); + io_addrs.emplace_back(fixed_addr); + } else { + io_addrs.emplace_back(reinterpret_cast(dst_)); + } + davinci_model_->SetTotalIOAddrs(io_addrs); GELOGI("MemcpyAsyncTaskInfo::UpdateArgs success."); return SUCCESS; diff --git a/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h b/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h index d7506840..7e74ab6f 100755 --- a/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h +++ b/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.h @@ -44,10 +44,11 @@ class MemcpyAsyncTaskInfo : public TaskInfo { uint8_t *src_; uint64_t count_; uint32_t kind_; - vector io_addrs_; + OpDescPtr op_desc_; int64_t fixed_addr_offset_; DavinciModel *davinci_model_ = nullptr; uint32_t args_offset_ = 0; + domi::MemcpyAsyncDef memcpy_async_; }; } // namespace ge #endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_