From 6cf3a44a9fde9955730397a7fcce396012870d49 Mon Sep 17 00:00:00 2001 From: weiyang Date: Thu, 24 Dec 2020 16:33:04 +0800 Subject: [PATCH] dynamic shape --- .../task_info/kernel_ex_task_info.cc | 22 +++--- .../task_info/kernel_ex_task_info.h | 2 + .../task_info/kernel_task_info.cc | 37 +++++----- .../task_info/kernel_task_info.h | 1 + .../compiledsubgraph/known_node_executor.cc | 4 ++ tests/ut/ge/graph/ge_executor_unittest.cc | 67 +++++++++++++++++++ 6 files changed, 106 insertions(+), 27 deletions(-) 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 c146912a..1f3a4592 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 @@ -97,9 +97,10 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin return FAILED;) // 2.3 Create session GE_CHECK_NOTNULL(ModelManager::GetInstance()); - GE_IF_BOOL_EXEC(ModelManager::GetInstance()->CreateAicpuSession(session_id) != SUCCESS, - GELOGE(FAILED, "CreateAicpuSession error. session id: %lu", session_id); - return FAILED;) + ret = ModelManager::GetInstance()->CreateAicpuSession(session_id); + GE_IF_BOOL_EXEC(ret != SUCCESS, + GELOGE(ret, "CreateAicpuSession error. session id: %lu", session_id); + return ret;) kernel_buf_size_ = sizeof(STR_FWK_OP_KERNEL); if (davinci_model_->IsKnownNode()) { @@ -128,6 +129,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin return RT_ERROR_TO_GE_STATUS(rt_ret);) SetIoAddrs(op_desc); + InitDumpTask(input_output_addr, op_desc); GELOGI("KernelExTaskInfo knonw node Init Success."); return SUCCESS; } @@ -162,11 +164,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMemcpy to input_output_addr_ error: 0x%X", rt_ret); return RT_ERROR_TO_GE_STATUS(rt_ret);) - if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(), - op_desc->GetName())) { - dump_flag_ = RT_KERNEL_DUMPFLAG; - dump_args_ = input_output_addr_; - } + InitDumpTask(input_output_addr_, op_desc); if (davinci_model_->GetOpDugReg()) { GELOGI("Op debug is open in kernel ex task info"); dump_args_ = input_output_addr_; @@ -196,6 +194,14 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin return SUCCESS; } +void KernelExTaskInfo::InitDumpTask(void *addr, const OpDescPtr &op_desc) { + if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(), + op_desc->GetName())) { + dump_flag_ = RT_KERNEL_DUMPFLAG; + dump_args_ = input_output_addr_; + } +} + Status KernelExTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) { auto kernel_ex_def = task_def.kernel_ex(); uint32_t op_index = kernel_ex_def.op_index(); 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 8820aadc..f6873c6c 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 @@ -61,6 +61,8 @@ class KernelExTaskInfo : public TaskInfo { Status CopyTaskInfo(const domi::KernelExDef &kernel_def, const RuntimeParam &rts_param, const OpDescPtr &op_desc); void SetIoAddrs(const OpDescPtr &op_desc); + void InitDumpTask(void *addr, const OpDescPtr &op_desc); + uint32_t task_id_; uint32_t stream_id_; uint32_t dump_flag_; 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 31725e11..83bf2779 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 @@ -566,6 +566,8 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne OpDescPtr op_desc = davinci_model_->GetOpByIndex(ctx_.opIndex); GE_CHECK_NOTNULL(op_desc); if (davinci_model_->IsKnownNode()) { + args_ = davinci_model_->GetCurrentArgsAddr(args_offset_); + InitDumpTask(offset); return SUCCESS; } @@ -630,15 +632,7 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne return FAILED; } skt_dump_args_ = static_cast(args_) + offset; - if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(), - op_desc->GetName())) { - if (IsL1FusionOp(op_desc)) { - dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG; - } else { - dump_flag_ = RT_KERNEL_DUMPFLAG; - } - dump_args_ = static_cast(args_) + offset; - } + InitDumpTask(offset); GE_CHK_BOOL_TRUE_EXEC_INFO(davinci_model_->GetOpDugReg(), dump_args_ = static_cast(args_) + offset, "Op debug is open in TVM task info"); @@ -904,6 +898,8 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k aicpu_param_head->extInfoLength = static_cast(ext_info.size()); if (davinci_model_->IsKnownNode()) { + args_ = davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_); + InitDumpTask(sizeof(aicpu::AicpuParamHead)); return SUCCESS; } const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); @@ -937,16 +933,7 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k GELOGE(RT_FAILED, "Call rt api(rtMemcpy) failed, ret: 0x%X", rt_ret); return RT_ERROR_TO_GE_STATUS(rt_ret); } - - if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(), - op_desc->GetName())) { - if (IsL1FusionOp(op_desc)) { - dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG; - } else { - dump_flag_ = RT_KERNEL_DUMPFLAG; - } - dump_args_ = static_cast(args_) + sizeof(aicpu::AicpuParamHead); - } + InitDumpTask(sizeof(aicpu::AicpuParamHead)); if (davinci_model_->GetOpDugReg()) { GELOGI("Op debug is open in aicpu task info"); dump_args_ = static_cast(args_) + sizeof(aicpu::AicpuParamHead); @@ -960,6 +947,18 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k return SUCCESS; } +void KernelTaskInfo::InitDumpTask(uint32_t offset) { + if (davinci_model_->GetDumpProperties().IsLayerNeedDump(davinci_model_->Name(), davinci_model_->OmName(), + op_desc_->GetName())) { + if (IsL1FusionOp(op_desc_)) { + dump_flag_ = RT_FUSION_KERNEL_DUMPFLAG; + } else { + dump_flag_ = RT_KERNEL_DUMPFLAG; + } + dump_args_ = static_cast(args_) + offset; + } +} + Status KernelTaskInfo::InitAicpuTaskExtInfo(const std::string &ext_info) { if (ext_info.empty()) { return SUCCESS; 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 493d3981..cea25320 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 @@ -128,6 +128,7 @@ class KernelTaskInfo : public TaskInfo { Status SuperKernelDistribute(); bool IsL1FusionOp(const OpDescPtr &op_desc); void SetIoAddrs(const OpDescPtr &op_desc); + void InitDumpTask(uint32_t offset); // For super kernel Status SaveSKTDumpInfo(); diff --git a/ge/hybrid/node_executor/compiledsubgraph/known_node_executor.cc b/ge/hybrid/node_executor/compiledsubgraph/known_node_executor.cc index f985a3d0..c914ac1b 100755 --- a/ge/hybrid/node_executor/compiledsubgraph/known_node_executor.cc +++ b/ge/hybrid/node_executor/compiledsubgraph/known_node_executor.cc @@ -123,6 +123,10 @@ Status KnownNodeTask::Init(TaskContext &context) { davinci_model_->GetRuntimeParam().mem_base, davinci_model_->GetRuntimeParam().mem_size); } if (!load_flag_) { + auto dump_properties = context.GetDumpProperties(); + if (dump_properties.IsDumpOpen()) { + davinci_model_->SetDumpProperties(dump_properties); + } GE_CHK_STATUS_RET(davinci_model_->Init(), "KnownNodeExecutor::InitDavinciModel failed."); load_flag_ = true; } else { diff --git a/tests/ut/ge/graph/ge_executor_unittest.cc b/tests/ut/ge/graph/ge_executor_unittest.cc index 94415759..3d04fd0c 100644 --- a/tests/ut/ge/graph/ge_executor_unittest.cc +++ b/tests/ut/ge/graph/ge_executor_unittest.cc @@ -36,6 +36,9 @@ #include "graph/load/new_model_manager/davinci_model.h" #include "graph/load/new_model_manager/davinci_model_parser.h" #include "graph/load/new_model_manager/model_manager.h" +#include "graph/load/new_model_manager/task_info/kernel_task_info.h" +#include "graph/load/new_model_manager/task_info/kernel_ex_task_info.h" +#include "ge/common/dump/dump_properties.h" #include "graph/manager/graph_mem_allocator.h" #include "graph/utils/graph_utils.h" #include "proto/ge_ir.pb.h" @@ -79,6 +82,33 @@ class DModelListener : public ge::ModelListener { shared_ptr g_label_call_back(new DModelListener()); +static ge::OpDescPtr CreateOpDesc(string name = "", string type = "") { + auto op_desc = std::make_shared(name, type); + op_desc->SetStreamId(0); + op_desc->SetId(0); + + ge::AttrUtils::SetFloat(op_desc, ge::ATTR_NAME_ALPHA, 0); + ge::AttrUtils::SetFloat(op_desc, ge::ATTR_NAME_BETA, 0); + + op_desc->SetWorkspace({}); + ; + op_desc->SetWorkspaceBytes({}); + op_desc->SetInputOffset({}); + op_desc->SetOutputOffset({}); + + ge::AttrUtils::SetListStr(op_desc, ge::ATTR_NAME_WEIGHT_NAME, {}); + ge::AttrUtils::SetInt(op_desc, ge::POOLING_ATTR_MODE, 0); + ge::AttrUtils::SetInt(op_desc, ge::POOLING_ATTR_PAD_MODE, 0); + ge::AttrUtils::SetInt(op_desc, ge::POOLING_ATTR_DATA_MODE, 0); + ge::AttrUtils::SetInt(op_desc, ge::POOLING_ATTR_CEIL_MODE, 0); + ge::AttrUtils::SetInt(op_desc, ge::POOLING_ATTR_NAN_OPT, 0); + ge::AttrUtils::SetListInt(op_desc, ge::POOLING_ATTR_WINDOW, {}); + ge::AttrUtils::SetListInt(op_desc, ge::POOLING_ATTR_PAD, {}); + ge::AttrUtils::SetListInt(op_desc, ge::POOLING_ATTR_STRIDE, {}); + ge::AttrUtils::SetListInt(op_desc, ge::ATTR_NAME_ACTIVE_STREAM_LIST, {1, 1}); + ge::AttrUtils::SetInt(op_desc, ge::ATTR_NAME_STREAM_SWITCH_COND, 0); + return op_desc; +} /* TEST_F(UtestGeExecutor, fail_UnloadModel_model_manager_stop_unload_error) { uint32_t model_id = 1; @@ -104,4 +134,41 @@ TEST_F(UtestGeExecutor, InitFeatureMapAndP2PMem_failed) { model.is_feature_map_mem_has_inited_ = true; EXPECT_EQ(model.InitFeatureMapAndP2PMem(nullptr, 0), PARAM_INVALID); } + +TEST_F(UtestGeExecutor, kernel_InitDumpTask) { + DavinciModel model(0, g_label_call_back); + model.om_name_ = "testom"; + model.name_ = "test"; + OpDescPtr op_desc = CreateOpDesc("test", "test"); + + std::map> model_dump_properties_map; + std::set s; + model_dump_properties_map[DUMP_ALL_MODEL] = s; + DumpProperties dp; + dp.model_dump_properties_map_ = model_dump_properties_map; + model.SetDumpProperties(dp); + + KernelTaskInfo kernel_task_info; + kernel_task_info.davinci_model_ = &model; + kernel_task_info.op_desc_ = op_desc; + kernel_task_info.InitDumpTask(0); +} + +TEST_F(UtestGeExecutor, kernel_ex_InitDumpTask) { + DavinciModel model(0, g_label_call_back); + model.om_name_ = "testom"; + model.name_ = "test"; + OpDescPtr op_desc = CreateOpDesc("test", "test"); + + std::map> model_dump_properties_map; + std::set s; + model_dump_properties_map[DUMP_ALL_MODEL] = s; + DumpProperties dp; + dp.model_dump_properties_map_ = model_dump_properties_map; + model.SetDumpProperties(dp); + + KernelExTaskInfo kernel_ex_task_info; + kernel_ex_task_info.davinci_model_ = &model; + kernel_ex_task_info.InitDumpTask(nullptr, op_desc); +} } \ No newline at end of file