diff --git a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc index 5b65f57e..8fc8cc6c 100644 --- a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc +++ b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc @@ -53,6 +53,7 @@ Status KernelExTaskInfo::InitTaskExtInfo(const std::string &ext_info, const OpDe "Parse kernel ext info failed, kernel_ext_info_size=%zu.", ext_info.size()); GE_CHK_STATUS_RET(ext_handle->UpdateExecuteMode(true), "UpdateExecuteMode failed."); GELOGD("Update aicpu_task ext_info bit_map execute mode to 1."); + topic_type_flag_ = ext_handle->GetTopicTypeFlag(); bool all_shape = false; (void)AttrUtils::GetBool(op_desc, kAicpuAllshape, all_shape); @@ -407,6 +408,14 @@ Status KernelExTaskInfo::CopyTaskInfo(const domi::KernelExDef &kernel_def, const Status KernelExTaskInfo::Distribute() { GELOGI("KernelExTaskInfo Distribute Start."); + // Use the fifth and sixth bits of dump_flag_ indicate the value of topic_type. + // xxxxxxxx xxxxxxxx xxxxxxxx xx00xxxx: DEVICE_ONLY + // xxxxxxxx xxxxxxxx xxxxxxxx xx01xxxx: DEVICE_FIRST + // xxxxxxxx xxxxxxxx xxxxxxxx xx10xxxx: HOST_ONLY + // xxxxxxxx xxxxxxxx xxxxxxxx xx11xxxx: HOST_FIRST + if (topic_type_flag_ > 0) { + dump_flag_ = dump_flag_ | topic_type_flag_; + } rtError_t rt_ret = rtKernelLaunchEx(kernel_buf_, kernel_buf_size_, dump_flag_, stream_); if (rt_ret != RT_ERROR_NONE) { REPORT_CALL_ERROR("E19999", "Call rtKernelLaunchEx failed, ret:0x%X", diff --git a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h index 71153c31..bcc17168 100644 --- a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h +++ b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h @@ -76,6 +76,7 @@ class KernelExTaskInfo : public TaskInfo { vector io_addrs_; uint32_t args_offset_ = 0; int64_t fixed_addr_offset_ = 0; + int32_t topic_type_flag_ = -1; }; } // namespace ge #endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_EX_TASK_INFO_H_ diff --git a/ge/graph/load/model_manager/task_info/kernel_task_info.cc b/ge/graph/load/model_manager/task_info/kernel_task_info.cc index 9ac77531..b60529d5 100755 --- a/ge/graph/load/model_manager/task_info/kernel_task_info.cc +++ b/ge/graph/load/model_manager/task_info/kernel_task_info.cc @@ -431,6 +431,14 @@ Status KernelTaskInfo::Distribute() { int64_t env_flag = (res == EN_OK) ? strtol(skt_enable_env, nullptr, kBaseInt) : kStrtolFail; bool call_skt = ((env_flag != 0) || is_l1_fusion_enable_); if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) { + if (topic_type_flag_ > 0) { + // Use the fifth and sixth bits of dump_flag_ indicate the value of topic_type. + // xxxxxxxx xxxxxxxx xxxxxxxx xx00xxxx: DEVICE_ONLY + // xxxxxxxx xxxxxxxx xxxxxxxx xx01xxxx: DEVICE_FIRST + // xxxxxxxx xxxxxxxx xxxxxxxx xx10xxxx: HOST_ONLY + // xxxxxxxx xxxxxxxx xxxxxxxx xx11xxxx: HOST_FIRST + dump_flag_ = dump_flag_ | topic_type_flag_; + } GELOGI("distribute task info kernel_type %d, flag %d", kernel_type_, dump_flag_); // blockDim is reserved parameter, set to 1 rt_ret = rtCpuKernelLaunchWithFlag(reinterpret_cast(so_name_.c_str()), @@ -1116,6 +1124,7 @@ Status KernelTaskInfo::InitAicpuTaskExtInfo(const std::string &ext_info) { GELOGD("Update aicpu_task ext_info session_info session_id is %lu", davinci_model_->GetSessionId()); GE_CHK_STATUS_RET(ext_handle->UpdateExecuteMode(true), "UpdateExecuteMode failed."); GELOGD("Update aicpu_task ext_info bit_map execute mode to 1."); + topic_type_flag_ = ext_handle->GetTopicTypeFlag(); bool all_shape = false; (void)AttrUtils::GetBool(op_desc_, kAicpuAllshape, all_shape); diff --git a/ge/graph/load/model_manager/task_info/kernel_task_info.h b/ge/graph/load/model_manager/task_info/kernel_task_info.h index 4156c511..83df8736 100644 --- a/ge/graph/load/model_manager/task_info/kernel_task_info.h +++ b/ge/graph/load/model_manager/task_info/kernel_task_info.h @@ -169,6 +169,7 @@ class KernelTaskInfo : public TaskInfo { uint16_t io_addr_offset_ = 0; bool l2_buffer_on_ = false; bool call_save_dump_ = false; + int32_t topic_type_flag_ = -1; // aicpu ext_info device mem void *aicpu_ext_info_addr_ = nullptr; diff --git a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc index b6c48157..17d5cf8b 100644 --- a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc +++ b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc @@ -24,6 +24,12 @@ namespace hybrid { namespace { // if dim count is not reach kMaxShapeDims(8), use INT64_MIN to mark dim end. constexpr int64_t kDimEndFlag = INT64_MIN; +const std::map kTopicTypeToRtsFlagMap { + {static_cast(aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_ONLY), 0}, + {static_cast(aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_FIRST), RT_KERNEL_DEVICE_FIRST}, + {static_cast(aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_ONLY), RT_KERNEL_HOST_ONLY}, + {static_cast(aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_FIRST), RT_KERNEL_HOST_FIRST} +}; } Status AicpuExtInfoHandler::Parse(const std::string &ext_info) { @@ -72,6 +78,9 @@ Status AicpuExtInfoHandler::Parse(const std::string &ext_info) { case aicpu::FWKAdapter::FWK_ADPT_EXT_UPDATE_ADDR: GE_CHK_STATUS_RET(ParseExtUpdateAddr(aicpu_ext_info), "[Parse][ExtUpdateAddr] failed."); break; + case aicpu::FWKAdapter::FWK_ADPT_EXT_TOPIC_TYPE: + GE_CHK_STATUS_RET(ParseExtTopicType(aicpu_ext_info), "[Parse][ExtTopicType] failed."); + break; default: GELOGD("Node[%s] ignore infoType=%d, infoLen=%u.", node_name_.c_str(), aicpu_ext_info->infoType, aicpu_ext_info->infoLen); @@ -207,6 +216,44 @@ Status AicpuExtInfoHandler::ParseExtUpdateAddr(AicpuExtInfo *aicpu_ext_info) { return SUCCESS; } +Status AicpuExtInfoHandler::ParseExtTopicType(AicpuExtInfo *aicpu_ext_info) { + if (aicpu_ext_info->infoLen != sizeof(int32_t)) { + REPORT_INNER_ERROR("E19999", + "Node[%s] parse topic_type info failed as infoLen must be %zu but %u.", + node_name_.c_str(), sizeof(int32_t), aicpu_ext_info->infoLen); + GELOGE(ACL_ERROR_GE_PARAM_INVALID, + "[Check][DataLen]Node[%s] parse topic_type info failed as infoLen must be %zu but %u.", + node_name_.c_str(), sizeof(int32_t), aicpu_ext_info->infoLen); + return ACL_ERROR_GE_PARAM_INVALID; + } + GE_CHECK_NOTNULL(aicpu_ext_info->infoMsg); + auto type = *reinterpret_cast(aicpu_ext_info->infoMsg); + + topic_type_flag_ = TopicTypeToRtsFlag(type); + if (topic_type_flag_ == -1) { + REPORT_INNER_ERROR("E19999", "Node[%s] parse ext topic type failed as need %d %d %d %d but %d.", + node_name_.c_str(), + aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_ONLY, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_FIRST, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_ONLY, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_FIRST, + type); + GELOGE(ACL_ERROR_GE_PARAM_INVALID, + "[Check][Type]Node[%s] parse ext shape type failed as need %d %d %d %d but %d.", + node_name_.c_str(), + aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_ONLY, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_DEVICE_FIRST, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_ONLY, + aicpu::FWKAdapter::FWK_ADPT_TOPIC_HOST_FIRST, + type); + return ACL_ERROR_GE_PARAM_INVALID; + } + + GELOGI("Node[%s] parse ext topic type info success infoLen=%u, topic_type=%d, rts_flag=%d.", + node_name_.c_str(), aicpu_ext_info->infoLen, type, topic_type_flag_); + return SUCCESS; +} + Status AicpuExtInfoHandler::UpdateExecuteMode(bool flag) { if (bit_map_ == nullptr) { GELOGD("There is no bit_map in ext_info, no need update."); @@ -341,5 +388,14 @@ void AicpuExtInfoHandler::GetShapeAndType(const AicpuShapeAndType *shape_and_typ data_type = static_cast(shape_and_type->type); shape = GeShape(dims); } + +int32_t AicpuExtInfoHandler::TopicTypeToRtsFlag(int32_t topic_type) { + auto it = kTopicTypeToRtsFlagMap.find(topic_type); + if (it != kTopicTypeToRtsFlagMap.end()) { + return it->second; + } + + return -1; +} } // namespace hybrid } // namespace ge diff --git a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h index 01092204..46fb7c05 100644 --- a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h +++ b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h @@ -62,6 +62,7 @@ class AicpuExtInfoHandler { Status GetOutputShapeAndType(uint32_t output_index, GeShape &shape, DataType &data_type); bool IsNeedRefreshIOAddr(); + int32_t GetTopicTypeFlag() const { return topic_type_flag_; } private: @@ -71,6 +72,7 @@ class AicpuExtInfoHandler { Status ParseExtSessionInfo(AicpuExtInfo *aicpu_ext_info); Status ParseExtBitMap(AicpuExtInfo *aicpu_ext_info); Status ParseExtUpdateAddr(AicpuExtInfo *aicpu_ext_info); + Status ParseExtTopicType(AicpuExtInfo *aicpu_ext_info); static Status UpdateShapeAndType(const GeShape &shape, DataType data_type, @@ -81,6 +83,8 @@ class AicpuExtInfoHandler { DataType &data_type); private: + int32_t TopicTypeToRtsFlag(int32_t topic_type); + const std::string node_name_; const uint32_t input_num_; const uint32_t output_num_; @@ -88,6 +92,7 @@ class AicpuExtInfoHandler { AicpuSessionInfo *session_info_ = nullptr; uint64_t *bit_map_ = nullptr; uint32_t *update_addr_ = nullptr; + int32_t topic_type_flag_ = -1; std::unique_ptr ext_info_; size_t ext_info_len_ = 0; diff --git a/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc b/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc index 63202a28..327dd248 100644 --- a/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc +++ b/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc @@ -155,4 +155,58 @@ TEST_F(UtestKernelExTaskInfo, parse_update_addr) { KernelExTaskInfo kernel_ex_task_info; EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); } + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_success_1) { + const string ext_info = {7,0,0,0,4,0,0,0,0,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_success_2) { + const string ext_info = {7,0,0,0,4,0,0,0,1,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_success_3) { + const string ext_info = {7,0,0,0,4,0,0,0,2,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_success_4) { + const string ext_info = {7,0,0,0,4,0,0,0,3,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_failed_1) { + const string ext_info = {7,0,0,0,4,0,0,0,4,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_NE(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} + +TEST_F(UtestKernelExTaskInfo, parse_topic_type_failed_2) { + const string ext_info = {7,0,0,0,2,0,0,0,2,0,0,0}; + const OpDescPtr op_desc = CreateOpDesc("FrameworkOp", "FrameworkOp"); + AttrUtils::SetBool(op_desc, "_AllShape", true); + + KernelExTaskInfo kernel_ex_task_info; + EXPECT_NE(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS); +} } // namespace ge diff --git a/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h b/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h index 7a2cbc50..df57c82e 100644 --- a/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h +++ b/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h @@ -61,9 +61,18 @@ enum FWKTaskExtInfoType { FWK_ADPT_EXT_OP_NAME, FWK_ADPT_EXT_SESSION_INFO, FWK_ADPT_EXT_BITMAP, + FWK_ADPT_EXT_TOPIC_TYPE, FWK_ADPT_EXT_INVALID }; +enum FWKExtTopicType { + FWK_ADPT_TOPIC_DEVICE_ONLY = 0, + FWK_ADPT_TOPIC_DEVICE_FIRST, + FWK_ADPT_TOPIC_HOST_ONLY, + FWK_ADPT_TOPIC_HOST_FIRST, + FWK_ADPT_TOPIC_INVALID +}; + enum FWKExtUpdateAddrType { FWK_ADPT_UPDATE_NULL = 0, FWK_ADPT_UPDATE_INPUT, diff --git a/third_party/fwkacllib/inc/runtime/kernel.h b/third_party/fwkacllib/inc/runtime/kernel.h index b4500e10..402fadef 100644 --- a/third_party/fwkacllib/inc/runtime/kernel.h +++ b/third_party/fwkacllib/inc/runtime/kernel.h @@ -191,6 +191,11 @@ typedef void (*rtCallback_t)(void *fnData); #define RT_FUSION_KERNEL_DUMPFLAG (0x04) #define RT_KERNEL_CUSTOM_AICPU (0x08) +// STARS topic scheduler sqe : topic_type +#define RT_KERNEL_DEVICE_FIRST (0X10) +#define RT_KERNEL_HOST_ONLY (0X20) +#define RT_KERNEL_HOST_FIRST (0X30) + /** * @ingroup rt_kernel * @brief kernel mode