From: @zhao_zhixuan Reviewed-by: @xchu42,@ji_chen Signed-off-by: @ji_chentags/v1.3.0
@@ -708,6 +708,7 @@ target_compile_definitions(ge_runner PRIVATE | |||
DAVINCI_CLOUD | |||
google=ascend_private | |||
FUNC_VISIBILITY | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_compile_options(ge_runner PRIVATE | |||
@@ -783,6 +784,7 @@ target_compile_definitions(ge_compiler PRIVATE | |||
COMPILE_OMG_PACKAGE | |||
google=ascend_private | |||
FUNC_VISIBILITY | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_compile_options(ge_compiler PRIVATE | |||
@@ -24,6 +24,7 @@ | |||
#include "graph/types.h" | |||
#include "runtime/base.h" | |||
#include "graph/load/model_manager/davinci_model.h" | |||
#include "mmpa/mmpa_api.h" | |||
namespace { | |||
const char *const kTrainingTrace = "training_trace"; | |||
@@ -46,6 +47,10 @@ const std::string kOptype = "op_type"; | |||
const std::string kBlockDim = "block_dims"; | |||
const std::string kTaskId = "task_id"; | |||
const std::string kStreamId = "stream_id"; | |||
const std::string kThreadId = "thread_id"; | |||
const std::string kIndexId = "index_id"; | |||
const std::string kTimeStamp = "time_stamp"; | |||
const std::string kTagId = "tag_id"; | |||
const std::string kShapeType = "shape_type"; | |||
const std::string kCurIterNum = "cur_iter_num"; | |||
const std::string kTaskType = "task_type"; | |||
@@ -286,6 +291,58 @@ FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::Profilin | |||
#endif | |||
} | |||
FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY Status ProfilingManager::ProfileStepInfo( | |||
uint64_t index_id, uint64_t model_id, uint16_t tag_id, rtStream_t stream, int32_t device_id) { | |||
#ifdef DAVINCI_SUPPORT_PROFILING | |||
rtError_t rt_ret = RT_ERROR_NONE; | |||
#ifndef ONLY_COMPILE_OPEN_SRC | |||
GELOGD("Profiling Step Info TraceTask execute async start, index_id = %lu, model_id = %lu, tag_id = %u", | |||
index_id, model_id, tag_id); | |||
rt_ret = rtProfilerTraceEx(index_id, model_id, tag_id, stream); | |||
if (rt_ret != RT_ERROR_NONE) { | |||
GELOGE(RT_FAILED, "[Call][rtProfilerTraceEx] failed, ret: 0x%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret); | |||
} | |||
GELOGD("Profiling Step Info TraceTask execute async success, index_id = %lu, model_id = %lu, tag_id = %u", | |||
index_id, model_id, tag_id); | |||
#endif | |||
mmTimespec timespec = mmGetTickCount(); | |||
// 1000 ^ 3 converts second to nanosecond | |||
int64_t time = timespec.tv_sec * 1000 * 1000 * 1000 + timespec.tv_nsec; | |||
uint32_t task_id = 0; | |||
uint32_t stream_id = 0; | |||
rt_ret = rtGetTaskIdAndStreamID(&task_id, &stream_id); | |||
if (rt_ret != RT_ERROR_NONE) { | |||
GELOGE(RT_FAILED, "[Get][RtsInfo] task_id and stream_id failed, ret: 0x%X.", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret); | |||
} | |||
GELOGD("Get profiling args, task_id[%u], stream_id[%u]", task_id, stream_id); | |||
Json step_info; | |||
step_info[kIndexId] = index_id; | |||
step_info[kModelId] = model_id; | |||
step_info[kTimeStamp] = time; | |||
step_info[kTagId] = tag_id; | |||
step_info[kTaskId] = task_id; | |||
step_info[kStreamId] = stream_id; | |||
step_info[kThreadId] = mmGetTid(); | |||
std::string reported_data; | |||
try { | |||
reported_data = step_info.dump(kInteval, ' ', false, Json::error_handler_t::ignore); | |||
} catch (std::exception &e) { | |||
GELOGE(FAILED, "Failed to convert JSON to string, reason: %s.", e.what()); | |||
} catch (...) { | |||
GELOGE(FAILED, "Failed to convert JSON to string."); | |||
} | |||
reported_data.append(",") | |||
.append("\n"); | |||
ReportData(device_id, reported_data, "step_info"); | |||
#endif | |||
return SUCCESS; | |||
} | |||
FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY void ProfilingManager::ReportData( | |||
const int32_t &device_id, const string &data, const string &tag_name) { | |||
#ifdef DAVINCI_SUPPORT_PROFILING | |||
@@ -27,6 +27,7 @@ | |||
#include "framework/common/ge_types.h" | |||
#include "external/register/register_types.h" | |||
#include "toolchain/prof_callback.h" | |||
#include "runtime/stream.h" | |||
using std::map; | |||
using std::string; | |||
@@ -97,6 +98,7 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ProfilingManager { | |||
void GetFpBpPoint(std::string &fp_point, std::string &bp_point); | |||
void GetOpInputOutputInfo(const OpDescPtr &op, TaskDescInfo &task_desc_info) const; | |||
void ReportData(const int32_t &device_id, const std::string &data, const std::string &tag_name); | |||
Status ProfileStepInfo(uint64_t index_id, uint64_t model_id, uint16_t tag_id, rtStream_t stream, int32_t device_id); | |||
private: | |||
Status InitFromOptions(const Options &options, MsprofGeOptions &prof_conf); | |||
Status ParseOptions(const std::string &options); | |||
@@ -179,6 +179,7 @@ target_compile_definitions(ge_executor PRIVATE | |||
google=ascend_private | |||
$<IF:$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>,OS_TYPE=WIN,OS_TYPE=0> | |||
$<$<STREQUAL:${TARGET_SYSTEM_NAME},Windows>:SECUREC_USING_STD_SECURE_LIB=0 NOMINMAX> | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
LOG_CPP | |||
) | |||
@@ -225,6 +226,7 @@ target_compile_definitions(ge_executor_shared PRIVATE | |||
DAVINCI_SUPPORT_PROFILING | |||
google=ascend_private | |||
FUNC_VISIBILITY | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_include_directories(ge_executor_shared PRIVATE | |||
@@ -3873,12 +3873,27 @@ Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputDa | |||
GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_PRE_PROC_END)); | |||
if (!task_list_.empty()) { | |||
uint64_t index_id = iterator_count_ + 1; | |||
uint64_t model_id = static_cast<uint64_t>(model_id_); | |||
int32_t device_id = static_cast<int32_t>(device_id_); | |||
// tag_id 0 means step begin, 1 meas step end. | |||
if (profiling_model_execute_on) { | |||
GE_CHK_STATUS_RET_NOLOG( | |||
ProfilingManager::Instance().ProfileStepInfo(index_id, model_id, 0, rt_model_stream_, device_id)); | |||
} | |||
GELOGD("rtModelExecute do"); | |||
GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_INFER_START)); | |||
rtError_t rt_ret = rtModelExecute(rt_model_handle_, rt_model_stream_, 0); | |||
GE_CHK_RT_EXEC(rt_ret, return RT_ERROR_TO_GE_STATUS(rt_ret)); | |||
GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_INFER_END)); | |||
GELOGD("rtModelExecute end"); | |||
if (profiling_model_execute_on) { | |||
GE_CHK_STATUS_RET_NOLOG( | |||
ProfilingManager::Instance().ProfileStepInfo(index_id, model_id, 1, rt_model_stream_, device_id)); | |||
} | |||
iterator_count_++; | |||
} | |||
if (!is_async_mode_) { | |||
@@ -18,6 +18,7 @@ | |||
#include "graph/ge_context.h" | |||
#include "graph/runtime_inference_context.h" | |||
#include "common/dump/dump_manager.h" | |||
#include "common/profiling/profiling_manager.h" | |||
namespace ge { | |||
namespace hybrid { | |||
@@ -77,10 +78,23 @@ Status HybridModelExecutor::ExecuteGraphInternal(SubgraphExecutor &executor, | |||
GE_CHK_STATUS_RET_NOLOG(ResetExecutionContext(context_)); | |||
RECORD_MODEL_EXECUTION_EVENT(&context_, "[InitContext] End"); | |||
uint64_t index_id = context_.iteration + 1; | |||
uint64_t model_id = static_cast<uint64_t>(model_->GetModelId()); | |||
int32_t device_id = static_cast<int32_t>(device_id_); | |||
auto &prof_mgr = ProfilingManager::Instance(); | |||
// tag_id 0 means step begin, 1 meas step end. | |||
if (prof_mgr.ProfilingModelExecuteOn()) { | |||
GE_CHK_STATUS_RET_NOLOG(prof_mgr.ProfileStepInfo(index_id, model_id, 0, stream_, device_id)); | |||
} | |||
HYBRID_CHK_STATUS_RET(executor.ExecuteAsync(args.inputs, args.input_desc, args.outputs), | |||
"Failed to execute partitioned call."); | |||
RECORD_MODEL_EXECUTION_EVENT(&context_, "[ExecuteAsync] End"); | |||
if (prof_mgr.ProfilingModelExecuteOn()) { | |||
GE_CHK_STATUS_RET_NOLOG(prof_mgr.ProfileStepInfo(index_id, model_id, 1, stream_, device_id)); | |||
} | |||
HYBRID_CHK_STATUS_RET(executor.Synchronize(), "Failed to sync root graph."); | |||
RECORD_MODEL_EXECUTION_EVENT(&context_, "[Synchronize] End"); | |||
@@ -313,6 +313,8 @@ rtError_t rtFlushCache(uint64_t base, uint32_t len) { return RT_ERROR_NONE; } | |||
rtError_t rtProfilerTrace(uint64_t id, bool notify, uint32_t flags, rtStream_t stream_) { return RT_ERROR_NONE; } | |||
rtError_t rtProfilerTraceEx(uint64_t id, uint64_t modelId, uint16_t tagId, rtStream_t stream) { return RT_ERROR_NONE; } | |||
rtError_t rtMemSetRC(const void *dev_ptr, uint64_t size, uint32_t read_count) { return RT_ERROR_NONE; } | |||
rtError_t rtStreamSwitch(void *ptr, rtCondition_t condition, int64_t value, rtStream_t true_stream, rtStream_t stream) { | |||
@@ -840,6 +840,7 @@ add_library(ge_ut_common STATIC ${COMMON_SRC_FILES} ${PROTO_HDRS}) | |||
target_compile_definitions(ge_ut_common PRIVATE | |||
google=ascend_private | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_compile_options(ge_ut_common PRIVATE | |||
@@ -853,6 +854,7 @@ target_link_libraries(ge_ut_common PRIVATE | |||
ascend_protobuf | |||
json | |||
ge_ut_graph | |||
runtime_stub | |||
) | |||
# build common format | |||
@@ -860,6 +862,7 @@ add_library(ge_ut_common_format STATIC ${COMMON_SRC_FILES} ${COMMON_FORMAT_SRC_F | |||
target_compile_definitions(ge_ut_common_format PRIVATE | |||
google=ascend_private | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_compile_options(ge_ut_common_format PRIVATE | |||
@@ -1012,6 +1015,7 @@ add_library(ge_single_op STATIC ${SINGLE_OP_SRC_FILES} ${PROTO_HDRS}) | |||
target_compile_definitions(ge_single_op PRIVATE | |||
google=ascend_private | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_compile_options(ge_single_op PRIVATE | |||
@@ -1108,6 +1112,7 @@ target_compile_options(ut_libge_distinct_load_utest PRIVATE | |||
target_compile_definitions(ut_libge_distinct_load_utest PRIVATE | |||
google=ascend_private | |||
$<$<STREQUAL:${ENABLE_OPEN_SRC},True>:ONLY_COMPILE_OPEN_SRC> | |||
) | |||
target_link_libraries(ut_libge_distinct_load_utest | |||
@@ -1022,6 +1022,10 @@ TEST_F(UtestDavinciModel, NnExecute) { | |||
EXPECT_EQ(outputs.size(), 1); | |||
input_data.blobs = output_data.blobs; | |||
EXPECT_EQ(input_data.blobs.size(), 1); | |||
ProfilingManager::Instance().prof_cb_.msprofReporterCallback = MsprofReport; | |||
ProfilingManager::Instance().device_id_.emplace_back(0); | |||
model.task_list_.resize(1); | |||
EXPECT_EQ(model.NnExecute(stream, false, input_data, output_data), SUCCESS); | |||
} | |||
} // namespace ge |