@@ -88,4 +88,8 @@ void OpTilingManager::LoadSo() { | |||
} | |||
} | |||
OpTilingManager &OpTilingManager::GetInstance() { | |||
static OpTilingManager instance; | |||
return instance; | |||
} | |||
} // namespace ge |
@@ -25,6 +25,7 @@ using SoToHandleMap = std::map<std::string, void *>; | |||
class OpTilingManager { | |||
public: | |||
OpTilingManager() = default; | |||
static OpTilingManager &GetInstance(); | |||
~OpTilingManager(); | |||
void LoadSo(); | |||
@@ -72,7 +72,89 @@ set(SRC_LIST | |||
"../single_op/task/tbe_task_builder.cc" | |||
"../single_op/task/aicpu_task_builder.cc" | |||
"../single_op/task/aicpu_kernel_task_builder.cc" | |||
"../hybrid/hybrid_davinci_model_stub.cc" | |||
"../hybrid/common/tensor_value.cc" | |||
"../hybrid/common/npu_memory_allocator.cc" | |||
"../hybrid/executor/rt_callback_manager.cc" | |||
"../hybrid/executor/node_state.cc" | |||
"../hybrid/executor/node_done_manager.cc" | |||
"../hybrid/executor/hybrid_profiler.cc" | |||
"../hybrid/executor/hybrid_model_executor.cc" | |||
"../hybrid/executor/hybrid_model_async_executor.cc" | |||
"../hybrid/executor/hybrid_execution_context.cc" | |||
"../hybrid/executor/subgraph_context.cc" | |||
"../hybrid/executor/subgraph_executor.cc" | |||
"../hybrid/executor/worker/task_compile_engine.cc" | |||
"../hybrid/executor/worker/shape_inference_engine.cc" | |||
"../hybrid/executor/worker/execution_engine.cc" | |||
"../hybrid/model/hybrid_model.cc" | |||
"../hybrid/model/hybrid_model_builder.cc" | |||
"../hybrid/model/node_item.cc" | |||
"../hybrid/model/graph_item.cc" | |||
"../hybrid/node_executor/aicore/aicore_node_executor.cc" | |||
"../hybrid/node_executor/aicore/aicore_op_task.cc" | |||
"../hybrid/node_executor/aicore/aicore_task_builder.cc" | |||
"../hybrid/node_executor/aicpu/aicpu_node_executor.cc" | |||
"../hybrid/node_executor/compiledsubgraph/known_node_executor.cc" | |||
"../hybrid/node_executor/ge_local/ge_local_node_executor.cc" | |||
"../hybrid/node_executor/host_cpu/host_cpu_node_executor.cc" | |||
"../hybrid/node_executor/host_cpu/kernel_factory.cc" | |||
"../hybrid/node_executor/host_cpu/kernel/no_op_kernel.cc" | |||
"../hybrid/node_executor/host_cpu/kernel/variable_kernel.cc" | |||
"../hybrid/node_executor/host_cpu/kernel/assign_kernel.cc" | |||
"../hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc" | |||
"../hybrid/node_executor/controlop/control_op_executor.cc" | |||
"../hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc" | |||
"../hybrid/node_executor/rts/rts_node_executor.cc" | |||
"../hybrid/node_executor/node_executor.cc" | |||
"../hybrid/node_executor/task_context.cc" | |||
"../hybrid/hybrid_davinci_model.cc" | |||
"../ge_local_engine/engine/host_cpu_engine.cc" | |||
"../graph/common/omg_util.cc" | |||
"../graph/manager/host_mem_manager.cc" | |||
"../graph/build/memory/var_mem_assign_util.cc" | |||
"../host_kernels/transpose_kernel.cc" | |||
"../host_kernels/add_kernel.cc" | |||
"../host_kernels/broadcast_args_kernel.cc" | |||
"../host_kernels/broadcast_gradient_args_kernel.cc" | |||
"../host_kernels/cast_kernel.cc" | |||
"../host_kernels/concat_offset_kernel.cc" | |||
"../host_kernels/concat_v2_kernel.cc" | |||
"../host_kernels/dynamic_stitch_kernel.cc" | |||
"../host_kernels/identity_kernel.cc" | |||
"../host_kernels/empty_kernel.cc" | |||
"../host_kernels/expanddims_kernel.cc" | |||
"../host_kernels/fill_kernel.cc" | |||
"../host_kernels/floordiv_kernel.cc" | |||
"../host_kernels/floormod_kernel.cc" | |||
"../host_kernels/gather_v2_kernel.cc" | |||
"../host_kernels/greater_kernel.cc" | |||
"../host_kernels/kernel_utils.cc" | |||
"../host_kernels/maximum_kernel.cc" | |||
"../host_kernels/mul_kernel.cc" | |||
"../host_kernels/pack_kernel.cc" | |||
"../host_kernels/permute_kernel.cc" | |||
"../host_kernels/range_kernel.cc" | |||
"../host_kernels/rank_kernel.cc" | |||
"../host_kernels/reduce_prod_kernel.cc" | |||
"../host_kernels/reshape_kernel.cc" | |||
"../host_kernels/rsqrt_kernel.cc" | |||
"../host_kernels/shape_kernel.cc" | |||
"../host_kernels/shape_n_kernel.cc" | |||
"../host_kernels/size_kernel.cc" | |||
"../host_kernels/slice_d_kernel.cc" | |||
"../host_kernels/slice_kernel.cc" | |||
"../host_kernels/squeeze_kernel.cc" | |||
"../host_kernels/unsqueeze_kernel.cc" | |||
"../host_kernels/ssd_prior_box_kernel.cc" | |||
"../host_kernels/strided_slice_kernel.cc" | |||
"../host_kernels/sub_kernel.cc" | |||
"../host_kernels/transdata_kernel.cc" | |||
"../host_kernels/unpack_kernel.cc" | |||
"../graph/passes/pass_utils.cc" | |||
"../graph/common/bcast.cc" | |||
"../common/fp16_t.cc" | |||
"../common/formats/format_transfers/format_transfer_transpose.cc" | |||
"../common/formats/utils/formats_trans_utils.cc" | |||
) | |||
######## libge_executor.a ######## | |||
@@ -105,9 +187,9 @@ target_include_directories(ge_executor PRIVATE | |||
${CMAKE_BINARY_DIR}/proto/ge | |||
#### yellow zone #### | |||
${GE_CODE_DIR}/../inc | |||
${GE_CODE_DIR}/../inc/cce | |||
${GE_CODE_DIR}/../inc/cce | |||
#### blue zone #### | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
) | |||
target_link_libraries(ge_executor PRIVATE | |||
@@ -147,9 +229,9 @@ target_include_directories(ge_executor_shared PRIVATE | |||
${CMAKE_BINARY_DIR}/proto/ge | |||
#### yellow zone #### | |||
${GE_CODE_DIR}/../inc | |||
${GE_CODE_DIR}/../inc/cce | |||
${GE_CODE_DIR}/../inc/cce | |||
#### blue zone #### | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
) | |||
target_link_libraries(ge_executor_shared PRIVATE | |||
@@ -158,7 +240,7 @@ target_link_libraries(ge_executor_shared PRIVATE | |||
-Wl,--no-as-needed | |||
ge_common | |||
runtime | |||
slog | |||
slog | |||
mmpa | |||
graph | |||
register | |||
@@ -39,6 +39,8 @@ | |||
#include "graph/manager/graph_var_manager.h" | |||
#include "graph/load/new_model_manager/davinci_model.h" | |||
#include "opskernel_manager/ops_kernel_builder_manager.h" | |||
#include "graph/opsproto_manager.h" | |||
#include "ge_local_engine/engine/host_cpu_engine.h" | |||
using std::string; | |||
using std::vector; | |||
@@ -221,6 +223,33 @@ class ModelListenerAdapter : public ModelListener { | |||
std::shared_ptr<ge::ModelListener> listener; | |||
}; | |||
static void InitOpsProtoManger() { | |||
string opsproto_path; | |||
const char *path_env = std::getenv("ASCEND_OPP_PATH"); | |||
if (path_env != nullptr) { | |||
string path = path_env; | |||
string file_path = RealPath(path.c_str()); | |||
if (file_path.empty()) { | |||
GELOGE(FAILED, "File path %s is invalid.", path.c_str()); | |||
return; | |||
} | |||
opsproto_path = (path + "/op_proto/custom/" + ":") + (path + "/op_proto/built-in/"); | |||
GELOGI("Get opsproto so path from env : %s", path.c_str()); | |||
} else { | |||
string path_base = PluginManager::GetPath(); | |||
GELOGI("path_base is %s", path_base.c_str()); | |||
path_base = path_base.substr(0, path_base.rfind('/')); | |||
path_base = path_base.substr(0, path_base.rfind('/') + 1); | |||
opsproto_path = (path_base + "ops/op_proto/custom/" + ":") + (path_base + "ops/op_proto/built-in/"); | |||
} | |||
GELOGI("Get opsproto path is %s", opsproto_path.c_str()); | |||
OpsProtoManager *manager = OpsProtoManager::Instance(); | |||
map<string, string> option_tmp; | |||
option_tmp.emplace(std::pair<string, string>(string("ge.opsProtoLibPath"), opsproto_path)); | |||
(void)manager->Initialize(option_tmp); | |||
} | |||
GeExecutor::GeExecutor() {} | |||
Status GeExecutor::Initialize() { | |||
@@ -230,6 +259,16 @@ Status GeExecutor::Initialize() { | |||
return ge::SUCCESS; | |||
} | |||
OpTilingManager::GetInstance().LoadSo(); | |||
Status initHostCpuEngineStatus = HostCpuEngine::GetInstance().Initialize(); | |||
if (initHostCpuEngineStatus != SUCCESS) { | |||
GELOGE(initHostCpuEngineStatus, "Failed to initialize HostCpuEngine"); | |||
return initHostCpuEngineStatus; | |||
} | |||
InitOpsProtoManger(); | |||
std::vector<rtMemType_t> mem_type(1, RT_MEMORY_HBM); | |||
mem_type.push_back(RT_MEMORY_P2P_DDR); | |||
auto ret = MemManager::Instance().Initialize(mem_type); | |||
@@ -600,10 +639,16 @@ Status GeExecutor::UnloadModel(uint32_t model_id) { | |||
return ACL_ERROR_GE_INTERNAL_ERROR; | |||
} | |||
std::shared_ptr<DavinciModel> davinci_model = ModelManager::GetInstance()->GetModel(model_id); | |||
if (davinci_model != nullptr) { | |||
uint64_t session_id = davinci_model->GetSessionId(); | |||
std::shared_ptr<hybrid::HybridDavinciModel> hybrid_davinci_model = ModelManager::GetInstance()->GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
uint64_t session_id = hybrid_davinci_model->GetSessionId(); | |||
VarManagerPool::Instance().RemoveVarManager(session_id); | |||
} else { | |||
std::shared_ptr<DavinciModel> davinci_model = ModelManager::GetInstance()->GetModel(model_id); | |||
if (davinci_model != nullptr) { | |||
uint64_t session_id = davinci_model->GetSessionId(); | |||
VarManagerPool::Instance().RemoveVarManager(session_id); | |||
} | |||
} | |||
ret = GraphLoader::UnloadModel(model_id); | |||
if (ret != SUCCESS) { | |||
@@ -933,6 +978,26 @@ Status GeExecutor::LoadModelWithQ(uint32_t &model_id, const ModelData &model_dat | |||
*/ | |||
Status GeExecutor::ExecModel(uint32_t model_id, void *stream, const ge::RunModelData &run_input_data, | |||
ge::RunModelData &run_output_data, bool async_mode) { | |||
std::vector<GeTensorDesc> input_desc = {}; | |||
std::vector<GeTensorDesc> output_desc = {}; | |||
return ExecModel(model_id, stream, run_input_data, input_desc, run_output_data, output_desc, async_mode); | |||
} | |||
/** | |||
* @ingroup ge | |||
* @brief Synchronous execution of offline model(Do not create thread) | |||
* @param [in] uint32_t model_id: Model ID to execute | |||
void* stream: stream to execute | |||
const domi::InputData *input_data: Model input data | |||
const std::vector<GeTensorDesc> &input_desc: Description of model input data | |||
bool async_mode: is asynchronize mode | |||
* @param [out] domi::OutputData *output_data: Model output data | |||
* @param [out] std::vector<GeTensorDesc> &output_desc: Description of model output data | |||
* @return SUCCESS handle successfully / others handle failed | |||
*/ | |||
Status GeExecutor::ExecModel(uint32_t model_id, void *stream, const ge::RunModelData &run_input_data, | |||
const std::vector<GeTensorDesc> &input_desc, ge::RunModelData &run_output_data, | |||
std::vector<GeTensorDesc> &output_desc, bool async_mode) { | |||
if (!isInit_) { | |||
GELOGE(ACL_ERROR_GE_EXEC_NOT_INIT, "GeExecutor has not been initialized!"); | |||
return ACL_ERROR_GE_EXEC_NOT_INIT; | |||
@@ -957,7 +1022,7 @@ Status GeExecutor::ExecModel(uint32_t model_id, void *stream, const ge::RunModel | |||
} | |||
} | |||
return GraphLoader::ExecuteModel(model_id, stream, async_mode, input_data, output_data); | |||
return GraphLoader::ExecuteModel(model_id, stream, async_mode, input_data, input_desc, output_data, output_desc); | |||
} | |||
/** | |||
@@ -61,9 +61,91 @@ local_ge_executor_src_files := \ | |||
../single_op/task/tbe_task_builder.cc \ | |||
../single_op/task/aicpu_task_builder.cc \ | |||
../single_op/task/aicpu_kernel_task_builder.cc \ | |||
../hybrid/hybrid_davinci_model_stub.cc\ | |||
../hybrid/node_executor/aicpu/aicpu_ext_info.cc \ | |||
../graph/common/local_context.cc \ | |||
../hybrid/common/tensor_value.cc \ | |||
../hybrid/common/npu_memory_allocator.cc \ | |||
../hybrid/executor/rt_callback_manager.cc \ | |||
../hybrid/executor/node_state.cc \ | |||
../hybrid/executor/node_done_manager.cc \ | |||
../hybrid/executor/hybrid_profiler.cc \ | |||
../hybrid/executor/hybrid_model_executor.cc \ | |||
../hybrid/executor/hybrid_model_async_executor.cc \ | |||
../hybrid/executor/hybrid_execution_context.cc \ | |||
../hybrid/executor/subgraph_context.cc \ | |||
../hybrid/executor/subgraph_executor.cc \ | |||
../hybrid/executor/worker/task_compile_engine.cc \ | |||
../hybrid/executor/worker/shape_inference_engine.cc \ | |||
../hybrid/executor/worker/execution_engine.cc \ | |||
../hybrid/model/hybrid_model.cc \ | |||
../hybrid/model/hybrid_model_builder.cc \ | |||
../hybrid/model/node_item.cc \ | |||
../hybrid/model/graph_item.cc \ | |||
../hybrid/node_executor/aicore/aicore_node_executor.cc \ | |||
../hybrid/node_executor/aicore/aicore_op_task.cc \ | |||
../hybrid/node_executor/aicore/aicore_task_builder.cc \ | |||
../hybrid/node_executor/aicpu/aicpu_node_executor.cc \ | |||
../hybrid/node_executor/compiledsubgraph/known_node_executor.cc \ | |||
../hybrid/node_executor/ge_local/ge_local_node_executor.cc \ | |||
../hybrid/node_executor/host_cpu/host_cpu_node_executor.cc \ | |||
../hybrid/node_executor/host_cpu/kernel_factory.cc \ | |||
../hybrid/node_executor/host_cpu/kernel/no_op_kernel.cc \ | |||
../hybrid/node_executor/host_cpu/kernel/variable_kernel.cc \ | |||
../hybrid/node_executor/host_cpu/kernel/assign_kernel.cc \ | |||
../hybrid/node_executor/host_cpu/kernel/random_uniform_kernel.cc \ | |||
../hybrid/node_executor/controlop/control_op_executor.cc \ | |||
../hybrid/node_executor/partitioned_call/partitioned_call_node_executor.cc \ | |||
../hybrid/node_executor/rts/rts_node_executor.cc \ | |||
../hybrid/node_executor/node_executor.cc \ | |||
../hybrid/node_executor/task_context.cc \ | |||
../hybrid/hybrid_davinci_model.cc \ | |||
../ge_local_engine/engine/host_cpu_engine.cc \ | |||
../graph/common/omg_util.cc \ | |||
../graph/manager/host_mem_manager.cc \ | |||
../graph/build/memory/var_mem_assign_util.cc \ | |||
../host_kernels/transpose_kernel.cc \ | |||
../host_kernels/add_kernel.cc \ | |||
../host_kernels/broadcast_args_kernel.cc \ | |||
../host_kernels/broadcast_gradient_args_kernel.cc \ | |||
../host_kernels/cast_kernel.cc \ | |||
../host_kernels/concat_offset_kernel.cc \ | |||
../host_kernels/concat_v2_kernel.cc \ | |||
../host_kernels/dynamic_stitch_kernel.cc \ | |||
../host_kernels/identity_kernel.cc \ | |||
../host_kernels/empty_kernel.cc \ | |||
../host_kernels/expanddims_kernel.cc \ | |||
../host_kernels/fill_kernel.cc \ | |||
../host_kernels/floordiv_kernel.cc \ | |||
../host_kernels/floormod_kernel.cc \ | |||
../host_kernels/gather_v2_kernel.cc \ | |||
../host_kernels/greater_kernel.cc \ | |||
../host_kernels/kernel_utils.cc \ | |||
../host_kernels/maximum_kernel.cc \ | |||
../host_kernels/mul_kernel.cc \ | |||
../host_kernels/pack_kernel.cc \ | |||
../host_kernels/permute_kernel.cc \ | |||
../host_kernels/range_kernel.cc \ | |||
../host_kernels/rank_kernel.cc \ | |||
../host_kernels/reduce_prod_kernel.cc \ | |||
../host_kernels/reshape_kernel.cc \ | |||
../host_kernels/rsqrt_kernel.cc \ | |||
../host_kernels/shape_kernel.cc \ | |||
../host_kernels/shape_n_kernel.cc \ | |||
../host_kernels/size_kernel.cc \ | |||
../host_kernels/slice_d_kernel.cc \ | |||
../host_kernels/slice_kernel.cc \ | |||
../host_kernels/squeeze_kernel.cc \ | |||
../host_kernels/unsqueeze_kernel.cc \ | |||
../host_kernels/ssd_prior_box_kernel.cc \ | |||
../host_kernels/strided_slice_kernel.cc \ | |||
../host_kernels/sub_kernel.cc \ | |||
../host_kernels/transdata_kernel.cc \ | |||
../host_kernels/unpack_kernel.cc \ | |||
../graph/passes/pass_utils.cc \ | |||
../graph/common/bcast.cc \ | |||
../common/fp16_t.cc \ | |||
../common/formats/format_transfers/format_transfer_transpose.cc \ | |||
../common/formats/utils/formats_trans_utils.cc \ | |||
local_ge_executor_c_include := \ | |||
proto/insert_op.proto \ | |||
@@ -195,7 +195,7 @@ set_target_properties(atc_ge_local_opskernel_builder PROPERTIES | |||
) | |||
############ libge_local_opskernel_builder.a ############ | |||
add_library(ge_local_opskernel_builder_static SHARED ${OPS_KERNEL_SRC_LIST} ${PROTO_HDRS}) | |||
add_library(ge_local_opskernel_builder_static STATIC ${OPS_KERNEL_SRC_LIST} ${PROTO_HDRS}) | |||
target_compile_options(ge_local_opskernel_builder_static PRIVATE | |||
-Werror | |||
@@ -95,8 +95,8 @@ Status GetDataNumber(const GeTensorDesc &out_desc, uint64_t &data_num) { | |||
void HostCpuEngine::CloseSo() { | |||
for (auto handle : lib_handles_) { | |||
if (dlclose(handle) != 0) { | |||
GELOGW("failed to close handle, message: %s", dlerror()); | |||
if (mmDlclose(handle) != 0) { | |||
GELOGW("failed to close handle, message: %s", mmDlerror()); | |||
} | |||
} | |||
lib_handles_.clear(); | |||
@@ -322,13 +322,13 @@ Status HostCpuEngine::LoadLibs(std::vector<std::string> &lib_paths) { | |||
Status HostCpuEngine::LoadLib(const std::string &lib_path) { | |||
GELOGI("To invoke dlopen on lib: %s", lib_path.c_str()); | |||
auto handle = dlopen(lib_path.c_str(), RTLD_NOW | RTLD_GLOBAL); | |||
auto handle = mmDlopen(lib_path.c_str(), MMPA_RTLD_NOW | MMPA_RTLD_GLOBAL); | |||
if (handle == nullptr) { | |||
GELOGE(INTERNAL_ERROR, "Failed to invoke dlopen. path = %s, error = %s", lib_path.c_str(), dlerror()); | |||
GELOGE(INTERNAL_ERROR, "Failed to invoke dlopen. path = %s, error = %s", lib_path.c_str(), mmDlerror()); | |||
return INTERNAL_ERROR; | |||
} | |||
auto initialize = (Status (*)(const HostCpuContext &))dlsym(handle, "Initialize"); | |||
auto initialize = (Status (*)(const HostCpuContext &))mmDlsym(handle, "Initialize"); | |||
if (initialize != nullptr) { | |||
GELOGI("Invoke function Initialize in lib: %s", lib_path.c_str()); | |||
if (initialize(HostCpuContext()) != SUCCESS) { | |||
@@ -20,7 +20,7 @@ | |||
#include "framework/common/ge_inner_error_codes.h" | |||
#include "graph/node.h" | |||
#include "graph/operator.h" | |||
#include "register/register.h" | |||
#include "external/../register/register.h" | |||
namespace ge { | |||
class HostCpuEngine { | |||
@@ -13,6 +13,9 @@ set(GE_SRC_LIST | |||
"task/hccl_task.cc" | |||
"task/memcpy_async_task.cc" | |||
"task/profiler_task.cc" | |||
"task/label_goto_task.cc" | |||
"task/label_set_task.cc" | |||
"task/label_switch_task.cc" | |||
) | |||
add_library(ge_runtime SHARED ${GE_SRC_LIST}) | |||
@@ -307,8 +307,8 @@ bool RuntimeModel::Run() { | |||
ret = rtStreamSynchronize(rt_model_stream_); | |||
if (ret != RT_ERROR_NONE) { | |||
if (ret == RT_ERROR_END_OF_SEQUENCE) { | |||
GELOGI("Model stream RT_ERROR_END_OF_SEQUENCE signal received, ret = 0x%X", ret); | |||
if (ret == ACL_ERROR_RT_END_OF_SEQUENCE) { | |||
GELOGI("Model stream ACL_ERROR_RT_END_OF_SEQUENCE signal received, ret = 0x%X", ret); | |||
return true; | |||
} | |||
GELOGE(RT_FAILED, "Model stream sync failed, ret = 0x%X", ret); | |||
@@ -24,6 +24,7 @@ | |||
#include "runtime/rt_model.h" | |||
#include "ge_runtime/model_context.h" | |||
#include "ge_runtime/task_info.h" | |||
#include "external/runtime/rt_error_codes.h" | |||
namespace ge { | |||
namespace model_runner { | |||
@@ -30,6 +30,7 @@ | |||
#include "model/ge_model.h" | |||
#include "graph/ge_context.h" | |||
#include "opskernel_manager/ops_kernel_builder_manager.h" | |||
#include "graph/utils/op_desc_utils.h" | |||
using domi::BuildMode; | |||
@@ -311,6 +312,53 @@ Status GraphBuilder::BuildForHostCpuGraph(ComputeGraphPtr &comp_graph, GeModelPt | |||
return BuildForUnknownShapeGraph(comp_graph, ge_model_ptr, session_id); | |||
} | |||
static Status InsertMemcpyNode(const ComputeGraphPtr &graph, const OutDataAnchorPtr &out_anchor, | |||
const std::vector<InDataAnchorPtr> &in_anchors, const std::string &name) { | |||
GE_CHECK_NOTNULL(out_anchor); | |||
NodePtr in_node = out_anchor->GetOwnerNode(); | |||
GE_CHECK_NOTNULL(in_node); | |||
OpDescBuilder op_desc_builder(name, MEMCPYADDRASYNC); | |||
OpDescPtr op_desc = op_desc_builder.AddInput("x", in_node->GetOpDesc()->GetOutputDesc(0)) | |||
.AddOutput("y", in_node->GetOpDesc()->GetOutputDesc(0)) | |||
.Build(); | |||
(void)AttrUtils::SetBool(op_desc, ATTR_NO_NEED_CONSTANT_FOLDING, false); | |||
if (GraphUtils::InsertNodeAfter(out_anchor, in_anchors, graph->AddNode(op_desc)) != GRAPH_SUCCESS) { | |||
GELOGE(FAILED, "Insert IDENTITY node %s after %s failed.", name.c_str(), in_node->GetName().c_str()); | |||
return FAILED; | |||
} | |||
return SUCCESS; | |||
} | |||
static Status GenerateTaskForConstant(const std::shared_ptr<ComputeGraph> &graph) { | |||
for (auto &node : graph->GetDirectNode()) { | |||
// CONSTANT not generate task, so insert IDENTITY between CONSTANT and NETOUTPUT | |||
auto op_desc = node->GetOpDesc(); | |||
if (op_desc == nullptr) { | |||
continue; | |||
} | |||
auto op_type = op_desc->GetType(); | |||
if (op_type == NETOUTPUT) { | |||
for (InDataAnchorPtr &in_data_anchor : node->GetAllInDataAnchors()) { | |||
const OutDataAnchorPtr &peer_out_anchor = in_data_anchor->GetPeerOutAnchor(); | |||
GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue); | |||
NodePtr in_node = peer_out_anchor->GetOwnerNode(); | |||
GE_CHECK_NOTNULL(in_node); | |||
std::string in_node_op_type = in_node->GetType(); | |||
if (in_node_op_type == CONSTANT) { | |||
GELOGD("Insert MemcpyAsync node between %s and %s.", in_node->GetName().c_str(), node->GetName().c_str()); | |||
std::string name = node->GetName() + "_input_" + std::to_string(in_data_anchor->GetIdx()) + "_Memcpy"; | |||
if (InsertMemcpyNode(graph, peer_out_anchor, {in_data_anchor}, name) != SUCCESS) { | |||
GELOGE(FAILED, "Insert memcpy between %s and %s failed.", in_node->GetName().c_str(), node->GetName().c_str()); | |||
return FAILED; | |||
} | |||
} | |||
} | |||
} | |||
} | |||
return SUCCESS; | |||
} | |||
Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph, | |||
std::vector<SubGraphInfoPtr> &subgraph_ptr_list, | |||
GeRootModelPtr &ge_root_model_ptr, GeModelPtr &ge_model_ptr, | |||
@@ -332,6 +380,9 @@ Status GraphBuilder::BuildForDynamicShapeGraph(ComputeGraphPtr &comp_graph, | |||
if (sub_graph->GetParentGraph() != comp_graph && !sub_graph->GetParentGraph()->GetGraphUnknownFlag()) { | |||
continue; | |||
} | |||
GE_CHK_STATUS_RET(GenerateTaskForConstant(sub_graph), "Generate task For constant node in subgraph failed."); | |||
if (sub_graph->GetGraphUnknownFlag()) { | |||
// unknown shape build flow | |||
GE_CHK_STATUS_RET(BuildForUnknownShapeGraph(sub_graph, ge_model_ptr, session_id), | |||
@@ -274,13 +274,16 @@ Status GraphLoader::LoadModelWithQ(uint32_t &model_id, const ModelData &model_da | |||
/// @param [in] stream stream to execute model on | |||
/// @param [in] async_mode is asynchronize mode. | |||
/// @param [in] input_data model input data | |||
/// @param [in] input_desc description of model input data | |||
/// @param [out] output_data model output data | |||
/// @param [out] output_desc description of model output data | |||
/// | |||
Status GraphLoader::ExecuteModel(uint32_t model_id, rtStream_t stream, bool async_mode, const InputData &input_data, | |||
OutputData &output_data) { | |||
const std::vector<GeTensorDesc> &input_desc, OutputData &output_data, | |||
std::vector<GeTensorDesc> &output_desc) { | |||
auto model_manager = ModelManager::GetInstance(); | |||
GE_CHECK_NOTNULL(model_manager); | |||
Status ret = model_manager->ExecuteModel(model_id, stream, async_mode, input_data, output_data); | |||
Status ret = model_manager->ExecuteModel(model_id, stream, async_mode, input_data, input_desc, output_data, output_desc); | |||
if (ret != SUCCESS) { | |||
GELOGE(ret, "Execute model failed, model_id:%u.", model_id); | |||
return ret; | |||
@@ -65,7 +65,8 @@ class GraphLoader { | |||
const std::vector<uint32_t> &output_queue_ids); | |||
static Status ExecuteModel(uint32_t model_id, rtStream_t stream, bool async_mode, const InputData &input_data, | |||
OutputData &output_data); | |||
const std::vector<GeTensorDesc> &input_desc, OutputData &output_data, | |||
std::vector<GeTensorDesc> &output_desc); | |||
static Status DestroyAicpuKernel(uint64_t session_id, uint32_t model_id); | |||
@@ -919,11 +919,11 @@ Status DataDumper::DumpExceptionInfo(const std::vector<rtExceptionInfo> exceptio | |||
ReplaceStringElem(op_name); | |||
ReplaceStringElem(op_type); | |||
string dump_file_path = | |||
"./" + op_type + "." + op_name + "." + to_string(op_desc_info.task_id) + "." + to_string(now_time); | |||
"./" + op_type + "." + op_name + "." + std::to_string(op_desc_info.task_id) + "." + std::to_string(now_time); | |||
GELOGI("The exception dump file path is %s", dump_file_path.c_str()); | |||
uint64_t proto_size = dump_data.ByteSizeLong(); | |||
unique_ptr<char[]> proto_msg(new (std::nothrow) char[proto_size]); | |||
std::unique_ptr<char[]> proto_msg(new (std::nothrow) char[proto_size]); | |||
bool ret = dump_data.SerializeToArray(proto_msg.get(), proto_size); | |||
if (!ret || proto_size == 0) { | |||
GELOGE(PARAM_INVALID, "Dump data proto serialize failed"); | |||
@@ -117,7 +117,8 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener | |||
load_end_time_(0), | |||
time_info_(), | |||
dataInputTid(0), | |||
is_model_has_inited_(false), | |||
is_weight_mem_has_inited_(false), | |||
is_feature_map_mem_has_inited_(false), | |||
model_id_(0), | |||
runtime_model_id_(0), | |||
version_(0), | |||
@@ -263,34 +264,65 @@ void DavinciModel::Shrink() { | |||
ge_model_.reset(); // delete object. | |||
} | |||
Status DavinciModel::InitModelMem(void *dev_ptr, size_t mem_size, void *weight_ptr, size_t weight_size) { | |||
if (is_model_has_inited_) { | |||
GELOGE(FAILED, "call InitModelMem more than once ."); | |||
Status DavinciModel::InitWeightMem(void *dev_ptr, void *weight_ptr, size_t weight_size) { | |||
if (is_weight_mem_has_inited_) { | |||
GELOGE(FAILED, "call InitWeightMem more than once."); | |||
return FAILED; | |||
} | |||
is_model_has_inited_ = true; | |||
is_weight_mem_has_inited_ = true; | |||
std::size_t data_size = TotalMemSize(); | |||
std::size_t p2p_data_size = P2PMemInfos().at(RT_MEMORY_P2P_DDR).memory_size; | |||
const Buffer &weights = ge_model_->GetWeight(); | |||
std::size_t weights_size = weights.GetSize(); | |||
GE_CHECK_LE(weights_size, ALLOC_MEMORY_MAX_SIZE); | |||
if ((dev_ptr != nullptr) && (mem_size < TotalMemSize())) { | |||
GELOGE(FAILED, "Invalid mem param: mem_size=%zu totalsize=%zu.", mem_size, TotalMemSize()); | |||
if ((weight_ptr != nullptr) && (weight_size < weights_size)) { | |||
GELOGE(FAILED, "Invalid mem param: weight_size=%zu totalsize=%zu.", weight_size, weights_size); | |||
return FAILED; | |||
} | |||
if ((weight_ptr != nullptr) && (weight_size < weights_size)) { | |||
GELOGE(FAILED, "Invalid mem param: weight_size=%zu totalsize=%zu.", weight_size, weights_size); | |||
weights_mem_base_ = static_cast<uint8_t *>(dev_ptr); | |||
is_inner_weight_base_ = false; | |||
if (weights_size != 0) { | |||
weights_mem_base_ = static_cast<uint8_t *>(weight_ptr); | |||
is_inner_weight_base_ = false; | |||
if (weight_ptr == nullptr) { | |||
weights_mem_base_ = MallocWeightsMem(weights_size); | |||
if (weights_mem_base_ == nullptr) { | |||
GELOGE(GE_EXEC_ALLOC_WEIGHT_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size); | |||
return GE_EXEC_ALLOC_WEIGHT_MEM_FAILED; | |||
} | |||
is_inner_weight_base_ = true; | |||
} | |||
GELOGI("[IMAS]InitWeightMem graph_%u MallocMemory type[W] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
weights_mem_base_, weights_size); | |||
GE_CHK_RT_RET(rtMemcpy(weights_mem_base_, weights_size, weights.GetData(), weights_size, RT_MEMCPY_HOST_TO_DEVICE)); | |||
GELOGI("copy weights data to device"); | |||
} | |||
runtime_param_.weight_base = weights_mem_base_; | |||
return SUCCESS; | |||
} | |||
Status DavinciModel::InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size) { | |||
if (is_feature_map_mem_has_inited_) { | |||
GELOGE(FAILED, "call InitFeatureMapMem more than once ."); | |||
return FAILED; | |||
} | |||
is_feature_map_mem_has_inited_ = true; | |||
std::size_t data_size = TotalMemSize(); | |||
std::size_t p2p_data_size = P2PMemInfos().at(RT_MEMORY_P2P_DDR).memory_size; | |||
if ((dev_ptr != nullptr) && (mem_size < TotalMemSize())) { | |||
GELOGE(FAILED, "Invalid mem param: mem_size=%zu totalsize=%zu.", mem_size, TotalMemSize()); | |||
return FAILED; | |||
} | |||
mem_base_ = static_cast<uint8_t *>(dev_ptr); | |||
p2p_mem_base_ = static_cast<uint8_t *>(dev_ptr); | |||
weights_mem_base_ = static_cast<uint8_t *>(dev_ptr); | |||
is_inner_mem_base_ = false; | |||
is_inner_weight_base_ = false; | |||
if (TotalMemSize() && mem_base_ == nullptr) { | |||
mem_base_ = MallocFeatureMapMem(data_size); | |||
@@ -298,12 +330,14 @@ Status DavinciModel::InitModelMem(void *dev_ptr, size_t mem_size, void *weight_p | |||
GELOGE(GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED, "Alloc feature map memory failed. size: %zu", data_size); | |||
return GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED; | |||
} | |||
GEEVENT("[IMAS]InitModelMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
GEEVENT("[IMAS]InitFeatureMapAndP2PMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
mem_base_, data_size); | |||
weights_mem_base_ = mem_base_; | |||
if (!is_inner_weight_base_) { | |||
weights_mem_base_ = mem_base_; | |||
is_inner_weight_base_ = true; | |||
} | |||
is_inner_mem_base_ = true; | |||
is_inner_weight_base_ = true; | |||
} | |||
if (p2p_data_size != 0) { | |||
@@ -312,27 +346,11 @@ Status DavinciModel::InitModelMem(void *dev_ptr, size_t mem_size, void *weight_p | |||
GELOGE(GE_EXEC_ALLOC_P2P_MEM_FAILED, "Alloc p2p memory failed,size: %zu", p2p_data_size); | |||
return GE_EXEC_ALLOC_P2P_MEM_FAILED; | |||
} | |||
GELOGI("InitModelMem graph_%u MallocMemory type[P] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
GELOGI("InitFeatureMapAndP2PMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
p2p_mem_base_, p2p_data_size); | |||
is_inner_p2p_mem_base_ = true; | |||
} | |||
if (weights_size != 0) { | |||
weights_mem_base_ = static_cast<uint8_t *>(weight_ptr); | |||
is_inner_weight_base_ = false; | |||
if (weight_ptr == nullptr) { | |||
weights_mem_base_ = MallocWeightsMem(weights_size); | |||
if (weights_mem_base_ == nullptr) { | |||
GELOGE(GE_EXEC_ALLOC_WEIGHT_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size); | |||
return GE_EXEC_ALLOC_WEIGHT_MEM_FAILED; | |||
} | |||
is_inner_weight_base_ = true; | |||
} | |||
GELOGI("[IMAS]InitModelMem graph_%u MallocMemory type[W] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id, | |||
weights_mem_base_, weights_size); | |||
GE_CHK_RT_RET(rtMemcpy(weights_mem_base_, weights_size, weights.GetData(), weights_size, RT_MEMCPY_HOST_TO_DEVICE)); | |||
} | |||
GE_CHK_STATUS_RET(InitVariableMem(), "Init variable memory failed."); | |||
runtime_param_.mem_base = mem_base_; | |||
runtime_param_.weight_base = weights_mem_base_; | |||
@@ -642,8 +660,9 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size | |||
GE_TIMESTAMP_START(InitModelMem); | |||
GELOGD("Known node is %d", known_node_); | |||
GE_CHK_STATUS_RET_NOLOG(InitWeightMem(dev_ptr, weight_ptr, weight_size)); | |||
if (!known_node_) { | |||
GE_CHK_STATUS_RET_NOLOG(InitModelMem(dev_ptr, mem_size, weight_ptr, weight_size)); | |||
GE_CHK_STATUS_RET_NOLOG(InitFeatureMapAndP2PMem(dev_ptr, mem_size)); | |||
data_inputer_ = new (std::nothrow) DataInputer(); | |||
GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, MEMALLOC_FAILED, "data_inputer_ is nullptr."); | |||
} | |||
@@ -1140,6 +1159,7 @@ Status DavinciModel::InitNetOutput(const NodePtr &node) { | |||
GE_IF_BOOL_EXEC(GetGearAndRealOutShapeInfo(input_count, op_desc) != SUCCESS, | |||
GELOGE(PARAM_INVALID, "Failed to get gear and real out shape info."); return PARAM_INVALID;); | |||
} | |||
return SUCCESS; | |||
} | |||
@@ -2780,7 +2800,7 @@ void *DavinciModel::Run(DavinciModel *model) { | |||
reinterpret_cast<int64_t *>(shape_data_buffer_data) + | |||
shape_data_buffer_length / sizeof(int64_t)); | |||
GELOGD("Data: cur dynamic dims is %s", formats::JoinToString(model->cur_dynamic_dims_).c_str()); | |||
delete[] (int64_t *)current_data.blobs.back().data; | |||
delete[] reinterpret_cast<int64_t *>(current_data.blobs.back().data); | |||
current_data.blobs.pop_back(); | |||
} | |||
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), model->SetProfileTime(MODEL_PRE_PROC_END)); | |||
@@ -584,7 +584,8 @@ class DavinciModel { | |||
Status SyncVarData(); | |||
Status InitModelMem(void *dev_ptr, size_t memsize, void *weight_ptr, size_t weightsize); | |||
Status InitWeightMem(void *dev_ptr, void *weight_ptr, size_t weight_size); | |||
Status InitFeatureMapAndP2PMem(void *dev_ptr, size_t mem_size); | |||
void CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input); | |||
@@ -850,7 +851,9 @@ class DavinciModel { | |||
Status GetRealOutputSizeOfMerge(size_t input_index, const NodePtr &merge_node); | |||
Status GetGearAndRealOutShapeInfo(size_t input_count, const OpDescPtr &op_desc); | |||
bool is_model_has_inited_; | |||
bool is_weight_mem_has_inited_; | |||
bool is_feature_map_mem_has_inited_; | |||
uint32_t model_id_; | |||
uint32_t runtime_model_id_; | |||
string name_; | |||
@@ -31,6 +31,7 @@ | |||
#include "model/ge_root_model.h" | |||
#include "graph/common/local_context.h" | |||
#include "common/formats/utils/formats_trans_utils.h" | |||
#include "hybrid/hybrid_davinci_model.h" | |||
namespace ge { | |||
thread_local uint32_t device_count = 0; | |||
@@ -204,6 +205,13 @@ void ModelManager::DestroyAicpuSession(uint64_t session_id) { | |||
ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) { | |||
std::lock_guard<std::mutex> lock(map_mutex_); | |||
auto hybrid_davinci_model = hybrid_model_map_.find(model_id); | |||
if (hybrid_davinci_model != hybrid_model_map_.end()) { | |||
uint64_t session_id = hybrid_davinci_model->second->GetSessionId(); | |||
DestroyAicpuSession(session_id); | |||
return SUCCESS; | |||
} | |||
auto it = model_map_.find(model_id); | |||
if (it == model_map_.end()) { | |||
GELOGE(GE_EXEC_MODEL_ID_INVALID, "model id %u does not exists.", model_id); | |||
@@ -216,7 +224,7 @@ ge::Status ModelManager::DestroyAicpuSessionForInfer(uint32_t model_id) { | |||
ge::Status ModelManager::DestroyAicpuKernel(uint64_t session_id, uint32_t model_id) { | |||
GELOGD("destroy aicpu kernel in session_id %lu, model_id %u.", session_id, model_id); | |||
std::lock_guard<std::mutex> lock(sess_ids_mutex_); | |||
std::lock_guard<std::mutex> lock(map_mutex_); | |||
std::string model_key = std::to_string(session_id) + "_" + std::to_string(model_id); | |||
if (model_aicpu_kernel_.find(model_key) != model_aicpu_kernel_.end()) { | |||
Status ret = KernelLaunchEx(aicpu::FWKAdapter::FWKOperateType::FWK_ADPT_KERNEL_DESTROY, session_id, model_id); | |||
@@ -229,7 +237,7 @@ ge::Status ModelManager::DestroyAicpuKernel(uint64_t session_id, uint32_t model_ | |||
} | |||
ge::Status ModelManager::CreateAicpuKernel(uint64_t session_id, uint32_t model_id, uint64_t kernel_id) { | |||
std::lock_guard<std::mutex> lock(sess_ids_mutex_); | |||
std::lock_guard<std::mutex> lock(map_mutex_); | |||
std::vector<uint64_t> v_aicpu_kernel; | |||
std::string model_key = std::to_string(session_id) + "_" + std::to_string(model_id); | |||
if (model_aicpu_kernel_.find(model_key) != model_aicpu_kernel_.end()) { | |||
@@ -925,6 +933,12 @@ Status ModelManager::GetInputOutputDescInfo(const uint32_t model_id, vector<Inpu | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &inputFormats, std::vector<uint32_t> &outputFormats, | |||
bool new_model_desc) { | |||
std::shared_ptr<hybrid::HybridDavinciModel> hybrid_davinci_model = GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
hybrid_davinci_model->SetModelDescVersion(new_model_desc); | |||
return hybrid_davinci_model->GetInputOutputDescInfo(input_desc, output_desc, inputFormats, outputFormats); | |||
} | |||
std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id); | |||
GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, GE_EXEC_MODEL_ID_INVALID, | |||
"GetInputOutputDescInfo Failed, Invalid model id %u!", model_id); | |||
@@ -943,6 +957,11 @@ Status ModelManager::GetInputOutputDescInfo(const uint32_t model_id, vector<Inpu | |||
/// | |||
Status ModelManager::GetDynamicBatchInfo(const uint32_t model_id, std::vector<std::vector<int64_t>> &batch_info, | |||
int32_t &dynamic_type) { | |||
std::shared_ptr<hybrid::HybridDavinciModel> hybrid_davinci_model = GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
return hybrid_davinci_model->GetDynamicBatchInfo(batch_info, dynamic_type); | |||
} | |||
std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id); | |||
GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID, | |||
"GetDynamicBatchInfo failed, Invalid model id %u!", model_id); | |||
@@ -975,6 +994,12 @@ Status ModelManager::GetCombinedDynamicDims(const uint32_t model_id, vector<vect | |||
/// | |||
Status ModelManager::GetUserDesignateShapeOrder(const uint32_t model_id, | |||
std::vector<std::string> &user_input_shape_order) { | |||
auto hybrid_davinci_model = GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
hybrid_davinci_model->GetUserDesignateShapeOrder(user_input_shape_order); | |||
return SUCCESS; | |||
} | |||
auto davinci_model = GetModel(model_id); | |||
GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, ACL_ERROR_GE_EXEC_MODEL_ID_INVALID, | |||
"GetUserDesignateShapeOrder Failed, Invalid Model ID %u!", model_id) | |||
@@ -990,6 +1015,12 @@ Status ModelManager::GetCurShape(const uint32_t model_id, std::vector<int64_t> & | |||
} | |||
Status ModelManager::GetModelAttr(uint32_t model_id, std::vector<string> &dynamic_output_shape_info) { | |||
std::shared_ptr<hybrid::HybridDavinciModel> hybrid_davinci_model = GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
hybrid_davinci_model->GetModelAttr(dynamic_output_shape_info); | |||
return SUCCESS; | |||
} | |||
std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id); | |||
GE_CHECK_NOTNULL(davinci_model); | |||
davinci_model->GetModelAttr(dynamic_output_shape_info); | |||
@@ -1201,10 +1232,25 @@ Status ModelManager::LoadModelWithQ(uint32_t &model_id, const ModelData &model_d | |||
/// @param [in] stream model stream | |||
/// @param [in] async_mode is asynchronize mode. | |||
/// @param [in] input_data input data | |||
/// @param [in] input_desc description of input data | |||
/// @param [out] output_data output data | |||
/// @param [out] output_desc description of output data | |||
/// | |||
Status ModelManager::ExecuteModel(uint32_t model_id, rtStream_t stream, bool async_mode, const InputData &input_data, | |||
OutputData &output_data) { | |||
const std::vector<GeTensorDesc> &input_desc, OutputData &output_data, | |||
std::vector<GeTensorDesc> &output_desc) { | |||
std::shared_ptr<hybrid::HybridDavinciModel> hybrid_davinci_model = GetHybridModel(model_id); | |||
if (hybrid_davinci_model != nullptr) { | |||
auto inputs = input_data.blobs; | |||
auto outputs = output_data.blobs; | |||
Status status = hybrid_davinci_model->Execute(inputs, input_desc, outputs, output_desc, stream); | |||
if (status == SUCCESS) { | |||
GELOGI("Execute model %u success.", model_id); | |||
} | |||
return status; | |||
} | |||
std::shared_ptr<DavinciModel> davinci_model = GetModel(model_id); | |||
GE_CHK_BOOL_RET_STATUS(davinci_model != nullptr, PARAM_INVALID, "Invalid model id %u.", model_id); | |||
@@ -1243,8 +1289,8 @@ Status ModelManager::CreateAicpuSession(uint64_t session_id) { | |||
return SUCCESS; | |||
} | |||
Status ModelManager::LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_name) { | |||
GELOGI("LoadCustAicpuSo in, op name %s, so name %s", op_desc->GetName().c_str(), so_name.c_str()); | |||
Status ModelManager::LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_name, bool &loaded) { | |||
GELOGD("LoadCustAicpuSo in, op name %s, so name %s", op_desc->GetName().c_str(), so_name.c_str()); | |||
std::lock_guard<std::mutex> lock(cust_aicpu_mutex_); | |||
CustAICPUKernelPtr aicpu_kernel = op_desc->TryGetExtAttr(OP_EXTATTR_CUSTAICPU_KERNEL, CustAICPUKernelPtr()); | |||
if (aicpu_kernel == nullptr) { | |||
@@ -1267,18 +1313,24 @@ Status ModelManager::LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_ | |||
std::map<string, CustAICPUKernelPtr> new_so_name; | |||
new_so_name.insert({so_name, aicpu_kernel}); | |||
cust_aicpu_so_[resource_id] = new_so_name; | |||
GELOGI("LoadCustAicpuSo new aicpu so resource id %lu", resource_id); | |||
loaded = false; | |||
GELOGD("LoadCustAicpuSo new aicpu so name %s, resource id %lu", so_name.c_str(), resource_id); | |||
return SUCCESS; | |||
} | |||
auto it_so_name = it->second.find(so_name); | |||
if (it_so_name == it->second.end()) { | |||
it->second.insert({so_name, aicpu_kernel}); | |||
GELOGI("LoadCustAicpuSo add aicpu so resource id %lu", resource_id); | |||
loaded = false; | |||
GELOGD("LoadCustAicpuSo add aicpu so name %s, resource id %lu", so_name.c_str(), resource_id); | |||
return SUCCESS; | |||
} | |||
loaded = true; | |||
GELOGD("LoadCustAicpuSo so name %s has been loaded.", so_name.c_str()); | |||
return SUCCESS; | |||
} | |||
Status ModelManager::LaunchKernelCustAicpuSo(const string &kernel_name) { | |||
GELOGD("Aicpu kernel launch task in, kernel name %s.", kernel_name.c_str()); | |||
std::lock_guard<std::mutex> lock(cust_aicpu_mutex_); | |||
if (cust_aicpu_so_.size() == 0) return SUCCESS; | |||
// get current context | |||
@@ -148,10 +148,13 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager { | |||
/// @param [in] stream model stream | |||
/// @param [in] async_mode is asynchronize mode. | |||
/// @param [in] input_data model input data | |||
/// @param [in] input_desc description of model input data | |||
/// @param [out] output_data model output data | |||
/// @param [out] output_desc description of model output data | |||
/// | |||
ge::Status ExecuteModel(uint32_t model_id, rtStream_t stream, bool async_mode, const InputData &input_data, | |||
OutputData &output_data); | |||
const std::vector<GeTensorDesc> &input_desc, OutputData &output_data, | |||
std::vector<GeTensorDesc> &output_desc); | |||
ge::Status SyncExecuteModel(uint32_t model_id, const std::vector<GeTensor> &inputs, std::vector<GeTensor> &outputs); | |||
@@ -286,7 +289,7 @@ class FMK_FUNC_HOST_VISIBILITY FMK_FUNC_DEV_VISIBILITY ModelManager { | |||
ge::Status DestroyAicpuSessionForInfer(uint32_t model_id); | |||
ge::Status LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_name); | |||
ge::Status LoadCustAicpuSo(const OpDescPtr &op_desc, const string &so_name, bool &loaded); | |||
ge::Status LaunchCustAicpuSo(); | |||
@@ -875,7 +875,9 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k | |||
} | |||
if (kernel_type_ == ccKernelType::CUST_AI_CPU) { | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc, so_name_), "launch cust aicpu so failed"); | |||
bool loaded = false; | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc, so_name_, loaded), | |||
"launch cust aicpu so failed"); | |||
} | |||
// copy args to new host memory | |||
@@ -41,7 +41,7 @@ class StreamSwitchTaskInfo : public TaskInfo { | |||
Status CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) override; | |||
private: | |||
void SetInputAndValuePtr(DavinciModel *davinci_model, const vector<void *> &input_data_addrs); | |||
void SetInputAndValuePtr(DavinciModel *davinci_model, const std::vector<void *> &input_data_addrs); | |||
void *input_ptr_; | |||
rtCondition_t cond_; | |||
void *value_ptr_; | |||
@@ -49,7 +49,7 @@ class StreamSwitchTaskInfo : public TaskInfo { | |||
uint32_t true_stream_id_; | |||
rtSwitchDataType_t data_type_; | |||
static const uint32_t kInputNum = 2; | |||
vector<int64_t> fixed_addr_offset_; | |||
std::vector<int64_t> fixed_addr_offset_; | |||
}; | |||
} // namespace ge | |||
#endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_STREAM_SWITCH_TASK_INFO_H_ |
@@ -25,10 +25,11 @@ Status SuperKernel::Launch(rtStream_t stream, uint32_t dump_flag) { | |||
const void *args[] = {this->GetNavTablePtr(), | |||
reinterpret_cast<const void *>(static_cast<uintptr_t>(this->GetNavTableSize()))}; | |||
rtError_t rt_ret = rtMalloc((void **)&(device_args_addr_), sizeof(args), RT_MEMORY_HBM); | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMalloc failied. error: 0x%X", rt_ret); return | |||
RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
rt_ret = rtMemcpy((void *)device_args_addr_, sizeof(args), (void *)args, sizeof(args), RT_MEMCPY_HOST_TO_DEVICE); | |||
rtError_t rt_ret = rtMalloc(reinterpret_cast<void **>(&device_args_addr_), sizeof(args), RT_MEMORY_HBM); | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMalloc failied. error: 0x%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
rt_ret = rtMemcpy(reinterpret_cast<void *>(device_args_addr_), sizeof(args), (void *)args, sizeof(args), | |||
RT_MEMCPY_HOST_TO_DEVICE); | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMemcpy failied. error: 0x%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
rt_ret = rtKernelLaunchWithFlag((void *const)func_stub_, block_dim_, device_args_addr_, sizeof(args), NULL, stream, | |||
@@ -87,7 +87,7 @@ Status SuperKernelFactory::FuseKernels(const std::vector<void *> &stub_func_list | |||
} | |||
GELOGI("SKT: superkernel start fuse, superkernel size %zu.", stub_func_list.size()); | |||
const size_t nav_table_len = 2 * stub_func_list.size(); | |||
std::unique_ptr<uint64_t[]> nav_table(new(std::nothrow) uint64_t[nav_table_len]); | |||
std::unique_ptr<uint64_t[]> nav_table(new (std::nothrow) uint64_t[nav_table_len]); | |||
GE_CHECK_NOTNULL(nav_table); | |||
uint64_t nav_table_size = 2 * stub_func_list.size() * sizeof(int64_t); | |||
@@ -106,16 +106,16 @@ Status SuperKernelFactory::FuseKernels(const std::vector<void *> &stub_func_list | |||
nav_table[i * 2 + 1] = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(args_addr_list[i])); | |||
GELOGD("SKT: fuseKernels args base address %lu", nav_table[i * 2 + 1]); | |||
} | |||
rt_ret = rtMalloc((void **)&hbm_nav_table_addr, nav_table_size, RT_MEMORY_HBM); | |||
rt_ret = rtMalloc(reinterpret_cast<void **>(&hbm_nav_table_addr), nav_table_size, RT_MEMORY_HBM); | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMalloc failed. error: 0x%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
rt_ret = | |||
rtMemcpy((void *)hbm_nav_table_addr, nav_table_size, (void *)nav_table.get(), nav_table_size, RT_MEMCPY_HOST_TO_DEVICE); | |||
rt_ret = rtMemcpy(reinterpret_cast<void *>(hbm_nav_table_addr), nav_table_size, | |||
reinterpret_cast<void *>(nav_table.get()), nav_table_size, RT_MEMCPY_HOST_TO_DEVICE); | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMemcpy failed. error: 0x%X", rt_ret); | |||
GE_CHK_RT(rtFree(hbm_nav_table_addr)); return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
// Create the necessary metadata for the super kernel | |||
h = std::unique_ptr<skt::SuperKernel>( | |||
new SuperKernel(this->func_stub_, hbm_nav_table_addr, nav_table_size, block_dim)); | |||
h = | |||
std::unique_ptr<skt::SuperKernel>(new SuperKernel(this->func_stub_, hbm_nav_table_addr, nav_table_size, block_dim)); | |||
return SUCCESS; | |||
} | |||
} // namespace skt | |||
@@ -63,8 +63,8 @@ struct RuntimeParam { | |||
}; | |||
typedef struct FusionOpInfo { | |||
vector<string> original_op_names; | |||
string op_name; | |||
std::vector<std::string> original_op_names; | |||
std::string op_name; | |||
uint32_t op_index; | |||
uint32_t stream_id; | |||
} FusionOpInfo; | |||
@@ -131,7 +131,7 @@ Status ZeroCopyTask::UpdateTaskParam(uintptr_t addr, void *buffer_addr, const ma | |||
auto dst_addr = static_cast<uint8_t *>(buffer_addr); | |||
GELOGI("[ZCPY] %s update task, args_addr: %p, size: %zu, offset: %zu, virtual_addr: 0x%lx, user_data_addr: %p", | |||
name_.c_str(), args_addr_, args_size_, offset, addr, buffer_addr); | |||
*(uintptr_t *)(args_info + offset) = reinterpret_cast<uintptr_t>(dst_addr); | |||
*reinterpret_cast<uintptr_t *>(args_info + offset)= reinterpret_cast<uintptr_t>(dst_addr); | |||
is_updated_ = true; | |||
} | |||
} | |||
@@ -26,6 +26,7 @@ | |||
#include <vector> | |||
#include "common/ge/ge_util.h" | |||
#include "framework/common/debug/ge_log.h" | |||
#include "framework/common/debug/log.h" | |||
#include "framework/common/types.h" | |||
#include "graph/debug/ge_attr_define.h" | |||
#include "graph/utils/graph_utils.h" | |||
@@ -72,7 +73,7 @@ Status DynamicShapePartitioner::Partition() { | |||
} | |||
REQUIRE(AttrUtils::SetBool(*root_graph_, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, true), | |||
"Failed set dynamic shape partitioned flag on root graph %s.", root_graph_->GetName().c_str()); | |||
REQUIRE_SUCCESS(CtrlEdgeTransfer(), "Failed do ctrl edge transfer!"); | |||
DumpGraph("_Before_DSP"); | |||
auto status = PartitionImpl(); | |||
GELOGD("%s.", DebugString().c_str()); | |||
@@ -86,6 +87,50 @@ Status DynamicShapePartitioner::Partition() { | |||
return status; | |||
} | |||
Status DynamicShapePartitioner::CtrlEdgeTransfer() { | |||
GELOGD("Do ctrl edge transfer start!"); | |||
GE_CHECK_NOTNULL(root_graph_); | |||
bool is_dynamic_shape = false; | |||
(void)AttrUtils::GetBool(root_graph_, ATTR_NAME_DYNAMIC_SHAPE_PARTITIONED, is_dynamic_shape); | |||
if (!is_dynamic_shape) { | |||
return SUCCESS; | |||
} | |||
for (auto &subgraph : root_graph_->GetAllSubgraphs()) { | |||
for (ge::NodePtr &n : subgraph->GetDirectNode()) { | |||
auto op_desc = n->GetOpDesc(); | |||
if (op_desc == nullptr) { | |||
continue; | |||
} | |||
auto op_type = op_desc->GetType(); | |||
if (op_type == CONSTANT || op_type == CONSTANTOP) { | |||
if (n->GetInAllNodes().empty()) { | |||
GELOGD("[CtrlEdgeTransferPass] node [%s] in nodes is empty", n->GetName().c_str()); | |||
continue; | |||
} | |||
GELOGD("start to tranfer ctrl edge for const node [%s]", n->GetName().c_str()); | |||
for (auto &in_control_node : n->GetInControlNodes()) { | |||
GE_CHECK_NOTNULL(in_control_node); | |||
GE_CHK_STATUS_RET(ge::GraphUtils::RemoveEdge(in_control_node->GetOutControlAnchor(), | |||
n->GetInControlAnchor()), "remove edge failed"); | |||
for (auto &out_node : n->GetOutNodes()) { | |||
if (out_node == nullptr) { | |||
continue; | |||
} | |||
GE_CHK_STATUS_RET(ge::GraphUtils::AddEdge(in_control_node->GetOutControlAnchor(), | |||
out_node->GetInControlAnchor()), "add edge failed."); | |||
} | |||
} | |||
} | |||
} | |||
} | |||
GELOGD("Do ctrl edge transfer end!"); | |||
return SUCCESS; | |||
} | |||
Status DynamicShapePartitioner::PartitionImpl() { | |||
REQUIRE_SUCCESS(root_graph_->TopologicalSorting(), "Graph topological sort failed."); | |||
REQUIRE_SUCCESS(InitClusters(), "Failed init cluster nodes."); | |||
@@ -151,6 +151,7 @@ class DynamicShapePartitioner { | |||
Status IsUnknownShapeGraph(ge::ComputeGraphPtr graph, bool &is_unknow); | |||
Status IsUnknownShapeNode(ge::NodePtr node, bool &is_unknow); | |||
bool IsUnknownShapeTensor(const ge::GeTensorDesc &tensor); | |||
Status CtrlEdgeTransfer(); | |||
ge::ComputeGraphPtr root_graph_; // The original graph to partition | |||
std::unordered_map<NodePtr, std::shared_ptr<Cluster>> node_2_cluster_; // Record nodes and the cluster it belongs to | |||
// topological sorted clusters, this field will change with the splitting. | |||
@@ -37,10 +37,6 @@ | |||
#include "graph/utils/type_utils.h" | |||
namespace ge { | |||
namespace { | |||
const uint32_t kShapeDimSize = 1; | |||
const uint32_t DIM_SIZE_TWO = 2; | |||
} // namespace | |||
Status PassUtils::ConstructTensorDescWithData(const GeTensorDesc &out_desc, std::vector<int64_t> &data, | |||
std::vector<GeTensorPtr> &v_output, const bool scalar_output) { | |||
@@ -63,7 +63,7 @@ std::string TransOpBreadthFusionPass::GetNodeId(const int anchor_index, const No | |||
GE_IF_BOOL_EXEC(node == nullptr || node->GetOpDesc() == nullptr, GELOGE(FAILED, "node is null"); return ""); | |||
if (node->GetType() == CAST) { | |||
trans_data_type = true; | |||
} else if (node->GetType() == TRANSPOSE || node->GetType() == TRANSPOSED) { | |||
} else if (node->GetType() == TRANSPOSE || node->GetType() == TRANSPOSED || node->GetType() == EXPANDDIMS) { | |||
trans_format = true; | |||
trans_shape = true; | |||
} else if (node->GetType() == TRANSDATA) { | |||
@@ -8,7 +8,7 @@ set(SRC_LIST | |||
"engine/host_cpu_engine.cc" | |||
"ops_kernel_store/host_cpu_ops_kernel_info.cc" | |||
"ops_kernel_store/op/op_factory.cc" | |||
"ops_kernel_store/op/host_op.cc" | |||
"ops_kernel_store/op/host_op.cc" | |||
) | |||
set(CPU_OPS_KERNEL_LIST | |||
@@ -98,7 +98,7 @@ target_link_libraries(atc_host_cpu_engine PRIVATE | |||
set_target_properties(atc_host_cpu_engine PROPERTIES | |||
OUTPUT_NAME host_cpu_engine | |||
LIBRARY_OUTPUT_DIRECTORY atclib | |||
LIBRARY_OUTPUT_DIRECTORY atclib | |||
) | |||
############ libhost_cpu_opskernel_builder.so ############ | |||
@@ -185,7 +185,7 @@ set_target_properties(atc_host_cpu_opskernel_builder PROPERTIES | |||
) | |||
############ libhost_cpu_opskernel_builder.a ############ | |||
add_library(host_cpu_opskernel_builder_static SHARED ${CPU_OPS_KERNEL_LIST}) | |||
add_library(host_cpu_opskernel_builder_static STATIC ${CPU_OPS_KERNEL_LIST}) | |||
target_compile_options(host_cpu_opskernel_builder_static PRIVATE | |||
-Werror | |||
@@ -112,8 +112,8 @@ void FloorDivKernel::ShapeCal(const std::vector<ge::ConstGeTensorPtr> &input, Ge | |||
template <typename T> | |||
T FloorDivKernel::DivCal(const T &x_i, const T &y_i) { | |||
if ((x_i < static_cast<T>(0)) != (y_i < static_cast<T>(0))) { | |||
T abs_x_i = std::abs(x_i); | |||
T abs_y_i = std::abs(y_i); | |||
T abs_x_i = x_i < 0 ? -x_i : x_i; | |||
T abs_y_i = y_i < 0 ? -y_i : y_i; | |||
return static_cast<T>(static_cast<int32_t>(-(abs_x_i + abs_y_i - 1) / abs_y_i)); | |||
} else { | |||
return static_cast<T>(static_cast<int32_t>(x_i / y_i)); | |||
@@ -40,10 +40,6 @@ class FloorDivKernel : public Kernel { | |||
template <typename T> | |||
Status DataCal(const std::vector<ConstGeTensorPtr> &input, ge::GeTensorPtr output_ptr); | |||
Status ComputeByDataType(DataType data_type, const std::vector<ConstGeTensorPtr> &input, GeTensorPtr output_ptr); | |||
int64_t axis_dim_; | |||
int64_t head_dim_; | |||
int64_t end_dim_; | |||
}; | |||
} // namespace ge | |||
@@ -187,7 +187,7 @@ Status SsdPriorboxKernel::GetNumPriorAndDimSize(uint32_t aspect_ratios_size, uin | |||
return PARAM_INVALID; | |||
} | |||
uint tmp_value = aspect_ratios_size * min_sizes_size; | |||
uint32_t tmp_value = aspect_ratios_size * min_sizes_size; | |||
if (ge::CheckUint32AddOverflow(tmp_value, max_sizes_size) != SUCCESS) { | |||
GELOGW("Failed to get list param."); | |||
return PARAM_INVALID; | |||
@@ -199,7 +199,7 @@ Status SsdPriorboxKernel::GetNumPriorAndDimSize(uint32_t aspect_ratios_size, uin | |||
return PARAM_INVALID; | |||
} | |||
num_priors = static_cast<int>(tmp_value); | |||
if (ge::CheckIntMulOverflow(layer_width, layer_height) != SUCCESS) { | |||
GELOGW("Failed to get list param."); | |||
return PARAM_INVALID; | |||
@@ -288,7 +288,7 @@ std::unique_ptr<float[]> SsdPriorboxKernel::BoundaryCalulate(int dim_size, int l | |||
} | |||
} | |||
return std::move(output_data); | |||
return output_data; | |||
} | |||
Status SsdPriorboxKernel::Compute(const NodePtr &node, std::vector<GeTensorPtr> &v_output) { | |||
@@ -77,7 +77,7 @@ do { \ | |||
RECORD_PROFILING_EVENT((context), HybridProfiler::EXECUTION, fmt, "Execution", name, ##__VA_ARGS__) | |||
#define RECORD_CALLBACK_EVENT(context, name, fmt, ...) \ | |||
RECORD_PROFILING_EVENT((context), HybridProfiler::CALLBACK, fmt, "Callback", name, ##__VA_ARGS__) | |||
RECORD_PROFILING_EVENT((context), HybridProfiler::CALLBACKS, fmt, "Callback", name, ##__VA_ARGS__) | |||
} // namespace hybrid | |||
} // namespace ge | |||
#endif // GE_HYBRID_EXECUTOR_HYBRID_EXECUTION_CONTEXT_H_ |
@@ -353,6 +353,44 @@ Status HybridModelAsyncExecutor::CopyOutputs(HybridModelExecutor::ExecuteArgs &a | |||
return SUCCESS; | |||
} | |||
Status HybridModelAsyncExecutor::Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc) { | |||
GELOGI("Start to execute model."); | |||
HybridModelExecutor::ExecuteArgs args; | |||
args.inputs.resize(inputs.size()); | |||
for (size_t i = 0; i < inputs.size(); ++i) { | |||
TensorValue tensor_value(inputs[i].data, inputs[i].length); | |||
args.inputs[i] = tensor_value; | |||
} | |||
GE_CHK_STATUS_RET(executor_->Execute(args), "Failed to execute model."); | |||
for (const auto &output_tensor_desc : args.output_desc) { | |||
output_desc.emplace_back(*output_tensor_desc); | |||
} | |||
for (size_t i = 0; i < args.outputs.size(); ++i) { | |||
int64_t output_real_size = 0; | |||
ge::graphStatus graph_status = TensorUtils::GetTensorSizeInBytes(output_desc[i], output_real_size); | |||
if (graph_status != GRAPH_SUCCESS) { | |||
GELOGE(FAILED, "Get tensor size in bytes failed."); | |||
return FAILED; | |||
} | |||
if (output_real_size > 0) { | |||
if (outputs[i].length < static_cast<uint64_t>(output_real_size)) { | |||
GELOGE(FAILED, "output idx[%zu], the memory size of output[%lu] given by user should be greater than or equal to the real size of output[%ld]", | |||
i, outputs[i].length, output_real_size); | |||
return FAILED; | |||
} | |||
GE_CHK_RT_RET(rtMemcpy(outputs[i].data, outputs[i].length, args.outputs[i].GetData(), output_real_size, RT_MEMCPY_DEVICE_TO_DEVICE)); | |||
} | |||
outputs[i].length = output_real_size; | |||
} | |||
return SUCCESS; | |||
} | |||
Status HybridModelAsyncExecutor::Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs) { | |||
GELOGD("Start to execute model."); | |||
// prepare inputs | |||
@@ -35,6 +35,11 @@ class HybridModelAsyncExecutor { | |||
Status Init(); | |||
Status Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc); | |||
Status Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs); | |||
Status Start(const std::shared_ptr<ModelListener> &listener); | |||
@@ -82,7 +82,7 @@ Status HybridModelExecutor::ExecuteGraphInternal(SubgraphExecutor &executor, | |||
Status HybridModelExecutor::Cleanup() { | |||
GELOGD("Start to cleanup."); | |||
context_.callback_manager->Destroy(); | |||
RuntimeInferenceContext::DestroyContext(to_string(context_.session_id)); | |||
RuntimeInferenceContext::DestroyContext(std::to_string(context_.session_id)); | |||
GELOGD("Cleanup successfully."); | |||
return SUCCESS; | |||
} | |||
@@ -33,7 +33,7 @@ class HybridProfiler { | |||
SHAPE_INFERENCE, | |||
COMPILE, | |||
EXECUTION, | |||
CALLBACK | |||
CALLBACKS | |||
}; | |||
struct Event { | |||
@@ -27,7 +27,7 @@ | |||
namespace ge { | |||
namespace hybrid { | |||
class NodeTask; | |||
class GraphExecutionContext; | |||
struct GraphExecutionContext; | |||
class SubgraphContext; | |||
class ShapeFuture { | |||
@@ -38,6 +38,14 @@ class HybridDavinciModel::Impl { | |||
return SUCCESS; | |||
} | |||
Status Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc, | |||
rtStream_t stream) { | |||
return executor_.Execute(inputs, input_desc, outputs, output_desc); | |||
} | |||
Status Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs) { | |||
return executor_.Execute(inputs, outputs); | |||
} | |||
@@ -68,6 +76,33 @@ class HybridDavinciModel::Impl { | |||
executor_.SetDeviceId(device_id); | |||
} | |||
uint64_t GetSessionId() { | |||
return model_.GetSessionId(); | |||
} | |||
Status GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type) { | |||
return model_.GetDynamicBatchInfo(batch_info, dynamic_type); | |||
} | |||
void GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) { | |||
model_.GetUserDesignateShapeOrder(user_input_shape_order); | |||
} | |||
void GetModelAttr(std::vector<std::string> &dynamic_output_shape_info) { | |||
model_.GetModelAttr(dynamic_output_shape_info); | |||
} | |||
Status GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &output_formats) { | |||
return model_.GetInputOutputDescInfo(input_desc, output_desc, input_formats, output_formats); | |||
} | |||
void SetModelDescVersion(bool is_new_model_desc) { | |||
model_.SetModelDescVersion(is_new_model_desc); | |||
} | |||
private: | |||
std::shared_ptr<ModelListener> listener_; | |||
HybridModel model_; | |||
@@ -78,8 +113,8 @@ HybridDavinciModel::~HybridDavinciModel() { | |||
delete impl_; | |||
} | |||
unique_ptr<HybridDavinciModel> HybridDavinciModel::Create(const GeRootModelPtr &ge_root_model) { | |||
auto instance = unique_ptr<HybridDavinciModel>(new (std::nothrow)HybridDavinciModel()); | |||
std::unique_ptr<HybridDavinciModel> HybridDavinciModel::Create(const GeRootModelPtr &ge_root_model) { | |||
auto instance = std::unique_ptr<HybridDavinciModel>(new (std::nothrow)HybridDavinciModel()); | |||
if (instance != nullptr) { | |||
instance->impl_ = new (std::nothrow) HybridDavinciModel::Impl(ge_root_model); | |||
if (instance->impl_ != nullptr) { | |||
@@ -95,6 +130,14 @@ Status HybridDavinciModel::Init() { | |||
return impl_->Init(); | |||
} | |||
Status HybridDavinciModel::Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc, rtStream_t stream) { | |||
GE_CHECK_NOTNULL(impl_); | |||
return impl_->Execute(inputs, input_desc, outputs, output_desc, stream); | |||
} | |||
Status HybridDavinciModel::Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs) { | |||
GE_CHECK_NOTNULL(impl_); | |||
return impl_->Execute(inputs, outputs); | |||
@@ -132,5 +175,41 @@ void HybridDavinciModel::SetDeviceId(uint32_t device_id) { | |||
impl_->SetDeviceId(device_id); | |||
} | |||
} | |||
Status HybridDavinciModel::GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type) { | |||
GE_CHECK_NOTNULL(impl_); | |||
return impl_->GetDynamicBatchInfo(batch_info, dynamic_type); | |||
} | |||
void HybridDavinciModel::GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) { | |||
if (impl_ != nullptr) { | |||
impl_->GetUserDesignateShapeOrder(user_input_shape_order); | |||
} | |||
} | |||
void HybridDavinciModel::GetModelAttr(std::vector<std::string> &dynamic_output_shape_info) { | |||
if (impl_ != nullptr) { | |||
impl_->GetModelAttr(dynamic_output_shape_info); | |||
} | |||
} | |||
Status HybridDavinciModel::GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &output_formats) { | |||
GE_CHECK_NOTNULL(impl_); | |||
return impl_->GetInputOutputDescInfo(input_desc, output_desc, input_formats, output_formats); | |||
} | |||
void HybridDavinciModel::SetModelDescVersion(bool is_new_model_desc) { | |||
if (impl_ != nullptr) { | |||
impl_->SetModelDescVersion(is_new_model_desc); | |||
} | |||
} | |||
uint64_t HybridDavinciModel::GetSessionId() { | |||
GE_CHECK_NOTNULL(impl_); | |||
return impl_->GetSessionId(); | |||
} | |||
} // namespace hybrid | |||
} // namespace ge |
@@ -37,6 +37,12 @@ class HybridDavinciModel { | |||
Status Init(); | |||
Status Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc, | |||
rtStream_t stream); | |||
Status Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs); | |||
Status ModelRunStart(); | |||
@@ -51,6 +57,21 @@ class HybridDavinciModel { | |||
void SetDeviceId(uint32_t device_id); | |||
uint64_t GetSessionId(); | |||
Status GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type); | |||
void GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order); | |||
void GetModelAttr(std::vector<std::string> &dynamic_output_shape_info); | |||
Status GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &output_formats); | |||
void SetModelDescVersion(bool is_new_model_desc); | |||
private: | |||
HybridDavinciModel() = default; | |||
class Impl; | |||
@@ -28,6 +28,14 @@ Status HybridDavinciModel::Init() { | |||
return UNSUPPORTED; | |||
} | |||
Status HybridDavinciModel::Execute(const std::vector<DataBuffer> &inputs, | |||
const std::vector<GeTensorDesc> &input_desc, | |||
std::vector<DataBuffer> &outputs, | |||
std::vector<GeTensorDesc> &output_desc, | |||
rtStream_t stream) { | |||
return UNSUPPORTED; | |||
} | |||
Status HybridDavinciModel::Execute(const vector<GeTensor> &inputs, vector<GeTensor> &outputs) { | |||
return UNSUPPORTED; | |||
} | |||
@@ -52,5 +60,29 @@ void HybridDavinciModel::SetModelId(uint32_t model_id) { | |||
void HybridDavinciModel::SetDeviceId(uint32_t device_id) { | |||
} | |||
uint64_t HybridDavinciModel::GetSessionId() { | |||
return 0; | |||
} | |||
Status HybridDavinciModel::GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type) { | |||
return UNSUPPORTED; | |||
} | |||
void HybridDavinciModel::GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) { | |||
} | |||
void HybridDavinciModel::GetModelAttr(std::vector<std::string> &dynamic_output_shape_info) { | |||
} | |||
Status HybridDavinciModel::GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &output_formats) { | |||
return UNSUPPORTED; | |||
} | |||
void HybridDavinciModel::SetModelDescVersion(bool is_new_model_desc) { | |||
} | |||
} // namespace hybrid | |||
} // namespace ge |
@@ -21,12 +21,18 @@ | |||
#include "graph/utils/graph_utils.h" | |||
#include "graph/utils/node_utils.h" | |||
#include "graph/utils/tensor_utils.h" | |||
#include "graph/utils/type_utils.h" | |||
#include "hybrid/common/npu_memory_allocator.h" | |||
#include "hybrid/model/hybrid_model_builder.h" | |||
#include "hybrid/node_executor/node_executor.h" | |||
#include "common/op/ge_op_utils.h" | |||
namespace ge { | |||
namespace hybrid { | |||
namespace { | |||
const int64_t kMemSizeUnknownShape = -1; // Unknown shape mem size | |||
} | |||
HybridModel::HybridModel(GeRootModelPtr ge_model) : ge_root_model_(std::move(ge_model)) { | |||
} | |||
@@ -128,7 +134,187 @@ const GraphItem *HybridModel::GetSubgraphItem(const ComputeGraphPtr &subgraph) c | |||
} | |||
const string &HybridModel::GetModelName() const { | |||
return model_name_; | |||
return model_name_; | |||
} | |||
Status HybridModel::GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type) { | |||
// dynamic shape do not need dynamic batch | |||
batch_info = {}; | |||
dynamic_type = -1; | |||
return SUCCESS; | |||
} | |||
void HybridModel::GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) { | |||
// dynamic shape do not need dynamic batch | |||
user_input_shape_order = {}; | |||
} | |||
void HybridModel::GetModelAttr(std::vector<std::string> &dynamic_output_shape_info) { | |||
dynamic_output_shape_info = {}; | |||
} | |||
Status HybridModel::GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &output_formats) { | |||
auto node_item_list = root_graph_item_->GetInputNodes(); | |||
if (node_item_list.empty()) { | |||
GELOGE(FAILED, "node item list is empty!"); | |||
return FAILED; | |||
} | |||
GE_CHECK_NOTNULL(node_item_list[0]->node); | |||
GE_CHECK_NOTNULL(node_item_list[0]->node->GetOpDesc()); | |||
if (node_item_list[0]->node->GetOpDesc()->GetInputsSize() != 1) { | |||
GELOGE(FAILED, "input size of op is not 1!"); | |||
return FAILED; | |||
} | |||
GE_CHK_STATUS_RET(GetInputDescInfo(input_desc, input_formats), "get input desc info failed"); | |||
GE_CHK_STATUS_RET(GetOutputDescInfo(output_desc, output_formats), "get ouput desc info failed"); | |||
return SUCCESS; | |||
} | |||
void HybridModel::SetInputDimsAndShapeRangesInfo(const vector<int64_t> &model_input_dims, std::vector<std::pair<int64_t,int64_t>> &shape_ranges, | |||
InputOutputDescInfo &input) { | |||
for (auto model_input_dim : model_input_dims) { | |||
input.shape_info.dims.push_back(model_input_dim); | |||
} | |||
input.shape_info.shape_ranges = shape_ranges; | |||
return; | |||
} | |||
void HybridModel::CreateInputDimsInfo(const OpDescPtr &op_desc, InputOutputDescInfo &input) { | |||
std::vector<std::pair<int64_t,int64_t>> shape_ranges; | |||
if (is_new_model_desc_ && op_desc->HasAttr(ATTR_NAME_INPUT_DIMS)) { | |||
// When static aipp is set, need to get the model input dims which processed by aipp | |||
vector<int64_t> model_input_dims; | |||
(void)AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_DIMS, model_input_dims); | |||
SetInputDimsAndShapeRangesInfo(model_input_dims, shape_ranges, input); | |||
return; | |||
} | |||
// judge if this data is linked dynamic aipp first, multiply batch has been considered | |||
if (op_desc->HasAttr("_dynamic_aipp_input_dims")) { | |||
vector<int64_t> dynamic_aipp_input_dims; | |||
(void)AttrUtils::GetListInt(op_desc, "_dynamic_aipp_input_dims", dynamic_aipp_input_dims); | |||
SetInputDimsAndShapeRangesInfo(dynamic_aipp_input_dims, shape_ranges, input); | |||
return; | |||
} else { | |||
vector<int64_t> input_dims = op_desc->GetInputDescPtr(0)->GetShape().GetDims(); | |||
op_desc->GetInputDescPtr(0)->GetShapeRange(shape_ranges); | |||
SetInputDimsAndShapeRangesInfo(input_dims, shape_ranges, input); | |||
return; | |||
} | |||
} | |||
Status HybridModel::GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, std::vector<uint32_t> &formats) { | |||
auto node_item_list = root_graph_item_->GetInputNodes(); | |||
for (auto &node_item : node_item_list) { | |||
InputOutputDescInfo input; | |||
GE_CHECK_NOTNULL(node_item->node); | |||
auto op_desc = node_item->node->GetOpDesc(); | |||
GE_CHECK_NOTNULL(op_desc); | |||
GE_CHECK_NOTNULL(op_desc->GetInputDescPtr(0)); | |||
Format format = op_desc->GetInputDescPtr(0)->GetFormat(); | |||
input.data_type = op_desc->GetInputDescPtr(0)->GetDataType(); | |||
input.name = op_desc->GetName(); | |||
int64_t input_size = 0; | |||
GE_CHK_STATUS_RET(TensorUtils::GetSize(*op_desc->GetInputDescPtr(0), input_size), "get input size failed."); | |||
// support dynamic shape | |||
if (input_size < 0) { | |||
GELOGD("dynamic shape scene, input size is unknown. " | |||
"format=%d, data_type=%d, input_size=%ld", | |||
format, input.data_type, input_size); | |||
input_size = kMemSizeUnknownShape; // -1 | |||
} | |||
// not support dynamic shape input for now, so input_size here will be not less than zero. | |||
input.size = input_size; | |||
CreateInputDimsInfo(op_desc, input); | |||
formats.push_back(format); | |||
input_desc.push_back(input); | |||
} | |||
is_new_model_desc_ = false; | |||
return SUCCESS; | |||
} | |||
void HybridModel::CreateOutput(ConstGeTensorDescPtr &output_desc, InputOutputDescInfo &output_desc_info, uint32_t &format_result) { | |||
GE_IF_BOOL_EXEC(output_desc == nullptr, GELOGE(FAILED, "output desc ptr is nullptr"); return ); | |||
Format format = output_desc->GetFormat(); | |||
GeShape shape = output_desc->GetShape(); | |||
std::vector<std::pair<int64_t,int64_t>> shape_ranges; | |||
output_desc->GetShapeRange(shape_ranges); | |||
DataType data_type = output_desc->GetDataType(); | |||
format_result = format; | |||
if (format == FORMAT_FRACTAL_Z) { // FraczToHWCK | |||
int64_t k = shape.GetDim(0); // 0: first dim | |||
int64_t c = shape.GetDim(1); // 1: second dim | |||
int64_t h = shape.GetDim(2); // 2: third dim | |||
int64_t w = shape.GetDim(3); // 3: forth dim | |||
output_desc_info.shape_info.dims.push_back(h); | |||
output_desc_info.shape_info.dims.push_back(w); | |||
output_desc_info.shape_info.dims.push_back(c); | |||
output_desc_info.shape_info.dims.push_back(k); | |||
if (shape_ranges.size() == 4) { // 4 dims | |||
output_desc_info.shape_info.shape_ranges.push_back(shape_ranges[2]); // h:2 | |||
output_desc_info.shape_info.shape_ranges.push_back(shape_ranges[3]); // w:3 | |||
output_desc_info.shape_info.shape_ranges.push_back(shape_ranges[1]); // c:1 | |||
output_desc_info.shape_info.shape_ranges.push_back(shape_ranges[0]); // k:0 | |||
} | |||
format_result = FORMAT_HWCN; | |||
} else { | |||
for (size_t j = 0; j < shape.GetDimNum(); j++) { | |||
output_desc_info.shape_info.dims.push_back(shape.GetDim(j)); | |||
} | |||
output_desc_info.shape_info.shape_ranges = shape_ranges; | |||
} | |||
int64_t tensor_size = 0; | |||
(void)TensorUtils::CalcTensorMemSize(shape, format, data_type, tensor_size); | |||
output_desc_info.size = static_cast<uint64_t>(tensor_size); | |||
output_desc_info.data_type = output_desc->GetDataType(); | |||
} | |||
Status HybridModel::GetOutputDescInfo(vector<InputOutputDescInfo> &output_desc, std::vector<uint32_t> &formats) { | |||
std::vector<ConstGeTensorDescPtr> output_desc_list; | |||
GE_CHK_STATUS_RET(root_graph_item_->GetOutputDescList(output_desc_list), "get output desc info failed"); // output_desc_list contains vaild input desc | |||
vector<std::string> out_node_names; | |||
(void)ge::AttrUtils::GetListStr(ge_root_model_->GetRootGraph(), ATTR_MODEL_OUT_NODES_NAME, out_node_names); | |||
GE_CHECK_NOTNULL(root_graph_item_->GetOutputNode()); | |||
auto op_desc = root_graph_item_->GetOutputNode()->op_desc; | |||
GE_CHECK_NOTNULL(op_desc); | |||
auto out_size = static_cast<uint32_t>(op_desc->GetInputsSize()); | |||
GE_CHK_BOOL_RET_STATUS(out_size == output_desc_list.size(), FAILED, "output size[%u] not match output_desc_list size[%zu]", out_size, output_desc_list.size()); | |||
for (uint32_t index = 0; index < out_size; ++index) { | |||
string output_name; | |||
std::vector<std::string> src_name = op_desc->GetSrcName(); | |||
std::vector<int64_t> src_index = op_desc->GetSrcIndex(); | |||
if (out_size == out_node_names.size()) { | |||
bool contains_colon = out_node_names[index].find(":") != std::string::npos; | |||
output_name = contains_colon ? out_node_names[index] : out_node_names[index] + ":" + std::to_string(src_index[index]); | |||
} else { | |||
output_name = std::string("output_") + std::to_string(index) + "_" + src_name[index] + "_" + std::to_string(src_index[index]); | |||
} | |||
InputOutputDescInfo output_desc_info; | |||
output_desc_info.name = output_name; | |||
uint32_t format_result; | |||
CreateOutput(output_desc_list[index], output_desc_info, format_result); | |||
output_desc.push_back(output_desc_info); | |||
formats.push_back(format_result); | |||
} | |||
return SUCCESS; | |||
} | |||
} // namespace hybrid | |||
} // namespace ge |
@@ -83,6 +83,30 @@ class HybridModel { | |||
const string &GetModelName() const; | |||
Status GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type); | |||
void GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order); | |||
void GetModelAttr(std::vector<std::string> &dynamic_output_shape_info); | |||
Status GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc, | |||
vector<InputOutputDescInfo> &output_desc, | |||
std::vector<uint32_t> &input_formats, | |||
std::vector<uint32_t> &outputFormats); | |||
Status GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, std::vector<uint32_t> &formats); | |||
void CreateOutput(ConstGeTensorDescPtr &output_desc, InputOutputDescInfo &output, uint32_t &format_result); | |||
Status GetOutputDescInfo(vector<InputOutputDescInfo> &output_desc, std::vector<uint32_t> &formats); | |||
void CreateInputDimsInfo(const OpDescPtr &op_desc, InputOutputDescInfo &input); | |||
void SetModelDescVersion(bool is_new_model_desc) { is_new_model_desc_ = is_new_model_desc; } | |||
void SetInputDimsAndShapeRangesInfo(const vector<int64_t> &model_input_dims, std::vector<std::pair<int64_t, int64_t>> &shape_ranges, | |||
InputOutputDescInfo &input); | |||
private: | |||
friend class HybridModelBuilder; | |||
friend class HybridModelAsyncExecutor; | |||
@@ -101,6 +125,8 @@ class HybridModel { | |||
std::map<std::string, std::unique_ptr<GraphItem>> subgraph_items_; | |||
std::map<NodePtr, std::unique_ptr<NodeItem>> node_items_; | |||
bool is_new_model_desc_ = false; // support aipp | |||
// runtime fields | |||
uint32_t device_id_ = 0; | |||
uint32_t model_id_ = 0; | |||
@@ -27,16 +27,41 @@ | |||
#include "graph/utils/graph_utils.h" | |||
#include "hybrid/common/npu_memory_allocator.h" | |||
#include "hybrid/node_executor/node_executor.h" | |||
#include "framework/common/debug/ge_log.h" | |||
#include "graph/utils/attr_utils.h" | |||
namespace ge { | |||
namespace hybrid { | |||
namespace { | |||
const uint32_t kSubgraphIndex = 0U; | |||
const uint32_t kVarOutputIndex = 0U; | |||
const uint32_t kAlignment = 32; | |||
const int kBytes = 8; | |||
const char *const kOwnerGraphIsUnknown = "OwnerGraphIsUnknown"; | |||
Status SetOutputNameAttr(ComputeGraph &graph) { | |||
vector<string> output_names; | |||
for (const auto &node : graph.GetDirectNode()) { | |||
auto op_desc = node->GetOpDesc(); | |||
if (op_desc == nullptr) { | |||
continue; | |||
} | |||
auto op_type = op_desc->GetType(); | |||
if (op_type == NETOUTPUT) { | |||
for (InDataAnchorPtr &in_data_anchor : node->GetAllInDataAnchors()) { | |||
const OutDataAnchorPtr &peer_out_anchor = in_data_anchor->GetPeerOutAnchor(); | |||
GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue); | |||
NodePtr in_node = peer_out_anchor->GetOwnerNode(); | |||
GE_CHECK_NOTNULL(in_node); | |||
output_names.push_back(in_node->GetName()); | |||
} | |||
} | |||
} | |||
GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListStr(&graph, ATTR_MODEL_OUT_NODES_NAME, output_names), | |||
GELOGE(FAILED, "SetListStr of ATTR_MODEL_OUT_NODES_NAME failed."); | |||
return FAILED); | |||
return SUCCESS; | |||
} | |||
int64_t CalcVarSizeInBytes(const GeTensorDesc &desc) { | |||
int64_t var_size = 0; | |||
auto data_type = desc.GetDataType(); | |||
@@ -939,6 +964,10 @@ Status HybridModelBuilder::LoadGeModel(ComputeGraph &sub_graph, const GeModelPtr | |||
Status HybridModelBuilder::IndexTaskDefs() { | |||
const auto &root_graph = ge_root_model_->GetRootGraph(); | |||
if (SetOutputNameAttr(*root_graph) != SUCCESS) { | |||
GELOGW("Set output name attr failed."); | |||
} | |||
for (auto &it : ge_root_model_->GetSubgraphInstanceNameToModel()) { | |||
auto &name = it.first; | |||
auto &ge_model = it.second; | |||
@@ -957,7 +986,7 @@ Status HybridModelBuilder::IndexTaskDefs() { | |||
// index task defs | |||
GELOGD("To index tasks for subgraph: %s", name.c_str()); | |||
unordered_map<int64_t, NodePtr> node_map; | |||
std::unordered_map<int64_t, NodePtr> node_map; | |||
for (const auto &node : sub_graph->GetDirectNode()) { | |||
GE_CHECK_NOTNULL(node); | |||
GE_CHECK_NOTNULL(node->GetOpDesc()); | |||
@@ -19,6 +19,7 @@ | |||
#include "framework/common/debug/log.h" | |||
#include "hybrid/executor/hybrid_execution_context.h" | |||
#include "hybrid/node_executor/aicore/aicore_task_builder.h" | |||
#include "graph/load/new_model_manager/tbe_handle_store.h" | |||
using optiling::OpRunInfo; | |||
@@ -36,6 +37,58 @@ Status AiCoreOpTask::Init(const OpDesc &op_desc, const domi::TaskDef &task_def) | |||
return SUCCESS; | |||
} | |||
Status AiCoreOpTask::RegisterTbeHandle(const OpDesc &op_desc) { | |||
auto op_desc_ptr = std::make_shared<OpDesc>(op_desc); | |||
GE_CHECK_NOTNULL(op_desc_ptr); | |||
auto tbe_kernel = op_desc_ptr->TryGetExtAttr(OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr()); | |||
if (tbe_kernel == nullptr) { | |||
GELOGE(INTERNAL_ERROR, "TBE: %s can't find tvm bin file!", op_desc_ptr->GetName().c_str()); | |||
return INTERNAL_ERROR; | |||
} | |||
TBEHandleStore &kernel_store = TBEHandleStore::GetInstance(); | |||
rtError_t rt_ret = rtQueryFunctionRegistered(stub_name_.c_str()); | |||
if (rt_ret != RT_ERROR_NONE) { | |||
void *bin_handle = nullptr; | |||
if (!kernel_store.FindTBEHandle(stub_name_.c_str(), bin_handle)) { | |||
GELOGI("TBE: can't find the kernel_name[%s] in HandleMap", stub_name_.c_str()); | |||
rtDevBinary_t binary; | |||
std::string json_string; | |||
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, TVM_ATTR_NAME_MAGIC, json_string), | |||
GELOGI("Get original type of session_graph_id.")); | |||
if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AICPU") { | |||
binary.magic = RT_DEV_BINARY_MAGIC_ELF_AICPU; | |||
} else if (json_string == "RT_DEV_BINARY_MAGIC_ELF") { | |||
binary.magic = RT_DEV_BINARY_MAGIC_ELF; | |||
} else if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AIVEC") { | |||
binary.magic = RT_DEV_BINARY_MAGIC_ELF_AIVEC; | |||
} else { | |||
GELOGE(PARAM_INVALID, "TBE: Invalid parameter magic number! json: %s", json_string.c_str()); | |||
return PARAM_INVALID; | |||
} | |||
binary.version = 0; | |||
binary.data = tbe_kernel->GetBinData(); | |||
binary.length = tbe_kernel->GetBinDataSize(); | |||
GELOGI("TBE: binary.length: %lu", binary.length); | |||
GE_CHK_RT_RET(rtDevBinaryRegister(&binary, &bin_handle)); | |||
std::string meta_data; | |||
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, TVM_ATTR_NAME_METADATA, meta_data), | |||
GELOGI("Get original type of json_string")); | |||
GELOGI("TBE: meta data: %s", meta_data.empty() ? "null" : meta_data.c_str()); | |||
GE_IF_BOOL_EXEC(!meta_data.empty(), GE_CHK_RT_RET(rtMetadataRegister(bin_handle, meta_data.c_str()))); | |||
kernel_store.StoreTBEHandle(stub_name_.c_str(), bin_handle, tbe_kernel); | |||
} else { | |||
GELOGI("TBE: find the kernel_name[%s] in HandleMap", stub_name_.c_str()); | |||
kernel_store.ReferTBEHandle(stub_name_.c_str()); | |||
} | |||
std::string kernel_name; | |||
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc_ptr, op_desc_ptr->GetName() + "_kernelname", kernel_name), | |||
GELOGI("Get original type of kernel_name")); | |||
GELOGI("TBE: binfile_key=%s, kernel_name=%s", stub_name_.c_str(), kernel_name.c_str()); | |||
GE_CHK_RT_RET(rtFunctionRegister(bin_handle, stub_name_.c_str(), stub_name_.c_str(), kernel_name.c_str(), 0)); | |||
} | |||
return SUCCESS; | |||
} | |||
Status AiCoreOpTask::InitWithTaskDef(const OpDesc &op_desc, const domi::TaskDef &task_def) { | |||
GE_CHK_STATUS_RET(ValidateTaskDef(task_def), | |||
"[%s] Failed to validate task def: [%s]", | |||
@@ -45,6 +98,9 @@ Status AiCoreOpTask::InitWithTaskDef(const OpDesc &op_desc, const domi::TaskDef | |||
const domi::KernelDef &kernel_def = task_def.kernel(); | |||
const domi::KernelContext &context = kernel_def.context(); | |||
stub_name_ = kernel_def.stub_func(); | |||
GE_CHK_STATUS_RET(RegisterTbeHandle(op_desc)); | |||
GE_CHK_RT_RET(rtGetFunctionByName(stub_name_.c_str(), &stub_func_)); | |||
args_size_ = kernel_def.args_size(); | |||
block_dim_ = kernel_def.block_dim(); | |||
@@ -62,6 +62,7 @@ class AiCoreOpTask { | |||
static Status ValidateTaskDef(const domi::TaskDef &task_def); | |||
Status InitWithTaskDef(const OpDesc &node, const domi::TaskDef &task_def); | |||
Status InitTilingInfo(const OpDesc &op_desc); | |||
Status RegisterTbeHandle(const OpDesc &op_desc); | |||
std::string stub_name_; | |||
void *stub_func_ = nullptr; | |||
@@ -26,7 +26,7 @@ namespace hybrid { | |||
class AiCoreTaskCompiler : public TaskCompiler { | |||
public: | |||
AiCoreTaskCompiler() = default; | |||
~AiCoreTaskCompiler() = default; | |||
~AiCoreTaskCompiler() override = default; | |||
Status CompileOp(const NodePtr &node, std::vector<domi::TaskDef> &tasks) override; | |||
Status Initialize() override; | |||
@@ -644,8 +644,12 @@ Status AicpuNodeTask::Init(const HybridModel &model) { | |||
const auto &context = kernel_def.context(); | |||
auto kernel_type = static_cast<ccKernelType>(context.kernel_type()); | |||
if (kernel_type == ccKernelType::CUST_AI_CPU) { | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc, so_name), "load cust aicpu so failed."); | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LaunchCustAicpuSo(), "Launch cust aicpu so failed."); | |||
bool loaded = false; | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc, so_name, loaded), | |||
"load cust aicpu so failed."); | |||
if (!loaded) { | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LaunchCustAicpuSo(), "Launch cust aicpu so failed."); | |||
} | |||
} | |||
GE_CHK_BOOL_RET_STATUS(args.size() == args_size_, FAILED, | |||
@@ -37,6 +37,8 @@ class AicpuNodeTaskBase : public NodeTask { | |||
~AicpuNodeTaskBase() override = default; | |||
using NodeTask::Init; | |||
virtual Status Init(const HybridModel &model) = 0; | |||
Status UpdateArgs(TaskContext &context) override; | |||
@@ -405,7 +405,7 @@ Status ControlOpNodeExecutor::LoadTask(const HybridModel &model, | |||
auto node_item = model.GetNodeItem(node); | |||
GE_CHECK_NOTNULL(node_item); | |||
unique_ptr<ControlOpNodeTask> node_task; | |||
std::unique_ptr<ControlOpNodeTask> node_task; | |||
auto node_type = node->GetType(); | |||
if (node_type == IF || node_type == STATELESSIF) { | |||
node_task.reset(new(std::nothrow) IfOpNodeTask()); | |||
@@ -25,6 +25,7 @@ namespace ge { | |||
namespace hybrid { | |||
class ControlOpNodeTask : public NodeTask { | |||
public: | |||
using NodeTask::Init; | |||
virtual Status Init(const NodePtr &node, const HybridModel &model) = 0; | |||
Status UpdateArgs(TaskContext &context) override; | |||
@@ -68,7 +68,7 @@ Status RefInputTask::RefOneByOne(TaskContext &context) { | |||
node_name_.c_str(), node_type_.c_str(), output_num, input_num); | |||
return INTERNAL_ERROR; | |||
} | |||
for (uint32_t out_index = 0; out_index < output_num; ++out_index) { | |||
for (uint32_t out_index = 0; out_index < static_cast<uint32_t>(output_num); ++out_index) { | |||
auto input = context.GetInput(out_index); | |||
GE_CHECK_NOTNULL(input); | |||
GE_CHK_STATUS_RET(context.SetOutput(out_index, *input)); | |||
@@ -20,7 +20,6 @@ | |||
#include "hybrid/node_executor/host_cpu/kernel_factory.h" | |||
namespace { | |||
const size_t kAssignInputNum = 2; | |||
const size_t kAssignRefInputIndex = 0; | |||
const size_t kAssignValueInputIndex = 1; | |||
const size_t kAssignRefOutputIndex = 0; | |||
@@ -34,7 +34,6 @@ const char *const kEngineNameAiCpuTf = "aicpu_tf_kernel"; | |||
const char *const kEngineNameHccl = "ops_kernel_info_hccl"; | |||
const char *const kEngineNameRts = "DNN_VM_RTS_OP_STORE"; | |||
const char *const kEngineNameHostCpu = "DNN_VM_HOST_CPU_OP_STORE"; | |||
const char *const kOwnerGraphIsUnknown = "OwnerGraphIsUnknown"; | |||
} | |||
Status NodeExecutor::PrepareTask(NodeTask &task, TaskContext &context) const { | |||
GE_CHK_STATUS_RET_NOLOG(context.AllocateOutputs()); | |||
@@ -41,7 +41,6 @@ class PartitionedCallNodeTask : public NodeTask { | |||
const GraphItem *graph_item_; | |||
std::unique_ptr<SubgraphExecutor> subgraph_executor_; | |||
GraphExecutionContext *context_ = nullptr; | |||
}; | |||
class PartitionedCallNodeExecutor : public NodeExecutor { | |||
@@ -29,7 +29,7 @@ | |||
namespace ge { | |||
namespace hybrid { | |||
class GraphExecutionContext; | |||
struct GraphExecutionContext; | |||
class SubgraphContext; | |||
class TaskContext { | |||
@@ -11,13 +11,13 @@ set(SRC_LIST | |||
"main.cc" | |||
"single_op_parser.cc" | |||
"../session/omg.cc" | |||
"../ir_build/atc_ir_common.cc" | |||
"../ir_build/atc_ir_common.cc" | |||
) | |||
############ atc ############ | |||
add_executable(atc ${SRC_LIST} ${PROTO_HDRS}) | |||
target_compile_options(atc PRIVATE | |||
target_compile_options(atc PRIVATE | |||
-Werror | |||
-O2 | |||
-Wno-deprecated-declarations | |||
@@ -74,10 +74,130 @@ target_link_libraries(atc PRIVATE | |||
-ldl | |||
) | |||
############ atc.bin ############ | |||
add_executable(atc.bin ${SRC_LIST} ${PROTO_HDRS}) | |||
target_compile_options(atc.bin PRIVATE | |||
-Werror | |||
-O2 | |||
-Wno-deprecated-declarations | |||
) | |||
target_compile_definitions(atc.bin PRIVATE | |||
PROTOBUF_INLINE_NOT_IN_HEADERS=0 | |||
COMPILE_OMG_PACKAGE | |||
google=ascend_private | |||
) | |||
target_include_directories(atc.bin PRIVATE | |||
${CMAKE_CURRENT_LIST_DIR} | |||
${GE_CODE_DIR} | |||
${GE_CODE_DIR}/ge | |||
${GE_CODE_DIR}/inc/external | |||
${GE_CODE_DIR}/common/inc/external | |||
${GE_CODE_DIR}/common/inc/external/graph | |||
${GE_CODE_DIR}/inc | |||
${GE_CODE_DIR}/inc/framework | |||
${METADEF_DIR}/inc | |||
${METADEF_DIR}/inc/graph | |||
${METADEF_DIR}/inc/register | |||
${METADEF_DIR}/inc/external | |||
${METADEF_DIR}/inc/external/graph | |||
${METADEF_DIR}/inc/external/register | |||
${PARSER_DIR} | |||
${CMAKE_BINARY_DIR} | |||
${CMAKE_BINARY_DIR}/proto/ge | |||
#### yellow zone #### | |||
${GE_CODE_DIR}/../inc | |||
${GE_CODE_DIR}/../inc/common | |||
#### blue zone #### | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc/toolchain | |||
) | |||
target_link_libraries(atc.bin PRIVATE | |||
$<BUILD_INTERFACE:intf_pub> | |||
ascend_protobuf | |||
ge_common | |||
register | |||
c_sec | |||
graph | |||
error_manager | |||
ge_compiler | |||
parser_common | |||
gflags | |||
json | |||
runtime_compile | |||
slog | |||
static_mmpa | |||
-lrt | |||
-ldl | |||
) | |||
############ fwk_atc.bin ############ | |||
add_executable(fwk_atc.bin ${SRC_LIST} ${PROTO_HDRS}) | |||
target_compile_options(fwk_atc.bin PRIVATE | |||
-Werror | |||
-O2 | |||
-Wno-deprecated-declarations | |||
) | |||
target_compile_definitions(fwk_atc.bin PRIVATE | |||
PROTOBUF_INLINE_NOT_IN_HEADERS=0 | |||
COMPILE_OMG_PACKAGE | |||
google=ascend_private | |||
) | |||
target_include_directories(fwk_atc.bin PRIVATE | |||
${CMAKE_CURRENT_LIST_DIR} | |||
${GE_CODE_DIR} | |||
${GE_CODE_DIR}/ge | |||
${GE_CODE_DIR}/inc/external | |||
${GE_CODE_DIR}/common/inc/external | |||
${GE_CODE_DIR}/common/inc/external/graph | |||
${GE_CODE_DIR}/inc | |||
${GE_CODE_DIR}/inc/framework | |||
${METADEF_DIR}/inc | |||
${METADEF_DIR}/inc/graph | |||
${METADEF_DIR}/inc/register | |||
${METADEF_DIR}/inc/external | |||
${METADEF_DIR}/inc/external/graph | |||
${METADEF_DIR}/inc/external/register | |||
${PARSER_DIR} | |||
${CMAKE_BINARY_DIR} | |||
${CMAKE_BINARY_DIR}/proto/ge | |||
#### yellow zone #### | |||
${GE_CODE_DIR}/../inc | |||
${GE_CODE_DIR}/../inc/common | |||
#### blue zone #### | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc | |||
${GE_CODE_DIR}/third_party/fwkacllib/inc/toolchain | |||
) | |||
target_link_libraries(fwk_atc.bin PRIVATE | |||
$<BUILD_INTERFACE:intf_pub> | |||
ascend_protobuf | |||
ge_common | |||
register | |||
c_sec | |||
graph | |||
error_manager | |||
ge_runner | |||
parser_common | |||
gflags | |||
json | |||
runtime | |||
slog | |||
static_mmpa | |||
-lrt | |||
-ldl | |||
) | |||
############ install ############ | |||
set(INSTALL_BASE_DIR "") | |||
set(INSTALL_LIBRARY_DIR lib) | |||
install(TARGETS atc OPTIONAL | |||
install(TARGETS atc atc.bin fwk_atc.bin OPTIONAL | |||
LIBRARY DESTINATION ${INSTALL_LIBRARY_DIR} | |||
) |
@@ -0,0 +1,20 @@ | |||
#!/bin/bash | |||
#------------------------------------------------------------------- | |||
# Purpose: | |||
# Copyright 2020 Huawei Technologies Co., Ltd. All rights reserved. | |||
#------------------------------------------------------------------- | |||
LOCAL_PATH=$(cd "$(dirname "$0")"; pwd) | |||
PKG_PATH=$(cd ${LOCAL_PATH}/..; pwd) | |||
LIB_P="/lib64" | |||
PYTHON_P="/python/site-packages" | |||
LIB64_PATH="${PKG_PATH}${LIB_P}" | |||
PYTHON_PATH="${PKG_PATH}${PYTHON_P}" | |||
export LD_LIBRARY_PATH="${LIB64_PATH}:${LD_LIBRARY_PATH}" | |||
export PYTHONPATH="${PYTHON_PATH}:${PYTHONPATH}" | |||
if [ -f "${PKG_PATH}/bin/atc.bin" ];then | |||
atc.bin $@ | |||
else | |||
fwk_atc.bin $@ | |||
fi |
@@ -54,3 +54,108 @@ LOCAL_LDFLAGS := -lrt -ldl | |||
include $(BUILD_HOST_EXECUTABLE) | |||
include $(CLEAR_VARS) | |||
LOCAL_MODULE := atc.bin | |||
LOCAL_CFLAGS += -Werror -Wno-deprecated-declarations | |||
LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -DCOMPILE_OMG_PACKAGE -O2 -Dgoogle=ascend_private | |||
LOCAL_SRC_FILES := \ | |||
main.cc \ | |||
single_op_parser.cc \ | |||
../session/omg.cc \ | |||
../ir_build/atc_ir_common.cc \ | |||
LOCAL_C_INCLUDES := \ | |||
$(LOCAL_PATH)/../ ./ \ | |||
$(TOPDIR)inc \ | |||
$(TOPDIR)metadef/inc \ | |||
$(TOPDIR)graphengine/inc \ | |||
$(TOPDIR)inc/external \ | |||
$(TOPDIR)metadef/inc/external \ | |||
$(TOPDIR)graphengine/inc/external \ | |||
$(TOPDIR)metadef/inc/external/graph \ | |||
$(TOPDIR)graphengine/inc/framework \ | |||
$(TOPDIR)libc_sec/include \ | |||
$(TOPDIR)metadef/inc/common/util \ | |||
$(TOPDIR)parser \ | |||
third_party/json/include \ | |||
third_party/gflags/include \ | |||
third_party/protobuf/include \ | |||
proto/om.proto \ | |||
proto/ge_ir.proto \ | |||
proto/task.proto \ | |||
proto/insert_op.proto \ | |||
LOCAL_SHARED_LIBRARIES := \ | |||
libc_sec \ | |||
libge_common \ | |||
libascend_protobuf \ | |||
libslog \ | |||
libgraph \ | |||
libregister \ | |||
liberror_manager \ | |||
libge_compiler \ | |||
libruntime_compile \ | |||
libparser_common \ | |||
liberror_manager \ | |||
LOCAL_STATIC_LIBRARIES := libgflags | |||
LOCAL_LDFLAGS := -lrt -ldl | |||
include $(BUILD_HOST_EXECUTABLE) | |||
include $(CLEAR_VARS) | |||
LOCAL_MODULE := fwk_atc.bin | |||
LOCAL_CFLAGS += -Werror -Wno-deprecated-declarations | |||
LOCAL_CFLAGS += -DPROTOBUF_INLINE_NOT_IN_HEADERS=0 -DCOMPILE_OMG_PACKAGE -O2 -Dgoogle=ascend_private | |||
LOCAL_SRC_FILES := \ | |||
main.cc \ | |||
single_op_parser.cc \ | |||
../session/omg.cc \ | |||
../ir_build/atc_ir_common.cc \ | |||
LOCAL_C_INCLUDES := \ | |||
$(LOCAL_PATH)/../ ./ \ | |||
$(TOPDIR)inc \ | |||
$(TOPDIR)metadef/inc \ | |||
$(TOPDIR)graphengine/inc \ | |||
$(TOPDIR)inc/external \ | |||
$(TOPDIR)metadef/inc/external \ | |||
$(TOPDIR)graphengine/inc/external \ | |||
$(TOPDIR)metadef/inc/external/graph \ | |||
$(TOPDIR)graphengine/inc/framework \ | |||
$(TOPDIR)libc_sec/include \ | |||
$(TOPDIR)metadef/inc/common/util \ | |||
$(TOPDIR)parser \ | |||
third_party/json/include \ | |||
third_party/gflags/include \ | |||
third_party/protobuf/include \ | |||
proto/om.proto \ | |||
proto/ge_ir.proto \ | |||
proto/task.proto \ | |||
proto/insert_op.proto \ | |||
LOCAL_SHARED_LIBRARIES := \ | |||
libc_sec \ | |||
libge_common \ | |||
libascend_protobuf \ | |||
libslog \ | |||
libgraph \ | |||
libregister \ | |||
liberror_manager \ | |||
libge_runner \ | |||
libruntime \ | |||
libparser_common \ | |||
liberror_manager \ | |||
LOCAL_STATIC_LIBRARIES := libgflags | |||
LOCAL_LDFLAGS := -lrt -ldl | |||
include $(BUILD_HOST_EXECUTABLE) |
@@ -202,7 +202,7 @@ Status CsaInteract::WriteFile(const std::string &file_name, const std::string &c | |||
} | |||
} | |||
mmSsize_t ret = mmWrite(fd, (void *)content.c_str(), content.length()); | |||
mmSsize_t ret = mmWrite(fd, reinterpret_cast<void *>(const_cast<char *>(content.c_str())), content.length()); | |||
if (ret == EN_ERROR) { | |||
GELOGE(INTERNAL_ERROR, "write file fail, errno is %d", errno); | |||
ret = mmClose(fd); | |||
@@ -33,6 +33,8 @@ const std::vector<std::string> kHcclBuilderLibs = { | |||
"libhvd_opskernel_builder.so", | |||
"libhcom_gradtune_opskernel_builder.so" | |||
}; | |||
const std::string kAicoreUtilsLib = "libaicore_utils_runtime.so"; | |||
} // namespace | |||
OpsKernelBuilderManager::~OpsKernelBuilderManager() { | |||
// it's OK to call Finalize multiply times | |||
@@ -45,13 +47,11 @@ OpsKernelBuilderManager &OpsKernelBuilderManager::Instance() { | |||
} | |||
Status OpsKernelBuilderManager::Initialize(const map<std::string, std::string> &options, bool is_train) { | |||
if (is_train) { | |||
std::string lib_paths; | |||
GE_CHK_STATUS_RET_NOLOG(GetLibPaths(options, lib_paths)); | |||
plugin_manager_.reset(new (std::nothrow)PluginManager()); | |||
GE_CHECK_NOTNULL(plugin_manager_); | |||
GE_CHK_STATUS_RET(plugin_manager_->LoadSo(lib_paths), "Failed to load libs"); | |||
} | |||
std::string lib_paths; | |||
GE_CHK_STATUS_RET_NOLOG(GetLibPaths(options, lib_paths, is_train)); | |||
plugin_manager_.reset(new (std::nothrow)PluginManager()); | |||
GE_CHECK_NOTNULL(plugin_manager_); | |||
GE_CHK_STATUS_RET(plugin_manager_->LoadSo(lib_paths), "Failed to load libs"); | |||
auto &kernel_builders = OpsKernelBuilderRegistry::GetInstance().GetAll(); | |||
GELOGI("Number of OpBuild = %zu", kernel_builders.size()); | |||
@@ -100,7 +100,8 @@ OpsKernelBuilderPtr OpsKernelBuilderManager::GetOpsKernelBuilder(const string &n | |||
return nullptr; | |||
} | |||
Status OpsKernelBuilderManager::GetLibPaths(const std::map<std::string, std::string> &options, std::string &lib_paths) { | |||
Status OpsKernelBuilderManager::GetLibPaths(const std::map<std::string, std::string> &options, std::string &lib_paths, | |||
bool is_train) { | |||
GELOGD("Start to execute GetLibPaths"); | |||
std::string path_base = PluginManager::GetPath(); | |||
std::string so_path = "plugin/opskernel/"; | |||
@@ -109,6 +110,9 @@ Status OpsKernelBuilderManager::GetLibPaths(const std::map<std::string, std::str | |||
for (const auto &lib_name : kBasicBuilderLibs) { | |||
all_lib_paths += (path + lib_name + ":"); | |||
} | |||
if (!is_train) { | |||
all_lib_paths += (path_base + kAicoreUtilsLib + ":"); | |||
} | |||
auto iter = options.find(OPTION_EXEC_HCCL_FLAG); | |||
if (iter == options.end() || iter->second != "0") { | |||
@@ -48,7 +48,7 @@ class OpsKernelBuilderManager { | |||
private: | |||
OpsKernelBuilderManager() = default; | |||
static Status GetLibPaths(const std::map<std::string, std::string> &options, std::string &lib_paths); | |||
static Status GetLibPaths(const std::map<std::string, std::string> &options, std::string &lib_paths, bool is_train); | |||
std::unique_ptr<PluginManager> plugin_manager_; | |||
std::map<std::string, OpsKernelBuilderPtr> ops_kernel_builders_{}; | |||
@@ -891,7 +891,7 @@ FMK_FUNC_HOST_VISIBILITY Status ConvertOmModelToJson(const char *model_file, con | |||
if (status != ge::GRAPH_SUCCESS) { | |||
GELOGE(ge::FAILED, "Om file init failed."); | |||
if (model.model_data != nullptr) { | |||
delete[](char *) model.model_data; | |||
delete[] reinterpret_cast<char *>(model.model_data); | |||
model.model_data = nullptr; | |||
} | |||
return status; | |||
@@ -902,7 +902,7 @@ FMK_FUNC_HOST_VISIBILITY Status ConvertOmModelToJson(const char *model_file, con | |||
if (status != ge::GRAPH_SUCCESS) { | |||
GELOGE(ge::FAILED, "Get model part failed."); | |||
if (model.model_data != nullptr) { | |||
delete[](char *) model.model_data; | |||
delete[] reinterpret_cast<char *>(model.model_data); | |||
model.model_data = nullptr; | |||
} | |||
return status; | |||
@@ -928,7 +928,7 @@ FMK_FUNC_HOST_VISIBILITY Status ConvertOmModelToJson(const char *model_file, con | |||
} | |||
if (model.model_data != nullptr) { | |||
delete[](char *) model.model_data; | |||
delete[] reinterpret_cast<char *>(model.model_data); | |||
model.model_data = nullptr; | |||
} | |||
return ret; | |||
@@ -62,8 +62,12 @@ Status AiCpuCCTaskBuilder::BuildTask(AiCpuCCTask &task, uint64_t kernel_id) { | |||
if (kernel_type == ccKernelType::CUST_AI_CPU) { | |||
task.is_custom_ = true; | |||
task.dump_flag_ |= RT_KERNEL_CUSTOM_AICPU; | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc_, so_name), "launch cust aicpu so failed"); | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LaunchCustAicpuSo(), "launch cust aicpu so failed."); | |||
bool loaded = false; | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LoadCustAicpuSo(op_desc_, so_name, loaded), | |||
"launch cust aicpu so failed"); | |||
if (!loaded) { | |||
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LaunchCustAicpuSo(), "launch cust aicpu so failed."); | |||
} | |||
} | |||
task.num_inputs_ = op_desc_->GetInputsSize(); | |||
@@ -1,3 +1,10 @@ | |||
#!/usr/bin/python3.7 | |||
# -*- coding: UTF-8 -*- | |||
#------------------------------------------------------------------- | |||
# Purpose: | |||
# Copyright 2020 Huawei Technologies Co., Ltd. All rights reserved. | |||
#------------------------------------------------------------------- | |||
import os | |||
import re | |||
import sys | |||
@@ -19,8 +19,6 @@ | |||
#include "runtime/rt.h" | |||
using namespace std; | |||
namespace ge { | |||
#define CC_FUSION_OP_MAX 32 | |||
@@ -234,6 +234,22 @@ class GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY GeExecutor { | |||
ge::Status ExecModel(uint32_t model_id, void *stream, const ge::RunModelData &input_data, | |||
ge::RunModelData &output_data, bool async_mode = false); | |||
/// | |||
/// @ingroup ge | |||
/// @brief Synchronous execution of offline model(Do not create thread) | |||
/// @param [in] uint32_t model_id: Model ID to execute | |||
/// @param [in] void* stream: stream to execute | |||
/// @param [in] bool async_mode: is asynchronize mode. | |||
/// @param [in] const domi::InputData *input_data: Model input data | |||
/// @param [in] const std::vector<GeTensorDesc> &input_desc: description of model input data | |||
/// @param [out] domi::OutputData *output_data: Model output data | |||
/// @param [out] std::vector<GeTensorDesc> &output_desc: description of model output data | |||
/// @return SUCCESS handle successfully / others handle failed | |||
/// | |||
ge::Status ExecModel(uint32_t model_id, void *stream, const ge::RunModelData &run_input_data, | |||
const std::vector<GeTensorDesc> &input_desc, ge::RunModelData &run_output_data, | |||
std::vector<GeTensorDesc> &output_desc, bool async_mode = false); | |||
/// | |||
/// @ingroup ge | |||
/// @brief Get weight memory size from model file | |||
@@ -1 +1 @@ | |||
Subproject commit 29c31bb87d8bbe6904ab6fa72034a803fb50a746 | |||
Subproject commit 5b9a7f84a4347f8816d492aa51f2414ccf8a0744 |
@@ -1 +1 @@ | |||
Subproject commit ba956d349d8ad3e864d27467f4f0119333cbadc6 | |||
Subproject commit 70369668abebed84942d9f355494a89e82cc1eac |
@@ -1,42 +0,0 @@ | |||
# Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
# | |||
# Licensed under the Apache License, Version 2.0 (the "License"); | |||
# you may not use this file except in compliance with the License. | |||
# You may obtain a copy of the License at | |||
# | |||
# http://www.apache.org/licenses/LICENSE-2.0 | |||
# | |||
# Unless required by applicable law or agreed to in writing, software | |||
# distributed under the License is distributed on an "AS IS" BASIS, | |||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
# See the License for the specific language governing permissions and | |||
# limitations under the License. | |||
# ============================================================================ | |||
cmake_minimum_required(VERSION 3.0) | |||
set(CMAKE_CXX_STANDARD 11) | |||
project(ge_st CXX C) | |||
set(CMAKE_CXX_FLAGS "-O1 -fPIC -Wl,-unresolved-symbols=ignore-in-shared-libs") | |||
file(GLOB_RECURSE RES50_TRAIN_SRCS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} | |||
"resnet50/resnet50_train.cc" | |||
"resnet50/common.cc" | |||
) | |||
include_directories(${GE_SOURCE_DIR}/inc) | |||
include_directories(${GE_SOURCE_DIR}/inc/graph) | |||
include_directories(${GE_SOURCE_DIR}/inc/framework) | |||
include_directories(${GE_SOURCE_DIR}/inc/external) | |||
include_directories(${GE_SOURCE_DIR}/inc/external/ge) | |||
include_directories(${GE_SOURCE_DIR}/inc/external/graph) | |||
include_directories(${GE_SOURCE_DIR}/third_party/fwkacllib/inc) | |||
include_directories(${GE_SOURCE_DIR}/third_party/fwkacllib/inc/ops) | |||
include_directories(/usr/local/HiAI/opp/op_proto/built-in/inc) | |||
add_executable(st_resnet50_train ${RES50_TRAIN_SRCS}) | |||
target_link_libraries(st_resnet50_train | |||
${PROTOBUF_LIBRARY} | |||
ge_client_train ge_memory | |||
) |
@@ -1,768 +0,0 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <math.h> | |||
#include <stdint.h> | |||
#include <stdio.h> | |||
#include <stdlib.h> | |||
#include <iostream> | |||
#include <vector> | |||
#include "common.h" | |||
#include "model.h" | |||
#define MAX_HEAD_SIZE 50 | |||
using namespace std; | |||
using namespace ge; | |||
void update_op_format(Operator ops, Format format) { | |||
printf("set format begin.........\n"); | |||
ge::TensorDesc tensor_desc_x = ops.GetInputDesc("x"); | |||
ge::TensorDesc tensor_desc_y = ops.GetOutputDesc("y"); | |||
Format f_x0 = tensor_desc_x.GetFormat(); | |||
Format f_y0 = tensor_desc_x.GetFormat(); | |||
printf("before set x format:%d \n", f_x0); | |||
printf("before set y format:%d \n", f_y0); | |||
printf("format to be set is :%d \n", format); | |||
tensor_desc_x.SetFormat(format); | |||
tensor_desc_y.SetFormat(format); | |||
ops.UpdateInputDesc("x", tensor_desc_x); | |||
ops.UpdateOutputDesc("y", tensor_desc_y); | |||
Format f_x = tensor_desc_x.GetFormat(); | |||
Format f_y = tensor_desc_y.GetFormat(); | |||
printf("after set x format:%d \n", f_x); | |||
printf("after set y format:%d \n", f_y); | |||
} | |||
/// getDimInfo: get dim info from data file | |||
/// param: | |||
/// fp: the testing datafile object | |||
/// | |||
/// return : | |||
/// dim_info: array to store the info of the dim in datafile, like [4,3,3,6,3,162(3*3*6*3)],4 is dim size,3,3,6,3 is the | |||
/// dim shape data_size: the size of the testing data including the data file | |||
void getDimInfo(FILE *fp, std::vector<uint64_t> &dim_info) { | |||
// get dim info from hisi testing data file | |||
uint32_t *dim_buffer = (uint32_t *)malloc(MAX_HEAD_SIZE * sizeof(uint32_t)); | |||
fread(dim_buffer, sizeof(uint32_t), MAX_HEAD_SIZE, fp); | |||
dim_info.push_back(*dim_buffer); // get dim size | |||
// get data shape to compute the datasize | |||
uint64_t data_size = 1; | |||
uint32_t i = 1; | |||
for (; i <= dim_info[0]; i++) { | |||
dim_info.push_back(*(dim_buffer + i)); | |||
data_size *= *(dim_buffer + i); | |||
} | |||
dim_info.push_back(data_size); | |||
free(dim_buffer); | |||
} | |||
/// readTestDataFile: read test date from hisi .t datafile | |||
/// param: | |||
/// infile: the path of hisi .t datafile | |||
/// return: | |||
/// dim_info: array to store the info of the dim in datafile, like [4,3,3,6,3],4 is dim size,3,3,6,3 is the dim shape | |||
void *readTestDataFile(std::string infile, std::vector<uint64_t> &dim_info) { | |||
FILE *fp; | |||
fp = fopen(infile.c_str(), "r"); | |||
if (fp == NULL) { | |||
printf("ERROR: cant't open file %s\n", infile.c_str()); | |||
return NULL; | |||
} else { | |||
getDimInfo(fp, dim_info); | |||
uint64_t data_size = dim_info[dim_info.size() - 1]; | |||
fclose(fp); | |||
fp = fopen(infile.c_str(), "r"); | |||
if (fp == NULL) { | |||
printf("ERROR: cant't open file %s\n", infile.c_str()); | |||
return NULL; | |||
} | |||
uint32_t *memory = (uint32_t *)malloc((dim_info[0] + 1 + data_size) * sizeof(uint32_t)); | |||
fread(memory, sizeof(uint32_t), (dim_info[0] + 1 + data_size), fp); | |||
fclose(fp); | |||
return memory + (dim_info[0] + 1); | |||
} | |||
} | |||
void *readUint8TestDataFile(std::string infile, int size) { | |||
FILE *fp; | |||
fp = fopen(infile.c_str(), "r"); | |||
if (fp == NULL) { | |||
printf("ERROR: cant't open file %s\n", infile.c_str()); | |||
return NULL; | |||
} | |||
uint8_t *memory = (uint8_t *)malloc((size) * sizeof(uint8_t)); | |||
fread(memory, sizeof(uint8_t), (size), fp); | |||
fclose(fp); | |||
return memory; | |||
} | |||
/// allclose | |||
/// param: | |||
/// a:compared file a | |||
/// b:compared file b | |||
/// count: the count size which will compare | |||
/// rtol: | |||
/// atol: | |||
/// return: | |||
/// true or false | |||
bool allclose(float *a, float *b, uint64_t count, float rtol = 1e-05, float atol = 1e-08) { | |||
uint32_t i = 0; | |||
for (; i < count; ++i) { | |||
if (fabs(a[i] - b[i]) > (atol + rtol * fabs(b[i]))) { | |||
printf("compara failed: i= %d, a[i]=%f, b[i]=%f,atol=%f,rtol=%f\n", i, a[i], b[i], atol, rtol); | |||
return false; | |||
} | |||
} | |||
return true; | |||
} | |||
/// compFp32WithTData: compare the data with the data in hisi .t file | |||
/// param: | |||
/// actual_output_data: the result of ge | |||
/// expected_data_file: the path of hisi .t result file | |||
/// rtol: | |||
/// atol: | |||
/// return: | |||
/// true of false | |||
bool compFp32WithTData(float *actual_output_data, std::string expected_data_file, float rtol = 1e-05, float atol = 1e-08) { | |||
std::vector<uint64_t> dim_info; | |||
float *expected_output_data = (float *)readTestDataFile(expected_data_file, dim_info); | |||
uint32_t i = 1; | |||
uint64_t data_size = 1; | |||
for (; i <= dim_info[0]; i++) { | |||
data_size *= dim_info[i]; | |||
} | |||
return allclose(actual_output_data, expected_output_data, data_size, rtol, atol); | |||
} | |||
int SwitchDatatype(DataType dt) { | |||
int size = 1; | |||
if (dt == ge::DT_FLOAT) size = 4; | |||
if (dt == ge::DT_INT32) size = 4; | |||
if (dt == ge::DT_FLOAT16) size = 2; | |||
if (dt == ge::DT_INT64) size = 8; | |||
return size; | |||
} | |||
ge::Tensor genTensor(std::vector<int64_t> tensor_shape, Format format, DataType dt) { | |||
int size = 1; | |||
for (int i = 0; i < tensor_shape.size(); i++) { | |||
size = size * tensor_shape[i]; | |||
} | |||
int data_type_size = SwitchDatatype(dt); | |||
size = abs(size * data_type_size); | |||
vector<uint8_t> data_value; | |||
if (size == 0) { | |||
TensorDesc input_tensor_desc = TensorDesc(ge::Shape(tensor_shape), format, dt); | |||
input_tensor_desc.SetRealDimCnt(tensor_shape.size()); | |||
Tensor gen_tensor = Tensor(input_tensor_desc, data_value); | |||
return gen_tensor; | |||
} | |||
for (int i = 0; i < size; i++) { | |||
data_value.push_back(1); | |||
} | |||
TensorDesc input_tensor_desc = TensorDesc(ge::Shape(tensor_shape), format, dt); | |||
input_tensor_desc.SetRealDimCnt(tensor_shape.size()); | |||
Tensor gen_tensor = Tensor(input_tensor_desc, data_value); | |||
return gen_tensor; | |||
} | |||
ge::Tensor genTensor_withVaule(std::vector<int64_t> tensor_shape, float value) { | |||
int size = 1; | |||
for (int i = 0; i < tensor_shape.size(); i++) { | |||
size = size * tensor_shape[i]; | |||
} | |||
float *data_value = new float[size]; | |||
for (int i = 0; i < size; i++) { | |||
*(data_value + i) = value; | |||
} | |||
Tensor gen_ge_tensor; | |||
TensorDesc input_tensor_desc = TensorDesc(ge::Shape(tensor_shape), FORMAT_NCHW); | |||
gen_ge_tensor.SetTensorDesc(input_tensor_desc); | |||
gen_ge_tensor.SetData((uint8_t *)data_value, size * 4); | |||
return gen_ge_tensor; | |||
} | |||
Tensor genTesnor_Shape_as_data(std::vector<int64_t> tensor_shape) { | |||
Format format = FORMAT_NCHW; | |||
DataType dt = DT_INT32; | |||
int size = tensor_shape.size(); | |||
int32_t *tensor_data = new int32_t[size]; | |||
std::cout << "shape tensor size:" << size << endl; | |||
for (int i = 0; i < size; i++) { | |||
*(tensor_data + i) = tensor_shape[i]; | |||
} | |||
Tensor gen_tensor; | |||
TensorDesc input_tensor_desc = TensorDesc(ge::Shape({size}), FORMAT_NCHW, DT_INT32); | |||
gen_tensor.SetData((uint8_t *)tensor_data, size * GetDatTypeSize(dt)); | |||
gen_tensor.SetTensorDesc(input_tensor_desc); | |||
return gen_tensor; | |||
} | |||
/// train_flag is 0 when infer; train_flag is 1 when train; train_flag is 0 default | |||
/// run_mode_path is not 0,1,2 when TBE; run_mode_path is 1 when FE; run_mode_path is 0 default | |||
/// run_mode_path is 2 now when AICPU, ge.enabledlocalFmkop is 1 | |||
ge::Status GEInitialize_api(string train_flag, string run_mode_path) { | |||
ge::Status ret; | |||
if (run_mode_path == "0") { | |||
const std::map<string, string> config = { | |||
{"device_id", "0,2,4,6"}, | |||
{"rank_table_file", "hccl from csa/paas"}, | |||
{"ge.graphRunMode", train_flag}, | |||
{"ge.aicpuFlag", "1"}, | |||
{"ge.feFlag", "1"}, | |||
{DDK_VERSION_FLAG, "1.60.T17.B830"}, | |||
{"ge.soLoadPath", | |||
"/usr/local/HiAI/runtime/lib64/plugin/opskernel/libfe.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/" | |||
"libaicpu_plugin.so"}}; | |||
ret = ge::GEInitialize(config); | |||
} else if (run_mode_path == "1") { | |||
const std::map<string, string> config = { | |||
{"device_id", "0,2,4,6"}, | |||
{"rank_table_file", "hccl from csa/paas"}, | |||
{"ge.graphRunMode", train_flag}, | |||
{"ge.feFlag", "1"}, | |||
{DDK_VERSION_FLAG, "1.60.T17.B830"}, | |||
{TBE_PLUGIN_PATH_FLAG, "/usr/local/HiAI/runtime/lib64/tbe_plugin/bert"}, | |||
{"ge.soLoadPath", "/usr/local/HiAI/runtime/lib64/plugin/opskernel/libfe.so"}}; | |||
ret = ge::GEInitialize(config); | |||
} else if (run_mode_path == "2") { | |||
const std::map<string, string> config = {{"device_id", "0,2,4,6"}, | |||
{"rank_table_file", "hccl from csa/paas"}, | |||
{"ge.graphRunMode", train_flag}, | |||
{LOCAL_FMKOP_FLAG, "1"}}; | |||
ret = ge::GEInitialize(config); | |||
} else { | |||
const std::map<string, string> config = { | |||
{"device_id", "0,2,4,6"}, | |||
{"rank_table_file", "hccl from csa/paas"}, | |||
{"ge.graphRunMode", train_flag}, | |||
{DDK_VERSION_FLAG, "1.60.T17.B830"}, | |||
{TBE_PLUGIN_PATH_FLAG, "/usr/local/HiAI/runtime/lib64/tbe_plugin/" + run_mode_path}}; | |||
ret = ge::GEInitialize(config); | |||
} | |||
std::cout << "GEInitialize_ret is " << ret << std::endl; | |||
return ret; | |||
} | |||
/// train_flag is infer default | |||
/// run_mode: is multi group of [fe,aicpu,bert,deeplabv3,mobilenetv2,single_path_nas,ssd] | |||
/// but bert,deeplabv3,mobilenetv2,single_path_nas,ssd can only set one value from array | |||
/// eg:"fe,aicpu,bert" or "fe", default is “fe” | |||
/// "fe,aicpu,bert" remain open fe aicpu and bert | |||
ge::Status GEInitialize_api_new(string train_flag, string run_mode) { | |||
ge::Status ret; | |||
vector<string> modes; | |||
char *strs = new char[run_mode.length() + 1]; | |||
strcpy(strs, run_mode.c_str()); | |||
const char *delim = ","; | |||
char *p = strtok(strs, delim); | |||
while (p) { | |||
string s = p; // transform substr to string | |||
modes.push_back(s); // save to result array | |||
p = strtok(NULL, delim); | |||
} | |||
std::map<string, string> config = { | |||
{"device_id", "0,2,4,6"}, | |||
{"rank_table_file", "hccl from csa/paas"}, | |||
{DDK_VERSION_FLAG, "1.60.T17.B830"}, | |||
{"ge.opsProtoLibPath", "/usr/local/HiAI/runtime/ops/op_proto/built-in/libopsproto.so"}}; | |||
if (train_flag == "infer") | |||
config.insert(pair<string, string>("ge.graphRunMode", "0")); | |||
else if (train_flag == "train") | |||
config.insert(pair<string, string>("ge.graphRunMode", "1")); | |||
else | |||
std::cout << "GeInitialize give the error param" << std::endl; | |||
for (int i = 0; i < modes.size(); i++) { | |||
if (modes[i] == "fe") { | |||
config.insert(pair<string, string>("ge.feFlag", "1")); | |||
if (config.find("ge.soLoadPath") != config.end()) { | |||
config["ge.soLoadPath"] = | |||
"/usr/local/HiAI/runtime/lib64/plugin/opskernel/libfe.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/" | |||
"libaicpu_plugin.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/libge_local_engine.so:/usr/local/HiAI/" | |||
"runtime/lib64/plugin/opskernel/librts_engine.so"; | |||
} else { | |||
config.insert(pair<string, string>( | |||
"ge.soLoadPath", | |||
"/usr/local/HiAI/runtime/lib64/plugin/opskernel/libfe.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/" | |||
"libge_local_engine.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/librts_engine.so")); | |||
} | |||
} else if (modes[i] == "aicpu") { | |||
config.insert(pair<string, string>("ge.aicpuFlag", "1")); | |||
if (config.find("ge.soLoadPath") != config.end()) { | |||
config["ge.soLoadPath"] = | |||
"/usr/local/HiAI/runtime/lib64/plugin/opskernel/libfe.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/" | |||
"libaicpu_plugin.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/libge_local_engine.so:/usr/local/HiAI/" | |||
"runtime/lib64/plugin/opskernel/librts_engine.so"; | |||
} else { | |||
config.insert(pair<string, string>( | |||
"ge.soLoadPath", | |||
"/usr/local/HiAI/runtime/lib64/plugin/opskernel/libaicpu_plugin.so:/usr/local/HiAI/runtime/lib64/plugin/" | |||
"opskernel/libge_local_engine.so:/usr/local/HiAI/runtime/lib64/plugin/opskernel/librts_engine.so")); | |||
} | |||
} else if (modes[i] == "bert" || modes[i] == "deeplabv3" || modes[i] == "mobilenetv2" || | |||
modes[i] == "single_path_nas" || modes[i] == "ssd") { | |||
config.insert(pair<string, string>(TBE_PLUGIN_PATH_FLAG, "/usr/local/HiAI/runtime/lib64/tbe_plugin/" + modes[i])); | |||
} else if (modes[i] == "plugin") { | |||
} else | |||
std::cout << "GeInitialize give the error param" << std::endl; | |||
} | |||
ret = ge::GEInitialize(config); | |||
std::cout << "GEInitialize_ret is " << ret << std::endl; | |||
return ret; | |||
} | |||
ge::Status GEFinalize_api() { | |||
ge::Status ret = ge::GEFinalize(); | |||
std::cout << "GEFinalize ret is " << ret << std::endl; | |||
return ret; | |||
} | |||
/// set train_flag | |||
/// if run_mode_path is "fe" remain FE process; "fe,plugin" is FE and TBE plugin process | |||
/// "aicpu" is open aicpu plugin | |||
int RunGraph_initData(Graph &graph, string op_name, map<string, std::vector<int64_t>> attr_test, string train_flag, | |||
string run_mode_path) { | |||
std::map<string, string> options = {{RUN_FLAG, "1"}}; | |||
uint32_t graph_id = 0; | |||
ge::Status ret = GEInitialize_api_new(train_flag, run_mode_path); | |||
EXPECT_EQ(ret, ge::SUCCESS); | |||
ge::Session *session = new Session(options); | |||
ASSERT_TRUE(session != NULL); | |||
std::vector<Tensor> input; | |||
if (attr_test.find("input1") != attr_test.end()) { | |||
Tensor input_tensor = genTensor(attr_test["input1"]); | |||
input.push_back(input_tensor); | |||
} | |||
if (attr_test.find("input2") != attr_test.end()) { | |||
Tensor input_tensor = genTensor(attr_test["input2"]); | |||
input.push_back(input_tensor); | |||
} | |||
if (attr_test.find("input3") != attr_test.end()) { | |||
Tensor input_tensor = genTensor(attr_test["input3"]); | |||
input.push_back(input_tensor); | |||
} | |||
std::vector<Tensor> output; | |||
ret = session->AddGraph(graph_id, graph); | |||
EXPECT_EQ(ret, ge::SUCCESS); | |||
if (train_flag == "1") { | |||
setenv("GE_TRAIN", "1", true); | |||
ret = session->RunGraph(graph_id, input, output); | |||
setenv("GE_TRAIN", "0", true); | |||
} else { | |||
ret = session->RunGraph(graph_id, input, output); | |||
} | |||
delete session; | |||
GEFinalize_api(); | |||
if (ret != ge::SUCCESS) { | |||
std::cout << " run graph failed" << std::endl; | |||
return -1; | |||
} else { | |||
return 0; | |||
} | |||
} | |||
ge::Status session_add_and_run_graph(ge::Session *session, uint32_t graph_id, Graph &graph, std::vector<Tensor> inputs, | |||
std::vector<Tensor> &outputs) { | |||
ge::Status ret = session->AddGraph(graph_id, graph); | |||
EXPECT_EQ(ret, ge::SUCCESS); | |||
ret = session->RunGraph(graph_id, inputs, outputs); | |||
return ret; | |||
} | |||
ge::Session *create_session() { | |||
// Init session | |||
std::map<string, string> options = {{"a", "b"}, {TRAIN_FLAG, "1"}}; | |||
ge::Session *session = new Session(options); | |||
ASSERT_TRUE(session != NULL); | |||
return session; | |||
} | |||
ge::Session *create_aipp_session() { | |||
// Init session | |||
std::map<string, string> options = {{"a", "b"}, {TRAIN_FLAG, "1"}, {"ge.insertOpFile", "/root/host/ge/aipp.cfg"}}; | |||
ge::Session *session = new Session(options); | |||
ASSERT_TRUE(session != NULL); | |||
return session; | |||
} | |||
int buildCheckPointGraph(Graph &graph, map<string, TensorDesc> variables) { | |||
std::vector<Operator> inputs{}; | |||
std::vector<Operator> outputs{}; | |||
for (map<string, TensorDesc>::iterator it = variables.begin(); it != variables.end(); ++it) { | |||
auto var = op::Variable(string(it->first)); | |||
var.update_output_desc_y(it->second); | |||
inputs.push_back(var); | |||
graph.AddOp(var); | |||
} | |||
auto save = op::Save().create_dynamic_input_tensors(inputs.size()); | |||
for (int i = 0; i < inputs.size(); i++) { | |||
save.set_dynamic_input_tensors(i, inputs[i]); | |||
} | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return 0; | |||
} | |||
int buildInitGraph(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var, | |||
std::vector<float> values_var) { | |||
std::vector<Operator> inputs{}; | |||
std::vector<Operator> outputs{}; | |||
for (int i = 0; i < desc_var.size(); i++) { | |||
desc_var[i].SetRealDimCnt(desc_var[i].GetShape().GetDimNum()); | |||
auto tensor_data = genTensor_withVaule(desc_var[i].GetShape().GetDims(), values_var[i]); | |||
auto var_constant = op::Constant().set_attr_value(tensor_data); | |||
var_constant.update_output_desc_y(desc_var[i]); | |||
auto var_init = op::Variable(string(name_var[i])); | |||
var_init.update_output_desc_y(desc_var[i]); | |||
auto var_assign = op::Assign().set_input_ref(var_init).set_input_value(var_constant); | |||
inputs.push_back(var_init); | |||
} | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return 0; | |||
} | |||
int buildInitGraph_other_dataType(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var) { | |||
std::vector<Operator> inputs{}; | |||
std::vector<Operator> outputs{}; | |||
for (int i = 0; i < desc_var.size(); i++) { | |||
desc_var[i].SetRealDimCnt(desc_var[i].GetShape().GetDimNum()); | |||
auto tensor_data = genTensor(desc_var[i].GetShape().GetDims(), desc_var[i].GetFormat(), desc_var[i].GetDataType()); | |||
auto var_constant = op::Constant().set_attr_value(tensor_data); | |||
var_constant.update_output_desc_y(desc_var[i]); | |||
auto var_init = op::Variable(string(name_var[i])); | |||
var_init.update_output_desc_y(desc_var[i]); | |||
auto var_assign = op::Assign().set_input_ref(var_init).set_input_value(var_constant); | |||
inputs.push_back(var_init); | |||
graph.AddOp(var_constant); | |||
graph.AddOp(var_init); | |||
graph.AddOp(var_assign); | |||
} | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return 0; | |||
} | |||
bool build_multi_input_multi_output_graph(Graph &graph) { | |||
auto data1 = op::Data("Data1").set_attr_index(0); | |||
auto data2 = op::Data("Data2").set_attr_index(1); | |||
vector<uint64_t> dim_info; | |||
auto relu1 = op::Relu("Relu1").set_input_x(data1); | |||
auto relu2 = op::Relu("Relu2").set_input_x(data2); | |||
auto eltwise = op::Eltwise("Eltwise") | |||
.create_dynamic_input_x(2) | |||
.set_dynamic_input_x(0, relu1) | |||
.set_dynamic_input_x(1, relu2) | |||
.set_attr_N(2) | |||
.set_attr_mode(1) | |||
.set_attr_coeff({1, 1}); | |||
auto eltwise1 = op::Eltwise("Eltwise1") | |||
.create_dynamic_input_x(2) | |||
.set_dynamic_input_x(0, eltwise) | |||
.set_dynamic_input_x(1, eltwise) | |||
.set_attr_N(2) | |||
.set_attr_mode(1) | |||
.set_attr_coeff({1, 1}); | |||
auto eltwise2 = op::Eltwise("Eltwise2") | |||
.create_dynamic_input_x(2) | |||
.set_dynamic_input_x(0, eltwise) | |||
.set_dynamic_input_x(1, eltwise) | |||
.set_attr_N(2) | |||
.set_attr_mode(1) | |||
.set_attr_coeff({1, 1}); | |||
std::vector<Operator> inputs{data1, data2}; | |||
std::vector<Operator> outputs{eltwise1, eltwise2}; | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return true; | |||
} | |||
void build_big_graph(Graph &graph, map<string, std::vector<int64_t>> attr) { | |||
auto data = op::Data("Data").set_attr_index(0); | |||
auto weight = op::Const("weight1").set_attr_value(genTensor(attr["weight"])); | |||
vector<int64_t> weight_shape(attr["weight"].begin(), attr["weight"].end()); | |||
TensorDesc weight_desc(ge::Shape(weight_shape), FORMAT_NCHW, DT_FLOAT); | |||
weight.update_output_desc_y(weight_desc); | |||
auto conv_1 = op::Conv2D("conv1").set_input_x(data).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_2 = op::Conv2D("conv2").set_input_x(conv_1).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_3 = op::Conv2D("conv3").set_input_x(conv_2).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_4 = op::Conv2D("conv4").set_input_x(conv_3).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_5 = op::Conv2D("conv5").set_input_x(conv_4).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_6 = op::Conv2D("conv6").set_input_x(conv_5).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_7 = op::Conv2D("conv7").set_input_x(conv_6).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_8 = op::Conv2D("conv8").set_input_x(conv_7).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_9 = op::Conv2D("conv9").set_input_x(conv_8).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_10 = op::Conv2D("conv10").set_input_x(conv_9).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_11 = op::Conv2D("conv11").set_input_x(conv_10).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_12 = op::Conv2D("conv12").set_input_x(conv_11).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_13 = op::Conv2D("conv13").set_input_x(conv_12).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_14 = op::Conv2D("conv14").set_input_x(conv_13).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_15 = op::Conv2D("conv15").set_input_x(conv_14).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_16 = op::Conv2D("conv16").set_input_x(conv_15).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_17 = op::Conv2D("conv17").set_input_x(conv_16).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_18 = op::Conv2D("conv18").set_input_x(conv_17).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_19 = op::Conv2D("conv19").set_input_x(conv_18).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_20 = op::Conv2D("conv20").set_input_x(conv_19).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_21 = op::Conv2D("conv21").set_input_x(conv_20).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_22 = op::Conv2D("conv22").set_input_x(conv_21).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_23 = op::Conv2D("conv23").set_input_x(conv_22).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_24 = op::Conv2D("conv24").set_input_x(conv_23).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_25 = op::Conv2D("conv25").set_input_x(conv_24).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_26 = op::Conv2D("conv26").set_input_x(conv_25).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_27 = op::Conv2D("conv27").set_input_x(conv_26).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_28 = op::Conv2D("conv28").set_input_x(conv_27).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_29 = op::Conv2D("conv29").set_input_x(conv_28).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_30 = op::Conv2D("conv30").set_input_x(conv_29).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_31 = op::Conv2D("conv31").set_input_x(conv_30).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_32 = op::Conv2D("conv32").set_input_x(conv_31).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_33 = op::Conv2D("conv33").set_input_x(conv_32).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_34 = op::Conv2D("conv34").set_input_x(conv_33).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_35 = op::Conv2D("conv35").set_input_x(conv_34).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_36 = op::Conv2D("conv36").set_input_x(conv_35).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_37 = op::Conv2D("conv37").set_input_x(conv_36).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_38 = op::Conv2D("conv38").set_input_x(conv_37).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_39 = op::Conv2D("conv39").set_input_x(conv_38).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_40 = op::Conv2D("conv40").set_input_x(conv_39).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_41 = op::Conv2D("conv41").set_input_x(conv_40).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_42 = op::Conv2D("conv42").set_input_x(conv_41).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_43 = op::Conv2D("conv43").set_input_x(conv_42).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_44 = op::Conv2D("conv44").set_input_x(conv_43).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_45 = op::Conv2D("conv45").set_input_x(conv_44).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_46 = op::Conv2D("conv46").set_input_x(conv_45).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_47 = op::Conv2D("conv47").set_input_x(conv_46).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_48 = op::Conv2D("conv48").set_input_x(conv_47).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_49 = op::Conv2D("conv49").set_input_x(conv_48).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_50 = op::Conv2D("conv50").set_input_x(conv_49).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_51 = op::Conv2D("conv51").set_input_x(conv_50).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_52 = op::Conv2D("conv52").set_input_x(conv_51).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_53 = op::Conv2D("conv53").set_input_x(conv_52).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_54 = op::Conv2D("conv54").set_input_x(conv_53).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_55 = op::Conv2D("conv55").set_input_x(conv_54).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_56 = op::Conv2D("conv56").set_input_x(conv_55).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_57 = op::Conv2D("conv57").set_input_x(conv_56).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_58 = op::Conv2D("conv58").set_input_x(conv_57).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_59 = op::Conv2D("conv59").set_input_x(conv_58).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_60 = op::Conv2D("conv60").set_input_x(conv_59).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_61 = op::Conv2D("conv61").set_input_x(conv_60).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_62 = op::Conv2D("conv62").set_input_x(conv_61).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_63 = op::Conv2D("conv63").set_input_x(conv_62).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_64 = op::Conv2D("conv64").set_input_x(conv_63).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_65 = op::Conv2D("conv65").set_input_x(conv_64).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_66 = op::Conv2D("conv66").set_input_x(conv_65).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_67 = op::Conv2D("conv67").set_input_x(conv_66).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_68 = op::Conv2D("conv68").set_input_x(conv_67).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_69 = op::Conv2D("conv69").set_input_x(conv_68).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_70 = op::Conv2D("conv70").set_input_x(conv_69).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_71 = op::Conv2D("conv71").set_input_x(conv_70).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_72 = op::Conv2D("conv72").set_input_x(conv_71).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_73 = op::Conv2D("conv73").set_input_x(conv_72).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_74 = op::Conv2D("conv74").set_input_x(conv_73).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_75 = op::Conv2D("conv75").set_input_x(conv_74).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_76 = op::Conv2D("conv76").set_input_x(conv_75).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_77 = op::Conv2D("conv77").set_input_x(conv_76).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_78 = op::Conv2D("conv78").set_input_x(conv_77).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_79 = op::Conv2D("conv79").set_input_x(conv_78).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_80 = op::Conv2D("conv80").set_input_x(conv_79).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_81 = op::Conv2D("conv81").set_input_x(conv_80).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_82 = op::Conv2D("conv82").set_input_x(conv_81).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_83 = op::Conv2D("conv83").set_input_x(conv_82).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_84 = op::Conv2D("conv84").set_input_x(conv_83).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_85 = op::Conv2D("conv85").set_input_x(conv_84).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_86 = op::Conv2D("conv86").set_input_x(conv_85).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_87 = op::Conv2D("conv87").set_input_x(conv_86).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_88 = op::Conv2D("conv88").set_input_x(conv_87).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_89 = op::Conv2D("conv89").set_input_x(conv_88).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_90 = op::Conv2D("conv90").set_input_x(conv_89).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_91 = op::Conv2D("conv91").set_input_x(conv_80).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_92 = op::Conv2D("conv92").set_input_x(conv_91).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_93 = op::Conv2D("conv93").set_input_x(conv_92).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_94 = op::Conv2D("conv94").set_input_x(conv_93).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_95 = op::Conv2D("conv95").set_input_x(conv_94).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_96 = op::Conv2D("conv96").set_input_x(conv_95).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_97 = op::Conv2D("conv97").set_input_x(conv_96).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_98 = op::Conv2D("conv98").set_input_x(conv_97).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_99 = op::Conv2D("conv99").set_input_x(conv_98).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_100 = op::Conv2D("conv100").set_input_x(conv_99).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_101 = op::Conv2D("conv101").set_input_x(conv_100).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_102 = op::Conv2D("conv102").set_input_x(conv_101).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_103 = op::Conv2D("conv103").set_input_x(conv_102).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_104 = op::Conv2D("conv104").set_input_x(conv_103).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_105 = op::Conv2D("conv105").set_input_x(conv_104).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_106 = op::Conv2D("conv106").set_input_x(conv_105).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_107 = op::Conv2D("conv107").set_input_x(conv_106).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_108 = op::Conv2D("conv108").set_input_x(conv_107).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_109 = op::Conv2D("conv109").set_input_x(conv_108).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_110 = op::Conv2D("conv110").set_input_x(conv_109).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_111 = op::Conv2D("conv111").set_input_x(conv_110).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_112 = op::Conv2D("conv112").set_input_x(conv_111).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_113 = op::Conv2D("conv113").set_input_x(conv_112).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_114 = op::Conv2D("conv114").set_input_x(conv_113).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_115 = op::Conv2D("conv115").set_input_x(conv_114).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_116 = op::Conv2D("conv116").set_input_x(conv_115).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_117 = op::Conv2D("conv117").set_input_x(conv_116).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_118 = op::Conv2D("conv118").set_input_x(conv_117).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_119 = op::Conv2D("conv119").set_input_x(conv_118).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_120 = op::Conv2D("conv120").set_input_x(conv_119).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_121 = op::Conv2D("conv121").set_input_x(conv_120).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_122 = op::Conv2D("conv122").set_input_x(conv_121).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_123 = op::Conv2D("conv123").set_input_x(conv_122).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_124 = op::Conv2D("conv124").set_input_x(conv_123).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_125 = op::Conv2D("conv125").set_input_x(conv_124).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_126 = op::Conv2D("conv126").set_input_x(conv_125).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_127 = op::Conv2D("conv127").set_input_x(conv_126).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_128 = op::Conv2D("conv128").set_input_x(conv_127).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_129 = op::Conv2D("conv129").set_input_x(conv_128).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
auto conv_130 = op::Conv2D("conv130").set_input_x(conv_129).set_input_filter(weight).set_attr_pads({0,0,0,0}).set_attr_strides({1,1,1,1}); | |||
std::vector<Operator> inputs{data}; | |||
std::vector<Operator> outputs{conv_130}; | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
} | |||
int GetDatTypeSize(DataType dt) { | |||
int dailation = 1; | |||
if (dt == ge::DT_FLOAT) | |||
dailation = 4; | |||
else if (dt == ge::DT_FLOAT16) | |||
dailation = 2; | |||
else if (dt == ge::DT_INT16) | |||
dailation = 2; | |||
else if (dt == ge::DT_UINT16) | |||
dailation = 2; | |||
else if (dt == ge::DT_INT32) | |||
dailation = 4; | |||
else if (dt == ge::DT_UINT32) | |||
dailation = 4; | |||
else if (dt == ge::DT_INT64) | |||
dailation = 8; | |||
else if (dt == ge::DT_UINT64) | |||
dailation = 8; | |||
else if (dt == ge::DT_INT8) | |||
dailation = 1; | |||
return dailation; | |||
} | |||
int buildConvGraph_new(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var, int flag, | |||
Format format) { | |||
auto data_x_shape = op::Data("xShape").set_attr_index(0); | |||
auto var = op::Variable(name_var[0]); | |||
auto var1 = op::Variable(name_var[1]); //add one seat of ApplyMomentum() | |||
auto label1 = op::Variable(name_var[2]); //add one seat of ApplyMomentum() | |||
auto conv2dgrad = op::Conv2DBackpropFilterD("output_1"); | |||
auto test2 = op::ApplyMomentum(); | |||
var.update_output_desc_y(desc_var[0]); | |||
var1.update_output_desc_y(desc_var[1]); | |||
label1.update_output_desc_y(desc_var[2]); | |||
graph.AddOp(var); | |||
graph.AddOp(var1); | |||
graph.AddOp(label1); | |||
auto conv2d = op::Conv2D().set_input_x(data_x_shape).set_input_filter(var).set_attr_strides({1, 1, 1, 1}).set_attr_pads({0,0,0,0}); | |||
update_op_format(conv2d, format); | |||
ge::TensorDesc tensor_desc_w = conv2d.GetInputDesc("filter"); | |||
tensor_desc_w.SetFormat(format); | |||
conv2d.UpdateInputDesc("filter", tensor_desc_w); | |||
if (flag >= 1) { | |||
conv2dgrad.set_input_x(data_x_shape) | |||
.set_attr_filter_size(desc_var[0].GetShape().GetDims()) | |||
.set_input_out_backprop(conv2d) | |||
.set_attr_strides({1, 1, 1, 1}) | |||
.set_attr_pads({0, 0, 0, 0}); | |||
update_op_format(conv2dgrad, format); | |||
graph.AddOp(conv2dgrad); | |||
} | |||
if (flag >= 2) { | |||
// set conv2dgrad var | |||
test2.set_input_accum(var1) | |||
.set_input_grad(conv2dgrad) | |||
.set_input_lr(label1) | |||
.set_input_momentum(label1) | |||
.set_input_var(var); | |||
graph.AddOp(test2); | |||
} | |||
std::vector<Operator> inputs{data_x_shape}; // set all val | |||
std::vector<Operator> outputs{conv2d}; | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
graph.AddOp(conv2d); | |||
return 0; | |||
} | |||
/// load bin data_fail | |||
/// input_path: path of bin data_file | |||
/// shapes: the shape of Tensor | |||
/// ft: the format of Tensor | |||
/// dt: the dataType of Tensor | |||
Tensor load_variable_input_data(string input_path, std::vector<int64_t> shapes, Format ft, DataType dt) { | |||
vector<uint64_t> dim_info1; | |||
uint8_t *input_data = (uint8_t *)readTestDataFile(input_path, dim_info1); // common.h | |||
TensorDesc input_tensor_desc = TensorDesc(ge::Shape(shapes), ft, dt); | |||
input_tensor_desc.SetRealDimCnt(shapes.size()); | |||
Tensor input_tensor = Tensor(input_tensor_desc, input_data, GetDatTypeSize(dt) * dim_info1[dim_info1[0] + 1]); | |||
return input_tensor; | |||
} |
@@ -1,102 +0,0 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#ifndef ST_RESNET50_GE_COMMON_H_ | |||
#define ST_RESNET50_GE_COMMON_H_ | |||
#include "common/ge_inner_error_codes.h" | |||
#include "utils/tensor_utils.h" | |||
#define MY_USER_GE_LOGI(...) GE_LOG_INFO(1, __VA_ARGS__) | |||
#define MY_USER_GE_LOGW(...) GE_LOG_WARN(1, __VA_ARGS__) | |||
#define MY_USER_GE_LOGE(...) GE_LOG_ERROR(1, 3, __VA_ARGS__) | |||
#ifndef USER_GE_LOGI | |||
#define USER_GE_LOGI MY_USER_GE_LOGI | |||
#endif // USER_GE_LOGI | |||
#ifndef USER_GE_LOGW | |||
#define USER_GE_LOGW MY_USER_GE_LOGW | |||
#endif // USER_GE_LOGW | |||
#ifndef USER_GE_LOGE | |||
#define USER_GE_LOGE MY_USER_GE_LOGE | |||
#endif // USER_GE_LOGE | |||
/// train_flag is 0 when infer, train_flag is 1 when train.this param is set for RunGranph_readData() and | |||
/// RunGraph_initData() | |||
#define TRAIN_FLAG_INFER "infer" | |||
#define TRAIN_FLAG_TRAIN "train" | |||
#include <string.h> | |||
#include <unistd.h> | |||
#include <algorithm> | |||
#include <chrono> | |||
#include <iostream> | |||
#include <thread> | |||
#include <vector> | |||
#include "ge_api.h" | |||
#include "graph.h" | |||
#include "ptest.h" | |||
#include "ops/all_ops.h" | |||
using namespace std; | |||
using namespace ge; | |||
// read bin file and compile result | |||
void update_op_format(Operator ops, Format format = ge::FORMAT_NCHW); | |||
void getDimInfo(FILE *fp, std::vector<uint64_t> &dim_info); | |||
void *readTestDataFile(std::string infile, std::vector<uint64_t> &dim_info); | |||
void *readUint8TestDataFile(std::string infile, int size); | |||
bool allclose(float *a, float *b, uint64_t count, float rtol, float atol); | |||
bool compFp32WithTData(float *actual_output_data, std::string expected_data_file, float rtol, float atol); | |||
Tensor load_variable_input_data(string input_path, std::vector<int64_t> shapes, Format ft = ge::FORMAT_NCHW, | |||
DataType dt = ge::DT_FLOAT); | |||
// constructor Tensor | |||
int GetDatTypeSize(DataType dt); | |||
ge::Tensor genTensor(std::vector<int64_t> tensor_shape, Format format = ge::FORMAT_NCHW, DataType dt = ge::DT_FLOAT); | |||
ge::Tensor genTensor_withVaule(std::vector<int64_t> tensor_shape, float value = 1); | |||
Tensor genTesnor_Shape_as_data(std::vector<int64_t> tensor_shape); | |||
// Init GE | |||
ge::Status GEInitialize_api(string train_flag = "0", string run_mode_path = "0"); | |||
ge::Status GEInitialize_api_new(string train_flag = "infer", string run_mode = "fe"); | |||
ge::Status GEFinalize_api(); | |||
// constructor session and build graph | |||
ge::Session *create_aipp_session(); | |||
ge::Session *create_session(); | |||
ge::Status session_add_and_run_graph(ge::Session *session, uint32_t graphId, Graph &graph, std::vector<Tensor> inputs, | |||
std::vector<Tensor> &outputs); | |||
// common interface for infer | |||
int RunGraph_initData(Graph &graph, string op_name, map<string, std::vector<int64_t>> attr_test, | |||
string train_flag = "infer", string run_mode_path = "fe"); | |||
void Inputs_load_Data(string op_name, std::vector<Tensor> &input, map<string, std::vector<int64_t>> attr_test, | |||
Format format = ge::FORMAT_NCHW, DataType dt = ge::DT_FLOAT); | |||
bool comparaData(std::vector<Tensor> &output, string op_name, map<string, std::vector<int64_t>> attr_test); | |||
int RunGraph_readData(Graph &graph, string op_name, map<string, std::vector<int64_t>> attr_test, | |||
string train_flag = "infer", string run_mode_path = "fe", Format format = ge::FORMAT_NCHW, | |||
DataType dt = ge::DT_FLOAT); | |||
// common interface for train | |||
int buildCheckPointGraph(Graph &graph, map<string, TensorDesc> variables); | |||
int buildInitGraph(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var, | |||
std::vector<float> values_var); | |||
int buildInitGraph_other_dataType(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var); | |||
bool build_multi_input_multi_output_graph(Graph &graph); | |||
void build_big_graph(Graph &graph, map<string, std::vector<int64_t>> attr); | |||
int buildConvGraph_new(Graph &graph, std::vector<TensorDesc> desc_var, std::vector<std::string> name_var, int flag = 2); | |||
#endif // ST_RESNET50_GE_COMMON_H_ |
@@ -1,225 +0,0 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#ifndef ST_RESNET50_PTEST_H_ | |||
#define ST_RESNET50_PTEST_H_ | |||
#include <stdarg.h> | |||
#include <string.h> | |||
#include <exception> | |||
#include <functional> | |||
#include <iostream> | |||
#include <list> | |||
#include <map> | |||
#include <memory> | |||
#include <string> | |||
namespace ptest { | |||
class assertion_error : public std::exception { | |||
public: | |||
const char *what() const throw() { return "Assertion Exception"; } | |||
}; | |||
class TestFixture { | |||
public: | |||
virtual void SetUp() {} | |||
virtual void TearDown() {} | |||
void Run() { _func(); } | |||
void BindFunction(std::function<void(void)> function) { _func = function; } | |||
void SetName(const std::string &name) { _name = name; } | |||
std::string Name() const { return _name; } | |||
virtual ~TestFixture() {} | |||
private: | |||
std::function<void(void)> _func; | |||
std::string _name; | |||
}; | |||
enum TestResult { SUCCESS, FAILED, UNAVAILABLE, UNKNOWN, NOCASEFOUND }; | |||
class TestManager { | |||
public: | |||
static TestManager &GetSingleton() { | |||
static TestManager instance; | |||
return instance; | |||
} | |||
void RegisterTest(const std::string &name, TestFixture *fixture) { _testfixtures[name] = fixture; } | |||
const std::string GetRunningTestcaseName() const { return _running_testcase_name; } | |||
const std::list<std::string> GetAllTestNames() const { | |||
std::list<std::string> result; | |||
for (auto &t : _testfixtures) { | |||
result.push_back(t.first); | |||
} | |||
return result; | |||
} | |||
TestResult RunTest(const std::string &name) { | |||
if (_testfixtures.find(name) == _testfixtures.end()) { | |||
return NOCASEFOUND; | |||
} | |||
_running_testcase_name = name; | |||
do { | |||
SetTestResult(name, UNKNOWN); | |||
_testfixtures[name]->SetUp(); | |||
if (_testresults[name] == FAILED) { | |||
_testresults[name] = UNAVAILABLE; | |||
break; | |||
} | |||
SetTestResult(name, SUCCESS); | |||
try { | |||
_testfixtures[name]->Run(); | |||
} catch (assertion_error &e) { | |||
// Do nothing as the error has been handled by the TestManager. | |||
} | |||
_testfixtures[name]->TearDown(); | |||
} while (0); | |||
return _testresults[name]; | |||
} | |||
void SetTestResult(const std::string &name, TestResult result) { _testresults[name] = result; } | |||
TestResult GetTestResult(const std::string &name) { return _testresults[name]; } | |||
private: | |||
std::map<std::string, TestFixture *> _testfixtures; | |||
std::map<std::string, TestResult> _testresults; | |||
std::string _running_testcase_name; | |||
}; | |||
class TestFixtureRegister { | |||
public: | |||
TestFixtureRegister(const std::string &name, TestFixture *fixture, std::function<void(void)> function) { | |||
fixture->BindFunction(function); | |||
fixture->SetName(name); | |||
TestManager::GetSingleton().RegisterTest(name, fixture); | |||
} | |||
}; | |||
} // namespace ptest | |||
#define _STR(x) #x | |||
#define _EMPTY_NAMESPACE | |||
#define _TEST(NAMESPACE, FIXTURECLASS, TESTNAME, CASENAME) \ | |||
void g_func_##TESTNAME##_##CASENAME(void); \ | |||
NAMESPACE::FIXTURECLASS g_fixture_##TESTNAME##_##CASENAME; \ | |||
ptest::TestFixtureRegister g_register_##TESTNAME##_##CASENAME( \ | |||
_STR(TESTNAME##_##CASENAME), &g_fixture_##TESTNAME##_##CASENAME, g_func_##TESTNAME##_##CASENAME); \ | |||
void g_func_##TESTNAME##_##CASENAME(void) | |||
#define TEST(TESTNAME, CASENAME) _TEST(ptest, TestFixture, TESTNAME, CASENAME) | |||
#define TEST_F(TESTFIXTURE, CASENAME) _TEST(_EMPTY_NAMESPACE, TESTFIXTURE, TESTFIXTURE, CASENAME) | |||
#define EXPECT_TRUE(X) \ | |||
do { \ | |||
if (!(X)) { \ | |||
std::string test_name = ptest::TestManager::GetSingleton().GetRunningTestcaseName(); \ | |||
ptest::TestManager::GetSingleton().SetTestResult(test_name, ptest::FAILED); \ | |||
std::cerr << #X << "Expectation Failed\n" \ | |||
<< "Testcase Name: " << test_name << "\n" \ | |||
<< "File: " __FILE__ << "\tLine:" << __LINE__ << std::endl; \ | |||
} \ | |||
} while (0); | |||
// With the macro definition ensures that the compiler can detect compiler warning. | |||
#define Max_Log_Len 1024 | |||
#define PRINT_ERR(lpszFormat, ...) \ | |||
do { \ | |||
char szTmpBuf[Max_Log_Len + 1] = {0}; \ | |||
snprintf(szTmpBuf, Max_Log_Len, lpszFormat, ##__VA_ARGS__); \ | |||
std::cerr << szTmpBuf << std::endl; \ | |||
} while (0) | |||
// Increase the content of print error messages and error to facilitate rapid analysis | |||
#define EXPECT_TRUE_C(X, ERR_TYPE, format, ...) \ | |||
do { \ | |||
if (!(X)) { \ | |||
std::string test_name = ptest::TestManager::GetSingleton().GetRunningTestcaseName(); \ | |||
ptest::TestManager::GetSingleton().SetTestResult(test_name, ptest::FAILED); \ | |||
std::cerr << #X << " Expectation Failed." \ | |||
<< "Testcase Name: " << test_name << " File:" __FILE__ << " Line:" << __LINE__ << std::endl; \ | |||
PRINT_ERR("[" ERR_TYPE "]" format, ##__VA_ARGS__); \ | |||
} \ | |||
} while (0) | |||
#define ASSERT_TRUE(X) \ | |||
do { \ | |||
if (!(X)) { \ | |||
std::string test_name = ptest::TestManager::GetSingleton().GetRunningTestcaseName(); \ | |||
ptest::TestManager::GetSingleton().SetTestResult(test_name, ptest::FAILED); \ | |||
std::cerr << #X << "Assertion Failed\n" \ | |||
<< "Testcase Name: " << test_name << "\n" \ | |||
<< "File: " __FILE__ << "\tLine:" << __LINE__ << std::endl; \ | |||
throw ptest::assertion_error(); \ | |||
} \ | |||
} while (0); | |||
// Add printing error information and error line content for quick analysis | |||
#define ASSERT_TRUE_C(X, ERR_TYPE, format, ...) \ | |||
do { \ | |||
if (!(X)) { \ | |||
std::string test_name = ptest::TestManager::GetSingleton().GetRunningTestcaseName(); \ | |||
ptest::TestManager::GetSingleton().SetTestResult(test_name, ptest::FAILED); \ | |||
std::cerr << #X << " Assertion Failed." \ | |||
<< "Testcase Name: " << test_name << " File:" __FILE__ << " Line:" << __LINE__ << std::endl; \ | |||
PRINT_ERR("[" ERR_TYPE "]" format, ##__VA_ARGS__); \ | |||
throw ptest::assertion_error(); \ | |||
} \ | |||
} while (0); | |||
#define CONFIG_ERR "CONFIG_ERR" | |||
#define LOAD_MODEL_ERR "LOAD_MODEL_ERR" | |||
#define FILE_READ_ERR "FILE_READ_ERR" | |||
#define RUN_ERROR "RUN_ERROR" | |||
#define MEM_ERROR "MEM_ERROR" | |||
#define RESULT_ERR "RESULT_ERR" | |||
#define EXPECT_FALSE(X) EXPECT_TRUE(!(X)) | |||
#define EXPECT_EQ(X, Y) EXPECT_TRUE(((X) == (Y))) | |||
#define EXPECT_NE(X, Y) EXPECT_TRUE(((X) != (Y))) | |||
#define EXPECT_GT(X, Y) EXPECT_TRUE(((X) > (Y))) | |||
#define EXPECT_GE(X, Y) EXPECT_TRUE(((X) >= (Y))) | |||
#define EXPECT_LT(X, Y) EXPECT_TRUE(((X) < (Y))) | |||
#define EXPECT_LE(X, Y) EXPECT_TRUE(((X) <= (Y))) | |||
#define EXPECT_FALSE_C(X, ERR_TYPE, format, ...) EXPECT_TRUE_C(!(X), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_EQ_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) == (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_NE_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) != (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_GT_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) > (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_GE_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) >= (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_LT_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) < (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define EXPECT_LE_C(X, Y, ERR_TYPE, format, ...) EXPECT_TRUE_C(((X) <= (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_FALSE(X) ASSERT_TRUE(!(X)) | |||
#define ASSERT_EQ(X, Y) ASSERT_TRUE(((X) == (Y))) | |||
#define ASSERT_NE(X, Y) ASSERT_TRUE(((X) != (Y))) | |||
#define ASSERT_GT(X, Y) ASSERT_TRUE(((X) > (Y))) | |||
#define ASSERT_GE(X, Y) ASSERT_TRUE(((X) >= (Y))) | |||
#define ASSERT_LT(X, Y) ASSERT_TRUE(((X) < (Y))) | |||
#define ASSERT_LE(X, Y) ASSERT_TRUE(((X) <= (Y))) | |||
#define ASSERT_FALSE_C(X, ERR_TYPE, format, ...) ASSERT_TRUE_C(!(X), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_EQ_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) == (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_NE_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) != (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_GT_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) > (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_GE_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) >= (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_LT_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) < (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#define ASSERT_LE_C(X, Y, ERR_TYPE, format, ...) ASSERT_TRUE_C(((X) <= (Y)), ERR_TYPE, format, ##__VA_ARGS__) | |||
#endif // ST_RESNET50_PTEST_H_ |
@@ -1,852 +0,0 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <assert.h> | |||
#include <sys/stat.h> | |||
#include <sys/types.h> | |||
#include <algorithm> | |||
#include <chrono> | |||
#include <ctime> | |||
#include <sstream> | |||
#include "common.h" | |||
#include "ge_api.h" | |||
#include "graph.h" | |||
#include "ops/all_ops.h" | |||
#include "types.h" | |||
#include "utils/tensor_utils.h" | |||
using namespace std; | |||
using namespace ge; | |||
using namespace op; | |||
typedef bool (*Func)(Graph &graph); | |||
#define PADDING_MODE 6 | |||
#define GRAD_PADDING_MODE 3 | |||
vector<int64_t> pad_1{1, 1, 1, 1}; | |||
vector<int64_t> pad_0{0, 0, 0, 0}; | |||
vector<int64_t> stride_1{1, 1}; | |||
vector<int64_t> stride_2{2, 2}; | |||
// (int out_channels, int h, int w, vector<uint_64> stride{1,1}, vector<uint_64> pad{1,1,1,1}, op::Data() input) | |||
#define GENERATE_CONV_VAR(LAYER, BLK, OPNUM, in_channels, out_channels, h, w, stride, pad, input) \ | |||
auto &LAYER##_##BLK##_##OPNUM##_input = input; \ | |||
\ | |||
TensorDesc LAYER##_##BLK##_##OPNUM##_desc(ge::Shape({out_channels, in_channels, h, w}), FORMAT_NCHW, DT_FLOAT); \ | |||
auto LAYER##_##BLK##_##OPNUM##_weight = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_weight"); \ | |||
LAYER##_##BLK##_##OPNUM##_weight.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_weight = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_weight"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_weight.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_weight.update_input_desc_x(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
cout << string(#LAYER) + string(#BLK) + string(#OPNUM) << "'s weight shape is:" << in_channels << out_channels << h \ | |||
<< w << endl; \ | |||
cout << string(#LAYER) + string(#BLK) + string(#OPNUM) \ | |||
<< "'s input_x op's shape is:" << input.GetOutputDesc("y").GetShape().GetDim(2) << endl; \ | |||
auto LAYER##_##BLK##_##OPNUM##_tmp_dims = input.GetOutputDesc("y").GetShape().GetDims(); \ | |||
for (auto LAYER##_##BLK##_##OPNUM##_tmp_it = LAYER##_##BLK##_##OPNUM##_tmp_dims.begin(); \ | |||
LAYER##_##BLK##_##OPNUM##_tmp_it != LAYER##_##BLK##_##OPNUM##_tmp_dims.end(); \ | |||
LAYER##_##BLK##_##OPNUM##_tmp_it++) { \ | |||
cout << *LAYER##_##BLK##_##OPNUM##_tmp_it; \ | |||
} \ | |||
cout << endl; \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM = op::Conv2D(string(#LAYER) + string(#BLK) + string(#OPNUM)) \ | |||
.set_input_x(input, "y") \ | |||
.set_input_filter(LAYER##_##BLK##_##OPNUM##_weight) \ | |||
.set_attr_strides({1, 1, stride[0], stride[1]}) \ | |||
.set_attr_pads(pad) \ | |||
.set_attr_data_format("NCHW"); \ | |||
update_op_format(LAYER##_##BLK##_##OPNUM); | |||
#define GENERATE_CONSTANT(LAYER, BLK, OPNUM, CONSTNAME) \ | |||
Tensor LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_tensor; \ | |||
float *LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_data = new float[LAYER##_##BLK##_##OPNUM##_size]; \ | |||
for (int i = 0; i < (int)LAYER##_##BLK##_##OPNUM##_size; i++) { \ | |||
*(LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_data + i) = 0.01; \ | |||
} \ | |||
LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_tensor.SetData((uint8_t *)LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_data, \ | |||
LAYER##_##BLK##_##OPNUM##_size * sizeof(float)); \ | |||
LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_tensor.SetTensorDesc(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_constant = \ | |||
op::Constant().set_attr_value(LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_tensor); \ | |||
LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_constant.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
delete[] LAYER##_##BLK##_##OPNUM##_##CONSTNAME##_data; | |||
#define GENERATE_CONV_VAR_VAR(LAYER, BLK, OPNUM, in_channels, out_channels, h, w, stride, pad, input) \ | |||
TensorDesc LAYER##_##BLK##_##OPNUM##_desc(ge::Shape({out_channels, in_channels, h, w}), FORMAT_NCHW, DT_FLOAT); \ | |||
uint32_t LAYER##_##BLK##_##OPNUM##_size = LAYER##_##BLK##_##OPNUM##_desc.GetShape().GetShapeSize(); \ | |||
auto LAYER##_##BLK##_##OPNUM##_weight = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_weight"); \ | |||
LAYER##_##BLK##_##OPNUM##_weight.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_weight = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_weight"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_weight.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, weight); \ | |||
auto LAYER##_##BLK##_##OPNUM##_weight_assign = op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_weight) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_weight_constant); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, mom_weight); \ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_weight_assign = \ | |||
op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_mom_weight) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_mom_weight_constant); \ | |||
\ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_weight); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_mom_weight); | |||
// (int out_channels, Operator& input) | |||
#define GENERATE_BN_VAR(LAYER, BLK, OPNUM, out_channels, input) \ | |||
auto &LAYER##_##BLK##_##OPNUM##_input = input; \ | |||
\ | |||
TensorDesc LAYER##_##BLK##_##OPNUM##_desc(ge::Shape({1, out_channels, 1, 1}), FORMAT_NCHW, DT_FLOAT); \ | |||
auto LAYER##_##BLK##_##OPNUM##_scale = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_scale"); \ | |||
LAYER##_##BLK##_##OPNUM##_scale.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_scale = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_scale"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_scale.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_b = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_b"); \ | |||
LAYER##_##BLK##_##OPNUM##_b.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_b = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_b"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_b.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mean = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mean"); \ | |||
LAYER##_##BLK##_##OPNUM##_mean.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
auto LAYER##_##BLK##_##OPNUM##_variance = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_variance"); \ | |||
LAYER##_##BLK##_##OPNUM##_variance.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM = op::FusedBatchNorm(string(#LAYER) + string(#BLK) + string(#OPNUM)) \ | |||
.set_input_x(input, "y") \ | |||
.set_input_scale(LAYER##_##BLK##_##OPNUM##_scale) \ | |||
.set_input_b(LAYER##_##BLK##_##OPNUM##_b) \ | |||
.set_input_mean(LAYER##_##BLK##_##OPNUM##_mean) \ | |||
.set_input_variance(LAYER##_##BLK##_##OPNUM##_variance) \ | |||
.set_attr_mode(1) \ | |||
.set_attr_epsilon(1e-5) \ | |||
.set_attr_is_training(true); | |||
#define GENERATE_BN_VAR_VAR(LAYER, BLK, OPNUM, out_channels, input) \ | |||
TensorDesc LAYER##_##BLK##_##OPNUM##_desc(ge::Shape({1, out_channels, 1, 1}), FORMAT_NCHW, DT_FLOAT); \ | |||
uint32_t LAYER##_##BLK##_##OPNUM##_size = LAYER##_##BLK##_##OPNUM##_desc.GetShape().GetShapeSize(); \ | |||
auto LAYER##_##BLK##_##OPNUM##_scale = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_scale"); \ | |||
LAYER##_##BLK##_##OPNUM##_scale.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_scale = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_scale"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_scale.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_b = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_b"); \ | |||
LAYER##_##BLK##_##OPNUM##_b.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_b = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mom_b"); \ | |||
LAYER##_##BLK##_##OPNUM##_mom_b.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mean = op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_mean"); \ | |||
LAYER##_##BLK##_##OPNUM##_mean.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
auto LAYER##_##BLK##_##OPNUM##_variance = \ | |||
op::Variable(string(#LAYER) + string(#BLK) + string(#OPNUM) + "_variance"); \ | |||
LAYER##_##BLK##_##OPNUM##_variance.update_output_desc_y(LAYER##_##BLK##_##OPNUM##_desc); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, scale); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_scale_assign = op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_scale) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_scale_constant); \ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, mom_scale); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_scale_assign = \ | |||
op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_mom_scale) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_mom_scale_constant); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, b); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_b_assign = \ | |||
op::Assign().set_input_ref(LAYER##_##BLK##_##OPNUM##_b).set_input_value(LAYER##_##BLK##_##OPNUM##_b_constant); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, mom_b); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mom_b_assign = op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_mom_b) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_mom_b_constant); \ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, mean); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_mean_assign = op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_mean) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_mean_constant); \ | |||
\ | |||
GENERATE_CONSTANT(LAYER, BLK, OPNUM, variance); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_variance_assign = op::Assign() \ | |||
.set_input_ref(LAYER##_##BLK##_##OPNUM##_variance) \ | |||
.set_input_value(LAYER##_##BLK##_##OPNUM##_variance_constant); \ | |||
\ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_scale); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_mom_scale); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_b); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_mom_b); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_mean); \ | |||
input.push_back(LAYER##_##BLK##_##OPNUM##_variance); | |||
// (int out_channels, Operator& input) | |||
#define GENERATE_RELU_VAR(LAYER, BLK, OPNUM, input) \ | |||
auto &LAYER##_##BLK##_##OPNUM##_input = input; \ | |||
auto LAYER##_##BLK##_##OPNUM = op::Relu(string(#LAYER) + string(#BLK) + string(#OPNUM)).set_input_x(input, "y"); | |||
// (int out_channels, Operator& input) | |||
#define GENERATE_MAXPOOL_VAR(LAYER, BLK, OPNUM, input) \ | |||
auto &LAYER##_##BLK##_##OPNUM##_input = input; \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM = op::MaxPoolWithArgmax(string(#LAYER) + string(#BLK) + string(#OPNUM)) \ | |||
.set_input_x(input, "y") \ | |||
.set_attr_ksize({1, 3, 3, 1}) \ | |||
.set_attr_padding("SAME") \ | |||
.set_attr_strides({1, 2, 2, 1}); | |||
// (int out_channels, Operator& input) | |||
#define GENERATE_ADD_VAR(LAYER, BLK, OPNUM, input_x1, input_x2) \ | |||
auto LAYER##_##BLK##_##OPNUM = \ | |||
op::Add(string(#LAYER) + string(#BLK) + string(#OPNUM)).set_input_x1(input_x1, "y").set_input_x2(input_x2, "y"); | |||
// (int in_channels, int out_channels,vector<int64_t> stride{1,1}, Operator& input) | |||
#define MAKE_RESIDUAL_BLOCK(LAYER, BLK, in_channels, out_channels, stride, input) \ | |||
auto &LAYER##_##BLK##_input = input; \ | |||
auto &LAYER##_##BLK##_stride = stride; \ | |||
int LAYER##_##BLK##_out_chls = out_channels / 4; \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv1, in_channels, LAYER##_##BLK##_out_chls, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn1, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_conv1); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu1, LAYER##_##BLK##_bn1); \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_out_chls, 3, 3, stride_1, pad_1, \ | |||
LAYER##_##BLK##_relu1); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_conv2); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu2, LAYER##_##BLK##_bn2); \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv3, LAYER##_##BLK##_out_chls, out_channels, 1, 1, stride_1, pad_0, \ | |||
LAYER##_##BLK##_relu2); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn3, out_channels, LAYER##_##BLK##_conv3); \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv4, in_channels, out_channels, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn4, out_channels, LAYER##_##BLK##_conv4); \ | |||
\ | |||
GENERATE_ADD_VAR(LAYER, BLK, add5, LAYER##_##BLK##_bn3, LAYER##_##BLK##_bn4); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu5, LAYER##_##BLK##_add5); \ | |||
\ | |||
auto &LAYER##_##BLK##_output = LAYER##_##BLK##_relu5; \ | |||
auto &LAYER##_##BLK##_output_label = "y"; | |||
#define MAKE_RESIDUAL_BLOCK_VAR(LAYER, BLK, in_channels, out_channels, stride, input) \ | |||
int LAYER##_##BLK##_out_chls = out_channels / 4; \ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv1, in_channels, LAYER##_##BLK##_out_chls, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn1, LAYER##_##BLK##_out_chls, input); \ | |||
\ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_out_chls, 3, 3, stride_1, pad_1, \ | |||
input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn2, LAYER##_##BLK##_out_chls, input); \ | |||
\ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv3, LAYER##_##BLK##_out_chls, out_channels, 1, 1, stride_1, pad_0, input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn3, out_channels, input); \ | |||
\ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv4, in_channels, out_channels, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn4, out_channels, input); | |||
// (int in_channels, int out_channels,vector<int64_t> stride{1,1}, Operator& input) | |||
#define MAKE_NORMAL_BLOCK(LAYER, BLK, in_channels, out_channels, stride, input) \ | |||
auto &LAYER##_##BLK##_input = input; \ | |||
auto &LAYER##_##BLK##_stride = stride; \ | |||
int LAYER##_##BLK##_out_chls = out_channels / 4; \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv1, in_channels, LAYER##_##BLK##_out_chls, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn1, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_conv1); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu1, LAYER##_##BLK##_bn1); \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_out_chls, 3, 3, stride_1, pad_1, \ | |||
LAYER##_##BLK##_relu1); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_conv2); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu2, LAYER##_##BLK##_bn2); \ | |||
\ | |||
GENERATE_CONV_VAR(LAYER, BLK, conv3, LAYER##_##BLK##_out_chls, out_channels, 1, 1, stride_1, pad_0, \ | |||
LAYER##_##BLK##_relu2); \ | |||
GENERATE_BN_VAR(LAYER, BLK, bn3, out_channels, LAYER##_##BLK##_conv3); \ | |||
\ | |||
GENERATE_ADD_VAR(LAYER, BLK, add5, LAYER##_##BLK##_bn3, input); \ | |||
GENERATE_RELU_VAR(LAYER, BLK, relu5, LAYER##_##BLK##_add5); \ | |||
\ | |||
auto &LAYER##_##BLK##_output = LAYER##_##BLK##_relu5; \ | |||
auto &LAYER##_##BLK##_output_label = "y"; | |||
#define MAKE_NORMAL_BLOCK_VAR(LAYER, BLK, in_channels, out_channels, stride, input) \ | |||
int LAYER##_##BLK##_out_chls = out_channels / 4; \ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv1, in_channels, LAYER##_##BLK##_out_chls, 1, 1, stride, pad_0, input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn1, LAYER##_##BLK##_out_chls, input); \ | |||
\ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv2, LAYER##_##BLK##_out_chls, LAYER##_##BLK##_out_chls, 3, 3, stride_1, pad_1, \ | |||
input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn2, LAYER##_##BLK##_out_chls, input); \ | |||
\ | |||
GENERATE_CONV_VAR_VAR(LAYER, BLK, conv3, LAYER##_##BLK##_out_chls, out_channels, 1, 1, stride_1, pad_0, input); \ | |||
GENERATE_BN_VAR_VAR(LAYER, BLK, bn3, out_channels, input); | |||
// (int in_channels, int out_channels,vector<int64_t> stride{1,1}, Operator& input) | |||
#define MAKE_RESIDUAL_LAYER(LAYER, in_channels, out_channels, stride, input) \ | |||
MAKE_RESIDUAL_BLOCK(LAYER, blk1, in_channels, out_channels, stride, input); \ | |||
\ | |||
auto &LAYER##_output = LAYER##_blk1_output; \ | |||
auto &LAYER##_output_label = LAYER##_blk1_output_label; | |||
#define MAKE_RESIDUAL_LAYER_VAR(LAYER, in_channels, out_channels, stride, input) \ | |||
MAKE_RESIDUAL_BLOCK_VAR(LAYER, blk1, in_channels, out_channels, stride, input); | |||
// (int in_channels, int out_channels,vector<int64_t> stride{1,1}, Operator& input) | |||
#define MAKE_NORMAL_LAYER(LAYER, in_channels, out_channels, stride, input) \ | |||
MAKE_NORMAL_BLOCK(LAYER, blk1, in_channels, out_channels, stride, input); \ | |||
\ | |||
auto &LAYER##_output = LAYER##_blk1_output; \ | |||
auto &LAYER##_output_label = LAYER##_blk1_output_label; | |||
#define MAKE_NORMAL_LAYER_VAR(LAYER, in_channels, out_channels, stride, input) \ | |||
MAKE_NORMAL_BLOCK_VAR(LAYER, blk1, in_channels, out_channels, stride, input); | |||
#define MAKE_RESNET50(input) \ | |||
MAKE_RESIDUAL_LAYER(layer1, 64, 256, stride_1, input) \ | |||
MAKE_NORMAL_LAYER(layer2, 256, 256, stride_1, layer1_output) \ | |||
MAKE_NORMAL_LAYER(layer3, 256, 256, stride_1, layer2_output) \ | |||
MAKE_RESIDUAL_LAYER(layer4, 256, 512, stride_2, layer3_output) \ | |||
MAKE_NORMAL_LAYER(layer5, 512, 512, stride_1, layer4_output) \ | |||
MAKE_NORMAL_LAYER(layer6, 512, 512, stride_1, layer5_output) \ | |||
MAKE_NORMAL_LAYER(layer7, 512, 512, stride_1, layer6_output) \ | |||
MAKE_RESIDUAL_LAYER(layer8, 512, 1024, stride_2, layer7_output) \ | |||
MAKE_NORMAL_LAYER(layer9, 1024, 1024, stride_1, layer8_output) \ | |||
MAKE_NORMAL_LAYER(layer10, 1024, 1024, stride_1, layer9_output) \ | |||
MAKE_NORMAL_LAYER(layer11, 1024, 1024, stride_1, layer10_output) \ | |||
MAKE_NORMAL_LAYER(layer12, 1024, 1024, stride_1, layer11_output) \ | |||
MAKE_NORMAL_LAYER(layer13, 1024, 1024, stride_1, layer12_output) \ | |||
MAKE_RESIDUAL_LAYER(layer14, 1024, 2048, stride_2, layer13_output) \ | |||
MAKE_NORMAL_LAYER(layer15, 2048, 2048, stride_1, layer14_output) \ | |||
MAKE_NORMAL_LAYER(layer16, 2048, 2048, stride_1, layer15_output) \ | |||
\ | |||
auto &resnet50_output = layer16_output; \ | |||
auto &resnet50_output_label = layer16_output_label; | |||
#define MAKE_RESNET50_VAR(inputs) \ | |||
MAKE_RESIDUAL_LAYER_VAR(layer1, 64, 256, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer2, 256, 256, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer3, 256, 256, stride_1, inputs) \ | |||
MAKE_RESIDUAL_LAYER_VAR(layer4, 256, 512, stride_2, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer5, 512, 512, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer6, 512, 512, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer7, 512, 512, stride_1, inputs) \ | |||
MAKE_RESIDUAL_LAYER_VAR(layer8, 512, 1024, stride_2, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer9, 1024, 1024, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer10, 1024, 1024, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer11, 1024, 1024, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer12, 1024, 1024, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer13, 1024, 1024, stride_1, inputs) \ | |||
MAKE_RESIDUAL_LAYER_VAR(layer14, 1024, 2048, stride_2, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer15, 2048, 2048, stride_1, inputs) \ | |||
MAKE_NORMAL_LAYER_VAR(layer16, 2048, 2048, stride_1, inputs) \ | |||
//--------------------------------------------------------------------------------------------- | |||
// (Operator& input) | |||
#define GENERATE_BIASADD_GRAD(LAYER, BLK, OPNUM, input) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = \ | |||
op::BiasAddGrad(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")) \ | |||
.set_input_x(input, input.name_out_dx()); | |||
// (Operator& input) | |||
#define GENERATE_MATMUL_GRAD(LAYER, BLK, OPNUM, input) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = \ | |||
op::MatMul(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")).set_input_x1(input); | |||
// (Operator& input) | |||
#define GENERATE_RESHAPE_GRAD(LAYER, BLK, OPNUM, input) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = \ | |||
op::Reshape(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")).set_input_tensor(input); | |||
// (Operator& input_grad, Operator& input_maxpool) | |||
#define GENERATE_MAXPOOL_GRAD(LAYER, BLK, OPNUM, input_grad, input_maxpool) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = \ | |||
op::MaxPoolGradWithArgmax(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")) \ | |||
.set_input_x(LAYER##_##BLK##_##OPNUM##_input, "y") \ | |||
.set_input_grad(input_grad) \ | |||
.set_input_argmax(input_maxpool, input_maxpool.name_out_argmax()) \ | |||
.set_attr_ksize({1, 1, 3, 3}) \ | |||
.set_attr_strides({1, 1, 2, 2}) \ | |||
.set_attr_padding("SAME"); | |||
// (Operator& input_dy) | |||
#define GENERATE_RELU_GRAD(LAYER, BLK, OPNUM, input_dy, dy_label) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = op::ReluGrad(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")) \ | |||
.set_input_gradients(input_dy, dy_label) \ | |||
.set_input_features(LAYER##_##BLK##_##OPNUM, "y"); | |||
// (Operator& input_dy) | |||
#define GENERATE_BN_GRAD(LAYER, BLK, OPNUM, input_dy) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = \ | |||
op::FusedBatchNormGrad(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")) \ | |||
.set_input_dy(input_dy, "backprops") \ | |||
.set_input_x(LAYER##_##BLK##_##OPNUM##_input, "y") \ | |||
.set_input_scale(LAYER##_##BLK##_##OPNUM##_scale) \ | |||
.set_input_save_mean(LAYER##_##BLK##_##OPNUM, "save_mean") \ | |||
.set_input_save_inv_variance(LAYER##_##BLK##_##OPNUM, "save_inv_variance") \ | |||
.set_attr_epsilon(0.0001); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_momentum_scale = \ | |||
op::ApplyMomentum() \ | |||
.set_input_accum(LAYER##_##BLK##_##OPNUM##_mom_scale) \ | |||
.set_input_grad(LAYER##_##BLK##_##OPNUM##_grad, LAYER##_##BLK##_##OPNUM##_grad.name_out_bn_scale()) \ | |||
.set_input_lr(label1) \ | |||
.set_input_momentum(label1) \ | |||
.set_input_var(LAYER##_##BLK##_##OPNUM##_scale); \ | |||
\ | |||
auto LAYER##_##BLK##_##OPNUM##_momentum_b = \ | |||
op::ApplyMomentum() \ | |||
.set_input_accum(LAYER##_##BLK##_##OPNUM##_mom_b) \ | |||
.set_input_grad(LAYER##_##BLK##_##OPNUM##_grad, LAYER##_##BLK##_##OPNUM##_grad.name_out_bn_bias()) \ | |||
.set_input_lr(label1) \ | |||
.set_input_momentum(label1) \ | |||
.set_input_var(LAYER##_##BLK##_##OPNUM##_b); | |||
// (Operator& input) | |||
#define GENERATE_CONV_PROP_FILTER(LAYER, BLK, OPNUM, input_bngrad, stride) \ | |||
auto LAYER##_##BLK##_##OPNUM##_propfilter = \ | |||
op::Conv2DBackpropFilterD(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("_propfilter")) \ | |||
.set_input_x(LAYER##_##BLK##_##OPNUM##_input, "y") \ | |||
.set_attr_filter_size(LAYER##_##BLK##_##OPNUM##_desc.GetShape().GetDims()) \ | |||
.set_input_out_backprop(input_bngrad, input_bngrad.name_out_dx()) \ | |||
.set_attr_strides(stride) \ | |||
.set_attr_pads({1, 1, 1, 1}); \ | |||
\ | |||
update_op_format(LAYER##_##BLK##_##OPNUM##_propfilter); \ | |||
auto LAYER##_##BLK##_##OPNUM##_momentum_weight = op::ApplyMomentum() \ | |||
.set_input_accum(LAYER##_##BLK##_##OPNUM##_mom_weight) \ | |||
.set_input_grad(LAYER##_##BLK##_##OPNUM##_propfilter) \ | |||
.set_input_lr(label1) \ | |||
.set_input_momentum(label1) \ | |||
.set_input_var(LAYER##_##BLK##_##OPNUM##_weight); | |||
///.set_attr_input_size({input_bngrad.name_out_dx().GetOutputDesc().GetShape().GetDim(0),LAYER##_##BLK##_##OPNUM##_weight.GetOutputDesc().GetShape().GetDim(1), | |||
///input_bngrad.name_out_dx().GetOutputDesc().GetShape().GetDim(2)*stride[2], | |||
///input_bngrad.name_out_dx().GetOutputDesc().GetShape().GetDim(3)*stride[3]}) | |||
#define GENERATE_CONV_PROP_INPUT(LAYER, BLK, OPNUM, input_bngrad, stride) \ | |||
auto LAYER##_##BLK##_##OPNUM##_propinput = \ | |||
op::Conv2DBackpropInputD(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("_propinput")) \ | |||
.set_attr_input_size(LAYER##_##BLK##_##OPNUM##_input.GetOutputDesc("y").GetShape().GetDims()) \ | |||
.set_input_filter(LAYER##_##BLK##_##OPNUM##_weight) \ | |||
.set_input_out_backprop(input_bngrad, input_bngrad.name_out_dx()) \ | |||
.set_attr_strides(stride) \ | |||
.set_attr_pads({1, 1, 1, 1}); \ | |||
cout << string(#LAYER) + string(#BLK) + string(#OPNUM) + "_propinput" \ | |||
<< "'s input_x op's shape is:" << input_bngrad.GetOutputDesc("dx").GetShape().GetDim(3) * stride[3] << endl; \ | |||
cout << string(#LAYER) + string(#BLK) + string(#OPNUM) + "_propinput" \ | |||
<< "'s input_x op's shape is:" << input_bngrad.GetOutputDesc("dx").GetShape().GetDim(2) * stride[2] << endl; \ | |||
\ | |||
update_op_format(LAYER##_##BLK##_##OPNUM##_propinput); \ | |||
auto &LAYER##_##BLK##_##OPNUM##_propinput_label = "y" | |||
// (int out_channels, Operator& input) | |||
#define GENERATE_ADD_GRAD(LAYER, BLK, OPNUM, input_x1, input_x1_label, input_x2, input_x2_label) \ | |||
auto LAYER##_##BLK##_##OPNUM##_grad = op::Add(string(#LAYER) + string(#BLK) + string(#OPNUM) + string("grad")) \ | |||
.set_input_x1(input_x1, input_x1_label) \ | |||
.set_input_x2(input_x2, input_x2_label); | |||
// (Operator& input) | |||
#define MAKE_RESIDUAL_BLOCK_GRAD(LAYER, BLK, input_dy, dy_label) \ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu5, input_dy, dy_label); \ | |||
\ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn4, LAYER##_##BLK##_relu5_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv4, LAYER##_##BLK##_bn4_grad, LAYER##_##BLK##_stride); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv4, LAYER##_##BLK##_bn4_grad, LAYER##_##BLK##_stride); \ | |||
\ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn3, LAYER##_##BLK##_relu5_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv3, LAYER##_##BLK##_bn3_grad, stride_1); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv3, LAYER##_##BLK##_bn3_grad, stride_1); \ | |||
\ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu2, LAYER##_##BLK##_conv3_propinput, "y"); \ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn2, LAYER##_##BLK##_relu2_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv2, LAYER##_##BLK##_bn2_grad, stride_1); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv2, LAYER##_##BLK##_bn2_grad, stride_1); \ | |||
\ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu1, LAYER##_##BLK##_conv2_propinput, "y"); \ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn1, LAYER##_##BLK##_relu1_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv1, LAYER##_##BLK##_bn1_grad, LAYER##_##BLK##_stride); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv1, LAYER##_##BLK##_bn1_grad, LAYER##_##BLK##_stride); \ | |||
\ | |||
GENERATE_ADD_GRAD(LAYER, BLK, add5, LAYER##_##BLK##_conv1_propinput, LAYER##_##BLK##_conv1_propinput_label, \ | |||
LAYER##_##BLK##_conv4_propinput, LAYER##_##BLK##_conv4_propinput_label); \ | |||
\ | |||
auto &LAYER##_##BLK##_grad_output = LAYER##_##BLK##_add5_grad; \ | |||
auto &LAYER##_##BLK##_grad_output_label = "y" | |||
// (Operator& input) | |||
#define MAKE_NORMAL_BLOCK_GRAD(LAYER, BLK, input_dy, dy_label) \ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu5, input_dy, dy_label); \ | |||
\ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn3, LAYER##_##BLK##_relu5_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv3, LAYER##_##BLK##_bn3_grad, stride_1); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv3, LAYER##_##BLK##_bn3_grad, stride_1); \ | |||
\ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu2, LAYER##_##BLK##_conv3_propinput, "y"); \ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn2, LAYER##_##BLK##_relu2_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv2, LAYER##_##BLK##_bn2_grad, stride_1); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv2, LAYER##_##BLK##_bn2_grad, stride_1); \ | |||
\ | |||
GENERATE_RELU_GRAD(LAYER, BLK, relu1, LAYER##_##BLK##_conv2_propinput, "y"); \ | |||
GENERATE_BN_GRAD(LAYER, BLK, bn1, LAYER##_##BLK##_relu1_grad); \ | |||
GENERATE_CONV_PROP_FILTER(LAYER, BLK, conv1, LAYER##_##BLK##_bn1_grad, LAYER##_##BLK##_stride); \ | |||
GENERATE_CONV_PROP_INPUT(LAYER, BLK, conv1, LAYER##_##BLK##_bn1_grad, LAYER##_##BLK##_stride); \ | |||
\ | |||
GENERATE_ADD_GRAD(LAYER, BLK, add5, LAYER##_##BLK##_conv1_propinput, LAYER##_##BLK##_conv1_propinput_label, \ | |||
input_dy, dy_label); \ | |||
\ | |||
auto &LAYER##_##BLK##_grad_output = LAYER##_##BLK##_add5_grad; \ | |||
auto &LAYER##_##BLK##_grad_output_label = "y" | |||
// (Operator& input_dy) | |||
#define MAKE_RESIDUAL_LAYER_GRAD(LAYER, input_dy, dy_label) \ | |||
MAKE_RESIDUAL_BLOCK_GRAD(LAYER, blk1, input_dy, dy_label); \ | |||
\ | |||
auto &LAYER##_grad_output = LAYER##_blk1_grad_output; \ | |||
auto &LAYER##_grad_output_label = LAYER##_blk1_grad_output_label; | |||
// (Operator& input_dy) | |||
#define MAKE_NORMAL_LAYER_GRAD(LAYER, input_dy, dy_label) \ | |||
MAKE_NORMAL_BLOCK_GRAD(LAYER, blk1, input_dy, dy_label); \ | |||
\ | |||
auto &LAYER##_grad_output = LAYER##_blk1_grad_output; \ | |||
auto &LAYER##_grad_output_label = LAYER##_blk1_grad_output_label; | |||
#define MAKE_RESNET50_GRAD(input_dy, dy_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer16, input_dy, dy_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer15, layer16_grad_output, layer16_grad_output_label) \ | |||
MAKE_RESIDUAL_LAYER_GRAD(layer14, layer15_grad_output, layer15_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer13, layer14_grad_output, layer14_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer12, layer13_grad_output, layer13_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer11, layer12_grad_output, layer12_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer10, layer11_grad_output, layer11_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer9, layer10_grad_output, layer10_grad_output_label) \ | |||
MAKE_RESIDUAL_LAYER_GRAD(layer8, layer9_grad_output, layer9_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer7, layer8_grad_output, layer8_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer6, layer7_grad_output, layer7_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer5, layer6_grad_output, layer6_grad_output_label) \ | |||
MAKE_RESIDUAL_LAYER_GRAD(layer4, layer5_grad_output, layer5_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer3, layer4_grad_output, layer4_grad_output_label) \ | |||
MAKE_NORMAL_LAYER_GRAD(layer2, layer3_grad_output, layer3_grad_output_label) \ | |||
MAKE_RESIDUAL_LAYER_GRAD(layer1, layer2_grad_output, layer2_grad_output_label) \ | |||
\ | |||
auto &resnet50_grad_output = layer1_grad_output; \ | |||
auto &resnet50_grad_output_label = layer1_grad_output_label; | |||
bool resnet50(Graph &graph) { | |||
auto data = op::Data().set_attr_index(0); | |||
auto data1 = op::Data().set_attr_index(1); | |||
TensorDesc shape_desc(ge::Shape({32, 3, 224, 224}), FORMAT_NCHW, DT_FLOAT); | |||
data.update_output_desc_y(shape_desc); | |||
TensorDesc desc(ge::Shape({64, 3, 7, 7}), FORMAT_NCHW, DT_FLOAT); | |||
auto var = op::Variable("conv2d_var"); | |||
var.update_output_desc_y(desc); | |||
var.update_input_desc_x(desc); | |||
auto varw1 = op::Variable("conv2d_varw1"); | |||
varw1.update_output_desc_y(desc); | |||
auto conv2d = op::Conv2D("Translate") | |||
.set_input_x(data) | |||
.set_input_filter(var) | |||
.set_attr_strides({1, 1, 2, 2}) | |||
.set_attr_pads({2, 3, 2, 3}) | |||
.set_attr_data_format("NCHW"); | |||
TensorDesc desc_y; | |||
desc_y.SetFormat(FORMAT_NCHW); // shape: 32 64 112 112 | |||
conv2d.update_output_desc_y(desc_y); | |||
TensorDesc desc1(ge::Shape({1, 64, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
auto var1 = op::Variable("bn_var1"); | |||
var1.update_output_desc_y(desc1); | |||
auto var2 = op::Variable("bn_var2"); | |||
var2.update_output_desc_y(desc1); | |||
auto var3 = op::Variable("bn_var3"); | |||
var3.update_output_desc_y(desc1); | |||
auto var4 = op::Variable("bn_var4"); | |||
var4.update_output_desc_y(desc1); | |||
TensorDesc desc2(ge::Shape({2048, 1001}), FORMAT_NCHW, DT_FLOAT); | |||
auto var5 = op::Variable("var5"); | |||
var5.update_output_desc_y(desc2); | |||
auto var6 = op::Variable("var6"); | |||
var6.update_output_desc_y(desc2); | |||
TensorDesc desclabel(ge::Shape({1, 1001, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
auto label1 = op::Variable("label1"); | |||
label1.update_output_desc_y(desclabel); | |||
TensorDesc descmatlabel(ge::Shape({1, 1001, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
auto matvar = op::Variable("matvar"); | |||
matvar.update_output_desc_y(descmatlabel); | |||
auto matvar1 = op::Variable("matvar1"); | |||
matvar1.update_output_desc_y(descmatlabel); | |||
auto bn = op::FusedBatchNorm() | |||
.set_input_x(conv2d, "y") | |||
.set_input_scale(var1) | |||
.set_input_b(var2) | |||
.set_input_mean(var3) | |||
.set_input_variance(var4) | |||
.set_attr_mode(1) | |||
.set_attr_epsilon(1e-5) | |||
.set_attr_is_training(true) | |||
.set_attr_is_training_fusion(true) | |||
.set_attr_moving_average_fraction(994352128); | |||
auto relu = op::Relu().set_input_x(bn, "y"); | |||
auto maxpool = op::MaxPoolWithArgmax() | |||
.set_input_x(relu, "y") | |||
.set_attr_ksize({1, 3, 3, 1}) | |||
.set_attr_padding("SAME") | |||
.set_attr_strides({1, 2, 2, 1}); | |||
MAKE_RESNET50(maxpool); | |||
std::vector<Operator> inputs{data}; //,var,var1,layer1_blk1_bn1_b,var3,var4}; | |||
std::vector<Operator> outputs{}; | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return true; | |||
} | |||
#define GENERATE_CONSTANT_USE_DESC(OPNUM, desc, val) \ | |||
uint32_t OPNUM##_size = desc.GetShape().GetShapeSize(); \ | |||
Tensor OPNUM##_tensor; \ | |||
OPNUM##_tensor.SetTensorDesc(desc); \ | |||
if (desc.GetDataType() == DT_FLOAT) { \ | |||
float *OPNUM##_data = new float[OPNUM##_size]; \ | |||
for (int i = 0; i < (int)OPNUM##_size; i++) { \ | |||
*(OPNUM##_data + i) = val; \ | |||
} \ | |||
OPNUM##_tensor.SetData((uint8_t *)OPNUM##_data, OPNUM##_size * sizeof(float)); \ | |||
delete[] OPNUM##_data; \ | |||
} \ | |||
if (desc.GetDataType() == DT_INT64) { \ | |||
int64_t *OPNUM##_data = new int64_t[OPNUM##_size]; \ | |||
for (int i = 0; i < (int)OPNUM##_size; i++) { \ | |||
*(OPNUM##_data + i) = val; \ | |||
} \ | |||
OPNUM##_tensor.SetData((uint8_t *)OPNUM##_data, OPNUM##_size * sizeof(int64_t)); \ | |||
delete[] OPNUM##_data; \ | |||
} \ | |||
auto OPNUM##_constant = op::Constant().set_attr_value(OPNUM##_tensor); \ | |||
OPNUM##_constant.update_output_desc_y(desc); | |||
#define GENERATE_VAR_LAYER(OPNUM, desc, input) \ | |||
auto OPNUM##_weight = op::Variable(string(#OPNUM)); \ | |||
OPNUM##_weight.update_output_desc_y(desc); \ | |||
auto OPNUM##_assign = op::Assign().set_input_ref(OPNUM##_weight).set_input_value(OPNUM##_constant); \ | |||
\ | |||
input.push_back(OPNUM##_weight); | |||
#define GENERATE_VAR_LAYER_1(OPNUM, desc, var_format, input, name) \ | |||
auto OPNUM##_weight = op::Variable(string(name)); \ | |||
OPNUM##_weight.update_output_desc_y(desc); \ | |||
auto OPNUM##_assign = op::Assign().set_input_ref(OPNUM##_weight).set_input_value(OPNUM##_constant); \ | |||
\ | |||
input.push_back(OPNUM##_weight); | |||
int BuildInitVarGraph(Graph &graph) { | |||
std::vector<Operator> inputs{}; | |||
std::vector<Operator> outputs{}; | |||
TensorDesc desc(ge::Shape({64, 3, 7, 7}), FORMAT_NCHW, DT_FLOAT); | |||
GENERATE_CONSTANT_USE_DESC(conv2d_var, desc, 0.01); | |||
GENERATE_VAR_LAYER(conv2d_var, desc, inputs); | |||
GENERATE_CONSTANT_USE_DESC(conv2d_varw1, desc, 0.01); | |||
GENERATE_VAR_LAYER(conv2d_varw1, desc, inputs); | |||
TensorDesc desc1(ge::Shape({1, 64, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
GENERATE_CONSTANT_USE_DESC(bn_var1, desc1, 0.01); | |||
GENERATE_VAR_LAYER(bn_var1, desc1, inputs); | |||
GENERATE_CONSTANT_USE_DESC(bn_var2, desc1, 0.01); | |||
GENERATE_VAR_LAYER(bn_var2, desc1, inputs); | |||
GENERATE_CONSTANT_USE_DESC(bn_var3, desc1, 0.01); | |||
GENERATE_VAR_LAYER(bn_var3, desc1, inputs); | |||
GENERATE_CONSTANT_USE_DESC(bn_var4, desc1, 0.01); | |||
GENERATE_VAR_LAYER(bn_var4, desc1, inputs); | |||
TensorDesc desc2(ge::Shape({2048, 1001}), FORMAT_NCHW, DT_FLOAT); | |||
GENERATE_CONSTANT_USE_DESC(var5, desc2, 0.01); | |||
GENERATE_VAR_LAYER(var5, desc2, inputs); | |||
GENERATE_CONSTANT_USE_DESC(var6, desc2, 0.01); | |||
GENERATE_VAR_LAYER(var6, desc2, inputs); | |||
TensorDesc desclabel(ge::Shape({1, 1001, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
GENERATE_CONSTANT_USE_DESC(label1, desclabel, 0.1); | |||
GENERATE_VAR_LAYER(label1, desclabel, inputs); | |||
TensorDesc descmatlabel(ge::Shape({1, 1001, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
GENERATE_CONSTANT_USE_DESC(matvar, descmatlabel, 0.01); | |||
GENERATE_VAR_LAYER(matvar, descmatlabel, inputs); | |||
GENERATE_CONSTANT_USE_DESC(matvar1, descmatlabel, 0.01); | |||
GENERATE_VAR_LAYER(matvar1, descmatlabel, inputs); | |||
MAKE_RESNET50_VAR(inputs); | |||
TensorDesc ctrl(ge::Shape({1, 1, 1, 1}), FORMAT_NCHW, DT_INT64); | |||
GENERATE_CONSTANT_USE_DESC(iterations_per_loop, ctrl, 100); | |||
GENERATE_VAR_LAYER_1(iterations_per_loop, ctrl, "4D", inputs, "npu_runconfig/iterations_per_loop"); | |||
GENERATE_CONSTANT_USE_DESC(loop_cond, ctrl, 0); | |||
GENERATE_VAR_LAYER_1(loop_cond, ctrl, "4D", inputs, "npu_runconfig/loop_cond"); | |||
GENERATE_CONSTANT_USE_DESC(one, ctrl, 1); | |||
GENERATE_VAR_LAYER_1(one, ctrl, "4D", inputs, "npu_runconfig/one"); | |||
GENERATE_CONSTANT_USE_DESC(zero, ctrl, 0); | |||
GENERATE_VAR_LAYER_1(zero, ctrl, "4D", inputs, "npu_runconfig/zero"); | |||
graph.SetInputs(inputs).SetOutputs(outputs); | |||
return 0; | |||
} | |||
int TestBuildGraphTest(Func fun, Graph &graph, vector<ge::Tensor> &inputs, vector<ge::Tensor> &outputs) { | |||
bool graph_ret = fun(graph); | |||
ge::Tensor shapeTensor; | |||
TensorDesc shape_desc(ge::Shape({32, 3, 224, 224}), FORMAT_NCHW, DT_FLOAT); | |||
uint32_t sizeshape = shape_desc.GetShape().GetShapeSize(); | |||
printf("[test] desc size filter shape:%u\n", sizeshape); | |||
shapeTensor.SetTensorDesc(shape_desc); | |||
vector<float> dataValuec; | |||
for (int i = 0; i < sizeshape; i++) { | |||
dataValuec.push_back(1); | |||
} | |||
shapeTensor.SetData((uint8_t *)dataValuec.data(), 4 * sizeshape); | |||
inputs.push_back(shapeTensor); | |||
ge::Tensor shapeTensor1; | |||
TensorDesc shape_desc1(ge::Shape({1, 32, 1, 1}), FORMAT_NCHW, DT_FLOAT); | |||
uint32_t sizeshape1 = shape_desc1.GetShape().GetShapeSize(); | |||
printf("[test] desc size filter shape:%u\n", sizeshape1); | |||
shapeTensor1.SetTensorDesc(shape_desc1); | |||
vector<int32_t> dataValuec1; | |||
for (int i = 0; i < sizeshape1; i++) { | |||
dataValuec1.push_back(1); | |||
} | |||
shapeTensor1.SetData((uint8_t *)dataValuec1.data(), 4 * sizeshape1); | |||
return 0; | |||
} | |||
int runTrainGraph(Func fun, int loopCount) { | |||
printf("GE BBIT begin...\n"); | |||
std::chrono::system_clock::time_point start = std::chrono::system_clock::now(); | |||
std::map<std::string, std::string> ge_options = { | |||
{"device_id", "0"}, {"rank_table_file", ""}, {"graphType", "1"}, {"ge.graphRunMode", "2"}}; | |||
std::map<std::string, std::string> session_options = {{"a", "b"}, {TRAIN_FLAG, "1"}}; | |||
ge::Status ret; | |||
// init ge | |||
ret = GEInitialize_api_new("train", "fe,plugin"); | |||
printf("ge::GEInitialize ret:%d\n", ret); | |||
// init session | |||
ge::Session session(session_options); | |||
int graphId_initvar = 1; | |||
ge::Graph graph_initvar("initVarGraph"); | |||
bool graph_ret = BuildInitVarGraph(graph_initvar); | |||
// session addgraph | |||
int graphId = 0; | |||
// build graph | |||
ge::Graph graph("bigGraph"); | |||
std::vector<ge::Tensor> inputs; | |||
ge::Tensor outputTensor; | |||
std::vector<ge::Tensor> outputs; | |||
graph_ret = TestBuildGraphTest(fun, graph, inputs, outputs); | |||
printf("TestReluGrad ret:%d\n", graph_ret); | |||
ret = session.AddGraph(graphId_initvar, graph_initvar); | |||
printf("session.AddVarGraph ret:%d\n", ret); | |||
if (ret) return ret; | |||
ret = session.AddGraph(graphId, graph); | |||
printf("session.AddGraph ret:%d\n", ret); | |||
if (ret) return ret; | |||
std::vector<ge::Tensor> inputs1; | |||
std::vector<ge::Tensor> outputs1; | |||
ret = session.RunGraph(graphId_initvar, inputs1, outputs1); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
// add loop for test of stabilty: | |||
for (int i = 0; i < loopCount; i++) { | |||
// session rungraph | |||
printf("loopCount:%d\n", loopCount); | |||
ret = session.RunGraph(graphId, inputs, outputs); | |||
printf("session.RunGraph ret:%d\n", ret); | |||
if (ret) return ret; | |||
// define 99999 as loop forever | |||
if (loopCount == 99999) i = 0; | |||
} | |||
std::chrono::system_clock::time_point end = std::chrono::system_clock::now(); | |||
auto millisecondsduration = std::chrono::duration_cast<std::chrono::milliseconds>(end - start); | |||
auto ms = millisecondsduration.count(); | |||
std::stringstream ss; | |||
ss << ms << "ms"; | |||
std::string run_time = ss.str(); | |||
printf("run time is : %s \n", run_time.c_str()); | |||
return 0; | |||
} | |||
int main(int argc, char *argv[]) { | |||
// add loop for test of stabilty: | |||
int loopCount = 1; | |||
if (argc >= 2) loopCount = atoi(argv[1]); | |||
Status ret = SUCCESS; | |||
ret = runTrainGraph(resnet50, loopCount); | |||
if (ret == SUCCESS) { | |||
std::cout << "[train resnet50 success]" << std::endl; | |||
} else { | |||
std::cout << "!!! train resnet50 fail !!!" << std::endl; | |||
} | |||
return ret; | |||
} |
@@ -1,56 +0,0 @@ | |||
# Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
# | |||
# Licensed under the Apache License, Version 2.0 (the "License"); | |||
# you may not use this file except in compliance with the License. | |||
# You may obtain a copy of the License at | |||
# | |||
# http://www.apache.org/licenses/LICENSE-2.0 | |||
# | |||
# Unless required by applicable law or agreed to in writing, software | |||
# distributed under the License is distributed on an "AS IS" BASIS, | |||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
# See the License for the specific language governing permissions and | |||
# limitations under the License. | |||
# ============================================================================ | |||
""" | |||
ge st test. | |||
""" | |||
import pytest | |||
import subprocess | |||
import os | |||
@pytest.mark.level0 | |||
@pytest.mark.platform_arm_ascend_training | |||
@pytest.mark.platform_x86_ascend_training | |||
@pytest.mark.env_card | |||
@pytest.mark.component_ge | |||
def test_resnet50_train(): | |||
ge_st_dir=os.environ.get('GE_ST_DIR', | |||
'/home/jenkins/workspace/release_pkg/gate/graphengine_lib') | |||
ge_lib_dir=os.environ.get('GRAPHENGINE_LIB', '/home/jenkins/workspace/release_pkg/gate/graphengine_lib') | |||
real_pythonpath=os.environ.get('REAL_PYTHONPATH') | |||
pythonpath=os.environ.get('PYTHONPATH') | |||
if real_pythonpath: | |||
if pythonpath: | |||
os.environ['PYTHONPATH']=real_pythonpath+':'+pythonpath | |||
else: | |||
os.environ['PYTHONPATH']=real_pythonpath | |||
print('PYTHONPATH: '+os.environ.get('PYTHONPATH')) | |||
os.environ['ASCEND_OPP_PATH']='/usr/local/Ascend/opp' | |||
os.environ['ASCEND_ENGINE_PATH']='/usr/local/Ascend/fwkacllib/lib64/plugin/opskernel/libaicpu_engine.so:' \ | |||
'/usr/local/Ascend/fwkacllib/lib64/plugin/opskernel/libfe.so:' \ | |||
'/usr/local/Ascend/fwkacllib/lib64/plugin/opskernel/librts_engine.so:'+ \ | |||
ge_lib_dir + '/libge_local_engine.so' | |||
print('ASCEND_OPP_PATH: '+os.environ.get('ASCEND_OPP_PATH')) | |||
print('ASCEND_ENGINE_PATH: '+os.environ.get('ASCEND_ENGINE_PATH')) | |||
print('LD_LIBRARY_PATH: '+os.environ.get('LD_LIBRARY_PATH')) | |||
cmd=ge_st_dir + '/st_resnet50_train' | |||
print('cmd: '+cmd) | |||
os.environ['SLOG_PRINT_TO_STDOUT']="1" | |||
ret=subprocess.call([cmd], shell=True) | |||
assert ret==0 | |||