Test Framework Guide【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/geBack to DT Case Development GuideUT/ST development has some test capabilities that are tedious and common. These tasks include: constructing compute graph etc. inputs, taking over rts etc. low-level interfaces, verifying a graphs correctness etc. These code arent difficult but verbose. If directly written in cases, affects case readability and causes lots of duplicate code between cases. Therefore in test framework, provided faker, stub, checker etc. mechanisms, separately for constructing inputs, stubbing low-level interfaces, verifying outputs.This section introduces various capabilities already provided in test framework, letting you quickly know if current test framework already provides this functionality when having a new public capability requirement. Therefore only overview introduction here. If need detailed usage instructions, please refer to corresponding framework capability header files and documentation.Test Base ClassesTest base classes inherit from gteststesting::Test, and do some common initialization and cleanup actions in SetUp and TearDown, to reduce case redundant code. Currently encapsulated test base classes:BgTest: No setup processing, TearDown automatically clears all ValueHolders FrameBgTestAutoCreateFrame: Setup automatically creates a root Frame, TearDown automatically clears all ValueHolders FrameBgTestAutoCreate3StageFrame: Setup automatically creates Init, Main, DeInit three node graphs, and selects current Frame as Main graph Frame; TearDown automatically clears all ValueHolders FrameAbove base class header files:#include common/bg_test.hFaker IntroductionFaker used to construct some data, these data can conveniently be used for UT or ST testing. Currently supported fakers:Graph DSL (Declarative Graph Construction)#include ge_graph_dsl/graph_dsl.hGraph DSL based on easy_graph library encapsulation, provides macro-driven graph building approach, replacing verbose manual graph construction code. Core capabilities:Core macros:DEF_GRAPH(name) { ... }defines graph,CHAIN(...)defines a node-edge-node chain,NODE(name, op_type)creates node,EDGE(src_out_idx, dst_in_idx)creates data edgeOperator configuration:OP_CFG(op_type)creates streaming operator description configuration, supports.TensorDesc(),.Weight(),.Attr(),.InputAttr(),.OutputAttr(),.InCnt(),.OutCnt()etc. chain callsGraph conversion:ToGeGraph(graph)→ge::Graph,ToComputeGraph(graph)→ComputeGraphPtr,ToExecuteGraph(graph)→ExecuteGraphPtrGraph assertion:DUMP_GRAPH_WHEN(phase1, ...)captures graph snapshot at specified compilation phase,CHECK_GRAPH(phase_id) { ... }executes assertion verification on snapshot (detailed usage see ST Development Guide)Usage example:#include ge_graph_dsl/graph_dsl.h #include ge_graph_dsl/assert/graph_assert.h TEST_F(MyUT, BuildAndCheckGraph) { // Configure operators auto data_cfg OP_CFG(DATA).TensorDesc(FORMAT_NCHW, DT_FLOAT, {1, 224, 224, 224}); auto add_cfg OP_CFG(ADD).TensorDesc(FORMAT_NCHW, DT_FLOAT, {1, 224, 224, 224}); // Declarative graph construction DEF_GRAPH(g) { CHAIN(NODE(data_0, data_cfg)-EDGE(0, 0)-NODE(add, add_cfg)); CHAIN(NODE(data_1, data_cfg)-EDGE(0, 1)-NODE(add, add_cfg)); CHAIN(NODE(add, add_cfg)-NODE(netoutput, NETOUTPUT)); }; // Convert to GE graph auto compute_graph ToComputeGraph(g); ASSERT_NE(compute_graph, nullptr); }Other FakersNodeFaker: Fake a ge::NodePtr object#include faker/node_faker.hTensorFaker: Fake one or multiple gert::Tensor#include faker/fake_value.hKernelRunContextFaker: Fake KernelContext, this is a series of faker, besides regular KernelContext, can also fake TilingContext, InferShapeContext#include faker/kernel_run_context_facker.hGeModelBuilder: Accept a compute graph object, fake it as Model#include faker/ge_model_builder.hModelDataFaker: Accept a model, fake it as ModelData#include faker/model_data_faker.hGlobalDataFaker: Fake a LoweringGlobalData#include faker/global_data_faker.hMagicOpFaker: Fake an operators complete implementation, including Lowering and complete kernel#include faker/magic_ops.hStub IntroductionStub stubs and takes over runtime environment, e.g., slog, rts etc., also provides interfaces for cases to verify GEs calling behavior correctness from log, rts perspective. For compile time, less dependency on low-level, therefore almost no need for stubbing.For execution time, stub has a package solution:GertRuntimeStub(#include stub/gert_runtime_stub.h). Once this class instantiated, automatically stubs all rts interfaces, takes over slog to listen logs, takes over node converter and kernel registry. Then in case can operate this instance (only affects this case behavior, wont affect global other cases).For example in some case, we expect to verify a log:TEST_F(KernelLogUT, KLog_Success_LogError) { GertRuntimeStub stub; auto context_holder KernelRunContextFaker().KernelName(tn).KernelType(tt).Build(); auto context context_holder.GetContextKernelContext(); stub.GetSlogStub().Clear(); KLOGE(Hello world); KLOGE(Hello world %d, 123); ASSERT_EQ(stub.GetSlogStub().GetLogs().size(), 2); ASSERT_EQ(stub.GetSlogStub().GetLogs().at(0).level, DLOG_ERROR); ASSERT_EQ(stub.GetSlogStub().GetLogs().at(1).level, DLOG_ERROR); ASSERT_ENDSWITH(stub.GetSlogStub().GetLogs().at(0).content, [tt][tn]Hello world); ASSERT_ENDSWITH(stub.GetSlogStub().GetLogs().at(1).content, [tt][tn]Hello world 123); }Checker IntroductionChecker mainly concentrated in graph verification. Currently provided checker capabilities:SummaryChecker: Verify a graphs overall info, e.g., how many nodes, count for each node type#include common/summary_checker.hTopoChecker: Verify topology info, can conveniently verify how many outputs or inputs some node has, can chain verify a string of nodes edge relationships#include common/topo_checker.hTest Framework ImprovementTest framework also maintained by developers. And currently test framework not assigned as separate responsibility field, therefore encourage anyone to actively improve test framework. Improvement items include but not limited to: new faker/stub/checker etc. framework capabilities, fix existing test framework bugs, improve usability etc.【免费下载链接】geGEGraph Engine是面向昇腾的图编译器和执行器提供了计算图优化、多流并行、内存复用和模型下沉等技术手段加速模型执行效率减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考