Reserved APIs

aclOpExecutor()

Indicates the context structure, used to record the running information of all APIs on the host. Almost all operations of the APIs on the host use this data structure as the medium. This function is the constructor.

CreateView(const aclTensor *tensor, const op::Shape &shape, int64_t offset)

Creates a view tensor for an existing aclTensor. Two tensors share the memory. You can specify the shape and offset of the view tensor.

UpdateTensorAddr(void *workspaceAddr, const size_t size)

Updates the address of each workspace after workspace addresses are allocated.

GetWorkspaceAddr()

Obtains the start pointer of a workspace.

GetWorkspaceSize()

Obtains the size of the required workspace.

GetLinearWorkspaceSize()

This API is deprecated and can be ignored.

GetWorkspaceOffsets()

Obtains the workspace offsets.

SetWorkspaceOffsets(const op::FVector<uint64_t> &workspaceOffsets)

Sets the workspace offsets to be recorded.

Run()

Runs aclOpExecutor to execute tasks in the queue.

GetStream()

Obtains the current stream.

GetInputTensors()

Obtains the input aclTensor of the API on the host.

GetOutputTensors()

Obtains the output aclTensor of the API on the host.

GetLogInfo()

Obtains log information stored in aclOpExecutor.

SetLogInfo(const op::internal::OpLogInfo &logInfo)

GetOpConfigInfo()

Obtains the configuration information related to operator running.

SetOpConfigInfo(const op::OpConfigInfo &opConfigInfo)

Sets the configuration information related to operator running.

SetStream(aclrtStream stream)

Sets the current stream.

AddTensorRelation(const aclTensor *tensorOut, const aclTensor *tensorMiddle)

Records the address equivalence relationship between two aclTensors for aclnn cache.

UpdateStorageAddr()

Updates the aclTensor address in the aclOpExecutor reuse scenario.

SetRepeatable()

Attempts to set the current aclOpExecutor to the reusable state.

IsRepeatable()

Checks whether the current aclOpExecutor is reusable.

FinalizeCache()

Saves the final data of the aclnn cache.

RepeatRunWithCache(void *workspaceAddr, const aclrtStream stream)

Runs the aclnn cache to execute tasks in the aclOpExecutor reuse scenario.

CheckLauncherRepeatable()

Checks whether aclOpExecutor reuse is allowed for each task in the aclOpExecutor task list.

AddCache()

Saves the aclnn cache in aclOpExecutor to global management.

DeleteCache()

Deletes the aclnn cache.

GetOpExecCache()

Obtains the aclnn cache recorded in aclOpExecutor.

SetIOTensorList()

Sets the input/output aclTensor of the API on the host.

GetGraph()

Obtains the execution graph of the API on the host.

GetMagicNumber()

Obtains the magic number to distinguish different objects.

UniqueExecutor(const char *funcName)

Indicates the constructor factory of aclOpExecutor.

UniqueExecutor()

get()

Obtains the aclOpExecutor pointer in UniqueExecutor.

ReleaseTo(aclOpExecutor **executor)

Passes the aclOpExecutor pointer in UniqueExecutor to the destination aclOpExecutor pointer.

UniqueExecutor(const UniqueExecutor &)

This API is deprecated and can be ignored.

GetOpExecCacheFromExecutor(aclOpExecutor *)

Converts an external aclOpExecutor to an aclnn cache object.

InitL2Phase1Context(const char *l2Name, [[maybe_unused]] aclOpExecutor **executor)

Initializes some DFX variable values in the first-phase API on the host.

InitL2Phase2Context([[maybe_unused]] const char* l2Name, aclOpExecutor* executor)

Initializes some DFX variable values in the second-phase API on the host.

InitL0Context(const char *profilingName, aclOpExecutor* executor)

Initializes some DFX variable values of the L0 API.

CreatAiCoreKernelLauncher([[maybe_unused]] const char *l0Name, uint32_t opType, aclOpExecutor *executor, op::OpArgContext *args)

Creates an AI Core task object.

CreatDSAKernelLauncher([[maybe_unused]] const char *l0Name, uint32_t opType, DSA_TASK_TYPE dsaTask, aclOpExecutor *executor, op::OpArgContext *args)

Creates a DSA task object.

InferShape(uint32_t optype, op::OpArgList &inputs, op::OpArgList &outputs, op::OpArgList &attrs)

Obtains the shape inference result of a specified operator.

ConvertToTensor(const aclIntArray *value, op::DataType dataType)

Converts different types of host data into a host aclTensor.

ConvertToTensor(const aclBoolArray *value, op::DataType dataType)

ConvertToTensor(const aclFloatArray *value, op::DataType dataType)

ConvertToTensor(const aclFp16Array *value, op::DataType dataType)

ConvertToTensor(const aclBf16Array *value, op::DataType dataType)

ConvertToTensor(const T *value, uint64_t size, op::DataType dataType)

ConvertToTensor(const aclScalar *value, op::DataType dataType)

Converts different types of host data into a host aclTensor.

AddToKernelLauncherList(op::KernelLauncher *obj)

Adds an AI Core task to an execution queue.

AddToKernelLauncherListDvpp(uint32_t opType, op::KernelLauncher *obj, op::OpArgContext *args)

Adds a DVPP task to an execution queue.

AddToKernelLauncherListCopyTask(uint32_t opType, op::KernelLauncher *obj, op::OpArgList &inputs, op::OpArgList &outputs, op::OpArgList &workspace)

Adds a data copy task to an execution queue.

AddToKernelLauncherListAiCpu(int32_t opType, op::KernelLauncher *obj, op::OpArgContext *args)

Adds an AI CPU task to an execution queue.

CommonOpExecutorRun(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)

Runs all tasks in the context based on the external aclOpExecutor, workspace, and stream.