aclSetInputTensorAddr

Description

After aclOpExecutor reuse is enabled by calling aclSetAclOpExecutorRepeatable, if the input device memory address changes, the device memory address recorded in the input aclTensor needs to be updated.

Prototype

aclnnStatus aclSetInputTensorAddr(aclOpExecutor *executor, const size_t index, aclTensor *tensor, void *addr)

Parameter Description

Parameter	Input/Output	Description
executor	Input	aclOpExecutor that is set to the reusable state.
index:	Input	Index of the input aclTensor to be updated. The value range is [0, total number of input tensors – 1].
tensor	Input	aclTensor pointer to be updated.
addr	Input	Device storage address to be updated to the specified aclTensor. The address must be 32-byte aligned. Otherwise, an undefined error may occur.

Return Value

0 on success; otherwise, failure. For details about the return codes, see Common API Return Codes.

Possible causes:

Error code 561103: executor or tensor is a null pointer.
Error code 161002: The index value is out of range.
Error code 161002: The input aclTensor when the first-phase API aclxxXxxGetWorkspaceSize is called for the first time is a null pointer. The address cannot be updated.

Constraints

None.

Example

The following sample code is for reference only. Do not copy and run it.

// Create the input and output aclTensor and aclTensorList.
std::vector<int64_t> shape = {1, 2, 3};
aclTensor tensor1 = aclCreateTensor(shape.data(), shape.size(), aclDataType::ACL_FLOAT,
nullptr, 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), nullptr);
aclTensor tensor2 = aclCreateTensor(shape.data(), shape.size(), aclDataType::ACL_FLOAT,
nullptr, 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), nullptr);
aclTensor tensor3 = aclCreateTensor(shape.data(), shape.size(), aclDataType::ACL_FLOAT,
nullptr, 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), nullptr);
aclTensor output = aclCreateTensor(shape.data(), shape.size(), aclDataType::ACL_FLOAT,
nullptr, 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), nullptr);
aclTensor *list[] = {tensor1, tensor2};
auto tensorList = aclCreateTensorList(list, 2);
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
// The AddCustom operator has two inputs (aclTensorList and aclTensor) and one output (aclTensor).
// Call the first-phase API.
aclnnAddCustomGetWorkspaceSize(tensorList, tensor3, output, &workspaceSize, &executor);     
// Set the executor to be reusable.
aclSetAclOpExecutorRepeatable(executor);                                                      
void *addr;
aclSetInputTensorAddr(executor, 0, tensor1, addr);   // Update the device address of the first aclTensor in the input tensor list.
aclSetInputTensorAddr(executor, 1, tensor2, addr);  // Update the device address of the second aclTensor in the input tensor list.
aclSetInputTensorAddr(executor, 2, tensor3, addr);  // Update the device address of the input aclTensor.
...
// Call the second-phase API.
aclnnAddCustom(workspace, workspaceSize, executor, stream);          
// Clear the executor.
aclDestroyAclOpExecutor(executor);                                  

Parent topic: Common APIs