Dynamic Shape Input (Setting the Shape Range)

//1. Initialize AscendCL.

//2. Allocate runtime resources.

//In this sample, assume that the first input of the model is dynamic with an index of 0, and the first output of the model is dynamic with also an index of 0.

//3. Load the model. After the model is successfully loaded, set the tensor description for dynamic input, mainly the shape information.
// ......

//4. Prepare the model description modelDesc_, and the model inputs input_ and model outputs output_.
//Pay attention to the following:
//If the size obtained by using aclmdlGetInputSizeByIndex is 0, the input shape is dynamic. A larger input buffer can be estimated based on the actual situation.
//If the size obtained by using aclmdlGetOutputSizeByIndex is 0, the output shape is dynamic. A larger output buffer can be estimated based on the actual situation.
// ......

//5. Customize a function to set the description of the dynamic input tensor.
void SetTensorDesc()
{
        // ......
    //Create the tensor description.
    //The shape must be the same as that of the specified input data.
	int64_t shapes = {1, 3, 224, 224};
	aclTensorDesc *inputDesc = aclCreateTensorDesc(ACL_FLOAT, 4, shapes, ACL_FORMAT_NCHW);
	//Set the description of the dynamic input tensor whose index is 0.
	aclError ret = aclmdlSetDatasetTensorDesc(input_, inputDesc, 0);
        // ......
}
//6. Customize a function, execute the model, and obtain the description of the dynamic output tensor.
void ModelExecute()
{
	aclError ret;
	//Call a custom API to set the description of the dynamic input tensor.
	SetTensorDesc();
        //Execute the model.
	ret = aclmdlExecute(modelId, input_, output_);
	//Obtain the description of the dynamic output tensor whose index is 0.
        aclTensorDesc *outputDesc = aclmdlGetDatasetTensorDesc(output_, 0);
	//Use the operation API of the aclTensorDesc type to obtain the outputDesc attribute. You need to obtain size (size occupied by tensor data) as an example: Then, read the data of the corresponding size from the memory.
	string outputFileName = ss.str();
        FILE *outputFile = fopen(outputFileName.c_str(), "wb");
	size_t outputDesc_size = aclGetTensorDescSize(outputDesc);
        aclDataBuffer *dataBuffer = aclmdlGetDatasetBuffer(output_, 0);
        void *data = aclGetDataBufferAddr(dataBuffer);
	void *outHostData = nullptr;
        //Call aclrtGetRunMode to obtain the run mode of the software stack and determine whether to transfer data, based on the run mode.
        aclrtRunMode runMode;
        ret = aclrtGetRunMode(&runMode);
        if (runMode == ACL_HOST) {
            ret = aclrtMallocHost(&outHostData, outputDesc_size);
	   //Since the memory allocated for the dynamic shape is large, the actual data size outputDesc_size is used to copy the memory.
	    ret = aclrtMemcpy(outHostData, outputDesc_size, data, outputDesc_size, ACL_MEMCPY_DEVICE_TO_HOST);
	    fwrite(outHostData, outputDesc_size, sizeof(char), outputFile);
            ret = aclrtFreeHost(outHostData);
	} else {
		// if app is running in host, write model output data into result file
                fwrite(data, outputDesc_size, sizeof(char), outputFile);
	}
	fclose(outputFile);
        // ......
}

//7. Process the model inference result.

//8. Deallocate runtime resources.

//9. Deinitialize the AscendCL.

// ......

//1. Initialize AscendCL.

//2. Allocate runtime resources.

//The sample code is for reference only. In actual application scenarios, verify the return value.
//In this sample, assume that the first input of the model is dynamic with an index of 0, and the first output of the model is dynamic with also an index of 0.

//3. Load the model. After the model is successfully loaded, set the tensor description for dynamic input, mainly the shape information.
// ...

//4. Prepare the model description modelDesc_, and the model inputs input_ and model outputs output_.
//Pay attention to the following:
//If the size obtained by using aclmdlGetInputSizeByIndex is 0, the input shape is dynamic. A larger input buffer can be estimated based on the actual situation.
//If the size obtained by using aclmdlGetOutputSizeByIndex is 0, the output shape is dynamic. A larger output buffer can be estimated based on the actual situation.
// ......

//5. Customize a function to set the description of the dynamic input tensor.
void SetTensorDesc()
{
        // ......
    //Create the tensor description.
    //The shape must be the same as that of the specified input data.
	int64_t shapes = {1, 3, 224, 224};
	aclTensorDesc *inputDesc = aclCreateTensorDesc(ACL_FLOAT, 4, shapes, ACL_FORMAT_NCHW);
	//Set the description of the dynamic input tensor whose index is 0.
	aclError ret = aclmdlSetDatasetTensorDesc(input_, inputDesc, 0);
        // ......
}

//6. Create the Allocator descriptor and register the Allocator.
//Assume that the allocator is the Allocator object that the user wants to register.
void ResgisterCustomAllocator(aclrtAllocator allocator, aclrtStream stream) {
    //6.1 Create AllocatorDesc.
    aclrtAllocatorDesc allocatorDesc = aclrtAllocatorCreateDesc();
    //6.2 Initialize AllocatorDesc and set callback functions related to Allocator memory allocation and deallocation. CustomMallocFunc, CustomFreeFunc, CustomMallocAdviseFunc, and CustomGetBlockAddrFunc are callback functions defined in the C style and are transferred to AllocatorDesc as function pointers.
    aclrtAllocatorSetObjToDesc(allocatorDesc, allocator);
    aclrtAllocatorSetAllocFuncToDesc(allocatorDesc, CustomMallocFunc);
    aclrtAllocatorSetFreeFuncToDesc(allocatorDesc, CustomFreeFunc);
    aclrtAllocatorSetAllocAdviseFuncToDesc(allocatorDesc, CustomMallocAdviseFunc);
    aclrtAllocatorSetGetAddrFromBlockFuncToDesc(allocatorDesc, CustomGetBlockAddrFunc);
    //Register the Allocator and bind it to the stream. The API creates the Allocator based on AllocatorDesc.
    aclrtAllocatorRegister(stream, allocatorDesc);
    //After the Allocator descriptor is used, it can be destroyed.
    aclrtAllocatorDestroyDesc(allocatorDesc);
}


//7. Customize a function, execute the model, and obtain the description of the dynamic output tensor.
void ModelExecute(aclrtStream stream)
{
	aclError ret;
	//Call a custom API to set the description of the dynamic input tensor.
	SetTensorDesc();
       //Execute the model. Only asynchronous APIs support the Allocator registered by the user, and the execution stream must be the same as the stream bound to the registered Allocator.
	ret = aclmdlExecuteAsync(modelId, input_, output_, stream);
       //If the asynchronous API is successfully called, the task is successfully delivered. Before obtaining the result, call the call synchronization API to ensure that the task has been executed.
        aclrtSynchronizeStream(stream);
       //Obtain the description of the dynamic output tensor whose index is 0.
        aclTensorDesc *outputDesc = aclmdlGetDatasetTensorDesc(output_, 0);
       //Use the operation API of the aclTensorDesc type to obtain the outputDesc attribute. You need to obtain size (size occupied by tensor data) as an example: Then, read the data of the corresponding size from the memory.
	string outputFileName = ss.str();
        FILE *outputFile = fopen(outputFileName.c_str(), "wb");
	size_t outputDesc_size = aclGetTensorDescSize(outputDesc);
        aclDataBuffer *dataBuffer = aclmdlGetDatasetBuffer(output_, 0);
        void *data = aclGetDataBufferAddr(dataBuffer);
	void *outHostData = nullptr;
       //Call aclrtGetRunMode to obtain the run mode of the software stack and determine whether to transfer data, based on the run mode.
        aclrtRunMode runMode;
        ret = aclrtGetRunMode(&runMode);
        if (runMode == ACL_HOST) {
            ret = aclrtMallocHost(&outHostData, outputDesc_size);
           //Since the memory allocated for the dynamic shape is large, the actual data size outputDesc_size is used to copy the memory.
	    ret = aclrtMemcpy(outHostData, outputDesc_size, data, outputDesc_size, ACL_MEMCPY_DEVICE_TO_HOST);
	    fwrite(outHostData, outputDesc_size, sizeof(char), outputFile);
            ret = aclrtFreeHost(outHostData);
	    } else {
		// if app is running in host, write model output data into result file
                fwrite(data, outputDesc_size, sizeof(char), outputFile);
	    }
	fclose(outputFile);
        // ......
}

//8. Process the model inference result.
// ...

//9. Deregister and destroy the registered Allocator.
void UnresgisterCustomAllocator(aclrtStream stream) {
     aclrtAllocatorUnregister(stream);
    //Destroy the customized Allocator.
}

//10. Deallocate runtime resources.

//11. Deinitialize AscendCL.

// ......