Vitis AI Examples - 1.4.1 English

Vitis AI provides several C++ and Python examples to demonstrate the use of the unified cloud-edge runtime programming APIs.

Note: The sample code helps you get started with the new runtime (VART). They are not meant for performance benchmarking.

To familiarize yourself with the unified APIs, use the VART examples. These examples are only to understand the APIs and do not provide high performance. These APIs are compatible between the edge and cloud, though cloud boards may have different software optimizations such as batching and on the edge would require multi-threading to achieve higher performance. If you desire higher performance, see the Vitis AI Library samples and demo software.

If you want to do optimizations to achieve high performance, here are some suggestions:

Rearrange the thread pipeline structure so that every DPU thread has its own "DPU" runner object.
Optimize display thread so that when DPU FPS is higher than display rate, skipping some frames. 200 FPS is too high for video display.
Pre-decoding. The video file might be H.264 encoded. The decoder is slower than the DPU and consumes a lot of CPU resources. The video file has to be first decoded and transformed into raw format.
The batch mode on Alveo boards needs special consideration as it may cause video frame jittering. ZCU102 has no batch mode support.
OpenCV cv::imshow is slow, so you need to use libdrm.so. This is only for local display, not through X server.

The following table below describes these Vitis AI examples.

Table 1. Vitis AI Examples
ID	Example Name	Models	Framework	Notes
1	resnet50	ResNet-50	Caffe	Image classification with Vitis AI unified C++ APIs.
2	resnet50_pt	ResNet-50	PyTorch	Image classification with Vitis AI unified extension C++ APIs.
3	resnet50_ext	ResNet-50	Caffe	Image classification with Vitis AI unified extension C++ APIs.
4	resnet50_mt_py	ResNet-50	TensorFlow	Multi-threading image classification with Vitis AI unified Python APIs.
5	inception_v1_mt_py	Inception-v1	TensorFlow	Multi-threading image classification with Vitis AI unified Python APIs.
6	pose_detection	SSD, Pose detection	Caffe	Pose detection with Vitis AI unified C++ APIs.
7	video_analysis	SSD	Caffe	Traffic detection with Vitis AI unified C++ APIs.
8	adas_detection	YOLOv3	Caffe	ADAS detection with Vitis AI unified C++ APIs.
9	segmentation	FPN	Caffe	Semantic segmentation with Vitis AI unified C++ APIs.
10	squeezenet_pytorch	Squeezenet	PyTorch	Image classification with Vitis AI unified C++ APIs.

The typical code snippet to deploy models with Vitis AI unified C++ high-level APIs is as follows:

// get dpu subgraph by parsing model file
auto runner = vart::Runner::create_runner(subgraph, "run");
// get input scale and output scale,
// they are used for fixed-floating point conversion
auto outputTensors = runner->get_output_tensors();
auto inputTensors = runner->get_input_tensors();
auto input_scale = get_input_scale(inputTensors[0]);
auto output_scale = get_output_scale(outputTensors[0]);
// do the image pre-process, convert float data to fixed point data
// populate input/output tensors
auto job_id = runner->execute_async(inputsPtr, outputsPtr);
runner->wait(job_id.first, -1);
// process outputs, convert fixed point data to float data

The typical code snippet to deploy models with Vitis AI unified extension C++ high-level APIs is as follows:

// get dpu subgraph by parsing model file
std::unique_ptr<vart::RunnerExt> runner =
          vart::RunnerExt::create_runner(subgraph, attrs.get());
// get input & output tensor buffers
auto input_tensor_buffers = runner->get_inputs();
auto output_tensor_buffers = runner->get_outputs();
// get input scale and output scale,
// they are used for fixed-floating point conversion
auto input_tensor = input_tensor_buffers[0]->get_tensor();
auto output_tensor = output_tensor_buffers[0]->get_tensor();
auto input_scale = get_input_scale(input_tensor);
auto output_scale = get_output_scale(output_tensor);
// do the image pre-process, convert float data to fixed point data
setImageBGR(images[batch_idx], (void*)data_in, input_scale);
// sync data for input
input->sync_for_write(0, input->get_tensor()->get_data_size() /
                         input->get_tensor()->get_shape()[0]); 
// populate input/output tensors
auto v = runner->execute_async(input_tensor_buffers, output_tensor_buffers);
auto status = runner->wait((int)v.first, -1);
// sync data for output
output->sync_for_read(0, output->get_tensor()->get_data_size() /
                         output->get_tensor()->get_shape()[0]);
// process outputs, conver fixed point data to float data

The typical code snippet to deploy models with Vitis AI unified Python high-level APIs is shown below:

dpu_runner = runner.Runner(subgraph，"run")
# populate input/output tensors
jid = dpu_runner.execute_async(fpgaInput, fpgaOutput)
dpu_runner.wait(jid)
# process fpgaOutput

Note: DPU processes only work with the input and output of fixed-point data. For improved performance and more efficient memory usage, use int8 data as input and run the float to fixed-point conversion along with preprocessing. If the input data is float, the VART converts the float data to fixed-point data which consumes more time.