A.L. Lohvinenko, I.Y.Shubin

Kharkov national university of radioelectronics

Acceleration of multi-core multimedia converters

Video converters are computer programs that can change the storage format of digital video. They may recompress the video to another format in a process called transcoding, or may simply change the container format without changing the video format. One of the disadvantages to transcoding is that the process is highly CPU intensive. As a result conversion may take hours.

Parallelization has become the solution of the too long conversion.

We all used to think that CPU is the only component of a computer that can execute any code. However, it appears that producing and constructing multi-core processors too difficult and expensive. Therefore, the increasing number of cores significantly slowed down. GPUs, devoid of their CPU deficiencies, have become the excellent and very fast counting machines.

Unlike the CPU is able to solve any problem including graphics, but with the averaged performance graphics processor is adapted to high-speed solution of one problem: converting the input polygons piles into a pile on the output pixels. Moreover, this problem can be solved in parallel on hundreds of relatively simple cores in the GPU. A schedule of productivity CPU and GPU is shown in Figure 1.

Figure 1 – CPU vs GPU

Theoretical performance video chip far superior CPU.

CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model created by NVIDIA and implemented by the graphics processing units (GPUs) that they produce. CUDA gives developers access to the virtual instruction set and memory of the parallel computational elements in CUDA GPUs. Using CUDA, the latest Nvidia GPUs become accessible for computation like CPUs.

CUDA work is this: the program is executed on the CPU as long as it appears the piece of code that can be executed in parallel. Then, instead of being carried out slowly in two (yes even if eight) nuclei the steepest CPU, it is transmitted to the hundreds of cores GPU. At execution time of this portion is reduced in times, and hence reduced time and the entire program.

Video - a set of frames that show consistently at a certain speed. The process of encoding video is divided into:

- reducing the redundancy of video in the time dimension, based on the fact that neighboring frames tend to be not much different;

- compression of individual images.

Compression algorithm:

- preparation of macroblocks. For each macroblocks determines how it will be compressed;

- translation macroblocks in the color space YUV. Getting the right amount of 8x8 matrices;

- for P-blocks and B-blocks produced by calculating the difference between the corresponding macroblocks in the reference frame;

- discrete cosine transform (DCT);

- quantization;

- zigzag scan;

- batch encoding;

- Huffman coding.

View this generalized algorithm can notice that it is subject to parallelization.

Image contains macroblocks, which can encode and decode whatever. Each macroblocks contains six blocks of data to DCT. Parallelize DCT is very important because, in addition to finding displacement vectors is very slow operation. The result is a parallel-pipelined processing circuit video stream.

Discrete cosine transform is the kind of Fourier transform. Code fast Fourier transform (FFT) for CPU is shown in Figure 2.

Figure 2 – FFT core code for CPU

Code fast Fourier transform (FFT) for GPU is shown in Figure 3.

Figure 3 – FFT core code for GPU

With the implementation for the GPU we have:

- GPU code makes two parallel inner loop of the CPU code. This results in increased performance for large data sequences.

- simplified algorithm, internal loops replaced by the condition IF-ELSE.

Using DCT algorithm on GPU get better performance for large matrices (> 256 * 256). Productivity of handle different amounts of data on the CPU and the GPU is shown in Figure 4.

Figure 4 - Productivity of handle different amounts of data on the CPU and the GPU

Dependence of the code on the CPU and GPU on the size of the matrix is shown in Figure 5.

Figure 5 - Dependence of the code on the CPU and GPU on the size of the matrix

Speed up is shown in Figure 6.

Figure 6 – Speed up (FFT based algorithm: GPU vs CPU)

GPU has reached a point of development where many real world applications are easily implemented on them and faster than on multi-core systems.