Performance of HEVC discrete cosine and sine transforms on GPU using CUDA
| Autor: | Moustafa Masoumi, HamidReza Ahmadifar |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
Lossless compression
Computer science 020208 electrical & electronic engineering 02 engineering and technology Parallel computing CUDA Discrete sine transform 0202 electrical engineering electronic engineering information engineering Benchmark (computing) Discrete cosine transform 020201 artificial intelligence & image processing SIMD Quantization (image processing) Encoder |
| Zdroj: | 2017 IEEE 4th International Conference on Knowledge-Based Engineering and Innovation (KBEI). |
| DOI: | 10.1109/kbei.2017.8324916 |
| Popis: | High Efficiency Video Coding (HEVC) is the most recent video encoding standard that achieves much higher compression while maintaining the same quality as its predecessor x264. This is done by a variety of upgrades for example a more diverse transform block size while some fundamental methods remain unchanged. Discrete Cosine Transform (DCT) along with an alternate Discrete Sine Transform (DST) is used by HEVC before quantization to achieve a lossless encoding of transform blocks. Both of these transforms are implemented in HEVC encoder using CPU. In this paper we implement these transforms on GPU using CUDA and analyze their performance compared to the default implementation. Using a batched kernel call 48% and 13% improvement is achieved on Inverse 32×32 and 16×16 DCT transform respectively compared to SIMD while our GPU implementation outperforms other CPU implementations on every transform. We also present a benchmark software for this very purpose which we will use to measure their performance real time. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|