Fraunhofer-Gesellschaft

Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Sample-parallel execution of EBCOT in fast mode

 
: Bruns, V.; Martinez-Del-Amor, M.Á.

:

Institute of Electrical and Electronics Engineers -IEEE-; IEEE Signal Processing Society:
Picture Coding Symposium, PCS 2016. Proceedings : 4th-7th December 2016, Nürnberg, Germany
Piscataway, NJ: IEEE, 2016
ISBN: 978-1-5090-5966-9
ISBN: 978-1-5090-5965-2
ISBN: 978-1-5090-5967-6 (Print)
S.393-397
Picture Coding Symposium (PCS) <32, 2016, Nürnberg>
Englisch
Konferenzbeitrag
Fraunhofer IIS ()

Abstract
JPEG 2000's most computationally expensive building block is the Embedded Block Coder with Optimized Truncation (EBCOT). This paper evaluates how encoders targeting a parallel architecture such as a GPU can increase their throughput in use cases where very high data rates are used. The compression efficiency in the less significant bit-planes is then often poor and it is beneficial to enable the Selective Arithmetic Coding Bypass style (fast mode) in order to trade a small loss in compression efficiency for a reduction of the computational complexity. More importantly, this style exposes a more finely grained parallelism that can be exploited to execute the raw coding passes, including bit-stuffing, in a sample-parallel fashion. For a latency-or memory critical application that encodes one frame at a time, EBCOT's tier-1 is sped up between 1.1x and 2.4x compared to an optimized GPU-based implementation. When a low GPU occupancy has already been addressed by encoding multiple frames in parallel, the throughput can still be improved by 5% for high-entropy images and 27% for low-entropy images. Best results are obtained when enabling the fast mode after the fourth significant bit-plane. For most of the test images the compression rate is within 1% of the original.

: http://publica.fraunhofer.de/dokumente/N-502659.html