Talk 5.4: 1:30PM–2:45PM
| Session Name: | Computer Architectures, Compilers and Programming Models |
| Session Time: | 1:30PM–2:45PM |
| Author Name: | Jeff Kingyens |
| Author Email: | jeff.kingyens@utoronto.ca |
| Talk Title: | Automatic Customization and Generation of an FPGA-based Stream Processor |
| Slides: | 5-4.ppt |
| Abstract: | Personal computers equipped with FPGA coprocessors such as Xtremedata Inc's XD1000 make it easier for developers to integrate FPGA-based acceleration into scientific, financial and data mining workloads. For this acceleration to be accessible to an even broader range of developers, soft IP should automatically be synthesized. While behavioral synthesis tools such as TridentC succeed in transforming C functions into hardware datapaths and control units, other necessary soft IP such as host interface and the memory hierarchy must still be manually designed. We propose to extend TridentC to automatically generate a complete, stream-based architecture targeting the XD1000 system. We implement a data-parallel streaming programming model similar to those used in GPU programming where a data-parallel kernel function to be executed on the coprocessor is described in C and an API executing on the host processor is used to control kernel execution and block memory transfers. |
| Research Group: | Computer |
| Degree Program: | M.A.Sc. |
| Author Bio: | Biography not submitted. |