Thursday, January 26, 2012
OpenCL/CUDA for FPGAs
I noticed Altera announcing OpenCL support for FPGAs. Here's a paper and some slides on it. I also noticed a blog pop up last week about doing OpenCL on Xilinx.
There's also been work on a CUDA->FPGA system called FCUDA from a group from UIUC and UCLA (here's the longer paper on FCUDA).
So hardening GPU-designed algorithms is now at least an idea, and possibly a good one. This will enable CMOS array sensors with integrated ASICs performing computer vision.
Subscribe to:
Post Comments (Atom)
2 comments:
Besides Altera a lot of people tried to implement OpenCl for FPGA
http://activedaily.blogspot.com
http://opencl4fpga.blogspot.com/
but I don't know of any success in this field. On the host side there are no problems, but when it comes to translating kernel code to FPGA a lot of problems arise. The best you can do up to date is to use commercial C to RTL compilers like AutoESL (Xilinx), Catapult C or other
The Altera OpenCL compiler works great. I've been doing reconfigurable computing since 1987 and I just started working in the OpenCL group here at Altera. The Altera OpenCL compiler will compile your kernels and build out the whole system as well. I've taken kernels off the net and compiled them with no changes, no problem. OpenCL is a parallel language and the Altera compiler produces parallel and pipelined hardware as well as managing all the data flow from the memories and Pcie interface. It currently compiles several Pcie based boards as well as to the SoC product. As far as I know there is no other system level compiler that does all this except the Altera OpenCL compiler. Cheers, Steve Casselman
Post a Comment