All-
I’m looking for advanced developers who are interested in taking our
OpenMP accelerator and creating a demo GNU radio
application.
The accelerator is a 2.5 Teraflop (32-core) PCIe card with a 1 GBe port
and is programmed via OpenMP. The objectives
are twofold:
-fast response / low latency: C/C++ critical
response code sections are marked by OpenMP
pragmas and run on the card; data doesn’t
reach host sw
-acceleration: designated code sections are
marked by OpenMP pragmas and accelerated
on the card (data has already reached host
sw)
The latter category would not provide performance advantage over a GPU
board, unless multiple blocks can be enclosed
with a pragma section and run together, and or multiple sections run
concurrently (main idea being to limit frequency
and increase size of PCIe bus data transfers).
From an Ettus hardware standpoint, the card would look like an Ethernet
pass-thru for data, so UHD / standard drivers
would still be used.
Yes I have heard there is OpenMP support for Python, but our “source
code parser” doesn’t not handle Python yet.
The card is half-size, single-width, 54 W.
I’m open for questions on this. Thanks.
-Jeff