[hpsdr] Regarding the Discussion on the TeamSpeak session on July 16th

Hermann hvh.net at gmail.com
Sun Jul 17 06:45:55 PDT 2011


Dear All,

I was just listening to July's 16th TeamSpeak session, and was happy
to hear your discussion about parallel DSP etc.

Just want to mention that in my last years experimental implementation
of CudaSharpDSP - at that time working still with Ozy - I indeed read
8k samples from Ozy and copied the raw data (24 Bit) directly onto the
GPU doing the conversion to float on the GPU. There was a major speed
up doing this, especially after I got a new Nvidia GTX460 (GT200 chip)
which has the new feature of "ZeroCopy". That is, you can read and
write directly to pinned system memory. That has a huge bandwidth.

After these experiments, I started working on doing all graphical
stuff (Panadapter, Waterfall, S-Meter and maybe more) directly on the
GPU using OpenGL. There I encountered some major problems with C#. Up
to now I don't know what I have done wrong, but I decided then to
switch to Qt/C++ and so I had to start all over again. This is still
going on. I now have a new implementation (called cuHPSDR) which is
surely not finished, but I managed to use Qt, OpenGL and Cuda
together, and I'm already doing the panadapter spectrum computation
solely on the GPU, that is, the CPU has just to do the job of copying
the raw data from Metis onto the GPU.

There is also some exciting news from Intel: Intel has just released a
SDK for using OpenCL on their new generation of Sandy Bridge CPUs
(using their vector processing units). Last week I finished updating
my Hardware with an Asus P8Z68 and an Intel i7 2600K Quadcore. Now I'm
playing in the Formula 1 league and I'm thinking seriously about
OpenCL again, Jeremy!

It will be interesting to see the competition of Intel's i7 2600K and
Nvidia's GT200 (of course, it heavily depends on the application), but
from what I see now it will be hard to defeat the i7 for our
application. Though it still makes sense I guess to put all the
computation regarding graphics onto the GPU using OpenGL.

I hope I can present some results soon. But first I have to finish the
ChirpWSPR decoder! I now got some great help from Kurt, DL9SM, which
has already the Chirp beacon running on his Penelope!

73, Hermann

DL3HVH

P.S. anyone interested in Google+ ? I could send some invitations then...



More information about the Hpsdr mailing list