Elvis- > On Dec 14, 2010, at 11:14 PM, Matt Ettus wrote: > >>>> Since the OMAP has a GPU incorporated does that mean that we could use it >>>> for processing? Is there a CUDA equivalent for this type of GPU? >> >> >> Doug has it pretty correct here. This is one of those areas I would call >> theoretically possible, but unlikely to be >> worth the trouble. You would need Imagination Tech and TI to get together >> to release an OpenCL implementation for >> the GPU, and even then it isn't the world's fastest. >> >> The way I look at it, you have 3 much better, easier to use options -- the >> ARM, the DSP, and the FPGA. If you put >> all those to good use and still need more power, the GPU is not likely to >> contribute appreciably. > > > A quick search on the internet yields some interesting info, but at the end, > you'd probably get much better results > off the FPGA. > > Some work done by Nokia: > > http://www.hotchips.org/archives/hc21/1_sun/HC21.23.2.OpenCLTutorial-Epub/HC21.23.270.Pulli-OpenCL-in-Handheld-Devices.pdf > > The OMAP4430 seems to have support for OpenCL as mentioned in this article: > http://e2e.ti.com/blogs_/b/mobile_momentum/archive/2010/02/15/omap-4-platform-ready-to-lead-the-human-device-interaction-revolution.aspx > > You can get a Pandaboard from DigiKey for USD$179 which has the OMAP4430 and > the required GPMC interface available on > the Pandaboard expansion header, to experiment with. > > A statement from TI from this link: > http://e2e.ti.com/support/dsp/omap_applications_processors/f/447/t/43845.aspx > > "TI does not support OpenCL for the SGX530. IMG does advertises GP-GPU > applications on the SGX, but TI does not > license these OpenCL drivers. The best available solution would be to use > the OpenGL ES 2.0 shading language (GLSL ES > 1.0) to do the single precision matrix operations that you need. Rather than > displaying the output results as pixels > in a framebuffer, your program on the ARM can use the results." > > Another statement from Imagination Technologies from this link: > http://www.imgtec.com/forum/forum_posts.asp?TID=194&PID=2668 > > "The SGX530 core design that's in an OMAP 3530 board is an example of one of > our products. This was licenced to TI and > is a design with the capability of supporting the OpenCL Embedded profile. > However, it requires the correct software > i.e. drivers to expose this functionality, just like drivers are required for > OpenGL ES, DirectX etc. Once we have > licenced a core like this we work with the customer (in this case TI) to > support them in exposing the functionality > that they wish to have available in their product. So if a customer wants to > expose OpenCL then this work happens, but > it requires time and the desire of the customer. This means that developers > don't always have full access to every > feature that our cores are capable of because the driver support is still > being added or the customer doesn't wish to > expose that feature. > > In summary: we advertise an ability of our core design and our customers get > a core design that can do this if they > choose to enable it. The developer gets a core with access to the features > that our customer has exposed. So > developers don't always have access to every feature of our core. > > Licence deals involve a certain amount of confidentiality so I can't talk > about future products or support from our > customers so I can't tell you when or if OpenCL will be enabled on specific > platforms on this forum at this time."
A future option is to use OpenMP to annotate sections of user C code that will run on the C64x+ core. We've made some preliminary mention of this on TI's e2e forum (CIM OpenMP site:ti.com). We use "OpenMP style" syntax, basically the same pragmas with omp replaced by sig. For compute-intensive C code, TI cores have decent SIMD capability (especially the new C66x series) and (unlike GPUs) are also good at irregular code sections. -Jeff _______________________________________________ Discuss-gnuradio mailing list [email protected] http://lists.gnu.org/mailman/listinfo/discuss-gnuradio
