• State
  • Locked Locked
  • Replies 3 replies
  • Subscribers 101 subscribers
  • Views 5837 views
  • Users 0 members are here
Support feedback
Options
Options
Similar topics

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

PowerVR SGX for GP computation

Massimiliano Cialdi
Expert 1995 points
Other Parts Discussed in Thread: OMAP3530

Hi,

we are starting with a project that is about signal processing. We need very high performance, because we need to processe 6~8 channels 24nit @192KHz, and 16K~24K-Taps FIR filter for each channel.

We read in many place about the possibility to use PowerVR SGX core to speed up elaboration (GP-GPU) that is reported also in SGX Series5 Factsheet (in Imagination Techonologies web site).
We wonder how can we use the SGX core for that aim. Where can we find OpenCL library (ore some other library to make GPGPU computations)? Or may we need to write our own library that uses SGX core? In that case where can we find some technical doc? Is it so hard to write our own GP-GPU library?

thanks

  • 0
    TI__Mastermind 43441 points

    Have you already thought about using the C6000 DSP on OMAP3530 for that task?

    Seems to be the natural block to implement a filter.

    Regards.

  • 0 in reply to Frank Walzer
    Expert 1995 points

    Yes, we must perform some benchmark, but the DSP embedded in omap3530 is only fixed point. For FIR filters we need single precision floating point or at least >= 40 bit fixed point.

    As you can see the number of MAC is so high. if one TAP coulb be executed with a single floating point MAC, we need about 8x192000x24000 = 37000 MMAC. Unfortunately DSP of omap has 3440 MMAC where a MAC is on integer 8bit.

    So we plan to use NEON technologies, and also DSP, but we doubt that we can get that kind of performace without also the help of SGX used as GP-GPU.

    We can relax some constrain such as 96KHz instead of 192, 20K o 16K TAPs instead of 24K for FIR filters, in that case we need about 12000 MMAC floating point.

    This new project is the evolution af an older one where there was a number of processor and DSP, and following that experience, now we want to have only one processor

     

    thanks

  • 0
    TI__Intellectual 1910 points

      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.

    Regards, Clay