Forum Discussion

Altera_Forum's avatar
Altera_Forum
Icon for Honored Contributor rankHonored Contributor
10 years ago

Bandwidth Efficiency values in the Altera Dynamic Profiler for OpenCL

Hi everyone,

can somebody explain me the values of bandwidth efficiency in the Altera Dynamic Profiler for OpenCL?

Like in the screens attached to this post, in a kernel that I'm working on I get an efficiency value of 1.5%. When, in reality, I read at ~6510 MB/s out of 12800 MB/s from each bank. What am I missing here? :confused:

Thanks,

Nicola.

3 Replies

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    10 years ago

    It depends on the memory access unit size you use and the number of memory load/store instructions.

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    10 years ago

    ah thanks wzk6_3_8!

    I also found this:

    --- Quote Start ---

    The overall efficiency is the percentage of total bytes, acquired from the global memory system, that the kernel program uses.

    --- Quote End ---

    So I guess I'm getting low efficiency values because the .aocx file contains multiple lead/store instructions.

    Thanks,

    Nicola.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    10 years ago

    --- Quote Start ---

    ah thanks wzk6_3_8!

    I also found this:

    So I guess I'm getting low efficiency values because the .aocx file contains multiple lead/store instructions.

    Thanks,

    Nicola.

    --- Quote End ---

    Yes, the load/store instructions share the memory bandwidth.