Forum Discussion

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

aoc generated .attrib file meaning

Hi All,

I am now using AOC to generate .aocx for MatrixMul example.

Two hours later, it generate .aocx file successfully.

It also generate some report files.

One of the files is matrixmult.attrib, content as follows, any one can tell me the meaning of these data.

Vectorization: 4

Max_vectorization: 4

Copies: 1

Max_copies: 256

Throughput: 0.648117

Copyfactor: 4

Sharing: 1

Max_sharing: 1

Unroll: 1

Max_unroll: 1

Throughput_unroll: 1

Aggressive_unroll: 1

Thanks

5 Replies

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    It is about the optimization methods applied to your OpenCL kernel.

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    Speaking of .attrib file:

    does someone know the meaning of the fields throughput and copyfactor?

    For what is throughput used here? Work item per seconds?
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    --- Quote Start ---

    Speaking of .attrib file:

    does someone know the meaning of the fields throughput and copyfactor?

    For what is throughput used here? Work item per seconds?

    --- Quote End ---

    Just forget about it. It is not that accurate since it is really hard to predict the OpenCL kernel performance.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    --- Quote Start ---

    Just forget about it. It is not that accurate since it is really hard to predict the OpenCL kernel performance.

    --- Quote End ---

    Seriously, "Just forget about it", estimating performance is a required feature for hardware OpenCL compiler. Do not just forget about it, open service request to Intel to put 13.1 throughput estimates back into the tools or do something better, so users can see impact of code changes before waiting 2-8 hors before hardware is built.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor

    --- Quote Start ---

    Seriously, "Just forget about it", estimating performance is a required feature for hardware OpenCL compiler. Do not just forget about it, open service request to Intel to put 13.1 throughput estimates back into the tools or do something better, so users can see impact of code changes before waiting 2-8 hors before hardware is built.

    --- Quote End ---

    In a conference I asked members of Altera's OpenCL team in person about that throughput estimation; they explicitly said they removed it because it was completely inaccurate. Until they come up with a cycle-accurate emulator, giving accurate performance predictions will be impossible.