Forum Discussion

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

estimate total pipeline latency for kernel

Hi, I'm new to using Intel's OpenCL SDK for FPGAs .

I'm developing a single work-item kernel and I'm experimenting with code changes and would like to see the effect of these changes on my total kernel pipeline latency (in terms of clock cycles )

As recommended in the programming guide I use : 

aoc -c mykernel.cl -report

this generates a report.html , under the System Viewer I also see the different Latencies for each block of code and loop initiation intervals. However if the workload I want to run is known at kernel compile time

I would like to estimate the total number of clock cycles needed to complete a given task. (for example I hardcode the number of iterations that a loop is supposed to execute so at compile time it is known how many iterations a loop will execute )

Would really appreciate if anyone can help share their knowledge about this or if there is any suggestion so that I can refine my question .

Thanks :D

5 Replies

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

    For a pipeline depth of P, a loop trip count of L, and an initiation interval of I, the execution time of the pipeline in clock cycles is:

    P + (L - 1) * I

    You can calculate the execution time of each block separately using this formula, and just add them up to get the total number of cycles. Dividing that by operating frequency will give you the total run time in seconds. Of course this does not take stalls from channel and global memory operations into account. It not easy to model such stalls.
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    8 years ago

    Hi @HRZ ,

    Really appreciate your quick response !!

    Thank you for the tip , I was hoping there was some way to do this automatically with the offline compiler/ quartus toolset as it gets really tricky for complex models (as you said) :(
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    8 years ago

    @HRZ really appreciate your quick response !

    I was kind of hoping there was a way to extract this from the offlince compiler / generated files/quartus project as this gets tricky for complex designs with stalls (as you said) :(
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    8 years ago

    @HRZ really appreciate your quick response !

    I was kind of hoping there was a way to extract this from the offlince compiler / generated files/quartus project as this gets tricky for complex designs with stalls (as you said) :(
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    8 years ago

    Altera does not provide an automated way of doing this but this can be automated manually using static/dynamiccode analysis and parsing the info from the report using a custom program.