Latency Information - 2021.2 English

Vitis Unified Software Platform Documentation: Application Acceleration Development (UG1393)

Document ID
UG1393
ft:locale
English (United States)
Release Date
2021-12-15
Version
2021.2 English

The latency information presents the execution profile of each CU in the binary container. When analyzing this data, it is important to recognize that all values are measured from the CU boundary through the custom logic. In-system latencies associated with data transfers to global memory are not reported as part of these values. Also, the latency numbers reported are only for CUs targeted at the FPGA fabric. The following is an example of the latency report:

Latency Information (clock cycles)
Compute Unit  Kernel Name  Module Name  Start Interval  Best Case  Avg Case  Worst Case  
------------  -----------  -----------  --------------  ---------  --------  ----------  
mmult_1       mmult        mmult        826 ~ 829       825        827       828        

The latency report is divided into the following fields:

  • Start interval
  • Best case latency
  • Average case latency
  • Worst case latency

The start interval defines the amount of time that has to pass between invocations of a CU for a given kernel.

The best, average, and worst case latency numbers refer to how much time it takes the CU to generate the results of one ND Range data tile for the kernel. For cases where the kernel does not have data dependent computation loops, the latency values will be the same. Data dependent execution of loops introduces data specific latency variation that is captured by the latency report.

The interval or latency numbers will be reported as "undef" for kernels with one or more conditions listed below:
  • OpenCL kernels that do not have explicit reqd_work_group_size(x,y,z)
  • Kernels that have loops with variable bounds
Note: The latency information reflects estimates based on the analysis of the loop transformations and exploited parallelism of the model. These advanced transformations such as pipelining and data flow can heavily change the actual throughput numbers. Therefore, latency can only be used as relative guides between different runs.