User Guide

Contents

Identify Code Regions to Offload

After you run the
Offload Modeling
perspective, you get an
Offload Modeling
report. Depending on a configuration chosen, the report shows a different level of details:
  • Examine regions recommended for offloading and view estimated performance of your application after offloading to a target platform assuming it is mostly bounded by compute limitations. You need to run at least the Survey, Trip Counts and FLOP (Characterization), and Performance Modeling analyses (
    Low
    accuracy) to collect this data.
  • Examine data transfers estimated for modeled regions and view estimated performance with data transfer estimations between host and target platforms for all memory levels and total data for loop/function. You need to run at least the Survey, Trip Counts and FLOP with callstacks, light data transfer simulation, and cache simulation (Characterization), and Performance Modeling analyses (
    Medium
    accuracy) to collect this data.
  • Check for dependencies issues and view a more accurate performance estimated considering loop/function dependencies. You need to run at least the Survey, Trip Counts and FLOP with callstacks, cache simulation, and medium data transfer simulation (Characterization), Dependencies, and Performance Modeling analyses (
    High
    accuracy) to collect this data.
  • Investigate non-offloaded code regions and understand why they are not profitable to run on a target platform. The higher accuracy level you run, the more accurate offload recommendations and non-offloaded reasons are.

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.