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.
- Explore Performance Gain from GPU-to-GPU Modeling (preview) and view how your DPC++, OpenMP* target, or OpenCL™ application can have a better performance if you run it on a different graphics processing unit (GPU) device.
- 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.
For a general overview of the report, see
Offload Modeling Report Overview.