Identify Kernels to Offload
To best utilize the compute cycles available on the devices of a
heterogeneous platform, it is important to identify the tasks that are
compute intensive and that can benefit from parallel execution. Consider
an application that executes solely on a CPU, but there may be some
tasks suitable to execute on a GPU. This can be determined using the
Offload Modeling perspective of the Intel®
Advisor.
Intel Advisor estimates performance characterizations of the workload as
it may execute on an accelerator. It consumes the information from
profiling the workload and provides performance estimates, speedup,
bottleneck characterization, and offload data transfer estimates and
recommendations.
Typically, kernels with high compute, a large dataset, and limited
memory transfers are best suited for offload to a device.
See Get Started: Identify High-impact Opportunities to Offload to
GPU
for quick steps to ramp up with the Offload Modeling perspective. For
more resources about modeling performance of your application on GPU
platforms, see Offload Modeling Resources for Intel® Advisor
Users.