User Copy Benchmark

This program is meant to measure the latency of performing copy operations with specific instruction set on arm64 platforms.

How to run:

If your arm64 platform is able to be served by a PM server (a.k.a. fx serve), then adding //zircon/kernel/arch/arm64/user-copy:benchmarks will be enough. If this is not the case, then adding //zircon/kernel/arch/arm64/user-copy:benchmarks-standalone to your base labels is enough.

Once you have built, then the following binaries will be available on your fuchsia instance:

• _arm64_user_copy : stp, ldp bidirectional
• _arm64_user_copy_to_user : ldp, sttr unidirectional
• _arm64_user_copy_from_user : ldtr, stp unidirectional

The three binaries share the same set of command line arguments:

• ‘--cpu, -c’ cpu id to bind to. (fuchsia specific).
• ‘--name,-n’ name of the cpu(e.g. cortex-A53).
• ‘--output,-o’ path to an output file where to write the csv contents.
• ‘--seed,-s’ random seed to fix the copied contents (reproduceability).
• ‘--profile,-p’ allows setting up a deadline profile, which in theory should reduce variance (fuchsia specific).
• ‘--logtostdout,-l’ instead of writing output to a file, will write straight into stdout.

Sampling

The sampling is performed as follows, for each block size, source alignment and destination alignment triplet a warm up is performed, which consist of an arbitrary round of copies of arbitrary values into the range.

To reduce variance and noise, the measurement is performed of kSampleCount copies in sequence which is then averaged (divided by kSampleCount).