Description

Fix the timing calculation in example 10.

Previously, the measured total runtime from timer.seconds() was divided by
options.iterations twice:

1. once when assigning cute_time
2. again when computing cute_average_time

In addition, cute_average_time was later divided by 1000.0 when calculating
GFLOPS, even though the previous value was already based on seconds.

This caused the reported average runtime to be too small and the reported GFLOPS
to be too large. With the default --iterations=100, the GFLOPS value could be
inflated by 1000 * iterations, i.e. 100,000x.

The fix converts total elapsed seconds to milliseconds first, then divides once
by the iteration count to report average runtime in ms. GFLOPS calculation then
converts the average runtime back to seconds.

Type

• Bug
• Feature
• Performance
• Refactor

Testing

• Tests pass
• Xe12
• Xe20

Performance

References

Fixes #

Checklist

• Copyright
• Co-pilot Review
• Deprecated APIs not used