| Age | Commit message (Collapse) | Author |
|---|
|
The new version uses a fixed time and provides variants for
single-thread, multi-thread, and multi-core.
A few results are included.
Signed-off-by: Burt P <pburt0@gmail.com>
|
|
If (end - start) / threads left a remainder, benchmark_parallel_for()
would start an extra thread for the leftover elements. Now, it has the
last thread process the few extra elements.
Added a note:
benchmark_parallel_for(): element [start] included, but [end] is excluded.
callback(): expected to processes elements [start] through [end] inclusive.
Signed-off-by: Burt P <pburt0@gmail.com>
|
|
New benchmark types:
* benchmark_crunch_for(): call function repeatedly for a number
of seconds; result is number of completions.
* benchmark_parallel(): one call for each available thread
up to n_threads; result is sum of return values.
Tweaks:
* Store return values from callback via
benchmark_parallel_for_dispatcher()
Signed-off-by: Burt P <pburt0@gmail.com>
|
|
Benchmark results store actual number of threads used by benchmark
when it was run. Previously, results assumed all available threads
were used.
Examples:
* CPU Fib only uses one
* FPU FFT uses 4, 2, or 1
* N-Queens uses 10, 5, 2, or 1
Signed-off-by: Burt P <pburt0@gmail.com>
|
|
|
|
gcc was (correctly) assuming that benchmark_parallel_for() returned
an integer, which was being to a double and messing up results. Added
the missing prototype.
Also, populate bench_results array with sane values on module init.
Some other minor cleanups in benchmark code.
|
|
|
|
|
