[MicroBenchmarks/LoopVectorization] Add Microbenchmark for Epilogue Vectorization #165
Conversation
fhahn
requested review from
nilanjana87,
fhahn,
davemgreen and
sjoerdmeijer
| } \ | ||
| BENCHMARK(benchReductionAutoVecFor##Ty##ForLoopTC##Itr); | ||
|
|
||
| ADD_BENCHMARK(uint8_t, 65) |
There was a problem hiding this comment.
Function specialization has an option to specialize on constants, but it is off by default. If however it would be enabled, I am wondering if an iteration count of 65 is small enough to trigger the full unroller, in which case we are no longer test what we want to test it. I know this is all a bit hypothetical and a nit, but perhaps to future prove this a little, just increase the iteration count a bit?
There was a problem hiding this comment.
It might be worth covering all possible trip counts for the epilogue, using something like
BENCHMARK_TEMPLATE(uint8_t)->DenseRange(65, 83, 1);
There was a problem hiding this comment.
Ah, neat, thanks, I was wondering if there's a better way; latest version is now using BENCHMARK_TEMPLATE(...)->DenseRange(...) to get all trip counts.
There was a problem hiding this comment.
Now also running only a small set of tests by default to avoid potentially overloading build bots that run the test suite.
…ilogue Vectorization
… for Epilogue Vectorization Only run a reduced set of tests by default
…chmarks for Epilogue Vectorization
This microbenchmark attempts to test the impact of epilogue vectorization on loops with varying epilogue lengths and vector widths for loops with and without reductions.