Re: [boost] [compute] Review

Subject: Re: [boost] [compute] Review
From: Kyle Lutz (kyle.r.lutz_at_[hidden])
Date: 2014-12-30 23:42:40

On Tue, Dec 30, 2014 at 8:33 PM, John Bytheway
<jbytheway+boost_at_[hidden]> wrote:
> On 2014-12-29 02:01, Kyle Lutz wrote:
>> On Sun, Dec 28, 2014 at 7:26 PM, John Bytheway
>> <jbytheway+boost_at_[hidden]> wrote:
>>> - It looks like the performance tests don't include the time for copying
>>> data to (and, where applicable, from) the device.  You should mention
>>> this on the Performance page.
>>
>> I'll add this, though it is very device/system specific. Also, there
>> is a performance benchmark available in Boost.Compute
>> ("perf_copy_to_device") which could be used to measure this on your
>> own system.
>
> More on the subject of performance:
>
> I ran some of the performance tests on my machine, and I observed
> something surprising.
>
> set_intersection and set_difference are *unbelievably* fast (two orders
> of magnitude faster than the std:: versions), and yet set_union and
> set_symmetric_difference are quite slow (maybe twice as slow as std::).
>  Looking at the code, the four algorithms are very similar.  Have you
> any idea what causes this huge discrepancy in their speed?

Very interesting, I'll look into this more. Also, I wasn't the author
of those particular algorithms. Could you open an issue on the issue
tracker [1] for this? Please also include the performance numbers you
got and the device/platform you used.

-kyle

[1] https://github.com/kylelutz/compute/issues