vetter/shoc

Transfer timings for CUDA Scan are wrong in results.csv database

grahamlopez opened this issue · 4 comments

grahamlopez commented

I fixed a bug today with these timings, which I believe makes those results in the database wrong. After my fix, the rates are now similar between cuda/opencl, whereas previously there was a 10^3 difference.
I'm throwing this out there for suggestions about what to do with the results database.

jsmeredith commented

Seems like it should be easy enough to detect which ones are wrong, with
that size of magnitude error? If so, I'd be fine with just changing the
database.

grahamlopez commented

I guess you're volunteering then to hunt down all of the devices that we
have run on and make updates? ;-)

On Fri, Sep 12, 2014 at 2:06 PM, Jeremy Meredith notifications@github.com
wrote:

Seems like it should be easy enough to detect which ones are wrong, with
that size of magnitude error? If so, I'd be fine with just changing the
database.

—
Reply to this email directly or view it on GitHub
#42 (comment).

jsmeredith commented

Yeah, no. (Not possible anyway.) But I'll estimate the right values since we know the math behind the the error and update results.csv (marking them as estimates). Assigning to myself.

grahamlopez commented

You the man.
On Sep 12, 2014 4:41 PM, "Jeremy Meredith" notifications@github.com wrote:

Yeah, no. (Not possible anyway.) But I'll estimate the right values since
we know the math behind the the error and update results.csv (marking them
as estimates). Assigning to myself.

—
Reply to this email directly or view it on GitHub
#42 (comment).