[parsec-users] Seems VIPS results differs in 1 thread and 2 threads?

jcupitt at gmail.com jcupitt at gmail.com
Mon Jul 16 04:27:53 EDT 2012


Hi, I'm the libvips maintainer.

On 16 July 2012 06:16, Zhunping Zhang <zhunping at csail.mit.edu> wrote:
> Hi, its weird that I got different output from VIPs using 1 thread and 2 threads, as follows:
>
>  > export IM_CONCURRENCY=1
>  > vips im_benchmark ../data/bigben_2662x5500.v output1.v
> ...
>> export IM_CONCURRENCY=2
>> vips im_benchmark ../data/bigben_2662x5500.v output2.v
>> diff output1.v output2.v
> Binary files output1.v and output2.v differ
>
> any ideas?

One of the operations in the benchmark does an error-diffusion dither,
so the exact numbers you get change as the set of tiles used to
compute the result change, and that of course changes with how the
image is divided up for processing.

I agree that it would be better to have a benchmark that always gave
exactly the same results no matter how many threads were used. Perhaps
this would be something to look at in the next revision of PARSEC.

John


More information about the parsec-users mailing list