You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It shouldn't surprise anyone that ISPC generated code in general won't compete with Clang if we compare it by calling the ISPC compiled SPMD with just 1 active lane. ISPC is not made for that purpose, it's made to compute a bunch (8, 16, ..) of values at once.
However if we like to use a single source for mixed SPSD/SPMD programming, then it would be nice if ISPC could provide a 'single lane' compile target.
In our case we need to support legacy code that can't easily be transformed into MD operation but we would rather not duplicate code and have separate SD and MD code bases for new code going forward.
Thanks for your consideration
The text was updated successfully, but these errors were encountered:
There is a "generic-1" target that might work for that--it's basically ispc
with a single lane. The only trick is that it is only supported by the C++
backend.
ispc --target=generic-1 foo.ispc --emit-c++ -o foo.cpp
(And then compile foo.cpp as usual with clang.)
I'm not sure how well maintained generic-1 is at this point and/or how
thoroughly it's been tested, FWIW.
Thanks,
Matt
On Wed, Jan 25, 2017 at 10:08 AM, mauricevanswaaij ***@***.*** > wrote:
It shouldn't surprise anyone that ISPC generated code in general won't
compete with Clang if we compare it by calling the ISPC compiled SPMD with
just 1 active lane. ISPC is not made for that purpose, it's made to compute
a bunch (8, 16, ..) of values at once.
However if we like to use a single source for mixed SPSD/SPMD programming,
then it would be nice if ISPC could provide a 'single lane' compile target.
In our case we need to support legacy code that can't easily be
transformed into MD operation but we would rather not duplicate code and
have separate SD and MD code bases for new code going forward.
Thanks for your consideration
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#1254>, or mute the thread
<https://github.com/notifications/unsubscribe-auth/AA0YxKMi9MaxWnv8MFmyG-C1qcmYYX4Kks5rV4-7gaJpZM4Lt0Da>
.
It shouldn't surprise anyone that ISPC generated code in general won't compete with Clang if we compare it by calling the ISPC compiled SPMD with just 1 active lane. ISPC is not made for that purpose, it's made to compute a bunch (8, 16, ..) of values at once.
However if we like to use a single source for mixed SPSD/SPMD programming, then it would be nice if ISPC could provide a 'single lane' compile target.
In our case we need to support legacy code that can't easily be transformed into MD operation but we would rather not duplicate code and have separate SD and MD code bases for new code going forward.
Thanks for your consideration
The text was updated successfully, but these errors were encountered: