Webrev for first version of HSA offload integer reduce
Caspole, Eric
Eric.Caspole at amd.com
Tue Jun 10 15:03:29 UTC 2014
I moved collectArgs into the cpp files for both classes here:
http://cr.openjdk.java.net/~ecaspole/graal_int_reduce1/02/webrev/
Thanks,
Eric
________________________________
From: Christian Thalinger [christian.thalinger at oracle.com]
Sent: Monday, June 09, 2014 2:15 PM
To: Caspole, Eric
Cc: graal-dev at openjdk.java.net; sumatra-dev at openjdk.java.net
Subject: Re: Webrev for first version of HSA offload integer reduce
On Jun 9, 2014, at 10:50 AM, Eric Caspole <eric.caspole at amd.com<mailto:eric.caspole at amd.com>> wrote:
On 06/09/2014 01:07 PM, Christian Thalinger wrote:
On Jun 9, 2014, at 9:40 AM, Caspole, Eric <Eric.Caspole at amd.com<mailto:Eric.Caspole at amd.com>
<mailto:Eric.Caspole at amd.com>> wrote:
Hi everybody,
I put up a new webrev that allows some IntStream.reduce() to be
offloaded to HSA:
http://cr.openjdk.java.net/~ecaspole/graal_int_reduce1/webrev/
src/gpu/hsail/vm/hsailKernelArguments.hpp
+ virtual void collectArgs() {
Does this really have to be in the .hpp file? I see that the base class
has it in the .hpp file too.
No it doesn't have to be there, it is just some historical thing I don't know why. I can change that.
That would be good. Thanks.
This works with a Sumatra webrev I will eventually add into the
Sumatra JDK repo:
http://cr.openjdk.java.net/~ecaspole/sumatra_int_reduce_1/webrev/
This combo allows offloading IntStream.parallel().reduce() for
Integer::sum, Integer::max, and Integer::min. It builds a kernel from
a hand made HSAIL string of code that gets called from the new Stream
API diversion code in the Sumatra part.
Eventually we hope to have regular compilation of user reduce
functions but this allows us to do experiments with workloads before
we get all the details of reduce compilation in place such as barrier
placement and group size nuances.
Presumably that will remove the hand-made code eventually?
Yes. I just want to be able to expose this interesting functionality in the mean time.
Makes sense.
In the current Kaveri hardware version of Okra, it uses group size 256
by default and this reduce code is designed to work only with 256.
Also, because it is a handmade kernel, it only runs with
-UseHSAILDeoptimization and -UseCompressedOops.
In the included reduce tests, they will skip if the correct flags are
not there for the tests to run correctly.
This webrev also changes the way the kernel args are processed to look
at the actual args type rather than by the type signature of the
kernel method.
Regards,
Eric
More information about the graal-dev
mailing list