Heterogeneous Memory Management Is Coming Along For The Linux Kernel

[phoronix]

Phoronix: Heterogeneous Memory Management Is Coming Along For The Linux Kernel

Jerome Glisse remains hard at work on readying his Heterogeneous Memory Management (HMM) patch-set for eventual integration into the mainline Linux kernel. This HMM memory management will benefit HSA/OpenCL workloads on Linux...

http://www.phoronix.com/vr.php?view=MTg3MDc

[Zan Lynx]

This will be very neat. I'm looking forward to it.

The only downside is that as time goes by programmers will come in that don't realize how expensive the access is behind the scenes. Just like programmers now who don't realize how expensive cache line updates are in threading.

When you have to make explicit function calls to get access to the data it is a lot more obvious what price you are paying.

[gutigen]

Are there any mainstream apps actually making use of HSA at this moment?

[darkbasic]

Are there any mainstream apps actually making use of HSA at this moment?

No (stupid character limit)

[AndyChow]

Are there any mainstream apps actually making use of HSA at this moment?

Java 9 should make use of it, and if that's the case, every java based apps will make use of HSA "automatically".

I wouldn't count on it though... It will probably end up slowing down the apps, lol. I'll believe it when I see cold hard benchmarks showing me that it works.

[smitty3268]

Java 9 should make use of it, and if that's the case, every java based apps will make use of HSA "automatically".

I wouldn't count on it though... It will probably end up slowing down the apps, lol. I'll believe it when I see cold hard benchmarks showing me that it works.

OpenCL will use it automatically, eventually. It's not ready yet.

[Ericg]

Are there any mainstream apps actually making use of HSA at this moment?

HandBrake has an HSA enabled backend, Idk if its released yet.
libjpeg has a backend I THINK, again, idk if its released.
LibreOffice calc prototyped having HSA handle a few more expensive cell operations, again idk if it got released.

[coder]

OpenCL will use it automatically, eventually. It's not ready yet.

AMD already has OpenCL 2.0 support for certain PCIe cards. I wonder what they're currently using, and whether they will switch when this is in.

I don't like the demand-paging aspect of what he describes. Sure, it's simple & easy to use. The only problem is that performance would suffer from the fact that data is only transferred when the CPU page faults & thus is blocked. That's exactly the kind of thing high-performance APIs try to avoid.

[coder]

libjpeg has a backend I THINK, again, idk if its released.

Eek! I hope not! Many JPEG images are so fast to decode that the overhead of sending them to the GPU would be worse. Furthermore, the slowest part of JPEG decoding is the variable length decode, which is inherently serial and would be much slower on a GPU than most CPUs. And MMX/SSE/AVX is nearly as good as a GPU, for the GPU-friendly parts of JPEG.

[nanonyme]

Eek! I hope not! Many JPEG images are so fast to decode that the overhead of sending them to the GPU would be worse. Furthermore, the slowest part of JPEG decoding is the variable length decode, which is inherently serial and would be much slower on a GPU than most CPUs. And MMX/SSE/AVX is nearly as good as a GPU, for the GPU-friendly parts of JPEG.

Wasn't the point of HSA supposed to be you're dealing with shared memory with zero copy ao not much sending involved?