Intel OpenCL NEO Compute Stack Moves To "Production" Quality OpenCL 2.1

[phoronix]

Phoronix: Intel OpenCL NEO Compute Stack Moves To "Production" Quality OpenCL 2.1

This year Intel open-sourced their "NEO" OpenCL compute stack included a new compute runtime, a new LLVM/Clang-based compiler, makes use of the Intel Graphics Memory Management Library (GMMLIB), etc. While we don't hear too much from the NEO effort on an ongoing basis, their OpenCL 2.1 support for recent hardware generations is now to production quality...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-OpenCL-NEO-Production-2.1

[TemplarGR]

Nice. This seems to be the year of OpenCL for Linux. Hoping for similar improvements on the AMD side of things.

[davidbepo]

any aur to test this?

[boxie]

Nice. This seems to be the year of OpenCL for Linux. Hoping for similar improvements on the AMD side of things.

with discrete GPU handling landing for ROCm (sans vega) and the raven ridge CPUs already supported they are getting there with 4.17 (if memory serves). i can't remember off the top of my head which GPUs are supported though, and there is a requirement for PCIe Atomics

[darkbasic]

Is there any AUR package for that?

[TemplarGR]

with discrete GPU handling landing for ROCm (sans vega) and the raven ridge CPUs already supported they are getting there with 4.17 (if memory serves). i can't remember off the top of my head which GPUs are supported though, and there is a requirement for PCIe Atomics

Yes i know, but IIRC ROCm 2.x support is not that good yet. Plus i think they will provide a solution for PCIE 2.0 users too, without atomics.

[Marc Driftmeyer]

Yes i know, but IIRC ROCm 2.x support is not that good yet. Plus i think they will provide a solution for PCIE 2.0 users too, without atomics.

It turns out the atomics issue is no longer a requirement [as if it ever needed to be in the first place], and with OpenCL C++/C 2.1 rolling into LLVM/Clang proper you can bet all but Vega dGPUs will be testing the stack with 4.17 and come 4.18 Vega dGPU are said to be rolled in.

Plenty of traffic on OpenCL C++ in Clang commits this week: http://lists.llvm.org/pipermail/cfe-...23/thread.html

[boxie]

It turns out the atomics issue is no longer a requirement [as if it ever needed to be in the first place], and with OpenCL C++/C 2.1 rolling into LLVM/Clang proper you can bet all but Vega dGPUs will be testing the stack with 4.17 and come 4.18 Vega dGPU are said to be rolled in.

Plenty of traffic on OpenCL C++ in Clang commits this week: http://lists.llvm.org/pipermail/cfe-...23/thread.html

hey sweet - i'd heard that they were going to investigate if they could get rid of the pcie atomics, but not that they could and would.

I will be eagerly awaiting some OpenCL testing goodness i think - my vega needs a workout

[boxie]

Yes i know, but IIRC ROCm 2.x support is not that good yet. Plus i think they will provide a solution for PCIE 2.0 users too, without atomics.

I am sure that this is just a temporary situation and OpenCL will be improving greatly - especially with SPIR-V and clang's OpenCL work