Hopefully this time the SDK and 11.12 run-time will include some serious bug-fixes and performance optimizations:
http://blogs.amd.com/developer/2011/11/28/amd-opencl%E2%84%A2-app-sdk-preview/
This release will also include preview of OpenCL 1.2 features. So far, so good!
and also more features for Linux side. like OVD and zero copy.
Originally posted by: vvelichkov
Hopefully this time the SDK and 11.12 run-time will include some serious bug-fixes and performance optimizations:
http://blogs.amd.com/developer...-sdk-preview/
This release will also include preview of OpenCL 1.2 features. So far, so good!
New SDK available for download from APP SDK webpage. New drivers aren't there yet, but hope they make it sometime during the afternoon.
Note that the link is currently wrong for the Windows 64 bit installer, and points to APP 2.5. If you want the beta, just modify the link to 2.6 instead of 2.5... or click here
Regards,
Chris
Not sure if it's because I'm still using Catalyst 11.11 but now I'm using SDK 2.6, AMD APP KernelAnalyzer reports the same Compiler statistics no matter what Function you select...
AMD APP Kernel Analyzer uses runtime installed, which ships with drivers and not with the SDK. It might work if you manually replace runtime binaries with those found inside the SDK, but keep in mind that this is not standard procedure.
Hm, offline compilation crashes for me. Some kernels always crash when building, others just crash when building for the new targets (e.g Tahiti).
I guess I should wait until 11.12 comes out...
yes, offline compilations is broken. it either crashes during compilation or the produced binary output fails after reading with clBuildProgram() -11
Patience is a virtue my friends... (although I myself do have the url of 11-12 linux drivers in a browser, and check if the drivers are there, even if the webpage isn't updated yet)
If most bugs would be fixed, OCL 1.2 conformance be introduced, and perhaps a few new features be implemented (Direct GPU-GPU transfer, or GDS/GWS, ...) that would be a real nice treat before the holidays.
Yep 🙂
BTW is it my proxy or the AMD APP SDK page is seriously inconsistent? Links to APP SDK 2.6 were replaced with those to 2.5, but 2.6 checksums remain. Recommended Catalyst driver is 11.12 which is not out yet. Documentation links are for 2.5, but some of the documents linked are written for 2.6. Rather weird...
The contents of the page do not change to the wave of a wand. It takes roughly one day until they change all the links and update the pages. SDK was said to be out Dec 13 (today), so if they update everything by midnight, they are still in time.
Do not be surprised if drivers will only make it 1-2 days late. There have been sooo many bugs lately, I can live with a few days delay if it results in more stable and bug-free drivers. (Although you can bet I'm as eager to get my hands on them as you are)
Originally posted by: quadboon yes, offline compilations is broken. it either crashes during compilation or the produced binary output fails after reading with clBuildProgram() -11
ignore this, was an problem on my side. all my kernel compile cleanly.
Thanks Micah, I was able to compile successfully without -O0.
Drivers are out. Not sure if they are listed yet though.
http://www2.ati.com/drivers/11-12_vista64_win7_32_dd_ccc_ocl.exe
http://www2.ati.com/drivers/11-12_vista64_win7_64_dd_ccc_ocl.exe
also for linux http://www2.ati.com/drivers/linux/ati-driver-installer-11-12-x86.x86_64.run
so i instaled catalyst 11.12 and i can say that there is at lest one new extension and that is cl_ext_atomic_counters_32
I installed 11.12 and GPU is no long available as CL device?
Offline compilation still broken with 11.12, BUT:
I provide -O0 to the list of build options and it's OK. Moreover, looks like -O0 has no effect on performance, I get exactly the same performance as with 2.5 and 11.11 without -O0.
100% CPU problem is finally gone forever. Not 30% or so, almost no CPU utilization as it was in 2.4. Nice.
I am going to get the ISA dump out of the Tahiti target now, I expect that to be interesting 🙂
P.S WOW! Looks like AMD is truly ditching off VLIW! When are those going to be documented?
userElementCount
NumVgprs
NumSgprs
etc?
I cannot provide a good test case right now, I can send you one of my kernels (it's rather long though, dunno if it's the proper place to paste).
Yes, I can confirm. The 100% CPU bug is gone on Linux and on Windows. Good Job @ AMD.