Now I am getting grey hairs.
When I returned to my computor, BM had exited.

When starting it again 2 Seti Beta tasks had error.

- exit code 194 (0xc2) both of them.

link to the 2 tasks.

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391253

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391259

Eric, we need to convince David to report OpenCL runtime version for OpenCL devices. "Driver" versions have nothing to do with computations BOINC performs.

I'm adding that support now, but I'm firmly of the belief that it shouldn't have been necessary. Either an environment supports OpenCL 1.2 or it doesn't. The verion of the runtime should be irrelevant once OpenCL 1.2 support has been advertised. Newer runtimes shouldn't break things. That AMD feels free to release new runtimes that don't support compilation of existing compliant code is probably evidence that AMD no longer has any real interest in supporting OpenCL or general purpose computing on its GPUs.

The initial reaction from other members of the team here is that SETI@home should drop support for AMD/ATI GPUs until AMD decides to get its act together. I've decided that I'm willing to make this change this time, but the next time AMD decides to break things (which it could easily do by changing the format of the runtime version string) I might agree to give up on them.

I'm sincerely hoping Intel will choose not to have problems like this with their runtime environments.

---

I think there must be a bug here. The opencl_ati app shoudn't be going to OpenCL 1.1 drivers.

Now I am getting grey hairs.
When I returned to my computor, BM had exited.

When starting it again 2 Seti Beta tasks had error.

- exit code 194 (0xc2) both of them.

link to the 2 tasks.

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391253

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391259

---

I think there must be a bug here. The opencl_ati app shoudn't be going to OpenCL 1.1 drivers.

I don't know if this is a bug too, but I'm getting both SETI@home v7 v7.02 (ati_opencl_sah), and SETI@home v7 v7.02 (ati_opencl_pre129), for my ATI HD7870, Driver CAL 1.4.1720 (VM), Catalyst 12.4, OpenCL 1.2 AMD-APP (923.1.

http://setiweb.ssl.berkeley.edu/beta/show_host_detail.php?hostid=57178

I think there must be a bug here. The opencl_ati app shoudn't be going to OpenCL 1.1 drivers.

Now I am getting grey hairs.
When I returned to my computor, BM had exited.

When starting it again 2 Seti Beta tasks had error.

- exit code 194 (0xc2) both of them.

link to the 2 tasks.

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391253

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391259

I just noticed opencl1.1 some minute ago too...

I use catalyst 12.1

I've made alot of succesful tasks though...
More then a 100 of them are valid. about 70 pending 3 inconclusive and 2 errors.
1% error so far. That is pretty ok.

21 more tasks to run, then I am out of tasks.

Ok, I just finished a bunch of AP 6.04's on main. I will now crunch 44 SETI@home v7 v7.02 (ati_opencl_pre129) and 81 SETI@home v7 v7.02 (ati_opencl_sah) here in Beta.

Let's see how that works.

How are folks testing this? I can't get any GPU tasks for for several days now.

---

Dublin, California
Team: SETI.USA

How are folks testing this? I can't get any GPU tasks for for several days now.

It's just the luck of the draw just now really. There's mostly VLAR's coming from the splitters now, and those are not sent to the GPU.

I'm going to put a new scheduler on in the next hour or so that should be able to check ATI OpenCL runtime revisions on machines that report driver revisions.

---

How are folks testing this? I can't get any GPU tasks for for several days now.

It's just the luck of the draw just now really. There's mostly VLAR's coming from the splitters now, and those are not sent to the GPU.

Got it. And as luck would have it, I just received tasks. :D

---

Dublin, California
Team: SETI.USA

I'm going to put a new scheduler on in the next hour or so that should be able to check ATI OpenCL runtime revisions on machines that report driver revisions.

A catch 22 then.

I can't get any tasks that I can crunch on my OpenCl 1.1 catalyst driver.
but they worked better then the old 7.01 tasks that gave driver restarts and blue screens. I didn't get any of thoose after downgrading to 12.1 though.
Only working ones of 7.02...

I just un-NNT'd and got 0 tasks.
So if the change is made, now it works with my computor.

Right now we are limited as the ATI developers recommend, OpenCL runtime 938.2 and below which means Catalyst 12.8 and below.

Unless someone can find me a newer driver that works reliably...

But I did reduce the OpenCL requirement to 1.1 just to see if it works reliably.

---

Got it. And as luck would have it, I just received tasks. :D

I lied. Those were S@H tasks, not S@H Beta tasks. D'oh!

---

Dublin, California
Team: SETI.USA

Okay, I finally have some tasks. Wow, it uses a LOT of CPU. Anything less than a full CPU thread per task really degrades the load on the GPU and the speed of completing tasks. And the only way to load up the GPU is to run multiple tasks. Running 2 tasks, with 1 CPU thread each, results in 60% GPU load. Reserve another thread, and you get about 80%. Looks like I need to run 3 tasks per GPU, with 4 CPU threads to make this work. Feeling kinda like POEM. This is on a 7970 on a SB i7 @ 4ghz. BOINC 7.0.62.

---

Dublin, California
Team: SETI.USA

I'll be looking for min and max driver revisions under which this version functions.

As i keep repeating again and a again, the Max OpenCL device version has to be 'OpenCL 1.2 AMD-APP (938.2)' which is what comes with Cat 12.8,
The OpenCL runtime that comes with Cat 12.10 produces compilations that cause the wrong number of pulses to be found, leading to the rash of AMD/ATI app inconclusives,
Drivers after that produce compilations that use less registers and cause driver restarts on some Wu's, and low GPU usage/High CPU usage on others.
The APP runtime with Cat 13.2 and Cat 13.3 Beta's fail to do the compilations at all.
It is O.K to do the compilations under Cat 12.8/APP runtime 938.2, then upgrade to a later driver, the app will work fine (the compilations are only done when first run)
What you can't do is do the compilations on a later APP runtime, then revert to an earlier driver, compilations are already done wrong, only manual deletion, or a detach/reattach will remove them.

ATI Driver Version Cheat Sheet

Claggy

I am confused (again). By "compilations" do you mean computations? And does the CAT version issue have a problem with the compiling of the app (project side), or the running of the app (user side)?

Which version of CAT do I need to be running for this test?

---

Dublin, California
Team: SETI.USA

Now I am getting grey hairs.
When I returned to my computor, BM had exited.

When starting it again 2 Seti Beta tasks had error.

- exit code 194 (0xc2) both of them.

link to the 2 tasks.

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391253

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=13391259

So, BOINC just killed fully completed task. And looks like it did that few times in a row.

Eric, we need to convince David to report OpenCL runtime version for OpenCL devices. "Driver" versions have nothing to do with computations BOINC performs.

I'm adding that support now, but I'm firmly of the belief that it shouldn't have been necessary. Either an environment supports OpenCL 1.2 or it doesn't. The verion of the runtime should be irrelevant once OpenCL 1.2 support has been advertised. Newer runtimes shouldn't break things. That AMD feels free to release new runtimes that don't support compilation of existing compliant code is probably evidence that AMD no longer has any real interest in supporting OpenCL or general purpose computing on its GPUs.

The initial reaction from other members of the team here is that SETI@home should drop support for AMD/ATI GPUs until AMD decides to get its act together. I've decided that I'm willing to make this change this time, but the next time AMD decides to break things (which it could easily do by changing the format of the runtime version string) I might agree to give up on them.

I'm sincerely hoping Intel will choose not to have problems like this with their runtime environments.

In ideal world - yes, they should not. But in reality big companies say something like: "Either eat that shit we throw to you and sit silently or PAY, PAY PAY". OpenCL support is free of charge service. Just look into license, they do no guaranties for any use (as most software companies do). So, either we will try to use what we have or we will have nothing. Not very nice choice, right?

For illustration: Urs found recently what summoned compilation failure for all latest Cat 13.3 betas. It was one of compiler switches I used to reduce binary size of cache. And this key was and still is listed in AMD manual. They just "forgot" to implement it in latest Catalyst. It costed us months of delay already. Nice? Well, it's capitalism ;) For what reason they will care if it bring no additional money to them? ...

I am confused (again). By "compilations" do you mean computations? And does the CAT version issue have a problem with the compiling of the app (project side), or the running of the app (user side)?

Which version of CAT do I need to be running for this test?

Compilation is the process of converting program sources to machine codes. Computation is the process to "convert" input data to useful output results.
So, it's _compilation_ that fails. That, in turn, leads to inability to do any computations.

But I did reduce the OpenCL requirement to 1.1 just to see if it works reliably.

Just FYI, both apps are OpenCL 1.0 compliant.
Issue not with OpenCL 1.2, 1.1 or 1.0 OpenCL support per se, issue ONLY with some AMD Catalyst drivers that unable to conform OpenCL specification (if specification implies that driver should work of course ;) )

I'll be looking for min and max driver revisions under which this version functions.

As i keep repeating again and a again, the Max OpenCL device version has to be 'OpenCL 1.2 AMD-APP (938.2)' which is what comes with Cat 12.8,
The OpenCL runtime that comes with Cat 12.10 produces compilations that cause the wrong number of pulses to be found, leading to the rash of AMD/ATI app inconclusives,
Drivers after that produce compilations that use less registers and cause driver restarts on some Wu's, and low GPU usage/High CPU usage on others.
The APP runtime with Cat 13.2 and Cat 13.3 Beta's fail to do the compilations at all.
It is O.K to do the compilations under Cat 12.8/APP runtime 938.2, then upgrade to a later driver, the app will work fine (the compilations are only done when first run)
What you can't do is do the compilations on a later APP runtime, then revert to an earlier driver, compilations are already done wrong, only manual deletion, or a detach/reattach will remove them.

ATI Driver Version Cheat Sheet

Claggy

I am confused (again). By "compilations" do you mean computations? And does the CAT version issue have a problem with the compiling of the app (project side), or the running of the app (user side)?

Which version of CAT do I need to be running for this test?

The files the compiler compiles for the app on running for the first time will be similar to these (But with your own Processor and GPU names):

r1779_IntelRCoreTMi72600KCPU340GHz.wisdom
MultiBeam_Kernels_r1779.clHD5_Capeverde.bin_V7
MB_clFFTplan_Capeverde_8_r1779.bin
MB_clFFTplan_Capeverde_16_r1779.bin
MB_clFFTplan_Capeverde_32_r1779.bin
MB_clFFTplan_Capeverde_64_r1779.bin
MB_clFFTplan_Capeverde_128_r1779.bin
MB_clFFTplan_Capeverde_256_r1779.bin
MB_clFFTplan_Capeverde_512_r1779.bin
MB_clFFTplan_Capeverde_1024_r1779.bin
MB_clFFTplan_Capeverde_2048_r1779.bin
MB_clFFTplan_Capeverde_4096_r1779.bin
MB_clFFTplan_Capeverde_8192_r1779.bin
MB_clFFTplan_Capeverde_16384_r1779.bin
MB_clFFTplan_Capeverde_32768_r1779.bin
MB_clFFTplan_Capeverde_65536_r1779.bin
MB_clFFTplan_Capeverde_131072_r1779.bin
MB_clFFTplan_Capeverde_524288_r1779.bin

It is the actual compilation where the problems are, you need to do the compilations on Cat 12.8/APP runtime 938.2 or earlier, once the compilations are done the app can be run on later drivers no problem (until the next time compilations are done),
Soon we'll have apps that compile O.K on Cats 13.2 & Cat 13.3 (Both on APP runtime 1124.2) once Alpha testing is complete.

Claggy