Intel integrated GPU

[Bdot]

Originally Posted by kracker

-d 11 output. Hmm.

Cool, we already got everything prepared, just the OpenCL code does not fit the Intel compiler.

Please use the two attached barrett files to replace the ones in the mfakto folder. I hope I added enough type-casts to satisfy the Intel compiler. If not, then feel free to add more (to reduce the turnaround times).

Once we got that running, it might be useful to check, if that platform runs mad24() at least at the same speed as mad() - I know that NVIDIA needs to add extra operations to mask out the upper 8 bits of the operands ...

[kracker]

Done. (Hmm..)

[Bdot]

Originally Posted by kracker

Done. (Hmm..)

Interesting. 4 tests succeeded!

So now it loads and compiles fine, just the code is failing

To see if it's the sieve or the tf, could you please run with
SieveOnGPU=0

Another test:
VectorSize=1

And I'd like to see the output of 2-3 minutes of mfakto -st (best would be with CPU sieving).

[kracker]

Originally Posted by Bdot

Interesting. 4 tests succeeded!

So now it loads and compiles fine, just the code is failing

To see if it's the sieve or the tf, could you please run with
SieveOnGPU=0

Another test:
VectorSize=1

And I'd like to see the output of 2-3 minutes of mfakto -st (best would be with CPU sieving).

Will do. Unfortunately later this evening.

[Bdot]

Originally Posted by kracker

Will do. Unfortunately later this evening.

Assuming it is the TF that actually fails, the next step is to edit mfakto_Kernels.cl, line 45

#define TRACE_KERNEL 0

Change that to 2 at first - later 3 or 4 will be needed.

Run that with
SieveOnGPU=1
VectorSize=1

Use the same settings for your AMD GPU and compare the output. There should be no differences at TRACE level 2. Higher levels also trace the intermediate results which may differ due to different rounding.

I guess that at some point, TRACE level 2 also shows differences. These need to be examined in the higher-level traces ...

If you send me the Intel-output of level 2, 3 and 4 (just 1 minute each), I can do the comparing and searching myself.


The fact that 4 test cases were successful makes me think that not all is lost

[TheMawn]

I don't suppose there would be an advantage to delegating a certain part of the whole process to the iGPU and the rest to a discrete GPU? Or to the actual cores of the processor?

[kracker]

Originally Posted by TheMawn

I don't suppose there would be an advantage to delegating a certain part of the whole process to the iGPU and the rest to a discrete GPU? Or to the actual cores of the processor?

Sorry, don't quite get it. Care to expand?

[kracker]

Hmmmm.
With GPU sieving off and VectorSize1:

number of tests          117
successful tests         53
no factor found          64

[kracker]

Sorry, triple post
some self test.
http://www.filedropper.com/output

[TheMawn]

Originally Posted by kracker

Sorry, don't quite get it. Care to expand?

What I meant was a while back the CPU did part of the work (sieving if I recall) while the GPU did the rest. Later, it was found that the GPU could do that as well. Is it possible that the iGPU might do some part of the work better or worse than another part which could be delegated to a different piece of hardware?

I.e. sieve on iGPU and TF on proper GPU if there is one available?

[Bdot]

Originally Posted by TheMawn

What I meant was a while back the CPU did part of the work (sieving if I recall) while the GPU did the rest. Later, it was found that the GPU could do that as well. Is it possible that the iGPU might do some part of the work better or worse than another part which could be delegated to a different piece of hardware?

I.e. sieve on iGPU and TF on proper GPU if there is one available?

CPU sieving is still an option, also once we get the IntelHD to work with mfakto. This has never been removed from mfakt[co], it's just not the default configuration anymore.

Edit: sieving on one GPU and TF on the other is less likely to be efficient as the speeds would need to adjust to each other, leading to the same issues that CPU sieving has: the slower part may not keep up, wasting the faster one's resources.