topic Something else; in OpenCL* for CPU

Compiler error

Polar01 — Thu, 28 Nov 2013 20:50:47 GMT

Hi,

I have some strange behavior with my OpenCL kernel, by example I have the following code that give me the wrong result :

if ( condition )
{
printf("Enter here ...");
my code;
}

Where 'condition = false' ..

When I remove the code enclosed into the brackets the results are correct.
I simply add some "useless" code and the result is really different.

I work on this since several days and have try with other SDKs to insure that I'm not wrong. It works perfectly with other SDK !

BTW, I cannot show the complete kernel, it is too big, also it is impossible to create a repro example because it is to difficult to handle.

What I can do:
1) Provide a remote control to my machine, allowing your team to debug and see the problem directly
2) Generate the assembly code to help you compare the difference and see the error

Please contact me asap to fix the problem together ;-)

BTW,

Polar01 — Thu, 28 Nov 2013 21:07:33 GMT

BTW,

I have compile the good kernel and the same one but with one addition (useless) line... and the code difference is huge !

Contact me to got the kernel code, the llvm and the asm one.

Thanks

Something else;

Polar01 — Thu, 28 Nov 2013 21:29:10 GMT

Something else;

If I disable the optimization, at least in some case (I have just do one tes), it seems to fix the problem.

I will do more tests tomorrow...

BAD NEWS !!!!!!!!!!!!!!!!!!!!

Polar01 — Fri, 29 Nov 2013 11:36:48 GMT

BAD NEWS !!!!!!!!!!!!!!!!!!!!!!

I have do the test on another machine (but still a i7 CPU) and I have different results, even with the -cl-opt-disable" option !

Also, it depends of the "test unit" I do... some works some does not work :-(

Please help me... it is CRITICAL !

Is this on the CPU or GPU?

Raghupathi_M_Intel — Sun, 01 Dec 2013 22:06:05 GMT

Is this on the CPU or GPU? Please provide details on the driver version, OS etc., that'll help us reproduce the problem.

Thanks,
Raghu

Hi Raghu,

Polar01 — Mon, 02 Dec 2013 08:10:18 GMT

Hi Raghu,

I have test it on 2 machine, one with i7 CPU and another with a i7 CPU+APU.

It seems to me impossible to create a "reproducible" version of it, or I have to send you all the application code, it is hughe and complex !

But, what I will propose is to organize a Team Viewer session to allow you to (a) see the bug in life (b) debug directly on the machine.

For this, I propose you to contact me directly by email, right ?

Krys