OpenCL compile fails about aparapi HOT 16 CLOSED

Interestingly, if you comment out these lines:

if(s_i == t_j) {
   cost = 0;
}
else {
    cost = 1;
}

...and then just initialise cost to 0 on its declaration, the code will compile 
and run (though obviously won't generate the correct result.) It seems to be 
the if statement causing the error which seems a bit odd?

I found a workaround in that the equivalent ternary statement works fine. That 
surprises me - I assumed that they'd compile the same way, is there a subtle 
difference?

Sorry, forgot to include the equivalent ternary statement above:

cost = s_i == t_j ? 0 : 1;

Actually, I get the following error:


clBuildProgram failed
************************************************
<program source>:45:70: error: use of undeclared identifier 'cost'
             int c = this->val$d[((firstSize * (i - 1)) + (j - 1))] + cost;
                                                                      ^

************************************************


Feb 6, 2012 12:49:52 PM com.amd.aparapi.KernelRunner warnFallBackAndExecute
WARNING: Reverting to Java Thread Pool (JTP) for class gov.pnnl.ia.gpu.test$1: 
OpenCL compile failed

Which makes sense since you cannot declare variables without initializing them 
first...we've run into this same limitation with our kernels as well.

FYI, if you use the Trunk code char support has been added :)

Ah I see! That does make a lot of sense. And thanks for the tip about the char 
support - don't suppose 2D arrays is in there as well?

Oh and how do you get the specific OpenCL error displayed? Is it an option I 
need to add? I'm just getting the "reverting to JTP" warning.

Thanks for reporting this. 

So for me when I initialize cost to 0 (line 25) I don't have to comment out the 
two lines or switch to the ternary form. 

So on line 25 (of your code) I just swapped 'int cost=0;' for 'int cost;'

Can you try this and just confirm. 

If you still see a compile error try setting 
-Dcom.amd.aparapi.enableShowGeneratedOpenCL=true this should dump the generated 
opencl and might help us diagnose. 

Are you using eclipse's java compiler or Oracles?

There is a subtle difference parsing the ternary form of "a = b?c:d;" instead 
of "if (b)c;else d; but I too would expect to have the same issue parsing the 
bytecodes. 

Essentially the stack is offset (by both conditional paths) after executing 
"b?c:d" wheres a normal if()..else.. we don't expect the stack to be effected 
by following either path.

The problem stems from not being able to determine the scope of the 'cost' 
variable unless we initialize it. 

Gary

Thanks for the quick feedback guys.

Gary, yup I can confirm that works if I change it as you describe (Oracle 
JDK7.) It does seem to be a scoping problem - I'm not a compiler guy at all, 
but seem to remember hearing a talk somewhere that explained determining the 
scope in the Java compiler for these sorts of cases was a decidedly complex 
problem!

On a related point, in the attached code I've tried to take the working example 
and apply it to work with a number of strings (so it has a practical use.) 
However, I end up with a generic NullPointerException rather than a specific 
OpenCL compile error:

Exception in thread "main" java.lang.NullPointerException
    at com.amd.aparapi.KernelWriter.convertType(KernelWriter.java:111)
    at com.amd.aparapi.KernelWriter.write(KernelWriter.java:260)
    at com.amd.aparapi.KernelRunner.execute(KernelRunner.java:1197)
    at com.amd.aparapi.Kernel.execute(Kernel.java:1523)
    at com.amd.aparapi.Kernel.execute(Kernel.java:1469)
    at com.amd.aparapi.Kernel.execute(Kernel.java:1454)
    at org.quelea.utils.opencl.LevenshteinDistance.compare(LevenshteinDistance.java:104)
    at org.quelea.utils.opencl.LevenshteinDistance.main(LevenshteinDistance.java:32)
Java Result: 1


I suspect this may be because there's 2D arrays involved, in the following line:

result[getGlobalId()] = ld(s1Arr[getGlobalId()], s2Arr[getGlobalId()], 
s1Length, s2Length);

...s1Arr and s2Arr are 2D arrays of integers. I "hoped" it'd just count as a 1D 
array since all I'm doing is retrieving the 1D array from the 2D array based on 
the global ID and using that, but is this not allowed either and hence is 
what's causing the issue?

Two dimensional arrays (well, arrays of arrays) will throw a NPE in the Aparapi 
code.

See http://code.google.com/p/aparapi/issues/detail?id=10

Sadly no we just can't access 2D arrays.

The NPE is actually bad coding on my behalf,  this code is only expecting so 
see primitives or single dimension arrays of primitives and should be more 
defensive.  I think it has been fixed in the main branch (it is time for me to 
create a new set of binary builds isn't it!) or at least it will generated an 
Aparapi exception complaining that we can't access 2D arrays.

As Ryan indicated (he is clearly faster than me at typing ;)) I think you are 
hitting issue #10.  http://code.google.com/p/aparapi/issues/detail?id=10

Moving 2D arrays to the GPU would be expensive due to the layout of Java arrays.

Although I think I know *how* to do it if someone wanted to try to implement 
it....

Gary

Just an FYI, when I receive the error above (uninitialized method local 
variables) and OpenCL dies, Aparapi never returns or fails over to JTP mode. 
Should I file a new bug on that condition and use the attachment on this ticket 
as the test case?

I also get that behaviour - it just seems to hang rather than falling over to a 
thread pool.

I just got around to trying to fix this.  I thought I had wrapped all tests in 
ClassParseExceptions, but this is not detected until OpenCL write time which is 
a shock.  I think this should be detected earlier.  

Worse case I will defend the KernelRunner from any exception and fall back.  I 
will also see if this particular error can't be detected earlier.  I need to 
create a junit case for this. 

Thanks 'berry120' 

I made the changes suggested above.  Now any exception encountered at codegen 
time is converted to CodeGenException so should revert to JTP.  Fix is in SVN 
R#287