ParModelica: Error setting kernel arguments: CL_INVALID_ARG_SIZE

[Gustaf Thorslund]

After applying my workaround in #3324 I try to call a small function. The test is almost same as in #3242, but here I've added a phony input argument too (see below). Now tries to run a kernel for me, but fails.

2 CL_DEVICE_NAME :		Intel(R) HD Graphics Haswell M
2 CL_DRIVER_VERSION :	0.8.0
2 CL_DEVICE_MAX_COMPUTE_UNITS :	64
2 CL_DEVICE_MAX_CLOCK_FREQUENCY :	1000
2 CL_DEVICE_LOCAL_MEM_SIZE :	                                                              64 KB
2 CL_DEVICE_GLOBAL_MEM_SIZE:                                                              128 MB
2 CL_DEVICE_MAX_MEM_ALLOC_SIZE:                                                              128 MB
2 CL_DEVICE_MAX_PARAMETER_SIZE: 1024 MB
2 CL_DEVICE_MAX_WORK_GROUP_SIZE: 512 
2 CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE  : 0


- Select your device:      2
	OpenCL initialization :        1349.381000 ms
Error setting kernel arguments:
CL_INVALID_ARG_SIZE 
Error: setting argument nr:  1

Here comes the code:

package oclTest
  
  constant Integer globalSizes = 10;
  constant Integer localSizes = 2;

  parkernel function Kernel
    input Integer S;
    parglobal input Integer i[10];
    parglobal output Integer groupId[globalSizes];
    parglobal output Integer localId[globalSizes];
  protected
    Integer g;
  algorithm
    g := oclGetGlobalId(1);
    groupId[g] := oclGetGroupId(1);
    localId[g] := oclGetLocalId(1);
  end Kernel;
  
  function test
    output Integer groupId[globalSizes];
    output Integer localId[globalSizes];
  protected
    parglobal Integer p_groupId[globalSizes];
    parglobal Integer p_localId[globalSizes];
    Integer I[10];
    parglobal Integer pI[10];
  algorithm
    for i loop
      I[i] := i;
    end for;
    pI := I;
    oclSetNumThreadsGlobalLocal1D({globalSizes}, {localSizes});
      (p_groupId, p_localId) := Kernel(7,pI);
      groupId := p_groupId;
      localId := p_localId;
  end test;
end oclTest;

Together with the script used to call it:

// name: oclTest
// cflags: +g=ParModelica
// status: correct

// setGrammarParModelica();
setCommandLineOptions("+d=noevalfunc +g=ParModelica -v=1");
loadFile("oclTest.mo");
getErrorString();

(x,y):=oclTest.test();
getErrorString();

Trying a parfor gives same kernel failure. For example:

package ParForTest
  constant Integer globalSizes = 10;
  constant Integer localSizes = 2;
  constant Integer elements = 10;

  function test
    input Integer a;
    output Integer result[elements];
  protected
    Integer v[elements] = {i for i in 1:elements};
    parglobal Integer pv[elements];
    parglobal Integer pr[elements];
    parglobal Integer pa;
  algorithm
    oclSetNumThreadsGlobalLocal1D({globalSizes}, {localSizes});
    pa := a;
    pv := v;
    parfor i in 1 : elements loop
      pr[i] := pa * pv[i];
    end parfor;
    result := pr;
  end test;
end ParForTest;

Called with the script:

setCommandLineOptions("+d=noevalfunc +g=ParModelica -v=1");
getErrorString();
loadFile("parfor.mo");
getErrorString();

x:=ParForTest.test(5);
getErrorString();

[Gustaf Thorslund]

I've now tried this on our test machine fermi. Using the oclTest.mos above, I get the following three errors for the available three devices:

1 CL_DEVICE_NAME :		Intel(R) Xeon(R) CPU           E5520  @ 2.27GHz
1 CL_DRIVER_VERSION :	2.0
...
- Select your device:      1
	 :Using flags -I"/home/gusth/src/OpenModelica/build/include/omc/c/"
	OpenCL initialization :        19955.674000 ms
	Kernel Execution      :        0.380000 ms
Error copying device to device:
CL_MEM_COPY_OVERLAP 

""

2 CL_DEVICE_NAME :		Tesla M2050
2 CL_DRIVER_VERSION :	304.108
...
- Select your device:      2
	 :Using flags -I"/home/gusth/src/OpenModelica/build/include/omc/c/"
	OpenCL initialization :        3559.964000 ms
Error setting kernel arguments:
CL_INVALID_ARG_SIZE 
Error: setting argument nr:  1

3 CL_DEVICE_NAME :		Intel(R) Xeon(R) CPU           E5520  @ 2.27GHz
3 CL_DRIVER_VERSION :	1.2
...
- Select your device:      3
	 :Using flags -I"/home/gusth/src/OpenModelica/build/include/omc/c/"
	OpenCL initialization :        6130.132000 ms
	Kernel Execution      :        0.406000 ms
Error copying device to device:
CL_MEM_COPY_OVERLAP 

""

The parfor.mos test, does however work for the two Intel devices:

- Select your device:      3
	 :Using flags -I"/home/gusth/src/OpenModelica/build/include/omc/c/"
	OpenCL initialization :        4126.450000 ms
	Kernel Execution      :        0.198000 ms
{5,10,15,20,25,30,35,40,45,50}
""

It gives same CL_INVALID_ARG_SIZE error for the Tesla device.

[Gustaf Thorslund]

Update of test script shows the test actually works on some CPU drivers for OpenCL.

test/parmodelica/oclTest.mos

@@ -9 +9 @@
 
 (x,y):=oclTest.test();
 getErrorString();
+x;
+y;

Error copying device to device:
CL_MEM_COPY_OVERLAP 

""
{1,1,2,2,3,3,4,4,5,5}
{1,2,1,2,1,2,1,2,1,2}

[Mahder Alemseged Gebremedhin]

Fixed in fec5e5b91d3986fe9c24f4a56cec79d7613960b5/OMCompiler.
Test added in a243f9eb9b3638c7b233edbe1ad27a479af3db58/OpenModelica-testsuite

[Dietmar Winkler]

Sorting these closed tickets away from "Future". Since they were closed after the last 1.9.3 release, it's very likely that they should have been part of the 1.9.4 release.

[Martin Sjölund]

Milestone renamed

[Martin Sjölund]

Milestone renamed