Driver Error?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

here I have a simple programm, that sometimes works and sometimes fails. More accurate, I have two GTX580 mounted on one board.

When I run the code on decive 0, the example fails.

When I run the code on decive 1, the example works.

Has anyone a idea what might be the issue?

The code shall write 1.0 in an array of size[750][56][56], thats it.

I use official OpenCL 1.1 C++ Bindings.

const unsigned int deviceID = 0;

	// Platform

	VECTOR_CLASS< cl::Platform > platformList;

	cl::Platform::get(&platformList);

	cl::Platform thePlatform = platformList[0];

	// Device

	VECTOR_CLASS< cl::Device > gpuList;

	thePlatform.getDevices(CL_DEVICE_TYPE_GPU, &gpuList);

	cl::Device theDevice = gpuList[deviceID];

	// Context & Q

	cl::Context theContext(	VECTOR_CLASS< cl::Device >(1,theDevice) );

	cl::CommandQueue theQueue(theContext,theDevice);

	// Kernel

	const string kernelString="\n\

	__kernel\n\

	void ErrorKernel(	__global float data[750][56][56] )\n\

	{\n\

		const size_t threadblocknr = get_group_id(0);\n\

		const size_t lid0 = get_local_id(0);\n\

		const size_t lid1 = get_local_id(1);\n\

		const size_t element_k = threadblocknr*1+lid0;\n\

		for (size_t col=0; col < 56; col++)\n\

		{\n\

			data[element_k][col][lid1] = 1.0f;\n\

		}\n\

	}";

	cout << kernelString << endl;

	// todo check +1

	cl::Program::Sources theSource (1,make_pair(kernelString.c_str(), kernelString.length()+1));

	cl::Program theProgram (theContext, theSource);

	theProgram.build( VECTOR_CLASS< cl::Device >(1,theDevice) );

	cl::Kernel theKernel(theProgram,"ErrorKernel");

	// Kernel Arguments

	cl::Buffer data(theContext,CL_MEM_READ_WRITE,sizeof(cl_float)*750*56*56);

	theKernel.setArg(0,data);

	// Enqueue Kernel

	const cl::NDRange offset = cl::NullRange;

			const cl::NDRange local (     1,56);

			const cl::NDRange global( 750*1,56);

	theQueue.finish();	// safety first

	theQueue.enqueueNDRangeKernel(theKernel,offset,global,local);

	theQueue.finish();

	// read back data

	float* ones = new float [750*56*56];

	theQueue.enqueueReadBuffer(data,CL_TRUE,0,sizeof(cl_float)*750*56*56,ones);

	theQueue.finish();	// safety first

	// sum over all entries

	float sum = 0.0;

	for(unsigned int i=0;i<750*56*56;i++) sum += ones[i];

	if (sum == 750*56*56) cout << "SUCCESS: sum = 2352000"<<endl;

	else cout << "ERROR: sum = " << sum << endl;

cudatoolkit_3.2.16_linux_64_ubuntu10.04

SDK Version: gpucomputingsdk_3.2.16_linux

Driver: 260.19.36 on a GTX580

Thanks.