Stuck on why my OpenCL kernel won't execute with particular parameters

I have an OpenCL kernel that I run in JOCL and it passes all my JUnit tests. I ported my code into C++ so I could profile the kernel under the same conditions. The driver works fine in all cases except one. It runs perfectly fine in JOCL so I believe something in my C++ code is wrong. My code is below, I have audited it to death. If anyone can help me spot what is wrong I'd appreciate it.

The driver code works fine with args 1 and 2 as 8192, arg 3 as 512. It also works fine with args 1 and 2 as 512 and arg 3 as 8192. Arg 4 is always just 1, which sets the kernel to real numbers. When I set args 1 and 2 to 262144 and arg 3 to 16, it executes, no errors are reported, no seg faults, but the kernel does not change the data in the end. Note that arg 1*3 in all cases above is equal to 2^22. I believe I am allocating the same amount of floats in all cases. I am stumped. I can't get OpenCL to tell me what is wrong :(

void HelperFunctions::callKernel(int windowSize, int primitivesPerDataFrame, int nInFramesThisCall, int realOrComplex)
{
// OpenCL Vars
cl_platform_id platform;       // OpenCL platform
cl_device_id device;           // OpenCL device
cl_context gpuContext;         // OpenCL context
cl_command_queue commandQueue; // OpenCL command queue
cl_program clProgram;           // OpenCL program
cl_kernel clkernel;             // OpenCL kernel
void *dataHostBuffer;        // Host buffer
void *windowDataHostBuffer;        // Host buffer
cl_mem inData;   // OpenCL device buffer
cl_mem windowData;  // OpenCL device source buffer
size_t szKernelLength;        // Byte size of kernel code
cl_int errCode;                // Error code var

long gridX = 256;
long gridY = 16384;
long gridZ = 1;
size_t global_work_size[] = {gridX, gridY, gridZ};
size_t local_work_size[] = {gridX, 1, 1};
const char* cSourceCL = NULL;     // Buffer to hold source for compilation

// Allocate and initialize host arrays
dataHostBuffer = (void *)malloc(sizeof(cl_float) * primitivesPerDataFrame * nInFramesThisCall);
windowDataHostBuffer = (void *)malloc(sizeof(cl_float) * windowSize);

//Populate the data buffers
dataHostBuffer = generateRampData(primitivesPerDataFrame * nInFramesThisCall);

windowDataHostBuffer = blackman(windowSize);

//Get an OpenCL platform
errCode = clGetPlatformIDs(1, &platform, NULL);
cout << "Error Code: " << errCode << endl;

//Get the devices
errCode = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &device, NULL);
cout << "Error Code: " << errCode << endl;

//Create the contex开发者_Python百科t
gpuContext = clCreateContext(0, 1, &device, NULL, NULL, &errCode);
cout << "Error Code: " << errCode << endl;

// Create a command-queue
commandQueue = clCreateCommandQueue(gpuContext, device, 0, &errCode);

// Read the OpenCL kernel in from source file
cSourceCL = oclLoadProgSource("/home/djkasht/workspaceBlueprint/bp/bp-trunk/bundles/CopperShark/src/coppershark/dsp/blocks/opencl/dsp/window/Window.cl", "", &szKernelLength);

szKernelLength = strlen(cSourceCL);
// Create the program
clProgram = clCreateProgramWithSource(gpuContext, 1, (const char **)&cSourceCL, &szKernelLength, &errCode);
cout << "Error Code: " << errCode << endl;

// Build the program
errCode = clBuildProgram(clProgram, 0, NULL, NULL, NULL, NULL);
cout << "Error Code: " << errCode << endl;

size_t log_size = 1000000 * sizeof(char);
char build_log[log_size];
size_t len;
errCode = clGetProgramBuildInfo(clProgram, device, CL_PROGRAM_BUILD_LOG, log_size, build_log, &len);
cout << build_log << endl;

// Create the kernel
clkernel = clCreateKernel(clProgram, "window", &errCode);
cout << "Error Code: " << errCode << endl;

// Allocate the OpenCL buffer memory objects
inData = clCreateBuffer(gpuContext, CL_MEM_READ_WRITE, sizeof(cl_float) * primitivesPerDataFrame * nInFramesThisCall, NULL, &errCode);
cout << "Error Code: " << errCode << endl;
windowData = clCreateBuffer(gpuContext, CL_MEM_READ_ONLY, sizeof(cl_float) * windowSize, NULL, &errCode);
cout << "Error Code: " << errCode << endl;

// Set the Argument values
errCode = clSetKernelArg(clkernel, 0, sizeof(cl_mem), (void*)&inData);
cout << "Error Code: " << errCode << endl;
errCode = clSetKernelArg(clkernel, 1, sizeof(cl_mem), (void*)&windowData);
cout << "Error Code: " << errCode << endl;
errCode = clSetKernelArg(clkernel, 2, sizeof(cl_int), (void*)&windowSize);
cout << "Error Code: " << errCode << endl;
errCode = clSetKernelArg(clkernel, 3, sizeof(cl_int), (void*)&primitivesPerDataFrame);
cout << "Error Code: " << errCode << endl;
errCode = clSetKernelArg(clkernel, 4, sizeof(cl_int), (void*)&nInFramesThisCall);
cout << "Error Code: " << errCode << endl;
errCode = clSetKernelArg(clkernel, 5, sizeof(cl_int), (void*)&realOrComplex);
cout << "Error Code: " << errCode << endl;

// Asynchronous write of data to GPU device
errCode = clEnqueueWriteBuffer(commandQueue, inData, CL_FALSE, 0, sizeof(cl_float) * primitivesPerDataFrame * nInFramesThisCall, dataHostBuffer, 0, NULL, NULL);
cout << "Error Code: " << errCode << endl;

// Synchronous/blocking read of results, and check accumulated errors
errCode = clEnqueueWriteBuffer(commandQueue, windowData, CL_FALSE, 0, sizeof(cl_float) * windowSize, windowDataHostBuffer, 0, NULL, NULL);
cout << "Error Code: " << errCode << endl;

errCode = clEnqueueNDRangeKernel(commandQueue, clkernel, 3, NULL, &(global_work_size[0]), &(local_work_size[0]), 0, NULL, NULL);
cout << "Error Code: " << errCode << endl;

void* dataHostBuffer2 = (void *)malloc(sizeof(cl_float) * primitivesPerDataFrame * nInFramesThisCall);
errCode = clEnqueueReadBuffer(commandQueue, inData, CL_TRUE, 0, sizeof(cl_float) * primitivesPerDataFrame * nInFramesThisCall, dataHostBuffer2, 0, NULL, NULL);

}

UPDATE, I figured it out! The problem is in my kernel. I use constant memory. My java code accounts for this and textually manipulates the code so that if my buffer size for arg 2 > 16384, it changes the __constant to __global. I should have known this, but I forgot...