Shared memory in OpenCL - c++

Question

I intend to perform vector manipulations and was trying a small dummy program with vector addition and multiplication. However, the code does not run due to limitations on my knowledge on shared memory. All the sources in the internet show 2D matrix operations which I cannot translate to my vector problems. Please try to explain where am I going wrong considering the fact I am a novice in OpenCL. The code is given below:
Host Code:
std::vector<cl::Platform> platforms;
std::vector<cl::Device> devices;
cl::Context context;
cl::CommandQueue queue;
cl::Program program;
cl::Kernel kernel;
cl::Platform::get(&platforms);
deviceUsed = 0;
cl_context_properties properties[] =
{ CL_CONTEXT_PLATFORM, (cl_context_properties)(platforms[0])(),0 };
context = cl::Context(CL_DEVICE_TYPE_ALL, properties);
devices = context.getInfo<CL_CONTEXT_DEVICES>();
queue = cl::CommandQueue(context, devices[deviceUsed]);
cl::Program::Sources source( 1, std::make_pair(kernel_source.c_str(), kernel_source.size()));
program = cl::Program(context, source);
program.build(devices);
std::vector < float > a;
std::vector < float > b;
std::vector < float > sum;
std::vector < float > prod;
int globalSize = 128;
int localSize = 16;
a.resize(globalSize);
b.resize(globalSize);
sum.resize(globalSize);
prod.resize(globalSize);
for (int i = 0; i < globalSize ; i++)
{
a[i] = 1.0f * i;
b[i] = 5.0f * i;
}
cl::Buffer buffer_A;
cl::Buffer buffer_B;
cl::Buffer buffer_sum;
cl::Buffer buffer_prod;
buffer_A = cl::Buffer (context, CL_MEM_READ_WRITE, sizeof(float) * globalSize);
buffer_B = cl::Buffer (context, CL_MEM_READ_WRITE, sizeof(float) * globalSize);
queue.enqueueWriteBuffer(buffer_A, CL_TRUE, 0, sizeof(float) * globalSize , &a[0]);
queue.enqueueWriteBuffer(buffer_B, CL_TRUE, 0, sizeof(float) * globalSize , &b[0]);
buffer_sum = cl::Buffer(context, CL_MEM_WRITE_ONLY, sizeof(float) * globalSize);
buffer_prod = cl::Buffer(context, CL_MEM_WRITE_ONLY, sizeof(float) * globalSize);
kernel.setArg(0, buffer_A);
kernel.setArg(1, buffer_B);
kernel.setArg(2, buffer_sum);
kernel.setArg(3, buffer_prod);
queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(globalSize/localSize), cl::NDRange(N), NULL);
queue.finish();
queue.enqueueReadBuffer(buffer_sum, CL_TRUE, 0, sizeof(float) * globalSize, &sum[0]);
queue.enqueueReadBuffer(buffer_prod, CL_TRUE, 0, sizeof(float) * globalSize, &prod[0]);
Kernel:
#define STRINGI(ker) #ker
std::string kernel_source = STRINGI(
__kernel void KernelAddMul(__global float* a, __global float* b, __global float* sum, __global float* prod)
{
unsigned int j = get_local_id(0);
int N = get_local_size(0);
unsigned int i = N * get_global_id(0) + j;
float locSum[N];
float locProd[N];
__local float Asub[N];
__local float Bsub[N];
for(int k = 0; k < N; k++){
Asub[k] = a[i];
Bsub[k] = b[i];
barrier(CLK_LOCAL_MEM_FENCE);
locSum[k] = Asub[k] + Bsub[k];
locProd[k] = Asub[k] * Bsub[k];
barrier(CLK_LOCAL_MEM_FENCE);
sum[i] = locSum[k];
prod[i] = locProd[k];
}
}
);

Answer 1

I suspect that your code does not run because your kernel does not compile.
The following lines are invalid:
int N = get_local_size(0);
float locSum[N];
float locProd[N];
__local float Asub[N];
__local float Bsub[N];
N must be a constant, you cannot dynamically size the arrays using get_local_size(0).
I strongly recommend that you use a standalone compiler to compile your kernels:
CodeXL is very good, as is the Intel
SDK for OpenCL.
Anything is better than trying to debug your kernel in an application!