CODEWITHSUNDEEP
c++openclgpgpu
Kasun
Kasun

Reputation: 307

How to transfer C++ std::vector to openCL kernels?

I have implement matrix class using two dimensional vectors in C++ (vector<vector<float>>()). I now want to optimize the code using GPGPU using openCL. But i am runing in to problems every miniute. So Please help me and give me so tips to do this.

My requirements are follows

  1. Since I want to use this matrix library to implement machine learning algo there may be huge matrix, like 1000*400.
  2. Can i use two dimensional vectors and transfer them to opencl kernels (because if I can use the vectors to implement this class it would be much more easy than implementing these from scratch using array).

One of my code segments to do this as follows, here in my kernal I try to add 10 to every element.

But output shows that it only change the values in frist vector[0][n] elemets.

This is the segment in my host program....

int in_vec_size = 100;
int out_vec_size = 100;

vector<vector<float>> in_vec(10,vector<float>(10));
vector<vector<float>> out_vec(10, vector<float>(10));
int k = 0;

//initialize the input vec
for (int i=0; i < 10;i++)
{
    for (int j = 0; j < 10;j++)
    {
        in_vec[i][j] = k++;
        out_vec[i][j] = 0;
    }
}

//creating bufferes
cl::Buffer inBuff(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, in_vec_size*4, &in_vec[0][0]);
cl::Buffer outBuff(context, CL_MEM_WRITE_ONLY, out_vec_size*4, NULL);

//set kernal args
kernal.setArg(0, inBuff);
kernal.setArg(1, outBuff);
kernal.setArg(2, in_vec_size);

cl::CommandQueue queue(context, devices_gpu[0]);

queue.enqueueTask(kernal);
queue.enqueueWriteBuffer(inBuff, CL_TRUE, 0, in_vec_size*4, &in_vec[0][0]);
queue.enqueueReadBuffer(outBuff, CL_TRUE, 0, out_vec_size*4, &out_vec[0][0]);

for (int i = 0; i < 10; i++)
{
    for (int j = 0; j < 10; j++)
    {
        cout << out_vec[i][j] << endl;
    }
}

__kernel void add(__global float*in,__global float*out,int x) 
{
   // i=get_global_id(0);
   for(int i=0;i<x;i++)
   {
      out[i] = in[i]+10;
   }
}

Upvotes: 0

Views: 1369

Answers (1)

Roman Arzumanyan
Roman Arzumanyan

Reputation: 1814

You're using multdimensonal vector.

It means, that outer vector contains inner vectors in continuous fashion. But content is a class, not a plain data. So, data, which you initialize OpenCL memory objects with, isn't continuous. You initialize cl_mem with insides of vector class implementation data, not matrix data.

Use single vector of size MxN instead. Take a look at this SO questionon.

Upvotes: 4

Related Questions