Hi everyone,

I am doing parallel programming(using cuda) over an year now and I am new to OpenCL. I was making my 1st openCL code (matrix multiplication ). I wrote the following code,

[#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <string.h>
#include <SDKCommon.hpp>
#include <SDKApplication.hpp>
#include <SDKCommandArgs.hpp>
#include <SDKFile.hpp>
#include <CL/cl.h>

#define MAX_SOURCE_SIZE (0x100000)
#define MATSIZE 16

void initmat(float *Aa,float *Bb,float *Cc,int row,int colrow,int col);

void initmat(float *Aa,float *Bb,float *Cc,int row,int colrow,int col)
unsigned int i;





int main(void)

// Load the kernel source code into the array source_str
FILE *fp;
char *source_str;
size_t source_size;

fp = fopen("matmul.cl", "r");
if (!fp) {
fprintf(stderr, "Failed to load kernel.\n");
source_str = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread( source_str, 1, MAX_SOURCE_SIZE, fp);
fclose( fp );

// matrix declaration
float *A;
float *B;
float *C;

// set dimesions
int Arow,AcolBrow,Bcol;


// no. of elements in matrix
int sizea, sizeb, sizec;

// Error code from opencl

int err;

// Setting up matrices
sizea= Arow*AcolBrow;
sizeb= AcolBrow*Bcol;
sizec= Arow*Bcol;

A = (float *) malloc(sizeof(float)*sizea);
B = (float *) malloc(sizeof(float)*sizeb);
C = (float *) malloc(sizeof(float)*sizec);


// get platform id & device id

cl_uint numplatform;
cl_platform_id platformid=NULL;
cl_device_id deviceid=NULL;

err= clGetPlatformIDs(1,&platformid,&numplatform);
err=clGetDeviceIDs(platformid,CL_DEVICE_TYPE_GPU,1 ,&deviceid,NULL);

cl_context_properties properties[]=
CL_CONTEXT_PLATFORM, (cl_context_properties)platformid,0

// create context
cl_context context= clCreateContext(properties,1,&deviceid,NULL,NULL,& err);

/* when more than one gpu is installed on the system than we make use of the approach as we stated in the notes !! */

// create command queue

cl_command_queue queue = clCreateCommandQueue(context,deviceid,0,&err); // I have disabled profiling option

// Allocate buffer object for Ad,Bd,Cd

cl_mem Ad = clCreateBuffer(context,CL_MEM_READ_ONLY,sizeof(cl_ float)*sizea,NULL,NULL);
cl_mem Bd = clCreateBuffer(context,CL_MEM_READ_ONLY,sizeof(cl_ float)*sizeb,NULL,NULL);
cl_mem Cd = clCreateBuffer(context,CL_MEM_WRITE_ONLY,sizeof(cl _float)*sizec,NULL,NULL);

// We are not explicitely making kernel. We are putting the kernel code here itself (see notes)

cl_program program= clCreateProgramWithSource(context,1,(const char **)&source_str, (const size_t *)&source_size,&err);

// Build program using program object just created

err = clBuildProgram(program,0,NULL,NULL,NULL,NULL);

if(err !=CL_SUCCESS)

size_t len;
char buffer[2048];
printf("ERROR: Failed to build executable \n ");
clGetProgramBuildInfo(program,deviceid,CL_PROGRAM_ BUILD_LOG, sizeof(buffer), buffer , &len);
printf("%s \n",buffer);
//return FAILURE;


// Create kernel object

cl_kernel kernel = clCreateKernel(program,"matmul",NULL);

// set kernel argument values

err= clSetKernelArg(kernel,0,sizeof(int),&Arow);
err|= clSetKernelArg(kernel,1,sizeof(int),&AcolBrow);
err|= clSetKernelArg(kernel,2,sizeof(int),&Bcol);
err|= clSetKernelArg(kernel,3,sizeof(cl_mem),&Ad);
err|= clSetKernelArg(kernel,4,sizeof(cl_mem),&Bd);
err|= clSetKernelArg(kernel,5,sizeof(cl_mem),&Cd);

// Write to device buffers. Ad=A and Bd=B : Equivalent to CUDAmemcpy

err=clEnqueueWriteBuffer(queue,Ad,CL_TRUE,0,sizeof (cl_float)*sizea,A,0,NULL,NULL);
err=clEnqueueWriteBuffer(queue,Bd,CL_TRUE,0,sizeof (cl_float)*sizeb,B,0,NULL,NULL);

// since we have set the copy as synchronous we will be creating event
cl_event event;

// Execute the kernel over entire range of C matrix

size_t global[2];
size_t local[2];
cl_uint * ndim; // no. of dimension in ND range. 3rd parameter in kernel call signifies the dimension.


* ndim=2; // because we want 2-D multiplication. Gives n

/* no local size declaration cause we are not making work groups ie blocks.We are just make making oneblock where everythread takes one element of A,B and computes C */

err = clEnqueueNDRangeKernel(queue,kernel,*ndim,NULL,glo bal,NULL,0,NULL,&event); // the NULL position after global is for passing local dimension. In this case we don't have one.
clFinish(queue); // wait for kernel to finish before we begin copying the result back on host

//read back the result

err=clEnqueueReadBuffer(queue,Cd,CL_TRUE,0,sizeof( cl_float)*sizec,C,0,NULL,NULL);

// free all memory


return 0;

My kernel code is as follows,
__kernel void matmul(const int Mdim, const int Ndim,const int Pdim,__global float* A,__global float* B,__global float* C)

float tmp;
//*int k;
// int i = get_global_id(0);
//int j = get_global_id(1);
//C[i*Ndim+j] = 3.0;


I could compile it successfully and it created a binary in ../bin/x86/ folder. When I try to run it using ./matmul it throes the following error,

93 > Sun Mar 17 : 04:22 PM : samkit@samkit:~/AMD/AMD-APP-SDK-v2.8-RC-lnx32/samples/opencl/bin/x86$ ./matmul
Segmentation fault (core dumped)

I know functionality of kernel is no where near that of matrix multiplication kernel but I did this to just check if these is some error in my kernel. Please give suggestions or advice that can help me make my code run.

Thanks in advance.