OpenCL筆記

OpenCL主要分成四個部份

Platform model -- device 和 host

Execution model -- kernel如何在device上執行

Memory model -- abstract memory hierarchy

Programming model -- 定義concurrency model如何map 到physical hardware


(1) Initialize platforms

clGetPlatformIDs(...,...,...);

(2)Initialize devices

clGetDeviceIDs(...,...,...,...,...);

(3) Create a context

clCreateContext(...,...,...,...,...,...,...);

(4) Create a command queue

clCreateCommandQueue(...);

(5) Create device buffers

clCreateBuffer(...);

(6) Write host data to device buffers

clEnqueueWriteBuffer(...);

(7) Create and compile the program

clCreateProgramWithSource(...);

clBuildProgram(...);

(8) Create the kernel

clCreateKernel(...);

(9) Set kernel arguments

clSetKernelArg(...);

(10) Configure the work-item structure

size_t globalWorkSize[1];

globalWorkSize[0]=2048;

(11) Enqueue the kernel for execution

clEnqueueNDRangeKernel(...);

(12) Read output results to host

clEnqueueReadBuffer(...);

(13) Release OpenCL Resources

clRealeaseKernel(...);
clRealeaseProgram(...);
...

OpenCL(1) vector add

//vector add 副程式  類似glsl中 fragment shader 和vertex shader呼叫方式


const char* programSource =
"__kernel                 \n"
"void vecadd(__global int *A,\n"
"            __global int *B,\n"
"            __global int *C) \n"
"{\n"

"   // Get the work-item unique ID                \n"
"   int idx = get_global_id(0);                   \n"
"\n"
"   // Add the corresponding locations of        \n"
"   // 'A' and 'B', and store the result in 'C'. \n"
"   C[idx] = A[idx] + B[idx];\n"
"}\n"
;

main主程式
...
kernel = clCreateKernel(program, "vecadd", &status);    //inline 方式call vecadd

...


Compile方法

g++ vectoradd.c -I/usr/local/cuda-5.0/include -L/usr/local/cuda-5.0/lib64 -lOpenCL