Question

如何加载文本文件的行，这些行＆＃39;长度和读取线的数量，以正确的方式投射它们并将它们传递给GPU？

含义：

输入文字文件：

Line1
Line2
..
LineN

app.cl

#define UP_LIMIT 256

typedef struct {
  uint bar; // amount of read lines
} foomatic;

typedef struct {
   uint len; // length of line
   uchar s[UP_LIMIT]; // line
} foo;

__kernel foobar(__global foo * kernel_in, __global foomatic kernel_in2){
 // do something
}

的main.cpp

#define LLEN 256
typedef struct {
   cl_uint bar; // amount of read lines
} foomatic;
typedef struct {
   cl_uint len; // length of line
   cl_uchar s[LLEN]; // line
} foo;
int main(){
// load from file
// count lines of file
foo * input = new foo[N]; // N is the amount of lines of source text file
// cast line to cl_uchar

// pass lines, their lengths and number of lines to ocl kernel
delete [] input;    
}

Answer 1

在OpenCL中使用struct似乎很棘手，因为可以在CPU和设备上以不同的方式打包字段。此外，它在GPU上效率不高，因为它会影响对内存访问进行分组的能力。更好的方法是使用多个数组。我有这样的事情：

//N is the number of strings.
cl_uint* strlens = new cl_uint[N];
cl_uchar* input = new cl_uchar[N * LLEN];
cl_int err_code = CL_SUCCESS;
//Remember to check error codes, omitted here for convenience.
cl_mem strlens_buffer = clCreateBuffer(context, CL_MEM_READ_ONLY, N*sizeof(cl_uint), NULL, &err_code);
cl_mem input_buffer = clCreateBuffer(context, CL_MEM_READ_ONLY, N*LLEN*sizeof(cl_uchar), NULL, &err_code);

//Some initialisation code...

//Send to device.
err_code = clEnqueueWriteBuffer(queue, strlens_buffer, CL_TRUE, 0, N*sizeof(cl_uint), strlens, 0, NULL, NULL);
err_code = clEnqueueWriteBuffer(queue, input_buffer, CL_TRUE, 0, N*LLEN*sizeof(cl_uchar), 0 NULL, NULL);

//Send work to the GPU...

//Clean up.
delete[] strlens;
delete[] input;

我已经将上下文和队列用于与您的设备关联的OpenCL上下文以及与该上下文关联的命令队列。

如何将字符串行作为cl_uint数组和文件详细信息传递给GPGPU？

1 个答案: