cuda生成汉宁窗代码

__global__ void hann(float * cos_window, int rows, int cols)//输入参数：数据指针，行数，列数
{
int tidx = threadIdx.x + blockDim.x*blockIdx.x;
int tidy = threadIdx.y + blockDim.y*blockIdx.y;
float row_charac = 0.5 - 0.5*(cos((2 * pi*tidx) / (rows - 1)));
float col_charac = 0.5 - 0.5*(cos((2 * pi*tidy) / (cols - 1)));
if (tidx<rows&&tidy<cols)
cos_window[tidx*cols + tidy] = row_charac*col_charac;
}