例如,我分配以下指针:
float *data_1, *data_2, *data_3, *data_4;
//Use malloc to allocate memory and fill out some data to these pointers
......
//Filling complete
float *data_d1,*data_d2,*data_d3,*data_d4;
cudaMalloc((void **)&data_d1,size1);
cudaMalloc((void **)&data_d2,size2);
cudaMalloc((void **)&data_d3,size3);
cudaMalloc((void **)&data_d4,size4);
cudaMemcpy(data_d1,data_1,size1,cudaMemcpyHostToDevice);
cudaMemcpy(data_d2,data_2,size2,cudaMemcpyHostToDevice);
cudaMemcpy(data_d3,data_3,size3,cudaMemcpyHostToDevice);
cudaMemcpy(data_d4,data_4,size4,cudaMemcpyHostToDevice);
在此之后,我应该已经获得了4个设备指针,其中包含与主机指针相同的确切数据.现在我想将这些指针存储到一个指针数组中,如下所示,
float *ptrs[4];
ptrs[0] = data_d1;
ptrs[1] = data_d2;
ptrs[2] = data_d3;
ptrs[3] = data_d4;
现在我想将这个指针数组传递给CUDA内核.但是,我知道由于ptrs [4]实际上是在主机内存上,我需要在设备上分配一个新指针.所以我这样做了,
float **ptrs_d;
size_t size = 4 * sizeof(float*);
cudaMalloc((void ***)&ptrs_d,size);
cudaMemcpy(ptrs_d,ptrs,size,cudaMemcpyHostToDevice);
然后调用内核:
kernel_test<<<dimGrid,dimBlock>>>(ptrs_d, ...);
//Declaration should be
//__global__ void kernel_test(float **ptrs_d, ...);
在kernel_test中,使用以下语法加载数据:
if (threadIdx.x < length_of_data_1d) …