在本地存储阵列定义变量的大小，使用CUDA [英] define variable size on array in local memory, using CUDA

问题描述

是不是有点可能使一个列表，数组，一些在设备与列表/数组中的呼叫beeing一个参数的大小...或全局变量的在呼叫时初始化函数？

Is it somewhat possible to make a list, array, something in a device function with the size of the list/array beeing a parameter in the call… or a global variable that's initialized at call time?

我想是这样，这些列表中的一个工作：

I would like something like one of these list to work:

unsigned int size1;

__device__ void function(int size2) {

    int list1[size1];
    int list2[size2];
}

是否有可能做一些聪明的做出这样的工作？

Is it possible to do something smart to make something like this work?

推荐答案

有是分配共享内存的动态量1的方式 - 使用第三个发射内核参数：

There is 1 way to allocate dynamic amount of shared memory - to use third launch kernel parameter:

__global__ void kernel (int * arr) 
{
    extern __shared__ int buf []; // size is not stated
    // copy data to shared mem:
    buf[threadIdx.x] = arr[blockIdx.x * blockDim.x + threadIdx.x];
    // . . . 
}
// . . . 
// launch kernel, set size of shared mem in bytes (k elements in buf):
kernel<<<grid, threads, k * sizeof(int)>>> (arr);

有是许多阵列的一个黑客：

There is a hack for many arrays:

__device__ void function(int * a, int * b, int k) // k elements in first list
{
    extern __shared__ int list1 [];
    extern __shared__ int list2 []; // list2 points to the same point as list1 does

    list1 [threadIdx.x] = a[blockIdx.x * blockDim.x + threadIdx.x];
    list2 [k + threadIdx.x] = b[blockIdx.x * blockDim.x + threadIdx.x];
    // . . .
}

您必须考虑：分配给所有的块内存

You must take into account: memory allocated to all block.

这篇关于在本地存储阵列定义变量的大小，使用CUDA的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！