将包含指向另一个结构体的指针的结构体传递给CUDA中的内核 [英] Passing an struct including a pointer to another struct, to kernel in CUDA

问题描述

我有两个结构

  struct collapsed {
 char ** seq; 
 int num; 
}; 
 
 
结构数据{
 collapsed * x; 
 int num; 
 int numblocks; 
 int * blocksizes; 
 float * regmult; 
 float * learnmult; 
}; 
  

我将其传递给我的内核;

  __ global__ void KERNEL（data * X，...）{
 ... 
 collapsed x = X-> x [0] // GIVES CUDA_EXPECTION_1：Lane Illegal Address 
} 
 
 data X; 
 // init X 
 data * X_dev; 
 cudaMalloc（（data **）& X_dev，sizeof（data））; 
 cudaMemcpy（X_dev，& X，sizeof（data），cudaMemcpyHostToDevice）; 
 KERNEL<<< ...>>>（X_dev，...）; 
  

此代码在内核代码中提供 CUDA_EXPECTION_1：Lane Illegal Address 什么是错误或什么是正确的方式做到吗？任何想法？

解决方案

您要解除引用设备上的主机指针。
X 是有效的设备指针。

但是当您复制 X 结构到设备，你复制了 x 和它，它包含一个主机指针。当取消引用该指针时：

 折叠x = X-> x [0] 
 ^这是解除引用x指针
  

设备代码引发错误。 >

有关详情，请参阅此处以及如何解决它的说明。

I have two structs as

struct collapsed {
    char **seq;
    int num;
};


struct data {
    collapsed *x;
    int num;
    int numblocks;
    int *blocksizes;
    float *regmult;
    float *learnmult;
};

I am passing it to my kernel as;

__global__ void KERNEL(data* X,...){
    ...
    collapsed x = X->x[0]; // GIVES CUDA_EXPECTION_1:Lane Illegal Address
}

data X;
//init X
data *X_dev;
cudaMalloc((data **) & X_dev, sizeof(data));
cudaMemcpy(X_dev, &X, sizeof(data), cudaMemcpyHostToDevice);
KERNEL<<<...>>>(X_dev,...);

This code gives CUDA_EXPECTION_1:Lane Illegal Address in the kernel code. What is wrong or what is the right way to do it ? Any idea?

解决方案

You're dereferencing a host pointer on the device. X is a valid device pointer.

But when you copied the X struct to the device, you copied x along with it, which contains a host pointer. When you dereference that pointer:

collapsed x = X->x[0];
                 ^ this is dereferencing the x pointer

the device code throws an error.

More detail is given here as well as instructions on how to fix it.

这篇关于将包含指向另一个结构体的指针的结构体传递给CUDA中的内核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！