如何将嵌套结构的成员复制到 CUDA 设备的内存空间？

Question

我试图将一些嵌套结构复制到设备内存中，以便在cuda加速的神经网络模拟器中使用内核。这段代码链接并运行，但它抛出一些异常和CUDA错误:

typedef struct rdLayer
{
    long NeuronQty ;
    long DendriteQty ;
    cuDoubleComplex *gpuWeights ;
    cuDoubleComplex *gpuZOutputs ;
    cuDoubleComplex *gpuDeltas ;
    cuDoubleComplex *gpuUnWeights ;
} rdLayer;
typedef struct rdNetwork
{
    long SectorQty;
    double K_DIV_TWO_PI;
    double two_pi_div_sect_qty;
    cuDoubleComplex *gpuSectorBdry;
    long LayerQty;
    rdLayer *rLayer;
} rdNetwork;
struct rdLearningSet 
{
    long EvalMode ;
    long SampleQty ;
    long InputQty ;
    long OutputQty ;
    long ContOutputs ;
    long SampleIdxReq ;
    cuDoubleComplex *gpuXInputs ;
    cuDoubleComplex *gpuDOutputs ;
    cuDoubleComplex *gpuYOutputs ;
    double *gpudSE1024 ;
    cuDoubleComplex *gpuOutScalar ;
};
[...]
    struct rdLearningSet * rdLearn;
    struct rdNetwork * rdNet;
[...]
    cudaMalloc(&rdNet, sizeof(rdNetwork));
    cudaMalloc(&rdLearn, sizeof(rdLearningSet));
[...]
    cuDoubleComplex * dummy;
    struct rdLayer rdlSource, * rdldummy;
[...]
    //rdLayer *rLayer;
    cudaMalloc(&rdldummy, sizeof(rdLayer)*rSes.rNet->LayerQty);
    cudaMemcpy( &rdNet->rLayer, &rdldummy, sizeof(rdLayer*), cudaMemcpyHostToDevice);
    for (int L=1; L<rSes.rNet->LayerQty; L++){
            // construct layer to be copied
            rdlSource.NeuronQty=rSes.rNet->rLayer[L].iNeuronQty 
            rdlSource.DendriteQty=rSes.rNet->rLayer[L].iDendriteQty 
            cudaMalloc( &rdlSource.gpuWeights, sizeof(cuDoubleComplex) * (rSes.rNet->rLayer[L].DendriteQty+1) * (rSes.rNet->rLayer[L].NeuronQty+1) ) 
                    mCheckCudaWorked
            cudaMalloc( &rdlSource.gpuZOutputs, sizeof(cuDoubleComplex) * (rSes.rNet->rLayer[L].DendriteQty+1) * (rSes.rNet->rLayer[L].NeuronQty+1) ) 
                    mCheckCudaWorked
            cudaMalloc( &rdlSource.gpuDeltas, sizeof(cuDoubleComplex) * (rSes.rNet->rLayer[L].iDendriteQty+1) * (rSes.rNet->rLayer[L].iNeuronQty+1) ) 
                    mCheckCudaWorked
            cudaMalloc( &rdlSource.gpuUnWeights, sizeof(cuDoubleComplex) * (rSes.rNet->rLayer[L].iDendriteQty+1) * (rSes.rNet->rLayer[L].iNeuronQty+1) ) 
                    mCheckCudaWorked
            //copy layer sructure to Device mem
            cudaMemcpyToSymbol( "rdNet->rLayer", &rdlSource, sizeof(rdLayer), sizeof(rdLayer) * L, cudaMemcpyHostToDevice );/*! 2D neuron cx weight matrix on GPU */
                    mCheckCudaWorked
    }
[...]   
    cudaMalloc(&dummy, sizeof(cuDoubleComplex) * (rSes.rLearn->SampleQty) * (rSes.rLearn->InputQty+1) ); /*! 2D complex input tuples in GPU. */
            cudaMemcpy( &rdLearn->gpuXInputs, &dummy, sizeof(cuDoubleComplex*), cudaMemcpyHostToDevice );
                    cudaMemcpy( &dummy, &rSes.rLearn->gpuXInputs, sizeof(cuDoubleComplex) * (rSes.rLearn->SampleQty) * (rSes.rLearn->InputQty+1), cudaMemcpyHostToDevice); 
                    mCheckCudaWorked        
    cudaMalloc(&dummy, sizeof(cuDoubleComplex) * (rSes.rLearn->SampleQty) * (rSes.rLearn->OutputQty+1) ); /*! 2D desired complex outputs in GPU. */
            cudaMemcpy( &rdLearn->gpuDOutputs, &dummy, sizeof(cuDoubleComplex*), cudaMemcpyHostToDevice );
                    cudaMemcpy( &dummy, &rSes.rLearn->gpuDOutputs, sizeof(cuDoubleComplex) * (rSes.rLearn->SampleQty) * (rSes.rLearn->OutputQty+1), cudaMemcpyHostToDevice); 
                    mCheckCudaWorked
[...]

不幸的是，cudaMemcpyToSymbol调用返回一个错误，mCheckCudaWorked宏说的是"无效的设备符号"，而最后(cudaMemcpy(&dummy， &rSes.rLearn->gpuDOutputs…)和倒数第三(cudaMemcpy(&dummy， &rSes.rLearn->gpuXInputs…)cudaMemcpy调用返回"无效参数"。

我在如何继续得到这些项目复制到设备内存和可寻址的内核代码的损失。dummy和rdldummy正作为指向设备内存地址的指针返回，其中分配的内存正在等待，并且我可以将这些指针写入设备内存，但是我无法将大部分成员值复制到指向的分配中。帮助吗?

Answer 1

字段gpuXInputs需要指向已分配给cudaMalloc的内存，因此它们是指向设备内存的有效指针。

通常，您需要数据结构的主机版本，其中您的分配使用malloc等，然后设备上的这些数据结构的镜像，已通过cudaMalloc分配。这些数据结构中的任何指针都需要指向正确的内存类型——你不能"混用和匹配"。