Question

I’m trying to copy some nested structs to device memory for kernel use in a CUDA-accelerated neural network simulator. This code links and runs, but it throws some exceptions and CUDA errors:

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
[...]

Unfortunately, the cudaMemcpyToSymbol call returns an error that the mCheckCudaWorked macro says is “invalid device symbol”, while the last (cudaMemcpy( &dummy, &rSes.rLearn->gpuDOutputs…) and third-from-last (cudaMemcpy( &dummy, &rSes.rLearn->gpuXInputs…) cudaMemcpy calls return “invalid argument”.

I am at a loss as to how to proceed to get these items copied to device memory and addressable from kernel code. &dummy and &rdldummy are positively being returned as the pointers to the device memory addresses where the allocated memory awaits, and I can write those pointers to the device memory, but I cannot coax the bulk of the member values into being copied to the pointed-at allocations. Help?

Answer 1

Fields like gpuXInputs need to be point at memory which has been allocated with cudaMalloc, so that they are valid pointers to device memory.

Typically you need a host version of your data structures, where your allocations use malloc etc, and then a mirror of these data structures on the device, which have been allocated via cudaMalloc. Any pointers within these data structures need to point to the right kind of memory – you can’t “mix and match”.