c++ - 为什么这个 CUDA 内核给出与原始代码不同的结果？

Question

我移植了这段代码:

    if(_layersCount > 1)
    {
        for(int i=_layersCount-2;i>=0;i--)
        {
            for(int j=0;j<_neuronsPerLayerCount[i];j++) // cuda kernel
            {
                localGradients[indexByLayerAndNeuron(i, j)] = 0;

                for(int k=0;k<_neuronsPerLayerCount[i+1];k++)
                {
                    localGradients[indexByLayerAndNeuron(i, j)] += _neuronsInputsWeights[indexByLayerNeuronAndInput(i+1, k, j)]
                                                                    * localGradients[indexByLayerAndNeuron(i+1, k)];
                }

                localGradients[indexByLayerAndNeuron(i, j)] *= derivatives[indexByLayerAndNeuron(i, j)];
            }
        }
    }

到 CUDA:

    if(_layersCount > 1)
    {
        for(int i=_layersCount-2;i>=0;i--)
        {
            // calculateLocalGradientsForAnotherLayers
            blocksCount = floor((double) _neuronsPerLayerCount[i] / threads.x) + 1;
            blocks = dim3(blocksCount, 1);

            calculateLocalGradientsForAnotherLayers <<<blocks, threads>>> (deviceLocalGradients, _neuronsInputsWeights, deviceDerivatives, _neuronsPerLayerCount[i], _neuronsInPreviousLayers[i], _neuronsInPreviousLayers[i+1], _neuronsPerLayerCount[i+1], _inputsInPreviousLayers[i], _inputsInCurrentLayer[i]);
        }
    }

calculateLocalGradientsForAnotherLayers 内核:

__global__ void calculateLocalGradientsForAnotherLayers(double * localGradients, double * neuronsInputsWeights, double * derivatives, int neuronsCount, int neuronsInPreviousLayers, int neuronsInPreviousLayersWithCurrent, int neuronsInNextLayer, int inputsInPreviousLayers, int inputsInCurrentLayer)
{
    int idx = blockIdx.x * blockDim.x + threadIdx.x;

    if(idx < neuronsCount)
    {
        int neuron = neuronsInPreviousLayers + idx;

        localGradients[neuron] = 0;

        // this to Kernel, then reduce localGradients.
        for(int k=0;k<neuronsInNextLayer;k++)
        {
            localGradients[neuron] += neuronsInputsWeights[inputsInPreviousLayers + k*inputsInCurrentLayer + idx]
                                                            * localGradients[neuronsInPreviousLayersWithCurrent + k];
        }

        localGradients[neuron] *= derivatives[neuron];
    }
}

但是我从小数点后第二位看到了结果的差异。为什么误差这么大？除此以外，所有内核都运行良好。

我的 GPU 是 NV GF555M。它支持 double 。

Answer 1

在内核主体中，您需要通过 localGradients 数组进行某种 block 间同步:

for(int k=0;k<neuronsInNextLayer;k++)
        {
            localGradients[neuron] += neuronsInputsWeights[inputsInPreviousLayers + k*inputsInCurrentLayer + idx]
                                                            * localGradients[neuronsInPreviousLayersWithCurrent + k];
        }

并发读/写访问可能会破坏 localGradients 元素的实际值。由于读/写没有同步，您可能会看到随机结果。