cuda c ++错误：__device__函数呼叫无法配置

Question

我想实现基本远期函数的CDP（我将远期功能同时称为多次（也来自CUDA功能），因此我想使用CDP）

这里是我要运行的代码;

__device__ void NNFeedForwardNormalMultiple(double* __restrict__ values, double* __restrict__ weigths, double* result, int inputsize, int outputsize) {
    int idx = threadIdx.x + blockIdx.x * blockDim.x;
    int outputidx = idx / outputsize;
    int inputidx = idx % outputsize;

    if (outputidx >= outputsize || inputidx >= inputsize) {
        return;
    }

    atomicAdd(&result[outputidx], values[inputidx] * weigths[outputsize*outputidx + inputidx]);
}

__device__ void NNFeedForwardNormalActivate(double* __restrict__ biases, double* result, int size) {
    int idx = threadIdx.x + blockIdx.x * blockDim.x;

    if (idx >= size) {
        return;
    }

    result[idx] = 1.0 / (1.0 + exp(-(result[idx] + biases[idx])));
}

__global__ void NNFeedForwardNormal(double* __restrict__ values, double* __restrict__ weigths, double* result, double* __restrict__ biases, int inputsize, int outputsize) {
    int blocksize = (inputsize * outputsize + THREADS_PER_BLOCK - 1)/THREADS_PER_BLOCK;
    NNFeedForwardNormalMultiple<<<blocksize, THREADS_PER_BLOCK>>>(values, weigths, result, inputsize, outputsize);
    cudaDeviceSynchronize();
    NNFeedForwardNormalActivate<<<(outputsize + THREADS_PER_BLOCK - 1)/THREADS_PER_BLOCK, THREADS_PER_BLOCK>>>(biases, result, outputsize);
}

我还尝试从这样的设备函数运行该函数，但仍然给了我同样的错误；

__device__ void NNFeedForwardNormalMultiple(double* __restrict__ values, double* __restrict__ weigths, double* result, int inputsize, int outputsize) {
    int idx = threadIdx.x + blockIdx.x * blockDim.x;
    int outputidx = idx / outputsize;
    int inputidx = idx % outputsize;

    if (outputidx >= outputsize || inputidx >= inputsize) {
        return;
    }

    atomicAdd(&result[outputidx], values[inputidx] * weigths[outputsize*outputidx + inputidx]);
}

__device__ void NNFeedForwardNormalActivate(double* __restrict__ biases, double* result, int size) {
    int idx = threadIdx.x + blockIdx.x * blockDim.x;

    if (idx >= size) {
        return;
    }

    result[idx] = 1.0 / (1.0 + exp(-(result[idx] + biases[idx])));
}

__device__ void NNFeedForwardNormal(double* __restrict__ values, double* __restrict__ weigths, double* result, double* __restrict__ biases, int inputsize, int outputsize) {
    int blocksize = (inputsize * outputsize + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    
    NNFeedForwardNormalMultiple<<<blocksize, THREADS_PER_BLOCK>>>(values, weigths, result, inputsize, outputsize);
    NNFeedForwardNormalActivate<<<(outputsize + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK, THREADS_PER_BLOCK>>>(biases, result, outputsize);
}

__global__ void NNFeedForwardNormalWrapper(double* __restrict__ values, double* __restrict__ weigths, double* result, double* __restrict__ biases, int inputsize, int outputsize) {
    NNFeedForwardNormal(values, weigths, result, biases, inputsize, outputsize);
}

还尝试了

cudaLaunchKernel

函数，并使用全局而不是设备，但它们也没有起作用。我也使用-rdc = true标志，我的拱门是SM_75，它应该支持CDP

thanks！

Answer 1

驱动器函数调用。调试它们感觉就像在捕获动态内容时对

HTML2CANVAS进行故障排除。您是否检查了函数是否正确标记并从正确的上下文中调用？