Cuda的：内核调用

Question

后，内存中的数据损坏请看看这个代码：

#include <stdlib.h> 
#include <stdio.h> 

int N, L, I; 
float * inputs; 
float * temp; 

// first kernel 
__global__ void mulKernel (float * output, float * inputs)///, float * weights) 
{ 
    int idx = blockIdx.x * blockDim.x + threadIdx.x; 

    output [idx] = inputs [idx] * 3;//weights [idx]; 
    //weights [idx] = 4; 

    //__syncthreads(); 
} 

//second kernel 
__global__ void sumKernel (float * output, float * input) 
{ 
    int idx = blockIdx.x * blockDim.x + threadIdx.x; 
    output [idx] = input[idx]*2; 

    __syncthreads(); 
} 

void printVector (const float *p, const int N) { 
    for (int i=0; i<N; i++) 
    printf("%f\n",p[i]); 
} 

int main(int argc, char *argv[]) 
{ 
    if(argc < 3) 
     printf("Usage: cuda <layers> <inputs>\n"); 
    else 
    { 
     L = atoi(argv[1]); 
     N = atoi(argv[2]); 
     I = atoi(argv[2]); 
     inputs = (float*)malloc(I*sizeof(float)); 
     float * weights = (float*)malloc(I*sizeof(float)); 

     // and fill with some arbitrary values 
     for (int i=0; i<I; i++) 
     { 
      inputs[i] = 1; 
     } 
     for (int i=0; i<I; i++) 
     { 
      weights[i] = 1.5; 
     } 

     // allocate device memory 
     float * devInputs = NULL; 
     float * devTemp = NULL; 
     float * devWeights = NULL; 

     cudaMalloc ((void**)&devInputs, I*sizeof(float)); 
     cudaMalloc ((void**)&devTemp, I*sizeof(float)); 
     cudaMalloc ((void**)&devWeights, I*sizeof(float)); 

     // set kernel launch configuration 
     dim3 threadsMul = dim3(512, 1); 
     int blocksCount = floor(I/threadsMul.x) + 1; 
     dim3 blocksMul = dim3(blocksCount, 1); 

     dim3 threadsSum = dim3(512, 1); 
     blocksCount = floor(I/threadsSum.x) + 1; 
     dim3 blocksSum = dim3(blocksCount, 1); 

     cudaMemcpy  (devInputs, inputs, I*sizeof(float), cudaMemcpyHostToDevice); 
     cudaMemcpy  (devWeights, weights,I*sizeof(float), cudaMemcpyHostToDevice); 

     //kernels calling in this cycle 
     for(int j=0;j<L;j++) 
     { 
      // copying data to see that's ok 
      cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 
      cudaMemcpy  (weights, devWeights, I*sizeof(float), cudaMemcpyDeviceToHost); 

      // print it 
      printf("inputs:\n"); 
      printVector (inputs, N); 
      printf("weights:\n"); 
      printVector (weights, N); 
      printf("\n"); 

      // running first kernel 
      mulKernel<<<blocksMul, threadsMul>>>(devTemp, devInputs);//, devWeights); 

      // copying and printing data. We can see thats array weights contains a wrong values 
      cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 
      cudaMemcpy  (weights, devWeights, I*sizeof(float), cudaMemcpyDeviceToHost); 

      printf("inputs:\n"); 
      printVector (inputs, N); 
      printf("weights:\n"); 
      printVector (weights, N); 
      printf("\n"); 

      if(cudaDeviceSynchronize() == cudaSuccess) 
      printf("threads syncronized\n"); 

      cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 
      cudaMemcpy  (weights, devWeights, I*sizeof(float), cudaMemcpyDeviceToHost); 

      printf("inputs:\n"); 
      printVector (inputs, N); 
      printf("weights:\n"); 
      printVector (weights, N); 
      printf("\n"); 

      sumKernel<<<blocksSum, threadsSum>>>(devInputs, devTemp); 

      cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 
      cudaMemcpy  (weights, devWeights, I*sizeof(float), cudaMemcpyDeviceToHost); 

      printf("inputs:\n"); 
      printVector (inputs, N); 
      printf("weights:\n"); 
      printVector (weights, N); 
      printf("\n\n"); 

      if(cudaDeviceSynchronize() == cudaSuccess) 
      printf("threads syncronized\n"); 

      cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 
      cudaMemcpy  (weights, devWeights, I*sizeof(float), cudaMemcpyDeviceToHost); 

      printf("inputs:\n"); 
      printVector (inputs, N); 
      printf("weights:\n"); 
      printVector (weights, N); 
      printf("\n\n"); 
     } 

     cudaMemcpy  (inputs, devInputs, I*sizeof(float), cudaMemcpyDeviceToHost); 

     cudaFree   (devInputs ); 
     cudaFree   (devTemp ); 
     cudaFree   (devWeights ); 

     printVector (inputs, N); 

     free(inputs); 
     free(weights); 
    } 
    return 0; 
} 

，并期待输出。在callig第一个内核之后，devWeights数组丢失了它的数据。但它并没有在任何地方使用。我只是将它复制到内存中，运行内核（不影响它）并将其复制回主机。在输出中，我看到它发生了变化。为什么？我究竟做错了什么？

在主函数中可以看到循环。其中我运行了两个内核：sumKernel和mulKernel。在运行内核之前，之后和同步线程之后，我将数组复制到主机并进行打印。所以，我在调用内核后看到错误的数据。请参阅代码中的评论。

我没有看到任何错误（只有cudaSuccess）。

Answer 1

哦，我发现了错误。我忘记在我的内核中使用if（idx < N），并且CUDA在输出数组维度时没有打印错误。所以，当我改变输入数组时，我也改变了输入后位于内存中的数据。