新手问下关于CUDA效率的问题

2015-04-09 06:53:14 555 阅读 AI 写同款 · GEO 优化AI 选题数据监测一键分发 ›

cuda中atomicAdd效率问题 atomicAdd遇到冲突是开销会比较大，最好先用共享变量先累加一部分，然后再把共享变量加到全局变量，减少冲突。 { int j = blockIdx.x; int i = blockIdx.y; int k2=i*gridDim.x; int f=threadIdx.x; //0~5 int k = threadIdx.y; //0~31 at 阅读详情

试了一下一个简单的矩阵相加的并行程序感觉效率非常低下主要是memcpy和malloc太费时间了求教一下各位大神这种问题怎么解决？是由于计算量太少还是我代码的问题？

__global__ void addKernel(int *c, const int *a, const int *b)

{

	int i = blockIdx.x;

    int j = threadIdx.x;

	int idx = i*blockDim.x+j;

    c[idx] = a[idx] * b[idx];

	c[idx] += a[idx] * b[idx];

}

const int arraySize = 100;

int main()

{

    int a[arraySize][arraySize];

    int b[arraySize][arraySize];

    int c[arraySize][arraySize];

	int c1[arraySize][arraySize];



	for(int i = 0;i<arraySize;i++)

	{

		for(int j = 0;j<arraySize;j++)

		{

			a[i][j] = i;

			b[i][j] = 2*i;

		}

	}

	cudaError_t cudaStatus;

	// Choose which GPU to run on, change this on a multi-GPU system.

    cudaStatus = cudaSetDevice(0);



	clock_t t1 = clock();



    int *dev_a = 0;

    int *dev_b = 0;

    int *dev_c = 0;

    // Allocate GPU buffers for three vectors (two input, one output)    .

    cudaStatus = cudaMalloc((void**)&dev_c, arraySize * arraySize * sizeof(int));



    cudaStatus = cudaMalloc((void**)&dev_a, arraySize * arraySize * sizeof(int));



    cudaStatus = cudaMalloc((void**)&dev_b, arraySize * arraySize * sizeof(int));



    // Copy input vectors from host memory to GPU buffers.

    cudaStatus = cudaMemcpy(dev_a, a, arraySize * arraySize * sizeof(int), cudaMemcpyHostToDevice);



    cudaStatus = cudaMemcpy(dev_b, b, arraySize * arraySize  * sizeof(int), cudaMemcpyHostToDevice);

	

    // Launch a kernel on the GPU with one thread for each element.

    addKernel<<<arraySize, arraySize>>>(dev_c, dev_a, dev_b);



	

    // Check for any errors launching the kernel

    cudaStatus = cudaGetLastError();

    

    // cudaDeviceSynchronize waits for the kernel to finish, and returns

    // any errors encountered during the launch.

    cudaStatus = cudaDeviceSynchronize();



    // Copy output vector from GPU buffer to host memory.

    cudaStatus = cudaMemcpy(c, dev_c, arraySize * arraySize * sizeof(int), cudaMemcpyDeviceToHost);



	clock_t t2 = clock();

	cout<<"time:"<<t2-t1<<endl;



	for(int i = 0;i<arraySize;i++)

	{

		for(int j = 0;j<arraySize;j++)

		{

			c1[i][j] = a[i][j] + b[i][j];

 			c1[i][j] += a[i][j] * b[i][j];

 			c1[i][j] += a[i][j] * b[i][j];

 			c1[i][j] += a[i][j] * b[i][j];

		}

	}

	t1 = clock();

	cout<<"time:"<<t1-t2<<endl;	

	cudaFree(dev_c);

    cudaFree(dev_a);

    cudaFree(dev_b);

    cudaStatus = cudaDeviceReset();



    return 0;

}