Question

我有一些用于纹理对象分配和“主机到设备”复制的代码。这只是对答案here的修改。我没有明确使用流，只是cudaSetDevice()

此代码工作正常，但是，当我运行Visual Profiler时，我看到从主机到阵列的内存副本不是异步的。它们分别分配给自己的设备流，但是直到第一个完成（在2个GPU上运行）后，第二个才开始。我已经尝试过使用大图像，因此我确定它不会占用CPU的开销。

我的猜测是代码中需要同步才能停止CPU，但是我不知道是什么。如何使此循环异步？

MCVE：

    void CreateTexture(int num_devices,float* imagedata, int nVoxelX, int nVoxelY, int nVoxelZ ,cudaArray** d_cuArrTex, cudaTextureObject_t *texImage);

int main(void)
{

int deviceCount =0 ;
cudaGetDeviceCount(&deviceCount);

int nVoxelX=512;
int nVoxelY=512;
int nVoxelZ=512;
float* image=(float*)malloc(nVoxelX*nVoxelY*nVoxelZ*sizeof(float));

cudaTextureObject_t *texImg =new cudaTextureObject_t[deviceCount];
cudaArray **d_cuArrTex = new cudaArray*[deviceCount];

CreateTexture(deviceCount,image, nVoxelX,nVoxelY, nVoxelZ,d_cuArrTex,texImg);


}

实际功能：

void CreateTexture(int num_devices, float* imagedata, int nVoxelX, int nVoxelY, int nVoxelZ ,cudaArray** d_cuArrTex, cudaTextureObject_t *texImage)
{
    //size_t size_image=nVoxelX*nVoxelY*nVoxelZ;
    for (unsigned int i = 0; i < num_devices; i++){
        cudaSetDevice(i);

        //cudaArray Descriptor
        const cudaExtent extent = make_cudaExtent(nVoxelX, nVoxelY, nVoxelZ);
        cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();
        //cuda Array
        cudaMalloc3DArray(&d_cuArrTex[i], &channelDesc, extent);
        //cudaCheckErrors("Texture memory allocation fail");
        cudaMemcpy3DParms copyParams = {0};


        //Array creation
        copyParams.srcPtr   = make_cudaPitchedPtr((void *)imagedata, extent.width*sizeof(float), extent.width, extent.height);
        copyParams.dstArray = d_cuArrTex[i];
        copyParams.extent   = extent;
        copyParams.kind     = cudaMemcpyHostToDevice;
        cudaMemcpy3DAsync(&copyParams);
        //cudaCheckErrors("Texture memory data copy fail");


        //Array creation End
        cudaResourceDesc    texRes;
        memset(&texRes, 0, sizeof(cudaResourceDesc));
        texRes.resType = cudaResourceTypeArray;
        texRes.res.array.array  = d_cuArrTex[i];
        cudaTextureDesc     texDescr;
        memset(&texDescr, 0, sizeof(cudaTextureDesc));
        texDescr.normalizedCoords = false;
        texDescr.filterMode = cudaFilterModePoint;
        texDescr.addressMode[0] = cudaAddressModeBorder;
        texDescr.addressMode[1] = cudaAddressModeBorder;
        texDescr.addressMode[2] = cudaAddressModeBorder;
        texDescr.readMode = cudaReadModeElementType;
        cudaCreateTextureObject(&texImage[i], &texRes, &texDescr, NULL);
        //cudaCheckErrors("Texture object creation fail");
    }
}

Answer 1

我在代码中看到的两个主要问题是：

您的主机分配是可分页的分配。目标之一是主机内存的CUDA中的复制操作异步要求对主机内存进行固定分配。
您的创建纹理循环中还有其他同步操作。根据我的经验，设备分配操作（在这种情况下为cudaMalloc3DArray）正在同步。我没有运行测试来确定cudaCreateTextureObject是否正在同步，但是如果同步了，我不会感到惊讶。因此，我一般对异步的建议是使同步操作脱离循环。

对于您而言，我们可以按以下方式重构您的代码，从nvprof的角度来看，这似乎允许操作重叠：

$ cat t399.cu
void CreateTexture(int num_devices, float* imagedata, int nVoxelX, int nVoxelY, int nVoxelZ ,cudaArray** d_cuArrTex, cudaTextureObject_t *texImage)
{
    //size_t size_image=nVoxelX*nVoxelY*nVoxelZ;

    const cudaExtent extent = make_cudaExtent(nVoxelX, nVoxelY, nVoxelZ);
    for (unsigned int i = 0; i < num_devices; i++){
        cudaSetDevice(i);

        //cudaArray Descriptor
        cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();
        //cuda Array
        cudaMalloc3DArray(&d_cuArrTex[i], &channelDesc, extent);
        //cudaCheckErrors("Texture memory allocation fail");
        }
    for (unsigned int i = 0; i < num_devices; i++){
        cudaSetDevice(i);
        cudaMemcpy3DParms copyParams = {0};
        //Array creation
        copyParams.srcPtr   = make_cudaPitchedPtr((void *)imagedata, extent.width*sizeof(float), extent.width, extent.height);
        copyParams.dstArray = d_cuArrTex[i];
        copyParams.extent   = extent;
        copyParams.kind     = cudaMemcpyHostToDevice;
        cudaMemcpy3DAsync(&copyParams);
        //cudaCheckErrors("Texture memory data copy fail");
        }
    for (unsigned int i = 0; i < num_devices; i++){
        cudaSetDevice(i);
        //Array creation End
        cudaResourceDesc    texRes;
        memset(&texRes, 0, sizeof(cudaResourceDesc));
        texRes.resType = cudaResourceTypeArray;
        texRes.res.array.array  = d_cuArrTex[i];
        cudaTextureDesc     texDescr;
        memset(&texDescr, 0, sizeof(cudaTextureDesc));
        texDescr.normalizedCoords = false;
        texDescr.filterMode = cudaFilterModePoint;
        texDescr.addressMode[0] = cudaAddressModeBorder;
        texDescr.addressMode[1] = cudaAddressModeBorder;
        texDescr.addressMode[2] = cudaAddressModeBorder;
        texDescr.readMode = cudaReadModeElementType;
        cudaCreateTextureObject(&texImage[i], &texRes, &texDescr, NULL);
        //cudaCheckErrors("Texture object creation fail");
    }
    for (unsigned int i = 0; i < num_devices; i++){
        cudaSetDevice(i);
        cudaDeviceSynchronize();
    }
}

int main(void)
{
  int deviceCount =0 ;
  cudaGetDeviceCount(&deviceCount);

  int nVoxelX=512;
  int nVoxelY=512;
  int nVoxelZ=512;
  float* image;

  cudaHostAlloc(&image, nVoxelX*nVoxelY*nVoxelZ*sizeof(float), cudaHostAllocDefault);

  cudaTextureObject_t *texImg =new cudaTextureObject_t[deviceCount];
  cudaArray **d_cuArrTex = new cudaArray*[deviceCount];

  CreateTexture(deviceCount,image, nVoxelX,nVoxelY, nVoxelZ,d_cuArrTex,texImg);
}


$ nvcc -o t399 t399.cu
$ cuda-memcheck ./t399
========= CUDA-MEMCHECK
========= ERROR SUMMARY: 0 errors
$ nvprof --print-gpu-trace ./t399
==19953== NVPROF is profiling process 19953, command: ./t399
==19953== Profiling application: ./t399
==19953== Profiling result:
   Start  Duration            Grid Size      Block Size     Regs*    SSMem*    DSMem*      Size  Throughput  SrcMemType  DstMemType           Device   Context    Stream  Name
1.55311s  90.735ms                    -               -         -         -         -  512.00MB  5.5106GB/s      Pinned       Array  Tesla P100-PCIE         1         7  [CUDA memcpy HtoA]
1.55316s  90.640ms                    -               -         -         -         -  512.00MB  5.5163GB/s      Pinned       Array   Tesla K40m (1)         2        18  [CUDA memcpy HtoA]
1.55318s  85.962ms                    -               -         -         -         -  512.00MB  5.8165GB/s      Pinned       Array  Tesla K20Xm (2)         3        29  [CUDA memcpy HtoA]
1.55320s  89.908ms                    -               -         -         -         -  512.00MB  5.5612GB/s      Pinned       Array  Tesla K20Xm (3)         4        40  [CUDA memcpy HtoA]

Regs: Number of registers used per CUDA thread. This number includes registers used internally by the CUDA driver and/or tools and can be more than what the compiler shows.
SSMem: Static shared memory allocated per CUDA block.
DSMem: Dynamic shared memory allocated per CUDA block.
SrcMemType: The type of source memory accessed by memory operation/copy
DstMemType: The type of destination memory accessed by memory operation/copy
$

我的系统是一个4-GPU系统，两个根端口分别挂有两个GPU。因此，从探查器的角度来看，PCIE Gen3上主机->设备固定的传输带宽大约为10GB / s，这是在每个端口的2个GPU之间分配的，但是仔细研究探查器的开始和传输的持续时间会发现所有这4个都重叠了从探查器的角度来看。

多GPU代码中的异步纹理对象分配

1 个答案: