cuda非法内存访问但有足够的内存

Question

我试图运行下面的代码并遇到内存访问错误。我试图运行cuda-memcheck并返回以下信息：

这对我来说很奇怪，因为相同的代码在另一台具有不同GPU卡的计算机上运行良好。我尝试使用cuda-memcheck并返回以下信息：

========= CUD的CUDA-MEMCHECK CUDA错误：/Users/s161901/Desktop/Liver_Bio_recon_study/IVFD_GPU_only/codes/kernel.cu:2010 code = 4（cudaErrorLaunchFailure）“cudaMemcpy（Iout，d_Iout，sizeof（ float）* voxelNumber，cudaMemcpyDeviceToHost）“=========程序命中cudaErrorLaunchFailure（错误4），因为CUDA API调用cudaThreadSynchronize时出现”未指定的启动失败“。 =========保存的主机回溯到错误的驱动程序入口点=========主机帧：C：\ WINDOWS \ system32 \ nvcuda.dll（cuDevicePrimaryCtxGetState + 0x2d242e）[0x2e006b] = ========主机框架：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ cudart64_70.dll（cudaThreadSynchronize + 0xf5）[0x1bba5] =========主机框架：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0。 exe [0x130ef] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0.exe [0xc17b] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0。 exe [0x1bedc] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0.exe [0x1bd2e] =========主机帧：C：\ WINDOWS \ System32 \ KERNEL32。 DLL（BaseThreadInitThunk + 0x14）[0x13034]

========= Host Frame:C:\WINDOWS\SYSTEM32\ntdll.dll (RtlUserThreadStart + 0x21) [0x71551]

=========由于CUDA API调用cudaMemcpy时出现“未指定的启动失败”，程序命中cudaErrorLaunchFailure（错误4）。 =========保存的主机回溯到错误的驱动程序入口点=========主机帧：C：\ WINDOWS \ system32 \ nvcuda.dll（cuDevicePrimaryCtxGetState + 0x2d242e）[0x2e006b] = ========主机框架：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ cudart64_70.dll（cudaMemcpy + 0x12f）[0x2711f] =========主机框架：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0。 exe [0x13113] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0.exe [0xc17b] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0。 exe [0x1bedc] =========主机帧：F：\ Liver_Bio_recon_study \ IVFD_GPU_only \ IVFD_GPU_0.exe [0x1bd2e] =========主机帧：C：\ WINDOWS \ System32 \ KERNEL32。 DLL（BaseThreadInitThunk + 0x14）[0x13034]

========= Host Frame:C:\WINDOWS\SYSTEM32\ntdll.dll (RtlUserThreadStart + 0x21) [0x71551]

=========错误摘要：2个错误

我的cuda内核是：

__global__ void backwardProj(float *dest, int NPROJ, float *d_prj3d, int nx, int nz, float vx, float vz, float *d_sine, float *d_cosine, float PIXSIZE_X, float PIXSIZE_Z, int NI_X, int NI_Z, float L1, float L2)
{
const int tid = (blockIdx.y*32768 + blockIdx.x)*blockDim.x + threadIdx.x;


if (tid >= nx*nx*nz)
return;


    float result = 0.0F;
    for(int iproj = 0; iproj < NPROJ; iproj++)
    {

        float sinTheta = d_sine[iproj];
        float cosTheta = d_cosine[iproj];
        //      setup rotation angle

        int ix = tid % nx;
        int iy = ((tid-ix)/nx) % nx;
        int iz = tid/(nx*nx);

        float xptemp = (ix - nx/2 + 0.5F) * vx;
        float yptemp = (iy - nx/2 + 0.5F) * vx;

        float xp = xptemp * cosTheta + yptemp * sinTheta;
        float yp = -xptemp * sinTheta + yptemp * cosTheta;
        float zp = (iz - nz/2 + 0.5F) * vz;
        //      coordinate of a point in the phantom in rotated coordinate

        float xs = -L1;
        float ys = 0.0;
        float zs = 0.0;
        //      coordinate of source in rotated coordinate

        float x = -(ys + (yp - ys)*(L2 - xs)/(xp - xs))/PIXSIZE_X + NI_X/2;
        float z = (zs + (zp - zs)*(L2 - xs)/(xp - xs))/PIXSIZE_Z + NI_Z/2;
        int xi = floor(x-0.50);
        int zi = floor(z-0.50);
        //      coordinate on the imager in unit of pixsize

        float factor1 = sqrt((xp-xs)*(xp-xs)+(yp-ys)*(yp-ys)+(zp-zs)*(zp-zs))/abs(xp-xs);
        float factor2 = (L1+L2)*(L1+L2)/(xp - xs)/(xp-xs);

        float v00 = (x>=0 && x<=NI_X && z>=0 && zi<=NI_Z) * 
            d_prj3d[ind3to1(xi+(xi<0), zi+(zi<0), iproj, NI_X, NI_Z, NPROJ) ];
        float v10 = (x>=0 && x<=NI_X && z>=0 && zi<=NI_Z) * 
            d_prj3d[ (xi+1-(xi+1>=NI_X)) +  (zi+(zi<0)) * NI_X +  iproj * NI_X * NI_Z];
        float v01 = (x>=0 && x<=NI_X && z>=0 && zi<=NI_Z) * 
            d_prj3d[ xi+(xi<0) + (zi+1-(zi+1>=NI_Z)) * NI_X + iproj * NI_X * NI_Z ];
        float v11 = (x>=0 && x<=NI_X && z>=0 && zi<=NI_Z) * 
            d_prj3d[ xi+1-(xi+1>=NI_X) + (zi+1-(zi+1>=NI_Z))*NI_X +  iproj * NI_X * NI_Z ];

        //      obtain values at four nearest neighbors

        x -= xi;
        z -= zi;
        float value = v00*(1-x)*(1-z) + v10*x*(1-z) + v01*(1-x)*z + v11*x*z;
        //      biliear interpolation

        result += factor1*factor2*value;

    }

    dest[tid] =(result*vx*vx*vz/PIXSIZE_X/PIXSIZE_Z)/vx;    

}

我的主持人电话是：

N = 256*256*217;
nblocks.x = 32768;
NTHREAD_PER_BLOCK=512;
nblocks.y =  ((1 + (N - 1)/NTHREAD_PER_BLOCK) - 1) / NBLOCKX + 1;       

    backwardProj<<<nblocks, NTHREAD_PER_BLOCK>>>(sk, nview, d_diff3d, nx, nz, VOXSIZE_X, VOXSIZE_Z, d_sine, d_cosin,  PIXSIZE_X,  PIXSIZE_Z, NI_X, NI_Z,  L1,  L2);
    cudaThreadSynchronize();
    cudaError_t error = cudaGetLastError();
    if(error!=cudaSuccess)
     {
    fprintf(stderr,"ERROR: %s\n", cudaGetErrorString(error) );
    exit(-1);
     }

Answer 1

非法内存访问与没有足够内存无关。这意味着您的代码正在尝试访问它不应该访问的内存。这意味着您的代码中存在错误。

不幸的是，你只显示了一些代码，而不是一个自包含的例子，所以对你的代码没有太多可说的。我最好的猜测是，最有可能的是，你的索引计算有问题......