在 Cuda Fortran 中使用其他文件中的托管模块变量的问题

Question

我在module a1 的文件p2.f95 中有一个托管变量。我将这个托管变量传递给文件p6.f95中的内核。我能够访问文件p6.f95中的变量，但无法从内核访问它，即；我无法从设备打印变量值。谁能解释一下为什么.....？

file1:p2.f95

module a1
    implicit none
    integer(kind=4),managed::lnode
end module a1

文件2：p6.f95

module kernelsubroutine
    contains
    attributes(global) subroutine kernel(lnode)
        implicit none
        integer(kind=4)::lnode
        print*,"on device",lnode
    end subroutine kernel
end module kernelsubroutine

program main
    use a1
    use cudafor
    use kernelsubroutine
    implicit none
    integer(kind=4)::n
    lnode = 1
    print*,"on host",lnode
    call kernel<<<1,1>>>(lnode)
    n = cudaDeviceSynchronize()
end program main

编译命令1：$pgf95 -Mcuda=rdc p2.f95 p6.f95

编译命令2：

$pgf95 -Mcuda=rdc -c p2.f95

$pgf95 -Mcuda=rdc p6.f95 p2.o
p6.f95:

使用两组编译命令，输出保持不变

输出命令：$ cuda-memcheck ./a.out

输出：

========= CUDA-MEMCHECK
 on host            1
========= Invalid __global__ read of size 4
=========     at 0x000010c8 in /home/vsriram/Documents/fortran_programs/p6.f95:11:kernelsubroutine_kernel_
=========     by thread (0,0,0) in block (0,0,0)
=========     Address 0x0060cf40 is out of bounds
=========     Saved host backtrace up to driver entry point at kernel launch time
=========     Host Frame:/usr/lib/x86_64-linux-gnu/libcuda.so [0x25428a]
=========     Host Frame:/opt/nvidia/hpc_sdk/Linux_x86_64/21.7/cuda/11.4/lib64/libcudart.so.11.0 [0x1402c]
=========     Host Frame:/opt/nvidia/hpc_sdk/Linux_x86_64/21.7/cuda/11.4/lib64/libcudart.so.11.0 (cudaLaunchKernel + 0x1d8) [0x67e58]
=========     Host Frame:/opt/nvidia/hpc_sdk/Linux_x86_64/21.7/compilers/lib/libcudafor.so (__pgiLaunchKernel + 0x1a6) [0x11194]
=========     Host Frame:./a.out [0x14f4]
=========     Host Frame:./a.out [0x1393]
=========
========= Program hit cudaErrorLaunchFailure (error 719) due to "unspecified launch failure" on CUDA API call to cudaDeviceSynchronize. 
=========     Saved host backtrace up to driver entry point at error
=========     Host Frame:/usr/lib/x86_64-linux-gnu/libcuda.so [0x355b43]
=========     Host Frame:/opt/nvidia/hpc_sdk/Linux_x86_64/21.7/cuda/11.4/lib64/libcudart.so.11.0 (cudaDeviceSynchronize + 0x127) [0x43217]
=========     Host Frame:/opt/nvidia/hpc_sdk/Linux_x86_64/21.7/compilers/lib/libcudafor_114.so (cudadevicesynchronize_ + 0x11) [0x7aca1]
=========     Host Frame:./a.out [0x14fb]
=========     Host Frame:./a.out [0x1393]
=========
========= ERROR SUMMARY: 2 errors```

Answer 1

CUDA 统一内存在使用“托管”属性时启用，目前仅可用于动态分配的内存。虽然我们希望将来支持静态数据，但您将无法在像“lnode”这样的静态变量上使用“托管”。

最简单（可能也是最好）的解决方案是按值传递 lnode。虽然您也可以使“lnode”可分配，或者使主机 lnode 和“设备”lnode。

% cat sol1.cuf
module a1
    implicit none
    integer(kind=4)::lnode
end module a1

module kernelsubroutine
    contains
    attributes(global) subroutine kernel(lnode)
        implicit none
        integer(kind=4),value::lnode
        print*,"on device",lnode
    end subroutine kernel
end module kernelsubroutine

program main
    use a1
    use cudafor
    use kernelsubroutine
    implicit none
    integer(kind=4)::n
    lnode = 1
    print*,"on host",lnode
    call kernel<<<1,1>>>(lnode)
    n = cudaDeviceSynchronize()
end program main
% nvfortran sol1.cuf; a.out
 on host            1
 on device            1
% cat sol2.cuf
module a1
    implicit none
    integer(kind=4),allocatable,managed::lnode
end module a1

module kernelsubroutine
    contains
    attributes(global) subroutine kernel(lnode)
        implicit none
        integer(kind=4)::lnode
        print*,"on device",lnode
    end subroutine kernel
end module kernelsubroutine

program main
    use a1
    use cudafor
    use kernelsubroutine
    implicit none
    integer(kind=4)::n
    allocate(lnode)
    lnode = 1
    print*,"on host",lnode
    call kernel<<<1,1>>>(lnode)
    n = cudaDeviceSynchronize()
    deallocate(lnode)
end program main
% nvfortran sol2.cuf ; a.out
 on host            1
 on device            1
% cat sol3.cuf
module a1
    implicit none
    integer(kind=4)::lnode_h
    integer(kind=4),device::lnode_d
end module a1

module kernelsubroutine
    contains
    attributes(global) subroutine kernel(lnode)
        implicit none
        integer(kind=4)::lnode
        print*,"on device",lnode
    end subroutine kernel
end module kernelsubroutine

program main
    use a1
    use cudafor
    use kernelsubroutine
    implicit none
    integer(kind=4)::n
    lnode_h = 1
    print*,"on host",lnode_h
    lnode_d=lnode_h
    call kernel<<<1,1>>>(lnode_d)
    n = cudaDeviceSynchronize()
end program main
% nvfortran sol3.cuf ; a.out
 on host            1
 on device            1