Jetson The call to cuMemHostRegister failed

Jetson The call to cuMemHostRegister(0xace67b80, 33792, 0) failed

Jetson 运行 mpi 报错：

--------------------------------------------------------------------------
The call to cuMemHostRegister(0xace67b80, 33792, 0) failed.
  Host:  tegra-ubuntu
  cuMemHostRegister return value:  801
  Memory Pool:  sm
--------------------------------------------------------------------------
[tegra-ubuntu:03031] 11 more processes have sent help message help-mpi-common-cuda.txt / cuMemHostRegister failed
[tegra-ubuntu:03031] Set MCA parameter "orte_base_help_aggregate" to 0 to see all help / error messages

这是因为 Jetson 不支持 cuMemHostRegister ，修改为 cuMemHostAlloc。

修改 openmpi-4.1.2/opal/mca/common/cuda/common_cuda.c :

• 方法1：将所有的 cuMemHostRegister 替换为 cuMemHostAlloc
• 方法2：使用宏定义 __aarch64__ 区分是电脑还是 jetson，例如：

#ifdef __aarch64__ // cuMemHostRegister cannot be supported on Jetson
	OPAL_CUDA_DLSYM(libcuda_handle, cuMemHostAlloc);
#else
    OPAL_CUDA_DLSYM(libcuda_handle, cuMemHostRegister);
#endif

将 openmpi-4.1.2/opal/mca/common/cuda/common_cuda.c 内所有使用了 cuMemHostRegister 都做类似替换。

修改 "openmpi-4.1.2/opal/mca/common/cuda/help-mpi-common-cuda.txt ，将：

#
[cuMemHostRegister during init failed]
The call to cuMemHostRegister(%p, %d, 0) failed.
  Host:  %s
  cuMemHostRegister return value:  %d
  Registration cache:  %s
#
[cuMemHostRegister failed]
The call to cuMemHostRegister(%p, %d, 0) failed.
  Host:  %s
  cuMemHostRegister return value:  %d
  Registration cache:  %s

修改为：

#
[cuMemHostRegister during init failed]
The call to cuMemHostRegister(%p, %d, 0) failed.
  Host:  %s
  cuMemHostRegister return value:  %d
  Registration cache:  %s
#
[cuMemHostRegister failed]
The call to cuMemHostRegister(%p, %d, 0) failed.
  Host:  %s
  cuMemHostRegister return value:  %d
  Registration cache:  %s
#
[cuMemHostAlloc during init failed]
The call to cuMemHostAlloc(%p, %d, 0) failed.
  Host:  %s
  cuMemHostAlloc return value:  %d
  Registration cache:  %s
#
[cuMemHostAlloc failed]
The call to cuMemHostAlloc(%p, %d, 0) failed.
  Host:  %s
  cuMemHostAlloc return value:  %d
  Registration cache:  %s

参考链接: Cuda-aware OpenMPI fails on Jetson K1