When running GPGPU-sim with NVIDIA driver, such an error occurs. about gpgpu-sim_distribution HOT 7 OPEN

Perhaps you are using the latest version of NVCC compiler, and GPGPU-sim has not fully implemented all the APIs of CUDA language; Or even if the NVCC version you use is not the latest, but your CUDA executable uses the API that GPGPU-Sim has not yet implemented.

List your experimental environment as much as possible, especially the .cu source code and compiler version.

from gpgpu-sim_distribution.

See also:

Our implementation of libcudart is not compatible with all versions of cuda because of different interfaces that NVIDIA uses between different versions.

From here.

from gpgpu-sim_distribution.

@nothingface0
Yes.

When generating the memory accesses in a warp, the sstarr_space's action is same to shared_space's.
I will treat it as a shared memory for the time being.

Thank you!

from gpgpu-sim_distribution.

@nothingface0

Hi!

May I ask you a problem about GPGPU-sim:

Among the types of memory space in the src/abstract_hardware_model.h​ file:

enum _memory_space_t { 
    undefined_space = 0, 
    reg_space, 
    local_space, 
    shared_space, 
    sstarr_space,  <========= HERE 
    param_space_unclassified, 
    param_space_kernel, /* global to all threads in a kernel : read-only */ 
    param_space_local,    /* local to a thread : read-writable */ 
    const_space, 
    tex_space, 
    surf_space, 
    global_space, 
    generic_space, 
    instruction_space 
}; 

, there is such a type called "sstarr_space​", but I can't find it in GPGPU-Sim manual and other materials.
Is it the abbreviation of one memory space?

Thanks for your favourable reply.

from gpgpu-sim_distribution.

@masa-laboratory
I'm not an expert, I'm just a user of GPGPU-Sim, plus I have not really delved into the memory part yet.

I can assume that, since this is used in the ptx part of the code (cuda-sim), it's PTX-specific. I found a list of PTX directives here. From the name of the variable, I would assume that it's referring to some array, possible "shared static array"? I have no idea why it would be different to normal shared memory, though.

from gpgpu-sim_distribution.

@masa-laboratory sstarr seems to be related with Tensor Cores, from what I saw here. They mention that it works similarly to shared memory.

from gpgpu-sim_distribution.

@nothingface0 Thanks!. Have you been reading the source code of GPGPU-Sim?

from gpgpu-sim_distribution.