search

ctarver / NR_GPU_LDPC

GNU General Public License v3.0
2 stars 0 forks source link

Question about CPU time #1

Open Jokeren opened 2 years ago

Jokeren commented 2 years ago

Hi @ctarver I finally got some time to look into this code before my thesis defense. Sorry for the delay.

I took a quick look and collected initial profile data.

If I understand it correct, the code below from Line 136 to Line 139 performs major computation.

https://github.com/ctarver/NR_GPU_LDPC/blob/238acf55eb2c3c7c509b701b6d10d7dafec5a93b/src/main.cpp#L136

I'd like to know if code in generate_new_codewords needs any optimization? In other words, should it be offloaded to GPU? I asked this question because the CPU code in generate_new_codewords took 5x time compared to GPU code in other routines.

1655930102 commented 5 months ago

Hello blogger, I have a question to consult with you. In line 193 of the setup_exp.cu file, there is a line of code: “copy_to_cuda_staged_memory(h->llr_quan, h->llr_cuda, h->mem_size_llr_cuda, i * l->N_before_puncture, p->n_streams); ” When it runs to this point, it shows a memory access conflict. After some investigation, I found that the size of “h->llr_quan” is 104448, while the size of “h->mem_size_llr_cuda” is 1044480. Is this the cause of the problem? If so, how should I correct it? Thank you!