deepseek-ai / deepseek-ai/DeepGEMM
question about tmem barrier
- Dominant language
- Cuda
- Stars
- 7.8k
- Forks
- 1.3k
- Avg merge
- 3d 7h
- Merged PRs (30d)
- 3
Description
I have a question about the comment here:
https://github.com/deepseek-ai/DeepGEMM/blob/54e22612409371d6364144b69086735beb54e98b/deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh#L269
Since TMEM is consumed by both CTAs, the next line uses 2 * kNumEpilogueThreads. Shouldn't this barrier therefore use arrive at all CTAs instead?
Conversely, for tmem_full_barrier above, it seems that only the leader CTA calls arrive. Is my understanding correct?
Contributor guide
No contributing guide indexed for this repository
Research direction
Start in deep_gemm/include/deep_gemm/impls/sm100_fp8_fp4_mega_moe.cuh around the cited line and compare the barrier calls there with tmem_full_barrier above. Read the surrounding TMEM and CTA synchronization flow to determine which CTAs should call arrive. Done means resolving whether the current barrier participation is correct and documenting or correcting the synchronization behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- hpc
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100