Fully remove barrier FIFO designs (#200)

* Fully remove FIFO slots

* Fully remove FIFO buffers

* Minor fix styles

* Fix some typos

* Bugs fixed

* Cleanup `ibgda_poll_cq`
This commit is contained in:
Chenggang Zhao
2025-06-10 16:23:20 +08:00
committed by GitHub
parent a16af40531
commit 8da2d7b38d
10 changed files with 121 additions and 181 deletions

View File

@@ -12,13 +12,13 @@ namespace deep_ep {
namespace intranode {
template<int kNumRanks>
__global__ void barrier(int** task_fifo_ptrs, int head, int rank) {
barrier_device<kNumRanks>(task_fifo_ptrs, head, rank);
__global__ void barrier(int** barrier_signal_ptrs, int rank) {
barrier_block<kNumRanks>(barrier_signal_ptrs, rank);
}
void barrier(int** task_fifo_ptrs, int head, int rank, int num_ranks, cudaStream_t stream) {
void barrier(int** barrier_signal_ptrs, int rank, int num_ranks, cudaStream_t stream) {
#define BARRIER_LAUNCH_CASE(ranks) \
LAUNCH_KERNEL(&cfg, barrier<ranks>, task_fifo_ptrs, head, rank); \
LAUNCH_KERNEL(&cfg, barrier<ranks>, barrier_signal_ptrs, rank); \
break
SETUP_LAUNCH_CONFIG(1, 32, stream);