reverting the (Reduce NPKit latency overhead in MSCCL kernel) PR #893 (#1525)

This commit is contained in:
Pedram Alizadeh
2025-02-14 11:03:43 -05:00
committed by GitHub
parent 824b81c034
commit 0e5f4d0662
3 changed files with 15 additions and 41 deletions
-13
View File
@@ -47,19 +47,6 @@ class NpKit {
}
}
static inline __device__ void CollectGpuEventLDS(uint8_t type, int64_t size, uint32_t rsvd, uint64_t timestamp) {
#if defined(ENABLE_NPKIT)
if (ncclShmem.event_buffer_head < LDS_NUM_EVENTS) {
NpKitEvent& event = ncclShmem.event_buffer[ncclShmem.event_buffer_head];
event.fields.type = type;
event.fields.size = size < 0 ? 0 : size;
event.fields.rsvd = rsvd;
event.fields.timestamp = timestamp;
ncclShmem.event_buffer_head++;
}
#endif
}
static void CollectCpuEvent(uint8_t type, int64_t size, uint32_t rsvd, uint64_t timestamp, int channel_id);
static uint64_t *GetCpuTimestamp();