mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-06-26 23:15:49 +00:00
comment more clear about Memory Consistency and Barrier Visibility
Memory Consistency and Barrier Visibility: Both __syncthreads() and cute::cluster_sync() serve as synchronization points, ensuring that all threads reach the barrier before any proceed. This guarantees that all prior memory operations, including barrier initialization, are visible to all threads within the synchronization scope.
This commit is contained in:
@@ -122,7 +122,7 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
|
||||
(kNumTMAMulticast > 1) ? cutlass::arch::fence_barrier_init() : void();
|
||||
}
|
||||
|
||||
// Synchronize all threads to make barrier visible in normal memory model
|
||||
// Synchronize threads to ensure barrier initialization is visible to all participating threads.
|
||||
(kNumTMAMulticast > 1) ? cute::cluster_sync() : __syncthreads();
|
||||
|
||||
// For pipeline unrolling
|
||||
|
||||
Reference in New Issue
Block a user