add lds barrier

This commit is contained in:
joye
2025-06-11 12:44:03 +08:00
parent 4c85e10281
commit 09a2e029b9

View File

@@ -511,6 +511,7 @@ __global__ void kernel_grouped_conv_bwd_data_optimized_v2(const ABDataType* __re
#pragma unroll
for(int batch_id = 0; batch_id < BatchPerBlock; batch_id += WaveNum)
{
block_sync_lds();
for(int i = tid; i < InLoopNum; i += BlockSize)
{
int group_id = i % GroupPerBlockInFP4;