youngkingdom/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Replace 0x1f with 0xffffffff in __shfl_sync (#1097)

Browse Source 
This fixes compatibility with H100 and resolves #1094
This commit is contained in:
Vadim Markovtsev
2023-09-19 01:58:19 +02:00
committed by GitHub
parent 6407bcdf0a
commit 8783c41851
3 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/44_multi_gemm_ir_and_codegen/ir_gen/gen_kernel.py
View File

@ -331,7 +331,7 @@ class gen_Kernel:
operator_code += " " + helper.var_idx("FusedAddBiasEpilogue", i ) + helper.var_idx(" epilogue_", i ) + ";\n"
operator_code += " " + "int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);\n"
operator_code += " " + "int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);\n"
operator_code += " " + "int lane_idx = threadIdx.x % 32;\n"
for i in range (self.b2bnum - 1):