From 9659efbdd5172503848aacf3c6d61817f75ce124 Mon Sep 17 00:00:00 2001 From: Binyang Li Date: Sat, 8 Feb 2025 08:32:34 +0000 Subject: [PATCH] remove unnecessary sync --- python/mscclpp/language/ir.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/mscclpp/language/ir.py b/python/mscclpp/language/ir.py index 3a1408acb..203dfa75e 100644 --- a/python/mscclpp/language/ir.py +++ b/python/mscclpp/language/ir.py @@ -131,8 +131,8 @@ def ir_to_json(program: Program): # Expand extra dependencies into nop operations nop = Op(Instruction.nop, -1, None, None, []) for i, dep in enumerate(op.depends): - # barrier already syncs all threads - if dep.inst != Instruction.barrier: + # barrier already syncs all threads, only sync within the same threadblock + if dep.inst != Instruction.barrier and dep.tb == op.tb: nop.depends.append(dep) if len(new_ops) > 0 and ( new_ops[-1].inst == Instruction.barrier or new_ops[-1].inst == Instruction.nop