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

v4.3 tag release update. (#2789)

Browse Source
This commit is contained in:
Junkai-Wu
2025-11-21 09:49:44 +08:00
committed by GitHub
parent 406e078b29
commit 8cd5bef43a
225 changed files with 23229 additions and 2813 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
python/cutlass_cppgen/backend/evt/passes/pass_preprocess_red.py
View File

@@ -51,15 +51,14 @@ class PassPreprocessRed(EVTPassBase):
# Step 1: find the compute nodes with op=red
red_compute_nodes = []
for node_meta in self.dag_ir.nodes_meta:
if isinstance(node_meta, ComputeNode):
if type(node_meta.fn) == tuple:
# To keep the frontend simple, the reduction nodes
# are parsed into compute nodes by default
# The simple heuristic to distinguish between compute
# and reduction node is that compute node is a single function,
# while the reduction node is a tuple of functions for
# in-register reduction and atomic global memory reduction
red_compute_nodes.append(node_meta.name)
if isinstance(node_meta, ComputeNode) and type(node_meta.fn) == tuple:
# To keep the frontend simple, the reduction nodes
# are parsed into compute nodes by default
# The simple heuristic to distinguish between compute
# and reduction node is that compute node is a single function,
# while the reduction node is a tuple of functions for
# in-register reduction and atomic global memory reduction
red_compute_nodes.append(node_meta.name)
# Step 2: for each compute, merge it with the succeeding store
for node in red_compute_nodes: