NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-04-19 22:38:56 +00:00
Code Issues 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

2
python/cutlass_cppgen/backend/evt/backend/sm90_emitter.py
View File

@@ -90,7 +90,7 @@ class Sm90Emitter:
tile_description=operation.tile_description,
schedule=operation.tile_description.epilogue_schedule,
element_c=operation.C.element,
element_d=operation.C.element,
element_d=operation.D.element,
fusion_callbacks=fusion_callbacks
)

2
python/cutlass_cppgen/backend/evt/frontend/python_ast.py
View File

@@ -140,7 +140,7 @@ class PythonASTFrontend(EVTFrontendBase, ast.NodeVisitor):
self.add_edge(rhs, name, weight=1)
return name
def visit_Assign(self, node: ast.BinOp):
def visit_Assign(self, node: ast.Assign):
target = self.visit(node.targets[0])
value = self.visit(node.value)
# Create the assign node

4
python/cutlass_cppgen/backend/evt/ir/layout_algorithm.py
View File

@@ -119,8 +119,8 @@ def _get_first_rhs_nonzero_stride(stride_list, idx):
for i in range(idx+1, len(stride_list)):
if stride_list[i] != 0:
return i
else:
return None
else:
return None
def reshape(layout, new_shape):
"""

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

@@ -99,11 +99,10 @@ class EVTGraphDrawer:
stride = node.tensor.stride
label += f"|shape={shape}|stride={stride}"
if hasattr(node, "store_tensor"):
if node.store_tensor is not None:
store_shape = node.store_tensor.shape
store_stride = node.store_tensor.stride
label += f"|store_shape={store_shape}|stride_stride={store_stride}"
if hasattr(node, "store_tensor") and node.store_tensor is not None:
store_shape = node.store_tensor.shape
store_stride = node.store_tensor.stride
label += f"|store_shape={store_shape}|store_stride={store_stride}"
label += "}"
return label
@@ -114,7 +113,7 @@ class EVTGraphDrawer:
name: str
):
import pydot
dot_graph = pydot.Dot(name, randir="TB")
dot_graph = pydot.Dot(name, rankdir="TB")
for node in graph.nodes_meta:
style = self._get_node_style(node)
label = self._get_node_label(node)
@@ -133,11 +132,11 @@ class EVTGraphDrawer:
return dot_graph
def get_dot_graph(self) -> pydot.Dot:
def get_dot_graph(self) -> "pydot.Dot":
return [(key, self.get_dot_graph_by_name(key)) for key in self._dot_graphs.keys()]
def get_dot_graph_by_name(self, name) -> pydot.Dot:
def get_dot_graph_by_name(self, name) -> "pydot.Dot":
return self._dot_graphs[name]
def get_main_dot_graph(self) -> pydot.Dot:
def get_main_dot_graph(self) -> "pydot.Dot":
return self._dot_graphs[self._name]

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: