Stable Diffusion CUDA Optimizations Part 5 (#14706)

Add a fusion to remove transpose in subgraph like  
```
--> Gemm --> Unsqueeze(axes=[2]) --> Unsqueeze(axes=[3]) --> Add --> Transpose([0,2,3,1]) --> GroupNorm
```
With this fusion, we can remove 22 Transpose nodes in UNet, and reduce
latency by 0.1 second per image in T4.

onnxruntime/python/tools/transformers/fusion_transpose.py

@@ -8,7 +8,7 @@ from typing import Dict, List
 
 from fusion_base import Fusion
 from fusion_utils import FusionUtils
-from onnx import NodeProto, helper
+from onnx import NodeProto, TensorProto, helper
 from onnx_model import OnnxModel
 
 logger = getLogger(__name__)
@@ -80,3 +80,83 @@ class FusionTranspose(Fusion):
                 self.nodes_to_remove.append(transpose_a)
         transpose_b.ClearField("attribute")
         transpose_b.attribute.extend([helper.make_attribute("perm", output_permutation)])
+
+
+class FusionInsertTranspose(Fusion):
+    def __init__(self, model: OnnxModel):
+        super().__init__(model, "", "GroupNorm")
+
+    def create_transpose_node(self, input_name: str, perm: List[int], output_name=None):
+        """Append a Transpose node after an input"""
+        node_name = self.model.create_node_name("Transpose")
+        if output_name is None:
+            output_name = node_name + "_out" + "-" + input_name
+        transpose_node = helper.make_node("Transpose", inputs=[input_name], outputs=[output_name], name=node_name)
+        transpose_node.attribute.extend([helper.make_attribute("perm", perm)])
+        return transpose_node
+
+    def fuse(
+        self,
+        group_norm_node: NodeProto,
+        input_name_to_nodes: Dict[str, List[NodeProto]],
+        output_name_to_node: Dict[str, NodeProto],
+    ):
+        """
+        This optimization will insert an Transpose, and onnxruntime transpose optimizer will remove it together with
+        another Transpose so that we can get effect of reducing one Transpose after onnxruntime optimization.
+        Before:
+            --> Gemm --> Unsqueeze(axes=[2]) --> Unsqueeze(axes=[3]) --> Add --> Transpose([0,2,3,1]) --> GroupNorm
+        After:
+            --> Gemm --> Unsqueeze(axes=[1]) --> Unsqueeze(axes=[2]) -->Transpose([0,3,1,2]) --> Add --> Transpose([0,2,3,1]) --> GroupNorm
+        """
+        gemm_path = self.model.match_parent_path(
+            group_norm_node, ["Transpose", "Add", "Unsqueeze", "Unsqueeze", "Gemm"], [0, 0, None, 0, 0]
+        )
+        if gemm_path is None:
+            return
+        transpose, add, unsqueeze_3, unsqueeze_2, gemm = gemm_path
+        if self.model.find_graph_output(unsqueeze_3.output[0]):
+            return
+
+        permutation = OnnxModel.get_node_attribute(transpose, "perm")
+        assert isinstance(permutation, list)
+        if permutation != [0, 2, 3, 1]:
+            return
+
+        if not (
+            self.model.get_constant_value(unsqueeze_3.input[1]) == 3
+            and self.model.get_constant_value(unsqueeze_2.input[1]) == 2
+            and len(self.model.get_children(gemm, input_name_to_nodes)) == 1
+            and len(self.model.get_children(unsqueeze_3, input_name_to_nodes)) == 1
+            and len(self.model.get_children(unsqueeze_2, input_name_to_nodes)) == 1
+        ):
+            return
+
+        # Here we use hard-coded name so that it could be shared for the whole model.
+        axes_1 = "ort_const_unsqueeze_axes_1"
+        if self.model.get_initializer(axes_1) is None:
+            axes_1_tensor = helper.make_tensor(
+                name=axes_1,
+                data_type=TensorProto.INT64,
+                dims=[1],
+                vals=[1],
+            )
+            self.model.add_initializer(axes_1_tensor, self.this_graph_name)
+
+        axes_2 = "ort_const_unsqueeze_axes_2"
+        if self.model.get_initializer(axes_2) is None:
+            axes_2_tensor = helper.make_tensor(
+                name=axes_2,
+                data_type=TensorProto.INT64,
+                dims=[1],
+                vals=[2],
+            )
+            self.model.add_initializer(axes_2_tensor, self.this_graph_name)
+
+        unsqueeze_3.input[1] = "ort_const_unsqueeze_axes_2"
+        unsqueeze_2.input[1] = "ort_const_unsqueeze_axes_1"
+        transpose_output_name = self.model.create_node_name("Transpose") + "_NCHW"
+        self.model.replace_input_of_all_nodes(unsqueeze_3.output[0], transpose_output_name)
+        new_transpose = self.create_transpose_node(unsqueeze_3.output[0], [0, 3, 1, 2], transpose_output_name)
+        self.model.add_node(new_transpose, self.this_graph_name)
+        self.increase_counter("Insert Transpose")

onnxruntime/python/tools/transformers/onnx_model_unet.py

@@ -12,7 +12,7 @@ from fusion_biassplitgelu import FusionBiasSplitGelu
 from fusion_group_norm import FusionGroupNorm
 from fusion_nhwc_conv import FusionNhwcConv
 from fusion_options import FusionOptions
-from fusion_transpose import FusionTranspose
+from fusion_transpose import FusionInsertTranspose, FusionTranspose
 from onnx import ModelProto
 from onnx_model import OnnxModel
 from onnx_model_bert import BertOnnxModel
@@ -131,6 +131,9 @@ class UnetOnnxModel(BertOnnxModel):
             group_norm_fusion = FusionGroupNorm(self)
             group_norm_fusion.apply()
 
+            insert_transpose_fusion = FusionInsertTranspose(self)
+            insert_transpose_fusion.apply()
+
         if (options is None) or options.enable_bias_splitgelu:
             bias_split_gelu_fusion = FusionBiasSplitGelu(self)
             bias_split_gelu_fusion.apply()