Qualcomm AI Engine Direct - xr model enablement (mld_f)

Differential Revision: D74590962

Pull Request resolved: #10546

Author: haowhsu-quic

backends/qualcomm/_passes/__init__.py

@@ -20,6 +20,7 @@
 from .expand_broadcast_tensor_shape import ExpandBroadcastTensorShape
 from .fixed_linear_keep_dim import FixedLinearKeepDim
 from .fold_qdq import FoldQDQ
+from .fuse_consecutive_cast import FuseConsecutiveCast
 from .fuse_consecutive_transpose import FuseConsecutiveTranspose
 from .i64_to_i32 import I64toI32
 from .insert_io_qdq import InsertIOQDQ
@@ -54,6 +55,7 @@
     ExpandBroadcastTensorShape,
     FixedLinearKeepDim,
     FoldQDQ,
+    FuseConsecutiveCast,
     FuseConsecutiveTranspose,
     I64toI32,
     InsertIOQDQ,

backends/qualcomm/_passes/fuse_consecutive_cast.py

@@ -0,0 +1,116 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
+from executorch.exir.passes import dead_code_elimination_pass
+
+
+class FuseConsecutiveCast(ExportPass):
+    """
+    This pass fuses consecutive cast into one or none to reduce runtime
+    overhead.
+    To simplify the fuse logic, we ensure each cast node's output has at most 1 cast node
+    by cloning cast.
+    Example:
+    Before clone cast:
+    relu -> cast1 ─> cast2
+            |──────> cast3
+
+    After clone cast:
+    relu ─> cast1 ──────> cast2
+      |───> cast4(new) ─> cast3
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.op_map = {
+            exir_ops.edge.dim_order_ops._to_dim_order_copy.default,
+            exir_ops.edge.aten._to_copy.default,
+        }
+        self.visited = set()
+        self.nodes = []
+
+    def _canonicalize_cast(
+        self, graph_module: torch.fx.GraphModule
+    ) -> torch.fx.GraphModule:
+        # replace all i64 cast nodes with i32 version
+        graph = graph_module.graph
+        for n in graph_module.graph.nodes:
+            if n.target in self.op_map and n.meta["val"].dtype == torch.int64:
+                users = list(n.users)
+                for user in users:
+                    # bypass graph output node to meet original convention
+                    if user.op == "output":
+                        continue
+
+                    with graph.inserting_after(n):
+                        cast_node = graph.create_node(
+                            "call_function",
+                            exir_ops.edge.aten._to_copy.default,
+                            n.args,
+                            kwargs={"dtype": torch.int32},
+                        )
+                        cast_node.meta = n.meta
+                        cast_node.meta["val"] = cast_node.meta["val"].to(torch.int32)
+                        user.replace_input_with(n, cast_node)
+
+        graph.eliminate_dead_code()
+
+        # clone nodes for future fusion
+        for n in graph_module.graph.nodes:
+            # make sure we're handling cast node instead of convert node
+            if n.target in self.op_map and n.kwargs.get("dtype", None) is not None:
+                users = [user for user in list(n.users) if user.target in self.op_map]
+                if len(users) > 1:
+                    for i in range(1, len(users)):
+                        with graph.inserting_after(n):
+                            clone_cast_node = graph.create_node(
+                                "call_function",
+                                exir_ops.edge.aten._to_copy.default,
+                                n.args,
+                                kwargs=n.kwargs,
+                            )
+                            clone_cast_node.meta = n.meta
+                            users[i].replace_input_with(n, clone_cast_node)
+
+    def _traverse(self, node):
+        if node in self.visited or node.target not in self.op_map:
+            return
+
+        self.nodes.append(node)
+        self.visited.add(node)
+        next_users = [n for n in list(node.users) if n.target in self.op_map]
+
+        assert (
+            len(next_users) <= 1
+        ), "Each cast node should have at most 1 cast output node after _clone_cast"
+        if not next_users:
+            return
+        else:
+            self._traverse(list(node.users)[0])
+
+    def _fuse(self, graph_module: torch.fx.GraphModule) -> torch.fx.GraphModule:
+        for n in graph_module.graph.nodes:
+            self._traverse(n)
+            # TODO: how to handle following scenario (won't happen for quantized graph)
+            #       fp -> to(i32) -> to(fp)
+            if len(self.nodes) > 1:
+                input_node, output_node = self.nodes[0], self.nodes[-1]
+                output_node.replace_input_with(output_node.args[0], input_node.args[0])
+
+            # clear current stack
+            self.nodes = []
+
+    def call(self, graph_module: torch.fx.GraphModule):
+        self._canonicalize_cast(graph_module)
+        self._fuse(graph_module)
+        graph_module.recompile()
+        dead_code_elimination_pass(graph_module)
+        return PassResult(graph_module, True)

backends/qualcomm/_passes/i64_to_i32.py

@@ -31,6 +31,14 @@ class I64toI32(ExportPass):
         exir_ops.edge.aten.full.default,
         exir_ops.edge.aten.scalar_tensor.default,
     }
+    # This dict is to ensure that the input of the OPs are int64 due to Pytorch restrictions.
+    # For example, scatter op can only accept args[2], the index, as int64.
+    # Key: Ops to cast input to i64
+    # Value: The args' indices to add casting op
+    I64_IN_OPS = {
+        exir_ops.edge.aten.gather.default: [2],
+        exir_ops.edge.aten.scatter.src: [2],
+    }
     copy_op = exir_ops.edge.aten._to_copy.default
 
     def __init__(
@@ -141,11 +149,32 @@ def _cast_constant_to_int32(self, graph_module: torch.fx.GraphModule):
                         n.replace_all_uses_with(to_dst_node)
                         to_dst_node.args = (n,)
 
+    def _cast_op_args_to_i64(self, graph_module: torch.fx.GraphModule):
+        # input will be cast to i32 during call_operator dtype propogation
+        # insert i64 cast node to prevent PyTorch's operator validation failure
+        for node in graph_module.graph.nodes:
+            if node.target in self.I64_IN_OPS:
+                with graph_module.graph.inserting_before(node):
+                    arg_indices = self.I64_IN_OPS[node.target]
+                    for arg_index in arg_indices:
+                        input_node = node.args[arg_index]
+                        cast_i64_node = graph_module.graph.create_node(
+                            "call_function",
+                            self.copy_op,
+                            (input_node,),
+                            {"dtype": torch.int64},
+                        )
+                        cast_i64_node.meta["val"] = node.meta["val"].to(torch.int64)
+                        args_list = list(node.args)
+                        args_list[arg_index] = cast_i64_node
+                        node.args = tuple(args_list)
+
     def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
         # Record original output dtype to ensure that if user expects int64 as output,
         # convert the output back to int64 if it is casted from int64->int32.
         self._record_original_output_dtype(graph_module)
         self._cast_constant_to_int32(graph_module)
+        self._cast_op_args_to_i64(graph_module)
         graph_module = super().call(graph_module).graph_module
         self._preserve_output_dtype(graph_module)
         graph_module.recompile()

backends/qualcomm/_passes/qnn_pass_manager.py

@@ -25,6 +25,7 @@
     ExpandBroadcastTensorShape,
     FixedLinearKeepDim,
     FoldQDQ,
+    FuseConsecutiveCast,
     FuseConsecutiveTranspose,
     I64toI32,
     InsertIOQDQ,
@@ -182,6 +183,7 @@ def transform_for_to_edge_pipeline(
 
     # Before quantizer
     def transform_for_annotation_pipeline(self, graph_module: GraphModule):
+        self.add_pass(RemoveRedundancy(quantization_capture=True))
         self.add_pass(ReduceDynamicRange())
         self.add_pass(RecomposePixelUnshuffle(quantization_capture=True))
         self.add_pass(ReplaceArangeArgs())
@@ -214,5 +216,6 @@ def transform_for_preprocess_pipeline(self, exported_program: ExportedProgram):
         self.add_pass(InsertRequantize())
         self.add_pass(InsertIOQDQ(exported_program))
         self.add_pass(LayoutTransform(exported_program, insert_permute=True))
+        self.add_pass(FuseConsecutiveCast())
         self.add_pass(FuseConsecutiveTranspose())
         return self._transform(exported_program.graph_module)

backends/qualcomm/_passes/remove_redundancy.py

@@ -14,9 +14,9 @@ class RemoveRedundancy(ExportPass):
     Trim certain operators to reduce unnecessary overhead.
     """
 
-    def __init__(self):
+    def __init__(self, quantization_capture=False):
         super(RemoveRedundancy, self).__init__()
-        self.redundant_ops = {
+        self.redundant_ops_general = {
             torch.clone: self._default_condition,
             torch.ops.aten.clone.default: self._default_condition,
             exir_ops.edge.aten.clone.default: self._default_condition,
@@ -28,7 +28,16 @@ def __init__(self):
             exir_ops.edge.dim_order_ops._to_dim_order_copy.default: self._dim_order_op_condition,
             # remove channel_last / contiguous _to_copy if '_skip_dim_order' is set to True
             exir_ops.edge.aten._to_copy.default: self._to_copy_op_condition,
+            torch.ops.aten._assert_tensor_metadata.default: self._default_condition,
         }
+        self.redundant_ops_annotation = {
+            torch.ops.aten._assert_tensor_metadata.default: self._default_condition,
+        }
+        self.redundant_ops = (
+            self.redundant_ops_annotation
+            if quantization_capture
+            else self.redundant_ops_general
+        )
 
     def _dim_order_op_condition(self, node):
         dim_order = node.kwargs.get("dim_order")
@@ -50,6 +59,10 @@ def _remove(self, graph_module: torch.fx.GraphModule) -> torch.fx.GraphModule:
                 continue
 
             to_be_remove = n
+            # assert_tensor_metadata op has no user
+            if len(n.users.keys()) == 0:
+                n.args = ()
+            # normal case
             for user_n in list(n.users.keys()):
                 user_n.replace_input_with(n, n.args[0])
             graph_module.graph.erase_node(to_be_remove)

backends/qualcomm/builders/__init__.py

@@ -32,6 +32,7 @@
     op_expand,
     op_full,
     op_full_like,
+    op_gather,
     op_ge,
     op_gelu,
     op_group_norm,
@@ -120,6 +121,7 @@
     op_expand,
     op_full,
     op_full_like,
+    op_gather,
     op_ge,
     op_gelu,
     op_group_norm,

backends/qualcomm/builders/op_argmin.py

@@ -10,8 +10,8 @@
 import torch
 from executorch.backends.qualcomm.utils.constants import QCOM_AXIS_ORDER, QCOM_DATA
 
-from .node_visitor import NodeVisitor, QNN_TENSOR_TYPE_MAP, register_node_visitor
-from .qnn_constants import OpArgmin, OpCast, QNN_OP_PACKAGE_NAME_QTI_AISW
+from .node_visitor import NodeVisitor, register_node_visitor
+from .qnn_constants import OpArgmin, QNN_OP_PACKAGE_NAME_QTI_AISW
 
 
 @register_node_visitor
@@ -26,7 +26,6 @@ def define_node(
         node: torch.fx.Node,
         nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
     ) -> PyQnnWrapper.PyQnnOpWrapper:
-        op_wrapper_list = []
         input_node = self.get_node(node.args[0])
         input_tensor = self.get_tensor(input_node, node)
         output_tensor = self.get_tensor(node, node)
@@ -38,26 +37,14 @@ def define_node(
             nodes_to_wrappers,
         )
         argmin_input_tensors = [argmin_inp_tensor_wrapper]
-
-        # arg output is index, do not quantize it.
-        node.meta.pop("quant_attrs", None)
-        input_quant_encoding, input_quant_configs = self.get_quant_encoding_conf(
-            input_node, node
-        )
-
-        argmin_intermediate_tensor_wrapper = self.define_custom_tensor_wrapper(
-            node_name=node.name + "_cast",
-            tensor_type=PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
-            dtype=QNN_TENSOR_TYPE_MAP[torch.int32],
-            quant_encoding=input_quant_encoding,
-            quant_configs=input_quant_configs,
-            dims=output_tensor.size(),
-            tensor=output_tensor,
-            is_fake_tensor=True,
-            nodes_to_wrappers=nodes_to_wrappers,
+        argmin_out_tensor_wrapper = self.define_tensor(
+            node,
+            node,
+            output_tensor.to(torch.int32),
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
         )
-
-        argmin_output_tensors = [argmin_intermediate_tensor_wrapper]
+        argmin_output_tensors = [argmin_out_tensor_wrapper]
 
         dim = cast(int, node.args[1])
         if dim < 0:
@@ -87,24 +74,4 @@ def define_node(
                 {QCOM_DATA: keep_dims},
             )
 
-        op_wrapper_list.append(argmin_op)
-
-        cast_op = PyQnnWrapper.PyQnnOpWrapper(
-            node.name + "_cast",
-            QNN_OP_PACKAGE_NAME_QTI_AISW,
-            OpCast.op_name,
-        )
-
-        output_tensor_wrapper = self.define_tensor(
-            node,
-            node,
-            output_tensor,
-            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
-            nodes_to_wrappers,
-        )
-
-        cast_op.AddInputTensors([argmin_intermediate_tensor_wrapper])
-        cast_op.AddOutputTensors([output_tensor_wrapper])
-        op_wrapper_list.append(cast_op)
-
-        return op_wrapper_list
+        return argmin_op