Add GELU support to XNNPACK (#11006)

### Summary
Add GELU support to XNNPACK

### Test plan
Wrote tests and asserted that xnnpack delegate would run for GELU op

Author: leafs1

backends/xnnpack/operators/__init__.py

@@ -20,6 +20,7 @@
     op_dynamic_quantize_ops,
     op_elu,
     op_floor,
+    op_gelu,
     op_hardswish,
     op_hardtanh,
     op_leaky_relu,

backends/xnnpack/operators/op_gelu.py

@@ -0,0 +1,52 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# 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.
+
+from typing import Dict
+
+import torch
+from executorch.backends.xnnpack.operators.node_visitor import (
+    NodeVisitor,
+    register_node_visitor,
+)
+from executorch.backends.xnnpack.serialization.xnnpack_graph_schema import (
+    XNNGelu,
+    XNNGraph,
+    XNode,
+)
+from executorch.backends.xnnpack.utils.utils import get_input_node
+
+
+@register_node_visitor
+class GeluVisitor(NodeVisitor):
+    target = "aten.gelu.default"
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        xnn_graph: XNNGraph,
+        vals_to_ids: Dict[torch.fx.Node, int],
+        debug_handle: int,
+    ) -> None:
+        self.define_nodes_tensor_inputs_outputs(node, xnn_graph, vals_to_ids)
+
+        # input
+        input_id = vals_to_ids[get_input_node(node, 0)]
+
+        # output
+        output_id = vals_to_ids[node]
+
+        ser_node = XNode(
+            xnode_union=XNNGelu(
+                input_id=input_id,
+                output_id=output_id,
+                flags=0,
+            ),
+            debug_handle=debug_handle,
+        )
+        xnn_graph.xnodes.append(ser_node)

backends/xnnpack/partition/config/__init__.py

@@ -26,6 +26,7 @@
     DeQuantizedPerTensorConfig,
     DivConfig,
     FloorConfig,
+    GeluConfig,
     HardswishConfig,
     # EluConfig,
     HardtanhConfig,
@@ -79,6 +80,7 @@
     DivConfig,
     # EluConfig, # Waiting for PyTorch Pin Update
     FloorConfig,
+    GeluConfig,
     HardtanhConfig,
     HardswishConfig,
     LeakyReLUConfig,

backends/xnnpack/partition/config/generic_node_configs.py

@@ -343,6 +343,13 @@ def supported_precision_types(self) -> List[ConfigPrecisionType]:
         return [ConfigPrecisionType.FP32]
 
 
+class GeluConfig(GenericNodePartitionerConfig):
+    target_name = "gelu.default"
+
+    def supported_precision_types(self) -> List[ConfigPrecisionType]:
+        return [ConfigPrecisionType.FP32]
+
+
 class HardswishConfig(GenericNodePartitionerConfig):
     target_name = "hardswish.default"
 

backends/xnnpack/partition/configs.py

@@ -65,6 +65,7 @@
     exir_ops.edge.aten.addmm.default,  # TODO(T163877189) add constraint for addmm
     exir_ops.edge.aten.rsqrt.default,
     exir_ops.edge.aten.log.default,
+    exir_ops.edge.aten.gelu.default,
 ]
 
 SUPPORTED_MODULES = [

backends/xnnpack/runtime/XNNCompiler.cpp

@@ -1448,6 +1448,36 @@ Error defineLogNode(
   return Error::Ok;
 }
 
+/*
+Define serialized gelu node into the subgraph, using the remapped ids
+to map the serialized ids, to the new ids generated when defining the
+tensor value
+*/
+Error defineGeluNode(
+    xnn_subgraph_t subgraph_ptr,
+    const std::unordered_map<uint32_t, uint32_t>& remapped_ids,
+    const NodePtr node,
+    const fb_xnnpack::XNNGraph* graph) noexcept {
+  MAYBE_UNUSED(graph);
+
+  auto graph_node = node->xnode_union_as_XNNGelu();
+
+  xnn_status status = xnn_define_gelu(
+      subgraph_ptr,
+      remapped_ids.at(graph_node->input_id()),
+      remapped_ids.at(graph_node->output_id()),
+      graph_node->flags());
+
+  ET_CHECK_OR_RETURN_ERROR(
+      status == xnn_status_success,
+      Internal,
+      "Failed to create gelu node %i with code: %s",
+      node->debug_handle(),
+      xnn_status_to_string(status));
+
+  return Error::Ok;
+}
+
 /*
 Define serialized ceiling node into the subgraph, using the remapped ids
 to map the serialized ids, to the new ids generated when defining the
@@ -2009,6 +2039,7 @@ DefineNodeFunc getDefineNodeFunc(fb_xnnpack::XNodeUnion nodeType) {
     _DEFINE(SquareRoot)
     _DEFINE(ReciprocalSquareRoot)
     _DEFINE(Ceiling)
+    _DEFINE(Gelu)
     _DEFINE(Hardswish)
     _DEFINE(LeakyReLU)
     _DEFINE(Log)

backends/xnnpack/serialization/runtime_schema.fbs

@@ -140,6 +140,7 @@ union XNodeUnion {
   XNNConvTranspose2d: _XNNNodeConv,
   XNNReciprocalSquareRoot: _XNNNode1x1,
   XNNLog: _XNNNode1x1,
+  XNNGelu: _XNNNode1x1,
 }
 
 union XValueUnion {

backends/xnnpack/serialization/schema.fbs

@@ -136,6 +136,7 @@ union XNodeUnion {
   XNNConvTranspose2d: _XNNNodeConv,
   XNNReciprocalSquareRoot: _XNNNode1x1,
   XNNLog: _XNNNode1x1,
+  XNNGelu: _XNNNode1x1,
 }
 
 union XValueUnion {

backends/xnnpack/serialization/xnnpack_graph_schema.py

@@ -291,6 +291,11 @@ class XNNCeiling(XNNNode1x1):
     pass
 
 
+@dataclass
+class XNNGelu(XNNNode1x1):
+    pass
+
+
 @dataclass
 class XNNHardswish(XNNNode1x1):
     pass
@@ -385,6 +390,7 @@ class XNNScaledDotProductAttention:
     XNNBatchMatrixMultiply,
     XNNReciprocalSquareRoot,
     XNNLog,
+    XNNGelu,
 ]
 
 

backends/xnnpack/test/ops/test_gelu.py

@@ -0,0 +1,44 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# 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 unittest
+
+import torch
+from executorch.backends.xnnpack.test.tester import Tester
+
+
+class TestGelu(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
+    class Gelu(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.gelu = torch.nn.GELU()
+
+        def forward(self, x):
+            return self.gelu(x)
+
+    def run_gelu_test(self, inputs):
+        (
+            Tester(self.Gelu(), inputs)
+            .export()
+            .check_count({"torch.ops.aten.gelu.default": 1})
+            .to_edge_transform_and_lower()
+            .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
+            .check_not(["executorch_exir_dialects_edge__ops_aten_gelu_default"])
+            .to_executorch()
+            .serialize()
+            .run_method_and_compare_outputs()
+        )
+
+    def test_fp16_gelu(self):
+        inputs = (torch.randn(20).to(torch.float16),)
+        self.run_gelu_test(inputs)
+
+    def test_fp32_gelu(self):
+        inputs = (torch.randn(20),)
+        self.run_gelu_test(inputs)