Support exp op in XNNPACK backend (pytorch#11373)

### Summary
Support exp in XNNPACK backend

### Test plan
Wrote test cases to see if appropriate xnnpack exp was called

Author: leafs1

backends/xnnpack/operators/__init__.py

@@ -19,6 +19,7 @@
     op_dynamic_dequantize_ops,
     op_dynamic_quantize_ops,
     op_elu,
+    op_exp,
     op_floor,
     op_gelu,
     op_hardswish,

backends/xnnpack/operators/op_exp.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 (
+    XNNExp,
+    XNNGraph,
+    XNode,
+)
+from executorch.backends.xnnpack.utils.utils import get_input_node
+
+
+@register_node_visitor
+class ExpVisitor(NodeVisitor):
+    target = "aten.exp.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=XNNExp(
+                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

@@ -25,10 +25,11 @@
     ConstantPadConfig,
     DeQuantizedPerTensorConfig,
     DivConfig,
+    # EluConfig,
+    ExpConfig,
     FloorConfig,
     GeluConfig,
     HardswishConfig,
-    # EluConfig,
     HardtanhConfig,
     LeakyReLUConfig,
     LogConfig,
@@ -79,6 +80,7 @@
     ClampConfig,
     DivConfig,
     # EluConfig, # Waiting for PyTorch Pin Update
+    ExpConfig,
     FloorConfig,
     GeluConfig,
     HardtanhConfig,

backends/xnnpack/partition/config/generic_node_configs.py

@@ -336,6 +336,13 @@ def get_original_aten(self) -> Optional[torch._ops.OpOverload]:
         return torch.ops.aten.upsample_bilinear2d.vec
 
 
+class ExpConfig(GenericNodePartitionerConfig):
+    target_name = "exp.default"
+
+    def supported_precision_types(self) -> List[ConfigPrecisionType]:
+        return [ConfigPrecisionType.FP32]
+
+
 class FloorConfig(GenericNodePartitionerConfig):
     target_name = "floor.default"
 

backends/xnnpack/partition/configs.py

@@ -59,6 +59,7 @@
     exir_ops.edge.aten.sigmoid.default,
     exir_ops.edge.aten._softmax.default,
     exir_ops.edge.aten.cat.default,
+    exir_ops.edge.aten.exp.default,
     exir_ops.edge.aten.elu.default,
     exir_ops.edge.aten.avg_pool2d.default,
     exir_ops.edge.aten.leaky_relu.default,

backends/xnnpack/runtime/XNNCompiler.cpp

@@ -1723,6 +1723,36 @@ Error defineELUNode(
   return Error::Ok;
 }
 
+/*
+Define serialized exp node into the subgraph, using the remapped ids
+to map the serialized ids, to the new ids generated when defining the
+tensor value
+*/
+Error defineExpNode(
+    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_XNNExp();
+
+  xnn_status status = xnn_define_exp(
+      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 exp node %i with code: %s",
+      node->debug_handle(),
+      xnn_status_to_string(status));
+
+  return Error::Ok;
+}
+
 /*
 Defines absolute value node into subgraph using the remapped ids to map the
 serialized ids to the new ids generated when defining the tensor value
@@ -2082,6 +2112,7 @@ DefineNodeFunc getDefineNodeFunc(fb_xnnpack::XNodeUnion nodeType) {
     _DEFINE(Negate)
     _DEFINE(Square)
     _DEFINE(ELU)
+    _DEFINE(Exp)
     _DEFINE(Abs)
     _DEFINE(PReLU)
     _DEFINE(Concatenate2)

backends/xnnpack/serialization/runtime_schema.fbs

@@ -132,6 +132,7 @@ union XNodeUnion {
   XNNNegate: _XNNNode1x1,
   XNNSquare: _XNNNode1x1,
   XNNELU,
+  XNNExp: _XNNNode1x1,
   XNNAbs: _XNNNode1x1,
   XNNPReLU: _XNNNode2x1,
   XNNConcatenate2: _XNNCat,

backends/xnnpack/serialization/schema.fbs

@@ -128,6 +128,7 @@ union XNodeUnion {
   XNNNegate: _XNNNode1x1,
   XNNSquare: _XNNNode1x1,
   XNNELU,
+  XNNExp: _XNNNode1x1,
   XNNAbs: _XNNNode1x1,
   XNNPReLU: _XNNNode2x1,
   XNNConcatenate2: _XNNCat,

backends/xnnpack/serialization/xnnpack_graph_schema.py

@@ -291,6 +291,11 @@ class XNNCeiling(XNNNode1x1):
     pass
 
 
+@dataclass
+class XNNExp(XNNNode1x1):
+    pass
+
+
 @dataclass
 class XNNGelu(XNNNode1x1):
     pass

backends/xnnpack/test/ops/test_exp.py

@@ -0,0 +1,43 @@
+# 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 TestExp(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
+    class Exp(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+
+        def forward(self, x):
+            return torch.exp(x)
+
+    def run_exp_test(self, inputs):
+        (
+            Tester(self.Exp(), inputs)
+            .export()
+            .check_count({"torch.ops.aten.exp.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_exp_default"])
+            .to_executorch()
+            .serialize()
+            .run_method_and_compare_outputs()
+        )
+
+    def test_fp16_exp(self):
+        inputs = (torch.randn(20).to(torch.float16),)
+        self.run_exp_test(inputs)
+
+    def test_fp32_exp(self):
+        inputs = (torch.randn(20),)
+        self.run_exp_test(inputs)