Add interpolate nearest composite support (#41)

* Add interpolate nearest composite support * fix fmt
google-ai-edge · Jun 8, 2024 · fe2afd5 · fe2afd5
1 parent db77cc4
commit fe2afd5
Show file tree

Hide file tree

Showing 4 changed files with 57 additions and 8 deletions.
diff --git a/ai_edge_torch/convert/conversion.py b/ai_edge_torch/convert/conversion.py
@@ -25,7 +25,7 @@
 from ai_edge_torch import model
 from ai_edge_torch.convert import conversion_utils as cutils
 from ai_edge_torch.convert.fx_passes import BuildAtenCompositePass
-from ai_edge_torch.convert.fx_passes import BuildUpsampleBilinear2DCompositePass  # NOQA
+from ai_edge_torch.convert.fx_passes import BuildInterpolateCompositePass  # NOQA
 from ai_edge_torch.convert.fx_passes import CanonicalizePass
 from ai_edge_torch.convert.fx_passes import InjectMlirDebuginfoPass
 from ai_edge_torch.convert.fx_passes import OptimizeLayoutTransposesPass
@@ -41,7 +41,7 @@ def _run_convert_passes(
   return run_passes(
       exported_program,
       [
-          BuildUpsampleBilinear2DCompositePass(),
+          BuildInterpolateCompositePass(),
           CanonicalizePass(),
           OptimizeLayoutTransposesPass(),
           CanonicalizePass(),

diff --git a/ai_edge_torch/convert/fx_passes/__init__.py b/ai_edge_torch/convert/fx_passes/__init__.py
@@ -24,7 +24,7 @@
 from ai_edge_torch.convert.fx_passes._pass_base import FxPassBase
 from ai_edge_torch.convert.fx_passes._pass_base import FxPassResult
 from ai_edge_torch.convert.fx_passes.build_aten_composite_pass import BuildAtenCompositePass  # NOQA
-from ai_edge_torch.convert.fx_passes.build_upsample_bilinear2d_composite_pass import BuildUpsampleBilinear2DCompositePass  # NOQA
+from ai_edge_torch.convert.fx_passes.build_interpolate_composite_pass import BuildInterpolateCompositePass  # NOQA
 from ai_edge_torch.convert.fx_passes.canonicalize_pass import CanonicalizePass
 from ai_edge_torch.convert.fx_passes.inject_mlir_debuginfo_pass import InjectMlirDebuginfoPass  # NOQA
 from ai_edge_torch.convert.fx_passes.optimize_layout_transposes_pass import OptimizeLayoutTransposesPass  # NOQA

diff --git a/...ild_upsample_bilinear2d_composite_pass.py → ...asses/build_interpolate_composite_pass.py b/...ild_upsample_bilinear2d_composite_pass.py → ...asses/build_interpolate_composite_pass.py
@@ -66,13 +66,34 @@ def attr_builder(graph_module, pattern, internal_match):
   return pattern
 
 
-class BuildUpsampleBilinear2DCompositePass(FxPassBase):
+@functools.cache
+def _get_interpolate_nearest2d_pattern():
+  pattern = mark_pattern.Pattern(
+      "tfl.resize_nearest_neighbor",
+      lambda x: torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest"),
+      export_args=(torch.rand(1, 3, 100, 100),),
+  )
+
+  @pattern.register_attr_builder
+  def attr_builder(pattern, graph_module, internal_match):
+    output = internal_match.returning_nodes[0]
+    output_h, output_w = output.meta["val"].shape[-2:]
+    return {
+        "size": (int(output_h), int(output_w)),
+        "is_nchw_op": True,
+    }
+
+  return pattern
+
+
+class BuildInterpolateCompositePass(FxPassBase):
 
   def __init__(self):
     super().__init__()
     self._patterns = [
         _get_upsample_bilinear2d_pattern(),
         _get_upsample_bilinear2d_align_corners_pattern(),
+        _get_interpolate_nearest2d_pattern(),
     ]
 
   def call(self, graph_module: torch.fx.GraphModule):

diff --git a/ai_edge_torch/convert/fx_passes/test/test_build_upsample_bilinear2d_composite_pass.py b/ai_edge_torch/convert/fx_passes/test/test_build_upsample_bilinear2d_composite_pass.py
@@ -19,7 +19,7 @@
 import torch
 import torch_xla
 
-from ai_edge_torch.convert.fx_passes import BuildUpsampleBilinear2DCompositePass  # NOQA
+from ai_edge_torch.convert.fx_passes import BuildInterpolateCompositePass  # NOQA
 from ai_edge_torch.convert.fx_passes import run_passes
 
 
@@ -38,9 +38,7 @@ def forward(self, *args, **kwargs):
     module = func
 
   exported_program = torch.export.export(module, export_args)
-  exported_program = run_passes(
-      exported_program, [BuildUpsampleBilinear2DCompositePass()]
-  )
+  exported_program = run_passes(exported_program, [BuildInterpolateCompositePass()])
 
   return torch_xla.stablehlo.exported_program_to_stablehlo(
       exported_program
@@ -192,6 +190,36 @@ def test_nn_functional_interpolate_bilinear_size_align_corners(self):
         1,
     )
 
+  def test_nn_functional_interpolate_nearest(self):
+    stablehlo = _export_to_stablehlo_with_composite(
+        lambda x: torch.nn.functional.interpolate(x, scale_factor=3.0, mode='nearest'),
+        (torch.rand(1, 3, 10, 10),),
+    )
+    self.assertTrue(
+        stablehlo.count('stablehlo.composite "tfl.resize_nearest_neighbor"'), 1
+    )
+    self.assertTrue(
+        stablehlo.count(
+            'composite_attributes = {is_nchw_op = true, size = dense<30> : tensor<2xi64>}'
+        ),
+        1,
+    )
+
+  def test_nn_functional_interpolate_nearest_size(self):
+    stablehlo = _export_to_stablehlo_with_composite(
+        lambda x: torch.nn.functional.interpolate(x, size=[15, 20], mode='nearest'),
+        (torch.rand(1, 3, 10, 10),),
+    )
+    self.assertTrue(
+        stablehlo.count('stablehlo.composite "tfl.resize_nearest_neighbor"'), 1
+    )
+    self.assertTrue(
+        stablehlo.count(
+            'composite_attributes = {is_nchw_op = true, size = dense<[15, 20]> : tensor<2xi64>}'
+        ),
+        1,
+    )
+
 
 if __name__ == '__main__':
   unittest.main()