Arm backend: Add TOSA support for gt.Scalar and lt.Scalar (#9908)

- Convert gt.Scalar and lt.Scalar to gt.Tensor and lt.Tensor
- Expand the scalar operands to match the shape of the tensor operands
- Rename the eq test names to include full aten op name

Signed-off-by: Yufeng Shi <[email protected]>

Author: YufengShi-dudu

backends/arm/_passes/match_arg_ranks_pass.py

@@ -48,6 +48,8 @@ def __init__(self, exported_program):
         exir_ops.edge.aten.bitwise_right_shift.Tensor,
         exir_ops.edge.aten.bitwise_left_shift.Tensor,
         exir_ops.edge.aten.eq.Tensor,
+        exir_ops.edge.aten.gt.Tensor,
+        exir_ops.edge.aten.lt.Tensor,
         exir_ops.edge.aten.pow.Tensor_Tensor,
         exir_ops.edge.aten.where.self,
     ]

backends/arm/_passes/replace_scalar_with_tensor_pass.py

@@ -26,13 +26,17 @@
     exir_ops.edge.aten.__rshift__.Scalar: exir_ops.edge.aten.bitwise_right_shift.Tensor,
     exir_ops.edge.aten.__lshift__.Scalar: exir_ops.edge.aten.bitwise_left_shift.Tensor,
     exir_ops.edge.aten.eq.Scalar: exir_ops.edge.aten.eq.Tensor,
+    exir_ops.edge.aten.gt.Scalar: exir_ops.edge.aten.gt.Tensor,
+    exir_ops.edge.aten.lt.Scalar: exir_ops.edge.aten.lt.Tensor,
     torch.ops.aten.add.Scalar: torch.ops.aten.add.Tensor,
     torch.ops.aten.sub.Scalar: torch.ops.aten.sub.Tensor,
     torch.ops.aten.mul.Scalar: torch.ops.aten.mul.Tensor,
     torch.ops.aten.div.Scalar: torch.ops.aten.div.Tensor,
     torch.ops.aten.__rshift__.Scalar: torch.ops.aten.bitwise_right_shift.Tensor,
     torch.ops.aten.__lshift__.Scalar: torch.ops.aten.bitwise_left_shift.Tensor,
     torch.ops.aten.eq.Scalar: torch.ops.aten.eq.Tensor,
+    torch.ops.aten.gt.Scalar: torch.ops.aten.gt.Tensor,
+    torch.ops.aten.lt.Scalar: torch.ops.aten.lt.Tensor,
 }
 
 

backends/arm/operator_support/ethos_u55_support.py

@@ -135,8 +135,10 @@ class EthosU55NotSupported(OperatorSupportBase):
         exir_ops.edge.aten.eq.Scalar,
         exir_ops.edge.aten.ge.Tensor,
         exir_ops.edge.aten.gt.Tensor,
+        exir_ops.edge.aten.gt.Scalar,
         exir_ops.edge.aten.le.Tensor,
         exir_ops.edge.aten.lt.Tensor,
+        exir_ops.edge.aten.lt.Scalar,
         exir_ops.edge.aten.flip.default,  # REVERSE
         exir_ops.edge.aten.grid_sampler_2d,  # GATHER
         exir_ops.edge.aten.scatter.src,

backends/arm/operator_support/tosa_supported_operators.py

@@ -176,8 +176,10 @@ def is_node_supported(
             exir_ops.edge.aten.full_like.default,
             exir_ops.edge.aten.ge.Tensor,
             exir_ops.edge.aten.gt.Tensor,
+            exir_ops.edge.aten.gt.Scalar,
             exir_ops.edge.aten.le.Tensor,
             exir_ops.edge.aten.lt.Tensor,
+            exir_ops.edge.aten.lt.Scalar,
             exir_ops.edge.aten.mul.Tensor,
             exir_ops.edge.aten.add.Scalar,
             exir_ops.edge.aten.sub.Scalar,

backends/arm/test/ops/test_eq.py

@@ -96,8 +96,16 @@ def test_eq_scalar_tosa_MI(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_tensor | test_data_scalar)
-def test_eq_tosa_BI(test_module):
+@common.parametrize("test_module", test_data_tensor)
+def test_eq_tensor_tosa_BI(test_module):
+    pipeline = TosaPipelineBI[input_t](
+        test_module, test_module.get_inputs(), Equal.aten_op_Tensor, Equal.exir_op
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_module", test_data_scalar)
+def test_eq_scalar_tosa_BI(test_module):
     pipeline = TosaPipelineBI[input_t](
         test_module, test_module.get_inputs(), Equal.aten_op_Tensor, Equal.exir_op
     )
@@ -133,15 +141,34 @@ def test_eq_scalar_u55_BI(test_module):
 
 @common.parametrize(
     "test_module",
-    test_data_tensor | test_data_scalar,
+    test_data_tensor,
     xfails={
         "eq_tensor_rank4_randn": "MLETORCH-847: Boolean eq result unstable on U85",
+    },
+    strict=False,
+)
+@common.XfailIfNoCorstone320
+def test_eq_tensor_u85_BI(test_module):
+    pipeline = EthosU85PipelineBI[input_t](
+        test_module,
+        test_module.get_inputs(),
+        Equal.aten_op_Tensor,
+        Equal.exir_op,
+        run_on_fvp=True,
+    )
+    pipeline.run()
+
+
+@common.parametrize(
+    "test_module",
+    test_data_scalar,
+    xfails={
         "eq_scalar_rank4_randn": "MLETORCH-847: Boolean eq result unstable on U85",
     },
     strict=False,
 )
 @common.XfailIfNoCorstone320
-def test_eq_u85_BI(test_module):
+def test_eq_scalar_u85_BI(test_module):
     pipeline = EthosU85PipelineBI[input_t](
         test_module,
         test_module.get_inputs(),

backends/arm/test/ops/test_gt.py

@@ -5,7 +5,6 @@
 
 from typing import Tuple
 
-import pytest
 import torch
 from executorch.backends.arm.test import common
 
@@ -16,13 +15,15 @@
     TosaPipelineMI,
 )
 
-aten_op = "torch.ops.aten.gt.Tensor"
-exir_op = "executorch_exir_dialects_edge__ops_aten_gt_Tensor"
 
 input_t = Tuple[torch.Tensor]
 
 
 class Greater(torch.nn.Module):
+    aten_op_tensor = "torch.ops.aten.gt.Tensor"
+    aten_op_scalar = "torch.ops.aten.gt.Scalar"
+    exir_op = "executorch_exir_dialects_edge__ops_aten_gt_Tensor"
+
     def __init__(self, input, other):
         super().__init__()
         self.input_ = input
@@ -31,106 +32,143 @@ def __init__(self, input, other):
     def forward(
         self,
         input_: torch.Tensor,
-        other_: torch.Tensor,
+        other_: torch.Tensor | int | float,
     ):
         return input_ > other_
 
     def get_inputs(self):
         return (self.input_, self.other_)
 
 
-op_gt_rank1_ones = Greater(
+op_gt_tensor_rank1_ones = Greater(
     torch.ones(5),
     torch.ones(5),
 )
-op_gt_rank2_rand = Greater(
+op_gt_tensor_rank2_rand = Greater(
     torch.rand(4, 5),
     torch.rand(1, 5),
 )
-op_gt_rank3_randn = Greater(
+op_gt_tensor_rank3_randn = Greater(
     torch.randn(10, 5, 2),
     torch.randn(10, 5, 2),
 )
-op_gt_rank4_randn = Greater(
+op_gt_tensor_rank4_randn = Greater(
     torch.randn(3, 2, 2, 2),
     torch.randn(3, 2, 2, 2),
 )
 
-test_data_common = {
-    "gt_rank1_ones": op_gt_rank1_ones,
-    "gt_rank2_rand": op_gt_rank2_rand,
-    "gt_rank3_randn": op_gt_rank3_randn,
-    "gt_rank4_randn": op_gt_rank4_randn,
+op_gt_scalar_rank1_ones = Greater(torch.ones(5), 1.0)
+op_gt_scalar_rank2_rand = Greater(torch.rand(4, 5), 0.2)
+op_gt_scalar_rank3_randn = Greater(torch.randn(10, 5, 2), -0.1)
+op_gt_scalar_rank4_randn = Greater(torch.randn(3, 2, 2, 2), 0.3)
+
+test_data_tensor = {
+    "gt_tensor_rank1_ones": op_gt_tensor_rank1_ones,
+    "gt_tensor_rank2_rand": op_gt_tensor_rank2_rand,
+    "gt_tensor_rank3_randn": op_gt_tensor_rank3_randn,
+    "gt_tensor_rank4_randn": op_gt_tensor_rank4_randn,
+}
+
+test_data_scalar = {
+    "gt_scalar_rank1_ones": op_gt_scalar_rank1_ones,
+    "gt_scalar_rank2_rand": op_gt_scalar_rank2_rand,
+    "gt_scalar_rank3_randn": op_gt_scalar_rank3_randn,
+    "gt_scalar_rank4_randn": op_gt_scalar_rank4_randn,
 }
 
 
-@common.parametrize("test_module", test_data_common)
-def test_gt_tosa_MI(test_module):
+@common.parametrize("test_module", test_data_tensor)
+def test_gt_tensor_tosa_MI(test_module):
+    pipeline = TosaPipelineMI[input_t](
+        test_module, test_module.get_inputs(), Greater.aten_op_tensor, Greater.exir_op
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_module", test_data_scalar)
+def test_gt_scalar_tosa_MI(test_module):
     pipeline = TosaPipelineMI[input_t](
-        test_module, test_module.get_inputs(), aten_op, exir_op
+        test_module, test_module.get_inputs(), Greater.aten_op_scalar, Greater.exir_op
+    )
+    pipeline.run()
+
+
+@common.parametrize("test_module", test_data_tensor)
+def test_gt_tensor_tosa_BI(test_module):
+    pipeline = TosaPipelineBI[input_t](
+        test_module, test_module.get_inputs(), Greater.aten_op_tensor, Greater.exir_op
     )
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_common)
-def test_gt_tosa_BI(test_module):
+@common.parametrize("test_module", test_data_scalar)
+def test_gt_scalar_tosa_BI(test_module):
     pipeline = TosaPipelineBI[input_t](
-        test_module, test_module.get_inputs(), aten_op, exir_op
+        test_module, test_module.get_inputs(), Greater.aten_op_tensor, Greater.exir_op
     )
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_common)
-def test_gt_u55_BI(test_module):
-    # GREATER is not supported on U55.
+@common.parametrize("test_module", test_data_tensor)
+@common.XfailIfNoCorstone300
+def test_gt_tensor_u55_BI(test_module):
+    # Greater is not supported on U55.
     pipeline = OpNotSupportedPipeline[input_t](
         test_module,
         test_module.get_inputs(),
         "TOSA-0.80+BI+u55",
-        {exir_op: 1},
+        {Greater.exir_op: 1},
     )
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_common)
-def test_gt_u85_BI(test_module):
-    pipeline = EthosU85PipelineBI[input_t](
+@common.parametrize("test_module", test_data_scalar)
+@common.XfailIfNoCorstone300
+def test_gt_scalar_u55_BI(test_module):
+    # Greater is not supported on U55.
+    pipeline = OpNotSupportedPipeline[input_t](
         test_module,
         test_module.get_inputs(),
-        aten_op,
-        exir_op,
-        run_on_fvp=False,
-        use_to_edge_transform_and_lower=True,
+        "TOSA-0.80+BI+u55",
+        {Greater.exir_op: 1},
+        n_expected_delegates=1,
     )
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_common)
-@pytest.mark.skip(reason="The same as test_gt_u55_BI")
-def test_gt_u55_BI_on_fvp(test_module):
-    # GREATER is not supported on U55.
-    pipeline = OpNotSupportedPipeline[input_t](
+@common.parametrize(
+    "test_module",
+    test_data_tensor,
+    xfails={
+        "gt_tensor_rank4_randn": "MLETORCH-847: Boolean eq result unstable on U85",
+    },
+)
+@common.XfailIfNoCorstone320
+def test_gt_tensor_u85_BI(test_module):
+    pipeline = EthosU85PipelineBI[input_t](
         test_module,
         test_module.get_inputs(),
-        "TOSA-0.80+BI+u55",
-        {exir_op: 1},
+        Greater.aten_op_tensor,
+        Greater.exir_op,
+        run_on_fvp=True,
     )
     pipeline.run()
 
 
 @common.parametrize(
     "test_module",
-    test_data_common,
-    xfails={"gt_rank4_randn": "4D fails because boolean Tensors can't be subtracted"},
+    test_data_scalar,
+    xfails={
+        "gt_scalar_rank4_randn": "MLETORCH-847: Boolean eq result unstable on U85",
+    },
 )
-@common.SkipIfNoCorstone320
-def test_gt_u85_BI_on_fvp(test_module):
+@common.XfailIfNoCorstone320
+def test_gt_scalar_u85_BI(test_module):
     pipeline = EthosU85PipelineBI[input_t](
         test_module,
         test_module.get_inputs(),
-        aten_op,
-        exir_op,
+        Greater.aten_op_tensor,
+        Greater.exir_op,
         run_on_fvp=True,
-        use_to_edge_transform_and_lower=True,
     )
     pipeline.run()