Revert "Add pass for replacing dq-q patterns with rescale" (#8480)

Revert "Arm backend: Add pass for replacing dq-q patterns with rescale (#8415)"

This reverts commit 0740a11, which broke some tests

Author: zingo

backends/arm/_passes/arm_pass_manager.py

@@ -52,7 +52,6 @@
 from executorch.backends.arm._passes.fuse_quantized_activation_pass import (  # type: ignore[import-not-found]
     FuseQuantizedActivationPass,
 )
-from executorch.backends.arm._passes.insert_rescales_pass import InsertRescalePass
 from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.backends.arm._passes.keep_dims_false_to_squeeze_pass import (
     KeepDimsFalseToSqueezePass,
@@ -76,7 +75,6 @@
     UnsqueezeScalarPlaceholdersPass,
 )
 from executorch.backends.arm.tosa_specification import TosaSpecification
-
 from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
 from executorch.exir.pass_manager import PassManager
@@ -121,7 +119,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertSqueezesToViewPass())
 
         self.add_pass(AnnotateChannelsLastDimOrder())
-        self.add_pass(InsertRescalePass())
+
         return self._transform(exported_program.graph_module)
 
     def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
@@ -159,7 +157,6 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertSqueezesToViewPass())
 
         self.add_pass(AnnotateChannelsLastDimOrder())
-        self.add_pass(InsertRescalePass())
 
         return self._transform(exported_program.graph_module)
 

backends/arm/_passes/fold_qdq_with_annotated_qparams_pass.py

@@ -131,9 +131,6 @@ def call(self, graph_module: GraphModule) -> PassResult:
             n = cast(Node, n)
             if n.op != "call_function":
                 continue
-            # Don't fold chains of quant-ops into each other.
-            if n.target in (q_op, dq_op):
-                continue
 
             # Make sure we haven't already set qparams meta information on the node
             assert "input_qparams" not in n.meta.keys()

backends/arm/_passes/insert_rescales_pass.py

@@ -31,7 +31,6 @@
     op_reciprocal,
     op_relu,
     op_repeat,
-    op_rescale,
     op_rshift,
     op_rsqrt,
     op_sigmoid,

backends/arm/operators/__init__.py

@@ -9,19 +9,13 @@
 from typing import Tuple
 
 import torch
-from executorch.backends.arm.arm_backend import get_tosa_version
-from executorch.backends.arm.quantizer import arm_quantizer
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.tester.test_pipeline import (
     EthosU55PipelineBI,
     EthosU85PipelineBI,
     TosaPipelineBI,
     TosaPipelineMI,
 )
-from executorch.backends.xnnpack.test.tester import Quantize
-from torch.ao.quantization.observer import HistogramObserver
-from torch.ao.quantization.quantizer import QuantizationSpec
-
 
 aten_op = "torch.ops.aten.add.Tensor"
 exir_op = "executorch_exir_dialects_edge__ops_aten_add_Tensor"
@@ -73,38 +67,6 @@ def test_add_tosa_BI(test_data: input_t1):
     pipeline.run()
 
 
-@common.parametrize("test_data", Add.test_data)
-def test_add_i32_tosa_BI(test_data: input_t1):
-    pipeline = TosaPipelineBI[input_t1](Add(), test_data, aten_op, exir_op)
-
-    # Create a  quantizer with int8 quantization on the input and output but int32 on everything else.
-    quantizer = arm_quantizer.ArmQuantizer(
-        get_tosa_version(common.get_tosa_compile_spec("TOSA-0.80+BI"))
-    )
-    quantizer.set_io(arm_quantizer.get_symmetric_quantization_config())
-    observer_options = {"eps": 2**-16}
-    observer = HistogramObserver.with_args(**observer_options)
-    input_act_qspec = QuantizationSpec(
-        torch.int32,
-        observer,
-        qscheme=torch.per_tensor_symmetric,
-        quant_max=2**31 - 1,
-        quant_min=-(2**31),
-    )
-    # This quantization_config will be set as global config.
-    quantization_config = arm_quantizer.QuantizationConfig(
-        input_act_qspec, None, None, None
-    )
-    quantize_stage = Quantize(quantizer, quantization_config)
-    pipeline.change_args("quantize", quantize_stage)
-
-    # Check that we get the additional (dq -> q
-    pipeline.add_stage_after(
-        "export", pipeline.tester.check_count, {"torch.ops.quantized_decomposed": 8}
-    )
-    pipeline.run()
-
-
 @common.parametrize("test_data", Add.test_data)
 def test_add_u55_BI(test_data: input_t1):
     pipeline = EthosU55PipelineBI[input_t1](