Arm backend: Remove CheckNeedsDecomposition

backends/arm/operator_support/tosa_profile_supported_op_lists.py

@@ -18,6 +18,7 @@
 
 
 # INT profile: ops supported via native TOSA ops, decompositions/transformations, precompute, TableOps, etc.
+# Note that ops supported via pre-quantization decompositions are not included here.
 TOSA_PRO_INT_SupportList: Final[Set] = {
     exir_ops.edge.aten.abs.default,
     exir_ops.edge.aten.add.Tensor,
@@ -46,8 +47,6 @@
     exir_ops.edge.aten.hardsigmoid.default,
     exir_ops.edge.aten.hardtanh.default,
     exir_ops.edge.aten.hardswish.default,
-    exir_ops.edge.aten.div.Tensor,
-    exir_ops.edge.aten.div.Tensor_mode,
     exir_ops.edge.aten.eq.Tensor,
     exir_ops.edge.aten.eq.Scalar,
     exir_ops.edge.aten.erf.default,
@@ -68,16 +67,7 @@
     exir_ops.edge.aten.lt.Tensor,
     exir_ops.edge.aten.lt.Scalar,
     exir_ops.edge.aten.mul.Tensor,
-    exir_ops.edge.aten.ne.Tensor,
-    exir_ops.edge.aten.ne.Scalar,
     exir_ops.edge.aten.neg.default,
-    exir_ops.edge.aten.add.Scalar,
-    exir_ops.edge.aten.sub.Scalar,
-    exir_ops.edge.aten.mul.Scalar,
-    exir_ops.edge.aten.div.Scalar,
-    exir_ops.edge.aten._native_batch_norm_legit_no_training.default,
-    exir_ops.edge.aten.native_layer_norm.default,
-    exir_ops.edge.aten.native_group_norm.default,
     exir_ops.edge.aten.sigmoid.default,
     exir_ops.edge.aten.mean.dim,
     exir_ops.edge.aten.mm.default,
@@ -86,19 +76,12 @@
     exir_ops.edge.aten.repeat.default,
     exir_ops.edge.aten.reciprocal.default,
     exir_ops.edge.aten.relu.default,
-    exir_ops.edge.aten.leaky_relu.default,
-    exir_ops.edge.aten.sqrt.default,
     exir_ops.edge.aten.rsqrt.default,
-    exir_ops.edge.aten.round.default,
-    exir_ops.edge.aten._softmax.default,
     exir_ops.edge.aten.select_copy.int,
-    exir_ops.edge.aten._log_softmax.default,
     exir_ops.edge.aten.sub.Tensor,
     exir_ops.edge.aten.tanh.default,
     exir_ops.edge.aten.upsample_bilinear2d.vec,
     exir_ops.edge.aten.upsample_nearest2d.vec,
-    exir_ops.edge.aten.var.correction,
-    exir_ops.edge.aten.var.dim,
     exir_ops.edge.aten.view_copy.default,
     exir_ops.edge.aten.unsqueeze_copy.default,
     exir_ops.edge.aten.squeeze_copy.dims,
@@ -127,12 +110,9 @@
     exir_ops.edge.aten.sign.default,
     exir_ops.edge.aten.asin.default,
     exir_ops.edge.aten.atanh.default,
-    exir_ops.edge.aten.addmm.default,
     exir_ops.edge.aten.masked_fill.Scalar,
     exir_ops.edge.aten.asinh.default,
     exir_ops.edge.aten.cosh.default,
-    exir_ops.edge.aten.glu.default,
-    exir_ops.edge.aten.logit.default,
     exir_ops.edge.aten.acos.default,
     exir_ops.edge.aten.elu.default,
     exir_ops.edge.aten.bitwise_not.default,

backends/arm/operator_support/tosa_supported_operators.py

@@ -135,7 +135,6 @@ def tosa_support_factory(
     ]
 
     if not tosa_spec.support_float():
-        negative_checks.append(NeedsDecompositionCheck(reporter))
         negative_checks.append(CheckProperQuantization(reporter))
     if tosa_spec.is_U55_subset:
         negative_checks.append(EthosU55NotSupported(reporter))
@@ -156,7 +155,8 @@ def tosa_support_factory(
 class TOSAProINTSupportList(OperatorSupportBase):
     """
     TOSA_PRO_INT_SupportList:
-        Ops supported in INT profile via native TOSA ops, decomposition/transformation, pre-compute, or TableOps
+        Ops supported in INT profile via native TOSA ops, decomposition/transformation, pre-compute, or TableOps.
+        Note that ops supported via pre-quantization decompositions are not included here.
     """
 
     def is_node_supported(
@@ -179,57 +179,6 @@ def is_node_supported(
         return node.op == "call_function" and node.target in TOSA_PRO_FP_SupportList
 
 
-class NeedsDecompositionCheck(OperatorSupportBase):
-    """
-    Targeted operators need to be decomposed prior to quantization in order to get a pair of q-dq-nodes surrounding
-    the operator, and to get optimal quantization parameters for each operator. This check will reject operators
-    that need to be decomposed.
-    """
-
-    def __init__(self, reporter: WhyNoPartitionReporter):
-        self.reporter = reporter
-
-    def is_node_supported(
-        self, submodules: typing.Mapping[str, torch.nn.Module], node: fx.Node
-    ) -> bool:
-
-        if node.op != "call_function":
-            return True
-
-        needs_decomp_dict = {
-            exir_ops.edge.aten.div.Tensor: None,
-            exir_ops.edge.aten._native_batch_norm_legit_no_training.default: "BatchNorm2D with track_running_stats==True not immediately following a convolution is not supported for quantized TOSA backends.",
-            exir_ops.edge.aten.native_layer_norm.default: None,
-            exir_ops.edge.aten.native_group_norm.default: None,
-            exir_ops.edge.aten._softmax.default: None,
-            exir_ops.edge.aten._log_softmax.default: None,
-            exir_ops.edge.aten.var.correction: None,
-            exir_ops.edge.aten.var.dim: None,
-            exir_ops.edge.aten.add.Scalar: None,
-            exir_ops.edge.aten.sqrt.default: None,
-            exir_ops.edge.aten.sub.Scalar: None,
-            exir_ops.edge.aten.mul.Scalar: None,
-            exir_ops.edge.aten.ne.Tensor: None,
-            exir_ops.edge.aten.ne.Scalar: None,
-            exir_ops.edge.aten.div.Scalar: None,
-            exir_ops.edge.aten.leaky_relu.default: None,
-            exir_ops.edge.aten.round.default: None,
-            exir_ops.edge.aten.addmm.default: None,
-            exir_ops.edge.aten.glu.default: None,
-            exir_ops.edge.aten.logit.default: None,
-        }
-
-        if node.target in needs_decomp_dict:
-            reject_message = needs_decomp_dict[node.target]
-            if reject_message is None:
-                reject_message = "Op needs to be decomposed into other ops before quantization to get quantized properly."
-
-            self.reporter.report_reject(node, reject_message)
-            return False
-        else:
-            return True
-
-
 class CheckProperQuantization(OperatorSupportBase):
     """
     For targeted nodes, check that it has been quantized as expected. In most cases this means that a pair of quantize