Arm backend: Decompose conv2d with 16 bit activation

per · per · commit 6dc9cda7ebb2 · 2025-09-16T16:17:29.000+02:00
Support quantization to 16a8w. Since the resulting TOSA operator
needs to have the bias in int48 which isn't avaiable as a type in
torch, the conv2d needs to be decomposed into a conv + add, where
the conv result is scaled down to 32 bit before the addition of the
bias is done.

Signed-off-by: Per Åstrand &lt;per.astrand@arm.com&gt;
Change-Id: Ib8cae694035796374a55a9909e501596e983abf5
diff --git a/backends/arm/_passes/__init__.py b/backends/arm/_passes/__init__.py
@@ -46,6 +46,9 @@
 from .decompose_glu_pass import DecomposeGluPass  # noqa
 from .decompose_grouped_conv import DecomposeGroupedConv  # noqa
 from .decompose_groupnorm_pass import DecomposeGroupNormPass  # noqa
+from .decompose_int16_activation_conv2d_pass import (  # noqa
+    DecomposeConv2dWithInt16ActivationPass,
+)
 from .decompose_layernorm_pass import DecomposeLayerNormPass  # noqa
 from .decompose_leaky_relu_pass import DecomposeLeakyReLUPass  # noqa
 from .decompose_linalg_vector_norm_pass import DecomposeLinearVectorNormPass  # noqa
diff --git a/backends/arm/_passes/arm_pass_manager.py b/backends/arm/_passes/arm_pass_manager.py
@@ -39,6 +39,7 @@
     DecomposeAtanPass,
     DecomposeAvgPool2d,
     DecomposeBatchNormNoStatsPass,
+    DecomposeConv2dWithInt16ActivationPass,
     DecomposeCoshPass,
     DecomposeCosineSimilarityPass,
     DecomposeCumsumPass,
@@ -154,6 +155,7 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(ComputeConstantOpsAOT(exported_program))
 
         self.add_pass(DecomposeGroupedConv())
+
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(UnsqueezeBeforeRepeatPass())
         self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
@@ -167,9 +169,14 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
 
         self.add_pass(FuseViewCopyTransform())
         self.add_pass(FuseConstantArgsPass(exported_program))
+        self.add_pass(InsertTableOpsPass(exported_program))
+        # If we have a conv2d with int16 activation split up into a convolution
+        # and an addition, to work-around the lack of support for int48 in torch
+        # needs to happen before AddBiasPass, but after the table ops are inserted
+        # to be able to validate that conv2d has right dtype arguments.
+        self.add_pass(DecomposeConv2dWithInt16ActivationPass())
         self.add_pass(AddBiasPass(exported_program))
 
-        self.add_pass(InsertTableOpsPass(exported_program))
         self.add_pass(FuseEqualPlaceholdersPass(exported_program))
         self.add_pass(ToTosaMemoryFormatPass(exported_program))
         self.add_pass(RemoveNoopPass())
diff --git a/backends/arm/_passes/decompose_int16_activation_conv2d_pass.py b/backends/arm/_passes/decompose_int16_activation_conv2d_pass.py
@@ -0,0 +1,145 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-unsafe
+
+from typing import cast
+
+import torch
+from executorch.backends.arm._passes.quant_args import QuantArgs
+
+from executorch.backends.arm.tosa.specification import get_context_spec, Tosa_1_00
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass
+
+
+class DecomposeConv2dWithInt16ActivationPass(ExportPass):
+    """
+    This pass decomposes a convolution with input dtype int16 and bias
+    into a convolution without bias followed by an addition of the bias
+    since the TOSA op requires the bias to be int48 which is hard to represent
+    in torch. Instead rescale the int48 output to int16 and add the bias in int16.
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        if op != exir_ops.edge.aten.convolution.default:
+            return super().call_operator(op, args, kwargs, meta)
+
+        tosa_spec = get_context_spec()
+        if not tosa_spec.support_integer():
+            return super().call_operator(op, args, kwargs, meta)
+
+        # return if no bias
+        if args[2] is None:
+            return super().call_operator(op, args, kwargs, meta)
+
+        if args[0].data.dtype == torch.int8:
+            return super().call_operator(op, args, kwargs, meta)
+        elif args[0].data.dtype == torch.int16:
+            if isinstance(tosa_spec, Tosa_1_00) and not tosa_spec.support_extension(
+                "int16"
+            ):
+                raise ValueError(
+                    "int16 activation for convolution requires TOSA int16 extension"
+                )
+        else:
+            raise NotImplementedError(
+                "Decomposition to conv+add only implemented for activation of int16 type"
+            )
+
+        # convolution with bias and activation is int16
+        # The bias is assumed to be quantized with the same quantization parameters as
+        # as the output of the convolution
+        bias = args[2]
+        assert (
+            meta.data["output_qparams"][0].dtype == bias.data.dtype
+        ), "Bias needs to have same type as quantized output type"
+        no_bias_args = list(args)
+        no_bias_args[2] = None
+        # split up to convolution + bias
+        convolution = super().call_operator(op, tuple(no_bias_args), kwargs, meta)
+
+        # create a copy of the meta without the qparams, to be used with the new nodes
+        new_meta = meta.copy()
+        new_meta.data.pop("output_qparams", None)
+        new_meta.data.pop("input_qparams", None)
+
+        # reshape the tensor to the same rank as the convolution output to add the bias to the channels
+        channel_bias = super().call_operator(
+            exir_ops.edge.aten.view_copy.default,
+            (bias, [1, len(bias.data), 1, 1]),
+            {},
+            new_meta,
+        )
+
+        output_dtype = meta.data["output_qparams"][0].dtype
+
+        if output_dtype == torch.int16:
+            # The conv will get the output int48 scaled to int32 in serialization step.
+            # To be able to add the bias we need to first scale (cast?) the output to int32.
+            # The resulting i32 sum will then need to be scaled back to the output dtype.
+
+            # calculate common rescale factor from convolution output and bias quantization
+            output_qparams = cast(QuantArgs, meta.data["output_qparams"][0])
+            conv_output_scale = output_qparams.scale
+            bias_qparams = cast(QuantArgs, meta.data["input_qparams"][2])
+            bias_scale = bias_qparams.scale
+
+            common_scale = max(bias_scale, conv_output_scale)
+
+            # calculate how we can rescale bias and conv to a common scale and maximize the output range
+            bias_rescale_factor = bias_scale / common_scale
+            conv_rescale_factor = conv_output_scale / common_scale
+
+            # Either of conv output or bias now covers the full int16 range and the other one a smaller range.
+            # Since we are upscaling to int32 we have 16 additional bits to work with to maximize the output range.
+            # Worst case here is that both bias and conv output covers the full int16 range so we leave one bit
+            # and then one for the sign bit.
+            bits_left_to_shift = 14
+
+            # update rescale factors
+            bias_rescale_factor *= 1 << bits_left_to_shift
+            conv_rescale_factor *= 1 << bits_left_to_shift
+
+            conv_output = super().call_operator(
+                exir_ops.backend.tosa.RESCALE.default,
+                (convolution, torch.int32, conv_rescale_factor, 0, 0),
+                {},
+                new_meta,
+            )
+
+            bias_rescaled = super().call_operator(
+                exir_ops.backend.tosa.RESCALE.default,
+                (channel_bias, torch.int32, bias_rescale_factor, 0, 0),
+                {},
+                new_meta,
+            )
+
+            add = super().call_operator(
+                exir_ops.edge.aten.add.Tensor,
+                (conv_output, bias_rescaled),
+                {},
+                new_meta,
+            )
+
+            res_rescale = super().call_operator(
+                exir_ops.backend.tosa.RESCALE.default,
+                (
+                    add,
+                    output_dtype,
+                    (common_scale / (conv_output_scale * (1 << bits_left_to_shift))),
+                    0,
+                    0,
+                ),
+                {},
+                new_meta,
+            )
+
+        else:
+            raise NotImplementedError(
+                f"Decomposition to conv+add only implemented for activation of int16 type, not for {output_dtype}"
+            )
+
+        return res_rescale
diff --git a/backends/arm/quantizer/quantization_config.py b/backends/arm/quantizer/quantization_config.py
@@ -89,29 +89,48 @@ def _derive_qparams_fn(
             torch.ops.aten.linear.default,
             torch.ops.aten.conv2d.padding,
         ]:
-            input_act = node.args[0]
-            weight = node.args[1]
-            # If the weights are quantized per_tensor, do the same with bias
-            qscheme = (
-                torch.per_tensor_symmetric
-                if self.weight is None
-                else self.weight.qscheme
-            )
-            ch_axis = None
-            if self.weight is not None:
-                if qscheme == torch.per_channel_symmetric:
-                    ch_axis = self.weight.ch_axis
+            if self.input_activation is None or self.weight is None:
+                raise ValueError(
+                    "Input activation and weight QuantizationConfig must be specified."
+                )
+            if self.input_activation.dtype == self.weight.dtype == torch.int8:
+                # This is the default int8 quantization which uses the derived quantization
+                # calculated from the activation and weight scale
+                input_act = node.args[0]
+                weight = node.args[1]
 
-            quantization_spec = DerivedQuantizationSpec(
-                derived_from=[(input_act, node), (weight, node)],  # type: ignore[list-item]
-                derive_qparams_fn=_derive_qparams_fn,
-                dtype=torch.int32,
-                quant_min=torch.iinfo(torch.int32).min,
-                quant_max=torch.iinfo(torch.int32).max - 1,
-                qscheme=qscheme,
-                ch_axis=ch_axis,
-            )
-            return quantization_spec  # type: ignore[return-value]
+                # If the weights are quantized per_tensor, do the same with bias
+                qscheme = (
+                    torch.per_tensor_symmetric
+                    if self.weight is None
+                    else self.weight.qscheme
+                )
+                ch_axis = None
+                if self.weight is not None:
+                    if qscheme == torch.per_channel_symmetric:
+                        ch_axis = self.weight.ch_axis
+
+                quantization_spec = DerivedQuantizationSpec(
+                    derived_from=[(input_act, node), (weight, node)],  # type: ignore[list-item]
+                    derive_qparams_fn=_derive_qparams_fn,
+                    dtype=torch.int32,
+                    quant_min=torch.iinfo(torch.int32).min,
+                    quant_max=torch.iinfo(torch.int32).max - 1,
+                    qscheme=qscheme,
+                    ch_axis=ch_axis,
+                )
+                return quantization_spec  # type: ignore[return-value]
+            elif (
+                self.input_activation.dtype == torch.int16
+                and self.weight.dtype == torch.int8
+            ):
+                # In case the activation is quantized to int16, the bias needs to be
+                # added after the convolution, so use the output quantization for this case.
+                return self.output_activation
+            else:
+                raise NotImplementedError(
+                    f"Bias quantization of types: i:{self.input_activation.dtype}, w:{self.weight.dtype} not implemented"
+                )
 
         if self.bias is None:
             return None
