Cortex-M: Fuse relu activation into quantized_add (#18462)

### Summary
ResNet8 has skip connections with relu(add(conv(x), skip(x))). The
ActivationFusionPass only fused relu into conv/linear, leaving 3 unfused
relu ops that fell through to portable aten::relu.out which incorrectly
clamps int8 tensors to literal 0 instead of the quantized zero_point,
causing numerical mismatches on the FVP.

Add fused activation patterns (relu, hardtanh, clamp) for add/add_ to
quantizer_support.py BINARY_OP_PATTERNS so the quantizer produces
activation-aware quantization bounds. Add aten.add.Tensor to
ActivationFusionPass FUSE_OPS. Update QuantizedOpFusionPass to read
activation bounds from output_qparams and pass them to quantized_add.
Update the quantized_add operator (schema, meta, impl, C++) to accept
activation_min/activation_max parameters.

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: rascani

backends/cortex_m/ops/op_quantized_add.cpp

@@ -26,6 +26,8 @@ Tensor& quantized_add_out(
     const int64_t output_zero_point,
     const int64_t output_multiplier,
     const int64_t output_shift,
+    const int64_t activation_min,
+    const int64_t activation_max,
     Tensor& out) {
   // Validate tensor types and dim order
   bool channel_broadcast = is_channel_broadcast(input1_int8, input2_int8);
@@ -69,8 +71,8 @@ Tensor& quantized_add_out(
 
   // Left shift to maximize precision
   const int32_t left_shift = 20;
-  const int32_t activation_min = std::numeric_limits<int8_t>::min();
-  const int32_t activation_max = std::numeric_limits<int8_t>::max();
+  const int32_t act_min = static_cast<int32_t>(activation_min);
+  const int32_t act_max = static_cast<int32_t>(activation_max);
 
   ET_LOG(
       Debug,
@@ -121,8 +123,8 @@ Tensor& quantized_add_out(
         static_cast<int32_t>(out_zp),
         output_mult,
         output_shift_val,
-        activation_min,
-        activation_max,
+        act_min,
+        act_max,
         adds_per_loop);
 
     if (status != ARM_CMSIS_NN_SUCCESS) {

backends/cortex_m/ops/operators.py

@@ -124,15 +124,16 @@ def dequantize_per_tensor_impl(
     "quantized_add("
     "Tensor self, int self_zero_point, int self_multiplier, int self_shift, "
     "Tensor other, int other_zero_point, int other_multiplier, int other_shift, "
-    "int output_zero_point, int output_multiplier, int output_shift) -> Tensor"
+    "int output_zero_point, int output_multiplier, int output_shift, "
+    "int activation_min, int activation_max) -> Tensor"
 )
 
-# Define the operator schema with multipliers and shifts (11 args + out tensor)
 lib.define(
     "quantized_add.out("
     "Tensor self, int self_zero_point, int self_multiplier, int self_shift, "
     "Tensor other, int other_zero_point, int other_multiplier, int other_shift, "
     "int output_zero_point, int output_multiplier, int output_shift, "
+    "int activation_min, int activation_max, "
     "*, Tensor(a!) out) -> Tensor(a!)"
 )
 
@@ -150,6 +151,8 @@ def quantized_add_meta(
     output_zero_point: int,
     output_multiplier: int,
     output_shift: int,
+    activation_min: int,
+    activation_max: int,
 ) -> torch.Tensor:
     assert self.shape == other.shape or is_channel_broadcast(self, other), (
         "Cortex-M quantized_add: broadcasting is not yet supported except for channel dim — "
@@ -175,6 +178,8 @@ def quantized_add_impl(
     output_zero_point: int,
     output_multiplier: int,
     output_shift: int,
+    activation_min: int,
+    activation_max: int,
 ) -> torch.Tensor:
     assert self.shape == other.shape or is_channel_broadcast(self, other), (
         "Cortex-M quantized_add: broadcasting is not yet supported except for channel dim — "
@@ -188,7 +193,9 @@ def quantized_add_impl(
 
     result_fp = self_fp + other_fp
     result_quantized = requantize_cmsis(result_fp, output_multiplier, output_shift)
-    result = torch.clamp(result_quantized + output_zero_point, -128, 127).to(torch.int8)
+    result = torch.clamp(
+        result_quantized + output_zero_point, activation_min, activation_max
+    ).to(torch.int8)
     return result
 
 

backends/cortex_m/ops/operators.yaml

@@ -17,7 +17,7 @@
     - arg_meta: null
       kernel_name: cortex_m::dequantize_per_tensor_out
 
-- func: cortex_m::quantized_add.out(Tensor self, int self_zero_point, int self_multiplier, int self_shift, Tensor other, int other_zero_point, int other_multiplier, int other_shift, int output_zero_point, int output_multiplier, int output_shift, *, Tensor(a!) out) -> Tensor(a!)
+- func: cortex_m::quantized_add.out(Tensor self, int self_zero_point, int self_multiplier, int self_shift, Tensor other, int other_zero_point, int other_multiplier, int other_shift, int output_zero_point, int output_multiplier, int output_shift, int activation_min, int activation_max, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels:
     - arg_meta: null

backends/cortex_m/passes/activation_fusion_pass.py

@@ -40,6 +40,7 @@ class ActivationFusionPass(ExportPass):
     FUSE_OPS = {
         exir_ops.edge.aten.linear.default,
         exir_ops.edge.aten.convolution.default,
+        exir_ops.edge.aten.add.Tensor,
     }
 
     def _get_validated_qparams(self, node, input_node):
@@ -85,7 +86,7 @@ def _get_validated_qparams(self, node, input_node):
                     else qmax
                 )
             case _:
-                raise RuntimeError("Unexpected target {node.target}.")
+                raise RuntimeError(f"Unexpected target {node.target}.")
 
         # If the minimal quantized value is larger than the qmin, it means that the quantized range contains
         # invalid values [qmin, ..., quantized_min_val-1], indicating bad quantization parameters.

backends/cortex_m/passes/quantized_op_fusion_pass.py

@@ -70,6 +70,9 @@ def _get_add_replacement(self, args, meta):
             max_scale_2x / (output_scale * (1 << SHIFT_INT8))
         )
 
+        activation_min = meta["output_qparams"][0].qmin
+        activation_max = meta["output_qparams"][0].qmax
+
         args = (
             args[0],
             zero_point1,
@@ -82,6 +85,8 @@ def _get_add_replacement(self, args, meta):
             output_zero_point,
             output_mult,
             output_shift,
+            activation_min,
+            activation_max,
         )
 
         return exir_ops.edge.cortex_m.quantized_add.default, args

backends/cortex_m/quantizer/quantizer_support.py

@@ -17,6 +17,12 @@
 
 BINARY_OP_PATTERNS = {
     (torch.ops.aten.add.Tensor,): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.relu.default): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.relu_.default): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.hardtanh.default): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.hardtanh_.default): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.clamp.default): CortexMAddMulCheck,
+    (torch.ops.aten.add.Tensor, torch.ops.aten.clamp_.default): CortexMAddMulCheck,
     (torch.ops.aten.add_.Tensor,): CortexMAddMulCheck,
     (torch.ops.aten.mul.Tensor,): CortexMAddMulCheck,
     (torch.ops.aten.mul_.Tensor,): CortexMAddMulCheck,

backends/cortex_m/test/models/test_nn_modules.py

@@ -188,9 +188,7 @@ def forward(self, x):
     ),
 }
 
-xfails: dict[str, xfail_type] = {
-    "conv_add_relu": "Activation fusion does not support relu after add",
-}
+xfails: dict[str, xfail_type] = {}
 
 
 @parametrize("test_case", test_cases, xfails=xfails, strict=False)

backends/cortex_m/test/ops/test_add.py

@@ -73,6 +73,50 @@ class CortexMAlphaAdd(ModelAlpha):
     }
 
 
+class CortexMAddReLU(torch.nn.Module):
+    ops_before_transforms = {
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor": 1,
+        "executorch_exir_dialects_edge__ops_aten_relu_default": 1,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_default": 3,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default": 3,
+    }
+
+    ops_after_transforms = {
+        "executorch_exir_dialects_edge__ops_cortex_m_quantized_add_default": 1,
+        "executorch_exir_dialects_edge__ops_cortex_m_quantize_per_tensor_default": 2,
+        "executorch_exir_dialects_edge__ops_cortex_m_dequantize_per_tensor_default": 1,
+    }
+
+    def __init__(self):
+        super().__init__()
+        self.relu = torch.nn.ReLU()
+
+    def forward(self, x, y):
+        return self.relu(x + y)
+
+
+class CortexMAddHardtanh(torch.nn.Module):
+    ops_before_transforms = {
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor": 1,
+        "executorch_exir_dialects_edge__ops_aten_hardtanh_default": 1,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_default": 3,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default": 3,
+    }
+
+    ops_after_transforms = {
+        "executorch_exir_dialects_edge__ops_cortex_m_quantized_add_default": 1,
+        "executorch_exir_dialects_edge__ops_cortex_m_quantize_per_tensor_default": 2,
+        "executorch_exir_dialects_edge__ops_cortex_m_dequantize_per_tensor_default": 1,
+    }
+
+    def __init__(self, min_val=-0.5, max_val=0.5):
+        super().__init__()
+        self.act = torch.nn.Hardtanh(min_val=min_val, max_val=max_val)
+
+    def forward(self, x, y):
+        return self.act(x + y)
+
+
 test_cases = {
     "self_rank_1": McuTestCase(
         CortexMSelfAdd(),
@@ -149,6 +193,34 @@ class CortexMAlphaAdd(ModelAlpha):
             ramp_tensor(-20, 20, (4, 5)),
         ),
     ),
+    "add_relu": McuTestCase(
+        CortexMAddReLU(),
+        (
+            ramp_tensor(-5, 5, (2, 4)),
+            ramp_tensor(-3, 3, (2, 4)),
+        ),
+    ),
+    "add_relu_channels_last": McuTestCase(
+        CortexMAddReLU(),
+        (
+            ramp_tensor(-5, 5, (1, 4, 8, 8)).to(memory_format=torch.channels_last),
+            ramp_tensor(-3, 3, (1, 4, 8, 8)).to(memory_format=torch.channels_last),
+        ),
+    ),
+    "add_hardtanh": McuTestCase(
+        CortexMAddHardtanh(min_val=-0.5, max_val=0.5),
+        (
+            ramp_tensor(-2, 2, (2, 4)),
+            ramp_tensor(-1, 1, (2, 4)),
+        ),
+    ),
+    "add_hardtanh_channels_last": McuTestCase(
+        CortexMAddHardtanh(min_val=-1.0, max_val=1.0),
+        (
+            ramp_tensor(-3, 3, (1, 4, 8, 8)).to(memory_format=torch.channels_last),
+            ramp_tensor(-2, 2, (1, 4, 8, 8)).to(memory_format=torch.channels_last),
+        ),
+    ),
 }