Fix conv1d QuantFusion issue (#18623)

Differential Revision: D98941237

Pull Request resolved: #18623

Author: mcremon-meta

backends/cadence/aot/quantizer/fusion_pass.py

@@ -40,7 +40,6 @@
     copy_node_metadata,
     create_zero_bias_int32,
     find_sequential_partitions_aten,
-    get_conv_args,
     quantize_tensor_multiplier,
 )
 from executorch.exir.pass_base import ExportPass
@@ -263,10 +262,10 @@ def get_args_and_kwargs_conv(
     weight_zero_point = dequants_weights[0].args[2]
     # pyre-fixme[58]: Unsupported operand types
     bias_scale = dequants_inputs[0].args[1] * weight_scale
-    stride = [1, 1] if len(op_node.args) < 4 else get_conv_args(op_node.args[3], 1)
-    padding = [0, 0] if len(op_node.args) < 5 else get_conv_args(op_node.args[4], 0)
-    dilation = [1, 1] if len(op_node.args) < 6 else get_conv_args(op_node.args[5], 1)
-    groups = 1 if len(op_node.args) < 7 else op_node.args[6]
+    stride = get_arg(op_node, "stride", list[int])
+    padding = get_arg(op_node, "padding", list[int])
+    dilation = get_arg(op_node, "dilation", list[int])
+    groups = get_arg(op_node, "groups", int)
 
     # If bias is not available, create a bias tensor with the shape of weight[0]
     if not bias_inputs:

backends/cadence/aot/quantizer/utils.py

@@ -170,10 +170,6 @@ def get_bias_qparams(
     return bias_scale, bias_zero_point
 
 
-def get_conv_args(arg, first_val: int) -> List[fx.Node]:
-    return arg if len(arg) == 2 else [first_val, arg[0]]
-
-
 def get_aten_node_target_partitions(
     graph: torch.fx.Graph,
     wanted_original_aten_op: List[OpOverload],

backends/cadence/hifi/operators/op_quantized_conv1d_nlc.cpp

@@ -262,9 +262,9 @@ void quantized_conv1d_nlc_per_tensor_out(
   ScalarType dtype = out.scalar_type();
 
   if (dtype == ScalarType::Char) {
-    // HiFi nnlib conv2d kernel produces incorrect results with stride > 1
-    // on some backends (e.g., Artemis HiFi4). Fall back to generic.
-    if (stride[0] > 1) {
+    // HiFi nnlib conv2d kernel does not support depthwise (groups > 1)
+    // or stride > 1. Fall back to generic implementation.
+    if (groups > 1 || stride[0] > 1) {
       impl::generic::native::quantized_conv1d_nlc_per_tensor_out(
           ctx,
           input,