Add feedback

Author: sfvaroglu

src/relay/qnn/op/op_common.h

@@ -261,8 +261,8 @@ static inline bool QnnBroadcastRel(const Array<Type>& types, int num_inputs, con
   auto lhs_rank = static_cast<int>(lhs_data->shape.size());
   auto rhs_rank = static_cast<int>(rhs_data->shape.size());
 
-  lhs_axis = (lhs_axis < 0) ? ((lhs_rank > 0) ? lhs_data->shape.size() + lhs_axis : 0) : lhs_axis;
-  rhs_axis = (rhs_axis < 0) ? ((rhs_rank > 0) ? rhs_data->shape.size() + rhs_axis : 0) : rhs_axis;
+  lhs_axis = (lhs_axis < 0) ? ((lhs_rank > 0) ? lhs_rank + lhs_axis : 0) : lhs_axis;
+  rhs_axis = (rhs_axis < 0) ? ((rhs_rank > 0) ? rhs_rank + rhs_axis : 0) : rhs_axis;
 
   // If zero point and scale are scalar then axis doesn't matter.
   bool lhs_scale_is_scalar = (types[2].as<TensorTypeNode>())->shape.size() == 0;
@@ -349,7 +349,6 @@ static inline bool QnnBroadcastRel(const Array<Type>& types, int num_inputs, con
       .add_argument("lhs_axis", "Tensor", "The channel quantization of the lhs tensor.")           \
       .add_argument("rhs_axis", "Tensor", "The channel quantization of the rhs tensor.")           \
       .add_type_rel("QnnBroadcast", QnnBroadcastRel)                                               \
-      .set_attr<TOpPattern>("TOpPattern", kOpaque)                                                 \
       .set_attr<TNonComputational>("TNonComputational", true)                                      \
       .set_attr<FInferCorrectLayout>("FInferCorrectLayout", QnnBinaryBroadcastLayout)
 

tests/python/relay/test_pass_fake_quantization_to_integer.py

@@ -611,17 +611,14 @@ def verify_binary_per_channel(lhs_scale, rhs_scale, lhs_zp, rhs_zp, out_zp, lhs_
             rhs_axis = lhs_axis  # TODO: Support different axes for per-channel quantized multiply
         else:
             out_scale = relay.const(0.1)
+
         x = relay.var("x", shape=[1, 3, 224, 224], dtype="int8")
         x = relay.qnn.op.dequantize(x, relay.const(lhs_scale), relay.const(lhs_zp), axis=lhs_axis)
 
         y = relay.var("y", shape=[1, 3, 224, 224], dtype="int8")
         y = relay.qnn.op.dequantize(y, relay.const(rhs_scale), relay.const(rhs_zp), axis=rhs_axis)
 
         op = operator(x, y)
-        if operator == relay.op.multiply:
-            out_scale = relay.const(2.0)
-        else:
-            out_scale = relay.const(0.1)
 
         op = relay.qnn.op.quantize(op, out_scale, relay.const(out_zp), out_dtype="int8")
         x_np = np.random.randint(-25, 25, size=[1, 3, 224, 224], dtype="int8")