Add feedback - round 2

Author: sfvaroglu

python/tvm/relay/qnn/op/qnn.py

@@ -633,10 +633,12 @@ def add(
        The zero point of output quantized expr.
 
     lhs_axis: int
-        The channel axis for lhs quantization. Default value is -1.
+        The channel axis for lhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     rhs_axis: int
-        The channel axis for rhs quantization. Default value is -1.
+        The channel axis for rhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     Returns
     -------
@@ -759,10 +761,12 @@ def mul(
        The zero point of output quantized expr.
 
     lhs_axis: int
-        The channel axis for lhs quantization. Default value is -1.
+        The channel axis for lhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     rhs_axis: int
-        The channel axis for rhs quantization. Default value is -1.
+        The channel axis for rhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     Returns
     -------
@@ -1035,10 +1039,12 @@ def subtract(
        The zero point of output quantized expr.
 
     lhs_axis: int
-        The channel axis for lhs quantization. Default value is -1.
+        The channel axis for lhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     rhs_axis: int
-        The channel axis for rhs quantization. Default value is -1.
+        The channel axis for rhs quantization. Default value is -1 which corresponds
+        to the last axis.
 
     Returns
     -------

src/relay/qnn/op/mul.cc

@@ -59,7 +59,7 @@ Expr QnnMulCanonicalize(const Attrs& attrs, const Array<Expr>& new_args,
   auto lhs_axis = broadcast_attrs->lhs_axis;
   auto rhs_axis = broadcast_attrs->rhs_axis;
 
-  if (lhs_axis == -1 && rhs_axis == -1) {
+  if (IsConstScalar(args.lhs_scale) && IsConstScalar(args.rhs_scale)) {
     /*
     This is per-tensor quantized multiply.
 

tests/python/relay/test_pass_fake_quantization_to_integer.py

@@ -654,6 +654,15 @@ def verify_binary_per_channel(lhs_scale, rhs_scale, lhs_zp, rhs_zp, out_zp, lhs_
         lhs_axis=1,
         rhs_axis=1,
     )
+    verify_binary_per_channel(
+        lhs_scale=np.random.uniform(1.0, 5.0, 224),
+        rhs_scale=np.random.uniform(1.0, 5.0, 224),
+        lhs_zp=np.random.randint(1, 3),
+        rhs_zp=np.random.randint(1, 3),
+        out_zp=np.random.randint(1, 3),
+        lhs_axis=-1,
+        rhs_axis=-1,
+    )
 
     # Different axes
     verify_binary_per_channel(