Add clarification for bias quantization in QlinearConv Op spec (#2464)

Author: askhade

docs/Changelog.md

@@ -9735,6 +9735,8 @@ This version of the operator has been available since version 10 of the default
   and computes the quantized output. Each scale and zero-point pair must have same shape.
   It means they must be either scalars (per tensor) or 1-D tensors (per output channel).
   Each input or output and its related zero point must have same type.
+  When bias is present it must be quantized using scale = input scale * weight scale and 
+  zero point as 0.
 
 #### Version
 
@@ -9777,7 +9779,7 @@ This version of the operator has been available since version 10 of the default
 <dt><tt>y_zero_point</tt> : T3</dt>
 <dd>Scale tensor for output 'y'. It's a scalar, which means a per-tensor/layer quantization.</dd>
 <dt><tt>B</tt> (optional) : T4</dt>
-<dd>Optional 1D bias to be added to the convolution, has size of M.</dd>
+<dd>Optional 1D bias to be added to the convolution, has size of M. Bias must be quantized using scale = x_scale * w_scale and zero_point = 0</dd>
 </dl>
 
 #### Outputs

docs/Operators.md

@@ -10700,6 +10700,8 @@ expect(node, inputs=[x, y], outputs=[z],
   and computes the quantized output. Each scale and zero-point pair must have same shape.
   It means they must be either scalars (per tensor) or 1-D tensors (per output channel).
   Each input or output and its related zero point must have same type.
+  When bias is present it must be quantized using scale = input scale * weight scale and 
+  zero point as 0.
 
 #### Version
 
@@ -10742,7 +10744,7 @@ This version of the operator has been available since version 10 of the default
 <dt><tt>y_zero_point</tt> : T3</dt>
 <dd>Scale tensor for output 'y'. It's a scalar, which means a per-tensor/layer quantization.</dd>
 <dt><tt>B</tt> (optional) : T4</dt>
-<dd>Optional 1D bias to be added to the convolution, has size of M.</dd>
+<dd>Optional 1D bias to be added to the convolution, has size of M. Bias must be quantized using scale = x_scale * w_scale and zero_point = 0</dd>
 </dl>
 
 #### Outputs

onnx/defs/nn/defs.cc

@@ -778,6 +778,8 @@ a quantized filter, its scale and zero point, and output's scale and zero point,
 and computes the quantized output. Each scale and zero-point pair must have same shape.
 It means they must be either scalars (per tensor) or 1-D tensors (per output channel).
 Each input or output and its related zero point must have same type.
+When bias is present it must be quantized using scale = input scale * weight scale and 
+zero point as 0.
 )DOC";
 
 ONNX_OPERATOR_SET_SCHEMA(
@@ -849,7 +851,8 @@ ONNX_OPERATOR_SET_SCHEMA(
         .Input(
             8,
             "B",
-            "Optional 1D bias to be added to the convolution, has size of M.",
+            "Optional 1D bias to be added to the convolution, has size of M. "
+            "Bias must be quantized using scale = x_scale * w_scale and zero_point = 0",
             "T4",
             OpSchema::Optional)
         .Output(