Check that zero point and scale of inputs and outputs match for Concatenate operator.

The initial checks within VisitConcatenationNode all pass as xnn_define_concatenate is not called when checking if a node may be delegated or not. It is during the second pass that the check fails, causing delegation to fail.

INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
ERROR: failed to delegate CONCATENATION node #31
ERROR: failed to delegate CONCATENATION node #36
ERROR: Node number 55 (TfLiteXNNPackDelegate) failed to prepare.
PiperOrigin-RevId: 514364419

Author: alankelly

tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc

@@ -2816,6 +2816,33 @@ class Subgraph {
     if (axis < 0) axis += NumDimensions(&output_tensor);
     int sum_axis = 0;
 
+    if (output_tensor.type == kTfLiteUInt8) {
+      const int32_t zero_point =
+          tensors[node->outputs->data[0]].params.zero_point;
+      const float scale = tensors[node->outputs->data[0]].params.scale;
+      for (int i = 0; i < num_inputs; i++) {
+        if (tensors[node->inputs->data[i]].params.zero_point != zero_point) {
+          TF_LITE_MAYBE_KERNEL_LOG(
+              logging_context,
+              "Mismatching quantization zero point across the %dth input "
+              "(%" PRId32 ") and the output (%" PRId32
+              ") for CONCATENATE operator #%d",
+              i, tensors[node->inputs->data[i]].params.zero_point, zero_point,
+              node_index);
+          return kTfLiteError;
+        }
+        if (tensors[node->inputs->data[i]].params.scale != scale) {
+          TF_LITE_MAYBE_KERNEL_LOG(
+              logging_context,
+              "Mismatching quantization scale across the %dth input (%f) "
+              "and the output (%f) for CONCATENATE operator #%d",
+              i, tensors[node->inputs->data[i]].params.scale, scale,
+              node_index);
+          return kTfLiteError;
+        }
+      }
+    }
+
     for (int i = 0; i < num_inputs; i++) {
       const TfLiteTensor& input_tensor = tensors[node->inputs->data[i]];
       TF_LITE_ENSURE_STATUS(CheckTensorFloat32OrQUInt8Type(