Enable optimized ADD/SUB for HiFi5 (#2399)

The HiFi5 nnlib has the same optimized kernels for ADD and SUB as the HiFi4 nnlib, but the proper defines were not added to the kernel implementation. This was resulting in falling back to reference kernels for both of those operators on HiFi5.

BUG=none

tensorflow/lite/micro/kernels/xtensa/add.cc

diff --git a/tensorflow/lite/micro/kernels/xtensa/add.cc b/tensorflow/lite/micro/kernels/xtensa/add.cc
index 423e798..d3c2371 100644
--- a/tensorflow/lite/micro/kernels/xtensa/add.cc
+++ b/tensorflow/lite/micro/kernels/xtensa/add.cc
@@ -113,11 +113,11 @@
   op_params.output_shift = data->output_shift;
   SetActivationParams(data->output_activation_min, data->output_activation_max,
                       &op_params);
-#if !(defined(HIFI3) || defined(HIFI4))
+#if !(defined(HIFI3) || defined(HIFI4) || defined(HIFI5))
   bool need_broadcast = reference_ops::ProcessBroadcastShapes(
       tflite::micro::GetTensorShape(input1),
       tflite::micro::GetTensorShape(input2), &op_params);
-#endif  // !defined(HIFI3) && !defined(HIFI4)
+#endif  // !defined(HIFI3) && !defined(HIFI4) && !defined(HIFI5)
 
   switch (output->type) {
     case kTfLiteInt8: {
@@ -126,7 +126,7 @@
           *(reinterpret_cast<XtensaAddOpData*>(node->user_data));
       AddEvalQuantizedVision(context, node, *params, op_data, input1, input2,
                              output);
-#elif defined(HIFI3) || defined(HIFI4)  // defined(VISION_P6)
+#elif defined(HIFI3) || defined(HIFI4) || defined(HIFI5)  // defined(VISION_P6)
       int err;
       const RuntimeShape extended_input1_shape =
           RuntimeShape::ExtendedShape(4, tflite::micro::GetTensorShape(input1));
@@ -150,7 +150,7 @@
           op_params.left_shift);
 
       TF_LITE_ENSURE(context, err == 0);
-#else                                   // defined(VISION_P6)
+#else                                                     // defined(VISION_P6)
       if (need_broadcast) {
         reference_integer_ops::BroadcastAdd4DSlow(
             op_params, tflite::micro::GetTensorShape(input1),
@@ -168,11 +168,11 @@
             tflite::micro::GetTensorShape(output),
             tflite::micro::GetTensorData<int8_t>(output));
       }
-#endif                                  // defined(VISION_P6)
+#endif                                                    // defined(VISION_P6)
       break;
     }
     case kTfLiteInt16: {
-#if defined(HIFI3) || defined(HIFI4)
+#if defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       int err;
       const RuntimeShape extended_input1_shape =
           RuntimeShape::ExtendedShape(4, tflite::micro::GetTensorShape(input1));
@@ -196,7 +196,7 @@
           op_params.left_shift);
 
       TF_LITE_ENSURE(context, err == 0);
-#else   // defined(HIFI3) || defined(HIFI4)
+#else   // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       if (need_broadcast) {
         reference_ops::BroadcastAdd4DSlow(
             op_params, tflite::micro::GetTensorShape(input1),
@@ -214,7 +214,7 @@
                            tflite::micro::GetTensorData<int16_t>(output),
                            false);
       }
-#endif  // defined(HIFI3) || defined(HIFI4)
+#endif  // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       break;
     }
     default:

tensorflow/lite/micro/kernels/xtensa/sub.cc

diff --git a/tensorflow/lite/micro/kernels/xtensa/sub.cc b/tensorflow/lite/micro/kernels/xtensa/sub.cc
index 3d287d8..898bfea 100644
--- a/tensorflow/lite/micro/kernels/xtensa/sub.cc
+++ b/tensorflow/lite/micro/kernels/xtensa/sub.cc
@@ -83,7 +83,7 @@
                       &op_params);
   // TODO(b/259724572): vision_p6 and hifi code path is getting very confusing.
   // Let's separate them into two different files.
-#if !(defined(HIFI3) || defined(HIFI4))
+#if !(defined(HIFI3) || defined(HIFI4) || defined(HIFI5))
   bool need_broadcast = reference_ops::ProcessBroadcastShapes(
       tflite::micro::GetTensorShape(input1),
       tflite::micro::GetTensorShape(input2), &op_params);
@@ -91,7 +91,7 @@
 
   switch (output->type) {
     case kTfLiteInt8: {
-#if defined(HIFI3) || defined(HIFI4)
+#if defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       int err;
       const RuntimeShape extended_input1_shape =
           RuntimeShape::ExtendedShape(5, tflite::micro::GetTensorShape(input1));
@@ -133,7 +133,7 @@
 
         TF_LITE_ENSURE(context, err == 0);
       }
-#else   // defined(HIFI3) || defined(HIFI4)
+#else   // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       if (need_broadcast) {
         tflite::reference_ops::BroadcastQuantSubSlow(
             op_params, tflite::micro::GetTensorShape(input1),
@@ -151,11 +151,11 @@
             tflite::micro::GetTensorShape(output),
             tflite::micro::GetTensorData<int8_t>(output));
       }
-#endif  // defined(HIFI3) || defined(HIFI4)
+#endif  // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       break;
     }
     case kTfLiteInt16: {
-#if defined(HIFI3) || defined(HIFI4)
+#if defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       int err;
       const RuntimeShape extended_input1_shape =
           RuntimeShape::ExtendedShape(5, tflite::micro::GetTensorShape(input1));
@@ -196,7 +196,7 @@
 
         TF_LITE_ENSURE(context, err == 0);
       }
-#else   // defined(HIFI3) || defined(HIFI4)
+#else   // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       if (need_broadcast) {
         tflite::reference_ops::BroadcastQuantSubSlow(
             op_params, tflite::micro::GetTensorShape(input1),
@@ -214,7 +214,7 @@
             tflite::micro::GetTensorShape(output),
             tflite::micro::GetTensorData<int16_t>(output));
       }
-#endif  // defined(HIFI3) || defined(HIFI4)
+#endif  // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
       break;
     }
     default:
@@ -256,4 +256,4 @@
   return tflite::micro::RegisterOp(SubInit, SubPrepare, SubEval);
 }
 
-}  // namespace tflite
\ No newline at end of file
+}  // namespace tflite