Fix REDUCE_PROD

Just creating this issue to document ongoing investigation.

```
DIM1
product: FAIL (1.52s)
✗ failed at test/nofib/Data/Array/Accelerate/Test/NoFib/Prelude/Fold.hs:84:3
after 1 test.
shrink path: 1:

━━━ Exception (ErrorCall) ━━━
edgetpu_compiler --show_operations --out_dir=/tmp/dpvanbalen/ /tmp/dpvanbalen/model3866256-1.tflite (exit 1)
ERROR: Didn't find op for builtin opcode 'REDUCE_PROD' version '2'. An older version of this builtin might be supported. Are you using an old TFLite binary with a newer model?

ERROR: Registration failed.

Invalid model: /tmp/dpvanbalen/model3866256-1.tflite
Model could not be parsed
```

Since TF 2.6, with [TF PR #49735](https://github.com/tensorflow/tensorflow/pull/49735), TFLite generates REDUCE_PROD version 2 in certain cases. The closed-source `edgetpu-compiler` (which we suspect is based on TF <2.6) apparently doesn't support REDUCE_PROD version 2. With the following diff on TF 2.10.1, the product test does seem to work:

```
diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc
index 83da3d319cd..8146888b4eb 100644
--- a/tensorflow/lite/kernels/register.cc
+++ b/tensorflow/lite/kernels/register.cc
@@ -224,7 +224,7 @@ BuiltinOpResolver::BuiltinOpResolver() {
/* max_version = */ 2);
AddBuiltin(BuiltinOperator_REDUCE_PROD, Register_REDUCE_PROD(),
/* min_version = */ 1,
- /* max_version = */ 2);
+ /* max_version = */ 1);
AddBuiltin(BuiltinOperator_REDUCE_MAX, Register_REDUCE_MAX(),
/* min_version = */ 1,
/* max_version = */ 3);
diff --git a/tensorflow/lite/kernels/register_ref.cc b/tensorflow/lite/kernels/register_ref.cc
index 8e6329b7caf..22089702d73 100644
--- a/tensorflow/lite/kernels/register_ref.cc
+++ b/tensorflow/lite/kernels/register_ref.cc
@@ -401,7 +401,7 @@ BuiltinRefOpResolver::BuiltinRefOpResolver() {
/* max_version = */ 2);
AddBuiltin(BuiltinOperator_REDUCE_PROD, Register_REDUCE_PROD_REF(),
/* min_version = */ 1,
- /* max_version = */ 2);
+ /* max_version = */ 1);
AddBuiltin(BuiltinOperator_REDUCE_MAX, Register_REDUCE_MAX_REF(),
/* min_version = */ 1,
/* max_version = */ 3);
diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc
index e9b960536da..09a19a167e1 100644
--- a/tensorflow/lite/tools/optimize/operator_property.cc
+++ b/tensorflow/lite/tools/optimize/operator_property.cc
@@ -1096,7 +1096,7 @@ OperatorProperty GetOperatorProperty(OpVariant op_variant) {
case BuiltinOperator_REDUCE_PROD:
property.inputs = {{0, {}}};
property.outputs = {{0, {}}};
- property.version = 2;
+ property.version = 1;
break;
case BuiltinOperator_REDUCE_MAX:
case BuiltinOperator_REDUCE_MIN:
diff --git a/tensorflow/lite/tools/versioning/op_version.cc b/tensorflow/lite/tools/versioning/op_version.cc
index b11a13354f8..87f40f150e2 100644
--- a/tensorflow/lite/tools/versioning/op_version.cc
+++ b/tensorflow/lite/tools/versioning/op_version.cc
@@ -805,10 +805,10 @@ int GetBuiltinOperatorVersion(const OpSignature& op_sig) {
return 1;

case BuiltinOperator_REDUCE_PROD:
- if (op_sig.inputs.at(0).type == kTfLiteInt8 ||
- op_sig.inputs.at(0).type == kTfLiteInt16) {
- return 2;
- }
+ // if (op_sig.inputs.at(0).type == kTfLiteInt8 ||
+ // op_sig.inputs.at(0).type == kTfLiteInt16) {
+ // return 2;
+ // }
return 1;

// The version one of broadcast to op won't be not supported since the
diff --git a/tensorflow/lite/tools/versioning/runtime_version.cc b/tensorflow/lite/tools/versioning/runtime_version.cc
index d5745c8d7c2..ce9c14b16b5 100644
--- a/tensorflow/lite/tools/versioning/runtime_version.cc
+++ b/tensorflow/lite/tools/versioning/runtime_version.cc
@@ -203,7 +203,7 @@ std::string FindMinimumRuntimeVersionForOp(tflite::BuiltinOperator op_code,
{{BuiltinOperator_REDUCE_MIN, 2}, "1.14.0"},
{{BuiltinOperator_REDUCE_MIN, 3}, "2.5.0"},
{{BuiltinOperator_REDUCE_PROD, 1}, "1.11.0"},
- {{BuiltinOperator_REDUCE_PROD, 2}, "2.6.0"},
+ // {{BuiltinOperator_REDUCE_PROD, 2}, "2.6.0"},
{{BuiltinOperator_REDUCE_ANY, 1}, "1.11.0"},
{{BuiltinOperator_RELU6, 1}, "1.5.0"},
{{BuiltinOperator_RELU6, 2}, "1.14.0"},
```

but just applying that without understanding the implications seems a very bad idea.