Add missing REQUIREs check to REDUCE_SUM in refmodel

And limit REDUCE_SUM test values to within int32

Signed-off-by: Jeremy Johnson <jeremy.johnson@arm.com>
Change-Id: I4d902b245d17eb343cfb2bbc23d9db28c1d1f4c3

4 files changed, 52 insertions, 2 deletions

diff --git a/reference_model/src/ops/op_factory.cc b/reference_model/src/ops/op_factory.cc
index d42c84a..6edd63f 100644
--- a/reference_model/src/ops/op_factory.cc
+++ b/reference_model/src/ops/op_factory.cc
@@ -273,7 +273,7 @@ GraphNode* OpFactory::newOp(SubgraphTraverser* sgt,
             break;
         case Op_REDUCE_SUM:
             DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FLOAT);
-            DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, INT32);
+            DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSumInt, INT32);
             break;
 
         // data layout
diff --git a/reference_model/src/ops/reduction.cc b/reference_model/src/ops/reduction.cc
index 8c1c4d0..18fac44 100644
--- a/reference_model/src/ops/reduction.cc
+++ b/reference_model/src/ops/reduction.cc
@@ -128,6 +128,31 @@ int OpReduceSum<Rank, Dtype>::eval()
     return GraphNode::eval();
 }
 
+struct SumRequiresReducer {
+    static const bool PacketAccess = false;
+    SumRequiresReducer(SubgraphTraverser* parent_sgt) : parent_sgt(parent_sgt) {}
+    void reduce(const int32_t val, int32_t* accum) {
+        int64_t res_in_64     = static_cast<int64_t>(*accum) + val;
+        int64_t i32_max_in_64 = static_cast<int64_t>(std::numeric_limits<int32_t>::max());
+        int64_t i32_min_in_64 = static_cast<int64_t>(std::numeric_limits<int32_t>::min());
+        REQUIRE(res_in_64 <= i32_max_in_64 && res_in_64 >= i32_min_in_64, "OpReduceSum: result not in i32 range");
+        *accum = static_cast<int32_t>(res_in_64);
+    }
+    int32_t initialize() const { return 0; }
+    int32_t finalize(const int32_t accum) const { return accum; }
+
+    private:
+    SubgraphTraverser* parent_sgt;
+};
+
+template <int Rank, DType Dtype>
+int OpReduceSumInt<Rank, Dtype>::eval()
+{
+    this->out->getTensor() = this->in->getTensor().reduce(this->dims, SumRequiresReducer(this->parent_sgt)).reshape(this->out->getTensor().dimensions());
+
+    return GraphNode::eval();
+}
+
 // template explicit instantiation
 DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceAll, BOOL);
 
@@ -146,4 +171,4 @@ DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT32);
 DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, FLOAT);
 
 DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FLOAT);
-DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, INT32);
+DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSumInt, INT32);
diff --git a/reference_model/src/ops/reduction.h b/reference_model/src/ops/reduction.h
index f4e29b9..f3407f4 100644
--- a/reference_model/src/ops/reduction.h
+++ b/reference_model/src/ops/reduction.h
@@ -104,6 +104,16 @@ public:
     virtual int eval();
 };
 
+template <int Rank, DType Dtype>
+class OpReduceSumInt : public ReduceNode<Rank, Dtype>
+{
+public:
+    OpReduceSumInt(SubgraphTraverser* sgt_, TosaAttributeBase* attribute_, TosaQuantInfoBase* qinfo_, uint64_t id_)
+        : ReduceNode<Rank, Dtype>(sgt_, Op_REDUCE_SUM, attribute_, id_)
+    {}
+    virtual int eval();
+};
+
 };    // namespace TosaReference
 
 #endif
diff --git a/verif/generator/tosa_test_gen.py b/verif/generator/tosa_test_gen.py
index 64f0c5e..80b2981 100644
--- a/verif/generator/tosa_test_gen.py
+++ b/verif/generator/tosa_test_gen.py
@@ -6063,6 +6063,21 @@ class TosaTestGen:
                 self.ser.addPlaceholder(shapeList[1], dtypeList[1], b_arr)
             )
             tens.extend(placeholders)
+        elif op["op"] == Op.REDUCE_SUM and dtypeList[0] == DType.INT32:
+            assert (
+                pCount == 1 and cCount == 0
+            ), "Op.REDUCE_SUM must have 1 placeholders, 0 consts"
+            # Limit values so that the sum cannot exceed the range of an int32 during
+            # summation of any axis
+            range_val = int((1 << 31) / max(shapeList[0]))
+            values_arr = np.int32(
+                self.rng.integers(low=-range_val, high=range_val, size=shapeList[0])
+            )
+            placeholders = []
+            placeholders.append(
+                self.ser.addPlaceholder(shapeList[0], dtypeList[0], values_arr)
+            )
+            tens.extend(placeholders)
         else:
             tens.extend(
                 self.buildPlaceholderTensors(shapeList[0:pCount], dtypeList[0:pCount])