Merged PR 4500055: Fix autopilot WinML::Engine::Test::BenchmarkProtobuf#onnxzoo_ssd#1.5#GPU

Model node "Slice-505" specifies a 64-bit INTMAX value (9223372036854775807) to mean essentially unbounded. The proper response in this case is to clamp it to 32-bit INTMAX, not bitwise truncate it, which yields -1.

Related work items: #24672220

onnxruntime/core/providers/dml/OperatorAuthorHelper/Common.h

@@ -20,6 +20,11 @@
     }\
   }
 
+template<typename T, typename I> T clamp_cast(I input)
+{
+    return static_cast<T>(std::clamp<I>(input, std::numeric_limits<T>::lowest(), std::numeric_limits<T>::max()));
+}
+
 namespace OperatorHelper
 {
     enum TensorAxis { N, C, H, W, DoNotCoerce = UINT_MAX, LeftAligned = INT_MAX, RightAligned = INT_MIN, NoPlacementAdjustment = 0 };

onnxruntime/core/providers/dml/OperatorAuthorHelper/OperatorHelper.cpp

@@ -56,9 +56,16 @@ namespace OperatorHelper
             {
                 const int64_t* data = tensor.GetData<int64_t>();
                 result.reserve(elementCount);
+
+                // Use clamped cast rather than static_cast/narrow_cast,
+                // because it's not uncommon for a model to specify a
+                // 64-bit INTMAX constant as a sentinel value to mean
+                // the largest possible value (even though the actual
+                // dimension values come nowhere close to that, far
+                // less than 32-bit INTMAX).
                 for (auto d : gsl::make_span(data, data + elementCount))
                 {
-                    result.push_back(gsl::narrow_cast<int32_t>(d));
+                    result.push_back(clamp_cast<int32_t>(d));
                 }
             }
             break;