Avoid integer overflow in EigenMetaKernel indexing

- The current implementation computes `size + total_threads`, which can
  overflow and cause CUDA_ERROR_ILLEGAL_ADDRESS when size is close to
  the maximum representable value.
- The num_blocks calculation can also overflow due to the implementation
  of divup().
- This patch prevents these overflows and allows the kernel to work
  correctly for the full representable range of tensor sizes.
- Also adds relevant tests.

unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h

@@ -553,11 +553,39 @@ class TensorExecutor<Expression, GpuDevice, Vectorizable, Tiling> {
 };
 
 #if defined(EIGEN_GPUCC)
+// Returns lhs + rhs, saturating to the highest/lowest representable value on
+// overflow/underflow respectively.
+template <typename Index>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index saturate_add(Index lhs, Index rhs) {
+  const Index highest = NumTraits<Index>::highest();
+  const Index lowest = NumTraits<Index>::lowest();
+  if (lhs > 0 && rhs > 0) {
+    return (lhs > highest - rhs) ? highest : lhs + rhs;
+  } else if (lhs < 0 && rhs < 0) {
+    return (lhs < lowest - rhs) ? lowest : lhs + rhs;
+  } else {
+    return lhs + rhs;
+  }
+}
+
+#if !defined(EIGEN_USE_HIP)
+// Specialization for int32 using PTX intrinsic.
+template <>
+__device__ EIGEN_ALWAYS_INLINE int32_t saturate_add<int32_t>(int32_t lhs,
+                                                             int32_t rhs) {
+  // add.sat is only supported for s32.
+  int32_t result;
+  asm("add.sat.s32 %0, %1, %2;" : "=r"(result) : "r"(lhs), "r"(rhs));
+  return result;
+}
+#endif
+
 template <typename Evaluator, typename StorageIndex, bool Vectorizable>
 struct EigenMetaKernelEval {
   static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   void run(Evaluator& eval, StorageIndex firstIdx, StorageIndex lastIdx, StorageIndex step_size) {
-    for (StorageIndex i = firstIdx; i < lastIdx; i += step_size) {
+    for (StorageIndex i = firstIdx; i < lastIdx;
+         i = saturate_add(i, step_size)) {
       eval.evalScalar(i);
     }
   }
@@ -573,10 +601,11 @@ struct EigenMetaKernelEval<Evaluator, StorageIndex, true> {
 
     // Use the vector path
     for (StorageIndex i = firstIdx * PacketSize; i < vectorized_size;
-         i += vectorized_step_size) {
+         i = saturate_add(i, vectorized_step_size)) {
       eval.evalPacket(i);
     }
-    for (StorageIndex i = vectorized_size + firstIdx; i < lastIdx; i += step_size) {
+    for (StorageIndex i = saturate_add(vectorized_size, firstIdx); i < lastIdx;
+         i = saturate_add(i, step_size)) {
       eval.evalScalar(i);
     }
   }
@@ -603,8 +632,11 @@ EIGEN_STRONG_INLINE void TensorExecutor<Expression, GpuDevice, Vectorizable, Til
   if (needs_assign) {
 
     const int block_size = device.maxGpuThreadsPerBlock();
-    const int max_blocks = device.getNumGpuMultiProcessors() *
-                           device.maxGpuThreadsPerMultiProcessor() / block_size;
+    const int max_blocks =
+        numext::mini<int64_t>(device.getNumGpuMultiProcessors() *
+                              device.maxGpuThreadsPerMultiProcessor(),
+                          NumTraits<StorageIndex>::highest()) /
+        block_size;
     const StorageIndex size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0.
     const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1);

unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h

@@ -30,13 +30,15 @@ const T2& choose(Cond<false>, const T1&, const T2& second) {
 template <typename T, typename X, typename Y>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T divup(const X x, const Y y) {
-  return static_cast<T>((x + y - 1) / y);
+  // Note: This form is used because it cannot overflow.
+  return static_cast<T>(x == 0 ? 0 : (x - 1) / y + 1);
 }
 
 template <typename T>
 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
 T divup(const T x, const T y) {
-  return static_cast<T>((x + y - 1) / y);
+  // Note: This form is used because it cannot overflow.
+  return static_cast<T>(x == 0 ? 0 : (x - 1) / y + 1);
 }
 
 template <size_t n> struct max_n_1 {