GPU: Raise CUDA/HIP minimum and remove legacy guards

- Raise CUDA minimum from 9.0 to 11.4 (sm_70/Volta).
- Raise HIP minimum to GFX906 (Vega 20/MI50) / ROCm 5.6.
- Remove EIGEN_HAS_{CUDA,HIP,GPU}_FP16 guards — FP16 is always available
  on sm_70+ and GFX906+.
- Remove obsolete __HIP_ARCH_HAS_* preprocessor branches.
- C++14 cleanup: remove pre-C++14 workarounds in GPU code.
- Fix NVCC warnings (deprecated register keyword, unreachable code,
  tautological comparisons).
- Fix HIP test execution on gfx1151.
- Update CI configuration for new minimum versions.

test/gpu_test_helper.h

@@ -6,10 +6,8 @@
 // Allow gpu** macros for generic tests.
 #include <Eigen/src/Core/util/GpuHipCudaDefines.inc>
 
-// std::tuple cannot be used on device, and there is a bug in cuda < 9.2 that
-// doesn't allow std::tuple to compile for host code either. In these cases,
-// use our custom implementation.
-#if defined(EIGEN_GPU_COMPILE_PHASE) || (defined(EIGEN_CUDACC) && EIGEN_CUDA_SDK_VER < 92000)
+// std::tuple cannot be used on device, so use our custom implementation there.
+#if defined(EIGEN_GPU_COMPILE_PHASE)
 #define EIGEN_USE_CUSTOM_TUPLE 1
 #else
 #define EIGEN_USE_CUSTOM_TUPLE 0
@@ -42,6 +40,12 @@ using tuple_impl::tuple;
 #undef EIGEN_USE_CUSTOM_TUPLE
 }  // namespace test_detail
 
+template <typename T>
+using decay_t = typename std::decay<T>::type;
+
+template <typename Func, typename... Args>
+using kernel_result_t = decltype(std::declval<Func>()(std::declval<Args>()...));
+
 template <size_t N, size_t Idx, typename OutputIndexSequence, typename... Ts>
 struct extract_output_indices_helper;
 
@@ -90,14 +94,15 @@ struct void_helper {
   // Non-void return value.
   template <typename Func, typename... Args>
   static EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC auto call(Func&& func, Args&&... args)
-      -> std::enable_if_t<!std::is_same<decltype(func(args...)), void>::value, decltype(func(args...))> {
+      -> std::enable_if_t<!std::is_same<kernel_result_t<Func&&, Args&&...>, void>::value,
+                          kernel_result_t<Func&&, Args&&...>> {
     return func(std::forward<Args>(args)...);
   }
 
   // Void return value.
   template <typename Func, typename... Args>
   static EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC auto call(Func&& func, Args&&... args)
-      -> std::enable_if_t<std::is_same<decltype(func(args...)), void>::value, Void> {
+      -> std::enable_if_t<std::is_same<kernel_result_t<Func&&, Args&&...>, void>::value, Void> {
     func(std::forward<Args>(args)...);
     return Void{};
   }
@@ -135,18 +140,18 @@ EIGEN_DEVICE_FUNC void run_serialized(std::index_sequence<Indices...>, std::inde
   const uint8_t* read_end = buffer + capacity;
   read_ptr = Eigen::deserialize(read_ptr, read_end, input_size);
   // Create value-type instances to populate.
-  auto args = make_tuple(typename std::decay<Args>::type{}...);
+  auto args = make_tuple(decay_t<Args>{}...);
   EIGEN_UNUSED_VARIABLE(args);  // Avoid NVCC compile warning.
   // NVCC 9.1 requires us to spell out the template parameters explicitly.
-  read_ptr = Eigen::deserialize(read_ptr, read_end, get<Indices, typename std::decay<Args>::type...>(args)...);
+  read_ptr = Eigen::deserialize(read_ptr, read_end, get<Indices, decay_t<Args>...>(args)...);
 
   // Call function, with void->Void conversion so we are guaranteed a complete
   // output type.
-  auto result = void_helper::call(kernel, get<Indices, typename std::decay<Args>::type...>(args)...);
+  auto result = void_helper::call(kernel, get<Indices, decay_t<Args>...>(args)...);
 
   // Determine required output size.
   size_t output_size = Eigen::serialize_size(capacity);
-  output_size += Eigen::serialize_size(get<OutputIndices, typename std::decay<Args>::type...>(args)...);
+  output_size += Eigen::serialize_size(get<OutputIndices, decay_t<Args>...>(args)...);
   output_size += Eigen::serialize_size(result);
 
   // Always serialize required buffer size.
@@ -157,7 +162,7 @@ EIGEN_DEVICE_FUNC void run_serialized(std::index_sequence<Indices...>, std::inde
   // Serialize outputs if they fit in the buffer.
   if (output_size <= capacity) {
     // Collect outputs and result.
-    write_ptr = Eigen::serialize(write_ptr, write_end, get<OutputIndices, typename std::decay<Args>::type...>(args)...);
+    write_ptr = Eigen::serialize(write_ptr, write_end, get<OutputIndices, decay_t<Args>...>(args)...);
     write_ptr = Eigen::serialize(write_ptr, write_end, result);
   }
 }
@@ -282,7 +287,7 @@ auto run_serialized_on_gpu(size_t buffer_capacity_hint, std::index_sequence<Indi
  * \return kernel(args...).
  */
 template <typename Kernel, typename... Args>
-auto run_on_cpu(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
+auto run_on_cpu(Kernel kernel, Args&&... args) -> internal::kernel_result_t<Kernel, Args&&...> {
   return kernel(std::forward<Args>(args)...);
 }
 
@@ -301,7 +306,7 @@ auto run_on_cpu(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
  * \return kernel(args...).
  */
 template <typename Kernel, typename... Args>
-auto run_on_gpu(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
+auto run_on_gpu(Kernel kernel, Args&&... args) -> internal::kernel_result_t<Kernel, Args&&...> {
   return internal::run_serialized_on_gpu<Kernel, Args...>(
       /*buffer_capacity_hint=*/0, std::make_index_sequence<sizeof...(Args)>{},
       internal::extract_output_indices<Args...>{}, kernel, std::forward<Args>(args)...);
@@ -322,7 +327,8 @@ auto run_on_gpu(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
  * \sa run_on_gpu
  */
 template <typename Kernel, typename... Args>
-auto run_on_gpu_with_hint(size_t buffer_capacity_hint, Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
+auto run_on_gpu_with_hint(size_t buffer_capacity_hint, Kernel kernel, Args&&... args)
+    -> internal::kernel_result_t<Kernel, Args&&...> {
   return internal::run_serialized_on_gpu<Kernel, Args...>(
       buffer_capacity_hint, std::make_index_sequence<sizeof...(Args)>{}, internal::extract_output_indices<Args...>{},
       kernel, std::forward<Args>(args)...);
@@ -409,7 +415,7 @@ void print_gpu_device_info() {
  * \return kernel(args...).
  */
 template <typename Kernel, typename... Args>
-auto run(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
+auto run(Kernel kernel, Args&&... args) -> internal::kernel_result_t<Kernel, Args&&...> {
 #ifdef EIGEN_GPUCC
   return run_on_gpu(kernel, std::forward<Args>(args)...);
 #else
@@ -432,7 +438,8 @@ auto run(Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
  * \sa run
  */
 template <typename Kernel, typename... Args>
-auto run_with_hint(size_t buffer_capacity_hint, Kernel kernel, Args&&... args) -> decltype(kernel(args...)) {
+auto run_with_hint(size_t buffer_capacity_hint, Kernel kernel, Args&&... args)
+    -> internal::kernel_result_t<Kernel, Args&&...> {
 #ifdef EIGEN_GPUCC
   return run_on_gpu_with_hint(buffer_capacity_hint, kernel, std::forward<Args>(args)...);
 #else