Merge #1742 support properly when disable half precision

This PR supports properly when disable half precision. It keeps the support for those GPU with compute capability less than 5.3 before cuda 12.2

Related PR: #1742

INSTALL.md

@@ -26,7 +26,8 @@ Ginkgo adds the following additional switches to control what is being built:
     Enabling this flag increases the library size, but improves performance of
     mixed-precision kernels.
 *   `-DGINKGO_ENABLE_HALF={ON, OFF}` enable half precision support in Ginkgo, default is `ON`.
-    It is `OFF` when the compiler is MSVC.
+    It is `OFF` when the compiler is MSVC. If compiling is done with the CUDA backend before CUDA 12.2, 
+    we only support half precision after compute capability 5.3. CUDA 12.2+ compilers waive the compute capbility limitation.
 *   `-DGINKGO_BUILD_TESTS={ON, OFF}` builds Ginkgo's tests
     (will download googletest), default is `ON`.
 *   `-DGINKGO_FAST_TESTS={ON, OFF}` reduces the input sizes for a few slow tests

common/cuda_hip/base/math.hpp

@@ -104,6 +104,9 @@ struct is_complex_or_scalar_impl<thrust::complex<T>>
 }  // namespace gko
 
 
+#if GINKGO_ENABLE_HALF
+
+
 GKO_THRUST_NAEMSPACE_PREFIX
 namespace thrust {
 
@@ -186,4 +189,7 @@ __device__ __forceinline__ bool is_finite(const thrust::complex<__half>& value)
 }  // namespace gko
 
 
+#endif  // GINKGO_ENABLE_HALF
+
+
 #endif  // GKO_COMMON_CUDA_HIP_BASE_MATH_HPP_