Optimize Jacobi eigensolver (WIP) (#847)

* use Amat, Jmat
* change order of Amat, Jmat
* snapshot before changing args to offd_rotate
* use Jsh shared memory
* development snapshot
* debug snapshot
* option to use Jsh in offd_rotate
* use lmemsizeOR
* change launch config for offd_rotate
* snapshot use Jsh for offd_kernel
* generalized offd_kernel
* update diag_rotate
* minor update to comments
* update gridOR, threadsOR for offd_rotate
* add option to use original offd_rotate()
* use original algorithm for small problems
* update idx2D for int32_t, add commonly used ceil function
* use std::min, std::max, remove lambda idx2D
* remove host idx2D
* update with improvement to syevj
* remove lambda for ceil()
* add __device__ for ceil()

CHANGELOG.md

@@ -17,6 +17,9 @@ Full documentation for rocSOLVER is available at the [rocSOLVER documentation](h
 ### Changed
 ### Removed
 ### Optimized
+
+* Improved the performance of SYEVJ
+
 ### Resolved issues
 ### Known issues
 ### Upcoming changes

library/src/include/lib_host_helpers.hpp

@@ -54,6 +54,16 @@ __device__ __host__ inline int64_t
     return j * lda + i * inca;
 }
 
+__device__ __host__ inline int32_t ceil(const int32_t n, const int32_t nb)
+{
+    return ((n - 1) / nb) + 1;
+}
+
+__device__ __host__ inline int64_t ceil(const int64_t n, const int64_t nb)
+{
+    return ((n - 1) / nb) + 1;
+}
+
 template <typename T>
 T const* cast2constType(T* array)
 {