Update elemental kernel to use ThreadVectorLoops

[Rohit-Kakodkar]

Currently we use ThreadThreadRange to distinguish elements inside elemental kernels. Limiting the number of threads per block we can use to 32 for 5 x 5 quadrature in 2D. A better way of doing this would be use ThreadVectorRange to distinguish elements - which would allow to use arbitary number of threads inside a block.

• Theoretical occupancy would increase from 50% -> 100% (not taking into account number of registers)
• Possibly would allow me to break the kernel up to reduce register usage

[Rohit-Kakodkar]

Final Commit for optimization

Summary of findings :

• Compute elemental calcualtions within ThreadVectorRanges

  • Move elemental calculations within ThreadVectorRange
  • This will maximize GPU usage

• Split compute forces routine in 3 subroutines

  • compute local gradients
  • compute stress integrands from local gradients
  • compute acceleration
  • Global memory accesses are not worth it for smaller kernels

• Reduce the number of registers used

  • Precompute indexing variables as much as possible
  • Using a single shared memory allocation and compute subviews from that allocation