updated SIMD discussion

README.md

@@ -173,7 +173,7 @@ Doppler broadening of resonance data requires evaluation of the complex error fu
 
 ## Optimized Kernels
 
-If using the event-based model, we will be executing the lookup kernel in RSBench across all particles at once. While SIMD execution is possible using this method, typically issues can arrise that greatly reduce SIMD efficiency. In particular, different materials in the simulation have very different numbers of nuclides in them. For instance, spent fuel has 300+ nuclides, while moderator regions only have 10 or so nuclides. This creates a significant load imbalance across lanes in a SIMD vector, as some particles may only need a few iterations to complete all nuclides while others would need hundreds. Furthermore, due to the random memory accesses involved in the binary search process, efficient SIMD execution of the event-based model is not possible without some optimizations.
+If using the event-based model, we will be executing the lookup kernel in RSBench across all particles at once. While SIMD execution is possible using this method, typically issues can arrise that greatly reduce SIMD efficiency. In particular, different materials in the simulation have very different numbers of nuclides in them. For instance, spent fuel has 300+ nuclides, while moderator regions only have 10 or so nuclides. This creates a significant load imbalance across lanes in a SIMD vector, as some particles may only need a few iterations to complete all nuclides while others would need hundreds. Therefore, efficient SIMD execution of the event-based model is not possible without some optimizations.
 
 One promising optimization for the event-based model is to perform a key-value sort of particles: first by material, and then by energy within each material. The first sort by material allows for adjacent particles in the vector to typically reside in the same type of material -- meaning that they will require the same number of nuclide lookup iterations. Then, the energy sort means that adjacent particles in the vector will be located close in energy space -- potentially allowing for many adjacent particles to access the same energy indices in each nuclide and therefore perform many or all of the same branching operations and read the same cache lines into memory at the same time. Once sorted, separate event kernels are then called for each material in the simulation. These two sorts can potentially boost both SIMD efficieny and cache efficiency, with effects being amplified as more particles are simulated at each event stage. The downside to this optimization is the introduction of the key-value particle sorting operations, which can be costly and potentially outweight any gains due to improved SIMD efficiency and cache performance.