In this presentation, we propose Pelemay Fast Parallel map (PelemayFp), which is a library to parallelize Elixir code, efficiently, based on skeletons for data parallelism. PelemayFp is implemented using only Elixir, like Flow, which is a library of previous works. In Flow, the order of the list after computation is not guaranteed, while in PelemayFp, the order of the list is guaranteed because it is sorting while collecting and merging. On the other hand, Pelemay Super Parallelism (Pelemay), which we proposed, generates native code using SIMD instructions and calls it by NIFs, which is one of FFIs that Erlang provides, without performing multi-core parallelism, guaranteeing the order of the list. We evaluated the integer arithmetic performance by logistic mapping of PelemayFp alone, Pelemay alone, the combination of PelemayFp and Pelemay, Flow, and Enum, which is in the standard library of Elixir. When run on an Intel Xeon W-2191B CPU with 18 cores and 36 threads, the PelemayFp alone is up to 2.1 times faster than Enum. It is also faster than Flow without sorting. On the other hand, the combination of PelemayFp and Pelemay is up to 1.1 times faster than Enum. We also estimated the percentage of parallel execution in the entire code based on Amdahl's law. That of PelemayFp is 48--66 percent, while that of the combination of PelemayFp and Pelemay is 6--46 percent. Further analysis revealed that this experimental results can be explained by assuming that when calling native code from Elixir with NIFs, the part that is not executed in parallel increases by about 30--40 percent. Therefore, when generating native code including SIMD instructions and adopting the approach of parallelizing with Elixir for speeding up, it will not be appropriate to adopt NIFs.