Add AniSOAP PyTorch speedup benchmarks and documentation

benchmarks/ospo_pytorch_speedups/combined_from_metrics.csv

@@ -0,0 +1,11 @@
+file,mode,wall_s,N,species
+ellipsoids,torch,0.21896767616271973,,
+one_species,numpy,0.49042439460754395,,1.0
+benzenes,torch,0.1308887004852295,,2.0
+three_species,numpy,0.39318180084228516,,3.0
+four_species,torch,0.22287511825561523,,4.0
+three_species,torch,0.18306541442871094,,3.0
+benzenes,numpy,0.24216747283935547,,2.0
+ellipsoids,numpy,0.21300196647644043,,
+one_species,torch,0.13603806495666504,,1.0
+four_species,numpy,0.2523961067199707,,4.0

benchmarks/ospo_pytorch_speedups/timings.csv

@@ -0,0 +1,10 @@
+benzenes_numpy,benzenes.xyz,numpy,0.242
+benzenes_torch,benzenes.xyz,torch,0.131
+ellipsoids_numpy,ellipsoids.xyz,numpy,0.213
+ellipsoids_torch,ellipsoids.xyz,torch,0.219
+four_species_numpy,four_species.xyz,numpy,0.252
+four_species_torch,four_species.xyz,torch,0.223
+one_species_numpy,one_species.xyz,numpy,0.490
+one_species_torch,one_species.xyz,torch,0.136
+three_species_numpy,three_species.xyz,numpy,0.393
+three_species_torch,three_species.xyz,torch,0.183

docs/performance_pytorch_speedups.md

@@ -0,0 +1,20 @@
+# AniSOAP PyTorch performance summary
+
+This document summarizes timing results for the PyTorch backend compared to the original NumPy implementation. The full raw metrics and figures are stored in `benchmarks/ospo_pytorch_speedups/`.
+
+Benchmarks cover:
+- Benzene systems with varying numbers of molecules.
+- Ellipsoid systems.
+- Multi-species systems (one, three, and four species).
+
+For each system, we compare:
+- NumPy (CPU)
+- PyTorch (CPU)
+- PyTorch (Apple MPS, where available)
+
+The associated CSV files are:
+
+- `timings.csv`: wall-clock timings for each configuration.
+- `combined_from_metrics.csv`: combined table of metrics for NumPy and PyTorch runs.
+
+These results are the same as those reported in my OSPO presentation and Zenodo record on AniSOAP performance optimization.