model3d/docs/performance.md

Model 3D SDK Performance Tests

For the tests I've used the Wuson model from Assimp's repository (Blitz version can be found in this repo's models directory too). For Wavefront OBJ, I've checked Assimp and tinyobjloader. I repeatedly loaded and decoded the models 32 times to mitigate the disk read overhead, and I've also run the test before the perfomance tests, so that the executable and the model file should be already in the Linux disk cache.

For the time measurement, I've used the standard Linux time tool, and took the real time, which includes both time spent in user space and the time spent serving kernel system calls.

NOTE: the uncompressed M3D format, which did not use deflate was actually smaller for this particular model. This mainly depends on how effective can zlib create the Huffman trees from the M3D chunks.

Model	Format	File Size	Library	Required Time
WusonOBJ.obj	text OBJ	258k	tinyobjloader	real 0m0.748s
WusonOBJ.obj	text OBJ	258k	libassimp	real 0m2.716s
WusonBlitz.b3d	binary Blitz	87k	libassimp	real 0m0.414s
WusonBlitz.m3d	uncompressed M3D	33k	libassimp	real 0m6.906s
WusonBlitz.a3d	ASCII Model 3D	141k	M3D SDK	real 0m0.258s
WusonBlitz.m3d	uncompressed M3D	33k	M3D SDK	real 0m0.073s
WusonCompr.m3d	deflated M3D	42k	M3D SDK	real 0m0.162s

Conclusion

The Assimp library was the slowest, except for loading the binary Blitz model in which case it was faster than the tinyobjloader which had to parse text. In all cases the fastest was the native M3D SDK, about 10 times faster than the second place tinyobjloader. Even with zlib deflate, it was still 4 times faster. To make it a fair compartition, comparing only the ASCII variant with tinyobjloader, the M3D SDK is still 2.5 times faster (and it was faster than the binary libassimp reader).