Sure enough I did this [here](https://github.com/nim-lang/Nim/pull/16055) (yeah, I know tests are failing at the moment..) and now (with a gcc-10.2 PGO build on Skylake) this simple program runs that benchmark over 3x faster (to within 1.4x of the SIMD optimized C++ (on demand), within 10% of the "D fast" and faster than all the Rust): import parsejson type Coordinate = tuple[x: float, y: float, z: float] proc calc(path = "/tmp/1.json"): Coordinate = var x = 0.0 var y = 0.0 var z = 0.0 var n = 0 for t in jsonTokens(path, false, false, false): if t.tok == tkFloat: # valid float case t.a # last string (strFloats,Ints) of "x": x = x + t.f # a[0] is actually slower! of "y": y = y + t.f of "z": z = z + t.f; n.inc else: discard result = (x: x/float(n), y: y/float(n), z: z/float(n)) echo calc() Run
To be honest, all that SIMD effort is nice and all, but 1.4x isn't _that_ huge a boost. Not duplicating work gets you almost all the way there. Parallelization as @mratsim mentioned could get much more. Unlike xsv/csv, however, json has this fluid hierarchical structure. So, in the general case it is likely a lot of figuring out how to divide work. Maybe not worth it. Truly big data usually has a regular format anyway.
