Rustify N3LO Ad

[giacomomagni]

No description provided.

[felixhekhorn]

Also here is another suggestion for the harmonics module: all functions called "S + positive integer" take as single argument $N$, all others take as single argument the cache - i.e. no messing with passing the correct arguments. (Since we are now in a compiled language I'm no longer worried of nesting things) we could even shift the "raw" S-es into the polygamma module ...

what do you think @giacomomagni @tgiani ?

[giacomomagni]

Also here is another suggestion for the harmonics module: all functions called "S + positive integer" take as single argument N , all others take as single argument the cache - i.e. no messing with passing the correct arguments. (Since we are now in a compiled language I'm no longer worried of nesting things) we could even shift the "raw" S-es into the polygamma module ...

what do you think @giacomomagni @tgiani ?

Sure this require a bit of restructure though...

[felixhekhorn]

Sure this require a bit of restructure though...

I made an attempt in 462f62c - so, do we want to do this? I can take care of the copy&paste if you wish ;-)

I also made some adjustments to make pre-commit happy

[felixhekhorn]

I'm also starting to think, that the log expressions are so complicated (look at lm14m2), that we should cache them - what do you think?

[giacomomagni]

I'm also starting to think, that the log expressions are so complicated (look at lm14m2), that we should cache them - what do you think?

Some of them might not be needed, because the Moch et al. expressions are simpler...
So I wanted to finish the conversion and then polish the harmonic file.

[felixhekhorn]

I made an attempt in 462f62c - so, do we want to do this?

I forgot to mention: the prize that we had to pay for this was

eko/crates/ekore/src/harmonics/w3.rs

Lines 36 to 37 in 462f62c

	let mut cp1 = Cache::new(c.n() + 1.);
	ZETA2 * hSm1 - 5. / 8. * ZETA3 + ZETA2 * LN_2 - g3(&mut cp1)

i.e. we had to create a new cache, which is potentially expensive ... there are in principle two strategies I can imagine to avoid that (if we want to avoid it):

1. math: i.e. maybe we can express $g_3(N+1)-g_3(N)$, like I was playing in the QED PR - but looking at the notes still laying on my desk, I'm not sure we can simplify this
2. more interface: we add a Cache::set method, which allows us to fill an entry into the cache. Then I know $g_3(N+1)$ will need $S_1(N+1)$ and that I can derive from $S_1(N)$ (instead of recalculating)

[felixhekhorn]

Any comment on the cache recreation issue?

Else: is this ready? if yes, let's mark it, thus make the benchmarks run and the we can quickly merge

[giacomomagni]

Else: is this ready? if yes, let's mark it, thus make the benchmarks run and the we can quickly merge

Not yet! I need to add the variation to the kernel. I'll do it asap.

[giacomomagni]

Any comment on the cache recreation issue?

Maybe all these optimisation can be addressed once the full ekore is in place ?
I guess that for the moment having a new cache in these function is fine as it will be mostly empty (filled only with g3 and related harmonics). But probably the math solution is also doable...

[felixhekhorn]

• if only we had #440 🙈
• do you have by chance a time (ratio) estimate?
• on second thought: should we also make w1::S1 take only the cache as argument?

besides these small issues the PR is good to go

[giacomomagni]

Just to have a quick feeling on the spped-up.
Running an eko with an x-grid of 30 points and 2 Q2s, in VFNS I get:

• Python (numba already compiled):

• order: (0,0), matching (0,0): 71.05s user 2.86s system 103% cpu 1:11.70 total
• order: (4,0), matching (2,0): 147.30s user 4.52s system 100% cpu 2:30.35 total

• Rust :

• order: (0,0), matching (0,0): 10.76s user 1.59s system 125% cpu 9.809 total
• order: (4,0), matching (2,0): 18.48s user 1.63s system 114% cpu 17.566 total

So the speed up ratio is more or less the same at both orders.

[felixhekhorn]

in VFNS I get

FFNS I assume? other it would be not fair

[giacomomagni]

FFNS I assume? other it would be not fair

VFNS, recall now we're able to decouple the 2 ptos!