Inline stuff

[riclarsson]

This moves a lot of cpp-code to headers. It also introduces nonstd::pow.

The former will likely make some compile-times a lot slower. The latter expresses std::pow(v, x) as std::exp(x * std::log(v)), since logarithms and natural exponentials are built-in commands. All tests complete without issues and the numerical error in the cases I tested was actually 0, so I guess there is simply some logic involved and we end up calling that expression regardless.

I ran some tests on all this. In the lbl_temperature_X tests, 10 million calls to the respective temperature models were done. The input is generated by RNG. There are four scenarios. The keyword "Original" means ARTS3 inlining today and "Inline" means the code is in the header. The other change, "using exp() and log() instead of pow()" means as above. Note that some methods below were already inlined as they are constexpr-able.

The short of all this is that there is a 10x speed-up in the most commonly used line-shape function. This is the function call that I often refers to as "as slow as 1000 frequency points", so hopefully, the linear growth in time complexity with frequency will now start at 100 frequency points rather than 1000.

Original:

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.94 ms	6.94 ms	6.94 ms	1
lbl_temperature_dpl	170.89 ms	170.89 ms	170.89 ms	1
lbl_temperature_poly	24.00 ms	24.00 ms	24.00 ms	1
lbl_temperature_t0	6.59 ms	6.59 ms	6.59 ms	1
lbl_temperature_t1	152.56 ms	152.56 ms	152.56 ms	1
lbl_temperature_t2	157.69 ms	157.69 ms	157.69 ms	1
lbl_temperature_t3	6.96 ms	6.96 ms	6.96 ms	1
lbl_temperature_t4	155.86 ms	155.86 ms	155.86 ms	1
lbl_temperature_t5	155.66 ms	155.66 ms	155.66 ms	1

Original using exp() and log() instead of pow():

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.94 ms	6.94 ms	6.94 ms	1
lbl_temperature_dpl	159.35 ms	159.35 ms	159.35 ms	1
lbl_temperature_poly	24.06 ms	24.06 ms	24.06 ms	1
lbl_temperature_t0	6.51 ms	6.51 ms	6.51 ms	1
lbl_temperature_t1	148.66 ms	148.66 ms	148.66 ms	1
lbl_temperature_t2	157.67 ms	157.67 ms	157.67 ms	1
lbl_temperature_t3	6.66 ms	6.66 ms	6.66 ms	1
lbl_temperature_t4	153.97 ms	153.97 ms	153.97 ms	1
lbl_temperature_t5	150.23 ms	150.23 ms	150.23 ms	1

Inline:

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.98 ms	6.98 ms	6.98 ms	1
lbl_temperature_dpl	117.90 ms	117.90 ms	117.90 ms	1
lbl_temperature_poly	14.92 ms	14.92 ms	14.92 ms	1
lbl_temperature_t0	6.57 ms	6.57 ms	6.57 ms	1
lbl_temperature_t1	57.71 ms	57.71 ms	57.71 ms	1
lbl_temperature_t2	61.02 ms	61.02 ms	61.02 ms	1
lbl_temperature_t3	6.97 ms	6.97 ms	6.97 ms	1
lbl_temperature_t4	57.97 ms	57.97 ms	57.97 ms	1
lbl_temperature_t5	58.04 ms	58.04 ms	58.04 ms	1

Inline using exp() and log() instead of pow():

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.94 ms	6.94 ms	6.94 ms	1
lbl_temperature_dpl	33.91 ms	33.91 ms	33.91 ms	1
lbl_temperature_poly	15.17 ms	15.17 ms	15.17 ms	1
lbl_temperature_t0	6.00 ms	6.00 ms	6.00 ms	1
lbl_temperature_t1	16.99 ms	16.99 ms	16.99 ms	1
lbl_temperature_t2	18.97 ms	18.97 ms	18.97 ms	1
lbl_temperature_t3	6.51 ms	6.51 ms	6.51 ms	1
lbl_temperature_t4	17.68 ms	17.68 ms	17.68 ms	1
lbl_temperature_t5	18.23 ms	18.23 ms	18.23 ms	1

The above is all clang so I tested GCC for the original and best scenario.

Original:

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.66 ms	6.66 ms	6.66 ms	1
lbl_temperature_dpl	126.01 ms	126.01 ms	126.01 ms	1
lbl_temperature_poly	24.94 ms	24.94 ms	24.94 ms	1
lbl_temperature_t0	6.70 ms	6.70 ms	6.70 ms	1
lbl_temperature_t1	67.43 ms	67.43 ms	67.43 ms	1
lbl_temperature_t2	94.36 ms	94.36 ms	94.36 ms	1
lbl_temperature_t3	6.73 ms	6.73 ms	6.73 ms	1
lbl_temperature_t4	73.14 ms	73.14 ms	73.14 ms	1
lbl_temperature_t5	67.67 ms	67.67 ms	67.67 ms	1

Inline using exp() and log() instead of pow():

Method	Average Time	Min Time	Max Time	# Calls
lbl_temperature_aer	6.66 ms	6.66 ms	6.66 ms	1
lbl_temperature_dpl	34.83 ms	34.83 ms	34.83 ms	1
lbl_temperature_poly	15.41 ms	15.41 ms	15.41 ms	1
lbl_temperature_t0	6.79 ms	6.79 ms	6.79 ms	1
lbl_temperature_t1	17.54 ms	17.54 ms	17.54 ms	1
lbl_temperature_t2	18.46 ms	18.46 ms	18.46 ms	1
lbl_temperature_t3	6.66 ms	6.66 ms	6.66 ms	1
lbl_temperature_t4	18.07 ms	18.07 ms	18.07 ms	1
lbl_temperature_t5	17.89 ms	17.89 ms	17.89 ms	1

[riclarsson]

@olemke I am concerned about the error in disort.

I was looking through the code to see if there were some errors from this optimization (and there were, so those are now fixed with a spiffed up version of Rational). But logically, this should not be happening by my change.

The only difference between LGPL and normal ARTS behavior for DISORT-CPP is that I rely on Lapack for the matrix eigenvalues instead of my own port of the code. This means there seem to be an update to OpenBLAS causing this issue - they released a new version 5 days ago. How do we deal with this? Should I reduce the test accuracy demand? Or open a bug towards OpenBLAS?

[olemke]

Something has changed in the packaging of libopenblas. Even though libopenblas is installed, cmake decided to link against libblas (the netlib version).

I'll commit a fix that worked locally by changing our dependency from libopenblas to openblas. Seems that openblas installs additional files that makes cmake detect it correctly again.

[riclarsson]

@olemke Thank you for the fix!