Performance issue in use of boost::math::digamma on aarch64 Linux

[dslarm]

Due to a current default in the boost library (https://github.com/boostorg/math/issues/1211) in boost::math::digamma, there is a performance hit on aarch64.

This happens on v1.10.3 of Salmon, with GNU compiler 13 on Linux aarch64.

A 4-thread quantization of one of the Salmon tutorials DRR0* series files spends ~15% of time in this routine (called within CollapsedEMOptimizer). On a larger example, we see 7% performance hit over a run that takes 1300 seconds on 4 cores. On x86 this time is small enough to be lost in the noise.

salmon quant -i athal_index -l A -1 DRR016125/DRR016125_1.fastq.gz -2 DRR016125/DRR016125_2.fastq.gz -p $threads --validateMappings -o quants/DRR016125_quant

There is a simple fix which is to ensure the CMake/Makefiles ensure salmon compiles with: -DBOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS or to add that to any file that brings in boost::math via adding #define BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS at the start.

With that change, a 1300 second runtime drops to 1212 for the larger test case, and for the tutorial case is 48 seconds down to 40 on a 4-core r8g.xlarge (Graviton4).

Whilst Boost may fix the issue soon - it's likely that older versions of the library will be found installed for some time. It would be helpful to add this define to cmake settings, or the sources.

[rdoeffinger]

Note that I think -DBOOST_MATH_PROMOTE_DOUBLE_POLICY=false is a better option than disabling long double functions in general

[dslarm]

I agree – I wasn’t aware of that one. I’ve tested that and it has the same effect as the other flag, performance looks good.