Staying with Float64, see if the runtime comes way down when you prescale the data using prescale(x) = x * 2.0^64
Guessing your values to be less than 10^15, and assuming the worst case smallest magnitude the base10 exponent of your largest data value is below 70 scale by constant is a good strategy when the largest of the data values is not large On Monday, July 13, 2015 at 12:04:32 PM UTC-4, Yichao Yu wrote: > > On Mon, Jul 13, 2015 at 11:39 AM, Jeffrey Sarnoff > <jeffrey...@gmail.com <javascript:>> wrote: > > Thanks for sharing your view about denormal values. I hope what I said > doesn't seem that I want to get rid of them completely (and if it did > sound like that, I didn't meant it...). I didn't read the more detail > analysis of their impact but I would believe you that they are > important in general. > > For my specific application, I'm doing time propagation on a > wavefunction (that can decay). For my purpose, there are many other > sources of uncertainty and I'm mainly interested in how the majority > of the wavefunction behave. Therefore I don't really care about the > actually value of something smaller than 10^-10 but I do want it to > run fast. Since this is a linear problem, I can also scale the values > by a constant factor to make underflow less of a problem. > > > I have not looked at the specifics of what is going on ... > > Dismissing denormals is particularly dicey when your functional data > flow is > > generating many denormalized values. > > > > Do you what it is causing many values of very small magnitude to occur > as > > you run this? > > > > Is the data holding them explicitly? If so, and you have access to > > preprocess the data, and you are sure that software > > cannot accumulate or reciprocate or exp etc them, clamp those values to > zero > > and then use the data. > > > > Does the code operate as a denormalized value generator? If so, a small > > alteration to the order of operations may help. > > > > > > > > On Monday, July 13, 2015 at 9:45:59 AM UTC-4, Jeffrey Sarnoff wrote: > >> > >> Cleve Moler's discussion is not quite as "contextually invariant" as > are > >> William Kahan's and James Demmel's. > >> In fact "the numerical analysis community" has made an overwhelmingly > >> strong case that, roughly speaking, > >> one is substantively better situated where denormalized floating point > >> values will be used whenever they may > >> arise than being free of those extra cycles at the mercy of an absent > >> smoothness shoving those values to zero. > >> And this holds widely for floating point centered applications or > >> libraries. > >> > >> If the world were remade with each sunrise by fixed bitwidth floating > >> point computations, supporting denormals > >> is to have made house-calls with few numerical vaccines to everyone who > >> will be relying on those computations > >> to inform expectations about non-trivial work with fixed bitwdith > floating > >> point types. It does not wipe out all forms > >> of numerical untowardness, and some will find the vaccinces more > >> prophylatic than others; still, the analogy holds. > >> > >> We vaccinate many babies against measles even though there are some who > >> would never have become exposed > >> to that disease .. and for those who forgot why, not long ago the news > was > >> about a Disney vaction disease nexus > >> and how far it spread -- then California changed its law to make it > more > >> difficult to opt-out of childhood vaccination. > >> Having denormals there when the values they cover arise brings benifit > >> that parallels the good in that law change. > >> The larger social environment gets better by growing stronger and that > >> can happen because somethat that had > >> been bringing weakness (disease or bad consequences from subtile > numbery > >> misadventures) no longer operates. > >> > >> There is another way denormals have been shown to be matter -- the way > >> above ought to help you feel at ease > >> with deciding not to move your work from Float64 to Float32 for the > >> purpose of avoiding values that hover around > >> smaller magnitudes realizable with Float64s. That sounds like a > headache, > >> and you would not have changed > >> the theory in a way that makes things work (or at all). Recasting the > >> approch to solving ot transforming at hand > >> to work with integer values would move the work away from any cost and > >> benefit that accompany denormals. > >> Other that that, thank your favorite floating point microarchitect for > >> giving you greater throughput with denormals > >> than everyone had a few design cycles ago. > >> > >> I would like their presence without measureable cost .. just not enough > to > >> dislike their availability. > >> > >> On Monday, July 13, 2015 at 8:02:13 AM UTC-4, Yichao Yu wrote: > >>> > >>> > As for doing it in julia, I found @simonbyrne's mxcsr.jl[1]. > However, > >>> > I couldn't get it working without #11604[2]. Inline assembly in > >>> > llvmcall is working on LLVM 3.6 though[3], in case it's useful for > >>> > others. > >>> > > >>> > >>> And for future references I find #789, which is not documented > >>> anywhere AFAICT.... (will probably file a doc issue...) > >>> It also supports runtime detection of cpu feature so it should be much > >>> more portable. > >>> > >>> [1] https://github.com/JuliaLang/julia/pull/789 > >>> > >>> > > >>> > [1] https://gist.github.com/simonbyrne/9c1e4704be46b66b1485 > >>> > [2] https://github.com/JuliaLang/julia/pull/11604 > >>> > [3] > >>> > > https://github.com/yuyichao/explore/blob/a47cef8c84ad3f43b18e0fd797dca9debccdd250/julia/array_prop/array_prop.jl#L3 > > >>> > >