Test-suite fails

[Fuuzetsu]

 sum: [Failed]
*** Failed! Falsifiable (after 6 tests): 
fromList [-2.85685358419018,0.9288729747988657,33.51582034523111,18.079680881481565,-4.872267803281682]
(used seed -931204248)
  product: [Failed]
*** Failed! Falsifiable (after 6 tests): 
fromList [12.798552827341206,-4.694175866482457,-3.27888675478366]
(used seed -330711891)

http://hydra.nixos.org/build/15860082/nixlog/1/raw

The x86_64 box seems to have tested it fine.

[cartazio]

likely https://ghc.haskell.org/trac/ghc/ticket/8195
given that the failure onlyhappens for Storable/Primitive/Unboxed Double vectors

[cartazio]

does the problem go away when you build / test after configure with the ghc option -msse2?

[Fuuzetsu]

I can't test, the log is from NixOS build bot.

[cartazio]

honestly theres 2-3 things going on

1. prior to the most recent releases of vector, the test suite wasn't even in the main cabal file! (so cabal test was a noop before)

2. I'm not sure if those tests should be using exact equality on floating point!

testNumFunctions :: forall a v. (COMMON_CONTEXT(a, v), Num a) => v a -> [Test]
testNumFunctions _ = $(testProperties ['prop_sum, 'prop_product])
where
prop_sum :: P (v a -> a) = V.sum `eq` sum
prop_product :: P (v a -> a) = V.product `eq` product

is the offending line in question

3. I suspect the tests would pass fine if compiled with
   cabal configure --enable-tests --ghc-options="-msse2" ; cabal test
   or the like

[Fuuzetsu]

cc @peti who has Hydra powers

if -msse2 is needed for proper function of vector on i686 then perhaps vector should impose the flag by default?

[cartazio]

the problem is more like the test should use some notation of approximate equality rather than exact,

though forcing -msse2 for the build would also work

[dolio]

-msse2 is needed for deterministic floating point arithmetic on 32-bit x86, period, it seems.

I'm not against sticking -msse2 on the test suite (or what have you) to force the issue, though (if we determine it's the best course of action). I just want to know that it actually fixes this problem before releasing a new package version. :)

[cartazio]

ermmm, we'd only want to enable that for 32bit intel (if at all). I'd slightly favor pushing for some sort of approximate equality test. a la https://hackage.haskell.org/package/ieee754-0.7.3/docs/src/Data-AEq.html#AEq

though because the test suite should be deterministic per se, maybe the -msse2 route is the right solution for nix.

[cartazio]

to repeat/clarify: the only way to get deterministic floating point computations done via an optimizing compiler on x86_32 is to compile all code to use the compiler's equivalent of -msse2. From that perspective, the test suite should certainly be run with -msse2 on applicable systems.

[peti]

Building the package with -msse2 fixes the test suite error on i686-linux: http://hydra.cryp.to/build/203733/log/raw

[Shimuuar]

I think this is one of few cases where we do want floating point equality. We don't test numerical calculations but rather fact that perform same calculation in the same order. So it does make sense to check that two implementation of same algorithm produce same results. (assuming determinitsic FP arithmetics)

AEq type class is in my opinion misguided. There's no such thing as these two floating point numbers are approximately equal. It's always: these numbers are equal in context of given problem. And same pair of numbers could be considered equal or not equal depending on problem. Furthemore there's no single algorithm for comparing two numbers. For example |x-y| / |x| < ε breaks down when x crosses zero, and it's practical problem.

[cartazio]

True, very true. Otoh, I think solution is basically just that on 32bit x86, the tests will fail if they're not built with -msse2. @dolio I think we can close this ticket, thoughts?

[Shimuuar]

Yeah. We did same for the math-functions test suite.

[dolio]

Do we want to ad -msse2 to the test suite so that it can't fail except on systems that don't support it?

Also, is it just the test suite that needs to be built with -msse2? Or do both the tests and the main package need it?

[cartazio]

yeah, i think its just the test suite that needs be built that way. (though maybe folks are using vector on realllly old systems, in which case ONLY the test suite should have that flag? I dont know how old the hardware people might have could be!)