Quad-precision math #4098

dalg24 · 2021-06-16T18:58:26Z

Add support for GCC's 128-bit floating-point type

Co-Authored-By: Irina Tezaur <ikalash@sandia.gov>

mhoemmen · 2021-06-16T20:30:29Z

I did this for Trilinos many years ago (but post adding Kokkos to Tpetra, I think?).

cmake/kokkos_arch.cmake

dalg24 · 2021-06-16T22:54:42Z

.jenkins

@@ -399,6 +399,8 @@ pipeline {
                                -DKokkos_ENABLE_TESTS=ON \
                                -DKokkos_ENABLE_OPENMP=ON \
                                -DKokkos_ENABLE_LIBDL=OFF \
+                                -DKokkos_ENABLE_LIBQUADMATH=ON \
+                                -DCMAKE_PREFIX_PATH=/usr/local/lib/gcc/x86_64-unknown-linux-gnu/5.3.0 \


This is a bit frustrating. This is only needed for CMake, g++ does not need to add anything to the include path to find <quadmath.h>. Also -lquadmath on the link line would do.
Unfortunately I am not sure how I would do that with our CMake files, especially if it has to behave the same with Trilinos.

core/src/impl/Kokkos_QuadPrecisionMath.hpp

dalg24 · 2021-06-16T23:15:26Z

core/src/impl/Kokkos_QuadPrecisionMath.hpp

+inline __float128 hypot(__float128 x, __float128 y) { return ::hypotq(x, y); }
+// Exponential functions
+inline __float128 exp(__float128 x) { return ::expq(x); }
+#if defined(KOKKOS_COMPILER_GNU) && (KOKKOS_COMPILER_GNU >= 910)


Should be checking the stdlibc++ version instead

Actually I am not sure... There does not seem to be any macro to check the libquadmath version.

Guard definition of infinity numeric trait for GCC version >= 7.1

masterleinad

Looks OK to me (maybe apart from the CMAKE_PREFIX_PATH weirdness)

ikalash · 2021-06-23T16:23:13Z

I did this for Trilinos many years ago (but post adding Kokkos to Tpetra, I think?).

Hi @mhoemmen ! hope you are doing well! You may be happy to hear I am reviving all the float128 stuff you put in awhile ago, as well as making extensions for 'long double' scalar types. Unfortunately there have been a number of issues getting the software stack to cooperate, sigh...

ikalash · 2021-06-23T16:52:05Z

Thanks for doing this @dalg24 ! Would Kokkos automatically detect if quadmath is enabled in Trilnos when being built as a part of Trilinos and set the configure flags appropriately?

dalg24 · 2021-06-23T16:56:48Z

Yes

kokkos/cmake/kokkos_tpls.cmake

Lines 70 to 75 in cfb94d9

    
           IF(Trilinos_ENABLE_Kokkos AND TPL_ENABLE_quadmath) 
        
             SET(LIBQUADMATH_DEFAULT ON) 
        
           ELSE() 
        
             SET(LIBQUADMATH_DEFAULT OFF) 
        
           ENDIF() 
        
           KOKKOS_TPL_OPTION(LIBQUADMATH ${LIBQUADMATH_DEFAULT} TRIBITS quadmath)

ikalash · 2021-06-23T17:03:40Z

Regarding adding the extended scalar types to Kokkos::rand - I understand the choice here. The problem is that Kokkos::rand is used within basic Trilinos routines with the implied scalar type, e.g., https://github.com/trilinos/Trilinos/blob/master/packages/tpetra/core/src/Tpetra_MultiVector_def.hpp#L2345-L2350 . I could try to change this if Trilnos folks are not opposed. What is the best way to do this - do a downcase only in the case of a 'long double' or '__float128'?

dalg24 · 2021-06-23T17:05:51Z

Use Kokkos::Experimental::finite_max<T>::value instead

ikalash · 2021-06-23T17:18:01Z

Use Kokkos::Experimental::finite_max<T>::value instead

This is equivalent to Kokkos::rand?

ikalash · 2021-06-23T17:18:55Z

Ah ok, never mind, I see what you are suggesting. Sorry, was looking in the wrong place. I'll try that.

ikalash · 2021-06-23T17:28:00Z

So I get the following error when I switch to Kokkos::Experimental::finite_max<T>::value:

/nightlyCDash/repos/Trilinos/packages/tpetra/core/src/Tpetra_MultiVector_def.hpp:2350:60: error: ‘value’ is not a member of ‘Kokkos::Experimental::finite_max<Kokkos::complex<double> >’
 2350 |     const IST max = Kokkos::Experimental::finite_max<IST>::value;
      |                                                            ^~~~~

I guess there are not complex implementations of finite_max?

dalg24 · 2021-06-23T17:44:10Z

No but you can refer to the discussion on #4033 and do something like 9a78864 where you'd return whatever you think is relevant for a complex number

mhoemmen · 2021-06-23T20:04:17Z

@ikalash wrote:

hope you are doing well!

I am, thank you! Hope you are too! : - )

Unfortunately there have been a number of issues getting the software stack to cooperate, sigh...

Par for the course ; - )

You may be happy to hear I am reviving all the float128 stuff you put in awhile ago, as well as making extensions for long double scalar types.

Excellent! I do hope it works out! These days I generally avoid long double, because on Windows it's the same as double (alas). It could be interesting to work on a more portable solution, that defines a type alias promising 128-bit floating-point arithmetic. This would align with the P1467r4 C++ Standard Committee proposal.

ikalash · 2021-06-23T20:41:21Z

Excellent! I do hope it works out! These days I generally avoid long double, because on Windows it's the same as double (alas). It could be interesting to work on a more portable solution, that defines a type alias promising 128-bit floating-point arithmetic. This would align with the P1467r4 C++ Standard Committee proposal.

Who uses Windows for scientific computing though?? (I know people do, I just always find it hard to believe). I have been lamenting that you are not around to discuss what I've been doing! It seems like there is a fair amount of interest in the extended scalar types within Trilinos, so it will be interesting to see if it goes anyway.

mhoemmen · 2021-06-24T16:42:16Z

@ikalash wrote:

Who uses Windows for scientific computing though??

Ha! ; - P Y'all need to remember that a lot of computation happens on the desktop. Visual Studio's C++ support really took off in the past few years -- these days, we're using a lot of C++20 even without requiring the latest compiler version.

I've seen long double have 64 bits, 80 bits, or 128 bits, depending on the hardware and compiler. It's not very nice if you're trying to write portable code. One possible portable fall-back would be QD. Trilinos had QD support at one time, though it wasn't kept up. These days, it might be better to look into ReproBLAS techniques. @jjellio knows more and has done GPU performance studies.

I have been lamenting that you are not around to discuss what I've been doing!

Likewise! : - ) (I'd be happy to talk more offline about contract work through my employer if you're interested; if not, no worries!)

It seems like there is a fair amount of interest in the extended scalar types within Trilinos, so it will be interesting to see if it goes anyway.

Best wishes with that work!

dalg24 and others added 4 commits June 14, 2021 16:12

Add libquadmath TPL to link against the GCC Quad-Precision Math Library

eaea4f7

Specialize numeric traits for __float128

9cf6485

Specialize reduction_identity for __float128

fb79bb6

Co-Authored-By: Irina Tezaur <ikalash@sandia.gov>

Add __float128 overloads to the common math functions

aaf25bd

crtrott reviewed Jun 16, 2021

View reviewed changes

cmake/kokkos_arch.cmake Outdated Show resolved Hide resolved

dalg24 commented Jun 16, 2021

View reviewed changes

core/src/impl/Kokkos_QuadPrecisionMath.hpp Outdated Show resolved Hide resolved

dalg24 commented Jun 16, 2021

View reviewed changes

dalg24 added 7 commits June 17, 2021 19:48

Guard __float128 stuff

e9d8e65

Drop -pedantic when using __float128

9db441e

Add unit test using __float128

ebeafe9

Enable libquadmath in GCC CI build on Jenkins

4c8b847

Workround GCC 5.3 use of qualified names in specialization

1b5a7f1

Workaround GCC bug

27c86eb

Guard definition of infinity numeric trait for GCC version >= 7.1

Workaround exp2q missing in older libstdc++

0728977

dalg24 force-pushed the libquadmath branch from 731886d to 0728977 Compare June 17, 2021 23:48

dalg24 marked this pull request as ready for review June 18, 2021 14:35

masterleinad approved these changes Jun 18, 2021

View reviewed changes

Rombur approved these changes Jun 22, 2021

View reviewed changes

crtrott approved these changes Jun 22, 2021

View reviewed changes

crtrott merged commit 905b426 into kokkos:develop Jun 22, 2021

dalg24 mentioned this pull request Jun 23, 2021

Support for extended floating-point types #4121

Open

dalg24 deleted the libquadmath branch June 23, 2021 16:57

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Quad-precision math #4098

Quad-precision math #4098

dalg24 commented Jun 16, 2021

mhoemmen commented Jun 16, 2021

dalg24 Jun 16, 2021 •

edited

Loading

dalg24 Jun 16, 2021

dalg24 Jun 16, 2021

masterleinad left a comment

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021 •

edited

Loading

mhoemmen commented Jun 23, 2021 •

edited

Loading

ikalash commented Jun 23, 2021

mhoemmen commented Jun 24, 2021

Quad-precision math #4098

Quad-precision math #4098

Conversation

dalg24 commented Jun 16, 2021

mhoemmen commented Jun 16, 2021

dalg24 Jun 16, 2021 • edited Loading

Choose a reason for hiding this comment

dalg24 Jun 16, 2021

Choose a reason for hiding this comment

dalg24 Jun 16, 2021

Choose a reason for hiding this comment

masterleinad left a comment

Choose a reason for hiding this comment

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

ikalash commented Jun 23, 2021

dalg24 commented Jun 23, 2021 • edited Loading

mhoemmen commented Jun 23, 2021 • edited Loading

ikalash commented Jun 23, 2021

mhoemmen commented Jun 24, 2021

dalg24 Jun 16, 2021 •

edited

Loading

dalg24 commented Jun 23, 2021 •

edited

Loading

mhoemmen commented Jun 23, 2021 •

edited

Loading