CUDA part fails (Typeinfo for ... undefined in device code)

[tdd11235813]

The error message is:

/usr/include/boost/mpl/for_each.hpp(78): error: identifier "Typeinfo for ::mpl_::int_<(int)1> " is undefined in device code

Tested with: gcc-6+ and Boost 1.65.1, 1.66.0, 1.69.0 and CUDA 9.0+ (which requires Boost 1.65.1+).

Older versions (gcc, boost, cuda) produced different errors, but I would prefer to work on with CUDA 9+ and Boost 1.65.1+.

The line in

# test/testWrappers.cpp
    using SliceViewArgs = boost::mpl::list<int_<0>, int_<1>, int_<2>>;                                                                                                                                                                                        
# if you remove the last two arguments it will work

[Flamefire]

Can you try to use std::integral_constant<int, x> instead of the int_?

[tdd11235813]

I already tried, but got the same error. I first need to understand where exactly the device information is missing, so Typeinfo (from `<typeinfo>`?) fails when called in device code.

…

On 10.04.19 18:15, Alexander Grund wrote: Can you try to use |std::integral_constant<int, x>| instead of the |int_|? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#31 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AP01KSpHYkr_4VOJWuwMLHt73zN25KeYks5vfg4WgaJpZM4cnqYO>.

[Flamefire]

I investigated a bit into where this usage of RTTI is coming from. As expected it is from Boost.Test. The macro used to instantiate the test case template expands like this:

• template_test_case_gen<TestCaseTemplate,TestTypesList>
  • Constructor calls generate_test_case_4_type<template_test_case_gen,TestCaseTemplate> via mpl::for_eachon each type inTestTypesList`
• generate_test_case_4_type constructs a test name based on the function name and the type used
• here typeid(TestType).name(); is used

All of this is done on the host (it is a test after all). However the boost.MPL functions (starting at for_each are decorated with BOOST_MPL_CFG_GPU_ENABLED which expands to __host__ __device__. It seems that NVCC now also tries to compile function templates for device code that are called from inside such a function even when those are not marked with __device__, probably to support STL.

I have no idea why it would work when only 1 type is used in the list but not for all types. I'd expect it to work either for all or for none.

I would suggest to report this as a bug to NVIDIA after all this can (probably) be easily reproduced using any boost test case template.

To workaround this try to define BOOST_NO_RTTI which makes the function less pretty but should work otherwise.

[tdd11235813]

Thanks for the quick response!
Just the test result:
I just added BOOST_NO_RTTI definition on top before any boost unit test header is loaded.
While the errors from above vanishes, another error pops up:

/opt/spack-modules/linux-ubuntu16.04-x86_64/gcc-5.4.0/boost-1.65.1-5ovw2kdz2fh44at4z3bb2d7n5n3bs4u7/include/boost/mpl/for_each.hpp(81): error: cannot use an entity undefined in device code

/usr/include/c++/5/bits/basic_string.h(271): error: cannot use an entity undefined in device code

Again, when reducing from three to one type in the list, it works.
(Boost 1.65.1, 1.68.0, CUDA 9.2.88, 10.0.130)

[Flamefire]

Interesting. I'm really unsure WHY this fails for multiple types. The resulting code should look something like this:

void foreach<int_0, int_1, int_2>(){
  run<int_0>();
  foreach<int_1, int2>();
}
void foreach<int_1, int_2>(){
  run<int_1>();
  foreach<int2>();
}
void foreach< int_2>(){
  run<int_2>();
}

All of it is generated at compile time so if it works for one, it should work for all... Again: I'd report this to nvidia.

Another solution would be to split this test case into a template function and manually invoking it for all params like the example here: https://www.boost.org/doc/libs/1_62_0/libs/test/doc/html/boost_test/tests_organization/test_cases/test_organization_templates.html#boost_test.tests_organization.test_cases.test_organization_templates.test_case_template_with_manual_r

[tdd11235813]

Ok, yes it does look strange. I also use mpl list and boost::mpl::for_each in gearshifft directly and not via boost test case template interface. Have to play around with that. Report with MWE to Nvidia is also on TODO.

…

On 11.04.19 16:09, Alexander Grund wrote: Interesting. I'm really unsure WHY this fails for multiple types. The resulting code should look something like this: ``` void foreach<int_0, int_1, int_2>(){ run<int_0>(); foreach<int_1, int2>(); } void foreach<int_1, int_2>(){ run<int_1>(); foreach<int2>(); } void foreach< int_2>(){ run<int_2>(); } ``` All of it is generated at compile time so if it works for one, it should work for all... Again: I'd report this to nvidia. Another solution would be to split this test case into a template function and manually invoking it for all params like the example here: https://www.boost.org/doc/libs/1_62_0/libs/test/doc/html/boost_test/tests_organization/test_cases/test_organization_templates.html#boost_test.tests_organization.test_cases.test_organization_templates.test_case_template_with_manual_r

[tdd11235813]

status:

• workaround is to use boost for_each directly and a templated function object.
• Could not reproduce the issue in an isolated setup, must be postponed now due to other priorities