Repository navigation
GPU compilation failure with pow.hpp from boost::multiprecision #1285
Description
Activity
I can confirm this was enough to get our code compiling again, but I'm guessing this could be a more widespread issue than just this one instance.
Yes, we need to check for BOOST_NO_EXCEPTIONS as well. I don't think multiprecision is well tested in that mode.
Yes, we need to check for
BOOST_NO_EXCEPTIONSas well. I don't think multiprecision is well tested in that mode.I'm actually in the process of expanding Multiprecision's tests for 1.91 (the release in November). We can discuss testing improvements in a separate, future ticket.
I think this is similar to boostorg/multiprecision#695 which is fixed on develop.
BTW I'm curious as to whether you're using multiprecision on a CUDA device? Interesting if so ;)
Reacted by George T KramerIt seems like
BOOST_MP_CATCHalready checks forBOOST_NO_EXCEPTIONS, which explains why we've got this macro to begin with. I guess only cases outside of the std library are affected here during preprocessing, such as withboost::math::rounding_error, in the multiprocessing module.BTW I'm curious as to whether you're using multiprecision on a CUDA device? Interesting if so ;)
I don't think we are intentionally XD. We have a complex code base where we are using boost/multiprecision/cpp_int.hpp mostly for the CPU, but we extended a part of it to compile with the GPU. This part got caught on the GPU compilation path due to dependencies and happened to expose an edge case it seems.
Looks like this was fixed in multiprecision-feel free to reopen if not.
Reacted by George T Kramer
I have some C++ code that compiled just fine with
nvccfrom CUDA 12.6.1 and Boost 1.84.0. After updating to Boost 1.88.0, which includes the GPU math changes from Boost 1.87.0, I'm now seeing the following error message:If I'm understanding Boost's macro and include header structure correctly, boost/multiprecision/detail/functions/pow.hpp is inlined with boost/multiprecision/detail/default_ops.hpp. pow.hpp is always included no matter what in default_ops.hpp.
boost::math::rounding_erroris defined in boost/math/policies/error_handling.hpp, which is only included by default_ops.hpp ifBOOST_MP_MATH_AVAILABLEis defined. This is the same macro that's used for guardingBOOST_MP_CATCH(const boost::math::rounding_error&) {}in pow.hpp (the same code not compiling), so all seems good. But,boost::math::rounding_erroris only defined whenBOOST_MATH_NO_EXCEPTIONSis NOT defined. And, when compiling withnvcc,BOOST_MATH_NO_EXCEPTIONSDOES get defined, according to boost/math/tools/config.hpp, because we've got__CUDACC__defined by simply usingnvcc.Should we be adjusting pow.hpp, and possibly any other cases like this in either
boost::mathorboost::multiprecision, to have something like this instead?That way we only define this catch condition when we have exception support and multiprecision math is available?