Trouble Running CPPPotential in the HPMC Module with GPU

[dchien79]

Hello! I am trying to implement a user-defined potential in HOOMD while accelerating my HPMC simulations with the GPU. I previously installed the HOOMD GPU-enabled binary for CUDA 12.2, and confirmed that HOOMD is picking up the GPU.

I copied the tutorial script in the following webpage (https://hoomd-blue.readthedocs.io/en/v4.5.0/howto/cpppotential.html) to a python script called test.py and obtained the expected output. However, when I changed the line sim = hoomd.Simulation(device=hoomd.device.CPU(), seed=1) to sim = hoomd.Simulation(device=hoomd.device.GPU(), seed=1), I was returned with the following error:

(hoomd2) [dchien@bhg0042 benchmark]$ python test.py
/gpfs/fs2/scratch/dchien/esc-hoomd/hoomd/benchmark/test.py:25: FutureWarning: CPPPotential is deprecated since 4.5.0. Use a hpmc.pair.Pair potential.
square_well = hoomd.hpmc.pair.user.CPPPotential(r_cut=r_interaction,
/gpfs/fs2/scratch/dchien/esc-hoomd/hoomd/benchmark/test.py:34: FutureWarning: pair_potential is deprecated since 4.5.0. Use pair_potentials.
mc.pair_potential = square_well
Traceback (most recent call last):
File "/gpfs/fs2/scratch/dchien/esc-hoomd/hoomd/benchmark/test.py", line 42, in
sim.run(0)
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/simulation.py", line 555, in run
self.operations._schedule()
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/operations.py", line 211, in _schedule
self.integrator._attach(sim)
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/operation.py", line 315, in _attach
self._attach_hook()
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/hpmc/integrate.py", line 457, in _attach_hook
self._pair_potential._attach(self._simulation)
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/operation.py", line 315, in _attach
self._attach_hook()
File "/software/miniconda3/4.12.0/envs/hoomd2/lib/python3.12/site-packages/hoomd/hpmc/pair/user.py", line 276, in _attach_hook
self._cpp_obj = _jit.PatchEnergyJITGPU(
^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Runtime compilation failed:
__cuda/cstdint_prelude.h(80): warning #47-D: incompatible redefinition of macro "UINT64_C" (declared at line 124 of nvccfeatures.h)
#define UINT64_C(X) ((uint_least64_t)(X))
^

Remark: The warnings can be suppressed with "-diag-suppress "

../limits(218): error: identifier "CHAR_BIT" is undefined
static _LIBCUDACXX_CONSTEXPR const int digits = static_cast(sizeof(type) * CHAR_BIT - is_signed);
^

../limits(319): error: identifier "FLT_MANT_DIG" is undefined
static _LIBCUDACXX_CONSTEXPR const int digits = FLT_MANT_DIG;
^

../limits(320): error: identifier "FLT_DIG" is undefined
static _LIBCUDACXX_CONSTEXPR const int digits10 = FLT_DIG;
^

../limits(328): error: identifier "FLT_RADIX" is undefined
static _LIBCUDACXX_CONSTEXPR const int radix = FLT_RADIX;
^

../limits(332): error: identifier "FLT_MIN_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int min_exponent = FLT_MIN_EXP;
^

../limits(333): error: identifier "FLT_MIN_10_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int min_exponent10 = FLT_MIN_10_EXP;
^

../limits(334): error: identifier "FLT_MAX_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int max_exponent = FLT_MAX_EXP;
^

../limits(335): error: identifier "FLT_MAX_10_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int max_exponent10 = FLT_MAX_10_EXP;
^

../limits(322): error: identifier "FLT_MIN" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type min() _NOEXCEPT {return FLT_MIN;}
^

../limits(323): error: identifier "FLT_MAX" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type max() _NOEXCEPT {return FLT_MAX;}
^

../limits(329): error: identifier "FLT_EPSILON" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type epsilon() _NOEXCEPT {return FLT_EPSILON;}
^

../limits(343): error: identifier "__builtin_huge_valf" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type infinity() _NOEXCEPT {return __builtin_huge_valf();}
^

../limits(344): error: identifier "__builtin_nanf" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type quiet_NaN() _NOEXCEPT {return __builtin_nanf("");}
^

../limits(345): error: identifier "__builtin_nansf" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type signaling_NaN() _NOEXCEPT {return __builtin_nansf("");}
^

../limits(351): error: identifier "FLT_DENORM_MIN" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type denorm_min() _NOEXCEPT {return FLT_DENORM_MIN;}
^

../limits(371): error: identifier "DBL_MANT_DIG" is undefined
static _LIBCUDACXX_CONSTEXPR const int digits = DBL_MANT_DIG;
^

../limits(372): error: identifier "DBL_DIG" is undefined
static _LIBCUDACXX_CONSTEXPR const int digits10 = DBL_DIG;
^

../limits(380): error: identifier "FLT_RADIX" is undefined
static _LIBCUDACXX_CONSTEXPR const int radix = FLT_RADIX;
^

../limits(384): error: identifier "DBL_MIN_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int min_exponent = DBL_MIN_EXP;
^

../limits(385): error: identifier "DBL_MIN_10_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int min_exponent10 = DBL_MIN_10_EXP;
^

../limits(386): error: identifier "DBL_MAX_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int max_exponent = DBL_MAX_EXP;
^

../limits(387): error: identifier "DBL_MAX_10_EXP" is undefined
static _LIBCUDACXX_CONSTEXPR const int max_exponent10 = DBL_MAX_10_EXP;
^

../limits(374): error: identifier "DBL_MIN" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type min() _NOEXCEPT {return DBL_MIN;}
^

../limits(375): error: identifier "DBL_MAX" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type max() _NOEXCEPT {return DBL_MAX;}
^

../limits(381): error: identifier "DBL_EPSILON" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type epsilon() _NOEXCEPT {return DBL_EPSILON;}
^

../limits(395): error: identifier "__builtin_huge_val" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type infinity() _NOEXCEPT {return __builtin_huge_val();}
^

../limits(396): error: identifier "__builtin_nan" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type quiet_NaN() _NOEXCEPT {return __builtin_nan("");}
^

../limits(397): error: identifier "__builtin_nans" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static type signaling_NaN() _NOEXCEPT {return __builtin_nans("");}
^

../limits(403): error: identifier "DBL_DENORM_MIN" is undefined
_LIBCUDACXX_INLINE_VISIBILITY static _LIBCUDACXX_CONSTEXPR type denorm_min() _NOEXCEPT {return DBL_DENORM_MIN;}
^

29 errors detected in the compilation of "default_program".

I am wondering, then, if the CPPPotential function is supported when using the GPU as a device. Is there a way to fix this issue or is there a different function in HOOMD that can run user-defined potentials while still using the GPU? Let me know if so. Thank you!

[joaander]

In the newly released HOOMD-blue 4.6.0, you can use hpmc.pair.Step and Union potentials without needing CPPPotential. The new potentials do not support the GPU, but that is for the best as the GPU implementation of CPPPotential was almost always much less efficient than the CPU. Furthermore, hpmc.pair.Step is much faster than an equivalent implementation with CPPPotential on the CPU.

I have not previously seen the specific error you report, but I have seen many like it. Fixing these usually requires many days of effort to replicate the error (matching compiler and CUDA environments) and then finding the right tweak to make to jitify.hpp.

As you can see in your output, CPPPotential is deprecated. I have no plans to further maintain it. In HOOMD-blue 4.6, hpmc.pair (excluding pair.user) covers almost all the cases I have seen people using CPPPotential for. In later 4.x releases, I plan to replace the external compiled potentials, provide a template component to help migrate custom CPPPotential cases, and update the tutorial. 5.0 will remove all CPPPotential classes.

[dchien79]

Okay, thank you for the assistance!