When using HIP as a platform, I encountered the HIP compiler errors below while trying to create a simple leapfrog integrator with a gaussian noise term.
CustomIntegrator Function | vvvvvvvvvvvvvvvv
def myintegrator(temperature,friction,dt): integrator=mm.CustomIntegrator(dt); integrator.addGlobalVariable("a", exp(-friction*dt)); integrator.addGlobalVariable("b", sqrt((1-exp(-2*friction*dt))); integrator.addUpdateContextState(); integrator.addComputePerDof("v", "a*v + b*sqrt(1/m)*gaussian"); integrator.addComputePerDof("x", "x + 0.5 * dt*v"); return integrator
ERRORS | vvvvvvvvvvvvvvvv
File "/usr/local/lib/python3.10/dist-packages/openmm/app/simulation.py", line 147, in step
self._simulate(endStep=self.currentStep+steps)
File "/usr/local/lib/python3.10/dist-packages/openmm/app/simulation.py", line 212, in _simulate
self.integrator.step(10) # Only take 10 steps at a time, to give Python more chances to respond to a control-c.
File "/usr/local/lib/python3.10/dist-packages/openmm/openmm.py", line 9185, in step
return _openmm.CustomIntegrator_step(self, steps)
openmm.OpenMMException: Error launching HIP compiler: 256
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:681:18: error: no matching function for call to '__frcp_rn'
double3 temp12 = RECIP(mass);
^~~~~
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:23:15: note: expanded from macro 'RECIP'
#define RECIP __frcp_rn
^~~~~~~~~
/opt/rocm-5.5.0/llvm/lib/clang/16.0.0/include/__clang_hip_math.h:700:7: note: candidate function not viable: no known conversion from 'TempType3' (aka 'HIP_vector_type<double, 3>') to 'float' for 1st argument
float __frcp_rn(float __x) { return 1.0f / _x; }
^
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:681:9: error: no viable conversion from 'float' to 'double3' (aka 'HIP_vector_type<double, 3>')
double3 temp12 = RECIP(mass);
^ ~~~~~~~~~~~
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:470:9: note: candidate constructor not viable: no known conversion from 'float' to 'const HIP_vector_type<double, 3> &' for 1st argument
HIP_vector_type(const HIP_vector_type&) = default;
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:473:9: note: candidate constructor not viable: no known conversion from 'float' to 'HIP_vector_type<double, 3> &&' for 1st argument
HIP_vector_type(HIP_vector_type&&) = default;
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:465:9: note: candidate template ignored: requirement 'sizeof...(Us) == 3U' was not satisfied [with Us = ]
HIP_vector_type(Us... xs) noexcept
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:456:9: note: explicit constructor is not a candidate
HIP_vector_type(U x) noexcept
^
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:692:8: warning: unused variable 'gaussian' [-Wunused-variable]
float4 gaussian = gaussianValues[gaussianIndex+index];
^
1 warning and 2 errors generated when compiling for gfx906.
This error went away when replacing the b*sqrt(1/m) expression with b/sqrt(m). Additionally, this error does not happen if the 1 in sqrt(1/m) is replaced by a global variable (possibly with an integer value) or when floats such as 0.99 or 1.01 are given. However, when passing floats that can easily be converted into an integer such as '1.0/m' or '1.2/1.2/m' in the place of '1/m', the above errors are obtained. This error is not reproducible using the OpenCL platform.
When using HIP as a platform, I encountered the HIP compiler errors below while trying to create a simple leapfrog integrator with a gaussian noise term.
CustomIntegrator Function | vvvvvvvvvvvvvvvv
def myintegrator(temperature,friction,dt): integrator=mm.CustomIntegrator(dt); integrator.addGlobalVariable("a", exp(-friction*dt)); integrator.addGlobalVariable("b", sqrt((1-exp(-2*friction*dt))); integrator.addUpdateContextState(); integrator.addComputePerDof("v", "a*v + b*sqrt(1/m)*gaussian"); integrator.addComputePerDof("x", "x + 0.5 * dt*v"); return integratorERRORS | vvvvvvvvvvvvvvvv
File "/usr/local/lib/python3.10/dist-packages/openmm/app/simulation.py", line 147, in step
self._simulate(endStep=self.currentStep+steps)
File "/usr/local/lib/python3.10/dist-packages/openmm/app/simulation.py", line 212, in _simulate
self.integrator.step(10) # Only take 10 steps at a time, to give Python more chances to respond to a control-c.
File "/usr/local/lib/python3.10/dist-packages/openmm/openmm.py", line 9185, in step
return _openmm.CustomIntegrator_step(self, steps)
openmm.OpenMMException: Error launching HIP compiler: 256
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:681:18: error: no matching function for call to '__frcp_rn'
double3 temp12 = RECIP(mass);
^~~~~
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:23:15: note: expanded from macro 'RECIP'
#define RECIP __frcp_rn
^~~~~~~~~
/opt/rocm-5.5.0/llvm/lib/clang/16.0.0/include/__clang_hip_math.h:700:7: note: candidate function not viable: no known conversion from 'TempType3' (aka 'HIP_vector_type<double, 3>') to 'float' for 1st argument
float __frcp_rn(float __x) { return 1.0f / _x; }
^
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:681:9: error: no viable conversion from 'float' to 'double3' (aka 'HIP_vector_type<double, 3>')
double3 temp12 = RECIP(mass);
^ ~~~~~~~~~~~
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:470:9: note: candidate constructor not viable: no known conversion from 'float' to 'const HIP_vector_type<double, 3> &' for 1st argument
HIP_vector_type(const HIP_vector_type&) = default;
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:473:9: note: candidate constructor not viable: no known conversion from 'float' to 'HIP_vector_type<double, 3> &&' for 1st argument
HIP_vector_type(HIP_vector_type&&) = default;
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:465:9: note: candidate template ignored: requirement 'sizeof...(Us) == 3U' was not satisfied [with Us = ]
HIP_vector_type(Us... xs) noexcept
^
/opt/rocm-5.5.0/include/hip/amd_detail/amd_hip_vector_types.h:456:9: note: explicit constructor is not a candidate
HIP_vector_type(U x) noexcept
^
/tmp/openmmTempKernel0x563223b219d0_94919.hip.cpp:692:8: warning: unused variable 'gaussian' [-Wunused-variable]
float4 gaussian = gaussianValues[gaussianIndex+index];
^
1 warning and 2 errors generated when compiling for gfx906.
This error went away when replacing the b*sqrt(1/m) expression with b/sqrt(m). Additionally, this error does not happen if the 1 in sqrt(1/m) is replaced by a global variable (possibly with an integer value) or when floats such as 0.99 or 1.01 are given. However, when passing floats that can easily be converted into an integer such as '1.0/m' or '1.2/1.2/m' in the place of '1/m', the above errors are obtained. This error is not reproducible using the OpenCL platform.