BorisCUDALib/Reduction.cuh(138): error: no instance of overloaded function "atomicAdd" matches the argument list

Robert Ryans

unread,

Oct 22, 2025, 12:09:15 PMOct 22

to Boris Computational Spintronics

Hi

I was asked to build Boris on one of our systems for a student. The PC has a 4080 and is running Ubuntu 22.04 LTS. I'm using arch=90 for building.

I have tried using both CUDA 12.9 and 13.0 and get the same error when starting to compile the cuda version - see below. The cpp version builds fine, it's a CUDA error on the first file it tries to compile. Google doesn't throw up much of use - other people see the overloaded function error but I'm not a C person so don't quite now what to try next.

Would be grateful for any suggestions!

Thanks

Robert

$ /usr/local/cuda-12.9/bin/nvcc -I/usr/local/cuda-12.9/targets/x86_64-linux/include/ -rdc=true -c -std=c++14 -IBorisLib -IBorisCUDALib -w -arch=sm_90 Boris/Atom_DiffEqCubic_EvalsCUDA_TEuler.cu -o Boris/Boris_cuo/Atom_DiffEqCubic_EvalsCUDA_TEuler.o

BorisCUDALib/Reduction.cuh(138): error: no instance of overloaded function "atomicAdd" matches the argument list

argument types are: (size_t *, size_t)

atomicAdd(&points_count, shared_memory_count[0]);

^

BorisCUDALib/atomics.cuh(183): note #3327-D: candidate function template "atomicAdd(cuVAL3<Type> *, const cuVAL3<Type> &)" failed deduction

__attribute__((device)) void atomicAdd(cuVAL3<Type>* result, const cuVAL3<Type>& value)

^

BorisCUDALib/atomics.cuh(176): note #3327-D: candidate function template "atomicAdd(cuVAL2<Type> *, const cuVAL2<Type> &)" failed deduction

__attribute__((device)) void atomicAdd(cuVAL2<Type>* result, const cuVAL2<Type>& value)

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/crt/sm_90_rt.h(112): note #3326-D: function "atomicAdd(float4 *, float4)" does not match because argument #1 does not match parameter

static __attribute__((device)) __inline__ float4 atomicAdd(float4 *__address, float4 val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/crt/sm_90_rt.h(109): note #3326-D: function "atomicAdd(float2 *, float2)" does not match because argument #1 does not match parameter

static __attribute__((device)) __inline__ float2 atomicAdd(float2 *__address, float2 val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/sm_60_atomic_functions.h(93): note #3326-D: function "atomicAdd(double *, double)" does not match because argument #1 does not match parameter

static __inline__ __attribute__((device)) double atomicAdd(double *address, double val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/sm_20_atomic_functions.h(88): note #3326-D: function "atomicAdd(float *, float)" does not match because argument #1 does not match parameter

static __inline__ __attribute__((device)) float atomicAdd(float *address, float val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/device_atomic_functions.h(169): note #3326-D: function "atomicAdd(unsigned long long *, unsigned long long)" does not match because argument #1 does not match parameter

static __inline__ __attribute__((device)) unsigned long long int atomicAdd(unsigned long long int *address, unsigned long long int val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/device_atomic_functions.h(91): note #3326-D: function "atomicAdd(unsigned int *, unsigned int)" does not match because argument #1 does not match parameter

static __inline__ __attribute__((device)) unsigned int atomicAdd(unsigned int *address, unsigned int val) { }

^

/usr/local/cuda-12.9/targets/x86_64-linux/include/device_atomic_functions.h(89): note #3326-D: function "atomicAdd(int *, int)" does not match because argument #1 does not match parameter

static __inline__ __attribute__((device)) int atomicAdd(int *address, int val) { }

^

1 error detected in the compilation of "Boris/Atom_DiffEqCubic_EvalsCUDA_TEuler.cu".

Robert Ryans

unread,

Oct 22, 2025, 12:28:48 PMOct 22

to Boris Computational Spintronics

OK - I found the previous post which included the advice to

Find the file atomics.cuh in BorisCUDALib directory. There, on line 163, replace the line "#if __CUDA_ARCH__ <= 700" with "#if __CUDA_ARCH__ <= 800".

So I just edited that from 800 to 900 and that seems to be letting it build. That may or may not have been a reasonable thing to do. But it built.

Serban Lepadatu

unread,

Oct 23, 2025, 2:35:31 AMOct 23

to Boris Computational Spintronics

Hi Robert,

Yes, that was what I was going to advise.

That being said, it's worth double checking you do need arch 900. I recently had the same problem for NVidia L40S GPU. This has compute capability 8.9. Compiling with architecture 900 did not work on this GPU, it still needs 800.