Build broken with CUDA 12.5.1

[asenzz]

I'm trying to build KBLAS using CUDA 12.5.82 and GCC 13.2 and getting the following error:
/mnt/slowstore/pub/kblas-gpu/include/kblas_operators.h(130): error: cannot overload functions distinguished by return type alone __attribute__((device)) static __inline__ void atomicAdd(cuFloatComplex* address, cuFloatComplex val)

[AlpacaMax]

I got the exact same error for multiple functions. @asenzz did you manage to solve this problem?

[asenzz]

No, I just don't use KBlas. regards, Zarko Asen ***@***.***

…

On Mon, Jun 23, 2025 at 8:14 PM Tian Yang ***@***.***> wrote: *AlpacaMax* left a comment (ecrc/kblas-gpu#27) <#27 (comment)> I got the exact same error for multiple functions. @asenzz <https://github.com/asenzz> did you manage to solve this problem? — Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACS4YPPLJUQMQXZZSCLDG2L3FBN2TAVCNFSM6AAAAAB76LKT52VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSOJXHAYTKOJUGY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[asenzz]

I think they stopped updating the project for the newest CUDA versions, so I switched to using Cublas. regards, Zarko Asen ***@***.***

…

On Mon, Jun 23, 2025 at 10:15 PM Žarko Asen ***@***.***> wrote: No, I just don't use KBlas. regards, Zarko Asen ***@***.*** On Mon, Jun 23, 2025 at 8:14 PM Tian Yang ***@***.***> wrote: > *AlpacaMax* left a comment (ecrc/kblas-gpu#27) > <#27 (comment)> > > I got the exact same error for multiple functions. @asenzz > <https://github.com/asenzz> did you manage to solve this problem? > > — > Reply to this email directly, view it on GitHub > <#27 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/ACS4YPPLJUQMQXZZSCLDG2L3FBN2TAVCNFSM6AAAAAB76LKT52VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSOJXHAYTKOJUGY> > . > You are receiving this because you were mentioned.Message ID: > ***@***.***> >

[AlpacaMax]

Got it thanks.

[caugonnet]

Hi, I'm stuck on this problem too when trying to compile ParallelVecchiaGP.

The reason is that CUDA has been shipping its own atomicAdd for complex numbers since CUDA 11+, so maybe you need to introduce an extra guard in this style

#if __CUDACC_VER_MAJOR__ < 11
__device__ static __inline__ void atomicAdd(cuFloatComplex* address, cuFloatComplex val)
{
    // custom implementation
}
#endif

The core issue is that this implementation also has a different signature from the CUDA one, which returns a value, so you cannot simply use an ifdef, you need to either add a wrapper on top CUDA's version that does not return a value, or you need to change your custom implementation to return a value too and match exactly the interface of CUDA.