CUDA Kernel fails at sqrt() or rsqrtf() operations

[Kirill Vasin]

I found a strange bug. Let me consider the simplest kernel

__global__ void makezeros(
    float4* spin
) {
    int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
    int gid = blockId * (blockDim.x * blockDim.y) + (threadIdx.y * blockDim.x) + threadIdx.x;
    
    spin[gid].x = 0.001;
}

It works absolutely perfect. However, it is enough to add a square root operation

__global__ void makezeros(
    float4* spin
) {
    int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
    int gid = blockId * (blockDim.x * blockDim.y) + (threadIdx.y * blockDim.x) + threadIdx.x;
    
    float empty = sqrt(3.3f);
    spin[gid].x = 0.001;
}

and the, the array spin will not be affected. However, It doesn't show any errors.
It looks like that this kernel code was not even called.

[Kirill Vasin]

Found an error code. It is

PTX JIT compilation failed
Error code 78

Every time at sqrt() function or others

[Simeon Ehrig]

I have to verify it with the latest cling version. The container is highly outdated.

My assumption: The path to the CUDA std library is not set for the device compiler.