support turing and ampere for CUDA

[forrestjgq]

In turing and ampere device, invalid device symbol will be reported in:

123 static __constant__ GpuData cData;
124 static GpuData cpuData;
125 
126 void SetKernelsGpuData(GpuData* pData)
127 {
128     cudaError_t status;
129     status = cudaMemcpyToSymbol(cData, pData, sizeof(GpuData));                                                         
130    >> RTERROR(status, "SetKernelsGpuData copy to cData failed");
131     memcpy(&cpuData, pData, sizeof(GpuData));
132 }

[syntesys87]

Tried in a system with a 4070ti and cuda 11.8. Working!

[TitouanCh]

works for rtx a4500

[atillack]

@forrestjgq Thank you for submitting your PR. While we won't go ahead and merge it - as adding to the static TARGETS list (and it existing in the first place) is not the best choice - it did inspire me to solve this particular pain point more generally.

To that end, I added PR #270 which by default will compile Cuda for every supported target of the installed Cuda version (larger than compute capability 50 as lower compute capabilities have a deprecation warning). In other words, Cuda 11 will compile up to compute capability 86, while Cuda 12 will go to compute capability 90 (and beyond if/when Nvidia adds it).

As with the current code, TARGETS can be used to override this during compilation, i.e. make DEVICE=GPU TARGETS=86 will only produce code optimized for compute architecture 86.