We compile for gpu architecture `sm_20` by default about brian2cuda HOT 4 CLOSED

the reason was simply compatibility and to make sure no older versions than sm_20 were used as some features from that version are required
i am fine with compiler default though an error/a warning should be shown for an unsupported old version (but this is not important for now -- compiler default is just fine)

from brian2cuda.

We haven't implemented the compiler default yet. When we do this, we also need to modify our _brian_atomic... operations depending on compute capability and cuda runtime version (see here).

from brian2cuda.

currently we compile with cc 3.5 as default. check whether we support older archs as well (e.g. cc 2.0) -- would be of no priority but if its free lunch we should not prevent it

from brian2cuda.

Commit 41bab5d implements automatic compute capability detection based on chosen GPU and adds a preference to manually set the compute capability (prefs.brian2cuda.compute_capability). It also detects if a user passes any architecture compile flags in prefs.codegen.cuda.extra_compile_args_nvcc. In that case, manually set compute capability is ignored and a warning is printed.

Commit 4007b84 implements correct atomicAdd support code generation depending on detected or manually set compute capability.

For the last comment of @moritzaugustin on minimal supported compute capability: It is currently set to 3.5 and the newest CUDA toolkit dropped support for anything below that. But checking for what we support (with older tookit versions) will be done in #5.

Closing.

from brian2cuda.