Hi. I've added you to the developers white list and moved this message to the OpenCL forum. You can now post in any of the developer forums.
The R9 380 has 1/16 double prerformance rate.
There are a few options:
- Get a new workstation gpu, those have good double performance, but also they aren't cheap.
- Get an older GCN gpu: As I remember, the HD7970 was at 1/4 rate.
- Get a fast new gaming card but use big numbers realized with the fast 32bit float instructions and their integer variants: v_mad_i32_i24 for example. Maybe the task can be done with this at the required precision, and if it is faster than the 1/16 rate, it can be a solution. (Note: If you need NAN/INF handling too, then this is a no go.)
I'm creating a neural network, and i was considering going for "fake decimal" but i suppose it's too much work for now since i haven't even finished the neural network