Will there be a windows binary with keccak? Please?
7 PM UTC
We're severly under register pressure on Compute 2.x and 3.0 devices
but Compute 3.5 supports about 255 registers max!
So guys, you can compile the keccak.cu module with -maxrregcount=112. That is what the kernel really wants under compute/sm_10. Then run keccak benchmarks with x16 launch configs and check if it's any faster. I do not have a Compute 3.5 device handy right now.
do NOT compile keccak.cu for sm_30 - in my testing the hashes were no longer validating. And a faster speed is no good if your submitted shares are broken.
I will try to reduce the register pressure in the kernel if possible - the above suggestion might me a good workaround until this is done.
Christian
