Christian a question.
I was testing around with keccak algo and I enabled checkCudaErrors macro in the nvkernel.cu.
Using K1024x32 with a lookup gap of 1024 (I have a GTX 660).
Everything seems to work but I do get a runtime error at: (nvkernel.cu:99 in repository)
at the call:


The c_V array on the device only supports up to TOTAL_WARP_LIMIT elements but since MAXWARPS under these settings
is higher the memcpy fails.
What I don't understand is, the pool is still accepting shares. Shouldn't this break the algo?
Is this a bug or can it be ignored?

Is this trick only required for windows? If so, you could try to set the time period to 1 ms. (I believe this is what flash does).
The API is: timeBeginPeriod
What this basically does is increase the (global) interrupt time of windows, which among other things allows waiting functions to be more accurate.
Windows has a default period of 15.6 ms. Meaning every WaitForX, Sleep or Timer function has no better accuracy then that.
When 1 process in the system (e.g chrome) requests a lower time period it affects all processes running on the system.  
Hence it might be boosting CudaMiner.
So if someone who gets major benefits of having chrome open and is able to compile the software would like to test this:
Add somewhere at the top of main a call to timeBeginPeriod(1).
At the top of .c file in which main is defined add:
#include <Mmsystem.h>
#pragma comment(lib, "winmm.lib")