When I run my program from a .bat file instead of ctrl+f5 in VS2005, the number of float error problems I have increases by a factor of 50.
I’m developing a path planning application using CUDA, and am currently trying to run performance tests with various parameters. When I was developing it, I did most of my testing using ctrl+f5 in Visual studio. I call the kernel function several times per cycle, and continue cycling for 5 minutes. In each cycle I only return a small subset of the calculations I run, so at the end of the cycle I send back the small subset to be recalculated and checked against the original calculation. Generally I complete around 250 cycles, and in one or two of these cycles the calculation wont match up. That’s fine for my purposes, though a little annoying.
I just started using a batch file to run the .exe multiple times with different inputs. When I run from the .bat instead of from VS2005, I’m getting those calculation errors 50 to 75 times per 250 cycles, instead of once or twice. 20% errors really isn’t acceptable for my purposes when I can get that below .5% by running it another way, but I really dont want to sit at my computer for 8 hours, running my program with a different input every 5 minutes.
Has anyone else run into this, and do you know a way around it?
I’ve tried running the .bat file in a cmd window using realtime priority, but that hasn’t changed anything. The application is part of a larger program, and I can’t just write a small wrapper to call the main function a bunch of times because most of the larger program doesn’t reinitialise correctly.