Status meeting April 12th, 2012¶
Update on implementing electron's physics process (Soon, Dongwook)¶
Dongwook is traveling to South Korea.
Update on CUDA development (Soon, Philippe)¶
Soon: I have finish translation the G4 transportation code into C that works on the device but not yet in GPU (some odd crashes). Some iterative process is involved in the stepping to reduce the error. To test in real situation, in some detector location might requires more steps, so need 'real' data. I would like to understand more of the pattern of this number of iteration.
Philippe: Finished G4MagErrorStepper__Stepper optimized implementation in CUDA. As it stands (error calculation are outside of the kernel), the kernel when using all 448 cores is up to 5.76 times faster than the CPU on cluck including uploading and 56 times faster if only taking computation in consideration. The fastest runs had a configuration where each block was 48 threads, and 14*256 blocks.