FERN: Fast Explicit Reaction Networks SVNCode

Modified explicitgpuDouble.cu to convert cuPrintf statements to printf statements in the kernel. Requires compute capability 2.0, so compile with nvcc -arch sm_20 explicitgpuDouble.cu. Without the -arch sm_20 flag the printf statements in the kernel will give compile errors.

Changed the three places where populations are updated to do serial. Now, at least for the first 100 timesteps, the CUDA and Java codes give almost exactly the same results. Thus, there seems to be an error in the parallel implementation of the population update algorithm.

update

update

Minor changes

Added print statements for each timestep in parallel with those in the java version for simpler debugging.

Current version for 100 steps

Corrected initial value of sumXLast to sum over actual mass fractions rather than 1. This seems to have been distorting the initial timestep correction versus the Java code. Now the first two timesteps give the same dt for CUDA and Java codes.

Print buffer size increased by factor of 10 over default in line 1064 to get all the diagnostic printf stuff from the gpu output. Only the last several lines are output if the default buffer size is used.

Minor revision