CHANGE LOG:
Version 3.1.0b versus 3.0.0b
- Added prints for L1 data cache statistics
- Added read-to-precharge constraint in DRAM
- Disabled Stream Manager's verbose output by default, use debug level 3 to 
  enable it
- Addresses returned by memory allocation are now 256 bytes aligned
- Ejection from the clock domain interface buffer between interconnection 
  network and L2 cache happens in the L2 clock domain instead of interconnect 
  clock domain.
- Testing the simulator on Ubuntu 10.04.3 LTS
- Added Doxygen generated documentation
- Bug fixes
   - Fixed the variation in instruction count seen under different cache
     configurations on the same workload
   - Fixed unnecessary flushing of instruction buffer
   - Fixed mislabeling of stall cycles recorded by scheduler
   - Fixed compilation issues when using gcc 4.5.1
   - Cleanup and bugfixes to build scripts
   - Fixed gpu_sim_cycle with sequential kernel launches

Version 3.0.0b versus 2.1.2b
- massive refactoring of code to C++ (most global variables eliminated)
- support for executing native SASS code via decuda
- added to timing model:
   - in order scoreboard allowing multiple instructions per warp in 
     pipeline at a time
   - operand collector for modeling access to banked register files
   - prefetching texture cache model (Igehy et al., Graphics Hardware 1998)
   - updated data and constant cache models with updated MSHR model
- other changes I don't remember right now

Version 2.1.2b versus 2.1.1b:
- TODO: fill this in!

Version 2.1.1b versus 2.1.0b:
- Added support for OpenCL (requires OpenCL driver from NVIDIA)
- Added performance visualizer tool
- Added manual documenting usage of the simulator (see doc directory)
- Added configuration file for Quadro FX5800 (see configs directory)
- Simultaneous release of benchmarks from ISPASS 2009 paper (separate download)
- Improved shared memory bank conflict modeling
- Improved default mapping of local memory accesses to global address space 
- Added interconnect concentration modeling (to approximate effect of a TPC)
- Added support for GPU-to-GPU memory copies (same GPU), timing not modeled
- SimpleScalar code (for command line parsing and functional simulation of 
  memory) removed.
- PTX support for vector operands in mov instruction
- Bug fixes
   - stack overwrite bug fix (with thanks to Jake Adriaens)
   - interconnect clock domain crossing "bottleneck" bug (was not removing
     traffic quickly enough going from memory to shader cores)
   - immediate post dominator detection (affected eigenvalues SDK benchmark)
   - CTA/core calculation error with PTX containing calls and returns
   - global symbol lookup error when building with multiple .cu files
- Other general improvements and bugfixes

Version 2.1.0b versus 2.0b:
- Added support for parsing and functionally simulating up to CUDA 2.2 
  generated PTX
- Added dynamic warp formation timing model (see MICRO'07, ACM TACO'09 papers)
- Updated gpgpusim.config and mesh in benchmark subdirectories to be similar to
  ISPASS 2009 paper baseline configurations
- Added OpenGL interoperability support
- Added support for parsing embedded PTX files without requiring recompilation 
  (suggested by Gregory Diamos)
- Improved support for texture filtering (linear mode for 2D, closer agreement
  to hardware for 2D point sampling where sample points are close to texel
  boundaries)
- Benchmark examples updated to version from CUDA 2.2 SDK (NOTE: these will not 
  compile with earlier CUDA installations. However, GPGPU-Sim should still work 
  with applications written for older versions of CUDA.)
- Fixed splash screen: was reporting incorrect version
- Fixed bug comparing strings in basic block formation code
- Fixed bug with additional basic block (which lead to incorrect postdominator 
  analysis on some benchmarks with recent versions of CUDA)
- Fixed "warning: "/*" within comment" when building.
- "Other general improvements"