How gpgpu-sim/gpgpu-sim_distribution works
Project type and purpose
• This is GPGPU-Sim, a cycle-level GPU simulator for CUDA/OpenCL workloads.
• It includes three major deliverables:
• The core simulator (PTX/SASS front-end, pipeline, interconnect, memory hierarchy)
• A power-model extension (AccelWattch)
• A performance-visualization tool (AerialVision)
File structure
Top-level directories:
• src/ → core simulation modules (gpgpu-sim, cuda-sim, accelwattch, intersim2)
• libcuda/, libopencl/ → runtime API wrappers for cuda and OpenCL calls
• cuobjdump_to_ptxplus/ → a standalone binary translator (SASS → PTX+)
• aerialvision/ → Python-based visualization front-end
• debug_tools/ → GUI and command-line debug harnesses (e.g., WatchYourStep)
• configs/ → static XML/Text files modeling GPU microarchitectures
• bin/ → top-level driver scripts (aerialvision.py, aerialvision startup)
Build & CI:
• CMakeLists.txt, Makefile, cmake/ → build orchestration
• .github/, .travis.yml, bitbucket-pipelines.yml → continuous integration
README for additional insights
• Simulator is invoked by setting CUDA_INSTALL_PATH and sourcing setup_environment.
• Supports PTX-level, SASS-trace-level, and hybrid simulation.
• AccelWattch is pluggable via XML config (power_simulation_enabled, mode, xml file).
• AerialVision reads simulator dump files via Python PLY; depends on numpy, matplotlib.
• Debug tools rely on Flex/Bison parsers for PTX/SASS.