hjunkim/CO-Optimizer

Code-level op-chip memory optimizer for your CUDA applications.

CO-Optimizer: Code-level On-chip memory Optimizer

This project gives an opportunity to optimize space-limited on-chip memories (L1, shared memory) of your CUDA applications.

If you use or build on this tool, please cite the following papers.

• Throttling for L1 cache contention reduction (ICPP'19)
• Preloading in the shared memory by using memory-inactive threads (CCPE'20)

Getting the Source Code and Building the On-chip Memory Optimizer

This is an example work-flow and configuration to get and build the Transpiler.

0. Tested with following setups

   • Ubuntu 18.04, cmake-3.10.2, gcc/g++-5
   • sudo apt install gcc-5 g++-5 cmake
   • sudo apt install libboost-all-dev
   • Benchmark -- PolyBench/GPU and Rodinia

1. Checkout llvm, clang, and co-optimizer

   • llvm

     • git clone https://github.com/llvm-mirror/llvm
     • cd llvm
     • git checkout release_80

   • clang

     • cd tools;
     • git clone https://github.com/llvm-mirror/clang
     • cd clang;
     • git checkout release_80

   • co-optimizer

     • cd tools
     • git clone https://github.com/hjunkim/CO-Optimizer.git

2. Build them

   • Add the CO-Optimizer repository to llvm/tools/clang/tools/CMakeLists.txt
     • add_clang_subdirectory(CO-Optimizer)
   • cd ../../../../;mkdir build;cd build
   • cmake -G "Unix Makefiles" ../llvm
   • make -j 16;sudo make install

Usage

Run

• {bin} {cuda_program}.cu [Run Options] -- --cuda-device-only --cuda-path={path/to/cuda} --cuda-gpu-arch={sm_xx}
  • {bin} --- ./build/bin/{throttling/preloading}
  • {cuda_program}.cu --- your target CUDA program
  • --cuda-device-only --- will run analysis/translate for the device code
  • --cuda-path= --- installed CUDA path (ex: /usr/local/cuda)
  • --cuda-gpu-arch=sm_xx --- CUDA architecture (ex: Titan V, V100: sm_70)

Run Options

• Throttling:

  • --csize=<int> - : L1 cache size of the GPU (default: 32 KB)
  • --nblks=<int> - : # of thread blocks per SM (default: 4 blks)
  • --tbsize=<int> - : thread block size (default: 8 warps)

• Preloading

  • --prdsize=<string> - : set preloading size (default: 1)
  • --tbsize=<string> - : set thread block size (default: 8 warps)