Runtime-Viz

With Runtime-Viz you can record memory accesses to global memory in your CUDA kernel and create an interactive visualization of the data.

NVBit is NVidias binary instrumentation tool. It works by dynamically loading a shared library with the LD_PRELOAD trick. Then any CUDA API call can be intercepted and custom code can be run. This project uses this functionality to intercept kernel launches and replace the memory access instructions in the kernel code with jumps to a method which saves the access addresses and sends them to the CPU. The addresses get mapped to buffers which must be manually tracked beforehand by calling TRACK_BUFFER<T>(T *buf, const char *name).

Usage

[TOOL_VERBOSE=1] [ONLY_WRITES=1] LD_PRELOAD=path_to_libmem_trace.so ./program_to_execute [program arguments]

Options:

• TOOL_VERBOSE=1 prints all kind of information
• ONLY_WRITES=1 only records memory accesses that perform a write operation

Executing this will generate a bunch of *.accesses.bin files and a buffers.json file. *.accesses.bin contain the raw access data for each kernel. Calling the same kernel multiple times overwrites the old file. So this should be avoided. buffers.json contains all the buffer information. An interactive visualization can then be made by calling python visualization/main.py buffers.json. Note that the visualization tool accumulates all buffer accesses. It can be configured to only show certain kernels. Use python visualization/main.py --help for more information.

Compiling

Dependencies for the Runtime-Viz are zlib and jsoncons. Both are included as submodules. Of course you will also need the CUDA compiler and libraries. CMake is used as the build system. For visualization you will need Python 3.x along with the python packages numpy and matplotlib. They can be installed with pip install -r visualization/requirements.txt.

First make sure that the submodules jsoncons and zlib are pulled. Then build zlib by executing make zlib in the root directory. Next build memtrack and mem_trace by executing:

make build
cd build
cmake ..
make

This will generate libmem_trace.so and all programs in test-apps.

Integrating a Custom Project

1. Create a directory in test-apps for your application.
2. In the accompanying CMakeLists.txt add the line target_link_libraries(name_to_target memtrack).
3. In the root directory CMakeLists.txt add the line add_subdirectory(${CMAKE_SOURCE_DIR}/test-apps/name_of_app).

Then cd into the build directory and excute:

cmake ..
make

This should create an executable in build/test-apps/name_of_app.