Toy GPU Compiler

A Toy Compiler to compile a C-style language to NVPTX. It is partially based on a User Guide by the LLVM Project:

Prerequisites

Under Ubuntu:

• CMake (sudo apt install cmake)
• A C++ compiler (will be picked up by CMake)
• Antlr 4 Runtime (sudo apt install libantlr4-runtime-dev)
• LLVM 15 (sudo apt install llvm-15)

To actually run compiled code, a CUDA installation is necessary:

• CUDA (sudo apt install ...)

Compilation

• Create a cmake build directory: mkdir build && cd build && cmake ..
• Compile the project: make

Examples

• CUDA needs to be installed
• Run make in the examples directory. This will compile and run two examples.
• They are compiled to .ptx files which are loaded and executed in main.cpp, which also contains code to set up a CUDA context and example data.
• The examples are:
  • addition.toy: Adds two vector and stores it in a third vector. It is not optimized, but only used to check correctness
  • saxpy.toy: Performs a scaled vector addition. It is optimized to achieve a decent memory bandwidth usage (for a fun project)

Example output

• Tested on a machine with an NVIDIA RTX 3090
• Note the achieved memory bandwidth of 388 GB/sec

Output logs:

Using CUDA Device [0]: NVIDIA GeForce RTX 3090
Device Compute Capability: 8.6

Starting Addition Example
Addition results (only the first 5 of 32):
0 + 0 = 0
2 + 4 = 6
4 + 8 = 12
6 + 12 = 18
8 + 16 = 24

Starting Saxpy Example
Time taken by saxpy kernel: 3235 microseconds
Total bytes read: 838860800
Total bytes written: 419430400
Achieved memory bandwidth: 388 GB/sec
Saxpy results (only the first 5):
12.345 * 180.429 + 84.6931 = 2312.09
12.345 * 168.169 + 171.464 = 2247.51
12.345 * 195.775 + 42.4238 = 2459.26
12.345 * 71.9885 + 164.976 = 1053.67
12.345 * 59.6517 + 118.964 = 855.364