How Does Rust Code Run on a GPU? Tracing cuda-oxide and PTX

Here is the story written from scratch, without reusing any of the source's phrasing:

I ran a vector addition kernel written in Rust on a GPU using the cuda-oxide tool. The calculation adds two arrays element-wise and returns the results. After building the example, I examined the generated code which is in a form called PTX. PTX stands for Parallel Thread Execution and is an instruction set for NVIDIA GPUs. It represents the Rust code before it is converted into actual machine code that the GPU can run.

I wanted to understand how the Rust thread indexing and array access worked in the PTX representation. cuda-oxide is part of NVIDIA's CUDA Rust project. It is a compiler that lets you write GPU kernels in Rust and then compile them. The name SIMT refers to Single Instruction, Multiple Threads which describes the programming model used.

Many threads execute the same kernel. Each thread computes the index of the element it is responsible for and performs that single element calculation. The thread index is provided by cuda-oxide and represents the thread's position. The generated PTX file starts with version, architecture and address width directives. It then includes the kernel entry point and parameter declarations.

The body of the PTX code contains the thread index extraction, bounds checking, and the actual addition operation. cuda-oxide provides tools to translate Rust code into PTX. It is an early-stage research project with features still under development. The project allows Rust programmers to write GPU kernels and compile them to run on NVIDIA GPUs.