/RTIOW_rust

https://raytracing.github.io/books/RayTracingInOneWeekend.html but rust and parallelized with rayon

Primary LanguageRust

RTIOW - Rust Version

So Far

  • I first made a naive implementation of Ray Tracing In One Weekend by basically 1:1 copying the C++ into Rust
  • I made changes to trade slower compilation for faster performance
    • link time optimization: fat
    • codegen units: 1
    • emit asm so llvm and further optimize
  • I made various tweaks to make my code more idiomatic (hopefully), to satisfy my craving for functional programming, improve performance, and for aesthetic/dogma reasons
    • Returning options instead of mutating C-like in/out parameters and returning a bool
    • Replacing imperative loops with map to express applying an operation on the whole collection simultaneously
    • Replacing if statements with match statements for FP reasons (much flimsier justification)
  • I added the rayon crate
    • Replaced .iter() and .into_iter() with .par_iter() and .into_par_iter()
    • tweaked rng generators and fold/reduce/collects until it compiled
    • Now with multithreaded execution, it was much faster

Current Steps

  • Figure out how to profile execution to find what are the slow parts
    • Possibly try timing runs with glam vs my own library to see which is faster
  • Figure out how to pass threadsafe rng or statically give each thread its own rng rather than making a new rng for each thing
  • Figure out what the best combination of parallel iterators are
    • Rows
    • Pixel in Row
    • Photon per Pixel
    • Which of these are better parallelized accounting for thread/rng overhead
  • General idiomatic code/performance improvements

Next Steps

  • Ray Tracing: The Next Week
  • Interactive Window
    • Move camera to find the best picture
    • Adjust camera effects like blur and fov
    • Redrawing the window takes less than 1s so low resolution and photon count
    • Load different scene
    • Fix Scene/Camera and Render high quality image to file
  • Png output
  • More shapes
    • Triangles, the vertex buffer default, allows for real polygons and general shapes
  • More materials
  • Possibly add distinction between vec3 and vec4 for homogeneous coordinates and normal vectors and such
  • Allow loading of object files

Benchmarking, Profiling, and Maximizing Performance

  • Increasing Performance:

    • --release vs debug
    • --emit=asm vs none
    • lto = fat vs thin vs none
    • codegen-units = 1
    • -C target-cpu=native
    • panic = abort

  • Using perf to profile

$ perf record --call-graph dwarf -- cargo run --release
$ perf report
  • Build
RUSTFLAGS="--emit=asm -C target-cpu=native" cargo build --release
  • Run
time target/release/./raytracer_weekend > dump.ppm