a2flo/floor_examples

example programs for https://github.com/a2flo/floor (dnn, nbody, warp, hlbvh, path tracer, other)

floor examples

dnn

• demonstrates inference on a VGG16 deep neural network (implements convolution, max-pooling, fully-connected and softmax layers)

• build with ./build.sh inside the folder

• cd data/nets/vgg16 && ./get_vgg16.sh to download the necessary VGG16 layers/net

• desktop/cli: run with --image <path/to/image.png> (image must be 224*224px 32-bit RGBA for now)

• iOS: press cheese and take a photo

warp

• image-based warping, implemented using a scatter-based approach and a gather-based approach (based on "Image-Based Bidirectional Scene Reprojection")

• also demonstrates use of: tessellation/displacement, argument buffers and indirect command pipelines

• build with ./build.sh inside the folder

• gather-based warping ref: Image-Based Bidirectional Scene Reprojection (original)

• NOTE: in addition to floor + dependencies, this requires SDL2_image with libpng, as well as libwarp

• video:
  

nbody

• N-body simulation to demonstrate local/shared memory buffers, local memory barriers, OpenGL and Metal buffer sharing, loop unrolling and that high performance computing is indeed possible with this toolchain

• build with ./build.sh inside the folder

• ref: http://http.developer.nvidia.com/GPUGems3/gpugems3_ch31.html

• video:
  

hlbvh

• triangle/triangle collision detection of animated meshes using HLBVH (constructed per-step/frame)

• build with ./build.sh inside the folder

• ref: https://research.nvidia.com/sites/default/files/publications/karras2012hpg_paper.pdf

path_tracer

• a simple path tracer to demonstrate compute code compilation and execution, as well as some mildly complex modern C++ code

• build with ./build.sh inside the folder

• default output:
  

• output with --with-textures:
  

reduction

• simple (WIP) reduction example that showcases 3 different reduce implementations: local/shared memory reduce, shuffle reduce and CUDA coop kernel + shuffle reduce

• inclusive/exclusive scan test

• build with ./build.sh inside the folder

img

• small program to demonstrate image functionality by doing a gaussian blur, implemented as both a single-stage blur with manual local memory caching, as well as a separable horizontal/vertical "dumb" blur w/o manual caching (also: a GLSL/OpenGL implementation that does the same)

• build with ./build.sh inside the folder

• example output:
  

occ

• offline-compute-compiler, compiles compute/graphics C++ code to CUDA/PTX, Metal/AIR, OpenCL/SPIR/SPIR-V, Vulkan/SPIR-V code/binaries (see --help for all options)

• build with ./build.sh inside the folder

libfloor and toolchain

• https://github.com/a2flo/floor