Octavian Rusu
present day. present time.
contact @octavian.work
github @gandalftea
discord @gandalfteaparty
blog
03.2024 > visualising ymm registers of VINSERTF128 transpose kernel
02.2024 > diving into a BLIS 8x8 sgemm kernel
now
building 5x5 convolution vector kernels for an efficient mnist example in c. afterwords, will add variable kernel size winograd conv2d's in tensorlib.
tensorlib
: tensor library with AVX256 and SSE vectorisation:
- 660 GFLOPS threaded cache-aware CPU sgemm kernel. [repo] [blog]
- VINSERTF128 8x8 transpose kernel. [code] [blog]
- VMOVNTDQ cache-less streaming for faster copy then memcpy.
- random number generators with kaiming normal/uniform, box-muller normal, chi-squared and uniform distributions + statistical tests with GSL (kolmogorov-smirnov, dagostino skewness/omnibus, etc.)
- CPUID asm calls and bit fields for SRAM size, associativity and TLB, including AMD 0x80000000 leafs.
- aligned data allocation and garbage collection
past work
MRCV : Simultanious Localisation and Mapping computer vision library:
- monocular/binocular environment mapping and position tracking.
- built using Eigen and OpenCV helper functions.
- ORB keypoints + descriptors and Brute-Force Hamming matcher.
- using Pangolin or OpenGL for point cloud map visualisation.
- had to manually implement a MovementHandler in Pangolin to allow for camera movement using keyboard and mouse.
- intended to be used when building DIY HMDs
- also built it again in Python with numpy and cv2.
enginehmw : procedural game engine with OpenGL rendering:
- procedural terrain generation with perlin and fractal noise.
- OpenGL + glut rendering and first-person controllable camera
- quaternions, euler angles and transformation matrices.
- generate or load tris or quad models from .obj files
- compute normals or sparse-normals
- variable resolution collision
opendataset : collaborative dataset creation tool:
- express.js webserver, react and cli frontends.
- PostgreSQL database, user authentification with pgcrypto hashing and express-session cookies.
- git-like object-based source control with database backend, binary contribution objects with custom byte headers, dataset delta files that allow for reversing of contributions, etc.
- automatic schema generation from .csv and .json files