deval deliwala
About
I am a rising master's student at Stanford ICME learning to write elegant but fast numerical code.
Contact
devald [at] stanford [dot] edu
linalg-kernels
a fully-safe, pure-Rust BLAS-like library competitive with optimized native implementations.
a safe, idiomatic Rust interface for contiguous BLAS kernels.
my level-1 BLAS. preventing code duplication using generics. when autovectorization is enough.
writing a safe and generic f32/f64 GEMV that outperforms OpenBLAS ARMv8.
a microkernel for TRMV.
GEMM notes
smaller notes documenting progress of GEMM.
thinking of a plan for GEMM. contiguous-only matrices makes things easier.
two GEMMs walk into a compiler. a difficult deep dive into inlining and compiler behavior.
relative to OpenBLAS ARMv8, BLIS, and faer.
all benchmarks. relative to Apple Accelerate, OpenBLAS, BLIS, and faer.
vector-functions
learning to write fast scalar and vectorized elementary functions in C
purpose of vex. the minimax polynomial and the Remez Algorithm.
writing an exponential function with the help of ARM Optimized Routines.
writing logf from scratch. reducing cancellation error from 2 million ULPs to 4.