About

Gagandeep Singh is a highly skilled Compiler Developer at GSI Technology, with a passion for open source technology. He spends most of his day on Github, opening PRs and fixing issues. He is currently working on LPython, a Python compiler for programming APUs. Gagandeep is responsible for adding all the Python language features in LLVM and C backends such as dict, list, numpy functionality, and array programming support. He also works on adding machine-independent code optimizations such as automatic vectorization of array operations, function inlining, loop unrolling, and fused-multiplication addition. Preliminary benchmarks reflect that with the recent addition of list, LPython beats C++ std::vector by a factor of 9x (without enabling optimization for C++ test code) and 2x (with optimized C++ test code). Similarly, for dict (hash map), LPython beats C++ std::unordered_map by a factor of 30x (without C++ test code optimized) and 6x (with C++ test code optimized). Gagandeep is implementing traditional Python features optimally in LPython, which will allow increasing the performance of existing Python code as is. In addition, the presence of multiple backends will allow users to transform their Python code into other high-level languages for their specific use cases without giving up on the comfortable development experience provided by Python syntax. Prior to his current role, Gagandeep worked as a Software Developer at Quansight, where he was responsible for improving the Ray project for Windows, implementing array standard APIs (NEP 47) in NumPy, and shifting the build system of SciPy from setuptools to meson. He also added iterative implementation of Dinic's algorithm for maximum flow and optimized check_grad by using random directional derivatives in SciPy. Gagandeep has also worked as a Student Software Developer at LFortran during his time with Google Summer of Code. He added semantic and backend support for arrays and allocatables in LFortran as per Fortran standards. He implemented various nodes like Allocate, Deallocate in the Abstract Semantic Representation (ASR) of the Fortran code. Operations over arrays are supported by implementing recursive algorithms to convert these operations into loops. Backend support is achieved by dynamically generating LLVM IR instructions for various low-level array functions in Fortran such as size, lbound, ubound, etc. Automatic deallocation of heap memory variables is achieved by adding an ImplicitDe