What you will do:

The Compiler Backend Team at d-Matrix is responsible for developing the software that performs the logical-to-physical mapping of a graph expressed in an IR dialect (like Tensor Operator Set Architecture (TOSA), MHLO or Linalg) to the physical architecture of the distributed parallel memory accelerator used to execute it. It performs multiple passes over the IR to apply operations like tiling, compute resource allocation, memory buffer allocation, scheduling and code generation. You will be joining a team of exceptional people enthusiastic about developing state-of-the-art ML compiler technology. This internship position is for 3 months.

In this role you will design, implement and evaluate a method for managing floating point data types in the compiler. You will work under the guidance of two members of the compiler backend team. One, is an experienced compiler developer based in the West Coast of the US.
You will engage and collaborate with engineering team in the US to understand the mechanisms made available by the hardware design to perform efficient floating point operations using reduced precision floating point data types.
Successful completion of the project will be demonstrated by a simple model output by the compiler incorporating the your code that executes correctly on the hardware instruction set architecture (ISA) simulator. This model incorporates various number format representations for reduced precision floating point.

What you will bring:

•    Bachelor’s degree in computer science or equivalent 3 years towards an Engineering degree with emphasis on computing and mathematics coursework.
•    Proficiency with C++ object-oriented programming is essential.
•    Understanding of fixed point and floating-point number representations, floating point arithmetic, reduced precision floating point representations and sparse matrix storage representations and the methods used to convert between them.
•    Some experience in applied computer programming (e.g. prior internship).
•    Understanding of basic compiler concepts and methods used in creating compilers (ideally via a compiler course).
•    Data structures and algorithms for manipulating directed acyclic graphs.

Desired:
•    Familiarity of sparse matrix storage representations.
•    Hands on experience with CNN, RNN, Transformer neural network architectures
•    Experience with programming GPUs and specialized HW accelerator systems for deep neural networks.
•    Passionate about learning new compiler development methodologies like MLIR.
•    Enthusiastic about learning new concepts from compiler experts in the US and a willingness to defeat the time zone barriers to facilitate collaboration.