Staff Compiler Engineer at AMD · LLVM/MLIR · C++ Systems
Download CVHello! I’m Akash, a staff compiler engineer at AMD in the UK. I work on OpenMP target offload in LLVM Flang and MLIR, from Fortran lowering and OpenMP dialect design through LLVM IR generation.
My main focus is data mapping: target-data constructs, user-defined and default declare mappers, and derived-type semantics. I have also helped move reusable offload code generation from Clang into OpenMPIRBuilder, allowing Clang and MLIR-based frontends to share the same implementation.
My background also includes performance engineering, program analysis and formal verification. I’m always happy to chat about compilers, open source or interesting systems problems.
When I’m not coding, you can find me losing at DOTA 2, staring at my fish tanks, or giving pats to Simba (OTRB). I’m also learning photography 🙂
Dominik Adamski, Sergio Afonso, Akash Banerjee, et al. Proceedings of the SC ’25 Workshops, pp. 1028–1038. The paper describes the architecture and implementation of OpenMP target offload in LLVM Flang.
Read on ACM Digital LibrarySelected compiler, low-level systems and verification work. Current LLVM projects are shown first; older research projects are retained where they show end-to-end systems work.
Built core pieces of the OpenMP mapping path for LLVM Flang and MLIR: map representation, target-region mapping, map-clause attachment, FIR conversion, LLVM IR translation and mapper emission. The work covers explicit user mappers, default mappers and derived-type component mapping.
LLVM contributionsImplemented AUTOMAP lowering and mapper-semantics fixes for partial maps, nested types, pointer captures and data-motion directives. Added diagnostics for unsafe descriptor mappings so incorrect offload cases fail clearly instead of producing misleading generated code.
LLVM contributionsImplemented AUTOMAP lowering and an FIR pass that converts inferred mappings into explicit target-data operations, then extended implicit mapping to nested allocatable components of Fortran derived types.
LLVM contributionsAdded AMDGPU handling for complex function arguments and results, together with complex operations and ROCDL lowerings required by Flang workloads.
LLVM contributionsMigrated reusable offload code generation from Clang into OpenMPIRBuilder, including target-data calls, mapping arrays and descriptors, device privatisation, GPU reductions and mapper emission. These paths are reused by Clang and MLIR-based frontends.
LLVM contributions
Added operations and end-to-end lowering for
target data, target enter data and
target exit data, spanning Flang lowering,
FIR-to-LLVM conversion and OpenMPIRBuilder translation, together
with map support for target regions.
Designed and built a compiler tool using polymorphic compilation and execution techniques to mitigate a class of side-channel attacks with low overhead. This was my master's thesis work and sits at the boundary between compiler transformations and low-level execution behaviour.
Proposed and implemented improvements to the Branch Probability Information pass in LLVM to allow better static profiling. The enhancements led to speed‑ups of up to 1.07× and were accepted as a poster at EuroLLVM 2020.
Source PosterAdded a loop-acceleration module for Pinaka, a symbolic-execution based C verifier used in SV-COMP. The module targets faster counterexample discovery for loops that simulate polynomial functions.
Source SV-COMPCreated a tool to translate LLVM IR to CBMC‑GOTO. This allows verification of programs that have LLVM front‑ends by translating their IR into GOTO IR and verifying them with CBMC.
SourceDesigned and implemented a compiler for the COOL language to generate LLVM IR as part of the Advanced Compiler Design course project.
Implemented DPLL, CDCL, and MaxSAT solvers with various heuristics and techniques such as MOMS, lazy data structures, and watch literals as part of the Constraint Programming course project.
SourceImplemented a hybrid mutual exclusion algorithm for distributed systems by combining Raymond’s and Maekawa’s algorithms, multiplexing between them based on load, latency, and throughput.
SourceImplemented thin‑slicing in CBMC‑GOTO to aid debugging of large programs by presenting only relevant sections of code, improving focus and efficiency.
Developed a Bitcoin wallet application capable of creating and managing BTC addresses with support for single and multisig authorization. Completed for the Blockchain Theory & Practice course.
SourceUnited Kingdom