Dayananda Sagar College of EngineeringJul 2024 to Jun 2028
B.E., Information Science and Engineering · GPA 9.8/10 · Bengaluru, India
Shreyas Hegde
Compiler enthusiast, systems geek.
I build compilers, GPU runtimes, and low-level tools, currently deep in LLVM, heterogeneous dispatch, and making GPU kernels fast in Julia.
GSoC'26 @ JuliaGPUResearch Intern @ IIT KanpurContributor @ GCC Rust
Experience
JuliaGPU: GSoC'26 Mentee / MaintainerMay 2026 to Present
Remote
- Maintainer of JuliaGPU, an organisation implementing vendor-specific and vendor-agnostic GPU/GPGPU kernels in Julia.
- Google Summer of Code mentee adding vendor-agnostic kernels in AcceleratedKernels.jl.
- Optimising KernelAbstractions.jl so vendor-agnostic kernels nearly match vendor-specific performance.
Indian Institute of Technology, Kanpur: Research InternMay 2025 to Present
Hybrid
- Heterogeneous dispatch systems for CPU / GPU / QPU.
- Cache warming and prediction-based dispatch.
Projects
A few things I've built or actively contribute to, pulled straight from GitHub with live stars.
SecureWipe46
pointblank-clubCross-platform, forensic-grade secure data wiping tool, built as an ISO on Tiny Core Linux. Implements device-level wipe workflows (ATA Secure Erase, NVMe Sanitize, SCSI SANITIZE, ADB/Fastboot) with verification to guarantee irrecoverable deletion.
C++
AcceleratedKernels.jl202
JuliaGPUCross-architecture parallel algorithms for Julia's CPU and GPU backends, targeting multithreaded CPUs and GPUs via Intel oneAPI, AMD ROCm, Apple Metal, and Nvidia CUDA. I contribute vendor-agnostic kernels here as part of GSoC'26.
Julia
Compromyler4
vishruth-thimmaiah/compromylrA tri-layer binary analysis system that detects compromised compilers, inspired by Ken Thompson's "Reflections on Trusting Trust." Uses Diverse Double Compilation, CFG similarity (Ghidra + BinDiff), and ptrace-based syscall tracing. 1st place, Aegis Sandbox 2.0.
C++
SiMG1
ShreyashSriA cross-platform Electron app that cryptographically verifies medical DICOM images before they reach an AI inference model, catching supply-chain attacks, adversarial converter compromises, and pipeline failures. Runner-up, HackNockturne 2.0.
TypeScriptOpen Source Contributions
JuliaGPU/juliagpu.orgAdd GSoC 2026 report: portable data-parallel primitivesmergedJuliaGPU/AcceleratedKernels.jlLet map and map! take several source arraysmergedJuliaGPU/AcceleratedKernels.jlSort along dims with RadixSortmergedJuliaGPU/GPUArrays.jlfeat: sort / sort! delegated to AcceleratedKernelsclosedJuliaGPU/GPUArrays.jlfeat: Add accumulate, cumsum and cumprod using AcceleratedKernelsclosedJuliaGPU/GPUArrays.jlfeat: Adding Base.reverse / reverse! support using AcceleratedKernels.jlclosedJuliaGPU/GPUArrays.jlAdd a sorting interfaceclosedJuliaGPU/AcceleratedKernels.jlAdd BitonicSort, a GPU sorting network for small arrays and short slicesmergedJuliaGPU/AcceleratedKernels.jlperf(sort): packed-key fast path for sort(A; dims=1)closedJuliaGPU/AMDGPU.jlGuard empty-array launch in out-of-place reversemergedJuliaGPU/AcceleratedKernels.jlAdd start/stop keywords to reverse and reverse!closedJuliaGPU/AcceleratedKernels.jlperf(reverse): use a dedicated kernel for the dims reversalclosedJuliaGPU/AcceleratedKernels.jlAdd dims support to sort, sort!, sortperm and sortperm!mergedJuliaGPU/AcceleratedKernels.jlFix DecoupledLookback with a device-scope memory fencemergedJuliaGPU/AcceleratedKernels.jlfeat(findall): Add findall kernelmergedJuliaGPU/AcceleratedKernels.jlAdd dims support to reverse and reverse!mergedJuliaGPU/AcceleratedKernels.jlfix(accumulate): keep GPU scans uniform and non-divergent across backendsclosedJuliaGPU/AcceleratedKernels.jlMake GPU scans process multiple items per threadmergedJuliaGPU/AcceleratedKernels.jlMake GPU reductions process multiple items per threadmergedJuliaGPU/AcceleratedKernels.jlReduce: vectorize contiguous by-block loadsmergedJuliaGPU/AcceleratedKernels.jlci(opencl): run POCL under --check-bounds=auto; skip scan on POCLclosedJuliaGPU/AcceleratedKernels.jlFix OpenCL/POCL CI: run under --check-bounds=auto, skip scan on POCLclosedJuliaGPU/AcceleratedKernels.jlAdd reverse! and reversemergedJuliaGPU/AcceleratedKernels.jlFix DecoupledLookback cross-block coherence (completes #91)closedJuliaGPU/AcceleratedKernels.jlOptimize Radix sortmergedJuliaGPU/AcceleratedKernels.jlOptimize GPU radix sort: ballot kernels, fused range, skip-pass, tuningclosedJuliaGPU/KernelAbstractions.jlfeat(intrinsics): add KI.vload / KI.vstore! for wide vector memory operationsclosedJuliaGPU/AcceleratedKernels.jlAdd opt-in GPU radix sort via sort alg keywordmergedJuliaGPU/GPUArrays.jlDelegate mapreducedim! to AcceleratedKernels.jlclosedJuliaGPU/AcceleratedKernels.jlExpand dimensional mapreduce / reducemergedJuliaGPU/cuTile.jlSubtype AbstractArray for TileArrayclosedJuliaGPU/GPUArrays.jlfeat: add GPU-native kron support for Diagonal matricesmergedJuliaGPU/cuTile.jlAdd alias-aware token threading for memory operationsmergedRust-GCC/gccrsgccrs: avoid ICE when canonical path record is missingopenJuliaGPU/GPUArrays.jlSpecialize ReshapedArray to resolve setindex! ambiguitiesmergedJuliaGPU/GPUArrays.jlfeat: Implement issorted for AbstractGPUArray without scalar indexingmergedJuliaLang/juliaFix OutOfMemory in arrayshow with unsigned indicesopenEmscripten/emscripten[memoryprofiler] Add CSS class to parent divmergedRust-GCC/gccrsgccrs: Fix ICE in no input fileopenRust-GCC/gccrsgcc: Prevent ICE on no input fileclosed
Posts
Writing
dev.toOct 2026
Google Summer of Code 2026, a review back
A look back at the GSoC'26 project with JuliaGPU: the proposal process, working with mentors, the sorting and reduction kernels that shipped, and advice for getting into the program.
juliagpu.orgOct 2026
GSoC 2026: Portable data-parallel primitives for the Julia GPU stack
The GSoC'26 final report: reverse, findall, scans, reductions, sorting, and multi-source map added to AcceleratedKernels.jl so every Julia GPU backend shares one implementation, benchmarked on an RTX 5080 against the previous release and CUDA.jl.
Talks
Achievements
Selected: Google Summer of Code 2026
The Julia Language.
1st place: Aegis Sandbox 2.0
Designed a compiler that injects vulnerabilities into every binary it compiles, plus a tool to test the integrity of the compiler.
Runner-up: HackNockturne 2.0
A solution to safely process DICOM files from medical imaging devices using SiMG, a custom binary, for comparison and processing.
Skills
Languages
Compilers
Tooling
Education
Contact
Want to build something together?
shreyashegde@acm.org