GPU Compute & MLIR Compiler Engineer — AI Workloads

Company: Qualcomm India Private Limited Job Area: Engineering Group, Engineering Group > Systems Engineering General Summary: We are looking for a highly skilled GPU compute, MLIR compiler, and kernel optimization engineer with deep expertise in GPU compute, MLIR-based code generation, and end-to-end performance optimization for AI workloads. In this role, you will design, optimize, and deploy high-performance GPU compute kernels, build and extend MLIR compiler backends, and collaborate closely with ML, runtime, and hardware teams to push the limits of performance on modern GPU architectures.

Develop and optimize GPU compute kernels targeting OpenCL and Vulkan compute backends for high-throughput AI/ML workloads. Design, build, and extend MLIR dialects across multiple abstraction levels — including frontend dialects, graph-level IR, tensor IR (e.g., Linalg, Tensor, Tosa), and runtime/low-level dialects — to enable efficient end-to-end model compilation. Implement and maintain MLIR-based compiler passes and transformations, including tiling, fusion, bufferization, vectorization, and lowering pipelines targeting OpenCL and Vulkan GPU backends. Conduct profiling and bottleneck analysis of compiled kernels using GPU counters and vendor-specific profilers, and drive performance improvements through compiler-level optimizations. Build and maintain GPU runtime infrastructure for both OpenCL and Vulkan, including memory management, pipeline setup, command buffer orchestration, and resource scheduling. Develop and extend code generation pipelines, enabling automatic lowering from tensor IR through MLIR to efficient OpenCL and Vulkan GPU kernels. Implement performance-critical schedules — tiling, loop fusion, parallelism, and caching strategies — within MLIR-based backends targeting OpenCL and Vulkan runtimes. Collaborate with framework teams to optimize end-to-end model lowering for computer vision and LLM workloads using MLIR compilation stacks. Design and implement robust compiler and runtime components using modern C/C++, leveraging advanced programming paradigms for high-performance systems.

Strong hands-on experience with MLIR framework, including authoring and extending custom dialects, writing compiler passes, and building end-to-end lowering pipelines. Deep expertise across MLIR abstraction levels: Frontend dialects — ingestion and representation of ML models (e.g., TOSA, StableHLO, ONNX-MLIR) Graph-level IR — high-level operation fusion, shape inference, and graph transformations Tensor IR level — structured op representation using Linalg, Tensor, and Vector dialects; tiling and fusion strategies Runtime/low-level dialects — bufferization, MemRef, SCF, GPU, and LLVM dialects for final code generation Strong hands-on experience in OpenCL programming, including kernel development, memory model, work-group/work-item optimization, and OpenCL runtime management. Solid understanding of Vulkan compute programming, including descriptor management, compute pipelines, synchronization primitives, and Vulkan runtime internals. Strong understanding of GPU architecture, memory hierarchies, and async compute. Proficiency in C/C++ for system-level development. Experience with kernel profiling and bottleneck analysis on GPU platforms. Strong background in machine learning fundamentals, covering both CV and LLM workloads. Good to Have Hands-on experience with IREE (Intermediate Representation Execution Environment) or other MLIR-based deployment frameworks such as TVM, XLA, or LLVM. Familiarity with IREE's compiler and runtime architecture — including HAL (Hardware Abstraction Layer), executable compilation, and dispatch mechanisms — particularly its OpenCL and Vulkan HAL backends. Experience contributing to open-source MLIR or IREE projects. Knowledge of quantization, mixed-precision inference, and model optimization techniques for edge and server GPU targets.