← Projects

Compute Systems Engineer

Vulkan Compute Accelerator

Compute Systems Engineer · 2023

VulkanComputeC++GPU
API
Vulkan 1.2
Architecture
Multi-queue, sparse memory
Pattern
Flowgraph abstraction
  • Vulkan 1.2 compute pipeline orchestrator using a flowgraph model: nodes are shader dispatches, edges are data dependencies, the scheduler resolves execution order and resource hazards automatically
  • Multi-queue dispatch with batching optimization and sparse memory management for large intermediate buffers
  • Synchronization on Vulkan timeline semaphores for fine-grained pipeline dependencies without over-synchronization; complex multi-stage workloads need no manual barrier insertion per shader
Outcome Measurable GPU throughput improvements through careful synchronization design and queue utilization.

A Vulkan 1.2 compute orchestrator that runs flows of prewritten compute shaders as composable pipelines. The core abstraction is a flowgraph: nodes represent shader dispatches, edges define data dependencies, and the scheduler resolves execution order and resource hazards automatically.

Multi-queue dispatch lets independent shader stages run concurrently across compute queues. Sparse memory reduces VRAM overhead for large intermediate buffers. Synchronization uses Vulkan timeline semaphores for fine-grained pipeline dependencies without over-synchronization, so complex multi-stage compute workloads need no manual barrier insertion per shader.