Concept: Async Compute

• Parent: PathTracer Learning - Phase 3 - GPU and Vulkan

What Is Async Compute?

• Running compute work in parallel with graphics work on the GPU
• Modern GPUs have separate compute queues that can run alongside the graphics queue
• Enables better GPU utilization by filling idle shader units

GPU Queue Types

• Graphics queue: supports all operations (graphics, compute, transfer)
• Compute queue: compute + transfer only (no rasterization)
• Transfer queue: DMA transfers only
• Multiple queues can run simultaneously on different hardware units

Why It Matters for Path Tracing

• BLAS builds are compute-heavy — can overlap with rendering
• Denoising passes can overlap with next frame’s ray tracing
• TLAS rebuild can overlap with shadow ray tracing
• Typical frame timeline without async
• With async compute

Vulkan Async Compute Setup

• Find a compute-only queue family

  for (auto& queueFamily : queueFamilies) {
      if ((queueFamily.queueFlags & VK_QUEUE_COMPUTE_BIT) &&
          !(queueFamily.queueFlags & VK_QUEUE_GRAPHICS_BIT)) {
          computeQueueFamily = index;
      }
  }

• Create separate command pools and queues for compute
• Submit compute work to compute queue, graphics to graphics queue

Synchronization

• Async compute requires careful synchronization
• Timeline semaphores (Vulkan 1.2) — preferred

  VkSemaphoreTypeCreateInfo typeInfo{};
  typeInfo.semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE;
  typeInfo.initialValue = 0;

  • Signal from compute queue, wait on graphics queue
• Pipeline barriers within a queue
• Queue ownership transfers for shared resources

Practical Considerations

• Not all GPUs benefit equally
  • Integrated GPUs: often single queue, no benefit
  • Discrete GPUs: multiple compute units, significant benefit
• Overhead: synchronization adds complexity and some latency
• Profile first: measure actual GPU utilization before optimizing
• NVIDIA NSight, AMD RGP — tools for visualizing queue utilization

In Godot Context

• Godot’s RenderingDevice exposes compute queues
• BLAS builds for skinned meshes are good candidates for async
• Denoising (OIDN compute) can run async with next frame’s RT

Related

• PathTracer Learning - Phase 3 - GPU and Vulkan
• PathTracer Learning - BLAS and TLAS