Skip to content

Latest commit

 

History

History
61 lines (53 loc) · 2.38 KB

vulkan_threading.md

File metadata and controls

61 lines (53 loc) · 2.38 KB
Raw

Threading in Vulkan

The Vulkan backend uses a dedicated concurrent worker pool that is created along with the creation of the Vulkan context.

Unlike other pools such as the IO worker pool, long running tasks may not be posted to this pool. This is because frame workloads can and often are distributed to workers in this pool. Having a potentially long running task (such as texture decompression) may potentially block frame critical tasks. The limitation of a separate pool for frame critical tasks is working around the separate limitation of not being able to specify a QoS to specific tasks and may be lifted in the future.

There is also a separate component called the fence waiter which operates on its own thread. The purpose of the fence waiter is to ensure that the resource reference count lives at least as long as the GPU command buffer(s) that access this resource.

Resource collection and pooling also happens on another thread called the resource manager. This is because touching the allocators is a potentially expensive operation and having the collection be done in a frame workload, or on the fence waiter thread may cause jank.

With this overview, the total number of thread used by the Impeller Vulkan backend is the number of workers in the concurrent worker pool, and the two threads for the fence waiter and resource manager respectively.

A summary of the interaction between the various threads is drawn below:

sequenceDiagram
    participant rt as Render Thread
    participant worker1 as Concurrent Worker 1
    participant worker2 as Concurrent Worker 2
    participant fence_waiter as Fence Waiter
    participant resource_manager as Resource Manager
    participant gpu as GPU
    rt->>+worker1: Setup PSO 1
    rt->>+worker2: Setup PSO n
    worker1-->>-rt: Done
    worker2-->>-rt: Done
    Note over rt,resource_manager: Application launch
    loop One Frame
        activate rt
        rt->>+worker2: Frame Workload
        activate fence_waiter
        rt->>fence_waiter: Resource 1 owned by GPU
        worker2-->>-rt: Done
        rt->>fence_waiter: Resource 2 owned by GPU
        rt->>gpu: Submit GPU Commands
        deactivate rt
    end
    activate gpu
    gpu-->>fence_waiter: GPU Work Done
    fence_waiter->>resource_manager: Collect/Pool Resources
    deactivate fence_waiter
    activate resource_manager
    deactivate gpu
    deactivate resource_manager
Loading