Command Data Flow

This part is going to help you to form an idea of how GL commands work.

You can think of a queue of commands to be executed for both CPU and GPU.

Notice:
OpenGL doesn’t expose queues like Vulkan. This is only meant to help you understand.

graph TD
    A[CPU Queue]
    A --> A1[A_1]
    A1 --> A2[A_2]
    A2 --> A3[A_3]
    A3 --> A4[...]
    B[GPU Queue]
    B --> B1[B_1]
    B1 --> B2[B_2]
    B2 --> B3[B_3]
    B3 --> B4[...]

Quick mental model check:

• Can I expect B_2 to perceive A_1? like, can I access data from A_1 safely when running B_2?

  No. A_1 may run after B_2; the data is not safe.

• CPU commands come from Java code, but where do GPU commands come from?

  Java code still. Your GL code like GL11.doSomething() to be exact.

• Code snippet:

  GL11.draw();
  func();

  What does GL11.draw() do? Is it going to be executed before func() or after?

  GL11.draw() adds a GPU-side command to the command buffer (or say the GPU queue)
  The action of enqueueing a command is going to be executed before func(), but the actual GPU-side action can be executed before OR after func()

Key point:
CPU and GPU timing mismatches.

However, at the end of every frame, a special command swap is going to be executed, which acts as an implicit sync point. Note that latency can still accumulate.

Command swap itself moves content from framebuffer A to framebuffer B (not necessarily memory copy) and display it to the screen, so we are free to write to framebuffer A while displaying B (not necessarily immediately displaying B; controlled by vsync).

Quick mental model check:

• Code snippet:

  GL11.drawA();
  GL11.drawB();

  Is the actual draw command B going to be run after A?

  Yes. See below

graph TD
   B[GPU Queue]
   B --> B1[Draw A]
   B1 --> B2[Draw B]

• Code snippet:

  GL11.computeA(); // compute some data on GPU and store in buffer A
  GL11.computeB(); // read buffer A

  Is the data safe?

  Unfortunately, no. You must insert a GL barrier to guarantee the data visibility. Although the GPU command order is respected, GPU execution order isn't equivalent to memory visibility.
  
  Note:
  Memory barrier is not a global synchronization indicator.

Contributors

Nullpinter

Last edited about 1 hour ago

View full history

16cf2d7-docs(en, wiki): add renderbook metadata

7ba1bad-docs(en, wiki): add renderbook from https://github.com/tttsaurus/Mc122RenderBook