Bug Report: Same-frame renderer dependencies

[harrand]

Describe the bug Right now, there is no concept of a renderer dependency. In my mind, and documentation's mind, two renderers running in a single frame have no guarantees at all about which one runs first/order etc.

That is actually not the case. https://github.com/harrand/Topaz/blob/71fe2485b8af632e9fe2741992e37da32c05cb5e/src/tz/gl/impl/vulkan/renderer.cpp#L1102 This will CPU block for any previously-submitted work that happened this frame. That means for two composite renderers A and B. if A runs first, B will block the thread until A finishes.

Without dependency semantics this is sensible -- if we can't specify a dependency then the only way to be safe is if all composite renderers are dependent on the previous having ran to completion. However, this is dreadful for perf if the renderers are doing unrelated work and do not need to run one-after-the-other serially.

This work is blocked by Feature Request: Renderer Dependencies #74 as to make the change beforehand now means there is no safety in the case of composite renderers in a single frame (i.e all debugui will break, for starters)

Reproduction

• Repros On: anything on a vulkan build
• Repro Type: unexpected behaviour
• Repro Rate: every time
• Repro Severity (how badly does it affect your bottom line): high Repro Steps Steps to reproduce the bug:
  1. Invoke any composite renderers on any vulkan debug build.
  2. See unexpected behaviour - dbgui renderer will CPU wait on any previously-submitted CPU work. dbgui renderer has render_wait which will mean it blocks the thread until its own work is finished. So right now, it: waits for previous work -> submits -> waits for its own work. This is a potentially huge amount of unnecessary waiting. Note that for safety reasons, dbgui renderers should be dependent on anything that renders into the window, so this change needs to be careful to add some system to ensure that still happens.

Expected behavior If renderer A is invoked before unrelated renderer B, B should not wait on A to finish - unless B has specified a renderer dependency on A; something only possible using the feature provided by Feature Request: Renderer Dependencies #74

Screenshots If applicable, add screenshots to help explain your problem.

Conditions (please complete the following information):

• OS: windows
• Render API: vulkan
• Build Config: debug
• Version: master
• Compiler: gcc

Additional context N/A

[harrand]

As of Topaz 4.0, you are expected to populate the render graph and timeline, telling the underlying renderer the necessary information about how your renderers should be scheduled.

As a result of doing this, there is no longer this weird implicit wait dependency on any and all composite renderers, unless explicitly specified.