Both Immediate/AsyncCompute passes are always executed in order on the render thread but other RDG passes can run in parallel.
With my previous change, RayTracingScene pass was changed to Compute/AsyncCompute to have the correct order on the GPU.
But when it was set to Compute (when AsyncCompute was disabled) it was running in parallel with passes it depends on which led to random crashes.
Because the dependency between passes was implicit, we needed a new flag to express the intent and get the desired order between execution lambdas.
On behalf of Zach Bethel.
#rb zach.bethel
#jira UE-141003
#preflight 61fb0d9e923ac18db7015160
#lockdown michal.valient
#ROBOMERGE-AUTHOR: aleksander.netzel
#ROBOMERGE-SOURCE: CL 18841970 in //UE5/Release-5.0/... via CL 18841984 via CL 18842341
#ROBOMERGE-BOT: UE5 (Release-Engine-Test -> Main) (v910-18824042)
[CL 18842374 by aleksander netzel in ue5-main branch]
* Increased limit for GPU profiler events on D3D12. A typical frame with 4 view families would easily exhaust the old limit of 1K events resulting in many events not showing up in the profile -- limit increased to 8K, and confirmed negligible difference in memory. Fixed bug where we were only using half the queries allocated as well (unnecessary multiplication by 2 when calling RHICreateRenderQueryPool).
* Added optional "ProfileDescription" to FSceneViewFamily, which is appended to the Unreal Insights profile event names for both CPU and GPU, so you can differentiate view families.
* Infrastructure added to CPU profiler to allow a verbose name to be added to a Stat timing block.
* Infrastructure added to GPU profiler to allow verbose names to be added to timing blocks in general, both Event and Stat timing blocks.
* Added stats display for Display Cluster listing the CPU and GPU cost, and GPU assignment per view family by name, with the goal of allowing clients to tune features enabled per view family to adjust performance, without needing to gather performance information from Unreal Insights.
* FSceneRenderer::DoCrossGPUTransfers uses push transfer without lockstep (logic copied from 4.27), which significantly improves perf.
#rb zach.bethel marc.audy mihnea.balta
#jira none
#rnx
#preflight 61df4d23484d866ec01f17cf
#ROBOMERGE-AUTHOR: jason.hoerner
#ROBOMERGE-SOURCE: CL 18591295 in //UE5/Release-5.0/... via CL 18591308 via CL 18591324
#ROBOMERGE-BOT: STARSHIP (Release-Engine-Test -> Main) (v899-18417669)
[CL 18591346 by jason hoerner in ue5-main branch]
Add support for emitting named events from timing stats without compiling them in. This is controlled via new commandline options, which can be used in combination:
-csvNamedEvents emits named events for exclusive stats
-csvNamedEventsTiming emits named events for standard timing stats
Instrumentation overhead reduction
-Make it possible to disable CSVPROFILERTRACE_ENABLED, since this increases CSV profiler instrumentation overhead by about 15%
- Remove LLM scopes from stat functions and move into stat block allocation function. Reduces overhead slightly
#ROBOMERGE-AUTHOR: ben.woodhouse
#ROBOMERGE-SOURCE: CL 18253615 via CL 18253616 via CL 18253618 via CL 18253622 via CL 18253625 via CL 18254405 via CL 18254407
#ROBOMERGE-BOT: STARSHIP (Release-Engine-Staging -> Release-Engine-Test) (v895-18170469)
[CL 18254408 by ben woodhouse in ue5-release-engine-test branch]
This represents UE4/Main @17774255, Release-5.0 @17791557 and Dev-PerfTest @17789485
[CL 17794212 by aurel cordonnier in ue5-release-engine-test branch]
- Refactored RDG to support free-threaded execution of passes.
- Refactored renderer to use specific RHI command list variants in pass lambda. Immediate command list passes are forced to stay on the render thread, while other variants can be parallelized.
#rb christopher.waters
#ROBOMERGE-SOURCE: CL 16838717 in //UE5/Main/...
#ROBOMERGE-BOT: STARSHIP (Main -> Release-Engine-Test) (v836-16769935)
[CL 16838724 by zach bethel in ue5-release-engine-test branch]
- New Drain() method on FRDGBuilder; will flush all pending work.
- Drained passes are not culled; resource lifetimes are extended; async compute fences are optimized as best as possible but fence joining may occur after the drain.
- Batch up and pre-build all resource transitions. This is a prerequisite for parallel command lists.
- Removed ServiceLocalQueue passes with built-in RDG AddDispatchHint().
#jira UE-114622
[CL 16393495 by zach bethel in ue5-main branch]
- Simplified texture subresource tracking.
- Removed map lookup for each resource in SetupPass.
- Improved Compile / CollectPassResources to reduce cache misses.
- Added some container reservations to reduce reallocation costs.
- Added snapping of buffers to page boundaries to improve re-use.
#rb none
[CL 16208311 by zach bethel in ue5-main branch]
#rnx
#rb none
#ROBOMERGE-OWNER: ryan.durand
#ROBOMERGE-AUTHOR: ryan.durand
#ROBOMERGE-SOURCE: CL 10869210 via CL 10869511 via CL 10869900
#ROBOMERGE-BOT: (v613-10869866)
[CL 10870549 by ryan durand in Main branch]
