GPU driven rendering

[Try]

Goal of this ticket to prototype, implement and test GPU-driven rendering.

GPU driven allowed for better quality of invisible geometry culling, and can potentially reduce CPU load.

Current idea:

1. Take task/mesh shader as baseline
2. Test on native hardware
3. Implement backport for Vulkan1.0 hardware, if idea works

Implementation detail (so far):

• [x] Pack object into instanced-calls (full bucket for task-capable, only visible for everyone else)
• [x] Repack landscape into clusters of 64 vertices + 126 primitives (NVIDIA guideline)
• [x] Use bounding sphere(one per meshlet) to cull geometry in main pass
• [ ] ~Add visibility cone culling for shadow-pass~
• [x] HiZ

Baseline:

Main-pass 1.51ms Shadows 0.35 + 0.85 RayTracing pass 2.27 ms

[Try]

Meshlet packing:

Note: non-landscape objects are black

[Try]

Main-pass 1.81 ms Shadows 0.32 + 0.71 Drawcall count decreased by ~700 calls.

With meshshader execution barrier becomes an issue: shadow -> main pass barrier takes 1ms For now works slightly slower than CPU culling,

[Try]

Note: It seems gl_PrimitiveIndicesNV can be limiting factor in a way:

In spec gl_PrimitiveIndicesNV[] must have size that matches exactly max_primitives*triangles, and unfortunately never more.

For mesh block-compression scheme it world be way better to have out-of-bounds access on this array(or option to redeclarate it with bigger size)

[Try]

With pack-algorithm fixup: Main-pass 1.36 ms Shadows 0.34 + 0.67

[Try]

Some progress on task shader:

Main pass actually became slower, but good part is draw-call count: now whole frame takes 85 draw-calls instead of 2.3k.

[Coockie]

Is this a optimisation only for newer hardware like Nvidia Turing and Ampere and AMD RDNA2?

[Try]

@Coockie this iteration - yes only for latest hardware. Once culling algorithm is more-less complete and numbers are good we will expand to older hardware.

The way I'm planning to backport mesh shaders to Vulkan1.0 is based on this idea: https://tellusim.com/mesh-shader-emulation/

Basically set of compute-shaders + draw-indirect + software binner

[Try]

Who needs task shading? Just draw more meshlets:

So far best iteration:

• 85 drawcalls in total
• main-pass 1.32
• shadow1 - 0.67
• shadow0 - 0.30

[Try]

Add meshlets for skinned meshes:

Not great, not worse, timings, except for draw-call count. Basically looking forward to HiZ support

[Try]

Initial HiZ: only 32x32 tile size generated, with no mip-map chain for now Main-pass: 1.1

[Try]

Progress in HiZ. Overdraw:

• Without:

• With:

Small waley behind the farm(center) and stuff occluded by mountain(left) was removed

Still there is performance issue: reading HiZ with textureGather(mip=0)/texelFetch*4 is not performant.

[Try]

Wrapping up for now: System-trace profiler shows very odd numbers: rendering is faster, but barriers are terribly large (~1ms see above). Frame profiler shows else picture. According to frame-profiler meshlets 12% faster.

In the end of the day: base-fps = 170, mesh fps = 204. So meshlets are good.

Meshlets: Drawcalls:

	Meshlets	Drawcalls
Fps	204	170
Triangles	102316	3184884
Drawcalls	19	4766

Sys-trace: (Note: no idea what nsight displays as draw-call count. it's definitely not 19. Probably it's pipeline switches)

[Try]

Overdraw visualization: Baseline: Meshlets: