EEVEE-Next: Performance Analysis #117246

New Issue

Miguel Pozo · 2024-01-17T18:35:09+01:00

Miguel Pozo commented

2024-01-17 18:35:09 +01:00

To compare EEVEE Legacy and EEVEE Next and look for bottlenecks, I’ve made a performance test with the Wanderer scene using #116304.
The camera is static and pointing slightly downwards so the full view is covered by geometry.

Results

EEVEE Legacy

 Group                          | GPU  | CPU  | Latency
--------------------------------|------|------|--------
 Total                          | 8.37 | 3.02 |
. Manager.end_sync              | 0.02 | 0.06 | 1.68
. EEVEE                         | 7.95 | 2.02 | 1.62
.. Probes Refresh               | 0.00 | 0.01 | 1.61
.. Shadows                      | 0.00 | 0.05 | 1.58
... Cube Shadow Maps            | 0.00 | 0.01 | 1.58
... Cascaded Shadow Maps        | 0.00 | 0.01 | 1.55
.. Prepass                      | 1.36 | 0.38 | 1.51
... psl->depth_ps               | 1.36 | 0.34 | 1.48
.. Main MinMax buffer           | 0.10 | 0.24 | 2.48
... Max buffer                  | 0.10 | 0.21 | 2.47
.... psl->maxz_copydepth_ps     | 0.03 | 0.02 | 2.45
.... psl->maxz_downlevel_ps     | 0.02 | 0.01 | 2.44
.... psl->maxz_downlevel_ps     | 0.01 | 0.01 | 2.44
.... psl->maxz_downlevel_ps     | 0.00 | 0.01 | 2.43
.... psl->maxz_downlevel_ps     | 0.00 | 0.01 | 2.41
.... psl->maxz_downlevel_ps     | 0.00 | 0.01 | 2.39
.... psl->maxz_downlevel_ps     | 0.00 | 0.01 | 2.37
.. GTAO Horizon Scan            | 0.67 | 0.04 | 2.34
... psl->ao_horizon_search      | 0.66 | 0.01 | 2.32
.. Shading                      | 4.90 | 0.43 | 2.95
... psl->background_ps          | 0.01 | 0.02 | 2.94
... psl->material_ps            | 4.88 | 0.38 | 2.93
.. SSR                          | 0.60 | 0.32 | 7.42
... psl->ssr_raytrace           | 0.10 | 0.01 | 7.40
... Downsample Radiance         | 0.12 | 0.22 | 7.48
.... psl->color_copy_ps         | 0.04 | 0.01 | 7.46
.... psl->color_downsample_ps   | 0.02 | 0.01 | 7.48
.... psl->color_downsample_ps   | 0.00 | 0.01 | 7.48
.... psl->color_downsample_ps   | 0.00 | 0.01 | 7.47
.... psl->color_downsample_ps   | 0.00 | 0.01 | 7.42
.... psl->color_downsample_ps   | 0.00 | 0.01 | 7.40
.... psl->color_downsample_ps   | 0.00 | 0.01 | 7.39
... psl->ssr_resolve            | 0.37 | 0.02 | 7.36
.. psl->probe_display           | 0.00 | 0.01 | 7.69
.. Opaque Refraction            | 0.00 | 0.01 | 7.67
.. Post FX                      | 0.21 | 0.35 | 7.65
... Bloom Blit                  | 0.04 | 0.02 | 7.63
... Bloom Downsample First      | 0.01 | 0.01 | 7.64
... Bloom Downsample            | 0.00 | 0.01 | 7.62
... Bloom Downsample            | 0.00 | 0.01 | 7.60
... Bloom Downsample            | 0.00 | 0.01 | 7.57
... Bloom Downsample            | 0.00 | 0.01 | 7.55
... Bloom Upsample              | 0.00 | 0.02 | 7.52
... Bloom Upsample              | 0.00 | 0.01 | 7.49
... Bloom Upsample              | 0.00 | 0.01 | 7.47
... Bloom Upsample              | 0.02 | 0.01 | 7.46
... Bloom Resolve               | 0.08 | 0.01 | 7.46
. Overlay                       | 0.22 | 0.66 | 7.54
.. psl->background_ps           | 0.03 | 0.01 | 7.52
.. psl->fade_ps[i]              | 0.00 | 0.01 | 7.51
.. psl->facing_ps[i]            | 0.00 | 0.01 | 7.49
.. psl->extra_blend_ps          | 0.00 | 0.01 | 7.47
.. psl->wireframe_ps            | 0.00 | 0.01 | 7.45
.. *p_armature_trans_ps         | 0.00 | 0.01 | 7.43
.. *p_armature_ps               | 0.00 | 0.02 | 7.40
.. psl->particle_ps             | 0.00 | 0.01 | 7.37
.. psl->metaball_ps[i]          | 0.00 | 0.01 | 7.35
.. *p_extra_ps                  | 0.00 | 0.08 | 7.33
.. psl->attribute_ps            | 0.00 | 0.01 | 7.25
.. psl->grid_ps                 | 0.04 | 0.01 | 7.22
.. psl->fade_ps[i]              | 0.00 | 0.01 | 7.23
.. psl->facing_ps[i]            | 0.00 | 0.01 | 7.22
.. *p_armature_trans_ps         | 0.00 | 0.01 | 7.19
.. *p_armature_ps               | 0.00 | 0.01 | 7.17
.. *p_extra_ps                  | 0.00 | 0.02 | 7.15
.. psl->metaball_ps[i]          | 0.00 | 0.01 | 7.11
.. psl->motion_paths_ps         | 0.00 | 0.01 | 7.09
.. psl->extra_grid_ps           | 0.00 | 0.01 | 7.07
.. psl->extra_centers_ps        | 0.00 | 0.01 | 7.05
.. psl->antialiasing_ps         | 0.10 | 0.01 | 7.03
. RegionInfo                    | 0.03 | 0.06 | 7.06

 Group                          | GPU  | CPU  | Latency
--------------------------------|------|------|--------
 Total                          | 12.2 | 6.86 |
. SPACE_TOPBAR                  | 0.00 | 0.02 | 0.25
. SPACE_STATUSBAR               | 0.00 | 0.00 | 0.21
. SPACE_VIEW3D                  | 11.9 | 6.67 | 0.19
.. Viewport                     | 11.9 | 6.65 | 0.19
. Window Redraw                 | 0.30 | 0.11 | 5.43

EEVEE Next

 Group                          | GPU  | CPU  | Latency
--------------------------------|------|------|--------
 Total                          | 21.9 | 5.80 |
. Manager.end_sync              | 0.06 | 0.14 | 2.32
. EEVEE                         | 21.5 | 4.58 | 2.18
.. negZ_view                    | 21.5 | 4.40 | 2.16
... LightCulling                | 0.04 | 0.10 | 2.13
.... Select                     | 0.00 | 0.01 | 2.12
.... Sort                       | 0.00 | 0.01 | 2.10
.... Zbin                       | 0.01 | 0.01 | 2.08
.... Tiles                      | 0.01 | 0.01 | 2.08
... Probe.Select                | 0.00 | 0.01 | 2.06
... World.Background            | 1.03 | 0.03 | 2.03
... Deferred.Opaque             | 18.9 | 2.93 | 3.03
.... View.compute_visibility    | 0.01 | 0.02 | 3.01
.... DrawMultiBuf.bind          | 0.01 | 0.03 | 2.99
.... Prepass                    | 2.24 | 0.36 | 2.96
..... DoubleSided.Static        | 2.21 | 0.27 | 2.95
..... SingleSided.Static        | 0.00 | 0.01 | 4.88
..... DoubleSided.Moving        | 0.02 | 0.01 | 4.86
..... SingleSided.Moving        | 0.00 | 0.00 | 4.86
.... HizUpdate                  | 0.06 | 0.03 | 4.83
.... Shadow                     | 5.55 | 0.80 | 4.86
..... TilemapSetup              | 0.02 | 0.09 | 4.85
...... ClearClipmap             | 0.00 | 0.01 | 4.84
...... DirectionalBounds        | 0.00 | 0.01 | 4.82
...... Init                     | 0.00 | 0.01 | 4.80
...... CasterUpdate             | 0.00 | 0.01 | 4.79
..... View.compute_visibility   | 0.01 | 0.01 | 4.76
..... DrawMultiBuf.bind         | 0.00 | 0.01 | 4.76
..... TagUsage                  | 0.13 | 0.06 | 4.74
...... Opaque                   | 0.13 | 0.01 | 4.73
...... Transparent              | 0.00 | 0.02 | 4.83
..... TilemapUpdate             | 0.12 | 0.15 | 4.80
...... MaskLod                  | 0.00 | 0.01 | 4.79
...... Free                     | 0.00 | 0.01 | 4.77
...... Defrag                   | 0.00 | 0.01 | 4.76
...... AllocatePages            | 0.00 | 0.01 | 4.74
...... Finalize                 | 0.01 | 0.01 | 4.73
...... RenderClear              | 0.07 | 0.01 | 4.72
..... View.compute_procedural_bounds | 0.01 | 0.01 | 4.77
..... View.compute_visibility   | 0.03 | 0.01 | 4.76
..... DrawMultiBuf.bind         | 0.01 | 0.02 | 4.78
..... Shadow.Surface            | 5.17 | 0.31 | 4.76
...... Shadow.Surface           | 5.17 | 0.30 | 4.75
....... Shadow.Surface.Double-Sided | 5.17 | 0.25 | 4.74
....... Shadow.Surface.Single-Sided | 0.00 | 0.01 | 9.64
.... View.compute_visibility    | 0.01 | 0.01 | 9.58
.... DrawMultiBuf.bind          | 0.01 | 0.02 | 9.58
.... Shading                    | 5.08 | 0.41 | 9.55
..... DoubleSided               | 0.00 | 0.00 | 9.54
..... SingleSided               | 0.00 | 0.01 | 9.53
..... DoubleSided               | 5.00 | 0.30 | 9.51
..... SingleSided               | 0.00 | 0.01 | 14.2
..... StencilClassify           | 0.07 | 0.02 | 14.1
.... Raytracing                 | 4.35 | 0.93 | 14.2
..... HorizonScan.Setup         | 0.07 | 0.01 | 14.2
..... TileClassify              | 0.08 | 0.01 | 14.2
..... Raytracing                | 2.47 | 0.23 | 14.3
...... TileCompact              | 0.01 | 0.01 | 14.2
...... RayGenerate              | 0.01 | 0.01 | 14.2
...... Trace.Screen             | 0.01 | 0.01 | 14.2
...... DenoiseSpatial           | 0.05 | 0.01 | 14.2
...... DenoiseTemporal          | 0.03 | 0.02 | 14.2
...... DenoiseBilateral         | 0.01 | 0.01 | 14.2
...... HorizonScan.Trace        | 1.63 | 0.01 | 14.2
...... HorizonScan.Denoise      | 0.69 | 0.01 | 15.8
..... Raytracing                | 1.65 | 0.38 | 16.5
...... TileCompact              | 0.01 | 0.01 | 16.5
...... RayGenerate              | 0.01 | 0.02 | 16.5
...... Trace.Screen             | 0.01 | 0.01 | 16.4
...... DenoiseSpatial           | 0.04 | 0.02 | 16.4
...... DenoiseTemporal          | 0.02 | 0.05 | 16.4
...... DenoiseBilateral         | 0.01 | 0.04 | 16.3
...... HorizonScan.Trace        | 1.06 | 0.01 | 16.3
...... HorizonScan.Denoise      | 0.46 | 0.02 | 17.3
..... Raytracing                | 0.06 | 0.22 | 17.7
...... TileCompact              | 0.01 | 0.01 | 17.7
...... RayGenerate              | 0.00 | 0.01 | 17.7
...... Trace.Screen             | 0.00 | 0.01 | 17.7
...... DenoiseSpatial           | 0.00 | 0.02 | 17.7
...... DenoiseTemporal          | 0.00 | 0.01 | 17.7
...... DenoiseBilateral         | 0.00 | 0.01 | 17.6
...... HorizonScan.Trace        | 0.00 | 0.01 | 17.6
...... HorizonScan.Denoise      | 0.00 | 0.01 | 17.6
.... EvalLights                 | 1.34 | 0.06 | 17.6
..... Eval.Light                | 1.34 | 0.04 | 17.5
.... Combine                    | 0.11 | 0.01 | 18.8
... Deferred.Refract            | 0.00 | 0.01 | 18.9
... Forward.Opaque              | 1.31 | 1.06 | 18.9
.... View.compute_visibility    | 0.01 | 0.01 | 18.9
.... DrawMultiBuf.bind          | 0.00 | 0.01 | 18.9
.... Prepass                    | 0.00 | 0.08 | 18.9
..... DoubleSided.Static        | 0.00 | 0.01 | 18.9
..... SingleSided.Static        | 0.00 | 0.01 | 18.9
..... DoubleSided.Moving        | 0.00 | 0.01 | 18.8
..... SingleSided.Moving        | 0.00 | 0.00 | 18.8
.... HizUpdate                  | 0.06 | 0.01 | 18.8
.... Shadow                     | 1.20 | 0.75 | 18.8
..... TilemapSetup              | 0.02 | 0.09 | 18.8
...... ClearClipmap             | 0.00 | 0.01 | 18.8
...... DirectionalBounds        | 0.00 | 0.01 | 18.8
...... Init                     | 0.00 | 0.01 | 18.8
...... CasterUpdate             | 0.00 | 0.01 | 18.8
..... View.compute_visibility   | 0.01 | 0.01 | 18.7
..... DrawMultiBuf.bind         | 0.00 | 0.01 | 18.7
..... TagUsage                  | 0.13 | 0.05 | 18.7
...... Opaque                   | 0.12 | 0.01 | 18.7
...... Transparent              | 0.00 | 0.01 | 18.8
..... TilemapUpdate             | 0.04 | 0.14 | 18.8
...... MaskLod                  | 0.00 | 0.01 | 18.8
...... Free                     | 0.00 | 0.01 | 18.8
...... Defrag                   | 0.00 | 0.01 | 18.7
...... AllocatePages            | 0.00 | 0.01 | 18.7
...... Finalize                 | 0.01 | 0.01 | 18.7
...... RenderClear              | 0.00 | 0.01 | 18.7
..... View.compute_procedural_bounds | 0.01 | 0.01 | 18.7
..... View.compute_visibility   | 0.03 | 0.01 | 18.7
..... DrawMultiBuf.bind         | 0.01 | 0.02 | 18.7
..... Shadow.Surface            | 0.91 | 0.28 | 18.7
...... Shadow.Surface           | 0.91 | 0.26 | 18.7
....... Shadow.Surface.Double-Sided | 0.90 | 0.22 | 18.6
....... Shadow.Surface.Single-Sided | 0.00 | 0.01 | 19.3
.... View.compute_visibility    | 0.01 | 0.01 | 19.3
.... DrawMultiBuf.bind          | 0.00 | 0.01 | 19.3
.... Shading                    | 0.00 | 0.04 | 19.2
..... SingleSided               | 0.00 | 0.00 | 19.2
..... DoubleSided               | 0.00 | 0.00 | 19.2
... View.compute_visibility     | 0.01 | 0.01 | 19.2
... DrawMultiBuf.bind           | 0.00 | 0.01 | 19.2
... Forward.Transparent         | 0.00 | 0.02 | 19.1
.... ResourceBind               | 0.00 | 0.01 | 19.1
... Film.Accumulate             | 0.18 | 0.02 | 19.1
.. Velocity Copy Pass           | 0.00 | 0.01 | 19.2
. Overlay                       | 0.23 | 0.80 | 19.1
.. psl->background_ps           | 0.04 | 0.02 | 19.1
.. psl->fade_ps[i]              | 0.00 | 0.01 | 19.1
.. psl->facing_ps[i]            | 0.00 | 0.01 | 19.1
.. psl->extra_blend_ps          | 0.00 | 0.01 | 19.0
.. psl->wireframe_ps            | 0.00 | 0.02 | 19.0
.. *p_armature_trans_ps         | 0.00 | 0.01 | 19.0
.. *p_armature_ps               | 0.00 | 0.02 | 19.0
.. psl->particle_ps             | 0.00 | 0.01 | 18.9
.. psl->metaball_ps[i]          | 0.00 | 0.01 | 18.9
.. *p_extra_ps                  | 0.00 | 0.09 | 18.9
.. psl->attribute_ps            | 0.00 | 0.01 | 18.8
.. psl->grid_ps                 | 0.04 | 0.02 | 18.8
.. psl->fade_ps[i]              | 0.00 | 0.02 | 18.8
.. psl->facing_ps[i]            | 0.00 | 0.01 | 18.7
.. *p_armature_trans_ps         | 0.00 | 0.01 | 18.7
.. *p_armature_ps               | 0.00 | 0.02 | 18.7
.. *p_extra_ps                  | 0.00 | 0.01 | 18.6
.. psl->metaball_ps[i]          | 0.00 | 0.01 | 18.6
.. psl->motion_paths_ps         | 0.00 | 0.01 | 18.6
.. psl->extra_grid_ps           | 0.00 | 0.02 | 18.6
.. psl->extra_centers_ps        | 0.00 | 0.01 | 18.5
.. psl->antialiasing_ps         | 0.10 | 0.03 | 18.5
. RegionInfo                    | 0.03 | 0.11 | 18.5

 Group                          | GPU  | CPU  | Latency
--------------------------------|------|------|--------
 Total                          | 27.8 | 25.0 |
. SPACE_TOPBAR                  | 0.00 | 0.01 | 14.6
. SPACE_STATUSBAR               | 0.00 | 0.01 | 14.6
. SPACE_VIEW3D                  | 27.5 | 24.8 | 14.5
.. Viewport                     | 27.5 | 24.8 | 14.5
. Window Redraw                 | 0.30 | 0.10 | 17.2

I’ve also tested other scenes and they show similar results.

(These are from an RTX 3060 Ti, it would be good to have a similar comparison made on AMD and on Apple hardware)

Issues

PrePass

PrePass is slower in EEVEE Next (2.24 vs 1.36ms).
~~No idea about the reason, I've tried removing all the extra stuff from the shaders (velocity, displacement, transparency) and the performance difference didn't change.~~

Update: There are thousands of rocks in the scene and many of them are unique meshes.
After merging all of them into one mesh or ensuring they all share the same mesh (so they can be batched), the pre-pass performance between Legacy and Next is identical.
So the performance difference seems caused by the draw call overhead, maybe due to EEVEE-Next using indirect draw calls? If that's the case we may want to look into using regular draw calls + CPU culling for single instance meshes.

Update 2: Fix: #117561

Shading

The Shading pass in EEVEE Next (which only outputs the closure data) takes a bit longer than the EEVEE Legacy pass (5.08 vs 4.90 ms) even if EEVEE Legacy is computing the lighting too.
So, in reality, Shading + Lighting is taking 6.53 ms in EEVEE Next (Shading+Eval Lights+Combine) vs 4.90 ms in EEVEE Legacy.
I guess the main bottleneck here is fillrate and we’ll have to eat the cost as long as we use Deferred Rendering.
Other types of scenes (with more light sources or very dense geometry) might compare more favorably.

Update: A good chunk of the performance difference seems to be caused by the same reason as the pre-pass difference. The shading pass (without lighting) in Next becomes faster than Legacy after #117561.

Shadows

EEVEE Legacy is able to fully skip the Shadow pass, but EEVEE Next Deferred.Opaque Shadows are always being re-rendered (5.55 ms) so the update detection is not working correctly.

Update: This doesn't seem to happen consistently, even in the same scene. Sometimes the update detection seems to work fine.

Running Shadows.set_view has a non-trivial cost (1.20 ms) even when it doesn’t render anything.
That’s up to 2.40 ms (Deferred.Refraction + Forward.Opaque) doing nothing.
We could avoid by it by rendering all pre-passes fist and running the shadow update/rendering only once per frame.
Forward.Opaque can probably be fully removed, though?

Update: Deferred.Refraction and Forward.Opaque passes are now skipped when they're empty. The issue is still there when they're used, though.

Update 2: While tagging still has a cost, the Surface.Shading overhead is gone after #117561.

ShadowsStatistics

The large difference between EEVEE Next (21.5 ms) and the full Viewport (27.5 ms) seems to be caused almost entirely by the Shadow::statistics_buf_ read (which is done at sync time).
Removing the read also increases the latency, which is oddly low at the moment.
I’m not sure if the cause is a bug on our side (Swapchain maybe?) or if it’s simply the driver using a different strategy.

Update: Fix: #117521

Raytracing

Raytracing seems quite costly (~~4.35 ms~~ edit: actually more (~5ms), there was a bug in the initial test, see #117159),
not sure how much the horizon passes could be optimized.
But I found something odd with the “regular” screen tracing. If I disable the horizon method (by setting Max Roughness to 1) this is what I get:

.... Raytracing                 | 4.99 | 0.55 | 49.1
..... TileClassify              | 0.08 | 0.01 | 49.1
..... Raytracing                | 2.51 | 0.17 | 49.2
...... TileCompact              | 0.01 | 0.01 | 49.1
...... RayGenerate              | 0.06 | 0.01 | 49.1
...... Trace.Screen             | 0.44 | 0.01 | 49.2
...... DenoiseSpatial           | 0.76 | 0.01 | 49.6
...... DenoiseTemporal          | 0.40 | 0.01 | 50.3
...... DenoiseBilateral         | 0.82 | 0.01 | 50.7
..... Raytracing                | 2.34 | 0.15 | 51.5
...... TileCompact              | 0.01 | 0.01 | 51.5
...... RayGenerate              | 0.04 | 0.01 | 51.5
...... Trace.Screen             | 0.34 | 0.01 | 51.5
...... DenoiseSpatial           | 0.49 | 0.01 | 51.8
...... DenoiseTemporal          | 0.25 | 0.01 | 52.3
...... DenoiseBilateral         | 1.18 | 0.01 | 52.5
..... Raytracing                | 0.05 | 0.15 | 53.7
...... TileCompact              | 0.00 | 0.01 | 53.7
...... RayGenerate              | 0.00 | 0.01 | 53.6
...... Trace.Screen             | 0.00 | 0.01 | 53.6
...... DenoiseSpatial           | 0.00 | 0.01 | 53.6
...... DenoiseTemporal          | 0.00 | 0.01 | 53.6
...... DenoiseBilateral         | 0.00 | 0.01 | 53.6

Most of the time seems to be eaten by denoising (4.01 ms out of 4.99 ms).
So, at the moment, it doesn’t seem like the current horizon scanning or the denoise implementation are worth its cost.

Update: Heavily improved by #118924

Conclusion

By optimizing these issues, I think we should get roughly on the same performance level as EEVEE Legacy.
I'll keep updating the task as I find more details.

To compare EEVEE Legacy and EEVEE Next and look for bottlenecks, I’ve made a performance test with the Wanderer scene using #116304. The camera is static and pointing slightly downwards so the full view is covered by geometry. ### Results <details> <summary>EEVEE Legacy</summary> ``` Group | GPU | CPU | Latency --------------------------------|------|------|-------- Total | 8.37 | 3.02 | . Manager.end_sync | 0.02 | 0.06 | 1.68 . EEVEE | 7.95 | 2.02 | 1.62 .. Probes Refresh | 0.00 | 0.01 | 1.61 .. Shadows | 0.00 | 0.05 | 1.58 ... Cube Shadow Maps | 0.00 | 0.01 | 1.58 ... Cascaded Shadow Maps | 0.00 | 0.01 | 1.55 .. Prepass | 1.36 | 0.38 | 1.51 ... psl->depth_ps | 1.36 | 0.34 | 1.48 .. Main MinMax buffer | 0.10 | 0.24 | 2.48 ... Max buffer | 0.10 | 0.21 | 2.47 .... psl->maxz_copydepth_ps | 0.03 | 0.02 | 2.45 .... psl->maxz_downlevel_ps | 0.02 | 0.01 | 2.44 .... psl->maxz_downlevel_ps | 0.01 | 0.01 | 2.44 .... psl->maxz_downlevel_ps | 0.00 | 0.01 | 2.43 .... psl->maxz_downlevel_ps | 0.00 | 0.01 | 2.41 .... psl->maxz_downlevel_ps | 0.00 | 0.01 | 2.39 .... psl->maxz_downlevel_ps | 0.00 | 0.01 | 2.37 .. GTAO Horizon Scan | 0.67 | 0.04 | 2.34 ... psl->ao_horizon_search | 0.66 | 0.01 | 2.32 .. Shading | 4.90 | 0.43 | 2.95 ... psl->background_ps | 0.01 | 0.02 | 2.94 ... psl->material_ps | 4.88 | 0.38 | 2.93 .. SSR | 0.60 | 0.32 | 7.42 ... psl->ssr_raytrace | 0.10 | 0.01 | 7.40 ... Downsample Radiance | 0.12 | 0.22 | 7.48 .... psl->color_copy_ps | 0.04 | 0.01 | 7.46 .... psl->color_downsample_ps | 0.02 | 0.01 | 7.48 .... psl->color_downsample_ps | 0.00 | 0.01 | 7.48 .... psl->color_downsample_ps | 0.00 | 0.01 | 7.47 .... psl->color_downsample_ps | 0.00 | 0.01 | 7.42 .... psl->color_downsample_ps | 0.00 | 0.01 | 7.40 .... psl->color_downsample_ps | 0.00 | 0.01 | 7.39 ... psl->ssr_resolve | 0.37 | 0.02 | 7.36 .. psl->probe_display | 0.00 | 0.01 | 7.69 .. Opaque Refraction | 0.00 | 0.01 | 7.67 .. Post FX | 0.21 | 0.35 | 7.65 ... Bloom Blit | 0.04 | 0.02 | 7.63 ... Bloom Downsample First | 0.01 | 0.01 | 7.64 ... Bloom Downsample | 0.00 | 0.01 | 7.62 ... Bloom Downsample | 0.00 | 0.01 | 7.60 ... Bloom Downsample | 0.00 | 0.01 | 7.57 ... Bloom Downsample | 0.00 | 0.01 | 7.55 ... Bloom Upsample | 0.00 | 0.02 | 7.52 ... Bloom Upsample | 0.00 | 0.01 | 7.49 ... Bloom Upsample | 0.00 | 0.01 | 7.47 ... Bloom Upsample | 0.02 | 0.01 | 7.46 ... Bloom Resolve | 0.08 | 0.01 | 7.46 . Overlay | 0.22 | 0.66 | 7.54 .. psl->background_ps | 0.03 | 0.01 | 7.52 .. psl->fade_ps[i] | 0.00 | 0.01 | 7.51 .. psl->facing_ps[i] | 0.00 | 0.01 | 7.49 .. psl->extra_blend_ps | 0.00 | 0.01 | 7.47 .. psl->wireframe_ps | 0.00 | 0.01 | 7.45 .. *p_armature_trans_ps | 0.00 | 0.01 | 7.43 .. *p_armature_ps | 0.00 | 0.02 | 7.40 .. psl->particle_ps | 0.00 | 0.01 | 7.37 .. psl->metaball_ps[i] | 0.00 | 0.01 | 7.35 .. *p_extra_ps | 0.00 | 0.08 | 7.33 .. psl->attribute_ps | 0.00 | 0.01 | 7.25 .. psl->grid_ps | 0.04 | 0.01 | 7.22 .. psl->fade_ps[i] | 0.00 | 0.01 | 7.23 .. psl->facing_ps[i] | 0.00 | 0.01 | 7.22 .. *p_armature_trans_ps | 0.00 | 0.01 | 7.19 .. *p_armature_ps | 0.00 | 0.01 | 7.17 .. *p_extra_ps | 0.00 | 0.02 | 7.15 .. psl->metaball_ps[i] | 0.00 | 0.01 | 7.11 .. psl->motion_paths_ps | 0.00 | 0.01 | 7.09 .. psl->extra_grid_ps | 0.00 | 0.01 | 7.07 .. psl->extra_centers_ps | 0.00 | 0.01 | 7.05 .. psl->antialiasing_ps | 0.10 | 0.01 | 7.03 . RegionInfo | 0.03 | 0.06 | 7.06 Group | GPU | CPU | Latency --------------------------------|------|------|-------- Total | 12.2 | 6.86 | . SPACE_TOPBAR | 0.00 | 0.02 | 0.25 . SPACE_STATUSBAR | 0.00 | 0.00 | 0.21 . SPACE_VIEW3D | 11.9 | 6.67 | 0.19 .. Viewport | 11.9 | 6.65 | 0.19 . Window Redraw | 0.30 | 0.11 | 5.43 ``` </details> <details> <summary>EEVEE Next</summary> ``` Group | GPU | CPU | Latency --------------------------------|------|------|-------- Total | 21.9 | 5.80 | . Manager.end_sync | 0.06 | 0.14 | 2.32 . EEVEE | 21.5 | 4.58 | 2.18 .. negZ_view | 21.5 | 4.40 | 2.16 ... LightCulling | 0.04 | 0.10 | 2.13 .... Select | 0.00 | 0.01 | 2.12 .... Sort | 0.00 | 0.01 | 2.10 .... Zbin | 0.01 | 0.01 | 2.08 .... Tiles | 0.01 | 0.01 | 2.08 ... Probe.Select | 0.00 | 0.01 | 2.06 ... World.Background | 1.03 | 0.03 | 2.03 ... Deferred.Opaque | 18.9 | 2.93 | 3.03 .... View.compute_visibility | 0.01 | 0.02 | 3.01 .... DrawMultiBuf.bind | 0.01 | 0.03 | 2.99 .... Prepass | 2.24 | 0.36 | 2.96 ..... DoubleSided.Static | 2.21 | 0.27 | 2.95 ..... SingleSided.Static | 0.00 | 0.01 | 4.88 ..... DoubleSided.Moving | 0.02 | 0.01 | 4.86 ..... SingleSided.Moving | 0.00 | 0.00 | 4.86 .... HizUpdate | 0.06 | 0.03 | 4.83 .... Shadow | 5.55 | 0.80 | 4.86 ..... TilemapSetup | 0.02 | 0.09 | 4.85 ...... ClearClipmap | 0.00 | 0.01 | 4.84 ...... DirectionalBounds | 0.00 | 0.01 | 4.82 ...... Init | 0.00 | 0.01 | 4.80 ...... CasterUpdate | 0.00 | 0.01 | 4.79 ..... View.compute_visibility | 0.01 | 0.01 | 4.76 ..... DrawMultiBuf.bind | 0.00 | 0.01 | 4.76 ..... TagUsage | 0.13 | 0.06 | 4.74 ...... Opaque | 0.13 | 0.01 | 4.73 ...... Transparent | 0.00 | 0.02 | 4.83 ..... TilemapUpdate | 0.12 | 0.15 | 4.80 ...... MaskLod | 0.00 | 0.01 | 4.79 ...... Free | 0.00 | 0.01 | 4.77 ...... Defrag | 0.00 | 0.01 | 4.76 ...... AllocatePages | 0.00 | 0.01 | 4.74 ...... Finalize | 0.01 | 0.01 | 4.73 ...... RenderClear | 0.07 | 0.01 | 4.72 ..... View.compute_procedural_bounds | 0.01 | 0.01 | 4.77 ..... View.compute_visibility | 0.03 | 0.01 | 4.76 ..... DrawMultiBuf.bind | 0.01 | 0.02 | 4.78 ..... Shadow.Surface | 5.17 | 0.31 | 4.76 ...... Shadow.Surface | 5.17 | 0.30 | 4.75 ....... Shadow.Surface.Double-Sided | 5.17 | 0.25 | 4.74 ....... Shadow.Surface.Single-Sided | 0.00 | 0.01 | 9.64 .... View.compute_visibility | 0.01 | 0.01 | 9.58 .... DrawMultiBuf.bind | 0.01 | 0.02 | 9.58 .... Shading | 5.08 | 0.41 | 9.55 ..... DoubleSided | 0.00 | 0.00 | 9.54 ..... SingleSided | 0.00 | 0.01 | 9.53 ..... DoubleSided | 5.00 | 0.30 | 9.51 ..... SingleSided | 0.00 | 0.01 | 14.2 ..... StencilClassify | 0.07 | 0.02 | 14.1 .... Raytracing | 4.35 | 0.93 | 14.2 ..... HorizonScan.Setup | 0.07 | 0.01 | 14.2 ..... TileClassify | 0.08 | 0.01 | 14.2 ..... Raytracing | 2.47 | 0.23 | 14.3 ...... TileCompact | 0.01 | 0.01 | 14.2 ...... RayGenerate | 0.01 | 0.01 | 14.2 ...... Trace.Screen | 0.01 | 0.01 | 14.2 ...... DenoiseSpatial | 0.05 | 0.01 | 14.2 ...... DenoiseTemporal | 0.03 | 0.02 | 14.2 ...... DenoiseBilateral | 0.01 | 0.01 | 14.2 ...... HorizonScan.Trace | 1.63 | 0.01 | 14.2 ...... HorizonScan.Denoise | 0.69 | 0.01 | 15.8 ..... Raytracing | 1.65 | 0.38 | 16.5 ...... TileCompact | 0.01 | 0.01 | 16.5 ...... RayGenerate | 0.01 | 0.02 | 16.5 ...... Trace.Screen | 0.01 | 0.01 | 16.4 ...... DenoiseSpatial | 0.04 | 0.02 | 16.4 ...... DenoiseTemporal | 0.02 | 0.05 | 16.4 ...... DenoiseBilateral | 0.01 | 0.04 | 16.3 ...... HorizonScan.Trace | 1.06 | 0.01 | 16.3 ...... HorizonScan.Denoise | 0.46 | 0.02 | 17.3 ..... Raytracing | 0.06 | 0.22 | 17.7 ...... TileCompact | 0.01 | 0.01 | 17.7 ...... RayGenerate | 0.00 | 0.01 | 17.7 ...... Trace.Screen | 0.00 | 0.01 | 17.7 ...... DenoiseSpatial | 0.00 | 0.02 | 17.7 ...... DenoiseTemporal | 0.00 | 0.01 | 17.7 ...... DenoiseBilateral | 0.00 | 0.01 | 17.6 ...... HorizonScan.Trace | 0.00 | 0.01 | 17.6 ...... HorizonScan.Denoise | 0.00 | 0.01 | 17.6 .... EvalLights | 1.34 | 0.06 | 17.6 ..... Eval.Light | 1.34 | 0.04 | 17.5 .... Combine | 0.11 | 0.01 | 18.8 ... Deferred.Refract | 0.00 | 0.01 | 18.9 ... Forward.Opaque | 1.31 | 1.06 | 18.9 .... View.compute_visibility | 0.01 | 0.01 | 18.9 .... DrawMultiBuf.bind | 0.00 | 0.01 | 18.9 .... Prepass | 0.00 | 0.08 | 18.9 ..... DoubleSided.Static | 0.00 | 0.01 | 18.9 ..... SingleSided.Static | 0.00 | 0.01 | 18.9 ..... DoubleSided.Moving | 0.00 | 0.01 | 18.8 ..... SingleSided.Moving | 0.00 | 0.00 | 18.8 .... HizUpdate | 0.06 | 0.01 | 18.8 .... Shadow | 1.20 | 0.75 | 18.8 ..... TilemapSetup | 0.02 | 0.09 | 18.8 ...... ClearClipmap | 0.00 | 0.01 | 18.8 ...... DirectionalBounds | 0.00 | 0.01 | 18.8 ...... Init | 0.00 | 0.01 | 18.8 ...... CasterUpdate | 0.00 | 0.01 | 18.8 ..... View.compute_visibility | 0.01 | 0.01 | 18.7 ..... DrawMultiBuf.bind | 0.00 | 0.01 | 18.7 ..... TagUsage | 0.13 | 0.05 | 18.7 ...... Opaque | 0.12 | 0.01 | 18.7 ...... Transparent | 0.00 | 0.01 | 18.8 ..... TilemapUpdate | 0.04 | 0.14 | 18.8 ...... MaskLod | 0.00 | 0.01 | 18.8 ...... Free | 0.00 | 0.01 | 18.8 ...... Defrag | 0.00 | 0.01 | 18.7 ...... AllocatePages | 0.00 | 0.01 | 18.7 ...... Finalize | 0.01 | 0.01 | 18.7 ...... RenderClear | 0.00 | 0.01 | 18.7 ..... View.compute_procedural_bounds | 0.01 | 0.01 | 18.7 ..... View.compute_visibility | 0.03 | 0.01 | 18.7 ..... DrawMultiBuf.bind | 0.01 | 0.02 | 18.7 ..... Shadow.Surface | 0.91 | 0.28 | 18.7 ...... Shadow.Surface | 0.91 | 0.26 | 18.7 ....... Shadow.Surface.Double-Sided | 0.90 | 0.22 | 18.6 ....... Shadow.Surface.Single-Sided | 0.00 | 0.01 | 19.3 .... View.compute_visibility | 0.01 | 0.01 | 19.3 .... DrawMultiBuf.bind | 0.00 | 0.01 | 19.3 .... Shading | 0.00 | 0.04 | 19.2 ..... SingleSided | 0.00 | 0.00 | 19.2 ..... DoubleSided | 0.00 | 0.00 | 19.2 ... View.compute_visibility | 0.01 | 0.01 | 19.2 ... DrawMultiBuf.bind | 0.00 | 0.01 | 19.2 ... Forward.Transparent | 0.00 | 0.02 | 19.1 .... ResourceBind | 0.00 | 0.01 | 19.1 ... Film.Accumulate | 0.18 | 0.02 | 19.1 .. Velocity Copy Pass | 0.00 | 0.01 | 19.2 . Overlay | 0.23 | 0.80 | 19.1 .. psl->background_ps | 0.04 | 0.02 | 19.1 .. psl->fade_ps[i] | 0.00 | 0.01 | 19.1 .. psl->facing_ps[i] | 0.00 | 0.01 | 19.1 .. psl->extra_blend_ps | 0.00 | 0.01 | 19.0 .. psl->wireframe_ps | 0.00 | 0.02 | 19.0 .. *p_armature_trans_ps | 0.00 | 0.01 | 19.0 .. *p_armature_ps | 0.00 | 0.02 | 19.0 .. psl->particle_ps | 0.00 | 0.01 | 18.9 .. psl->metaball_ps[i] | 0.00 | 0.01 | 18.9 .. *p_extra_ps | 0.00 | 0.09 | 18.9 .. psl->attribute_ps | 0.00 | 0.01 | 18.8 .. psl->grid_ps | 0.04 | 0.02 | 18.8 .. psl->fade_ps[i] | 0.00 | 0.02 | 18.8 .. psl->facing_ps[i] | 0.00 | 0.01 | 18.7 .. *p_armature_trans_ps | 0.00 | 0.01 | 18.7 .. *p_armature_ps | 0.00 | 0.02 | 18.7 .. *p_extra_ps | 0.00 | 0.01 | 18.6 .. psl->metaball_ps[i] | 0.00 | 0.01 | 18.6 .. psl->motion_paths_ps | 0.00 | 0.01 | 18.6 .. psl->extra_grid_ps | 0.00 | 0.02 | 18.6 .. psl->extra_centers_ps | 0.00 | 0.01 | 18.5 .. psl->antialiasing_ps | 0.10 | 0.03 | 18.5 . RegionInfo | 0.03 | 0.11 | 18.5 Group | GPU | CPU | Latency --------------------------------|------|------|-------- Total | 27.8 | 25.0 | . SPACE_TOPBAR | 0.00 | 0.01 | 14.6 . SPACE_STATUSBAR | 0.00 | 0.01 | 14.6 . SPACE_VIEW3D | 27.5 | 24.8 | 14.5 .. Viewport | 27.5 | 24.8 | 14.5 . Window Redraw | 0.30 | 0.10 | 17.2 ``` </details> I’ve also tested other scenes and they show similar results. *(These are from an RTX 3060 Ti, it would be good to have a similar comparison made on AMD and on Apple hardware)* ### Issues #### PrePass PrePass is slower in EEVEE Next (2.24 vs 1.36ms). ~~No idea about the reason, I've tried removing all the extra stuff from the shaders (velocity, displacement, transparency) and the performance difference didn't change.~~ *Update:* There are thousands of rocks in the scene and many of them are unique meshes. After merging all of them into one mesh or ensuring they all share the same mesh (so they can be batched), the pre-pass performance between Legacy and Next is identical. So the performance difference seems caused by the draw call overhead, maybe due to EEVEE-Next using indirect draw calls? If that's the case we may want to look into using regular draw calls + CPU culling for single instance meshes. *Update 2:* Fix: #117561 #### Shading The Shading pass in EEVEE Next (which only outputs the closure data) takes a bit longer than the EEVEE Legacy pass (5.08 vs 4.90 ms) even if EEVEE Legacy is computing the lighting too. So, in reality, Shading + Lighting is taking 6.53 ms in EEVEE Next (Shading+Eval Lights+Combine) vs 4.90 ms in EEVEE Legacy. I guess the main bottleneck here is fillrate and we’ll have to eat the cost as long as we use Deferred Rendering. Other types of scenes (with more light sources or very dense geometry) might compare more favorably. *Update:* A good chunk of the performance difference seems to be caused by the same reason as the pre-pass difference. The shading pass (without lighting) in Next becomes faster than Legacy after #117561. #### Shadows EEVEE Legacy is able to fully skip the Shadow pass, but EEVEE Next Deferred.Opaque Shadows are always being re-rendered (5.55 ms) so the update detection is not working correctly. *Update:* This doesn't seem to happen consistently, even in the same scene. Sometimes the update detection seems to work fine. Running `Shadows.set_view` has a non-trivial cost (1.20 ms) even when it doesn’t render anything. That’s up to 2.40 ms (Deferred.Refraction + Forward.Opaque) doing nothing. We could avoid by it by rendering all pre-passes fist and running the shadow update/rendering only once per frame. Forward.Opaque can probably be fully removed, though? *Update:* Deferred.Refraction and Forward.Opaque passes are now skipped when they're empty. The issue is still there when they're used, though. *Update 2:* While tagging still has a cost, the `Surface.Shading` overhead is gone after #117561. #### ShadowsStatistics The large difference between EEVEE Next (21.5 ms) and the full Viewport (27.5 ms) seems to be caused almost entirely by the `Shadow::statistics_buf_` read (which is done at sync time). Removing the read also increases the latency, which is oddly low at the moment. I’m not sure if the cause is a bug on our side (Swapchain maybe?) or if it’s simply the driver using a different strategy. *Update:* Fix: #117521 #### Raytracing Raytracing seems quite costly (~~4.35 ms~~ *edit: actually more (~5ms), there was a bug in the initial test, see #117159*), not sure how much the horizon passes could be optimized. But I found something odd with the “regular” screen tracing. If I disable the horizon method (by setting Max Roughness to 1) this is what I get: ```` .... Raytracing | 4.99 | 0.55 | 49.1 ..... TileClassify | 0.08 | 0.01 | 49.1 ..... Raytracing | 2.51 | 0.17 | 49.2 ...... TileCompact | 0.01 | 0.01 | 49.1 ...... RayGenerate | 0.06 | 0.01 | 49.1 ...... Trace.Screen | 0.44 | 0.01 | 49.2 ...... DenoiseSpatial | 0.76 | 0.01 | 49.6 ...... DenoiseTemporal | 0.40 | 0.01 | 50.3 ...... DenoiseBilateral | 0.82 | 0.01 | 50.7 ..... Raytracing | 2.34 | 0.15 | 51.5 ...... TileCompact | 0.01 | 0.01 | 51.5 ...... RayGenerate | 0.04 | 0.01 | 51.5 ...... Trace.Screen | 0.34 | 0.01 | 51.5 ...... DenoiseSpatial | 0.49 | 0.01 | 51.8 ...... DenoiseTemporal | 0.25 | 0.01 | 52.3 ...... DenoiseBilateral | 1.18 | 0.01 | 52.5 ..... Raytracing | 0.05 | 0.15 | 53.7 ...... TileCompact | 0.00 | 0.01 | 53.7 ...... RayGenerate | 0.00 | 0.01 | 53.6 ...... Trace.Screen | 0.00 | 0.01 | 53.6 ...... DenoiseSpatial | 0.00 | 0.01 | 53.6 ...... DenoiseTemporal | 0.00 | 0.01 | 53.6 ...... DenoiseBilateral | 0.00 | 0.01 | 53.6 ```` Most of the time seems to be eaten by denoising (4.01 ms out of 4.99 ms). So, at the moment, it doesn’t seem like the current horizon scanning or the denoise implementation are worth its cost. *Update:* Heavily improved by #118924 ### Conclusion By optimizing these issues, I think we should get roughly on the same performance level as EEVEE Legacy. I'll keep updating the task as I find more details.

❤️ 5 🎉 1 🚀 2

Miguel Pozo added the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

EEVEE-Next: Performance Analysis #117246