Eevee preview nodes performance drop #71214

New Issue

Danila · 2019-10-30T11:03:28+01:00

Danila commented

2019-10-30 11:03:28 +01:00

System Information
Operating system:Windows 10
Graphics card: GeForce GTX 1060

Blender Version
Broken: Blender 2.81 Beta aff6446e06
Worked: 2.8

Short description of error
Complex shader with large amount of math nodes slow down performance
in 2.8 60 fps
in 2.81 30 fps

Exact steps for others to reproduce the error
Create material with 30 node groups contains 30 math nodes. Set viewport to material preview and watch performance droping NodesPerformanceDrop.blend

**System Information** Operating system:Windows 10 Graphics card: GeForce GTX 1060 **Blender Version** Broken: Blender 2.81 Beta aff6446e064f Worked: 2.8 **Short description of error** Complex shader with large amount of math nodes slow down performance in 2.8 60 fps in 2.81 30 fps **Exact steps for others to reproduce the error** Create material with 30 node groups contains 30 math nodes. Set viewport to material preview and watch performance droping [NodesPerformanceDrop.blend](https://archive.blender.org/developer/F7871941/NodesPerformanceDrop.blend)

Danila commented

2019-10-30 11:03:28 +01:00

Added subscriber: @DanilaPryamonosov

Germano Cavalcante commented

2019-10-31 16:25:34 +01:00

Added subscribers: @fclem, @Sergey, @dfelinto, @mano-wii

Clément Foucault was assigned by Germano Cavalcante

2019-10-31 16:25:34 +01:00

Germano Cavalcante commented

2019-10-31 16:25:35 +01:00

I can confirm, but I don't know if the solution is simple.
The problem arose because now all the shader tree is flattened before evaluation.

But the main reason for slowness is because of animated materials (see 9d59d209).
Blender doesn't differentiate whether a material is animated or not, so the BKE_material_eval simply clears all gpumaterials references with each frame change.
Then DRW_manager needs to recreate the gpumaterial, but first check for similar passes.
And to do this check, it has to go through the entire node tree to create a hash.
And so finds the pass of the original gpumaterial.

Removing this line, the problem is solved:

diff --git a/source/blender/blenkernel/intern/material.c b/source/blender/blenkernel/intern/material.c
index 73c278a0ab6..2e87f04a4bc 100644
--- a/source/blender/blenkernel/intern/material.c
+++ b/source/blender/blenkernel/intern/material.c
@@ -1613,5 +1613,4 @@ void paste_matcopybuf(Main *bmain, Material *ma)
 void BKE_material_eval(struct Depsgraph *depsgraph, Material *material)
 {
   DEG_debug_print_eval(depsgraph, __func__, material->id.name, material);
-  GPU_material_free(&material->gpumaterial);
 }

But then it would bring a problem for the animated materials.
It would be nice to find another solution for animated materials (updating only uniforms for example).

Assigning to @fclem, but @dfelinto and @Sergey may also have some remarks about it.

I can confirm, but I don't know if the solution is simple. The problem arose because now all the shader tree is flattened before evaluation. But the main reason for slowness is because of animated materials (see 9d59d209). Blender doesn't differentiate whether a material is animated or not, so the `BKE_material_eval` simply clears all `gpumaterials` references with each frame change. Then `DRW_manager` needs to recreate the `gpumaterial`, but first check for similar passes. And to do this check, it has to go through the entire node tree to create a hash. And so finds the pass of the original `gpumaterial`. Removing this line, the problem is solved: ``` diff --git a/source/blender/blenkernel/intern/material.c b/source/blender/blenkernel/intern/material.c index 73c278a0ab6..2e87f04a4bc 100644 --- a/source/blender/blenkernel/intern/material.c +++ b/source/blender/blenkernel/intern/material.c @@ -1613,5 +1613,4 @@ void paste_matcopybuf(Main *bmain, Material *ma) void BKE_material_eval(struct Depsgraph *depsgraph, Material *material) { DEG_debug_print_eval(depsgraph, __func__, material->id.name, material); - GPU_material_free(&material->gpumaterial); } ``` But then it would bring a problem for the animated materials. It would be nice to find another solution for animated materials (updating only uniforms for example). Assigning to @fclem, but @dfelinto and @Sergey may also have some remarks about it.

Sergey Sharybin commented

2019-10-31 17:02:39 +01:00

Part of the issue here is that material is not supposed to be re-evaluating on playback here.
The reason it was is because of some rather uncommon (in my experience) situation when ID has action but no f-curve. This can be fixed with something like P1149.

But there should be a way to speed up things even for the case when there is a real animation involved.

Not sure about current implementation, but we can calculate hash of every node group from dependency graph (which will happen in nice and threaded environment). We can even do things like constant folding and removal of unconnected nodes from there.

Now, even if that is still by-passed (imagine, that all the groups are connected and such) we can re-use hashes calculated for every group and use those to save on calculating hashes.

Part of the issue here is that material is not supposed to be re-evaluating on playback here. The reason it was is because of some rather uncommon (in my experience) situation when ID has action but no f-curve. This can be fixed with something like [P1149](https://archive.blender.org/developer/P1149.txt). But there should be a way to speed up things even for the case when there is a real animation involved. Not sure about current implementation, but we can calculate hash of every node group from dependency graph (which will happen in nice and threaded environment). We can even do things like constant folding and removal of unconnected nodes from there. Now, even if that is still by-passed (imagine, that all the groups are connected and such) we can re-use hashes calculated for every group and use those to save on calculating hashes.

Germano Cavalcante commented

2019-10-31 17:29:17 +01:00

I was a bit hasty, the hash is not the problem (obtained from GPU_generate_pass),
but it is GPU_material_from_nodetree in general (it should not even be called again),
especially ntreeGPUMaterialNodes (68,66% of the frame time).

I was a bit hasty, the hash is not the problem (obtained from `GPU_generate_pass`), but it is `GPU_material_from_nodetree` in general (it should not even be called again), especially `ntreeGPUMaterialNodes` (68,66% of the frame time).

Sergey Sharybin commented

2019-11-01 10:42:40 +01:00

If the material changes then i don't see how you can avoid GPU_material_from_nodetree.

If the material changes then i don't see how you can avoid `GPU_material_from_nodetree`.

Stanislav Ovcharov commented

2019-11-12 13:45:05 +01:00

Added subscriber: @StanislavOvcharov