Cycles OpenCL: Research per node compilation time #61461

New Issue

Jeroen Bakker · 2019-02-12T14:00:48+01:00

Jeroen Bakker commented

2019-02-12 14:00:48 +01:00

We identifier that svn_eval_nodes is the function that compiles slow on AMD. This task will research if there are problematic nodes, or that it is just the amount of nodes.
The result of this research will point us in a direction what nodes needs to be optimized

We identifier that `svn_eval_nodes` is the function that compiles slow on AMD. This task will research if there are problematic nodes, or that it is just the amount of nodes. The result of this research will point us in a direction what nodes needs to be optimized

Jeroen Bakker self-assigned this 2019-02-12 14:00:48 +01:00

Jeroen Bakker commented

2019-02-12 14:00:48 +01:00

Added subscriber: @Jeroen-Bakker

Jeroen Bakker changed title from ~~Research: Per node compilation time~~ to Cycles OpenCL: Research per node compilation time

2019-02-12 14:04:43 +01:00

Jeroen Bakker commented

2019-02-13 13:57:13 +01:00

Here are the measurements of how much effort it takes to compile a certain node.

node_times.pdf

node_times.ods

Nodes that take the most time are:

NODE_ENTER_BUMP_EVAL: 20 seconds inside primitive_attribute_float3 as it does kernel_tex_image_interp_3d texture
NODE_TEX_VOXEL: kernel_tex_image_interp_3d
NODE_VECTOR_DISPLACEMENT: primitive_attribute_float3 -> kernel_tex_image_interp_3d
NODE_NORMAL_MAP: primitive_attribute_float3 -> kernel_tex_image_interp_3d To retrieve tangent and one for normal smoothing
NODE_ATTR: primitive_attribute_float3 -> kernel_tex_image_interp_3d (twice) seems to be a fallback implementation. Can be written to float, float3, float.
NODE_ATTR_BUMP_DY: primitive_attribute_float3 -> kernel_tex_image_interp_3d (twice) seems to be a fallback implementation. Can be written to float, float3, float.
NODE_ATTR_BUMP_DX: primitive_attribute_float3 -> kernel_tex_image_interp_3d (twice) seems to be a fallback implementation. Can be written to float, float3, float.
NODE_PRINCIPLED_VOLUME: primitive_attribute_float3 -> kernel_tex_image_interp_3d for color attribute.
NODE_CLOSURE_BSDF: Is just a huge node. Would expect to put this in some compile directive as it has a big switch statement for all different kind of BSDF's.
NODE_AMBIENT_OCCLUSION: path_branched_rng_2D We could split the NODE_AO_ONLY_LOCAL in a separate node but that could also make the compilation time worse as you might need both in a single scene.
NODE_GEOMETRY: primitive_tangent -> primitive_attribute_float3 -> kernel_tex_image_interp_3d for generated coordinates.
NODE_TEX_VORONOI: 27 * cellnoise3 * 3 * hash?
NODE_TANGENT: two primitive_attribute_float3 that can be merged into a single call.
NODE_GEOMETRY_BUMP_DX: internally calls node_geometry
NODE_GEOMETRY_BUMP_DY: internally calls node_geometry
NODE_BEVEL: path_branched_rng_2D

Note: Is it possible to add a primitive_attribute_float3 without volume as function or parameter, I can imagine that some attributes never actually call kernel_tex_image_interp_3d as they are always stored somewhere else.

Here are the measurements of how much effort it takes to compile a certain node. [node_times.pdf](https://archive.blender.org/developer/F6599353/node_times.pdf) [node_times.ods](https://archive.blender.org/developer/F6599352/node_times.ods) Nodes that take the most time are: - NODE_ENTER_BUMP_EVAL: 20 seconds inside `primitive_attribute_float3` as it does `kernel_tex_image_interp_3d texture` - NODE_TEX_VOXEL: `kernel_tex_image_interp_3d` - NODE_VECTOR_DISPLACEMENT: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` - NODE_NORMAL_MAP: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` To retrieve tangent and one for normal smoothing - NODE_ATTR: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` (twice) seems to be a fallback implementation. Can be written to float, float3, float. - NODE_ATTR_BUMP_DY: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` (twice) seems to be a fallback implementation. Can be written to float, float3, float. - NODE_ATTR_BUMP_DX: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` (twice) seems to be a fallback implementation. Can be written to float, float3, float. - NODE_PRINCIPLED_VOLUME: `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` for color attribute. - NODE_CLOSURE_BSDF: Is just a huge node. Would expect to put this in some compile directive as it has a big switch statement for all different kind of BSDF's. - NODE_AMBIENT_OCCLUSION: `path_branched_rng_2D` We could split the NODE_AO_ONLY_LOCAL in a separate node but that could also make the compilation time worse as you might need both in a single scene. - NODE_GEOMETRY: `primitive_tangent` -> `primitive_attribute_float3` -> `kernel_tex_image_interp_3d` for generated coordinates. - NODE_TEX_VORONOI: 27 * cellnoise3 * 3 * hash? - NODE_TANGENT: two `primitive_attribute_float3` that can be merged into a single call. - NODE_GEOMETRY_BUMP_DX: internally calls node_geometry - NODE_GEOMETRY_BUMP_DY: internally calls node_geometry - NODE_BEVEL: `path_branched_rng_2D` Note: Is it possible to add a `primitive_attribute_float3` without volume as function or parameter, I can imagine that some attributes never actually call `kernel_tex_image_interp_3d` as they are always stored somewhere else.

Jeroen Bakker commented

2019-02-14 12:20:14 +01:00

Script used for this research. Needs to be run blender containing D2264 patch.

node_compile_times.py

And the updated svm.h

svm.h

Script used for this research. Needs to be run blender containing [D2264](https://archive.blender.org/developer/D2264) patch. [node_compile_times.py](https://archive.blender.org/developer/F6608353/node_compile_times.py) And the updated svm.h [svm.h](https://archive.blender.org/developer/F6608425/svm.h)

Jeroen Bakker commented

2019-02-15 13:38:40 +01:00

Changed status from 'Open' to: 'Resolved'

Jeroen Bakker closed this issue

2019-02-15 13:38:40 +01:00

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles OpenCL: Research per node compilation time #61461