Fix: Too complex to be optimized by compiler #109222

Iliya Katushenock · 2023-06-22T00:55:33+02:00

Iliya Katushenock commented

2023-06-22 00:55:33 +02:00

This node split all faces to duplicate each random number time.
Each loops, edges, vertices is just ordered sequences of indices.
Most likely, due to historical reasons, this was overlooked and
as a result the code became so complex with potential simplicity.
If it's a simple loop, the compiler might have output the span
data as just a sequence of indexes. Not sure about the sizing.
But if it was a lambda, then the compiler could not make a
prediction about the index sequence.
The problem was found in a benchmark of #109174.
The code in the lambda could not be optimized. Although for a
regular loops it was possible (?): (Now) 154.0 ms -> (with PR) 119.6 ms.
If this is in a lambda (selection.foreach_index_optimized<int>),
then the difference is: (in #109174) 193.8 ms -> (with PR) 116.9 ms.

Lambda loop	Regular loop	Lambda loop	Regular loop
In #109174 is `193.8 ms`	In main is `154.0 ms`	In #109174 is `116.9 ms`	In main is `119.6 ms`
Without compiler optimization	With compiler optimization	With manual PR's optimization	With PR's manual optimization

This node split all faces to duplicate each random number time. Each loops, edges, vertices is just ordered sequences of indices. Most likely, due to historical reasons, this was overlooked and as a result the code became so complex with potential simplicity. If it's a simple loop, the compiler might have output the span data as just a sequence of indexes. Not sure about the sizing. But if it was a lambda, then the compiler could not make a prediction about the index sequence. The problem was found in a benchmark of https://projects.blender.org/blender/blender/pulls/109174. The code in the lambda could not be optimized. Although for a regular loops it was possible (?): (Now) `154.0 ms` -> (with PR) `119.6 ms`. If this is in a lambda (`selection.foreach_index_optimized<int>`), then the difference is: (in #109174) `193.8 ms` -> (with PR) `116.9 ms`. | Lambda loop | Regular loop | Lambda loop | Regular loop | | -- | -- | -- | -- | | In #109174 is `193.8 ms` | In main is `154.0 ms` | In #109174 is `116.9 ms` | In main is `119.6 ms` | | Without compiler optimization | With compiler optimization | With manual PR's optimization | With PR's manual optimization |

duplicate benchmark.blend

877 KiB

Iliya Katushenock added 1 commit 2023-06-22 00:55:46 +02:00

dbe3461a32 init

Iliya Katushenock changed title from ~~Fix: Too complex to compiler to be optimized~~ to Fix: Too complex to be optimized by compiler

2023-06-22 01:22:18 +02:00

Hans Goudey approved these changes 2023-06-22 13:26:06 +02:00

Hans Goudey left a comment

Even disregarding the performance numbers, I think this is better because it's clearer.

Iliya Katushenock commented

2023-06-22 13:36:01 +02:00

I'm sure there's a lot more of this kind of stuff in this node as well as in other parts of blender, so when I have time I'll add other fixes of this kind here if I find it.

Hans Goudey merged commit a711c5004c into main

2023-06-22 13:45:49 +02:00

Hans Goudey referenced this issue from a commit

2023-06-22 13:45:50 +02:00

Geometry Nodes: Simplify code in duplicate elements face mode

Iliya Katushenock deleted branch fix_too_complex_fro_compile_optimizations

2023-06-22 13:47:15 +02:00

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Fix: Too complex to be optimized by compiler #109222