Leon Brands - Methodical GPU Profiling (Graphics Programming Conference 2025)

BEHAVIOUR
INTERACTIVE
©2024 and BEHAVIOUR and other related trademarks and logos belong to Behaviour Interactive Inc.
All rights reserved. Other trademarks or copyrights are the property of their respective owners.
HIGHLY CONFIDENTIAL – DO NOT REDISTRIBUTE
Methodical GPU Profiling

2
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
Introduction
Leon Brands
Senior Graphics Programmer
Behaviour Interactive
Socials at the end

3
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
Who are we?

4
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
Agenda
• Motivation
• Strategy
• Investigate
• Planning
• Takeaways

6
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
Motivation
• WFH
• 3+ years w/ 6-10 prog
• Graphics support on consoles
• Optimizations 
Scope

Motivation
• Inherently undefined
• Bottlenecks are always changing
• Needed a strategy
Workload
7
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Familiarity with code base,
articles, talks, discussions
Ideas
Observed performance issues by
other developers & playtesters
Reports
Hard work!
Profiling
Optimization
Origins
1
2
3
Origi
n
8
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Familiarity with code base,
articles, talks, discussions
Ideas
Observed performance issues by
other developers & playtesters
Reports
Hard work!
Profiling
Optimization
Origins
1
2
3
Origi
n Unreliable
Might not
Happen
Scientific
9
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• Optimizing is scientific
• Clear Goal
• Defined & Provable Results
• Iterative workflow:
• Analyze
• Optimize
• Confirm
• Repeat
while(frameTime > 16.6)
11
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• Goals
• Find optimization ideas
• Re-evaluate status
• Adjust to current phase
• Do you need more data?
• Don’t waste time
Monthly Profiling Suite
12
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• Our Profiling Suite
• Divide into steps
• Work our way downwards
• Scope 1: Game
• Scope 2: Level
• Scope 3: POI
• Scope 4: Investigate
Narrow it Down
13
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• 60 FPS
• Test all levels
Scope 1: Game
14
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

15
HIGHLY
CONFIDENTIAL
–
DO
NOT
REDISTRIBUTE
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
©2024
and
BEHAVIOUR
and
other
related
trademarks
and
logos
belong
to
Behaviour
Interactive
Inc.
All
rights
reserved.
Other
trademarks
or
copyrights
are
the
property
of
their
respective
owners.
Strategy
• 60 FPS
• Test all levels
• UGC…
Scope 1: Game

Strategy
• Fine, only official levels
• Still tons
• Quick tool for QA
Scope 1: Game
16
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• Selected some levels
• Look deeper
• Playthrough Profiler
• Collect detailed frame data
• GPU frame-time
• Detailed breakdown into passes
• Exported as .tsv files, imported in excel
• Lot of data
• Process w/ Graphs/Algorithms
• Worst frames -> POI
Scope 2: Level
17
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
1 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 154 163 172 181 190 199 208 217 226 235 244 253 262 271 280 289 298 307 316 325 334 343 352 361 370 379 388 397 406
10
20
30
40
50
60
70
GPU Frame Time
GPU Frame Time Target
Capture (Every 30 frames)
Time
(ms)

Strategy
• List of POIs
• Enough to start diving deeper?
• Narrow down
• Match with gameplay video using game time
• Find new performance issues
• Skip
• Similar spikes
• Non-prog work
• Known issues
Scope 3: Point of Interest
1.56
1.12
0.79
2.21
0.81
9.64
1.76
0
.
0
5
1.50
1.17
0.06 0.00
0.62
0.08
0.62
Frame Breakdown
18
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Strategy
• Look at POI
• Poor performance on a render pass?
• How to optimize?
Scope 4: Investigate
19
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigat
e
• How do you optimize something?
• Depends on what you’re optimizing
• Rasterized Meshes
• Full-screen Shader
• Compute Shader
How?
21
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• GPU deals with large quantities
• Draw call spawns vert shaders, spawns …
• Take a top-down approach
• What can you reduce?
Rasterized Meshes
22
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Draw calls
• Frustum Culling
• Occlusion Culling
• Distance Culling
Rasterized Meshes
23
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Draw calls
• Vertices
• Optimize Assets
• Create LODs (and implement)
• Mesh Clusters
• Triangle Culling
Rasterized Meshes
24
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Draw calls
• Vertices
• Pixels
• Upscaling
• VRS
Rasterized Meshes
25
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Draw calls
• Vertices
• Pixels
• Lights
• Spawns even more work!
• Lights, decals
• We use clustered rendering
• With a simple z-range filter
• (short sight-lines)
Rasterized Meshes
26
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Draw calls
• Vertices
• Pixels
• Lights
• Don’t do everything blindly
Rasterized Meshes
27
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Main pass slow
• Bottleneck should be pixel shaders
• Pixel shaders low occupancy
• Vertex shaders also low
• Bottleneck in VAF
Learn From Our Mistakes
28
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Main pass slow
• Tried everything we thought we could
• Don’t read attributes you don’t need
• Separate vertex attribute streams
• Compress vertex attributes
29
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Main pass slow
• Tried everything we thought we could
• Nothing worked
• Goal: Unlock pixel shader
• Overzealous shader optimizations
• Vertex workload vs pixel workload
• Rebalance, change the input
30
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Can we optimize shaders now?
• After you’ve done everything else
• (usually)
• Optimize (pixel) shaders
• FS pass / Compute
• No larger pipeline
• 1 consideration
Shader Optimizations
31
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Not entirely in the scope of this talk
• Quick Thoughts:
• Waste less!
• Reduce
• Reuse
• Recycle
Shader Optimizations
32
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Consider this
• You know your project the best
• Optimizations are rarely without a sacrifice
• What are you willing to give up?
• Less Sacrifice = More Effort
Know Your Project
33
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Investigate
• Sacrifices were made
• Don’t keep them blindly!
• Prove impact & stability:
• Profile before/after
• Screenshot before/after
Post-Optimization
34
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Planning
• Very agile
• We had a lot of autonomy
• Status update
• Re-evaluate monthly
• Sprint-like planning
Agile
36
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Planning
• How do you determine priority?
• Discuss:
• Impact
• Complexity
• Risk
• Sacrifice
• Every idea should be considered
Prioritize
37
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Planning
• How to stay on track?
• Leave leeway
• Determine goal
• Set periodic targets
Longer Term
38
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Planning
• Example
• 45ms -> 30ms in 4 months
• 5ms/month
• Profile, analyze, etc. as discussed
• Tasks with little visual impact
• Fallbacks with worse tradeoffs
• Only use fallbacks to hit targets
Longer Term
39
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Takeaways
• Our decisions are backed by hard facts along the way
• This helps us make the best decisions we can
• Monthly results provided a record
Optimization is Scientific
41
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Takeaways
• This workflow evolved over time
• You can’t force yourself into a single workflow and never
reconsider
• We constantly adjusted for our needs, the project’s state, etc.
Don’t Let It Get Stale
42
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Takeaways
• This approach works for us
• Measurable results
• Backlog full of potential work
• Most valuable work is prioritized
• Consistent and significant performance improvements
It Worked
43
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

Sources
• Occlusion Culling:
https://blog.selfshadow.com/publications/practical-visibility/
• Practical Clustered Culling
https://www.humus.name/Articles/PracticalClusteredShading
.pdf
• Siggraph 2015: Advances in Real-Time Rendering in Games
https://advances.realtimerendering.com/s2015/aaltonenhaar
_siggraph2015_combined_final_footer_220dpi.pdf
• Octahedral Mapping
https://knarkowicz.wordpress.com/2014/04/16/octahedron-n
ormal-vector-encoding/
• Harnessing Wave Intrinsics For Good (And Evil)
https://youtu.be/U6t33RLa0XM?si=_KIA-EnzYVgS3ZxV
For Reading?
45
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E

46
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
We’re Hiring
BHVR Rotterdam
https://www.bhvr.com/rotterdam/
• Senior Programmer
• Principal Programmer

47
B
E
H
A
V
I
O
U
R
I
N
T
E
R
A
C
T
I
V
E
Let’s Connect
Leon Brands
Scan QR code for LinkedIn

Leon Brands - Methodical GPU Profiling (Graphics Programming Conference 2025)

More Related Content

Similar to Leon Brands - Methodical GPU Profiling (Graphics Programming Conference 2025)

Recently uploaded

Leon Brands - Methodical GPU Profiling (Graphics Programming Conference 2025)

Editor's Notes