-Based Workload Estimation for Mobile 3D Graphics Bren Mochocki*†, Kanishka Lahiri*, Srihari Cadambi*, Xiaobo Sharon Hu† *NEC Laboratories America, †University of Notre Dame DAC 2006

Mobile Graphics Technology Increasing resource load Performance (Speed) Lifetime (Energy) Graphics Technology Advanced 3D Basic 3D Video clips 2D color 1997 2000 2001 2002 2003 2004 2005 2006 2007 Time

Meeting Performance/Lifetime Requirements Hardware Solutions Woo, 04 low-power 3D ASIC Kameyama, 03 low-power 3D ASIC Gu, Chakraborty, Ooi, 06 Games are up for DVFS Akenine-Moller, 03 Texture compression for mobile terminals Mochocki, Lahiri, Cadambi, 06 DVFS for mobile 3D graphics Keep short Stick to title Some can use workload prediction System - Level Optimizations Graphics Algorithms Tack, 04 LoD control for mobile terminals Accurate workload prediction is critical

Mobile 3D Workload Estimation Why? Adapt architectural parameters Adapt application parameters Better on-line resource management Desirable properties Speed – must be performed on-line Accuracy Compact

Workload-Estimation Spectrum General Purpose Simplicity Application specific Accuracy History-Based Predictors Analytical Predictors General purpose history-based predictors provide poor prediction accuracy for rapidly changing workloads Highly accurate analytical schemes are too complex for use at run time

Workload-Estimation Spectrum General Purpose Simplicity Application specific Accuracy Signature-Based Predictor Uses combination of history and application-specific parameters (the signature) to predict future workload Strikes a balance between simplicity and accuracy Preserves both cause AND effect Preserves substantial history

Outline Introduction and Motivation Background 3D-pipeline Basics Challenges in workload Estimation Signature-Based Workload Prediction Experimental Results Conclusions

3D Pipeline Basics 3D representation  2D image Geometry Setup Texturing Geometry Setup Rendering High level simplified view of 3D graphics pipeline Practice describing rendering stage quickly (hidden surface removal) World View Camera View Raster View Frame Buffer Transformations Lighting Clipping Scan-line conversion Pixel rendering Texturing

Workload Across Applications 12 TexCube RoomRev 10 8 Execution Cycles (ARM, x107) 6 4 2 Motivate the source of the pipeline imbalance Benchmark Workload varies significantly between applications Prediction scheme must be flexible

Workload Within an Application Workload can change rapidly between frames 1 2 3 4 5 6 Race geometry render Execution Cycles (ARM, x107) setup Start animation right away This motivates DVFS, but first Transition: what impacts the workload variation? 1 16 31 46 61 76 91 106 121 136 151 166 181 196 Frame

Outline Introduction and Motivation Background Signature-Based Workload Prediction Signature Generation Method Overview Pipeline Modifications Experimental Results Conclusions

Example Signature: <vertex count, avg. area> 3D Pipeline extract end frame Frame Buffer Application extract extract signature measure workload <6, 2.5> 1.0e4 -Workload we want to predict (last box) -Default prediction (Static analysis) Success slide after this one Signature Workload Default Signature Table Workload Prediction

Example Signature: <vertex count, avg. area> 3D Pipeline end frame Frame Buffer Application extract extract signature measure workload <6, 2.5> 1.0e4 1. More parameters implies fewer collisions 2. Exact match not necessary – Distance measurements can be used 3. Initial signature table can be included with the application Signature Workload <6, 2.5> 1.0e4 1.0e4 Signature Table Workload Prediction

Example Signature: <vertex count, avg. area> 3D Pipeline end frame Frame Buffer Application No overlap (render all pixels) extract extract signature measure workload <6, 2.5> 1.2e4 1. More parameters implies fewer collisions 2. Exact match not necessary – Distance measurements can be used 3. Initial signature table can be included with the application Signature Workload <6, 2.5> 1.0e4 1.0e4 Signature Table Workload Prediction

Partitioning the 3D pipeline Bulk of 3D workload ORIGINAL GEOMETRY SETUP RENDER Application Display Transform Transform Lighting Lighting Clipping Clipping Scan-line conversion Scan-line conversion Per-pixel Operations Per-pixel Operations Generally small workload Provides necessary signature elements Pre buffer, post buffer PARTITIONED Application Display Transform + Clipping Buffer Lighting Scan-line conversion Per-pixel Operations Pre-Buffer Post Buffer

Pipeline Workload Pre-buffer workload is less than 10% of the total workload Pre-buffer variation is small Post-buffer workload is large with significant variation post-buffer pre-buffer Red lines are min-max range

Signature Composition Can vary by application May include: Average Triangle Area Average Triangle Height Total vertex count Lit vertex count Number of lights Any measurable parameter Larger signatures  more accurate Smaller signatures  less time & space

Outline Introduction & Background Experimental Framework Signature-Based Workload Prediction Experimental Results Evaluation Framework Signature length vs. accuracy Frame Rate Energy Conclusions

System-level Communication Architecture Architectural View pre-buffer signature extraction post-buffer Prog. Voltage Regulator Prog. PLL V, F Applications Processor Programmable 3D Graphics Engine Performance counter System-level Communication Architecture Animate Voltage regulation (for example, we could…) Memory Frame Buffer measure workload buffer signature table output

Evaluation Framework OpenGL/ES library Instrumented with pipeline stage triggers Hans-Martin Will Vincent Fast, cycle-accurate Simulation W. Qin Cross Compiler ARM — g++ 3D application Simit-ARM OpenGL/ES 1.0 3D – application Trace simulator of mobile 3D pipeline Triangle, Instruction, & Trigger traces Workload prediction scheme Architecture Model Trace Simulator Processor Energy Model 3D pipeline Performance/power Simulation output

Workload Accuracy > 2 fps error at peaks Average Error (normalized) Peaks < 1 fps Even simple signatures can give a good prediction (3 or 4 four well chosen signatures) Signature must be chosen well <a> 2 bytes <a,b> 6 bytes <a,b,c> 10 bytes <a,b,c,d> 14 bytes Signature Complexity <a> triangle count, <b> avg. area, <c> avg. height, <d> vertex count

Frame Rate High peaks result in wasted energy Target Describe why peaks are bad (above target) and valleys are bad (below bad) Low valleys result in poor visual quality

Workload prediction for DVFS Before DVFS DVFS using signature-based workload Prediction 32% energy reduction One application of workload prediction….

Outline Introduction & Background Experimental Framework Signature-Based Workload Prediction Experimental Results Conclusions

Conclusions Accurate 3D workload prediction critical for mobile platforms. Proposed signature-based method Outperforms conventional history methods Trade accuracy for time & space Can be used to meet real time constraints and conserve energy.

Future Work Automatic selection of signature elements More sophisticated data structures for signature storage Faster comparison and replacement algorithms

-Based Workload Estimation for Mobile 3D Graphics Questions? -Based Workload Estimation for Mobile 3D Graphics Bren Mochocki*†, Kanishka Lahiri*, Srihari Cadambi*, Xiaobo Sharon Hu† *NEC Laboratories America, †University of Notre Dame DAC 2006