teapot: a toolset for evaluating performance, power and...

International Conference on SupercomputingJune 2013

TEAPOT: A Toolset for Evaluating Performance, Power and Image Quality on

Mobile Graphics Systems

Jose-Maria ArnauJoan-Manuel Parcerisa

Polychronis Xekalakis

Computer Architecture DepartmentUniversitat Politecnica de Catalunya

Intel LabsIntel Corporation

Jose-Maria Arnau

MotivationMobile GPU Simulator Requirements

2

Jose-Maria Arnau

Motivation

1. Support for Mobile Applications

Mobile GPU Simulator Requirements

2

Jose-Maria Arnau

Angry BirdsAndroid

Advertisement

% of GPU time

Motivation


2. Full-System GPU Simulation


2

Jose-Maria Arnau

Angry BirdsAndroid

Advertisement

% of GPU time

Motivation



3. GPU & Screen Power Models

System Memory

GPU

Screen

% of energy


2

Jose-Maria Arnau

Angry BirdsAndroid

Advertisement

% of GPU time

Motivation




4. Flexible GPU Timing Simulator

System Memory

GPU

Screen

% of energy

Tile-Based Deferred Rendering

Immediate Mode Rendering

GPU GPU

On-Chip Memory

External Memory


2

Jose-Maria Arnau

Angry BirdsAndroid

Advertisement

% of GPU time

Motivation





System Memory

GPU

Screen

% of energy



GPU GPU

On-Chip Memory

External Memory

Not supported by any publiclyavailable GPU simulator


2

Jose-Maria Arnau

Angry BirdsAndroid

Advertisement

% of GPU time

Motivation





System Memory

GPU

Screen

% of energy



GPU GPU

On-Chip Memory

External Memory

Not supported by any publiclyavailable GPU simulator

Tailored towards desktop-likepower hungry GPUs


2

Jose-Maria Arnau

Outline

1. Motivation

2. Simulation Infrastructure

2.1. OpenGL ES Trace Generation

2.2. GPU Functional Simulation

2.3. Cycle-Accurate Timing Simulation

2.4. Power Model

2.5. Image Quality Assessment

3. Conclusions

3

Jose-Maria Arnau

OpenGLTrace

Generation

Simulation Infrastructure - Overview

Mobile Applications

Android 4.2 Jelly Bean

Android Emulator

Virtual GPUOpenGL ES

Trace Generator

Desktop GPU Driver

OpenGL ES Trace

Vertex/Fragment programs (GLSL)Textures

Geometry

Tools unmodified Tools adapted Tools created from scratch Trace files Statistics

4

Jose-Maria Arnau

GPUFunctionalSimulation

OpenGLTrace

Generation


Mobile Applications


Android Emulator


Trace Generator

Desktop GPU Driver

OpenGL ES Trace


Geometry

InstrumentedGallium3D

(softpipe driver)Frames

GPU Trace

GPU assembly instructionsMemory addresses


4

Jose-Maria Arnau

GPUTimingSimulation


OpenGLTrace

Generation


Mobile Applications


Android Emulator


Trace Generator

Desktop GPU Driver

OpenGL ES Trace


Geometry



GPU Trace


Cycle-AccurateGPU Simulator McPAT

ScreenPowerModel

ImageQuality

Assessment


4

Jose-Maria Arnau

GPUTimingSimulation


OpenGLTrace

Generation


Mobile Applications


Android Emulator


Trace Generator

Desktop GPU Driver

OpenGL ES Trace


Geometry



GPU Trace


Cycle-AccurateGPU Simulator McPAT

ScreenPowerModel

ImageQuality

Assessment

GPU Execution Time

GPU EnergyScreenEnergy

ImageQuality


4

Jose-Maria Arnau

OpenGL ES Trace GeneratorMobile Game


Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)

OpenGL ESTrace Generator

void glDrawArrays(...) { saveCommandInfo(...); real_glDrawArrays(...);}

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)



OpenGL ES TraceThread: Mobile GameOpenGL Context: A

- OpenGL ES Command List

- Shaders

- Geometry

- Textures

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)





- Shaders

- Geometry

- Textures

Virtual Buttons

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)





- Shaders

- Geometry

- Textures

Virtual Buttons

Thread: Virtual ButtonsOpenGL Context: B


- Shaders

- Geometry

- Textures

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)





- Shaders

- Geometry

- Textures

Virtual Buttons



- Shaders

- Geometry

- Textures

SurfaceFlinger

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)





- Shaders

- Geometry

- Textures

Virtual Buttons



- Shaders

- Geometry

- Textures

SurfaceFlinger

Thread: Surface FlingerOpenGL Context: C- OpenGL ES Command List

- Shaders

- Geometry

- Textures

5

Jose-Maria Arnau



Android Emulator

Virtual GPU

GPU Driver

Desktop GPU Driver

Desktop GPU

glDrawArrays(...)





- Shaders

- Geometry

- Textures

Virtual Buttons



- Shaders

- Geometry

- Textures

SurfaceFlinger

Thread: Surface FlingerOpenGL Context: C- OpenGL ES Command List

- Shaders

- Geometry

- Textures

Support for multiple applications and

OpenGL ES contexts

5

Jose-Maria Arnau

GPU Functional Simulation

OpenGL ES Trace


Geometry

Gallium3D

OpenGL ES front-end

Intermediate Representation: TGSITungsten Graphics Shader Infrastructure

Instrumented Softpipe Driver- Software Rasterizer- TGSI Emulator GPU Trace

Information stored per GPU command:

- Thread ID

- OpenGL ES Context ID

- GPU Assembly Instructions (TGSI)

- Memory addresses referenced for fetchingvertices, texels and pixels

...

6

Jose-Maria Arnau

Geometry Unit

GPU Timing Simulator● Immediate-Mode Rendering

Command Processor

Vertex Fetcher

VertexCache

Vertex Processor

PrimitiveAssembly

L2Cache

MemoryController

Fixed-Function Stage

Programmable Stage

Memory Hierarchy

GPU Trace

GPU Command 0

GPU Command 1

GPU Command 2

7

Jose-Maria Arnau

Raster Unit 1

Raster Unit 0Fragment Processor

Geometry Unit

GPU Timing Simulator● Immediate-Mode Rendering

Command Processor

Vertex Fetcher

VertexCache

Vertex Processor

PrimitiveAssembly

RasterizerEarly

Depth Test

PixelCachePixel

CacheTextureCache

ALU LD/ST

L2Cache

MemoryController


Programmable Stage

Memory Hierarchy

GPU Trace

GPU Command 0

GPU Command 1

GPU Command 2

7

Jose-Maria Arnau

GPU Timing Simulator● Tile-Based Deferred Rendering

Geometry Unit

GPU Trace

GPU Command 0

GPU Command 1

GPU Command 2

Command Processor

Vertex Fetcher

VertexCache

Vertex Processor

PrimitiveAssembly

L2Cache

MemoryController


Memory Hierarchy

Programmable Stage

8

Jose-Maria Arnau

Tiling Engine


Geometry Unit

GPU Trace

GPU Command 0

GPU Command 1

GPU Command 2

Command Processor

Vertex Fetcher

VertexCache

Vertex Processor

PrimitiveAssembly

L2Cache

MemoryController


Memory Hierarchy

Programmable Stage

PolygonList Builder

TileScheduler

TileCache

8

Jose-Maria Arnau

Tiling Engine


Raster Unit 1

Raster Unit 0Fragment Processor

Geometry Unit

GPU Trace

GPU Command 0

GPU Command 1

GPU Command 2

Command Processor

Vertex Fetcher

VertexCache

Vertex Processor

PrimitiveAssembly

RasterizerEarly

Depth Test

Z-BufferColorBuffer

TextureCache

ALU LD/ST

L2Cache

MemoryController


Memory Hierarchy

Programmable Stage

PolygonList Builder

TileScheduler

TileCache

8

Jose-Maria Arnau

GPU Timing Simulator● Fragment/Vertex Processors

● Simple in-order 4-stage pipeline● Multi-warp execution● Vectorial ISA● Texture Sampling Units

Warpscheduler

InstructionMemory

ConstantRegister

File

InputRegister

File

TemporalRegister

File

OutputRegister

File

Operand buffering& routing

SIMDALU

SFU

MEMUNIT

TEXUNIT

PixelCachePixel

Cache

TextureCache

TextureCache

Instruction Fetch Instruction Decode Execution WriteBack

9

Jose-Maria Arnau

GPU Power Model● Based on McPAT

– TEAPOT extensions:● Multiple data caches per core● Read-only caches● Specialized graphics hardware (texture sampling units...)● Output file in JSON format

– Directly called from timing simulator

Configuration filenum_raster_unitsnum_geometry_unitsnum_fragment_procsnum_vertex_procsnum_warps_per_proc...

Configuration filenum_raster_unitsnum_geometry_unitsnum_fragment_procsnum_vertex_procsnum_warps_per_proc...

Cycle-Accurate GPU Simulator

McPAT

GPU description

Area, Leakage

Activity Factors

Dynamic Power

10

Jose-Maria Arnau

Screen Power Model● OLED displays

– Consume different energy depending on the colors– Screen energy depends on the output generated by the GPU

● OLED-based displays power model– Provides three functions, f(R), f(G), f(B), that map pixel

intensity into energy consumption

M. Dong, Y.-S. K. Choi, and L. Zhong. “Power Modeling of Graphical User Interfaces on OLED Displays”. In Proc. of DAC, pages 652–657, 2009.

11

Jose-Maria Arnau

Image Quality Assessment

● Image Quality Metrics– Based on per-pixel errors

● MSE (Mean-Squared Error)● PSNR (Peak Signal-to-Noise Ratio)

– Based on the human visual perception system● MSSIM (Mean Structural SIMilarity Index)

– Require reference noise-free image for comparison

– Evaluate distortion when trading quality for energy

Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli. “Image Quality Assessment: from Error Visibility to Structural Similarity”. IEEE Transactions on Image Processing, 2004.

12

Jose-Maria Arnau

Outline

1. Motivation

2. Simulation Infrastructure

2.1. OpenGL ES Trace Generation

2.2. GPU Functional Simulation

2.3. Cycle-Accurate Timing Simulation

2.4. Power Model

2.5. Image Quality Assessment

3. Conclusions

13

Jose-Maria Arnau

Conclusions

● The TEAPOT toolset is tailored towards the mobile segment since it:– Runs unmodified Android applications

– Estimates performance, energy, area and image quality of mobile graphics systems

– Provides a flexible timing simulator, supporting Immediate-Mode and Tile-Based Deferred Rendering

– Reports statistics:● Per Application, including Android OS (full-system)● Per Frame● Per Component: GPU, System Memory and Screen

14

International Conference on SupercomputingJune 2013

TEAPOT: A Toolset for Evaluating Performance, Power and Image Quality on

Mobile Graphics Systems

Jose-Maria ArnauJoan-Manuel Parcerisa

Polychronis Xekalakis

Computer Architecture DepartmentUniversitat Politecnica de Catalunya

Intel LabsIntel Corporation

teapot: a toolset for evaluating performance, power and...

Documents