Get documents about "Slide 1
Shared by: pX7oV0
-
Stats
- views:
- 0
- posted:
- 1/29/2013
- language:
- English
- pages:
- 32
Document Sample
Emergent Game Technologies
Gamebryo Element Engine
Thread for Performance
Goals for Cross-Platform Threading
•Play well with others
•Take advantage of platform-specific
performance features
•For engines/middleware, be adaptable to
the needs of customers
2
Write Once, Use Everywhere
•Underlying multi-threaded primitives are
replicated on all platforms
– Define cross-platform wrappers for these
•Processing models can be applied on
different architectures
– Define cross-platform systems for these
•Typical developer writes once, yet code
performs well on all platforms
3
Emergent's Gamebryo Element
•A foundation for easing cross-platform and
multi-core development
–Modular, customizable
–Suite of content pipeline tools
–Supports PC, Xbox, PS3 and Wii
•Booth # 5716 - North Hall
4
Cross-Platform Threading Requires
Common Primitives
•Threads
– Something that executes code
– Sub issues: local storage, priorities
•Data Locks / Critical sections
– Manage contention for a resource
•Atomic operations
– An operation that is guaranteed to complete
without interruption from another thread
5
Choosing a Processing Model
•Architectural features drive choice
– Cache coherence
– Prefetch on Xbox
– SPUs on PS3
– Many processing units
– General purpose GPU
•Stream Processing fits these properties
– Provide infrastructure to compute this way
– Shift engine work to this model
6
Stream Processing (Formal)
Wikipedia: Given a set of input and output
data (streams), the principle essentially
defines a series of computer-intensive
operations (kernel functions) to be applied
for each element in the stream.
Input 1
Kernel 1
Kernel 2
Output
Input 2
7
Generalized Stream Processing
•Improve for general purpose computing
– Partition streams into chunks
– Kernels have access to entire chunk
– Parameters for kernels (fixed inputs)
•Advantages
– Reduce need for strict data locality
– Enables loops, non-SIMD processing
– Maps better onto hardware
8
Morphing+Skinning Example
Morph Kernel Skin Vertices
(MK)
Bone Matrices Skinning Kernel (SK)
Blend Weights
Morph Weights
Morph Target 1 Vertices
Morph Target 2 Vertices
Vertex Locations
9
Morphing+Skinning Example
MT 1 V Part 1
Skin V Part 1
Verts Part 1 Verts Part 2
MT 2 V Part 1 MK SK
Instance 1 Instance 1
Weights Fixed
MW Fixed
Matrices Fixed
MK SK
MT 1 V Part 2 Instance 2 Instance 2
Skin V Part 2
MT 2 V Part 2
10
Floodgate
•Cross platform stream processing library
•Optimized per-platform implementation
•Documented API for customer use
•Engine uses the same API for built in
functionality
– Skinning, Morphing, Particles, Instance Culling,
...
11
Floodgate Basics
•Stream: A buffer of varying or fixed data
– A pointer, length, stride, locking
•Kernel: An operation to perform on streams
of data
– Code implementing “Execute” function
•Task: Wrapper a kernel and IO streams
•Workflow: A collection of Tasks processed
as a unit
12
Kernel Example: Times2
// Include Kernel Definition macros
#include <NiSPKernelMacros.h>
// Declare the Timer2Kernel
NiSPDeclareKernel(Times2Kernel)
13
Kernel Example: Times2
#include "Times2Kernel.h"
NiSPBeginKernelImpl(Times2Kernel)
{
// Get the input stream
float *pInput = kWorkload.GetInput<float>(0);
// Get the output stream
float *pOutput = kWorkload.GetOutput<float>(0);
// Process data
NiUInt32 uiBlockCount = kWorkload.GetBlockCount();
for (NiUInt32 ui = 0; ui < uiBlockCount; ui++)
{
pOutput[ui] = pInput[ui] * 2;
}
}
NiSPEndKernelImpl(Times2Kernel)
14
Life of a Workflow
•1. Obtain Workflow from Floodgate
•2. Add Task(s) to Workflow
•3. Set Kernel
•4. Add Input Streams
•5. Add Output Streams
•6. Submit Workflow
•… Do something else …
•7. Wait or Poll when results are needed
15
Example Workflow
// Setup input and output streams from existing buffers
NiTSPStream<float> inputStream(SomeInputBuffer, MAX_BLOCKS);
NiTSPStream<float> outputStream(SomeOutputBuffer, MAX_BLOCKS);
// Get a Workflow and setup a new task for it
NiSPWorkflow* pWorkflow = NiStreamProcessor::Get()-
>GetFreeWorkflow();
NiSPTask* pTask = pWorkflow->AddNewTask();
// Set the kernel and streams
pTask->SetKernel(&Times2Kernel);
pTask->AddInput(&inputStream);
pTask->AddOutput(&outputStream);
// Submit workflow for execution
NiStreamProcessor::Get()->Submit(pWorkflow);
// Do other operations...
// Wait for workflow to complete
NiStreamProcessor::Get()->Wait(pWorkflow);
16
Floodgate Internals
•Partitioning streams for Tasks
•Task Dependency Analysis
•Platform specific Workflow preparation
•Platform specific execution
•Platform specific synchronization
17
Overview of Workflow Analysis
•Task dependencies defined by streams
•Sort tasks into stages of execution
– Tasks that use results from other tasks run in
later stages
– Stage N+1 tasks depend on output of Stage N
tasks
•Tasks in a given stage can run concurrent
•Once a stage has completed, the next stage
can run
18
Analysis: Workflow with many
Tasks
Stream A Task 1 Stream B Stream C Task 2 Stream D Stream E Task 3 Stream F
Stream G Task 5 Stream H
Task 4 Stream G Task 6 Stream I
Task 7 Sync
19
Analysis: Dependency Graph
Stage 0 Stage 1 Stage 2 Stage 3
Stream A Task 1
Task 4 Stream G Task 5 Stream H
Stream C Task 2
Sync Sync
Task
Stream E Task 3 Stream F Task 6 Stream I
20
Performance Notes
•Data is broken into blocks -> Locality
– Good cache performance
– Optimize size for prefetch or DMA transfers
– Fits in limited local storage (PS3)
•Easily adapt to #cores
– Can manage interplay with other systems
•Kernels encapsulate processing
– Good target for optimization, platform-specific
– Clean solution without #if
21
Usability Notes
•Automatically manage data dependency and
simplify synchronization
•Hide nasty platform-specific details
– Prefetch, DMA transfers, processor detection, ...
•Learn one API, use it across platforms
– Productivity gains
–Helps us produce quality documentation and
samples
– Eases debugging
22
Exploiting Floodgate in the Engine
•Find tasks that operate on a single object
– Skinning, morphing, particle systems, ...
• Move these to Floodgate: Mesh Modifiers
– Launch at some point during execution
–After updating animation and bounds
–After determining visibility
–After physics finishes ...
– Finish them when needed
–Culling
–Render
–etc
23
Same applications, new
performance ...
Before After
Skinning Objects 42fps 62fps
Morphing Objects 12fps 38fps
•The big win is out-of-the-box
performance
– Same results could be achieved
with much developer time
– Hides details on different
platforms (esp. PS3)
24
Example CPU Utilization, Morphing
Before
After
25
Thread profiling, Morphing Before
•Some parallelization through hand-coded
parallel update
– Note high overhead and 85% or so in serial
execution
26
Thread profiling, Morphing After
•Automatic parallelism in engine
–4 threads for Floodgate (4 CPUs)
–Roughly, 50% of old serial time replaced
with 4x parallelism
27
New Issues
•Within the engine, resource usage peaks at
certain times
– e.g. Between visibility culling and rendering
– Application-level work might fill in the empty
spaces
–Physics, global illumination, ...
•What about single processor machines?
•What about variable sized output?
– Instance culling, for example
28
Ongoing Improvements
•Improved workflow scheduling
– Mechanisms to enhance application control
•Optimizing when tasks change
– Stream lengths change
– Inputs/outputs are changed
•More platform specific improvements
•Off-loading more engine work
29
Using Floodgate in a game
•Identify stream processing opportunities
– Places where lots of data is processed with local
access patterns
– Places where work can be prepared early but
results are not needed until later
•Re-factor to use Floodgate
– Depending on task, could be as little as a few
hours.
– Hard part is enforcing locality
30
Future proofed?
•Both CPUs and GPUs can function as stream
processors
•Easily extends to more processing units
•Potential snags are in application changes
31
Questions?
•Ask Stephen!
•Visit Emergent's booth at the show.
– Booth 5716, North Hall, opposite Intel on the
central aisle
32
