Software Performance Analysis Using CodeAnalyst for Windows

1
Software Performance Analysis UsingCodeAnalyst
for Windows

• Sherry Hurwitz
• SW Applications ManagerSRDsherry.hurwitz_at_amd.com
  
• Advanced Micro Devices

Lei Yu Member Technical StaffSRDlei.yu_at_amd.com A
dvanced Micro Devices
2
Session Outline

• Exploiting Performance Opportunities
• Obvious Performance Potential
• Hidden Performance Potential
• Exposing Untapped Performance Potential
• Analyzing Performance Improvement Trials
• AMD CodeAnalyst Performance Analysis Tool
• Capabilities of CodeAnalyst
• Functionality of CodeAnalyst
• Profile Capabilities
• Thread Analysis
• Pipeline Simulation

3
Obvious Performance Potential

• Processor Architecture
• x64 Processors
• Extended Memory Addressing
• Additional Registers
• Deeper Execution Pipeline
• Multi-Core Processors
• Multiprocessing for the desktop system
• Multiple processor platforms
• 64-bit Windows operating systems
• Compiler optimization switches
• Optimized libraries (for example AMD ACML)

4
Hidden Performance Potential

• Efficient algorithms
• Cache friendly memory access
• Branch Prediction friendly conditionals
• Parallel work through Threads
• Object Synchronization

5
Expose Untapped Performance Potential

• Profile your application with the AMD CodeAnalyst
  Performance Analyzer
• Timer-based sampling - identify time consuming or
  frequently executed code possibly pointing to
  algorithm issues (Hot Spots)
• Opteron and Athlon 64 processor performance
  events - evaluate the applications use of
  architectural features
• Thread View - evaluate effective use of multiple
  processors
• Pipeline Simulation - understand how data
  dependencies can stall the processor execution
• Iterate - between profiling and code
  modifications testing if there are performance
  benefits

6
Analyzing Performance Improvement Trials
7
Capabilities of AMD CodeAnalyst

• CodeAnalyst CAN
• Assist in optimizing your application
• Identify program bottlenecks
• Monitor and Analyze software performance
• CodeAnalyst CANNOT
• Identify defects in your program
• (Profile a functioning stable application.)
• CodeAnalyst RUNS ON
• Windows WinNT, Win2K , WinXP, 64-bit Windows
  operating systems

8
Key Functionality of AMD CodeAnalyst

• Profiling
• Timer-based sampling
• Event-based sampling
• Thread analysis
• Execution Pipeline Simulation

9
Profile Capabilities

• Low overhead system-wide profile
• Timer-based profile
• 0.1 ms resolution on APIC enabled systems
• 1.0ms resolution on APIC disabled systems
• Event-based profile
• 32 AMD Athlon and AMD Athlon XP performance
  events
• 78 AMD Opteron and AMD Athlon 64 performance
  events
• Simultaneously profile up to 4 user selected
  performance events.
• Profiles multiple processor systems up to 16
  processor cores

10
Profile Analysis

• Identifies all active Process Names, Process IDs,
  Thread IDs
• Identifies the Process CPU affinity
• Identifies performance event per CPU
• Maps sample addresses to Process, Module,
  Function, Source Line, Assembly Instruction, Code
  Byte

11
Hierarchical Navigation of Data Views

• System Data View
• System Graph View
• Module Data View
• Module Graph View
• Source View
• Disassembly View
• Demo will show the details of each of these views
  and the navigation between the views.

12
Timer-based Profiling - the First Level of
Analysis

• Exposes areas of intense activity
• Identifies the most likely suspects
• Provides a sample distribution chart
• Ability to drill down through several data views
• View the source code on and around the sample
• Algorithmic issues may be evident from the hot
  spot code
• Hot spot code might suggest particular events to
  profile in next level of Analysis

13
Common Hot Spots

• Loops
• Large content and large loop counts are natural
  hot spots but not bad for performance
• Small content with small fixed loop counts should
  be unrolled
• Remove redundant constant calculations from inner
  loops, including from inner control structures
• Long Logical Expressions in If Statements
• Long data dependent expressions
• Complicated Floating Point expressions

14
Event-based Profile - Second Level of Analysis

• Useful Events to Identify Memory Issues
• Data Cache Access and Data Cache Misses
  simultaneously
• use the ratio of Misses to Access
• Count Misaligned Data Reference
• Useful Events to Identify Branching Issues
• Retired branch mispredicted and Retired taken
  branches
• use the ratio of mispredicted to branch taken

15
Examples of Memory Issues

• Large data structures with variable size members
  not sorted by size
• Use of pointer notation in manipulating large
  data arrays
• Dereferenced pointer arguments inside a function
• Large declarations of local variables declared
  randomly with respect to size
• Memory buffers shared between threads

16
Examples of Branch Prediction Issues

• Order of the expressions in compound branch
  conditions
• Order of operands in Logical expressions
• Large switch statements with noncontiguous
  expressions
• Large switch statements cases out of order in
  respect to probability

17
Thread Analysis

• Identities threads in the target application.
• Shows Thread creation and termination
• Monitors CPU affinity of each thread
• Identifies Non-local memory access
• Graphs thread activity on each CPU

18
Thread Analysis Data View
19
Pipeline Simulation Capabilities

• CodeAnalyst can simulate a user specified block
  of code on AMD microprocessors and provide
  cycle-precise execution info.
• Requirement
• Defining a code block to simulate, requires the
  user to provide debug info for the target module.
• Limitation
• Cannot simulate instructions inside system space
• Cannot simulate multi-thread

20
Some Assumptions in the Simulator

• Assumes perfect memory subsystem
• All Load/Store Micro-ops hit in the Data Cache
• Assumes that 1 misaligned load 2 back-to-back
  aligned loads (64-bit)
• Assumes no cache bank conflicts
• 100 Instruction cache hit rate
• Assumes perfect branch prediction
• Assumes all schedulers are of infinite size

21
Pipeline Data View
22
CodeAnalyst Simulation Analysis

• User specifies Simulation configuration
• User sets Trace Point Start, Trace point End, and
  trace trigger
• Pipeline Data View
• Pipeline stage
• Penalty
• Dependency
• Delta completion
• IPC
• User can view Simulation History

23
Call to Action

• Download CodeAnalyst
• Improve Your Software!

24
Additional Resources

• Web Resources at
• http//www.developwithamd.com
• Download CodeAnalyst
• Software Optimization Guide for AMD Athlon 64 and
  AMD Opteron
• AMD64 Architecture Programmer's Manual Volume 1
  Application Programming
• AMD64 Architecture Programmer's Manual Volume 2
  System Programming
• AMD64 Architecture Programmer's Manual Volume 3
  General-Purpose and System Instructions
• AMD64 Architecture Programmer's Manual Volume 4
  128-Bit Media Instructions
• AMD64 Architecture Programmer's Manual Volume 5
  64-Bit Media and x87 Floating-Point Instructions
• http//www.devx.com
• Optimizing Your C/C Applications, Part 1 2
• Whitepapers
• Porting and Optimizing Applications on 64-bit
  Windows for AMD64 Architecture, Winhec 2004 paper
  by Mike Wall