United States-English

Parallel Programming Guide for HP-UX Systems

»

Technical documentation


	Complete book in PDF


»	Feedback

» Table of Contents

» Glossary

» Index

HP Part Number: B3909-90015

Edition: Sixth Edition

Published: September 2003

Table of Contents

Notational conventions

Command syntax

Associated documents

1 Introduction

HP SMP architectures

Bus-based systems
Hyperplane Interconnect systems

Parallel programming model

The shared-memory paradigm
The message-passing paradigm

Overview of HP optimizations

Basic scalar optimizations
Advanced scalar optimizations
Parallelization

2 Architecture overview

System architectures

Data caches
Cache thrashing

Memory Systems

Physical memory
Virtual memory
Interleaving
Variable-sized pages on HP-UX

3 Optimization levels

HP optimization levels and features

Cumulative Options

Using the Optimizer

General guidelines

4 Standard optimization features

Machine instruction level optimizations (+O0)

Constant folding
Partial evaluation of test conditions
Simple register assignment
Data alignment on natural boundaries

Block level optimizations (+O1)

Branch optimization
Dead code elimination
Faster register allocation
Instruction scheduling
Peephole optimizations

Routine level optimizations (+O2)

Advanced constant folding and propagation
Common subexpression elimination
Global register allocation (GRA)
Loop-invariant code motion
Loop unrolling
Register reassociation
Software pipelining
Strength reduction of induction variables and constants
Store and copy optimization
Unused definition elimination

5 Loop and cross-module optimization features

Strip mining

Inlining within a single source file

Cloning within a single source file

Data localization

Conditions that inhibit data localization

Loop blocking

Data reuse

Loop distribution

Loop fusion

Loop interchange

Loop unroll and jam

Preventing loop reordering

Test promotion

Cross-module cloning

Global and static variable optimizations
Inlining across multiple source files

6 Parallel optimization features

Levels of parallelism

Loop-level parallelism

Loop transformations

Idle thread states

Determining idle thread states

Parallel optimizations

Dynamic selection

Inhibiting parallelization

Loop-carried dependences (LCDs)

Preventing parallelization

Parallelism in the aC++ compiler

Cloning across multiple source files

7 Controlling optimization

Command-line optimization options

Invoking command-line options

C aliasing options

Optimization directives and pragmas

Rules for usage
block_loop[(block_factor=n)]
dynsel[(trip_count=n)]
no_block_loop
no_distribute
no_dynsel
no_loop_dependence(namelist)
no_loop_transform
no_parallel
no_side_effects(funclist)
unroll_and_jam[(unroll_factor=n)]

8 Optimization Report

Optimization Report contents

Loop Report

Supplemental tables

9 Parallel programming techniques

Parallelizing directives and pragmas

Parallelizing loops

prefer_parallel
loop_parallel
prefer_parallel, loop_parallel attributes
Combining the attributes
Comparing prefer_parallel, loop_parallel
Stride-based parallelism
critical_section, end_critical_section
Disabling automatic loop thread-parallelization

Parallelizing tasks

Parallelizing regions

Reentrant compilation

Setting thread default stack size

Modifying thread stack size

Collecting parallel information

Number of processors
Number of threads
Thread ID
Stack memory type

10 OpenMP Parallel Programming Model

What is OpenMP?

HP’s implementation of OpenMP

Command-line option
Optimization levels and parallelism
Arrays
Portable timing routines
Nested lock routines
Additional features

New library

Implementation-defined behavior

From HP Programming Model to OpenMP

Syntax
HP Programming Model directives

More information on OpenMP

11 Data privatization

Directives and pragmas for data privatization

Privatizing loop variables

loop_private
save_last[(list)]

Privatizing task variables

task_private

Privatizing region variables

parallel_private

12 Memory classes

Porting multinode applications to single-node servers

Private versus shared memory

thread_private
node_private

Memory class assignments

C and C++ data objects
Static assignments

13 Parallel synchronization

Thread-parallelism

Thread ID assignments

Synchronization tools

Using gates and barriers
Synchronization functions
sync_routine
loop_parallel(ordered)
Critical sections
Ordered sections

Synchronizing code

Using critical sections
Using ordered sections
Manual synchronization

14 Troubleshooting

ANSI algorithm
Type-safe algorithm
Specifying aliasing modes
Iteration and stop values
Global variables

False cache line sharing

Aligning data to avoid false sharing
Distributing iterations on cache line boundaries
Thread-specific array elements
Scalars sharing a cache line
Working with unaligned arrays
Working with dependences

Floating-point imprecision

Enabling sudden underflow

Invalid subscripts

Misused directives and pragmas

Loop-carried dependences
Reductions
Nondeterminism of parallel execution

Triangular loops

Parallelizing the outer loop
Parallelizing the inner loop
Examining the code

Compiler assumptions

Incrementing by zero
Trip counts that may overflow

A Porting CPSlib functions to pthreads

Introduction

Accessing pthreads

Mapping CPSlib functions to pthreads

Environment variables

Using pthreads

Symmetric parallelism
Asymmetric parallelism
Synchronization using high-level functions
Synchronization using low-level functions

List of Figures

1-1 Symmetric multiprocessor system
1-2 Message-passing programming model
2-1 K-Class bus configuration
2-2 V2250 Hyperplane Interconnect view
2-3 Array layouts—cache-thrashing
2-4 Array layouts—non-thrashing
2-5 V2250 interleaving
2-6 V2250 interleaving of arrays A and B
5-1 LCDs in original and interchanged loops
5-2 Values read into array A
5-3 Blocked array access
5-4 Spatial reuse of A and B
6-1 One-dimensional parallelism in threads
6-2 Conceptual strip mine for parallelization
6-3 Parallelized loop
9-1 Stride-parallelized loop
13-1 Ordered parallelization
13-2 LOOP_PARALLEL(ORDERED) synchronization
14-1 Data ownership by CHUNK and NTCHUNK blocks

List of Tables

1 Title not available (Notational conventions)
3-1 Locations of HP compilers
3-2 Optimization levels and features
5-1 Loop transformations affecting data localization
5-2 Form of no_loop_dependence directive and pragma
5-3 Computation sequence of A(I,J): original loop
5-4 Computation sequence of A(I,J): interchanged loop
5-5 Forms of block_loop, no_block_loop directives and pragmas
5-6 Form of no_distribute directive and pragma
5-7 Forms of unroll_and_jam, no_unroll_and_jam directives and pragmas
5-8 Form of no_loop_transform directive and pragma
6-1 Form of MP_IDLE_THREADS_WAIT environment variable
6-2 Form of dynsel directive and pragma
6-3 Form of reduction directive and pragma
6-4 Form of no_parallel directive and pragma
7-1 Command-line optimization options
7-2 +O[no]fltacc and floating-point optimizations
7-3 Optimization Report contents
7-4 +tm target and +DA/+DS
7-5 Directive-based optimization options
7-6 Form of optimization directives and pragmas
8-1 Optimization Report contents
8-2 Loop Report column definitions
8-3 Reordering transformation values in the Loop Report
8-4 Optimizing/special transformations values in the Loop Report
8-5 Analysis Table column definitions
8-6 Privatization Table column definitions
8-7 Variable Name Footnote Table column definitions
9-1 Parallel directives and pragmas
9-2 Forms of prefer_parallel and loop_parallel directives and pragmas
9-3 Attributes for loop_parallel, prefer_parallel
9-4 Comparison of loop_parallel and prefer_parallel
9-5 Iteration distribution using chunk_size = 1
9-6 Iteration distribution using chunk_size = 5
9-7 Forms of critical_section/end_critical_section directives and pragmas
9-8 Forms of task parallelization directives and pragmas
9-9 Attributes for task parallelization
9-10 Forms of region parallelization directives and pragmas
9-11 Attributes for region parallelization
9-12 Forms of CPS_STACK_SIZE environment variable
9-13 Number of processors functions
9-14 Number of threads functions
9-15 Thread ID functions
9-16 Stack memory type functions
10-1 Parallel and work-shared directives
10-2 OpenMP and HPPM Directives/Clauses
11-1 Data Privatization Directives and Pragmas
11-2 Form of loop_private directive and pragma
11-3 Form of save_last directive and pragma
11-4 Form of task_private directive and pragma
11-5 Form of parallel_private directive and pragma
12-1 Form of memory class directives and variable declarations
13-1 Forms of gate and barriers variable declarations
13-2 Forms of allocation functions
13-3 Forms of deallocation functions
13-4 Forms of locking functions
13-5 Form of unlocking functions
13-6 Form of wait functions
13-7 Form of sync_routine directive and pragma
13-8 Forms of critical_section, end_critical_section directives and pragmas
13-9 Forms of ordered_section, end_ordered_section directives and pragmas
14-1 Initial mapping of array to cache lines
14-2 Default distribution of the I loop
A-1 CPSlib library functions to pthreads mapping
A-2 CPSlib environment variables

List of Examples

2-1 Cache thrashing
2-2 Cache padding
2-3 Interleaving
4-1 Data alignment on natural boundaries
4-2 Conditional/unconditional branches
4-3 Advanced constant folding and propagation
4-4 Common subexpression elimination
4-5 Loop-invariant code motion
4-6 Loop unrolling
4-7 Loop unrolling
4-8 Register allocation
4-9 Software pipelining
4-10 Strength reduction of induction variables and constants
4-11 Unused definition elimination
5-1 Strip mining
5-2 Inlining within single source file
5-3 Loop-carried dependences
5-4 Avoid loop interchange
5-5 Aliasing
5-6 Aliasing
5-7 I/O statements
5-8 Multiple loop entries or exits
5-9 Simple loop blocking
5-10 Matrix multiply blocking
5-11 Loop blocking
5-12 Loop distribution
5-13 Loop fusion
5-14 Loop fusion
5-15 Loop peeling
5-16 Loop interchange
5-17 Unroll and jam
5-18 Test promotion
6-1 Loop-level parallelism
6-2 Loop transformations
6-3 Parallel-inhibiting LCDs
6-4 Parallel-inhibiting LCDs
6-5 Output LCDs
6-6 Apparent LCDs
6-7 Reduction
6-8 no_parallel
7-1 Data type interaction
7-2 Unsafe type cast
8-1 Optimization Report
8-2 Optimization Report
8-3 Optimization Report
9-1 prefer_parallel, loop_parallel
9-2 prefer_parallel, loop_parallel
9-3 critical_section
9-4 Parallelizing tasks
9-5 Parallelizing tasks
9-6 Region parallelization
11-1 loop_private
11-2 Using loop_private with loop_parallel
11-3 Denoting induction variables in parallel loops
11-4 Denoting induction variables in parallel loops
11-5 Secondary induction variables
11-6 Secondary induction variables
11-7 save_last
11-8 save_last
11-9 task_private
11-10 parallel_private
12-1 thread_private
12-2 thread_private
12-3 thread_private COMMON blocks in parallel subroutines
12-4 node_private
13-1 sync_routine
13-2 sync_routine
13-3 loop_parallel, ordered
13-4 critical_section
13-5 Gated critical sections
13-6 Ordered sections
13-7 Ordered section limitations
13-8 Ordered section limitations
13-9 Critical sections and gates
13-10 Conditionally lock critical sections

Printable version


Privacy statement	Using this site means you accept its terms	Feedback to webmaster

© Hewlett-Packard Development Company, L.P.