TMS320C55x DSP

Programmer’s Guide

Preliminary Draft

This document contains preliminary data

current as of the publication date and is
subject to change without notice.

SPRU376A
August 2001
 IMPORTANT NOTICE
Texas Instruments and its subsidiaries (TI) reserve the right to make changes to their products
or to discontinue any product or service without notice, and advise customers to obtain the latest
version of relevant information to verify, before placing orders, that information being relied on
is current and complete. All products are sold subject to the terms and conditions of sale supplied
at the time of order acknowledgment, including those pertaining to warranty, patent infringement,
and limitation of liability.
TI warrants performance of its products to the specifications applicable at the time of sale in
accordance with TI’s standard warranty. Testing and other quality control techniques are utilized
to the extent TI deems necessary to support this warranty. Specific testing of all parameters of
each device is not necessarily performed, except those mandated by government requirements.
Customers are responsible for their applications using TI components.
In order to minimize risks associated with the customer’s applications, adequate design and
operating safeguards must be provided by the customer to minimize inherent or procedural
hazards.
TI assumes no liability for applications assistance or customer product design. TI does not
warrant or represent that any license, either express or implied, is granted under any patent right,
copyright, mask work right, or other intellectual property right of TI covering or relating to any
combination, machine, or process in which such products or services might be or are used. TI’s
publication of information regarding any third party’s products or services does not constitute TI’s
approval, license, warranty or endorsement thereof.
Reproduction of information in TI data books or data sheets is permissible only if reproduction
is without alteration and is accompanied by all associated warranties, conditions, limitations and
notices. Representation or reproduction of this information with alteration voids all warranties
provided for an associated TI product or service, is an unfair and deceptive business practice,
and TI is not responsible nor liable for any such use.
Resale of TI’s products or services with statements different from or beyond the parameters stated
by TI for that products or service voids all express and any implied warranties for the associated
TI product or service, is an unfair and deceptive business practice, and TI is not responsible nor
liable for any such use.
Also see: Standard Terms and Conditions of Sale for Semiconductor Products.
www.ti.com/sc/docs/stdterms.htm

Mailing Address:

Texas Instruments
Post Office Box 655303
Dallas, Texas 75265

Copyright  2001, Texas Instruments Incorporated

 Preface

 
 

About This Manual

This manual describes ways to optimize C and assembly code for the
TMS320C55x DSPs and recommends ways to write TMS320C55x code for
specific applications.

Notational Conventions
This document uses the following conventions.
- The device number TMS320C55x is often abbreviated as C55x.

- Program listings, program examples, and interactive displays are shown

in a special typeface similar to a typewriter’s. Examples use a bold
version of the special typeface for emphasis; interactive displays use a
bold version of the special typeface to distinguish commands that you
enter from items that the system displays (such as prompts, command
output, error messages, etc.).
Here is a sample program listing:
0011 0005 0001 .field 1, 2
0012 0005 0003 .field 3, 4
0013 0005 0006 .field 6, 3
0014 0006 .even
Here is an example of a system prompt and a command that you might
enter:
C: csr −a /user/ti/simuboard/utilities
- In syntax descriptions, the instruction, command, or directive is in a bold
typeface font and parameters are in an italic typeface. Portions of a syntax
that are in bold should be entered as shown; portions of a syntax that are
in italics describe the type of information that should be entered. Here is
an example of a directive syntax:
.asect “section name”, address
.asect is the directive. This directive has two parameters, indicated by sec-
tion name and address. When you use .asect, the first parameter must be
an actual section name, enclosed in double quotes; the second parameter
must be an address.

Read This First iii

Notational Conventions

- Some directives can have a varying number of parameters. For example,

the .byte directive can have up to 100 parameters. The syntax for this di-
rective is:
.byte value1 [, ... , valuen ]
This syntax shows that .byte must have at least one value parameter, but
you have the option of supplying additional value parameters, separated
by commas.

- In most cases, hexadecimal numbers are shown with the suffix h. For ex-
ample, the following number is a hexadecimal 40 (decimal 64):
40h
Similarly, binary numbers usually are shown with the suffix b. For example,
the following number is the decimal number 4 shown in binary form:
0100b

- Bits are sometimes referenced with the following notation:

Notation Description Example

Register(n−m) Bits n through m of Register AC0(15−0) represents the 16

least significant bits of the regis-
ter AC0.

iv
 Related Documentation From Texas Instruments

Related Documentation From Texas Instruments

The following books describe the TMS320C55x devices and related support
tools. To obtain a copy of any of these TI documents, call the Texas
Instruments Literature Response Center at (800) 477-8924. When ordering,
please identify the book by its title and literature number.

TMS320C55x Technical Overview (literature number SPRU393). This over-

view is an introduction to the TMS320C55x digital signal processor
(DSP). The TMS320C55x is the latest generation of fixed-point DSPs in
the TMS320C5000 DSP platform. Like the previous generations, this
processor is optimized for high performance and low-power operation.
This book describes the CPU architecture, low-power enhancements,
and embedded emulation features of the TMS320C55x.

TMS320C55x DSP CPU Reference Guide (literature number SPRU371)

describes the architecture, registers, and operation of the CPU.

TMS320C55x DSP Mnemonic Instruction Set Reference Guide (literature

number SPRU374) describes the mnemonic instructions individually. It
also includes a summary of the instruction set, a list of the instruction
opcodes, and a cross-reference to the algebraic instruction set.

TMS320C55x DSP Algebraic Instruction Set Reference Guide (literature

number SPRU375) describes the algebraic instructions individually. It
also includes a summary of the instruction set, a list of the instruction
opcodes, and a cross-reference to the mnemonic instruction set.

TMS320C55x Optimizing C Compiler User’s Guide (literature number

SPRU281) describes the C55x C compiler. This C compiler accepts
ANSI standard C source code and produces assembly language source
code for TMS320C55x devices.

TMS320C55x Assembly Language Tools User’s Guide (literature number

SPRU280) describes the assembly language tools (assembler, linker,
and other tools used to develop assembly language code), assembler
directives, macros, common object file format, and symbolic debugging
directives for TMS320C55x devices.

TMS320C55x DSP Library Programmer’s Reference (literature number

SPRU422) describes the optimized DSP Function Library for C program-
mers on the TMS320C55x DSP.

The CPU, the registers, and the instruction sets are also described in online
documentation contained in Code Composer Studio.

Read This First v

Trademarks

Trademarks

Code Composer Studio, TMS320C54x, C54x, TMS320C55x, and C55x are

trademarks of Texas Instruments.

vi
 Contents



1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1
Lists some key features of the TMS320C55x DSP architecture and recommends a process for
code development.
1.1 TMS320C55x Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-2
1.2 Code Development Flow for Best Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
2 Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
Uses example code to walk you through the code development flow for the TMS320C55x DSP.
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.2 Writing Assembly Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3
2.2.1 Allocate Sections for Code, Constants, and Variables . . . . . . . . . . . . . . . . . . . . 2-5
2.2.2 Processor Mode Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.2.3 Setting up Addressing Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-8
2.3 Understanding the Linking Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10
2.4 Building Your Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.4.1 Creating a Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.4.2 Adding Files to the Workspace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-14
2.4.3 Modifying Build Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-17
2.4.4 Building the Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-18
2.5 Testing Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-19
2.6 Benchmarking Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21
3 Optimizing C Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
Describes how you can maximize the performance of your C code by using certain compiler
options, C code transformations, and compiler intrinsics.
3.1 Introduction to Writing C/C++ Code for a C55x DSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.1.1 Tips on Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.1.2 How to Write Multiplication Expressions Correctly in C Code . . . . . . . . . . . . . . 3-3
3.1.3 Memory Dependences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3.1.4 Analyzing C Code Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
3.2 Compiling the C/C++ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
3.2.1 Compiler Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
3.2.2 Performing Program-Level Optimization (−pm Option) . . . . . . . . . . . . . . . . . . . . 3-9
3.2.3 Using Function Inlining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-11
3.3 Profiling Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.3.1 Using the clock() Function to Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.3.2 Using CCS 2.0 to Profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14

vii
Contents

3.4 Refining the C/C++ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15

3.4.1 Generating Efficient Loop Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-16
3.4.2 Efficient Use of MAC hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-21
3.4.3 Using Intrinsics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
3.4.4 Using Long Data Accesses for 16-Bit Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-34
3.4.5 Simulating Circular Addressing in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-35
3.4.6 Generating Efficient Control Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-39
3.4.7 Summary of Coding Idioms for C55x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-40
3.5 Memory Management Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-42
3.5.1 Avoiding Holes Caused by Data Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-42
3.5.2 Local vs. Global Symbol Declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-43
3.5.3 Stack Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-43
3.5.4 Allocating Code and Data in the C55x Memory Map . . . . . . . . . . . . . . . . . . . . . 3-44
3.5.5 Allocating Function Code to Different Sections . . . . . . . . . . . . . . . . . . . . . . . . . 3-48

4 Optimizing Assembly Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1

Describes some of the opportunities for optimizing TMS320C55x assembly code and provides
corresponding code examples.
4.1 Efficient Use of the Dual-MAC Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
4.1.1 Implicit Algorithm Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-4
4.1.2 Loop Unrolling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-6
4.1.3 Multichannel Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-14
4.1.4 Multi-Algorithm Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-15
4.2 Using Parallel Execution Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-16
4.2.1 Built-In Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-16
4.2.2 User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-16
4.2.3 Architectural Features Supporting Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . 4-17
4.2.4 User-Defined Parallelism Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-20
4.2.5 Process for Implementing User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . 4-22
4.2.6 Parallelism Tips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-24
4.2.7 Examples of Parallel Optimization Within CPU Functional Units . . . . . . . . . . 4-25
4.2.8 Example of Parallel Optimization Across the A-Unit, P-Unit, and D-Unit . . . . 4-35
4.3 Implementing Efficient Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-42
4.3.1 Nesting of Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-42
4.3.2 Efficient Use of repeat(CSR) Looping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-46
4.3.3 Avoiding Pipeline Delays When Accessing Loop-Control Registers . . . . . . . . 4-48
4.4 Minimizing Pipeline and IBQ Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-49
4.4.1 Process to Resolve Pipeline and IBQ Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . 4-53
4.4.2 Recommendations for Preventing Pipeline Delays . . . . . . . . . . . . . . . . . . . . . . 4-54
4.4.3 Memory Accesses and the Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-72
4.4.4 Recommendations for Preventing IBQ Delays . . . . . . . . . . . . . . . . . . . . . . . . . . 4-79

viii
 Contents

5 Fixed-Point Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1

Explains important considerations for doing fixed-point arithmetic with the TMS320C55x DSP.
Includes code examples.
5.1 Fixed-Point Arithmetic − a Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2
5.1.1 2s-Complement Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2
5.1.2 Integers Versus Fractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
5.1.3 2s-Complement Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5.2 Extended-Precision Addition and Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9
5.2.1 A 64-Bit Addition Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9
5.2.2 A 64-Bit Subtraction Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-11
5.3 Extended-Precision Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
5.4 Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-17
5.4.1 Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-17
5.4.2 Fractional Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-23
5.5 Methods of Handling Overflows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-24
5.5.1 Hardware Features for Overflow Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-24
6 Bit-Reversed Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
Introduces bit-reverse addressing and its implementation on the TMS320C55x DSP. Includes
code examples.
6.1 Introduction to Bit-Reverse Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6.2 Using Bit-Reverse Addressing In FFT Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.3 In-Place Versus Off-Place Bit-Reversing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6.4 Using the C55x DSPLIB for FFTs and Bit-Reversing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-7
7 Application-Specific Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1
Explains how to implement some common DSP algorithms using specialized TMS320C55x
instructions. Includes code examples.
7.1 Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN) . . . . . . . . . . . . . . . . . . . . . . . . . 7-2
7.1.1 Symmetric FIR Filtering With the firs Instruction . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7.1.2 Antisymmetric FIR Filtering With the firsn Instruction . . . . . . . . . . . . . . . . . . . . . 7-4
7.1.3 Implementation of a Symmetric FIR Filter on the TMS320C55x DSP . . . . . . . 7-4
7.2 Adaptive Filtering (LMS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6
7.2.1 Delayed LMS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-7
7.3 Convolutional Encoding (BFXPA, BFXTR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-10
7.3.1 Bit-Stream Multiplexing and Demultiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-12
7.4 Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF) . . . . . . . . . 7-16
8 TI C55x DSPLIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-1
Introduces the features and the C functions of the TI TMS320C55x DSP function library.
8.1 Features and Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.2 DSPLIB Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.3 DSPLIB Arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.4 Calling a DSPLIB Function from C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3
8.5 Calling a DSPLIB Function from Assembly Language Source Code . . . . . . . . . . . . . . . 8-4
8.6 Where to Find Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
8.7 DSPLIB Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5

Contents ix
Contents

A Special D-Unit Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1

Lists the D-unit instructions where A-unit registers are read (“pre-fetched”) in the Read phase
of the execution pipeline.

B Algebraic Instructions Code Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-1

Shows the algebraic instructions code examples that correspond to the mnemonic instructions
code examples shown in Chapters 2 through 7.

x
 Figures


1−1 Code Development Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3
2−1 Section Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2−2 Extended Auxiliary Registers Structure (XARn) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9
2−3 Project Creation Dialog Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-14
2−4 Add tutor.asm to Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-15
2−5 Add tutor.cmd to Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-16
2−6 Build Options Dialog Box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-17
2−7 Rebuild Complete Screen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-18
4−1 Data Bus Usage During a Dual-MAC Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
4−2 Computation Groupings for a Block FIR (4-Tap Filter Shown) . . . . . . . . . . . . . . . . . . . . . . . 4-7
4−3 Computation Groupings for a Single-Sample FIR With an
Even Number of TAPS (4-Tap Filter Shown) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-12
4−4 Computation Groupings for a Single-Sample FIR With an
Odd Number of TAPS (5-Tap Filter Shown) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13
4−5 Matrix to Find Operators That Can Be Used in Parallel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-18
4−6 CPU Operators and Buses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-20
4−7 Process for Applying User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-23
4−8 First Segment of the Pipeline (Fetch Pipeline) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-49
4−9 Second Segment of the Pipeline (Execution Pipeline) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-50
3−1 Dependence Graph for Vector Sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
5−1 4-Bit 2s-Complement Integer Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-4
5−2 8-Bit 2s-Complement Integer Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-4
5−3 4-Bit 2s-Complement Fractional Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5−4 8-Bit 2s-Complement Fractional Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
5−5 Effect on CARRY of Addition Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-11
5−6 Effect on CARRY of Subtraction Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-13
5−7 32-Bit Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
6−1 FFT Flow Graph Showing Bit-Reversed Input and In-Order Output . . . . . . . . . . . . . . . . . . 6-4
7−1 Symmetric and Antisymmetric FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7−2 Adaptive FIR Filter Implemented With the Least-Mean-Squares (LMS) Algorithm . . . . . . 7-6
7−3 Example of a Convolutional Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-10
7−4 Generation of an Output Stream G0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-11
7−5 Bit Stream Multiplexing Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-12
7−6 Butterfly Structure for K = 5, Rate 1/2 GSM Convolutional Encoder . . . . . . . . . . . . . . . . . 7-17

Contents xi
Tables


3−1 Compiler Options to Avoid on Performance-Critical Code . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
3−2 Compiler Options for Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3−3 Compiler Options That May Degrade Performance and Improve Code Size . . . . . . . . . . . 3-8
3−4 Compiler Options for Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
3−5 Summary of C/C++ Code Optimization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15
3−6 TMS320C55x C Compiler Intrinsics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-32
3−7 C Coding Methods for Generating Efficient C55x Assembly Code . . . . . . . . . . . . . . . . . . 3-40
3−8 Section Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-45
3−9 Possible Operand Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
4−1 CPU Data Buses and Constant Buses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-19
4−2 Basic Parallelism Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-21
4−3 Advanced Parallelism Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-22
4−4 Steps in Process for Applying User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-23
4−5 Pipeline Activity Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-51
4−6 Recommendations for Preventing Pipeline Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-54
4−7 Bit Groups for STx Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-63
4−8 Pipeline Register Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-64
4−9 Memory Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-72
4−10 C55x Data and Program Buses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-73
4−11 Half-Cycle Accesses to Dual-Access Memory (DARAM) and the Pipeline . . . . . . . . . . . . 4-73
4−12 Memory Accesses and the Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-74
4−13 Cross-Reference Table Documented By Software Developers to Help
Software Integrators Generate an Optional Application Mapping . . . . . . . . . . . . . . . . . . . 4-78
6−1 Syntaxes for Bit-Reverse Addressing Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6−2 Bit-Reversed Addresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6−3 Typical Bit-Reverse Initialization Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
7−1 Operands to the firs or firsn Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4

xii
 Examples


2−1 Final Assembly Code of tutor.asm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4
2−2 Partial Assembly Code of tutor.asm (Step 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-6
2−3 Partial Assembly Code of tutor.asm (Step 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2−4 Partial Assembly Code of tutor.asm (Part3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9
2−5 Linker command file (tutor.cmd) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10
2−6 Linker map file (test.map) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-11
2−7 x Memory Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-20
3−1 Generating a 16x16−>32 Multiply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
3−2 C Code for Vector Sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3−3 Main Function File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3−4 Sum Function File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
3−5 Assembly Code Generated With −o3 and −pm Options . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-11
3−6 Assembly Generated Using −o3, −pm, and −oi50 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-12
3−7 Using the clock() Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3−8 Simple Loop That Allows Use of localrepeat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-16
3−9 Assembly Code for localrepeat Generated by the Compiler . . . . . . . . . . . . . . . . . . . . . . . . 3-17
3−10 Inefficient Loop Code for Loop Variable and Constraints (C) . . . . . . . . . . . . . . . . . . . . . . . 3-18
3−11 Inefficient Loop Code for Variable and Constraints (Assembly) . . . . . . . . . . . . . . . . . . . . . 3-19
3−12 Using the MUST_ITERATE Pragma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
3−13 Assembly Code Generated With the MUST_ITERATE Pragma . . . . . . . . . . . . . . . . . . . . . 3-20
3−14 Use Local Rather Than Global Summation Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-22
3−15 Returning Q15 Result for Multiply Accumulate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-22
3−16 C Code for an FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
3−17 FIR C Code After Unroll-and-Jam Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-25
3−18 FIR Filter With MUST_ITERATE Pragma and restrict Qualifier . . . . . . . . . . . . . . . . . . . . . 3-27
3−19 Generated Assembly for FIR Filter Showing Dual-MAC . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
3−20 Implementing Saturated Addition in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-29
3−21 Inefficient Assembly Code Generated by C Version of Saturated Addition . . . . . . . . . . . . 3-30
3−22 Single Call to _sadd Intrinsic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-31
3−23 Assembly Code Generated When Using Compiler Intrinsic for Saturated Add . . . . . . . . 3-31
3−24 Using ETSI Functions to Implement sadd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-31
3−25 Block Copy Using Long Data Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-34
3−26 Simulating Circular Addressing in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-35
3−27 Assembly Output for Circular Addressing C Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-36
3−28 Circular Addressing Using Modulus Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-37
3−29 Assembly Output for Circular Addressing Using Modulus Operator . . . . . . . . . . . . . . . . . 3-38

Contents xiii
Examples

3−30 Considerations for Long Data Objects in Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-42

3−31 Declaration Using DATA_SECTION Pragma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
3−32 Sample Linker Command File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-47
4−1 Complex Vector Multiplication Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-6
4−2 Block FIR Filter Code (Not Optimized) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
4−3 Block FIR Filter Code (Optimized) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10
4−4 A-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-26
4−5 A-Unit Code in Example 4−4 Modified to Take Advantage of Parallelism . . . . . . . . . . . . . 4-28
4−6 P-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-31
4−7 P-Unit Code in Example 4−6 Modified to Take Advantage of Parallelism . . . . . . . . . . . . . 4-32
4−8 D-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-34
4−9 D-Unit Code in Example 4−8 Modified to Take Advantage of Parallelism . . . . . . . . . . . . . 4-35
4−10 Code That Uses Multiple CPU Units But No User-Defined Parallelism . . . . . . . . . . . . . . . 4-36
4−11 Code in Example 4−10 Modified to Take Advantage of Parallelism . . . . . . . . . . . . . . . . . . 4-39
4−12 Nested Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-43
4−13 Branch-On-Auxiliary-Register-Not-Zero Construct
(Shown in Complex FFT Loop Code) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-44
4−14 Use of CSR (Shown in Real Block FIR Loop Code) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-47
4−15 A-Unit Register (Write in X Phase/Read in AD Phase) Sequence . . . . . . . . . . . . . . . . . . . 4-56
4−16 A-Unit Register Read/(Write in AD Phase) Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-58
4−17 Register (Write in X Phase)/(Read in R Phase) Sequence . . . . . . . . . . . . . . . . . . . . . . . . . 4-59
4−18 Good Use of MAR Instruction (Write/Read Sequence) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-60
4−19 Bad Use of MAR Instruction (Read/Write Sequence) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-61
4−20 Solution for Bad Use of MAR Instruction (Read/Write Sequence) . . . . . . . . . . . . . . . . . . . 4-61
4−21 Stall During Decode Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-62
4−22 Unprotected BRC Write . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-65
4−23 BRC Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-66
4−24 CSR Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-67
4−25 Condition Evaluation Preceded by a X-Phase Write to the Register
Affecting the Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-68
4−26 Making an Operation Conditional With XCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-69
4−27 Making an Operation Conditional With execute(D_unit) . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-70
4−28 Conditional Parallel Write Operation Followed by an
AD-Phase Write to the Register Affecting the Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-71
4−29 A Write Pending Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-76
4−30 A Memory Bypass Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-77
3−33 Allocation of Functions Using CODE_SECTION Pragma . . . . . . . . . . . . . . . . . . . . . . . . . . 3-48
5−1 Signed 2s-Complement Binary Number Expanded to Decimal Equivalent . . . . . . . . . . . . 5-3
5−2 Computing the Negative of a 2s-Complement Number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-3
5−3 Addition With 2s-Complement Binary Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-6
5−4 Subtraction With 2s-Complement Binary Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-7
5−5 Multiplication With 2s-Complement Binary Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-8
5−6 64-Bit Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-10
5−7 64-Bit Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-12

xiv
 Examples

5−8 32-Bit Integer Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-15

5−9 32-Bit Fractional Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-16
5−10 Unsigned, 16-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-18
5−11 Unsigned, 32-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-19
5−12 Signed, 16-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-21
5−13 Signed, 32-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-22
6−1 Sequence of Auxiliary Registers Modifications in Bit-Reversed Addressing . . . . . . . . . . . 6-2
6−2 Off-Place Bit Reversing of a Vector Array (in Assembly) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6−3 Using DSPLIB cbrev() Routine to Bit Reverse a Vector Array (in C) . . . . . . . . . . . . . . . . . . 6-7
7−1 Symmetric FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
7−2 Delayed LMS Implementation of an Adaptive Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-9
7−3 Generation of Output Streams G0 and G1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-11
7−4 Multiplexing Two Bit Streams With the Field Expand Instruction . . . . . . . . . . . . . . . . . . . . 7-13
7−5 Demultiplexing a Bit Stream With the Field Extract Instruction . . . . . . . . . . . . . . . . . . . . . . 7-15
7−6 Viterbi Butterflies for Channel Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-19
7−7 Viterbi Butterflies Using Instruction Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-20
B−1 Partial Assembly Code of test.asm (Step 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-2
B−2 Partial Assembly Code of test.asm (Step 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-2
B−3 Partial Assembly Code of test.asm (Part3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-3
B−4 Assembly Code Generated With −o3 and −pm Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-4
B−5 Assembly Generated Using −o3, −pm, and −oi50 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-5
B−6 Assembly Code for localrepeat Generated by the Compiler . . . . . . . . . . . . . . . . . . . . . . . . . B-5
B−7 Inefficient Loop Code for Variable and Constraints (Assembly) . . . . . . . . . . . . . . . . . . . . . . B-6
B−8 Assembly Code Generated With the MUST_ITERATE Pragma . . . . . . . . . . . . . . . . . . . . . . B-6
B−9 Generated Assembly for FIR Filter Showing Dual-MAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-7
B−10 Inefficient Assembly Code Generated by C Version of Saturated Addition . . . . . . . . . . . . . B-8
B−11 Assembly Code Generated When Using Compiler Intrinsic for Saturated Add . . . . . . . . . B-9
B−12 Assembly Output for Circular Addressing C Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-9
B−13 Assembly Output for Circular Addressing Using Modulo . . . . . . . . . . . . . . . . . . . . . . . . . . . B-10
B−14 Complex Vector Multiplication Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-11
B−15 Block FIR Filter Code (Not Optimized) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-12
B−16 Block FIR Filter Code (Optimized) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-13
B−17 A-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-14
B−18 A-Unit Code in Example B−17 Modified to Take Advantage of Parallelism . . . . . . . . . . . B-16
B−19 P-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-18
B−20 P-Unit Code in Example B−19 Modified to Take Advantage of Parallelism . . . . . . . . . . . B-19
B−21 D-Unit Code With No User-Defined Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-20
B−22 D-Unit Code in Example B−21 Modified to Take Advantage of Parallelism . . . . . . . . . . . B-21
B−23 Code That Uses Multiple CPU Units But No User-Defined Parallelism . . . . . . . . . . . . . . . B-22
B−24 Code in Example B−23 Modified to Take Advantage of Parallelism . . . . . . . . . . . . . . . . . B-25
B−25 Nested Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-28
B−26 Branch-On-Auxiliary-Register-Not-Zero Construct
(Shown in Complex FFT Loop Code) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-28
B−27 64-Bit Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-30

Contents xv
Examples

B−28 64-Bit Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-31

B−29 32-Bit Integer Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-32
B−30 32-Bit Fractional Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-33
B−31 Unsigned, 16-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-33
B−32 Unsigned, 32-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-34
B−33 Signed, 16-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-35
B−34 Signed, 32-Bit By 16-Bit Integer Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-36
B−35 Off-Place Bit Reversing of a Vector Array (in Assembly) . . . . . . . . . . . . . . . . . . . . . . . . . . . B-37
B−36 Delayed LMS Implementation of an Adaptive Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-38
B−37 Generation of Output Streams G0 and G1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-39
B−38 Viterbi Butterflies for Channel Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-39
B−39 Viterbi Butterflies Using Instruction Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-40

xvi
 Chapter 1



This chapter lists some of the key features of the TMS320C55x (C55x) DSP
architecture and shows a recommended process for creating code that runs
efficiently.

Topic Page

1.1 TMS320C55x Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-2

1.2 Code Development Flow for Best Performance . . . . . . . . . . . . . . . . . . 1-3

1-1
TMS320C55x Architecture

1.1 TMS320C55x Architecture

The TMS320C55x device is a fixed-point digital signal processor (DSP). The
main block of the DSP is the central processing unit (CPU), which has the fol-
lowing characteristics:

- A unified program/data memory map. In program space, the map contains

16M bytes that are accessible at 24-bit addresses. In data space, the map
contains 8M words that are accessible at 23-bit addresses.

- An input/output (I/O) space of 64K words for communication with peripher-

als.

- Software stacks that support 16-bit and 32-bit push and pop operations.
You can use these stack for data storage and retreival. The CPU uses
these stacks for automatic context saving (in response to a call or inter-
rupt) and restoring (when returning to the calling or interrupted code se-
quence).

- A large number of data and address buses, to provide a high level of paral-
lelism. One 32-bit data bus and one 24-bit address bus support instruction
fetching. Three 16-bit data buses and three 24-bit address buses are used
to transport data to the CPU. Two 16-bit data buses and two 24-bit address
buses are used to transport data from the CPU.

- An instruction buffer and a separate fetch mechanism, so that instruction

fetching is decoupled from other CPU activities.

- The following computation blocks: one 40-bit arithmetic logic unit (ALU),
one 16-bit ALU, one 40-bit shifter, and two multiply-and-accumulate units
(MACs). In a single cycle, each MAC can perform a 17-bit by 17-bit multi-
plication (fractional or integer) and a 40-bit addition or subtraction with op-
tional 32-/40-bit saturation.

- An instruction pipeline that is protected. The pipeline protection mecha-

nism inserts delay cycles as necessary to prevent read operations and
write operations from happening out of the intended order.

- Data address generation units that support linear, circular, and bit-reverse
addressing.

- Interrupt-control logic that can block (or mask) certain interrupts known as
the maskable interrupts.

- A TMS320C54x-compatible mode to support code originally written for a

TMS320C54x DSP.

1-2
 Code Development Flow for Best Performance

1.2 Code Development Flow for Best Performance

The following flow chart shows how to achieve the best performance and code-
generation efficiency from your code. After the chart, there is a table that de-
scribes the phases of the flow.

Figure 1−1. Code Development Flow

Step 1: Write C code

Write C Code

Compile

Profile

Efficient Yes
Done
enough?

No
Optimize C code
Step 2:
Optimize
C Code Compile

Profile

Efficient Yes
Done
enough?

No
Yes More C
optimization?

No

To Step 3 (next page)

Introduction 1-3
Code Development Flow for Best Performance

Figure 1−1. Code Development Flow (Continued)

From Step 2 (previous page)

Identify time-critical portions of C code

Step 3:
Write
Assembly Write them in assembly code
Code

Profile

Efficient Yes
Done
enough?

No
Optimize assembly code
Step 4:
Optimize
Assembly Profile
Code

No
Efficient
enough?

Yes
Done

1-4
 Code Development Flow for Best Performance

Step Goal

1 Write C Code: You can develop your code in C using the ANSI-
compliant C55x C compiler without any knowledge of the C55x DSP.
Use Code Composer Studio to identify any inefficient areas that
you might have in your C code. After making your code functional,
you can improve its performance by selecting higher-level optimiza-
tion compiler options. If your code is still not as efficient as you would
like it to be, proceed to step 2.

2 Optimize C Code: Explore potential modifications to your C code

to achieve better performance. Some of the techniques you can ap-
ply include (see Chapter 3):
- Use specific types (register, volatile, const).
- Modify the C code to better suit the C55x architecture.
- Use an ETSI intrinsic when applicable.
- Use C55x compiler intrinsics.
After modifying your code, use the C55x profiling tools again, to
check its performance. If your code is still not as efficient as you
would like it to be, proceed to step 3.

3 Write Assembly Code: Identify the time-critical portions of your C

code and rewrite them as C-callable assembly-language functions.
Again, profile your code, and if it is still not as efficient as you would
like it to be, proceed to step 4.

4 Optimize Assembly Code: After making your assembly code func-

tional, try to optimize the assembly-language functions by using
some of the techniques described in Chapter 4, Optimizing Your As-
sembly Code. The techniques include:
- Place instructions in parallel.
- Rewrite or reorganize code to avoid pipeline protection delays.
- Minimize stalls in instruction fetching.

Introduction 1-5
1-6
 Chapter 2



This tutorial walks you through the code development flow introduced in Chap-
ter 1, and introduces you to basic concepts of TMS320C55x (C55x) DSP pro-
gramming. It uses step-by-step instructions and code examples to show you
how to use the software development tools integrated under Code Composer
Studio (CCS).

Installing CCS before beginning the tutorial allows you to edit, build, and debug
DSP target programs. For more information about CCS features, see the CCS
Tutorial. You can access the CCS Tutorial within CCS by choosing
Help!Tutorial.

The examples in this tutorial use instructions from the mnemonic instruction
set, but the concepts apply equally for the algebraic instruction set.

Topic Page

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2

2.2 Writing Assembly Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3
2.3 Understanding the Linking Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10
2.4 Building Your Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.5 Testing Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-19
2.6 Benchmarking Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21

2-1
Introduction

2.1 Introduction
This tutorial presents a simple assembly code example that adds four num-
bers together (y = x0 + x3 + x1 + x2). This example helps you become familiar
with the basics of C55x programming.

After completing the tutorial, you should know:

- The four common C55x addressing modes and when to use them.

- The basic C55x tools required to develop and test your software.

This tutorial does not replace the information presented in other C55x docu-
mentation and is not intended to cover all the topics required to program the
C55x efficiently.

Refer to the related documentation listed in the preface of this book for more
information about programming the C55x DSP. Much of this information has
been consolidated as part of the C55x Code Composer Studio online help.

For your convenience, all the files required to run this example can be down-
loaded with the TMS320C55x Programmer’s Guide (SPRU376) from
http://www.ti.com/sc/docs/schome.htm. The examples in this chapter can be
found in the 55xprgug_srccode\tutor directory.

2-2
 Writing Assembly Code

2.2 Writing Assembly Code

Writing your assembly code involves the following steps:

- Allocate sections for code, constants, and variables.

- Initialize the processor mode.

- Set up addressing modes and add the following values: x0 + x1 + x2 + x3.

The following rules should be considered when writing C55x assembly code:

- Labels

The first character of a label must be a letter or an underscore ( _ ) fol-

lowed by a letter, and must begin in the first column of the text file. Labels
can contain up to 32 alphanumeric characters.

- Comments

When preceded by a semicolon ( ; ), a comment may begin in any column.

When preceded by an asterisk ( * ), a comment must begin in the first
column.

The final assembly code product of this tutorial is displayed in Example 2−1,
Final Assembly Code of tutor.asm. This code performs the addition of the ele-
ments in vector x. Sections of this code are highlighted in the three steps used
to create this example.

For more information about assembly syntax, see the TMS320C55x Assembly
Language Tools User’s Guide (SPRU280).

Tutorial 2-3
Writing Assembly Code

Example 2−1. Final Assembly Code of tutor.asm

* Step 1: Section allocation

* −−−−−−
.def x,y,init
x .usect ”vars”,4 ; reserve 4 uninitalized 16-bit locations for x
y .usect ”vars”,1 ; reserve 1 uninitialized 16-bit location for y

.sect ”table” ; create initialized section ”table” to

init .int 1,2,3,4 ; contain initialization values for x

.text ; create code section (default is .text)

.def start ; define label to the start of the code
start

* Step 2: Processor mode initialization

* −−−−−−
BCLR C54CM ; set processor to ’55x native mode instead of
; ’54x compatibility mode (reset value)
BCLR AR0LC ; set AR0 register in linear mode
BCLR AR6LC ; set AR6 register in linear mode

* Step 3a: Copy initialization values to vector x using indirect addressing

* −−−−−−−
copy
AMOV #x, XAR0 ; XAR0 pointing to variable x
AMOV #init, XAR6 ; XAR6 pointing to initialization table

MOV *AR6+, *AR0+ ; copy starts from ”init” to ”x”

MOV *AR6+, *AR0+
MOV *AR6+, *AR0+
MOV *AR6, *AR0

* Step 3b: Add values of vector x elements using direct addressing

* −−−−−−−
add
AMOV #x, XDP ; XDP pointing to variable x
.dp x ; and the assembler is notified

MOV @x, AC0

ADD @(x+3), AC0
ADD @(x+1), AC0
ADD @(x+2), AC0

* Step 3c. Write the result to y using absolute addressing

* −−−−−−−
MOV AC0, *(#y)

end
NOP
B end

2-4
 Writing Assembly Code

2.2.1 Allocate Sections for Code, Constants, and Variables

The first step in writing this assembly code is to allocate memory space for the
different sections of your program.

Sections are modules consisting of code, constants, or variables needed to

successfully run your application. These modules are defined in the source file
using assembler directives. The following basic assembler directives are used
to create sections and initialize values in the example code.

- .sect “section_name” creates initialized name section for code/data. Ini-

tialized sections are sections defining their initial values.

- .usect “section_name”, size creates uninitialized named section for data.

Uninitialized sections declare only their size in 16-bit words, but do not de-
fine their initial values.

- .int value reserves a 16-bit word in memory and defines the initialization
value

- .def symbol makes a symbol global, known to external files, and indicates
that the symbol is defined in the current file. External files can access the
symbol by using the .ref directive. A symbol can be a label or a variable.

As shown in Example 2−2 and Figure 2−1, the example file tutor.asm contains
three sections:

- vars, containing five uninitialized memory locations

J The first four are reserved for vector x (the input vector to add).
J The last location, y, will be used to store the result of the addition.

- table, to hold the initialization values for x. The init label points to the begin-
ning of section table.

- .text, which contains the assembly code

Example 2−2 shows the partial assembly code used for allocating sections.

Tutorial 2-5
Writing Assembly Code

Example 2−2. Partial Assembly Code of tutor.asm (Step 1)

* Step 1: Section allocation

* −−−−−−
.def x, y, init
x .usect “vars”, 4 ; reserve 4 uninitialized 16−bit locations for x
y .usect “vars”, 1 ; reserve 1 uninitialized 16−bit location for y

.sect “table” ; create initialized section “table” to

init .int 1, 2, 3, 4 ; contain initialization values for x

.text ; create code section (default is .text)

.def start ; define label to the start of the code
start

Note: The algebraic instructions code example for Partial Assembly Code of tutor.asm (Step 1) is shown in Example B−1 on
page B-2.

Figure 2−1. Section Allocation

Init 1
2
3
4

Start
Code

2-6
 Writing Assembly Code

2.2.2 Processor Mode Initialization

The second step is to make sure the status registers (ST0_55, ST1_55,
ST2_55, and ST3_55) are set to configure your processor. You will either need
to set these values or use the default values. Default values are placed in the
registers after processor reset. You can locate the default register values after
reset in the TMS320C55x DSP CPU Reference Guide (SPRU371).

As shown in Example 2−3:

- The AR0 and AR6 registers are set to linear addressing (instead of circular
addressing) using bit addressing mode to modify the status register bits.

- The processor has been set in C55x native mode instead of C54x-compat-
ible mode.

Example 2−3. Partial Assembly Code of tutor.asm (Step 2)

* Step 2: Processor mode initialization

* −−−−−−
BCLR C54CM ; set processor to ’55x native mode instead of
; ’54x compatibility mode (reset value)
BCLR AR0LC ; set AR0 register in linear mode
BCLR AR6LC ; set AR6 register in linear mode

Note: The algebraic instructions code example for Partial Assembly Code of tutor.asm (Step 2) is shown in Example B−2 on
page B-2.

Tutorial 2-7
Writing Assembly Code

2.2.3 Setting up Addressing Modes

Four of the most common C55x addressing modes are used in this code:

- ARn Indirect addressing (identified by *), in which you use auxiliary regis-
ters (ARx) as pointers.

- DP direct addressing (identified by @), which provides a positive offset ad-

dressing from a base address specified by the DP register. The offset is
calculated by the assembler and defined by a 7-bit value embedded in the
instruction.

- k23 absolute addressing (identified by #), which allows you to specify the
entire 23-bit data address with a label.

- Bit addressing (identified by the bit instruction), which allows you to modify
a single bit of a memory location or MMR register.

For further details on these addressing modes, refer to the TMS320C55x DSP
CPU Reference Guide (SPRU371). Example 2−4 demonstrates the use of the
addressing modes discussed in this section.

In Step 3a, initialization values from the table section are copied to vector x (the
vector to perform the addition) using indirect addressing. Figure 2−2 illustrates
the structure of the extended auxiliar registers (XARn). The XARn register is
used only during register initialization. Subsequent operations use ARn be-
cause only the lower 16 bits are affected (ARn operations are restricted to a
64k main data page). AR6 is used to hold the address of table, and AR0 is used
to hold the address of x.

In Step 3b, direct addressing is used to add the four values. Notice that the
XDP register was initialized to point to variable x. The .dp assembler directive
is used to define the value of XDP, so the correct offset can be computed by
the assembler at compile time.

Finally, in Step 3c, the result was stored in the y vector using absolute address-
ing. Absolute addressing provides an easy way to access a memory location
without having to make XDP changes, but at the expense of an increased code
size.

2-8
 Writing Assembly Code

Example 2−4. Partial Assembly Code of tutor.asm (Part3)

* Step 3a: Copy initialization values to vector x using indirect addressing

* −−−−−−−
copy
AMOV #x, XAR0 ; XAR0 pointing to variable x
AMOV #init, XAR6 ; XAR6 pointing to initialization table

MOV *AR6+, *AR0+ ; copy starts from ”init” to ”x”

MOV *AR6+, *AR0+
MOV *AR6+, *AR0+
MOV *AR6, *AR0

* Step 3b: Add values of vector x elements using direct addressing

* −−−−−−−
add
AMOV #x, XDP ; XDP pointing to variable x
.dp x ; and the assembler is notified

MOV @x, AC0

ADD @(x+3), AC0
ADD @(x+1), AC0
ADD @(x+2), AC0

* Step 3c: Write the result to y using absolute addressing

* −−−−−−−
MOV AC0, *(#y)

end
NOP
B end

Note: The algebraic instructions code example for Partial Assembly Code of tutor.asm (Part3) is shown in Example B−3 on
page B-3.

Figure 2−2. Extended Auxiliary Registers Structure (XARn)

22−16 15−0

XARn ARnH ARn

Note: ARnH (upper 7 bits) specifies the 7-bit main data page. ARn (16-bit register) specifies a
16-bit offset to the 7-bit main data page to form a 23-bit address.

Tutorial 2-9
Understanding the Linking Process

2.3 Understanding the Linking Process

The linker (lnk55.exe) assigns the final addresses to your code and data sec-
tions. This is necessary for your code to execute.

The file that instructs the linker to assign the addresses is called the linker com-
mand file (tutor.cmd) and is shown in Example 2−5. The linker command file
syntax is covered in detail in the TMS320C55x Assembly Language Tools
User’s Guide (SPRU280).

- All addresses and lengths given in the linker command file uses byte ad-
dresses and byte lengths. This is in contrast to a TMS320C54x linker com-
mand file that uses 16-bit word addresses and word lengths.

- The MEMORY linker directive declares all the physical memory available
in your system (For example, a DARAM memory block at location 0x100
of length 0x8000 bytes). Memory blocks cannot overlap.

- The SECTIONS linker directive lists all the sections contained in your input
files and where you want the linker to allocate them.

When you build your project in Section 2.4, this code produces two files, tu-
tor.out and a tutor.map. Review the test.map file, Example 2−6, to verify the
addresses for x, y, and table. Notice that the linker reports byte addresses for
program labels such as start and .text, and 16-bit word addresses for data la-
bels like x, y, and table. The C55x DSP uses byte addressing to acces variable
length instructions. Instructions can be 1-6 bytes long.

Example 2−5. Linker command file (tutor.cmd)

MEMORY /* byte address, byte len */

{
DARAM: org= 000100h, len = 8000h
SARAM: org= 010000h, len = 8000h
}

SECTIONS /* byte address, byte len */

{
vars :> DARAM
table: > SARAM
.text:> SARAM
}

2-10
 Understanding the Linking Process

Example 2−6. Linker map file (test.map)

******************************************************************************
TMS320C55xx COFF Linker
******************************************************************************
>> Linked Mon Feb 14 14:52:21 2000

OUTPUT FILE NAME: <tutor.out>

ENTRY POINT SYMBOL: ”start” address: 00010008

MEMORY CONFIGURATION

name org (bytes) len (bytes) used (bytes) attributes fill

−−−− −−−−−−−−−−− −−−−−−−−−−− −−−−−−−−−−−− −−−−−−−−−− −−−−
DARAM 00000100 000008000 0000000a RWIX
SARAM 00010000 000008000 00000040 RWIX

SECTION ALLOCATION MAP

output attributes/
section page orgn(bytes) orgn(words) len(bytes) len(words) input sections
−−−−−−−− −−−− −−−−−−−−−−− −−−−−−−−−−− −−−−−−−−−− −−−−−−−−−− −−−−−−−−−−−−−−
vars 0 00000080 00000005 UNINITIALIZED
00000080 00000005 test.obj (vars)

table 0 00008000 00000004

00008000 00000004 test.obj
(table)

.text 0 00010008 00000038

00010008 00000037 test.obj
(.text)
0001003f 00000001 −−HOLE−− [fill
= 2020]

.data 0 00000000 00000000 UNINITIALIZED

00000000 00000000 test.obj
(.data)

.bss 0 00000000 00000000 UNINITIALIZED

00000000 00000000 test.obj (.bss)

Tutorial 2-11
Understanding the Linking Process

Example 2−6. Linker map file (test.map), (Continued)

GLOBAL SYMBOLS: SORTED ALPHABETICALLY BY Name

abs. value/
byte addr word addr name
−−−−−−−−− −−−−−−−−− −−−−
00000000 .bss
00000000 .data
00010008 .text
00000000 ___bss__
00000000 ___data__
00000000 ___edata__
00000000 ___end__
00010040 ___etext__
00010008 ___text__
00000000 edata
00000000 end
00010040 etext
00008000 init
00010008 start
00000080 x
00000084 y

GLOBAL SYMBOLS: SORTED BY Symbol Address

abs. value/
byte addr word addr name
−−−−−−−−− −−−−−−−−− −−−−
00000000 ___end__
00000000 ___edata__
00000000 end
00000000 edata
00000000 ___data__
00000000 .data
00000000 .bss
00000000 ___bss__
00000080 x
00000084 y
00008000 init
00010008 start
00010008 .text
00010008 ___text__
00010040 ___etext__
00010040 etext

[16 symbols]

2-12
 Building Your Program

2.4 Building Your Program

At this point, you should have already successfully installed CCS and selected
the C55x Simulator as the CCS configuration driver to use. You can select the
configuration driver to be used in the CCS setup.

Before building your program, you must set up your work environment and
create a .pjt file. Setting up your work environment involves the following tasks:

- Creating a project

- Adding files to the work space

- Modifying the build options

- Building your program

2.4.1 Creating a Project

Create a new project called tutor.pjt.

1) From the Project menu, choose New and enter the values shown in
Figure 2−3.

2) Select Finish.

You have now created a project named tutor.pjt and saved it in the new
c:\ti\myprojects\tutor folder.

Tutorial 2-13
Building Your Program

Figure 2−3. Project Creation Dialog Box

2.4.2 Adding Files to the Workspace

Copy the tutorial files (tutor.asm and tutor.cmd) to the tutor project direc-
tory.

1) Navigate to the directory where the tutorial files are located (the
55xprgug_srccode\tutor directory) and copy them into the c:\ti\mypro-
jects\tutor directory. As an alternative, you can create your own source
files by choosing File!New!Source File and typing the source code from
the examples in this book.

2) Add the two files to the tutor.pjt project. Highlight tutor.pjt, right-click the
mouse, select Add Files, browse for the tutor.asm file, select it, and click
Open, as shown in Figure 2−4. Do the same for tutor.cmd, as shown in
Figure 2−5.

2-14
 Building Your Program

Figure 2−4. Add tutor.asm to Project

Tutorial 2-15
Building Your Program

Figure 2−5. Add tutor.cmd to Project

2-16
 Building Your Program

2.4.3 Modifying Build Options

Modify the Linker options.

1) From the Project menu, choose Build Options.

2) Select the Linker tab and enter fields as shown in Figure 2−6.

3) Click OK when finished.

Figure 2−6. Build Options Dialog Box

Tutorial 2-17
Building Your Program

2.4.4 Building the Program

From the Project menu, choose Rebuild All. After the Rebuild process
completes, the screen shown in Figure 2−7 should display.

When you build your project, CCS compiles, assembles, and links your code
in one step. The assembler reads the assembly source file and converts C55x
instructions to their corresponding binary encoding. The result of the assembly
processes is an object file, tutor.obj, in industry standard COFF binary format.
The object file contains all of your code and variables, but the addresses for
the different sections of code are not assigned. This assignment takes place
during the linking process.

Because there is no C code in your project, no compiler options were used.

Figure 2−7. Rebuild Complete Screen

2-18
 Testing Your Code

2.5 Testing Your Code

To test your code, inspect its execution using the C55x Simulator.

Load tutor.out

1) From the File menu, choose Load program.

2) Navigate to and select tutor.out (in the \debug directory), then choose
Open.

CCS now displays the tutor.asm source code at the beginning of the start label
because of the entry symbol defined in the linker command file (-e start).
Otherwise, it would have shown the location pointed to by the reset vector.

Display arrays x, y, and init by setting Memory Window options

1) From the View menu, choose Memory.

2) In the Title field, type x.

3) In the Address field, type x.

4) Repeat 1−3 for y.

5) Display the init array by selecting View→ Memory.

6) In the Title field, type Table.

7) In the Address field, type init.

8) Display AC0 by selecting View→CPU Registers→CPU Registers.

The labels x, y, and init are visible to the simulator (using View→ Memory) be-
cause they were exported as symbols (using the .def directive in tutor.asm).
The -g option was used to enable assembly source debugging.

Now, single-step through the code to the end label by selecting Debug→Step
Into. Examine the X Memory window to verify that the table values populate
x and that y gets the value 0xa (1 + 2 + 3 + 4 = 10 = 0xa), as shown in
Example 2−7.

Tutorial 2-19
Testing Your Code

Example 2−7. x Memory Window

2-20
 Benchmarking Your Code

2.6 Benchmarking Your Code

After verifying the correct functional operation of your code, you can use CCS
to calculate the number of cycles your code takes to execute.

Reload your code

From the File menu, choose Reload Program.

Enable clock for profiling

1) From the Profiler menu, choose Enable Clock.

2) From the Profiler menu, choose View Clock.

Set breakpoints

1) Select the tutor.asm window.

2) Set one breakpoint at the beginning of the code you want to benchmark
(first instruction after start): Right-click on the instruction next to the copy
label and choose Toggle Breakpoint.

3) Set one breakpoint marking the end: Right-click on the instruction next to
the end label and choose Toggle Breakpoint.

Benchmark your code

1) Run to the first breakpoint by selecting Debug→ Run.

2) Double-click in the Clock Window to clear the cycle count.

3) Run to the second breakpoint by selecting Debug→ Run.

4) The Clock Window displays the number of cycles the code took to execute
between the breakpoints, which was approximately 17.

Tutorial 2-21
2-22
 Chapter 3

 

You can maximize the performance of your C code by using certain compiler
options, C code transformations, and compiler intrinsics. This chapter dis-
cusses features of the C language relevant to compilation on the
TMS320C55x (C55x) DSP, performance-enhancing options for the compiler,
and C55x-specific code transformations that improve C code performance. All
assembly language examples were generated for the large memory model via
the −ml compiler option.

Topic Page

3.1 Introduction to Writing C/C++ Code for a C55x DSP . . . . . . . . . . . . . . 3-2

3.2 Compiling the C/C++ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
3.3 Profiling Your Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.4 Refining the C/C++ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15
3.5 Memory Management Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-42

3-1
Introduction to Writing C/C++ Code for a C55x DSP

3.1 Introduction to Writing C/C++ Code for a C55x DSP

This section describes some general issues to keep in mind when writing
C/C++ code for the TMS320C55x (C55x) architecture (or any DSP architec-
ture). Keep this information in mind when working in Step 1 of code develop-
ment as described in Chapter 1. Refer to TMS320C55x Optimizing C/C++
Compiler User’s Guide (SPRU281) for additional language issues.

3.1.1 Tips on Data Types

Give careful consideration to the data type size when writing your code. The
C55x compiler defines a size for each C data type (signed and unsigned):
char 16 bits
short 16 bits
int 16 bits
long 32 bits
long long 40 bits
float 32 bits
double 64 bits

Floating point values are in the IEEE format. Based on the size of each data
type, follow these guidelines when writing your code:

- Avoid code that assumes that int and long types are the same size.

- Use the int data type for fixed-point arithmetic (especially multiplication)
whenever possible. Using type long for multiplication operands will result
in calls to a run-time library routine.

- Use int or unsigned int types rather than long for loop counters.
The C55x has mechanisms for efficient hardware loops, but hardware
loop counters are only 16 bits wide.

- Avoid code that assumes char is 8 bits or long long is 64 bits.

When writing code to be used on multiple DSP targets, it may be wise to define
“generic” types for the standard C types. For example, one could use the types
Int16 and Int32 for a 16 bit integer type and 32 bit integer type respectively.
When compiling for the C55x DSP, these types would be type defined to int
and long, respectively.

In general it is best to use the type int for loop index variables and other inte-
ger variables where the number of bits is unimportant as int typically repre-
sents the most efficient integer type for the target to manipulate, regardless of
architecture.

3-2
 Introduction to Writing C/C++ Code for a C55x DSP

3.1.2 How to Write Multiplication Expressions Correctly in C Code

Writing multiplication expressions in C code so that they are both correct and
efficient can be confusing, especially when technically illegal expressions can,
in some circumstances, generate the code you wanted in the first place. This
section will help you choose the correct expression for your algorithm.

The correct expression for a 16x16−>32 multiplication on a C55x DSP is:

long res = (long)(int)src1 * (long)(int)src2;

According to the C arithmetic rules,this is actually a 32x32−>32 multiplication,

but the compiler will notice that each operand fits in 16 bits, so it will issue an
efficient single-instruction multiplication.

A 16-bit multiplication with a 32-bit result is an operation which does not direct-
ly exist in the C language, but does exist on C55x hardware, and is vital for mul-
tiply-and-accumulate (MAC)-like algorithm performance.

Example 3−1 shows two incorrect ways and a correct way to write such a multi-
plication in C code.

Example 3−1. Generating a 16x16−>32 Multiply

long mult(int a, int b)

{
long result;
/* incorrect */
result = a * b;
/* incorrect */
result = (long)(a * b);
/* correct */
result = (long)a * b;
return result;
}

Note that the same rules also apply for other C arithmetic operators. For exam-
ple, if you want to add two 16-bit numbers and get a full 32 bit result, the correct
syntax is:
(long) res = (long)(int)src1 + (long)(int)src;

Optimizing C Code 3-3

Introduction to Writing C/C++ Code for a C55x DSP

3.1.3 Memory Dependences

To maximize the efficiency or your code, the C55x compiler reorders instruc-
tions to minimize pipeline stalls, puts certain assembly instructions in parallel,
and generates dual multiply-and-accumulate (dual-MAC) instructions. These
transformations require the compiler to determine the relationships, or depen-
dences, between instructions. Dependence means that one instruction must
occur before another. For example, a variable may need to be loaded from
memory before it can be used. Because only independent instructions can be
scheduled in parallel or reordered, dependences inhibit parallelism and code
movement. If the compiler cannot prove that two instructions are independent,
it must assume that instructions must remain in the order they originally ap-
peared and must not be scheduled in parallel.

Often it is difficult for the compiler to determine whether instructions that ac-
cess memory are independent. The following techniques help the compiler de-
termine which instructions are independent:

- Use the restrict keyword to indicate that a pointer is the only pointer
than can point to a particular object in the scope in which the pointer is de-
clared.

- Use the –pm option which gives the compiler global access to the whole
program and allows it to be more aggressive in ruling out dependences.

To illustrate the concept of memory dependences, it is helpful to look at the al-

gorithm code in a dependence graph. Example 3−2 shows code for a simple
vector sum. Figure 3−1 shows a simplified dependence graph for that piece of
code.

Example 3−2. C Code for Vector Sum

void vecsum(int *sum, short *in1, short *in2, int N)

{
int i;
for (i = 0; i < N; i++)
sum[i] = in1[i] + in2[i];
}

3-4
 Introduction to Writing C/C++ Code for a C55x DSP

Figure 3−1. Dependence Graph for Vector Sum

Load Load
in1[i] in2[i]

Add

Store
sum[i]

The dependence graph in Figure 3−1 shows that:

- The paths from the store of sum[i] back to the loads of in1[i] and
in2[i] indicate that writing to sum may have an effect on the memory
pointed to by either in1 or in2.

- A read from in1 or in2 cannot begin until the write to sum finishes, which
creates an aliasing problem. Aliasing occurs when two pointers can point
to the same memory location. For example, if vecsum() is called in a pro-
gram with the following statements, in1 and sum alias each other be-
cause they both point to the same memory location:
short a[10], b[10];
vecsum(a, a, b, 10);

To help the compiler resolve memory dependences, you can qualify a pointer
or array with the restrict keyword. Its use represents a guarantee by the
programmer that within the scope of the pointer declaration, the object pointed
to can be accessed only by that pointer. Any violation of this guarantee renders
the behavior of the program undefined. This practice helps the compiler opti-
mize certain sections of code because aliasing information can be more easily
determined.

In the declaration of the vector sum function you can use the restrict key-
word to tell the compiler that sum is the only pointer that points to that object:
void vecsum(int * restrict sum, int *in1, int *in2, int N)

Optimizing C Code 3-5

Introduction to Writing C/C++ Code for a C55x DSP

(Likewise, you could add restrict to in1 and in2 as well.) The next piece of
code shows how to use restrict with an array function parameter instead of
a pointer:
void vecsum(int sum[restrict], int *in1, int *in2, int N)

Caution must be exercised when using restrict. Consider this call of vecsum()
(with the sum parameter qualified by restrict):
vecsum(a, a, b, 10);

Undefined behavior would result because sum and in1 would point to the
same object, which violates sum’s declaration as restrict.

3.1.4 Analyzing C Code Performance

Use the following techniques to analyze the performance of specific code re-
gions:

- Use the clock() and printf() functions in C/C++ code to time and dis-
play the performance of specific code regions. You can use the stand-
alone simulator (load55) for this purpose. Remember to subtract out the
overhead time of calling the clock() function.

- Enable the clock and use profile points and the RUN command in the Code
Composer Studio debugger to track the number of CPU clock cycles con-
sumed by a particular section of code.

- Put each loop into a separate file that can be rewritten, recompiled, and
run with the stand-alone simulator (load55).The critical performance
areas in your code are most often loops.

As you use the techniques described in this chapter to optimize your C/C++
code, you can then evaluate the performance results by running the code and
looking at the instructions generated by the compiler. More detail on perfor-
mance analysis can be found in section 3.3.

3-6
 Compiling the C/C++ Code

3.2 Compiling the C/C++ Code

The C55x compiler offers high-level language support by transforming your
C/C++ code into assembly language source code. The compiler tools include
a shell program (cl55), which you use to compile, optimize, assemble, and link
programs in a single step. To invoke the compiler shell, enter:

cl55 [options] [filenames] [−z [linker options] [object files]]

For a complete description of the C/C++ compiler and the options discussed in
this section, see the TMS320C55x Optimizing C Compiler User’s Guide
(SPRU281).

3.2.1 Compiler Options

Options control the operation of the compiler. This section introduces you to
the recommended options for performance, information gathering, and code
size.

First make note of the options to avoid using on performance critical code.The
options described in Table 3−1 are intended for debugging, and could poten-
tially decrease performance and increase code size.

Table 3−1. Compiler Options to Avoid on Performance-Critical Code

Option Description
−g, −s, −ss, −gp These options are intended for debugging and can limit the
amount of optimization across C statements leading to larger
code size and slower execution.

−o1, −o0 Always use −o2/−o3 to maximize compiler analysis and opti-
mization

−mr Prevents generation of hardware loops to reduce context

save/restore for interrupts. As hardware loops greatly improve
performance of loop code, avoid this option on performance
critical code.

The options in Table 3−2 can be used to improve performance. The options
−o3, −pm, −mb, −oi50, and −op2 are recommended for maximum perfor-
mance.

Optimizing C Code 3-7

Compiling the C/C++ Code

Table 3−2. Compiler Options for Performance

Option Description
−o3 Represents the highest level of optimization available. Various loop
optimizations are performed, and various file-level characteristics
are also used to improve performance.

−pm Combines source files to perform program-level optimization by

allowing the compiler visibility to the entire application source.

−oi<size> Enables inlining of functions based on a maximum size. (Enabled

with −o3.) Size here is determined internally by the optimizer and
does not correspond to bytes or any other known standard unit.
Use a −onx option to check sizes of individual functions.

−mb Asserts to the compiler that all data is on-chip. This option is used
to enable the compiler to generate dual-MAC. See section 3.4.2.2
for more details.

−op2 When used with −pm, this option allows the compiler to assume
that the program being compiled does not contain any functions or
variables called or modified from outside the current file. The com-
piler is free to remove any functions or variables that are unused in
the current file.

−mn Re-enables optimizations disabled when using −g option (symbolic

debugging). Use this option when it is necessary to debug opti-
mized code.

The options described in Table 3−3, can be used to improve code size with a
possible degradation in performance.

Table 3−3. Compiler Options That May Degrade Performance and Improve Code Size

Option Description
−ms Encourages the compiler to optimize for code space. (Default is to
optimize for performance.)

−oi0 Disables all automatic size-controlled inlining enabled by −o3.

User specified inlining of functions is still allowed.

The options described in Table 3−4 provide information to the programmer.

Some of them may negatively affect performance and/or code size.

3-8
 Compiling the C/C++ Code

Table 3−4. Compiler Options for Information

Option Description
−k The assembly file is not deleted. This allows you to inspect the
generated code. This option has no impact on performance or
code size.

−s/−ss Interlists optimizer comments/C source in the assembly. The −s

option may show minor performance degradation. The –ss option
may show more severe performance degradation.

−mg Generate algebraic assembly code. (Default is mnemonic.) There

is no performance or code size impact.

−onx When x is 1, the optimizer generates an information file (.nfo file-

name extension). When x is 2, a more verbose information file is
generated. There is no performance or code size impact.

3.2.2 Performing Program-Level Optimization (−pm Option)

You can specify program-level optimization by using the −pm option with the
−o3 option. With program-level optimization, all your source files are compiled
into one intermediate file giving the compiler complete program view during
compilation. Because the compiler has access to the entire program, it per-
forms several optimizations that are rarely applied during file-level optimiza-
tion:
- If the number of iterations of a loop is determined by a value passed into
the function and the compiler can determine what the value is from the call-
ing function, the compiler will have more information about the number of
iterations of the loop, resulting in more efficient loop code.
- If a particular argument to a function always has the same value, the com-
piler replaces the occurrences of the argument in the function with the
value.
- If a return value of a function is never used, the compiler deletes the return
code in the function.
- If a function is not called, directly or indirectly, the compiler removes the
code in the function.

Program-level optimization increases compilation time because the compiler

performs more complex optimizations on a larger amount of code. For this rea-
son you may not want to use this option for every build of large programs.

Example 3−3 and Example 3−4 show the content of two files. One file contains
the source for the main function and the second file contains source for a small
function called sum.

Optimizing C Code 3-9

Compiling the C/C++ Code

Example 3−3. Main Function File

extern int sum(const int *a, int n);

const int a[10] = {1,2,3,4,5,6,7,8,9,10};
const int b[10] = {11,12,13,14,15,16,17,18,19,20};
int sum1, sum2;
int main(void)
{
sum1 = sum(a,10);
sum2 = sum(b,10);
return(0);
}

Example 3−4. Sum Function File

int sum(const int *a, int n)

{
int total = 0;
int i;
for(i=0; i<n; i++)
{
total += a[i];
}
return total;
}

When this code is compiled with −o3 and −pm options, the optimizer has
enough information about the calls to sum to determine that the same loop
count is used for both calls. It therefore eliminates the argument n from the call
to the function and explicitly uses the count in the repeat single instruction as
shown in Example 3−5.

3-10
 Compiling the C/C++ Code

Example 3−5. Assembly Code Generated With −o3 and −pm Options

_sum:
;** Parameter deleted n == 9u
MOV #0, T0 ; |3|
RPT #9
ADD *AR0+, T0, T0

return ; |11|
_main:
AADD #−1, SP
AMOV #_a, XAR0 ; |9|
call #_sum ; |9|
; call occurs [#_sum] ; |9|
MOV T0, *(#_sum1) ; |9|
AMOV #_b, XAR0 ; |10|
call #_sum ; |10|
; call occurs [#_sum] ; |10|
MOV T0, *(#_sum2) ; |10|
AADD #1, SP
return
; return occurs

Note: The algebraic instructions code example for Assembly Code Generated With −o3 and −pm Options is shown in
Example B−4 on page B-4.

Caution must be exercised when using program mode (−pm) compilation on

code that consists of a mixture of C/C++ and assembly language functions.
These issues are described in detail in the TMS320C55x Optimizing C Compil-
er User’s Guide (SPRU281).

3.2.3 Using Function Inlining

There are two ways to enable the compiler to inline functions:
- Inlining controlled by the inline keyword. To enable this mode you must
run the optimizer (that is, you must choose at least −o0.)
- Automatic inlining of small functions that are not declared as inline in
your C/C++ code. To enable this mode use the −o3 and −oi<size> com-
piler options.

Optimizing C Code 3-11

Compiling the C/C++ Code

The −oi<size> option may be used to specify automatic inlining of small

functions even if they have not been declared with the inline keyword. The
size of a function is an internal compiler notion of size. To see the size of a par-
ticular function use the −onx options described in Table 3−4 on page 3-9.

Example 3−6 shows the resulting assembly instructions when the code in
Example 3−3 and Example 3−4 is compiled with −o3, −pm, and −oi50 op-
tions.

In main, the function calls to sum have been inlined. However, code for the
body of function sum has still been generated. The compiler must generate this
code because it does not have enough information to eliminate the possibility
that the function sum may be called by some other externally defined function.
If no external function calls sum, it can be declared as static inline. The
compiler will then be able to eliminate the code for sum after inlining.

Example 3−6. Assembly Generated Using −o3, −pm, and −oi50

_sum:
MOV #0, T0 ; |3|
RPT #9
ADD *AR0+, T0, T0
return ; |11|
_main:
AMOV #_a, XAR3 ; |9|
RPT #9
|| MOV #0, AR1 ; |3|
ADD *AR3+, AR1, AR1

MOV AR1, *(#_sum1) ; |11|

MOV #0, AR1 ; |3|
AMOV #_b, XAR3 ; |10|
RPT #9
ADD *AR3+, AR1, AR1

MOV AR1, *(#_sum2) ; |11|

return

Note: The algebraic instructions code example for Assembly Generated Using −o3, −pm, and −oi50 is shown in Example B−5
on page B-5.

3-12
 Profiling Your Code

3.3 Profiling Your Code

In large applications, it makes sense to optimize the most important sections of

code first. You can use the information generated by profiling options to get
started. This section describes profiling methods to determine whether to
move from Step 1 to Step 2 of the code development flow as described in
Chapter 1 (or from Step 2 to Step 3). You can use several different methods to
profile your code.

3.3.1 Using the clock() Function to Profile

To get cycle count information for a function or region of code with the stand-
alone simulator, embed the clock() function in your C code. Example 3−7
demonstrates this technique.

Example 3−7. Using the clock() Function

#include <stdio.h>
#include <time.h> /* Need time.h in order to call clock() */
int main()
{
clock_t start, stop, overhead;
start = clock(); /* Calculate the overhead of calling clock */
stop = clock(); /* and subtract this amount from the results. */
overhead = stop − start;
start = clock();
/* Function or Code Region to time goes here */
stop = clock();
printf(”cycles: %ld\n”,(long)(stop − start – overhead));
return(0);
}

Caution: Using clock() to time a region of code could increase the cycle
count of that region due to the extra variables needed to hold the timing infor-
mation (the stop, start, and overhead variables above). Wrapping
clock() around a function call should not affect the cycle count of that func-
tion.

Optimizing C Code 3-13

Profiling Your Code

3.3.2 Using CCS 2.0 to Profile

Code Composer Studio (CCS) 2.0 has extensive profiling options that can be
used to profile your C code. First you must enable the clock by selecting En-
able Clock from the Profiler menu. Selecting Start New Session from the Profil-
er menu starts a new profiling session. To profile all functions, click on the Pro-
file All Functions button in the profiler session window. To profile certain func-
tions or regions of code, click the Create Profile Area and enter the starting and
ending line numbers of the code you wish to profile. (Note that you must build
your code for debugging (−g option) to enable this feature.) Then, run your pro-
gram and the profile information will be updated in the profiler session window.

More information on profiling with CCS 2.0 can be found in the online docu-
mentation.

3-14
 Refining the C/C++ Code

3.4 Refining the C/C++ Code

This section describes C55x-specific optimization techniques that you can use
to improve your C/C++ code performance. These techniques should be used
in Step 2 of the code development flow as described in Chapter 1. Consider
these tips when refining your code:

- Create loops that efficiently use C55x hardware loops, MAC hardware,
and dual-MAC hardware.

- Use intrinsics to replace complicated C/C++ code

- Avoid the modulus operator when simulating circular addressing

- Use long accesses to reference 16-bit data in memory

- Write efficient control code

It is recommended that the following optimization techniques be applied in the

order presented here. The code can be profiled after implementing the opti-
mization described in each section to determine if further optimization is need-
ed. If so, proceed to the next optimization. The six techniques presented in this
section are summarized in Table 3−5. The indications (high, moderate, low,
easy, many, some, few) in the table apply to typical DSP code. Potential per-
formance gain estimates the performance improvement over no modifica-
tions to the code. Ease of implementation reflects both the required amount
of change to the code and the complexity of the optimization. Opportunities
are the number of places the optimization can be applied in typical DSP code.

Table 3−5. Summary of C/C++ Code Optimization Techniques

Potential Ease of
Optimization Technique Performance Gain Implementation Opportunities Issues
Generate efficient loop High Easy Many
code

Use MAC hardware High Moderate Many

efficiently

Use Intrinsics High Moderate Many Reduces portability

Avoid modulus in circular Moderate Easy Some

addressing

Use long accesses for Low Moderate Few

16-bit data

Generate efficient control Low Easy Few

code

Optimizing C Code 3-15

Refining the C/C++ Code

3.4.1 Generating Efficient Loop Code

You can realize substantial gains from the performance of your C/C++ loop
code by refining your code in the following areas:
- Avoid function calls within the body of repeated loops. This enables the
compiler to use very efficient hardware looping constructs (repeat,
localrepeat, and blockrepeat, or RPT, RPTBLOCAL, and RPTB in
mnemonic syntax).
- Keep loop code small to enable the compiler to use localrepeat.

- Analyze trip count issues.

- Use the MUST_ITERATE pragma.

- Use the −o3 and −pm compiler options.

3.4.1.1 Avoid Function Calls within Loops

Whenever possible avoid using function calls within loops. Because repeat la-
bels and counts would have to be preserved across calls, the compiler decides
never to generate hardware loops that contain function calls. This leads to inef-
ficient loop code.

3.4.1.2 Keep Loops Small to Enable localrepeat

Keeping loop code small enables the compiler to make use of the native
localrepeat instruction. The compiler will generate localrepeat for
small loops that do not contain any control flow structures other than forward
conditionals. Localrepeat loops consume less power than other looping
constructs. An example of a small loop that can use localrepeat is shown in
Example 3−8 and Example 3−9. Example 3−8 shows C code and
Example 3−9 shows the assembly code generated by the compiler.

Example 3−8. Simple Loop That Allows Use of localrepeat

void vecsum(const short *a, const short *b, short *c, unsigned int n)
{
unsigned int i;

for (i=0; i<=n−1; i++)

{
*c++ = *a++ + *b++;
}
}

3-16
 Refining the C/C++ Code

Example 3−9. Assembly Code for localrepeat Generated by the Compiler

_vecsum:
SUB #1, T0, AR3
MOV AR3, BRC0
RPTBLOCAL L2−1
ADD *AR0+, *AR1+, AC0 ; |7|
MOV HI(AC0), *AR2+ ; |7|
L2:
return

Note: The algebraic instructions code example for Assembly Code for localrepeat Generated by the Compiler is shown in
Example B−6 on page B-5.

3.4.1.3 Trip Count Issues

A trip count is the number of times that a loop executes; the trip counter is the
variable used to count each iteration. When the trip counter reaches the limit
equal to the trip count, the loop terminates. Maximum performance for loop
code is gained when the compiler can determine the exact minimum and maxi-
mum for the trip count. To this end, use the following techniques to convey trip
count information to the compiler:

- Use int (or unsigned int) type for trip counter variable, whenever
possible.

- Use the MUST_ITERATE pragma to eliminate code to skip around loops

and help the compiler generate efficient hardware loops. This pragma can
also be used to aid in loop unrolling.

- Be sure to use the −o3 and −pm compiler options to allow the optimizer
access to the whole program or large parts of it and to characterize the be-
havior of loop trip counts.

Using int Type. Using the type int for the trip counter is important to allow
the compiler to generate hardware looping constructs.

In Example 3−10, consider this simple for loop:

for(i = 0; i<n; i++)

If, for example, i and n were declared to be of type long, no hardware loop
could be generated. This is because the C55x internal loop iteration count reg-
ister is only 16 bits wide. If i and n are declared as type int, then the compiler
will generate a hardware loop.

Optimizing C Code 3-17

Refining the C/C++ Code

3.4.1.4 Using the MUST_ITERATE Pragma

The MUST_ITERATE pragma is used to convey programmer knowledge about

loops to the compiler. It should be used as much as possible to aid the compiler
in the optimization of loops.

Example 3−10 shows code to compute the sum of a vector. The corresponding
assembly code is shown in Example 3−11. Notice the conditional branch that
jumps around the loop body in the generated assembly code. The compiler
must insert this additional code if there is any possibility that the loop could ex-
ecute zero times. In this particular case the loop upper bound n is an integer.
Thus, n could be zero or negative in which case C semantics would dictate that
the for loop body would not execute. A hardware loop must execute at least
once, so the jump around code ensures correct execution in cases where
n <= 0.

Example 3−10. Inefficient Loop Code for Loop Variable and Constraints (C)

int sum(const short *a, int n)

{
int sum = 0;
int i;
for(i=0; i<n; i++)
{
sum += a[i];
}
return sum;
}

3-18
 Refining the C/C++ Code

Example 3−11. Inefficient Loop Code for Variable and Constraints (Assembly)

_sum:
MOV #0, AR1 ; |3|
BCC L2,T0 <= #0 ; |6|
; branch occurs ; |6|
SUB #1, T0, AR2
MOV AR2, CSR
RPT CSR
ADD *AR0+, AR1, AR1

MOV AR1, T0 ; |11|

return ; |11|

Note: The algebraic instructions code example for Inefficient Loop Code for Variable and Constraints (Assembly) is shown in
Example B−7 on page B-6.

If it is known that the loop always executes at least once, this fact can be com-
municated to the compiler via the MUST_ITERATE pragma. Example 3−12
shows how to use the pragma for this piece of code. Example 3−13 shows the
more efficient assembly code that can now be generated because of the prag-
ma.

Example 3−12. Using the MUST_ITERATE Pragma

int sum(const short *a, int n)

{
int sum = 0;
int i;

#pragma MUST_ITERATE(1)
for(i=0; i<n; i++)
{
sum += a[i];
}
return sum;
}

(Note that the same effect could be generated by using an _nassert, to as-
sert to the compiler that n is greater than zero: _nassert(n>0)).

Optimizing C Code 3-19

Refining the C/C++ Code

Example 3−13. Assembly Code Generated With the MUST_ITERATE Pragma

_sum:
SUB #1, T0, AR2
MOV AR2, CSR
MOV #0, AR1 ; |3|
RPT CSR
ADD *AR0+, AR1, AR1

MOV AR1, T0 ; |12|

return ; |12|

Note: The algebraic instructions code example for Assembly Code Generated With the MUST_ITERATE Pragma is shown in
Example B−8 on page B-6.

MUST_ITERATE can be used to communicate several different pieces of infor-

mation to the compiler. The format of the pragma is:
#pragma MUST_ITERATE(min, max, mult)

All fields are optional. min is the minimum number of iterations of the loop, max
is the maximum number of iterations of the loop, and mult tells the compiler
that the loop always executes a multiple of mult times. If some of these values
are not known until run time, do not include them in the pragma. Incorrect infor-
mation communicated via the pragma could result in undefined program beha-
vior. The MUST_ITERATE pragma must appear immediately before the loop
that it is meant to describe in the C code. MUST_ITERATE can be used in the
following ways:
- It can convey that the trip count will be greater than some minimum value.
/* This loop will always execute at least 30 times */
#pragma MUST_ITERATE(30)
for(j=0; j<x; j++)
- It can convey the maximum trip count.
/* The loop will execute no more than 100 times */
#pragma MUST_ITERATE(,100)
for (j=0; j<x; j++)
- It can convey that the trip count is always divisible by a value.
/* The loop will execute some multiple of 4 times */
#pragma MUST_ITERATE(,,4)
for (j=0; j<x; j++)

3-20
 Refining the C/C++ Code

Consider the following loop header (from the ETSI gsmefr benchmark):
for(i=a[0]; i < 40; i +=5)

To generate a hardware loop, the compiler would need to emit code that would
determine the number of loop iterations at run time. This code would require an
integer division. Since this is computationally expensive, the compiler will not
generate such code and will not generate a hardware loop. However, if the pro-
grammer knows that, for example, a[0] is always less than or equal to 4, then
the loop always executes exactly eight times. This can be communicated via a
MUST_ITERATE pragma enabling the compiler to generate an efficient hard-
ware loop:
#pragma MUST_ITERATE(8,8)

3.4.2 Efficient Use of MAC hardware

Multiply-and-accumulate (MAC) is a very common paradigm in DSP algo-

rithms, and a C55x DSP has hardware to perform MAC operations efficiently. It
can perform a single MAC (or multiply, multiply and subtract) operation or two
MAC operations in a single cycle (a dual-MAC operation). The next section de-
scribes how to write efficient, small loops that use a single MAC operation.
Section 3.4.2.2 describes how to enable the compiler to generate dual-MAC
operations from your C/C++ code.

3.4.2.1 Special Considerations When Using MAC Constructs

The compiler can generate a very efficient single repeat MAC construct (that
is, a repeat (RPT) loop with a MAC as its only instruction.) To facilitate the
generation of single repeat MAC constructs, use local rather than global vari-
ables for the summation, as shown in Example 3−14. If a global variable is
used, the compiler is obligated to perform an intervening storage to the global
object. This prevents it from generating a single repeat.

In the case where Q15 arithmetic is being simulated, the result of the MAC op-
eration may be accumulated into a long object. The result may then be shifted
and truncated before the return, as shown in Example 3−15.

Optimizing C Code 3-21

Refining the C/C++ Code

Example 3−14. Use Local Rather Than Global Summation Variables

/* Not recommended */
int gsum=0;
void dotp1(const int *x, const int *y, unsigned int n)
{
unsigned int i;
for(i=0; i<=n−1; i++)
gsum += x[i] * y[i];
}
/* Recommended */
int dotp2(const int *x, const int *y, unsigned int n)
{
unsigned int i;
int lsum=0;
for(i=0; i<=n−1; i++)
lsum += x[i] * y[i];
return lsum;
}

Example 3−15. Returning Q15 Result for Multiply Accumulate

int dotp(const int *x, const int *y, unsigned int n)

{
unsigned int i;
long sum=0;
for(i=0; i<=n−1; i++)
sum += (long)x[i] * y[i];
return (int)((sum>>15) & 0x0000FFFFL);
}

3.4.2.2 Generating Dual-MAC Operations

A dual-MAC operation (2 multiply-and-accumulate/subtract instructions in a
single cycle) is one of the most important hardware features of a C55x DSP.
(Note, the term dual-MAC will be used to refer to dual multiplies, dual multiply-
and-accumulates and dual multiply-and-subtracts.) You must follow several
guidelines in your C code to help the compiler generate dual-MAC operations.

3-22
 Refining the C/C++ Code

In order for the compiler to generate a dual-MAC operation, the code must
have two consecutive MAC (or MAS/multiply) instructions that get all their mul-
tiplicands from memory and share one multiplicand. The two operations must
not write their results to the same variable or location. The compiler can easily
turn this example into a dual-MAC:
int *a,*b, onchip *c;
long s1,s2;
[...]
s1 = s1 + (*a++ * *c);
s2 = s2 + (*b++ * *c++);

This is a sequence of two MAC instructions that share the *c memory referen-
ce. Intrinsics can also be transformed into dual-MACs:
s1 = _smac(s1,*a++,*c);
s2 = _smac(s2,*b++,*c++);

You must inform the compiler that the memory pointed to by the shared dual-
MAC operand is on chip (a requirement for the addressing mode used for the
shared operand). There are two ways to do this. The first (and preferred) way
involves the use of the onchip type qualifier. It is used like this:
void foo(int onchip *a)
{
int onchip b[10];
...
}

This keyword can be applied to any pointer or array and indicates that the
memory pointed to by that pointer or array is always on chip.

The second technique is to compile with the −mb switch (passed to cl55). This
asserts to the compiler that all data pointed to by the shared dual-MAC pointer
will be on chip. This switch is a shortcut. Instead of putting many onchip quali-
fiers into the code, −mb can be used instead. You must ensure that all required
data will be on chip. If −mb is used and some data pointed to by a shared dual-
MAC pointer is not on chip, undefined behavior may result. Remember, this is a
shortcut. The onchip keyword should be used to enable dual-MAC opera-
tions in most circumstances. Using −mb could result in dual-MACs being gen-
erated in unexpected or undesirable places.

Unfortunately, a lot of C code that could benefit from using dual-MACs is not
written in such a way as to enable the compiler to generate them. However, the
compiler can sometimes transform the code in such a way to generate a dual-
MAC. For example, look at Example 3−16 which shows a C version of a simple

Optimizing C Code 3-23

Refining the C/C++ Code

FIR filter. (Notice the onchip keyword used for the pointer parameter h.) In
order to generate a dual-MAC in this case, the compiler must somehow gener-
ate two consecutive MAC operations from the single MAC operation in the
code. This is done via a loop transformation called unroll-and-jam. This trans-
formation replicates the outer loop and then fuses the two resulting inner loops
back together. Example 3−17 shows what the code in Example 3−16 would
look like if unroll-and-jam were applied manually.

Example 3−16. C Code for an FIR Filter

void fir(short onchip *h, short *x, short * y, short m, short n)

{
short i,j;
long y0;
for (j = 0; j < m; j++)
{
y0 = 0;

for (i = 0; i < n; i++)

y0 += (long)x[i + j] * h[i];

y[j] = (short) (y0 >> 16);

}
}

3-24
 Refining the C/C++ Code

Example 3−17. FIR C Code After Unroll-and-Jam Transformation

void fir(short onchip *h, short *x, short *y, short m, short n)
{
short i,j;
long y0,y1;
for (j = 0; j < m; j+=2)
{
y0 = 0;
y1 = 0;

for (i = 0; i < n; i++)

{
y0 += (long)x[i + j] * h[i];
y1 += (long)x[i + j+1] * h[i];
}

y[j] = (short) (y0 >> 16);

y[j+1] = (short) (y1 >> 16);
}
}

Notice that now we are computing two separate sums (y0 and y1) for each
iteration of the outer loop. If this C code were fed to the compiler, it would gen-
erate a dual-MAC in the inner loop. The compiler can perform the unroll-and-
jam transformation automatically, but the programmer must provide additional
information to ensure that the transformation is safe.

- The compiler must determine that the outer loop repeats an even number
of times. If the loop bounds are provably constant, the compiler can deter-
mine this automatically. Otherwise, if the user knows that the loop always
repeats an even number of times, a MUST_ITERATE pragma can be
used immediately preceding the outer loop:
#pragma MUST_ITERATE(1,,2)
(Note that the first parameter (1) indicates that the outer loop always exe-
cutes at least once. This is to eliminate loop jump-around code as de-
scribed in section 3.4.1.4 on page 3-18.)

Optimizing C Code 3-25

Refining the C/C++ Code

- The compiler must also know that the inner loop executes at least once.
This can be specified by inserting the following MUST_ITERATE pragma
just before the for statement of the inner loop:
#pragma MUST_ITERATE(1)

- The compiler must also know that there are no memory conflicts in the loop
nest. In our example, that means the compiler must know that all the writes
to array y cannot affect the values in array x or h. Consider the code in
Example 3−17 on page 3-25. We have changed the order of memory ac-
cesses by performing unroll-and-jam. In the transformed code, twice as
many reads from x (and h) occur before any writes to y. If writes to y could
affect the data pointed to by x (or h), the transformed code could produce
different results. If these three arrays were locally declared arrays, the
compiler would not have a problem. In this case we pass the arrays into
the function via pointer parameters. If the programmer is sure that writes
to y will not affect the arrays x and h within the function, the restrict
keyword can be used in the function declaration:
void fir(short onchip *h, short *x, short * restrict
y, short m, short n)
The restrict keyword tells the compiler that no other variable will point
at the memory that y points to. (See section 3.1.3 for more information on
memory dependences and restrict.) The final C code is shown in
Example 3−18, and the corresponding assembly code in Example 3−19.
Even using the MUST_ITERATE pragma and restrict qualifiers, some
loops may still be too complicated for the compiler to generate as dual-
MACs. If there is a piece of code you feel could benefit from dual-MAC op-
erations, it may be necessary to transform the code by hand. This process
is similar to the transformations described for writing dual-MAC operations
in assembly code as described in section 4.1.

3-26
 Refining the C/C++ Code

Example 3−18. FIR Filter With MUST_ITERATE Pragma and restrict Qualifier

void fir(short onchip *h, short *x, short * restrict y, short m,

short n)
{
short i,j;
long y0;
#pragma MUST_ITERATE(1,,2)
for (j = 0; j < m; j++)
{
y0 = 0;
#pragma MUST_ITERATE(1)
for (i = 0; i < n; i++)
y0 += (long)x[i + j] * h[i];

y[j] = (short) (y0 >> 16);

}
}

Optimizing C Code 3-27

Refining the C/C++ Code

Example 3−19. Generated Assembly for FIR Filter Showing Dual-MAC

_fir:
ADD #1, T0, AR3
SFTS AR3, #−1
SUB #1, AR3
MOV AR3, BRC0
PSH T3, T2
MOV #0, T3 ; |6|
|| MOV XAR0, XCDP
AADD #−1, SP
RPTBLOCAL L4−1

SUB #1, T1, T2

MOV XAR1, XAR3
MOV T2, CSR
ADD T3, AR3
MOV XAR3, XAR4
ADD #1, AR4
MOV #0, AC0 ; |8|
RPT CSR
|| MOV AC0, AC1 ; |8|
MAC *AR4+, *CDP+, AC0 :: MAC *AR3+, *CDP+, AC1
; loop ends
L3:
MOV XCDP, XAR0
ADD #2, T3
SUB T1, AR0
|| MOV HI(AC0), *AR2(short(#1))
ADD #2, AR2
|| MOV HI(AC1), *AR2
MOV XAR0, XCDP

AADD #1, SP
POP T3,T2
return

Note: The algebraic instructions code example for Generated Assembly for FIR Filter Showing Dual-MAC is shown in
Example B−9 on page B-7.

3-28
 Refining the C/C++ Code

3.4.3 Using Intrinsics

The C55x compiler provides intrinsics, special functions that map directly to
inlined C55x instructions, to optimize your C code quickly. Intrinsics are speci-
fied with a leading underscore ( _ ) and are accessed by calling them as you
would call a function.

For example, without intrinsics, saturated addition can only be expressed in C

code by writing a multicycle function, such as the one in Example 3−20.

Example 3−21 shows the resultant inefficient assembly language code gener-
ated by the compiler.

Example 3−20. Implementing Saturated Addition in C

int sadd(int a, int b)

{
int result;
result = a + b;

// Check to see if ’a’ and ’b’ have the same sign

if (((a^b) & 0x8000) == 0)
{
// If ’a’ and ’b’ have the same sign, check for underflow
// or overflow

if ((result ^ a) & 0x8000)

{
// If the result has a different sign than ’a’
// then underflow or overflow has occurred.
// if ’a’ is negative, set result to max negative
// If ’a’ is positive, set result to max positive
// value
result = ( a < 0) ? 0x8000 : 0x7FFF;
}
}
return result;
}

Optimizing C Code 3-29

Refining the C/C++ Code

Example 3−21. Inefficient Assembly Code Generated by C Version of Saturated Addition

_sadd:
MOV T1, AR1 ; |5|
XOR T0, T1 ; |9|
BTST @#15, T1, TC1 ; |9|
ADD T0, AR1
BCC L2,TC1 ; |9|
; branch occurs ; |9|
MOV T0, AR2 ; |9|
XOR AR1, AR2 ; |9|
BTST @#15, AR2, TC1 ; |9|
BCC L2,!TC1 ; |9|
; branch occurs ; |9|
BCC L1,T0 < #0 ; |22|
; branch occurs ; |22|
MOV #32767, T0 ; |22|
B L3 ; |22|
; branch occurs ; |22|
L1:
MOV #−32768, AR1 ; |22|
L2:
MOV AR1, T0 ; |25|
L3:
return ; |25|
; return occurs ; |25|

Note: The algebraic instructions code example for Inefficient Assembly Code Generated by C Version of Saturated Addition is
shown in Example B−10 on page B-8.

The code for the C simulated saturated addition can be replaced by a single
call to the _sadd intrinsic as is shown in Example 3−22. The assembly code
generated for this C source is shown in Example 3−23.

3-30
 Refining the C/C++ Code

Note that using compiler intrinsics reduces the portability of your code. You
may consider using ETSI functions instead of intrinsics. These functions can
be mapped to intrinsics for various targets. For C55x code, the file gsm.h de-
fines the ETSI functions using compiler intrinsics. (The actual C code ETSI
functions can be used when compiling on the host or other target without intrin-
sics.) For example, the code in Example 3−22 could be rewritten to use the
ETSI add function as shown in Example 3−24. The ETSI add function is
mapped to the _sadd compiler intrinsic in the header file gsm.h. (Of course,
you probably want to replace calls to the sadd function with calls to the ETSI
add function.)

Table 3−6 lists the intrinsics supported by the C55x compiler. For more infor-
mation on using intrinsics, please refer to the TMS320C55x Optimizing C
Compiler User’s Guide (SPRU281).

Example 3−22. Single Call to _sadd Intrinsic

int sadd(int a, int b)

{
return _sadd(a,b);
}

Example 3−23. Assembly Code Generated When Using Compiler Intrinsic for
Saturated Add

_sadd:
BSET ST3_SATA
ADD T1, T0 ; |3|
BCLR ST3_SATA
return ; |3|
; return occurs ; |3|

Note: The algebraic instructions code example for Assembly Code Generated When Using Compiler Intrinsic for Saturated
Add is shown in Example B−11 on page B-9.

Example 3−24. Using ETSI Functions to Implement sadd

#include <gsm.h>
int sadd(int a, int b)
{
return add(a,b);
}

Optimizing C Code 3-31

Refining the C/C++ Code

Table 3−6. TMS320C55x C Compiler Intrinsics

Intrinsic C Compiler Intrinsic Description

int _sadd(int src1, int src2); Adds two 16-bit integers, producing a saturated 16-bit re-
sult (SATA bit set)

long _lsadd(long src1, long src2); Adds two 32-bit integers, producing a saturated 32-bit re-
sult (SATD bit set)

long long _llsadd(long long src1, long long src2); Adds two 40-bit integers, producing a saturated 40-bit re-
sult (SATD bit set)

int _ssub(int src1, int src2); Subtracts src2 from src1, producing a saturated 16-bit
result (SATA bit set)

long _lssub(long src1, long src2); Subtracts src2 from src1, producing a saturated 32-bit
result (SATD bit set)

long long _llssub(long long src1, long long src2); Subtracts src2 from src1, producing a saturated 40-bit
result (SATD bit set)

int _smpy(int src1, int src2); Multiplies src1 and src2, and shifts the result left by 1. Pro-
duces a saturated 16-bit result. (SATD and FRCT bits set)

long _lsmpy(int src1, int src2); Multiplies src1 and src2, and shifts the result left by 1. Pro-
duces a saturated 32-bit result. (SATD and FRCT bits set)

long _smac(long src, int op1, int op2); Multiplies op1 and op2, shifts the result left by 1, and adds
it to src. Produces a saturated 32-bit result. (SATD, SMUL,
and FRCT bits set)

long _smas(long src, int op1, int op2); Multiplies op1 and op2, shifts the result left by 1, and sub-
tracts it from src. Produces a 32-bit result. (SATD, SMUL
and FRCT bits set)

int _abss(int src); Creates a saturated 16-bit absolute value.

_abss(8000h) results in 7FFFh (SATA bit set)

long _labss(long src); Creates a saturated 32-bit absolute value.

_labss(8000000h) results in 7FFFFFFFh (SATD bit set)

long long _llabss(long long src); Creates a saturated 40-bit absolute value.
_llabss(800000000h) results in 7FFFFFFFFFh (SATD bit
set)

int _sneg(int src); Negates the 16-bit value with saturation.

_sneg(8000h) results in 7FFFh

long _lsneg(long src); Negates the 32-bit value with saturation.

_lsneg(80000000h) results in 7FFFFFFFh

long long _llsneg(long long src); Negates the 40-bit value with saturation.
_llsneg(8000000000h) results in 7FFFFFFFFFh

3-32
 Refining the C/C++ Code

Table 3−6. TMS320C55x C Compiler Intrinsics (Continued)

Intrinsic C Compiler Intrinsic Description

long _smpyr(int src1, int src2); Multiplies src1 and src2, shifts the result left by 1, and
rounds by adding 215 to the result and zeroing out the low-
er 16 bits. (SATD and FRCT bits set)
long _smacr(long src, int op1, int op2); Multiplies op1 and op2, shifts the result left by 1, adds the
result to src, and then rounds the result by adding 215 and
zeroing out the lower 16 bits. (SATD , SMUL, and FRCT
bits set)

long _smasr(long src, int op1, int op2); Multiplies op1 and op2, shifts the result left by 1, subtracts
the result from src, and then rounds the result by adding
215 and zeroing out the lower 16 bits. (SATD , SMUL and
FRCT bits set)

int _norm(int src); Produces the number of left shifts needed to normalize
16-bit value.
int _lnorm(long src); Produces the number of left shifts needed to normalize
32-bit value.
long _rnd(long src); Rounds src by adding 215. Produces a 32-bit saturated
result with the lower 16 bits zeroed out. (SATD bit set)
int _sshl(int src1, int src2); Shifts src1 left by src2 and produces a 16-bit result. The
result is saturated if src2 is less than or equal to 8. (SATD
bit set)
long _lsshl(long src1, int src2); Shifts src1 left by src2 and produces a 32-bit result. The
result is saturated if src2 is less than or equal to 8. (SATD
bit set)
int _shrs(int src1, int src2); Shifts src1 right by src2 and produces a 16-bit result. Pro-
duces a saturated 16-bit result. (SATD bit set)
long _lshrs(long src1, int src2); Shifts src1 right by src2 and produces a 32-bit result. Pro-
duces a saturated 32-bit result. (SATD bit set)

Optimizing C Code 3-33

Refining the C/C++ Code

3.4.4 Using Long Data Accesses for 16-Bit Data

The primary use of treating 16-bit data as long is to transfer data quickly from
one memory location to another. Since 32-bit accesses also can occur in a
single cycle, this could reduce the data-movement time by half. The only limita-
tion is that the data must be aligned on a double word boundary (that is, an
even word boundary). The code is even simpler if the number of items trans-
ferred is a multiple of 2. To align the data use the DATA_ALIGN pragma:
short x[10];
#pragma DATA_ALIGN(x,2)

Example 3−25 shows a memory copy function that copies 16-bit data via
32-bit pointers.

Example 3−25. Block Copy Using Long Data Access

void copy(const short *a, const short *b, unsigned short n)

{
unsigned short i;
unsigned short na;
long *src, *dst;
// This code assumes that the number of elements to transfer ’n’
// is a multiple of 2. Divide the number of 1 word transfers
// by 2 to get the number of double word transfers.
na = (n>>1) −1;
// Set beginning address of SRC and DST for long transfer.
src = (long *)a;
dst = (long *)b;
for (i=0; i<= na; i++)
{
*dst++ = *src++;
}
}

3-34
 Refining the C/C++ Code

3.4.5 Simulating Circular Addressing in C

When simulating circular addressing in C, avoid using the modulus operator

(%). Modulus can take several cycles to implement and often results in a call to
a run-time library routine. Instead, use the macros shown in Example 3−26.
Use of these macros is only valid when the index increment amount (inc) is
less than the buffer size (size), and when the code of CIRC_UPDATE is al-
ways used to update the array index. Example 3−28 shows the same example
using modulus.

Notice that CIRC_REF simply expands to (var). In the future, using modulus
will be the more efficient way to implement circular addressing in C. The com-
piler will be able to transform certain uses of modulus into efficient C55x circu-
lar addressing code. At that time, the CIRC_UPDATE and CIRC_REF macros
can be updated to use modulus. Use of these macros will improve current per-
formance and minimize future changes needed to take advantage of improved
compiler functionality with regards to circular addressing.

Example 3−27 displays the resulting assembly code generated by the

compiler.

The (much less efficient) resulting assembly code is shown in Example 3−29.

Example 3−26. Simulating Circular Addressing in C

#define CIRC_UPDATE(var,inc,size)\
(var) +=(inc); if ((var)>=(size)) (var)−=(size);
#define CIRC_REF(var,size) (var)
long circ(const int *a, const int *b, int nb, int na)
{
int i,x=0;
long sum=0;
for(i=0; i<na; i++)
{
sum += (long)a[i] * b[CIRC_REF(x,nb)];
CIRC_UPDATE(x,1,nb)
}
return sum;
}

Optimizing C Code 3-35

Refining the C/C++ Code

Example 3−27. Assembly Output for Circular Addressing C Code

_circ:
MOV #0, AC0 ; |7|
BCC L2,T1 <= #0 ; |9|
; branch occurs ; |9|
SUB #1, T1, AR3
MOV AR3, BRC0
MOV #0, AR2 ; |6|
RPTBLOCAL L2−1
; loop starts
L1:
MACM *AR1+, *AR0+, AC0, AC0 ; |11|
|| ADD #1, AR2
CMP AR2 < T0, TC1 ; |12|
XCCPART !TC1 ||
SUB T0, AR1
XCCPART !TC1 ||
SUB T0, AR2
; loop ends ; |13|
L2:
return ; |14|
; return occurs ; |14|

Note: The algebraic instructions code example for Assembly Output for Circular Addressing C Code is shown in Example B−12
on page B-9.

3-36
 Refining the C/C++ Code

Example 3−28. Circular Addressing Using Modulus Operator

long circ(const int *a, const int *b, int nb, int na)
{
int i,x=0;
long sum=0;
for(i=0; i<na; i++)
{
sum += (long)a[i] * b[x % mb];
x++;
}
return sum;
}

Optimizing C Code 3-37

Refining the C/C++ Code

Example 3−29. Assembly Output for Circular Addressing Using Modulus Operator

_circ:
PSH T3, T2
AADD #−7, SP
MOV XAR1, dbl(*SP(#0))
|| MOV #0, AC0 ; |4|
MOV AC0, dbl(*SP(#2)) ; |4|
|| MOV T0, T2 ; |2|
BCC L2,T1 <= #0 ; |6|
; branch occurs ; |6|
MOV #0, T0 ; |3|
MOV T1, T3
|| MOV XAR0, dbl(*SP(#4))
L1:
MOV dbl(*SP(#0)), XAR3
MOV *AR3(T0), T1 ; |8|
MOV dbl(*SP(#4)), XAR3
MOV dbl(*SP(#2)), AC0 ; |8|
ADD #1, T0, T0
MACM *AR3+, T1, AC0, AC0 ; |8|
MOV AC0, dbl(*SP(#2)) ; |8|
MOV XAR3, dbl(*SP(#4))
call #I$$MOD ; |9|
|| MOV T2, T1 ; |9|
; call occurs [#I$$MOD] ; |9|
SUB #1, T3
BCC L1,T3 != #0 ; |10|
; branch occurs ; |10|
L2:
MOV dbl(*SP(#2)), AC0
AADD #7, SP ; |11|
POP T3,T2
return ; |11|
; return occurs ; |11|

Note: The algebraic instructions code example for Assembly Output for Circular Addressing Using Modulo is shown in
Example B−13 on page B-10.

3-38
 Refining the C/C++ Code

3.4.6 Generating Efficient Control Code

Control code typically tests a number of conditions to determine the appropri-

ate action to take.

The compiler generates similar constructs when implementing nested if-then-

else and switch/case constructs when the number of case labels is fewer than
eight. Because the first true condition is executed with the least amount of
branching, it is best to allocate the most often executed conditional first. When
the number of case labels exceeds eight, the compiler generates a .switch la-
bel section. In this case, it is still optimal to place the most often executed code
at the first case label.

In the case of single conditionals, it is best to test against zero. For example,
consider the following piece of C code:
if (a!=1) /* Test against 1 */
<inst1>
else
<inst2>

If the programmer knows that a is always 0 or 1, the following more efficient C

code can be used:
if (a==0) /* Test against 0 */
<inst1>
else
<inst2>

In most cases this test against zero will result in more efficient compiled code.

Optimizing C Code 3-39

Refining the C/C++ Code

3.4.7 Summary of Coding Idioms for C55x

Table 3−7 shows the C coding methods that should be used for some basic
DSP operations to generate the most efficient assembly code for the C55x.

Table 3−7. C Coding Methods for Generating Efficient C55x Assembly Code

Operation Recommended C Code Idiom

16bit * 16bit => 32bit (multiply) int a,b;

long c;
c = (long)a * b;

Q15 * Q15 => Q15 (multiply) int a,b,c;

Fractional mode with saturation c = _smpy(a,b);

Q15 * Q15 => Q31 (multiply) int a,b;

Fractional mode with saturation long c;
c = _lsmpy(a,b);

32bit + 16bit * 16bit => 32 bit (MAC) int a,b;

long c;
c = c + ((long)a * b));

Q31 + Q15 * Q15 => Q31 (MAC) int a,b;

Fractional mode with saturation long c;
c = _smac(c,a,b);

32bit – 16bit * 16bit => 32 bit (MAS) int a,b;

long c;
c = c – ((long)a * b));

Q31 – Q15 * Q15 => Q31 (MAS) int a,b;

Fractional mode with saturation long c;
c = _smas(c,a,b);

16bit +/− 16bit => 16bit <int, long, long long> a,b,c;
32bit +/− 32bit => 32bit c = a + b;
40bit +/− 40bit => 40bit (addition or subtraction) /* or */
c = a – b;

16bit + 16bit => 16bit (addition) int a,b,c;

with saturation c = _sadd(a,b);

32bit + 32bit => 32bit (addition) long a,b,c;

with saturation c = _lsadd(a,b);

40bit + 40bit => 40bit (addition) long long a,b,c;

with saturation c = _llsadd(a,b);

3-40
 Refining the C/C++ Code

Table 3−7. C Coding Methods for Generating Efficient C55x Assembly Code (Continued)
Operation Recommended C Code Idiom
16bit – 16bit => 16bit (subtraction) int a,b,c;
with saturation c = _ssub(a,b);

32bit – 32bit => 32bit (subtraction) long a,b,c;

with saturation c = _lssub(a,b);

40bit – 40bit => 40bit (subtraction) long long a,b,c;

with saturation c = _llssub(a,b);

|16bit| => 16bit <int, long, long long> a,b;

|32bit| => 32bit b = abs(a); /* or */
|40bit| => 40bit (absolute value) b = labs(a); /* or */
b = llabs(a);

|16bit| => 16bit <int, long, long long> a,b;

|32bit| => 32bit b = _abss(a); /* or */
|40bit| => 40bit (absolute value) b = _labss(a); /* or */
with saturation b = _llabss(a);

round(Q31) = > Q15 (rounding towards infinity) long a;

with saturation int b;
b = _rnd(a)>>16;

Q39 => Q31 (format change) long long a;

long b;
b = a >> 8;

Q30 = > Q31 (format change) long a;

with saturation long b;
b = _lsshl(a,1);

40bit => 32bit both Q31 (size change) long long a;

long b;
b = a;

Optimizing C Code 3-41

Memory Management Issues

3.5 Memory Management Issues

This section provides a brief discussion on managing data and code in

memory. Memory usage and subsequent code speed may be affected by a
number of factors. The discussion in this section will focus on the following
areas that affect memory usage. The information in this section is valid regard-
less of object code format (COFF or DWARF).

- Avoiding holes caused by data alignment

- Local versus global symbol declarations

- Stack configuration

- Allocating code and data in the C55x memory map

3.5.1 Avoiding Holes Caused by Data Alignment

The compiler requires that all values of type long be stored on an even word
boundary. When declaring data objects (such as structures) that may contain a
mixture of multi-word and single-word elements, place variables of type long
in the structure definition first to avoid holes in memory. The compiler automati-
cally aligns structure objects on an even word boundary. Placing these items
first takes advantage of this alignment.

Example 3−30. Considerations for Long Data Objects in Structures

/* Not recommended */
typedef struct abc{
int a;
long b;
int c;
} ABC;
/* Recommended */
typedef struct abc{
long a;
int b,c;
} ABC;

3-42
 Memory Management Issues

3.5.2 Local vs. Global Symbol Declarations

Locally declared symbols (symbols declared within a C function), are allocated

space by the compiler on the software stack. Globally declared symbols (sym-
bols declared at the file level) and static local variables are allocated space in
the compiler generated .bss section by default. The C operating environment
created by the C boot routine, _c_int00, places the C55x DSP in CPL mode.
CPL mode enables stack-based offset addressing and disables DP offset ad-
dressing. The compiler accesses global objects via absolute addressing
modes. Because the full address of the global object is encoded as part of the
instruction in absolute addressing modes, this can lead to larger code size and
potentially slower code. CPL mode favors the use of locally declared objects,
since it takes advantage of stack offset addressing. Therefore, if at all possible,
it is better to declare and manipulate local objects rather than global objects.
When function code requires multiple uses of a non-volatile global object, it is
better to declare a local object and assign it the appropriate value:
extern int Xflag;
int function(void)

{
int lflag = Xflag;

.
x = lflag ? lflag & 0xfffe : lflag;
.
.
return x;
}

3.5.3 Stack Configuration

The C55x has two software stacks: the data stack (referenced by the pointer
SP) and the system stack (referenced by the pointer SSP). These stacks can
be indexed independently or simultaneously depending on the chosen operat-
ing mode. There are three possible operating modes for the stack:

- Dual 16-bit stack with fast return

- Dual 16-bit stack with slow return

- 32-bit stack with slow return

Optimizing C Code 3-43

Memory Management Issues

In the 32-bit mode, SSP is incremented whenever SP is incremented. The pri-

mary use of SSP is to hold the upper 8 bits of the return address for context
saving. It is not used for data accesses. Because the C compiler allocates
space on the data stack for all locally declared objects, operating in this mode
doubles the space allocated for each local object. This can rapidly increase
memory usage. In the dual 16-bit modes, the SSP is only incremented for con-
text saving (function calls, interrupt handling). Allocation of memory for local
objects does not affect the system stack when either of the dual 16-bit modes is
used.

Additionally, the selection of fast return mode enables use of the RETA and
CFCT registers to effect return from functions. This potentially increases exe-
cution speed because it reduces the number of cycles required to return from a
function.

It is recommended to use dual 16-bit fast return mode to reduce memory space
requirements and increase execution speed. The stack operating mode is se-
lected by setting bits 28 and 29 of the reset vector address to the appropriate
values. Dual 16-bit fast return mode may be selected by using the .ivec assem-
bler directive when creating the address for the reset vector. For example:

.ivec reset_isr_addr, USE_RETA

(This is the default mode for the compiler as setup by the supplied runtime sup-
port library.) The assembler will automatically set the correct value for bits 28
and 29 when encoding the reset vector address. For more information on stack
modes, see the TMS320C55x DSP CPU Reference Guide (SPRU371).

3.5.4 Allocating Code and Data in the C55x Memory Map

The compiler groups generated code and data into logical units called sec-
tions. Sections are the building blocks of the object files created by the assem-
bler. They are the logical units operated on by the linker when allocating space
for code and data in the C55x memory map.

The compiler/assembler can create any of the sections described in

Table 3−8.

3-44
 Memory Management Issues

Table 3−8. Section Descriptions

Section Description
.cinit Initialization record table for global and static C variables

.pinit A list of constructor function pointers called at boot time

.const Explicitly initialized global and static const symbols

.text Executable code and constants

.bss Global and static variables

.ioport Uninitialized global and static variables of type ioport

.stack Data stack (local variables, lower 16 bits of return address, etc.)

.sysstack System stack (upper 8 bits of 24 bit return address)

.sysmem Memory for dynamic allocation functions

.switch Labels for switch/case

.cio For CIO Strings and buffers

These sections are encoded in the object file produced by the assembler.
When linking the objects, it is important to pay attention to where these sec-
tions are linked in memory to avoid as many memory conflicts as possible. Fol-
lowing are some recommendations:

- Allocate .stack and .sysstack in DARAM (dual-access RAM): the .stack

and .sysstack sections are often accessed at the same time when a func-
tion call/return occurs. If these sections are allocated in the same SARAM
(single-access RAM) block, then a memory conflict will occur, adding addi-
tional cycles to the call/return operation. If they are allocated in DARAM
or separate SARAM blocks, this will avoid such a conflict.

- The start address of the .stack and .sysstack sections are used to initialize
the data stack pointer (SP) and the system stack pointer (SSP), respec-
tively. Because these two registers share a common data page pointer
register (SPH) these sections must be allocated on the same 64K-word
memory page.

- Allocate the .bss and .stack sections in a single DARAM or separate SA-
RAM memory spaces. Local variable space is allocated on the stack. It is
possible that there may be conflicts when global variables, whose alloca-
tion is in .bss section, are accessed within the same instruction as a locally
declared variable.

Optimizing C Code 3-45

Memory Management Issues

- Use the DATA_SECTION pragma: If an algorithm uses a set of coefficients

that is applied to a known data array, use the DATA_SECTION pragma to
place these variables in their own named section. Then explicitly allocate
these sections in separate memory blocks to avoid conflicts.
Example 3−31 shows sample C source for using the DATA_SECTION
pragma to place variables in a user defined section.

Most of the preceding memory allocation recommendations are based on the

assumption that the typical operation accesses at most two operands.
Table 3−9 shows the possible operand combinations.

Example 3−31. Declaration Using DATA_SECTION Pragma

#pragma DATA_SECTION(h, ”coeffs”)

short h[10];
#pragma DATA_SECTION(x, ”mydata”)
short x[10];

Table 3−9. Possible Operand Combinations

Operand 1 Operand 2 Comment

Local var (stack) Local var (stack) If stack is in DARAM then no memory conflict will
occur

Local var(stack) Global var(.bss) If stack is in separate SARAM block or is in same

DARAM block, then no conflict will occur

Local var(stack) Const symbol (.const) If .const is located in separate SARAM or same DA-
RAM no conflict will occur

Global var(.bss) Global var(.bss) If .bss is allocated in DARAM, then no conflict will
occur

Global var(.bss) Const symbol(.const) If .const and .bss are located in separate SARAM or
same DARAM block, then no conflict will occur

When compiling with the small memory model (compiler default) allocate all
data sections, .data, .bss, .stack, .sysmem, .sysstack, .cio, and .const, on the
first 64K word page of memory (Page 0).

Example 3−32 contains a sample linker command file for the small memory
model. For extensive documentation on the linker and linker command files,
see the TMS320C55x Assembly Language Tools User’s Guide (SPRU280).

3-46
 Memory Management Issues

Example 3−32. Sample Linker Command File

/*********************************************************
LINKER command file for LEAD3 memory map.
Small memory model
**********************************************************/
−stack 0x2000 /* Primary stack size */
−sysstack 0x1000 /* Secondary stack size */
−heap 0x2000 /* Heap area size */
−c /* Use C linking conventions: auto−init vars at runtime */
−u _Reset /* Force load of reset interrupt handler */
MEMORY
{
PAGE 0: /* −−−− Unified Program/Data Address Space −−−− */
RAM (RWIX) : origin = 0x000100, length = 0x01ff00 /* 128Kb page of RAM */
ROM (RIX) : origin = 0x020100, length = 0x01ff00 /* 128Kb page of ROM */
VECS (RIX) : origin = 0xffff00, length = 0x000100 /*256−byte int vector*/
PAGE 1: /* −−−−−−−− 64K−word I/O Address Space −−−−−−−− */
IOPORT (RWI) : origin = 0x000000, length = 0x020000
}

SECTIONS
{
.text > ROM PAGE 0 /* Code */
/* These sections must be on same physical memory page */
/* when small memory model is used */
.data > RAM PAGE 0 /* Initialized vars */
.bss > RAM PAGE 0 /* Global & static vars */
.const > RAM PAGE 0 /* Constant data */
.sysmem > RAM PAGE 0 /* Dynamic memory (malloc) */
.stack > RAM PAGE 0 /* Primary system stack */
.sysstack > RAM PAGE 0 /* Secondary system stack */
.cio > RAM PAGE 0 /* C I/O buffers */
/* These sections may be on any physical memory page */
/* when small memory model is used */
.switch > RAM PAGE 0 /* Switch statement tables */
.cinit > RAM PAGE 0 /* Auto−initialization tables */
.pinit > RAM PAGE 0 /* Initialization fn tables */
vectors > VECS PAGE 0 /* Interrupt vectors */
.ioport > IOPORT PAGE 1 /* Global & static IO vars */

Optimizing C Code 3-47

Memory Management Issues

3.5.5 Allocating Function Code to Different Sections

The compiler provides a pragma to allow the placement of a function’s code

into a separate user defined section. The pragma is useful if it is necessary to
have some granularity in the placement of code in memory.

The pragma, in Example 3−33, defines a new section called .myfunc. The
code for the function myfunction() will be placed by the compiler into this
newly defined section. The section name can then be used within the SEC-
TIONS directive of a linker command file to explicitly allocate memory for this
function. For details on how to use the SECTIONS directive, see the
TMS320C55x Assembly Language Tools User’s Guide (SPRU280).

Example 3−33. Allocation of Functions Using CODE_SECTION Pragma

#pragma CODE_SECTION(myfunction, ”.myfunc”)

void myfunction(void)
{
.
.
}

3-48
 Chapter 4−

   

This chapter offers recommendations for producing TMS320C55x (C55x) as-

sembly code that:

- Makes good use of special architectural features, like the dual multiply-
and-accumulate (MAC) hardware, parallelism, and looping hardware.

- Produces no pipeline conflicts, memory conflicts, or instruction-fetch stalls

that would delay CPU operations.

This chapter shows ways you can optimize TMS320C55x assembly code, so
that you have highly-efficient code in time-critical portions of your programs.

Topic Page

4.1 Efficient Use of the Dual MAC Hardware . . . . . . . . . . . . . . . . . . . . . . . . 4-2

4.2 Using Parallel Execution Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-16
4.3 Implementing Efficient Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-42
4.4 Minimizing Pipeline and IBQ Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-49

4-1
Efficient Use of the Dual-MAC Hardware

4.1 Efficient Use of the Dual-MAC Hardware

Two of the most common operations in digital signal processing are the Multi-
ply and Accumulate (MAC) and the Multiply and Substract (MAS). The C55x
architecture can implement two multiply/accumulate (or two multiply/sub-
stract) operations in one cycle as shown in the typical C55x dual-MAC instruc-
tion below:
MAC *AR2+, *CDP+, AC0
:: MAC *AR3+, *CDP+, AC1

that performs

AC0 = AC0 + xmem * cmem

AC1 = AC1 + ymem * cmem

where xmem, ymem, and cmem are operands in memory pointed by registers
AR2, AR3, and CDP, respectively. Notice the following characteristics of C55x
dual-MAC instructions:

1) The dual-MAC/MAS operation can be performed using three oper-

ands only, which implies that one of the operands (cmem) should be com-
mon to both MAC/MAS operations.

The two MAC units on the C55x DSP are economically fed data via three
independent data buses: BB (the B bus), CB (the C bus), and DB (the D
bus). During a dual-MAC operation, each MAC unit requires two data op-
erands from memory (four operands total). However, the three data buses
are capable of providing at most three independent operands. To obtain
the required fourth operand, the data value on the B bus is used by both
MAC units. This is illustrated in Figure 4−1. With this structure, the fourth
data operand is not independent, but rather is dependent on one of the
other three operands.

Figure 4−1. Data Bus Usage During a Dual-MAC Operation

D bus

C bus

B bus

MAC unit #1 MAC unit #2

4-2
 Efficient Use of the Dual-MAC Hardware

In the most general case of two multiplications, one would expect a re-
quirement of four fully independent data operands. While this is true on the
surface, in most cases one can get by with only three independent oper-
ands and avoid degrading performance by specially structuring the DSP
code at either the algorithm or application level. The special structuring,
covered in sections 4.1.1 through 4.1.4, can be categorized as follows:
- Implicit algorithm symmetry (e.g., symmetric FIR, complex vector
multiply)
- Loop unrolling (e.g., block FIR, single-sample FIR, matrix multiply)
- Multi-channel applications
- Multi-algorithm applications

2) The common operand (cmem) has to be addressed using XCDP (Co-

efficient Data Pointer) and should be kept in internal memory since
the bus used to fetch this operand (B bus) is not connected to external
memory. For xmem and ymem operands, any of the eight auxiliary regis-
ters (XAR0−XAR7) can be used.

3) In order to perform a dual-MAC/MAS operation in one cycle, the com-

mon operand (cmem) should not be residing in the same memory
block with respect to the other two operands, as the maximum band-
width of a C55x memory block is two accesses in one cycle (internal DA-
RAM block).
If the cmem, xmem, and ymem operands point to the same data (for exam-
ple, during the autocorrelation of a signal), one option is to temporarily
copy the cmem data to a different block.
The programmer should make the appropriate decision whether the dual-
MAC cycle savings compensate the extra cycles and extra memory re-
quired by the data-copy process. If many functions need extra buffers in
order to use the dual-MAC, then an excess amount of data memory can be
consumed. One possibility for alleviating this problem is to allocate one
buffer (of maximum required size) and use it commonly across all the func-
tions.

Optimizing Assembly Code 4-3

Efficient Use of the Dual-MAC Hardware

4.1.1 Implicit Algorithm Symmetry

When an algorithm has internal symmetry, it can sometimes be exploited for
efficient dual-MAC implementation. One such example is a symmetric FIR fil-
ter. This filter has coefficients that are symmetrical with respect to delayed val-
ues of the input signal. The mathematical expression for a symmetric FIR filter
can be described by the following discrete-time difference equation:
N
Ă*Ă1

ȍ a ĂƪxĂǒk * j Ǔ ) xĂǒk ) j * N ) 1 Ǔƫ
2
y(k) + j
j=0

where

N = Number of filter taps (even)

x() = Element in the vector of input values
y() = Element in the vector of output values
k = Time index

Similar in form to the symmetrical FIR filter is the anti-symmetrical FIR filter:
N
Ă*Ă1

ȍ a ĂƪxĂǒk * j Ǔ * xĂǒk ) j * N ) 1 Ǔƫ
2
y(k) + j
j=0

Both the symmetrical and anti-symmetrical FIR filters can be implemented

using a dual-MAC approach because only three data values need be fetched
per inner loop cycle: aj, x(k − j), and x(k + j − N + 1). The coefficient aj is deliv-
ered to the dual-MAC units using the B bus and using XCDP as the pointer.
The C bus and the D bus are used along with two XARx registers to access
the independent elements x(k − j) and x(k + j − N + 1).

A second example of an algorithm with implicit symmetry is an element-by-ele-

ment complex vector multiply. Let {A}, {B}, and {C} be vectors of length N, and
let j be the imaginary unit value (i.e., square root of −1). The complex compo-
nents of {A}, {B}, and {C} can be expressed as
RE IM RE IM RE IM
a i + a i ) ja i ą b i + b i ) jb i ą c i + c i ) jc i ą for 1 v i v N

and the expression for each element in {C} is computed as

ci + ai * bi
ǒ RE
+ a i ) ja i
IM
Ǔ * ǒb RE
i
IM
) jb i Ǔ ąą
+ ǒa Ǔ ) jǒa Ǔ
RE RE IM IM RE IM IM RE
i b i *a i b i i b i )a i b i for 1 v i v N

4-4
 Efficient Use of the Dual-MAC Hardware

The required four multiplications in the above expression can be implemented

with two dual-MAC instructions by grouping the multiplications as follows:
RE RE IM RE
- 1st multiplication group: a i b i and a i b i

IM IM RE IM
- 2nd multiplication group: a i b i and a i b i

Each dual-multiply grouping requires only three independent operands. An as-

sembly code example for the complex vector multiply is given in Example 4−1
(showing mnemonic instructions). Note that this particular code assumes the
following arrangement in memory for a complex vector:

x RE
1
Lowest memory address

x IM
1

x RE
2

x IM
2

In addition, the code stores both portions of the complex result to memory at
the same time. This requires that the results vector be long-word aligned in me-
mory. One way to achieve this is through use of the alignment flag option with
the .bss directive, as was done with this code example. Alternatively, one could
place the results array in a separate uninitialized named section using a .usect
directive, and then use the linker command file to force long-word alignment
of that section.

Optimizing Assembly Code 4-5

Efficient Use of the Dual-MAC Hardware

Example 4−1. Complex Vector Multiplication Code

N .set 3 ; Length of each complex vector

.data
A .int 1,2,3,4,5,6 ; Complex input vector #1
B .int 7,8,9,10,11,12 ; Complex input vector #2

;Results are: 0xfff7, 0x0016, 0xfff3, 0x0042, 0xffef, 0x007e

.bss C, 2*N, ,1 ; Results vector, long−word aligned

.text
BCLR ARMS ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

cplxmul:
AMOV #A, XAR0 ; Pointer to A vector
AMOV #B, XCDP ; Pointer to B vector
AMOV #C, XAR1 ; Pointer to C vector
MOV #(N−1), BRC0 ; Load loop counter
MOV #1, T0 ; Pointer offset
MOV #2, T1 ; Pointer increment

RPTBLOCAL endloop ; Start the loop

MPY *AR0, *CDP+, AC0

:: MPY *AR0(T0), *CDP+, AC1

MAS *AR0(T0), *CDP+, AC0

:: MAC *(AR0+T1), *CDP+, AC1

endloop:
MOV pair(LO(AC0)), dbl(*AR1+) ; Store complex result
; End of loop

Note: The algebraic instructions code example for Complex Vector Multiplication is shown in Example B−14 on page B-11.

4.1.2 Loop Unrolling

Loop unrolling involves structuring computations to exploit the reuse of data
among different time or geometric iterations of the algorithm. Many algorithms
can be structured computationally to provide for such reuse and allow a dual-
MAC implementation.

In filtering, input and/or output data is commonly stored in a delay chain buffer.
Each time the filter is invoked on a new data point, the oldest value in the delay
chain is discarded from the bottom of the chain, while the new data value is
added to the top of the chain. A value in the chain will get reused (for example,
multiplied by a coefficient) in the computations over and over again as succes-

4-6
 Efficient Use of the Dual-MAC Hardware

sive time-step outputs are computed. The reuse will continue until such a time
that the data value becomes the oldest value in the chain and is discarded.
Dual-MAC implementation of filtering should therefore employ a time-based
loop unrolling approach to exploit the reuse of the data. This scenario is pre-
sented in sections 4.1.2.1 and 4.1.2.2.

An application amenable to geometric based loop unrolling is matrix computa-

tions. In this application, successive rows in a matrix get multiplied and accu-
mulated with the columns in another matrix. In order to obtain data reuse within
the loop kernel, the computations using two different rows of data should be
handled in parallel. This will be presented in section 4.1.2.3.

4.1.2.1 Temporal Loop Unrolling: Block FIR Filter

To efficiently implement a block FIR filter with the two MAC units, loop unrolling
must be applied so that two time-based iterations of the algorithm are com-
puted in parallel. This allows reuse of the coefficients.

Figure 4−2 illustrates the coefficient reuse for a 4-tap block FIR filter with
constant, real-value coefficients. The implementation computes two sequen-
tial filter outputs in parallel so that only a single coefficient, ai, is used by both
MAC units. Consider, for example, the computation of outputs y(k) and
y(k − 1). For the first term in each of these two rows, one MAC unit computes
a0x(k), while the second MAC unit computes a0x(k − 1). These two computa-
tions combined require only three different values from memory: a0, x(k), and
x(k − 1). Proceeding to the second term in each row, a1x(k − 1) and a1x(k − 2)
are computed similarly, and so on with the remaining terms. After fully comput-
ing the outputs y(k) and y(k − 1), the next two outputs, y(k − 2) and y(k − 3),
are computed in parallel. Again, the computation begins with the first two terms
in each of these rows. In this way, DSP performance is maintained at two MAC
operations per clock cycle.

Figure 4−2. Computation Groupings for a Block FIR (4-Tap Filter Shown)

y(k) = a0x(k) a1x(k−1) a2x(k−2) a3x(k−3)

y(k−1) = a0x(k−1) a1x(k−2) a2x(k−3) a3x(k−4)

y(k−2) = a0x(k−2) a1x(k−3) a2x(k−4) a3x(k−5)

y(k−3) = a0x(k−3) a1x(k−4) a2x(k−5) a3x(k−6)

Optimizing Assembly Code 4-7

Efficient Use of the Dual-MAC Hardware

Note that filters with either an even or odd number of taps are handled equally
well by this method. However, this approach does require one to compute an
even number of outputs y(). In cases where an odd number of outputs is de-
sired, one can always zero-pad the input vector x() with one additional zero
element, and then discard the corresponding additional output.

Note also that not all of the input data must be available in advance. Rather,
only two new input samples are required for each iteration through the algo-
rithm, thereby producing two new output values.

A non-optimized assembly code example for the block FIR filter is shown in
Example 4−2 (showing mnemonic instructions). An optimized version of the
same code is found in Example 4−3 (showing mnemonic instructions). The fol-
lowing optimizations have been made in Example 4−3:

- The first filter tap was peeled out of the inner loop and implemented using
a dual-multiply instruction (as opposed to a dual-multiply-and-accumulate
instruction). This eliminated the need to clear AC0 and AC1 prior to enter-
ing the inner loop each time.

- The last filter tap was peeled out of the inner loop. This allows for the use
of different pointer adjustments than in the inner loop, and eliminates the
need to explicitly rewind the CDP, AR0, and AR1 pointers.
The combination of these first two optimizations results in a requirement
that N_TAPS be a minimum of 3.

- Both results are now written to memory at the same time using a double
store instruction. Note that this requires the results array (OUT_DATA) to
be long-word aligned. One way to achieve this is through use of the align-
ment flag option with the .bss directive, as was done in this code example.
As an alternative, you could place the results array in a separate uninitial-
ized named section using a .usect directive, and then use the linker com-
mand file to force long-word alignment of that section.

- The outer loop start instruction, RPTBLOCAL, has been put in parallel with
the instruction that preceded it.

4-8
 Efficient Use of the Dual-MAC Hardware

Example 4−2. Block FIR Filter Code (Not Optimized)

N_TAPS .set 4 ; Number of filter taps

N_DATA .set 11 ; Number of input values

.data
COEFFS .int 1,2,3,4 ; Coefficients
IN_DATA .int 1,2,3,4,5,6,7,8,9,10,11 ; Input vector

;Results are: 0x0014, 0x001E, 0x0028, 0x0032,

; 0x003C, 0x0046, 0x0050, 0x005A

.bss OUT_DATA, N_DATA − N_TAPS + 1 ; Output vector

.text
BCLR ARMS ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

bfir:
AMOV #COEFFS, XCDP ; Pointer to coefficient array
AMOV #(IN_DATA + N_TAPS − 1), XAR0 ; Pointer to input vector
AMOV #(IN_DATA + N_TAPS), XAR1 ; 2nd pointer to input vector
AMOV #OUT_DATA, XAR2 ; Pointer to output vector
MOV #((N_DATA − N_TAPS + 1)/2 − 1), BRC0
; Load outer loop counter
MOV #(N_TAPS − 1), CSR ; Load inner loop counter

RPTBLOCAL endloop ; Start the outer loop

MOV #0, AC0 ; Clear AC0

MOV #0, AC1 ; Clear AC1

RPT CSR ; Start the inner loop

MAC *AR0−, *CDP+, AC0 ; All taps
:: MAC *AR1−, *CDP+, AC1

MOV AC0, *AR2+ ; Write 1st result

MOV AC1, *AR2+ ; Write 2nd result
MOV #COEFFS, CDP ; Rewind coefficient pointer
ADD #(N_TAPS + 2), AR0 ; Adjust 1st input vector
; pointer
endloop:
ADD #(N_TAPS + 2), AR1 ; Adjust 2nd input vector
; pointer
; End of outer loop

Note: The algebraic instructions code example for Block FIR Filter Code (Not Optimized) is shown in Example B−15 on
page B-12.

Optimizing Assembly Code 4-9

Efficient Use of the Dual-MAC Hardware

Example 4−3. Block FIR Filter Code (Optimized)

N_TAPS .set 4 ; Number of filter taps

N_DATA .set 11 ; Number of input values

.data
COEFFS .int 1,2,3,4 ; Coefficients
IN_DATA .int 1,2,3,4,5,6,7,8,9,10,11 ; Input vector

;Results are: 0x0014, 0x001E, 0x0028, 0x0032,

; 0x003C, 0x0046, 0x0050, 0x005A

.bss OUT_DATA, N_DATA − N_TAPS + 1, ,1

; Output vector, long word
aligned

.text
BCLR ARMS ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

bfir:
AMOV #COEFFS, XCDP ; Pointer to coefficient array
AMOV #(IN_DATA + N_TAPS − 1), XAR0 ; Pointer to input vector
AMOV #(IN_DATA + N_TAPS), XAR1 ; 2nd pointer to input vector
AMOV #OUT_DATA, XAR2 ; Pointer to output vector
MOV #((N_DATA − N_TAPS + 1)/2 − 1), BRC0
; Load outer loop counter
MOV #(N_TAPS − 3), CSR ; Load inner loop counter
MOV #(−(N_TAPS − 1)), T0 ; CDP rewind increment

MOV #(N_TAPS + 1), T1 ; ARx rewind increment

||RPTBLOCAL endloop ; Start the outer loop

MPY *AR0−, *CDP+, AC0 ; 1st tap

:: MPY *AR1−, *CDP+, AC1

RPT CSR ; Start the inner loop

MAC *AR0−, *CDP+, AC0 ; Inner taps
:: MAC *AR1−, *CDP+, AC1

MAC *(AR0+T1), *(CDP+T0), AC0 ; Last tap

:: MAC *(AR1+T1), *(CDP+T0), AC1

endloop:
MOV pair(LO(AC0)), dbl(*AR2+) ; Store both results
; End of outer loop

Note: The algebraic instructions code example for Block FIR Filter Code (Optimized) is shown in Example B−16 on page B-13.

4-10
 Efficient Use of the Dual-MAC Hardware

4.1.2.2 Temporal Loop Unrolling: Single-Sample FIR Filter

The temporally unrolled block FIR filter described in section 4.1.2.1 maintains
dual-MAC throughput by sharing a common coefficient between the two MAC
units. In some algorithms, the loop unrolling needs to be performed so that a
common data variable is shared instead. The single-sample FIR filter is an ex-
ample of such an algorithm. In the single-sample FIR filter, the calculations for
the current sample period are interlaced with those of the next sample period
in order to achieve a net performance of two MAC operations per cycle.

Figure 4−3 shows the needed computation groupings for a 4-tap FIR filter. At
any given time step, one multiplies and accumulates every other partial prod-
uct in the corresponding row, beginning with the first partial product in the row.
In addition, one also multiplies and accumulates every other term in the next
row (that is, the row above the current row) in advance of that time step, begin-
ning with the second partial product in the next row. In this way, each row is
fully computed over the course of two sample periods.

For example, at time step k, it is desired to compute y(k). The first term in the
y(k) row is a0x(k), which is computed using one of the two MAC units. In addi-
tion, the second MAC unit is used to compute the second term in the y(k+1)
row, a1x(k), in advance of time step k + 1. These two computations combined
require only three different values from memory: a0, a1, and x(k). Note that the
term x(k) is not available until time k. This is why calculations at each time step
must begin with the first term in the corresponding row.

The second term in the y(k) row is a1x(k − 1). However, this would have been
already computed during the first computation at time step k − 1 (similar to how
a1x(k) was just pre-computed for time step k+1) , so it can be skipped here.
The third term in the y(k) row, a2x(k − 2), is computed next, and at the same
time, the term a3x(k − 2) is computed in the y(k + 1) row in advance of time
step k+1.

Optimizing Assembly Code 4-11

Efficient Use of the Dual-MAC Hardware

Figure 4−3. Computation Groupings for a Single-Sample FIR With an

Even Number of TAPS (4-Tap Filter Shown)

y(k+2) = a0x(k+2) a1x(k+1) a2x(k) a3x(k−1)

y(k+1) = a0x(k−1) a1x(k) a2x(k−1) a3x(k−2)

y(k) = a0x(k) a1x(k−1) a2x(k−2) a3x(k−3)

y(k−1) = a0x(k−1) a1x(k−2) a2x(k−3) a3x(k−4)

y(k−2) a0x(k−2) a1x(k−3) a2x(k−4) a3x(k−5)

Notice that two separate running sums are maintained, one with partial prod-
ucts for the current time step, the other with pre-computed terms for the next
time step. At the next time step, the pre-computed running sum becomes the
current running sum, and a new pre-computed running sum is started from
zero. At the end of each sample period, the current running sum contains the
current filter output, which can be dispatched as required by the application.

The above approach is not limited to the 4-tap filter illustrated in Figure 4−3.
Any other filter with an even number of taps is a straightforward extension. For
filters with an odd number of taps, the computation groupings become prob-
lematic, in that the last grouping in each row is missing the pre-calculation term
in the row above it.

Figure 4−4 depicts this problem for a 5-tap filter. To overcome this problem,
one should pad the filter to the next higher even number of taps by using a zero
coefficient for the additional term. For example, the five tap filter is augmented
to

y(k) + a 0Ăx(k) ) a 1Ăx(k * 1) ) a 2Ăx(k * 2) ) a 3Ăx(k * 3) ) a 4Ăx(k * 4) ) 0[x(k * 5)]

In this way, any filter with an odd number of taps can be implemented as a filter
with an even number of taps but retain the frequency response of the original
odd-number-tap filter.

4-12
 Efficient Use of the Dual-MAC Hardware

Figure 4−4. Computation Groupings for a Single-Sample FIR With an

Odd Number of TAPS (5-Tap Filter Shown)
Incomplete
groupings

y(k+2) = a0x(k+2) a1x(k+1) a2x(k) a3x(k−1) a4x(k−2)

y(k+1) = a0x(k+1) a1x(k) a2x(k−1) a3x(k−2) a4x(k−3)

y(k) = a0x(k) a1x(k−1) a2x(k−2) a3x(k−3) a4x(k−4)

y(k−1) = a0x(k−1) a1x(k−2) a2x(k−3) a3x(k−4) a4x(k−5)

y(k−2) a0x(k−2) a1x(k−3) a2x(k−4) a3x(k−5) a4x(k−6)

4.1.2.3 Geometric Loop Unrolling: Matrix Mathematics

Matrix mathematics typically involves considerable data reuse. Consider the
general case of multiplying two matrices:

[C ] + [A ] [B ]

where
[A] = m × n matrix
[B] = n × p matrix
[C] = m × p matrix
m ≥ 1, n ≥ 1, p ≥ 1
The expression for each element in matrix C is given by:

ȍa
n
c i, j + i, k Ăb k, j ǒ1 v i v m,Ă 1 v j v pǓ
k=1

where the conventional notation xi,j is being used to represent the element of
matrix X in the ith row and jth column. There are basically two different options
for efficient dual-MAC implementation. First, one could compute ci,j and ci,j + 1
in parallel. The computations made are:

ȍa ȍa
n n
c i, j + i, k Ăb k, j c i, j+1 + i, k Ăb k, j+1
k=1 k=1

The element ai,k is common to both expressions. The computations can there-
fore be made in parallel, with the common data ai,k delivered to the dual-MAC
units using the B bus and using XCDP as the pointer. The C bus and the D bus
are used along with two XARx registers to access the independent elements
bk,j and bk,j+1.

Optimizing Assembly Code 4-13

Efficient Use of the Dual-MAC Hardware

Alternatively, one could compute ci,j and ci+1,j in parallel. The computations
made are then:

ȍa ȍa
n n
c i, j + i, kĂb k, j c i+1, j + i+1, k Ăb k, j
k=1 k=1

In this case, the element bk,j is common to both expressions. They can there-
fore be made in parallel, with the common data bk,j delivered to the dual-MAC
units using the B bus and using XCDP as the pointer. The C bus and the D bus
are used along with two XARx registers to access the independent elements
ai,k and ai+1,k.

The values of m and p determine which approach one should take. Because
the inner loop will compute two elements in matrix C each iteration, clearly it
is most efficient if an even number of elements can be computed. Therefore,
if p is even, one should implement the first approach: compute ci,j and ci,j+1 in
parallel. Alternatively, if m is even, the second approach is more efficient: com-
pute ci,j and ci+1,j in parallel. If both m and p are even, either approach is ap-
propriate. Finally, if neither m nor p is even, there will be an extra element c that
will need to be computed individually each time through the inner loop. One
could add additional single-MAC code to handle the final element in the inner
loop. Alternatively, one could pad either matrix A or matrix B with a row or col-
umn or zeros (as appropriate) to make either m or p even. The elements in ma-
trix C computed using the pad row or column should then be discarded after
computation.

4.1.3 Multichannel Applications

In multichannel applications, the same signal processing is often done on two

or more independent data streams. Depending on the specific type of process-
ing being performed, it may be possible to process two channels of data in par-
allel, one channel in each MAC unit. In this way, a common set of constant co-
efficients can be shared.

An application example readily amenable to this approach is non-adaptive fil-

tering where the same filter is being applied to two different data streams. Both
channels are processed in parallel, one channel in each of the two MAC units.
For example, the same FIR filter applied to two different data streams can be
represented mathematically by the following expressions:

N-1 N-1

y 1(k) + ȍ a x ǒk * j Ǔ
j 1 y 2(k) + ȍ a x ǒk * j Ǔ
j 2
j=0 j=0

4-14
 Efficient Use of the Dual-MAC Hardware

where

N = Number of filter taps

aj = Element in the coefficient array
xi() = Element in the ith vector of input values
yi() = Element in the ith vector of output values
k = Time index

The value aj is common to both calculations. The two calculations can there-
fore be performed in parallel, with the common aj delivered to the dual-MAC
units using the B bus and using XCDP as the pointer. The C bus and the D bus
are used along with two XARx registers to access the independent input ele-
ments x1(k − j) and x2(k − j).

A second example is the correlation computation between multiple incoming

data streams and a fixed data vector. Suppose it is desired to compute the cor-
relation between the vectors X1 and Y, and also between the vectors X2 and
Y. One would need to compute the following for each element in the correlation
vectors R1 and R2:

r x1ĂyĂǒ j Ǔ + ȍ x Ăǒk ) j ǓĂy(k)

1 r x2ĂyĂǒ j Ǔ + ȍ x Ăǒk ) j ǓĂy(k)
2
k k

The element y(k) is common to both calculations. The two calculations can
therefore be performed in parallel, with the common data y(k) delivered to the
dual-MAC units via the B bus with XCDP as the address pointer. The C bus
and the D bus are used along with two XARx registers to access the indepen-
dent elements x1(k + j) and x2(k + j).

4.1.4 Multi-Algorithm Applications

When two or more different processing algorithms are applied to the same
data stream, it may be possible to process two such algorithms in parallel. For
example, consider a statistical application that computes the autocorrelation
of a vector X, and also the correlation between vector X and vector Y. One
would need to compute the following for each element in the correlation vec-
tors Rxx and Rxy:

r xxĂǒ j Ǔ + ȍ xĂǒk ) j ǓĂx(k) r xyĂǒ j Ǔ + ȍ xĂǒk ) j ǓĂy(k)

k k

The element x(k + j) is common to both calculations. The two calculations can
therefore be made in parallel, with the common data x(k + j) delivered to the
dual-MAC units using the B bus and using XCDP as the pointer. The C bus and
the D bus are used along with two XARx registers to access the independent
elements x(k) and y(k).

Optimizing Assembly Code 4-15

Using Parallel Execution Features

4.2 Using Parallel Execution Features

The C55x architecture allows programmers to place two operations or instruc-
tions in parallel to reduce the total execution time. There are two types of paral-
lelism: Built-in parallelism within a single instruction and user-defined parallel-
ism between two instructions. Built-in parallelism (see section 4.2.1) is auto-
matic; as soon as you write the instruction, it is put in place. User-defined paral-
lelism is optional and requires decision-making. Sections 4.2.2 through 4.2.8
present the rules and restrictions associated with the use of user-defined par-
allelism, and give examples of using it.

4.2.1 Built-In Parallelism

Instructions that have built-in parallelism perform two different operations in

parallel. In the algebraic syntax, they can be identified by the comma that sepa-
rates the two operations, as in the following example:
AC0 = *AR0 * coef(*CDP), ; The data referenced by AR0 is multiplied by
AC1 = *AR1 * coef(*CDP) ; a coefficient referenced by CDP. At the same time
; the data referenced by AR1 is multiplied by the
; same coefficient.

In the mnemonic syntax, they can be identified by a double colon (::) that sepa-
rates the two operations. The preceding example in the mnemonic syntax is:
MPY *AR0, *CDP, AC0 ; The data referenced by AR0 is multiplied by
:: MPY *AR1, *CDP, AC1 ; a coefficient referenced by CDP. At the same time
; the data referenced by AR1 is multiplied by the
; same coefficient.

4.2.2 User-Defined Parallelism

Two instructions may be placed in parallel to have them both execute in a

single cycle. The two instructions are separated by the || separator. One of the
two instructions may have built-in parallelism. The following algebraic code ex-
ample shows a user-defined parallel instruction pair. One of the instructions
in the pair also features built-in parallelism.
AC0 = AC0 + (*AR3+ * coef(*CDP+)), ; 1st instruction (has built−in parallelism)
AC1 = AC1 + (*AR4+ * coef(*CDP+))
|| repeat(CSR) ; 2nd instruction

The equivalent mnemonic code example is:

MPY *AR3+, *CDP+, AC0 ; 1st instruction (has built−in parallelism)
:: MPY *AR4+, *CDP+, AC1
|| RPT CSR ; 2nd instruction

4-16
 Using Parallel Execution Features

4.2.3 Architectural Features Supporting Parallelism

The C55x architecture provides three main, independent computation units
that are controlled by the instruction buffer unit (I unit):
- Program flow unit (P unit)
- Address-data flow Unit (A unit)
- Data computation unit (D unit)

The C55x instructions make use of dedicated operative resources (or opera-
tors) within each of the units. In total, there are 14 operators available across
the three computation units, and the parallelism rules enable the use of two
independent operators in parallel within the same cycle. If all other rules are
observed, two instructions that independently use any two of the independent
operators may be placed in parallel.

Figure 4−5 shows a matrix that reflects the 14 operators mentioned and the
possible operator combinations that may be used in placing instructions in par-
allel. The operators are ordered from rows 1 through 14 as well as columns
1 though 14. A blank cell in any given position (row I, column J) in the matrix
indicates that operator I may be placed in parallel with operator J, and an X in
any given position indicates that the two operators cannot be placed in parallel.
For example, a D-Unit MAC operation (row 7) may be placed in parallel with
a P-Unit Load operation (column 13) but cannot be placed in parallel with a
D-Unit ALU operation (column 5).

Optimizing Assembly Code 4-17

Using Parallel Execution Features

Figure 4−5. Matrix to Find Operators That Can Be Used in Parallel

D-unit Shift, Store

P-unit Control
D-unit Shifter
A-unit Swap

D-unit Swap
A-unit Store

D-unit Store

P-unit Store
A-unit Load

D-unit Load

P-unit Load
A-unit ALU

D-unit MAC
D-unit ALU
1 2 3 4 5 6 7 8 9 10 11 12 13 14

A-unit ALU 1 X

A-unit Swap 2 X

A-unit Load 3

A-unit Store 4

D-unit ALU 5 X X X

D-unit Shifter 6 X X X X

D-unit MAC 7 X X X

D-unit Load 8

D-unit Store 9

D-unit Shift, Store 10 X X

D-unit Swap 11 X

P-unit Control 12 X

P-unit Load 13

P-unit Store 14

Note: X in a table cell indicates that the operator in that row and the operator in that
column cannot be used in parallel with each other. A blank table cell indicates
that the operators can be used in parallel.

4-18
 Using Parallel Execution Features

Bus resources also play an important part in determining whether two instruc-
tions may be placed in parallel. Typically, programmers should be concerned
with the data buses and the constant buses. Table 4−1 lists and describes the
main CPU buses of interest and gives examples of instructions that use the
different buses. These may also be seen pictorially in Figure 4−6. Figure 4−6
also shows all CPU buses and the registers/operators in each of the three
functional units.

Table 4−1. CPU Data Buses and Constant Buses

Example:
Bus Type Bus(es) Description of Bus(es) Instruction That Uses The Bus(es)

Data BB Special coefficient read bus MPY *AR1+, *CDP+, AC0

:: MPY *AR3+, *CDP+, AC1
The operand referenced by CDP is carried
to the CPU on BB.

CB, DB Data-read buses MOV *AR3+, AC0

The operand referenced by AR3 is carried
to the low half of AC0 on DB.

EB, FB Data-write buses MOV AC0, *AR3

The low half of AC0 is carried on EB to
the location referenced by AR3.

Constant KAB Constant bus used in the address P-unit use:

phase of the pipeline to carry ad-
B #Routine2
dresses:
The constant Routine2 is carried on KPB
- The P unit uses KAB to gener-
to the P unit, where it is used for a pro-
ate program-memory address-
gram-memory address.
es.
A-unit use:
- The A unit uses KAB to gener-
ate data-memory addresses. MOV *SP(#7), BRC0
The immediate offset (7) is carried to the
data-address generation unit (DAGEN) on
KAB.

KDB Constant bus used by the A unit or ADD #1234h, AC0

D unit for computations. This bus is
The constant 1234h is carried on KDB to
used in the execute phase of the in-
the D-unit ALU, where it is used in the
struction pipeline.
addition.

Optimizing Assembly Code 4-19

Using Parallel Execution Features

Figure 4−6. CPU Operators and Buses

FB EB DB CB KDB KAB KAB KDB CB DB EB FB BB CB DB KDB

D-unit

A-unit A-unit D-unit D-unit

register file DAGEN register file MAC1

A-unit P-unit

A-unit ALU P-unit D-unit shifter

D-unit D-unit
with shift (including register file) (embeds dedicated
ALU MAC2
capability ALU)

EB DB KDB FB EB DB CB KAB FB EB DB CB KDB FB EB DB CB KDB BB CB DB KDB

4.2.4 User-Defined Parallelism Rules

This section describes the rules that a programmer must follow to place two
instructions in parallel. It is essential to note here that all the rules must be ob-
served for the parallelism to be valid. However, this section begins with a set
of four basic rules (Table 4−2) that a programmer may use when implementing
user-defined parallelism. If these are not sufficient, the set of advanced rules
(Table 4−3) needs to be considered.

4-20
 Using Parallel Execution Features

Table 4−2. Basic Parallelism Rules

Consideration Rule
Hardware resource conflicts Two instructions in parallel cannot compete for operators
(see Figure 4−5, page 4-18) or buses (see Table 4−1, page
4-19).

Maximum instruction length The combined length of the instruction pair cannot exceed 6
bytes.

Parallel enable bit If either of the following cases is true, the instructions can be
OR placed in parallel:
Soft dual encoding
- Parallel enable bit is present: At least one of two instruc-
tions in parallel must have a parallel enable bit in its instruc-
tion code. The instruction set reference guides (see Re-
lated Documentation from Texas Instruments in the pref-
ace) indicate whether a given instruction has a parallel en-
able bit.

- Soft dual encoding is present: For parallel instructions

that use Smem or Lmem operands, each instruction must
use one of the indirect operands allowed for the dual AR in-
direct addressing mode:

*ARn
*ARn+
*ARn−
*(ARn + T0) (Available if C54CM bit = 0)
*(ARn + AR0) (Available if C54CM bit = 1)
*(ARn − T0) (Available if C54CM bit = 0)
*(ARn − AR0) (Available if C54CM bit = 1)
*ARn(T0) (Available if C54CM bit = 0)
*ARn(AR0) (Available if C54CM bit = 1)
*(ARn + T1)
*(ARn − T1)

Optimizing Assembly Code 4-21

Using Parallel Execution Features

Table 4−3. Advanced Parallelism Rules

Consideration Rule
Byte extensions for constants An instruction that uses one of the following addressing-
mode operands cannot be placed in parallel with another
instruction. The constant (following the # symbol) adds 2 or 3
bytes to the instruction.
*abs16(#k16)
*port(#k16) (algebraic syntax)
port(#k16) (mnemonic syntax)
*(#k23)
*ARn(#K16)
*+ARn(#K16)
*CDP(#K16)
*+CDP(#K16)

mmap() and port() qualifiers An instruction that uses the mmap() qualifier to indicate an
access to a memory-mapped register or registers cannot be
placed in parallel with another instruction. The use of the
mmap() qualifier is a form of parallelism already.
Likewise an instruction that uses a port() qualifier to indicate
an access to I/O space cannot be placed in parallel with
another instruction. The use of a port() qualifier is a form of
parallelism already.

Parallelism among A unit, D unit, and P unit Parallelism among the three computational units is allowed
without restriction (see Figure 4−5).
An operation executed within a single computational unit can
be placed in parallel with a second operation executed in one
of the other two computational units.

Parallelism within the P unit Two program-control instructions cannot be placed in parallel.
However, other parallelism among the operators of the P unit
is allowed.

Parallelism within the D unit Certain restrictions apply to using operators of the D unit in
parallel (see Figure 4−5).

Parallelism within the A unit Two A-unit ALU operations or two A-unit swap operations
cannot be performed in parallel. However, other parallelism
among the operators of the A unit is allowed.

4.2.5 Process for Implementing User-Defined Parallelism

This section presents a process that may be used to simplify the process of
using user-defined parallelism to produce optimized assembly language code.
Figure 4−7 is a flow chart outlining this process, and the steps are also de-
scribed in Table 4−4.

4-22
 Using Parallel Execution Features

Figure 4−7. Process for Applying User-Defined Parallelism

Start

Step 1: Write assembly code Step 3

without parallel optimization, Step 4: Use basic rules (Table 4−2)
Assemble code and advanced rules (Table 4−3) to
and test code for functional
correctness to check validity of parallel pairs,
and make changes

Step 2: Identify and place in Assembler Yes

parallel all potential parallel errors on parallel
pairs using the basic rules pairs
?
in Table 4−2
No

Step 5: Test code for functional

correctness

End

Table 4−4. Steps in Process for Applying User-Defined Parallelism

Step Description
1 Write assembly code without the use of user-defined parallelism, and verify the functionality of the code.
Note that in this step, you may take advantage of instructions with built-in parallelism.

2 Identify potential user-defined parallel instruction pairs in your code, and, using the basic rules outlined
in Table 4−2 as guidelines, place instructions in parallel. Start by focusing on heavily used kernels of
the code.

3 Run the optimized code through the assembler to see if the parallel instruction pairs are valid. The as-
sembler will indicate any invalid parallel instruction pairs. If you have invalid pairs, go to step 4; otherwise
go to step 5.

4 Refer to the set of parallelism rules in section 4.2.4 to determine why failing parallel pairs may be invalid.
Make necessary changes and return to step 3.

5 Once all your parallel pairs are valid, make sure your code still functions correctly.

Optimizing Assembly Code 4-23

Using Parallel Execution Features

4.2.6 Parallelism Tips

As you try to optimize your code with user-defined parallelism, you might find
the following tips helpful (these tips use mnemonic instructions):

- Place all load and store instructions in parallel. For example:

MOV *AR2, AC1 ; Load AC1
||MOV BRC0, *AR3 ; Store BRC0
- The A-Unit ALU can handle (un)saturated 16-bit processing in parallel
with the D-Unit ALU, MAC, and shift operators. For example:
; Modify AR1 in A unit, and perform an accumulator
; shift, saturate, and store operation in the D unit.
ADD T0, AR1
||MOV uns(rnd(HI(saturate(AC1 << #1)))), *AR2
- Accumulator shift, saturate, and store operations can be placed in parallel
with D-Unit ALU or MAC operations. For example:
; Shift, saturate, and store AC1 while
; modifying AC2
MOV uns(rnd(HI(saturate(AC1 << #1)))), *AR2
||ADD AC1, AC2
- Control operations can be placed in parallel with DSP operations. For ex-
ample:
; Switch control to a block−repeat loop, and
; Perform the first computation of the loop
||MACMR *AR1+, *AR3, AC0
RPTBLOCAL loop−1
- Instructions with built-in parallelism increase the bandwidth of instructions
paired by user-defined parallelism. For example:
; Place parallel accumulator load operations
; in parallel with an auxiliary register store
; operation
ADD dual(*AR2), AC0, AC1
||MOV AC2, dbl(*AR6)
- You can fill a buffer with a constant value efficiently. For example:
MOV #0, AC0 ; Clear AC0
||RPT #9 ; Switch control to repeat loop
MOV AC0, dbl(*AR1+) ; Store 32−bit constant to buffer
; and increment pointer
- Instructions to be executed conditionally can be placed in parallel with the
if instruction. For example:
XCC check, T0 == #0 ; If T0 contains 0, ...
||MOV #0, AR0 ; ... Load AR0 with 0
check

4-24
 Using Parallel Execution Features

4.2.7 Examples of Parallel Optimization Within CPU Functional Units

This section provides examples to show how to make code more efficient by
using parallelism within the A unit, the P unit, and D unit.

4.2.7.1 A-Unit Example of Parallel Optimization

Example 4−4 shows a simple Host-DSP application in which a host sends a

single command to tell the DSP which set of coefficients to use for a multiply-
and-accumulate (MAC) operation. The DSP calls a COMPUTE function to per-
form the computation and returns the result to the host. The communication
is based on a very simple handshaking, with the host and DSP exchanging
flags (codes). The code in Example 4−4 does not use parallelism.
Example 4−5 shows the code optimized through the use of parallel instruction
pairs.

As mentioned, Example 4−5 shows the optimized code for Example 4−4. In
Example 4−5, the parallel instruction pairs are highlighted. Notice the follow-
ing points:

- The first four instructions (ARn loads) are immediate loads and cannot be
placed in parallel due to constant bus conflicts and total instruction sizes.

- The first parallel pair shows an immediate load of CSR through the bus
called KDB. This load is executed in parallel with the setting of the SXMD
mode bit, which is handled by the A-unit ALU.

- The second parallel pair is a SWAP instruction in parallel with an if instruc-

tion. Despite the fact that the SWAP instruction is executed conditionally,
it is valid to place it in parallel with the if instruction.

- The third parallel pair stores AR4 to memory via the D bus (DB), and stores
a constant (BUSY) to memory via the bus called KDB.

- The fourth parallel pair loads AC1 with a constant that is carried on the bus
called KDB and, in parallel, switches program control to a single-repeat
loop.

- The last parallel pair stores AC1 to AR4 via a cross-unit bus and, in paral-
lel, returns from the COMPUTE function.

Optimizing Assembly Code 4-25

Using Parallel Execution Features

Example 4−4. A-Unit Code With No User-Defined Parallelism

; Variables
.data

COEFF1 .word 0x0123 ; First set of coeffiecients

.word 0x1234
.word 0x2345
.word 0x3456
.word 0x4567

COEFF2 .word 0x7654 ; Second set of coefficients

.word 0x6543
.word 0x5432
.word 0x4321
.word 0x3210

HST_FLAG .set 0x2000 ; Host flag address

HST_DATA .set 0x2001 ; Host data address

CHANGE .set 0x0000 ; “Change coefficients” command from host

READY .set 0x0000 ; “REARY” Flag from Host
BUSY .set 0x1111 ; “BUSY” Flag set by DSP

.global start_a1

.text

start_a1:
MOV #HST_FLAG, AR0 ; AR0 points to Host Flag
MOV #HST_DATA, AR2 ; AR2 points to Host Data
MOV #COEFF1, AR1 ; AR1 points to COEFF1 buffer initially
MOV #COEFF2, AR3 ; AR3 points to COEFF2 buffer initially
MOV #4, CSR ; Set CSR = 4 for repeat in COMPUTE
BSET FRCT ; Set fractional mode bit
BSET SXMD ; Set sign−extension mode bit

Note: The algebraic instructions code example for A-Unit Code With No User-Defined Parallelism is shown in Example B−17
on page B-14.

4-26
 Using Parallel Execution Features

Example 4−4. A-Unit Code With No User-Defined Parallelism (Continued)

LOOP:
MOV *AR0, T0 ; T0 = Host Flag
BCC PROCESS, T0 == #READY ; If Host Flag is “READY”, continue
B LOOP ; process − else poll Host Flag again

PROCESS:
MOV *AR2, T0 ; T0 = Host Data

XCC Check, T0 == #CHANGE

; The choice of either set of
; coefficients is based on the vlaue
; of T0. COMPUTE uses AR3 for
; computation, so we need to
; load AR3 correctly here.

SWAP AR1, AR3 ; Host message was “CHANGE”, so we

; need to swap the two coefficient
; pointers.

Check:
CALL COMPUTE ; Compute subroutine
MOV AR4, *AR2 ; Write result to Host Data
MOV #BUSY, *AR0 ; Set Host Flag to Busy
B LOOP ; Infinite loop continues
END

COMPUTE:
MOV #0, AC1 ; Initialize AC1 to 0
RPT CSR ; CSR has a value of 4
MACM *AR2, *AR3+, AC1 ; This MAC operation is performed
; 5 times
MOV AC1, AR4 ; Result is in AR4
RET

HALT:
B HALT

Note: The algebraic instructions code example for A-Unit Code With No User-Defined Parallelism is shown in Example B−17
on page B-14.

Optimizing Assembly Code 4-27

Using Parallel Execution Features

Example 4−5. A-Unit Code in Example 4−4 Modified to Take Advantage of Parallelism

; Variables
.data

COEFF1 .word 0x0123 ; First set of coeffiecients

.word 0x1234
.word 0x2345
.word 0x3456
.word 0x4567

COEFF2 .word 0x7654 ; Second set of coefficients

.word 0x6543
.word 0x5432
.word 0x4321
.word 0x3210

HST_FLG .set 0x2000 ; Host flag address

HST_DATA .set 0x2001 ; Host data address

CHANGE .set 0x0000 ; “Change coefficients” command from host

READY .set 0x0000 ; “READY” Flag from Host
BUSY .set 0x1111 ; “BUSY” Flag set by DSP

.global start_a2

.text

start_a2:
MOV #HST_FLAG, AR0 ; AR0 points to Host Flag
MOV #HST_DATA, AR2 ; AR2 points to Host Data
MOV #COEFF1, AR1 ; AR1 points to COEFF1 buffer initially
MOV #COEFF2, AR3 ; AR3 points to COEFF2 buffer initially
MOV #4, CSR ; Set CSR = 4 for repeat in COMPUTE
||BSET FRCT ; Set fractional mode bit
BSET SXMD ; Set sign−extension mode bit

LOOP:
MOV *AR0, T0 ; T0 = Host Flag
BCC PROCESS, T0 == #READY ; If Host Flag is “READY”, continue
B LOOP ; process − else poll Host Flag again

Note: The algebraic instructions code example for A-Unit Code in Example 4−4 Modified to Take Advantage of Parallelism is
shown in Example B−18 on page B-16.

4-28
 Using Parallel Execution Features

Example 4−5. A-Unit Code in Example 4−4 Modified to Take Advantage of Parallelism
(Continued)

PROCESS:
MOV *AR2, T0 ; T0 = Host Data

XCC Check, T0 == #CHANGE

; The choice of either set of
; coefficients is based on the vlaue
; of T0. COMPUTE uses AR3 for
; computation, so we need to
; load AR3 correctly here.

||SWAP AR1, AR3 ; Host message was “CHANGE”, so we

; need to swap the two coefficient
; pointers.

Check:
CALL COMPUTE ; Compute subroutine
MOV AR4, *AR2 ; Write result to Host Data
||MOV #BUSY, *AR0 ; Set Host Flag to Busy
B LOOP ; Infinite loop continues

END

COMPUTE:
MOV #0, AC1 ; Initialize AC1 to 0
||RPT CSR ; CSR has a value of 4
MACM *AR2, *AR3+, AC1 ; This MAC operation is performed
; 5 times
MOV AC1, AR4 ; Result is in AR4
||RET

HALT:
B HALT

Note: The algebraic instructions code example for A-Unit Code in Example 4−4 Modified to Take Advantage of Parallelism is
shown in Example B−18 on page B-16.

Optimizing Assembly Code 4-29

Using Parallel Execution Features

4.2.7.2 P-Unit Example of Parallel Optimization

Example 4−6 demonstrates a very simple nested loop and some simple con-
trol operations with the use of P-unit registers. This example shows the unopti-
mized code, and Example 4−7 shows the code optimized through the use of
the P-unit parallel instruction pairs (parallel instruction pairs are highlighted).

Notice the following about Example 4−7:

- The first three register loads are immediate loads using the same constant
bus and, therefore, cannot be placed in parallel.

- The fourth register load (loading BRC0) can be placed in parallel with the
next load (loading BRC1), which does not use the constant bus, but the
data bus DB to perform the load.

- The next instruction, which is a control instruction (blockrepeat) is placed

in parallel with the load of AC2. There are no conflicts and this pair is valid.
The loading of AC2 is not part of the blockrepeat structure.

- The final parallel pair is a control instruction (localrepeat) and a load of

AR1. The loading of AR1 is not part of the localrepeat structure but is part
of the outer blockrepeat structure.

4-30
 Using Parallel Execution Features

Example 4−6. P-Unit Code With No User-Defined Parallelism

.CPL_off ; Tell assembler that CPL bit is 0

; (Direct addressing is done with DP)

; Variables
.data

var1 .word 0x0004

var2 .word 0x0000

.global start_p1

.text

start_p1:
AMOV #var1, XDP
MOV #var2, AR3

MOV #0007h, BRC0 ; BRC0 loaded using KPB

MOV *AR3, BRC1 ; BRC1 loaded using DB

MOV #0006h, AC2

RPTB Loop1−1
MOV AC2, AC1
MOV #8000h, AR1
RPTBLOCAL Loop2−1
SUB #1, AC1
MOV AC1, *AR1+
Loop2:
ADD #1, AC2
Loop1:

MOV BRC0, @AC0L ; AC0L loaded using EB

MOV BRC1, @AC1L ; AC1L loaded using EB

BCC start_p1, AC0 >= #0

BCC start_p1, AC1 >= #0
end_p1

Note: The algebraic instructions code example for P-Unit Code With No User-Defined Parallelism is shown in Example B−19
on page B-18.

Optimizing Assembly Code 4-31

Using Parallel Execution Features

Example 4−7. P-Unit Code in Example 4−6 Modified to Take Advantage of Parallelism

.CPL_off ; Tell assembler that CPL bit is 0

; (Direct addressing is done with DP)

; Variables
.data

var1 .word 0x0004

var2 .word 0x0000

.global start_p2

.text

start_p2:
AMOV #var1, XDP
MOV #var2, AR3

MOV #0007h, BRC0 ; BRC0 loaded using KPB

||MOV *AR3, BRC1 ; BRC1 loaded using DB

MOV #0006h, AC2

||RPTB Loop1−1
MOV AC2, AC1
MOV #8000h, AR1
||RPTBLOCAL Loop2−1
SUB #1, AC1
MOV AC1, *AR1+
Loop2:
ADD #1, AC2
Loop1:

MOV BRC0, @AC0L ; AC0L loaded using EB

MOV BRC1, @AC1L ; AC1L loaded using EB

BCC start_p1, AC0 >= #0

BCC start_p1, AC1 >= #0
end_p2

Note: The algebraic instructions code example for P-Unit Code in Example 4−6 Modified to Take Advantage of Parallelism is
shown in Example B−20 on page B-19.

4-32
 Using Parallel Execution Features

4.2.7.3 D-Unit Example of Parallel Optimization

Example 4−8 demonstrates a very simple load, multiply, and store function
with the use of D-unit registers. Example 4−9 shows this code modified to take
advantage of user-defined parallelism (parallel instruction pairs are high-
lighted).

The following information determined the optimizations made in Example 4−9:

- As in the P-unit example on page 4-32, we cannot place the immediate

register loads in parallel due to constant-bus conflicts.

- The instructions that load AC0 and AC2 have been placed in parallel be-
cause they are not both immediate loads and as such, there are no
constant-bus conflicts.

- It is not possible to place the two single-MAC instructions in parallel since

the same operator is required for both and as such a conflict arises. How-
ever, placing the second MAC instruction in parallel with the SWAP in-
struction is valid.

- The two 16-bit store operations at the end of the code are placed in parallel
because there are two 16-bit write buses available (EB and FB).

Optimizing Assembly Code 4-33

Using Parallel Execution Features

Example 4−8. D-Unit Code With No User-Defined Parallelism

; Variables
.data

var1 .word 0x8000

var2 .word 0x0004

.global start_d1

.text

start_d1:
MOV #var1, AR3
MOV #var2, AR4

MOV #0004h, AC0 ; AC0 loaded using KDB

MOV *AR3, AC2 ; AC2 loaded using DB

MOV #5A5Ah, T0 ; T0 loaded with constant, 0x5A5A

MAC AC0, T0, AC2 ; MAC

MAC AC2, T0, AC1 ; MAC
SWAP AC0, AC2 ; SWAP

MOV HI(AC1), *AR3 ; Store result in AC1

MOV HI(AC0), *AR4 ; Store result in AC0
end_d1

Note: The algebraic instructions code example for D-Unit Code With No User-Defined Parallelism is shown in Example B−21
on page B-20.

4-34
 Using Parallel Execution Features

Example 4−9. D-Unit Code in Example 4−8 Modified to Take Advantage of Parallelism

; Variables
.data

var1 .word 0x8000

var2 .word 0x0004

.global start_d2

.text

start_d2:
MOV #var1, AR3
MOV #var2, AR4

MOV #0004h, AC0 ; AC0 loaded using KDB

||MOV *AR3, AC2 ; AC2 loaded using DB

MOV #5A5Ah, T0 ; T0 loaded with constant, 0x5A5A

MAC AC0, T0, AC2 ; MAC

MAC AC2, T0, AC1 ; MAC
||SWAP AC0, AC2 ; SWAP

MOV HI(AC1), *AR3 ; Store result in AC1

||MOV HI(AC0), *AR4 ; Store result in AC0
end_d2

Note: The algebraic instructions code example for D-Unit Code in Example 4−8 Modified to Take Advantage of Parallelism is
shown in Example B−22 on page B-21.

4.2.8 Example of Parallel Optimization Across the A-Unit, P-Unit, and D-Unit
Example 4−10 shows unoptimized code for an FIR (finite impulse response)
filter. Example 4−11 is the result of applying user-defined parallelism to the
same code. It is important to notice that the order of instructions has been al-
tered in a number of cases to allow certain instruction pairs to be placed in par-
allel. The use of parallelism in this case has saved about 50% of the cycles
outside the inner loop.

Optimizing Assembly Code 4-35

Using Parallel Execution Features

Example 4−10. Code That Uses Multiple CPU Units But No User-Defined Parallelism

.CPL_ON ; Tell assembler that CPL bit is 1

; (SP direct addressing like *SP(0) is enabled)

; Register usage
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

.asg AR0, X_ptr ; AR0 is pointer to input buffer − X_ptr

.asg AR1, H_ptr ; AR1 is pointer to coefficients − H_ptr
.asg AR2, R_ptr ; AR2 is pointer to result buffer − R_ptr
.asg AR3, DB_ptr ; AR3 is pointer to delay buffer − DB_ptr

FRAME_SZ .set 2

.global _fir

.text

; *****************************************************************

_fir

; Create local frame for temp values

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AADD #−FRAME_SZ, SP

; Turn on fractional mode

; Turn on sign−extension mode
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BSET FRCT ; Set fractional mode bit

BSET SXMD ; Set sign extension mode bit

; Set outer loop count by subtracting 1 from nx and storing into

; block−repeat counter
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

SUB #1, T1, AC1 ; AC1 = number of samples (nx) − 1

MOV AC1, *SP(0) ; Top of stack = nx − 1
MOV *SP(0), BRC0 ; BRC0 −= nx − 1 (outer loop counter)

Note: The algebraic instructions code example for Code That Uses Multiple CPU Units But No User-Defined Parallelism is
shown in Example B−23 on page B-22.

4-36
 Using Parallel Execution Features

Example 4−10. Code That Uses Multiple CPU Units But No User-Defined Parallelism
(Continued)

; Store length of coefficient vector/delay buffer in BK register

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BSET AR1LC ; Enable AR1 circular configuration

MOV *(#0011h), BSA01 ;Set buffer (filter) start address
; AR1 used as filter pointer

BSET AR3LC ; Enable AR3 circular configuration]

MOV DB_ptr, *SP(1) ; Save pointer to delay buffer pointer
MOV *DB_ptr, AC1 ; AC1 = delay buffer pointer
MOV AC1, DB_ptr ; AR3 (DB_ptr) = delay buffer pointer
MOV *(#0013h), BSA23 ;Set buffer (delay buffer) start address
; AR3 used as filter pointer
MOV T0, *SP(0) ; Save filter length, nh − used as buffer
; size
MOV *SP(0), BK03 ; Set circular buffer size − size passed
; in T0
SUB #3, T0, AC1 ; AC1 = nh − 3
MOV AC1, *SP(0)
MOV *SP(0), CSR ; Set inner loop count to nh − 3
MOV #0, H_ptr ; Initialize index of filter to 0
MOV #0, DB_ptr ; Initialize index of delay buffer to 0

; Begin outer loop on nx samples

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

RPTB Loop1−1

; Move next input sample into delay buffer

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MOV *X_ptr+, *DB_ptr

; Sum h * x for next y value

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MPYM *H_ptr+, *DB_ptr+, AC0

RPT CSR
MACM *H_ptr+, *DB_ptr+, AC0

MACMR *H_ptr+, *DB_ptr, AC0 ; Round result

Note: The algebraic instructions code example for Code That Uses Multiple CPU Units But No User-Defined Parallelism is
shown in Example B−23 on page B-22.

Optimizing Assembly Code 4-37

Using Parallel Execution Features

Example 4−10. Code That Uses Multiple CPU Units But No User-Defined Parallelism
(Continued)

; Store result
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MOV HI(AC0), *R_ptr+

Loop1:

; Clear FRCT bit to restore normal C operating environment

; Return overflow condition of AC0 (shown in ACOV0) in T0
; Restore stack to previous value, FRAME, etc.
; Update current index of delay buffer pointer
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

END_FUNCTION:

MOV *SP(1), AR0 ; AR0 = pointer to delay buffer pointer

AADD #FRAME_SZ, SP ; Remove local stack frame
MOV DB_ptr, *AR0 ; Update delay buffer pointer with current
; index

BCLR FRCT ; Clear fractional mode bit

MOV #0, T0 ; Make T0 = 0 for no overflow (return value)

XCC Label, overflow(AC0)
MOV #1, T0 ; Make T0 = 1 for overflow (return value)
Label:

RET
; **********************************************************************

Note: The algebraic instructions code example for Code That Uses Multiple CPU Units But No User-Defined Parallelism is
shown in Example B−23 on page B-22.

In Example 4−11, parallel pairs that were successful are shown in bold type;
potential parallel pairs that failed are shown in italic type. The first failed due
to a constant bus conflict, and the second failed due to the fact that the com-
bined size is greater than 6 bytes. The third pair failed for the same reason,
as well as being an invalid soft-dual encoding instruction. This last pair in italics
failed because neither instruction has a parallel enable bit. Some of the load/
store operations that are not in parallel were made parallel in the first pass opti-
mization process; however, the parallelism failed due to bus conflicts and had
to be removed.

4-38
 Using Parallel Execution Features

Note:
Example 4−11 shows optimization only with the use of the parallelism fea-
tures. Further optimization of this FIR function is possible by employing other
optimizations.

Example 4−11. Code in Example 4−10 Modified to Take Advantage of Parallelism

.CPL_ON ; Tell assembler that CPL bit is 1

; (SP direct addressing like *SP(0) is enabled)

; Register usage
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

.asg AR0, X_ptr ; AR0 is pointer to input buffer − X_ptr

.asg AR1, H_ptr ; AR1 is pointer to coefficients − H_ptr
.asg AR2, R_ptr ; AR2 is pointer to result buffer − R_ptr
.asg AR3, DB_ptr ; AR3 is pointer to delay buffer − DB_ptr

FRAME_SZ .set 2

.global _fir

.text

; *****************************************************************

_fir

; Create local frame for temp values

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AADD #−FRAME_SZ, SP ; (Attempt to put this in parallel with

; the following AC1 modification failed)

; Set outer loop count by subtracting 1 from nx and storing into

; block−repeat counter
; Turn on fractional mode
; Turn on sign−extension mode
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

SUB #1, T1, AC1 ; AC1 = number of samples (nx) − 1

BSET FRCT ; Set fractional mode bit

||MOV AC1, *SP(0) ; Top of stack = nx − 1

MOV *SP(0), BRC0 ; BRC0 = nx − 1 (outer loop counter)

||BSET SXMD ; Set sign−extension mode bit

Note: The algebraic instructions code example for Code in Example 4−10 Modified to Take Advantage of Parallelism is shown
in Example B−24 on page B-25.

Optimizing Assembly Code 4-39

Using Parallel Execution Features

Example 4−11. Code in Example 4−10 Modified to Take Advantage of Parallelism

(Continued)

; Store length of coefficient vector/delay buffer in BK register

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BSET AR1LC ; Enable AR1 circular configuration

MOV *(#0011h), BSA01 ;Set buffer (filter) start address
; AR1 used as filter pointer

BSET AR3LC ; Enable AR3 circular configuration

||MOV DB_ptr, *SP(1) ; Save pointer to delay buffer pointer

MOV *DB_ptr, AC1 ; AC1 = delay buffer pointer

MOV AC1, DB_ptr ; AR3 (DB_ptr) = delay buffer pointer
||MOV T0, *SP(0) ; Save filter length, nh − used as buffer
; size
MOV *(#0013h), BSA23 ;Set buffer (delay buffer) start address
; AR3 used as filter pointer

MOV *SP(0), BK03 ; Set circular buffer size − size passed

; in T0

SUB #3, T0, AC1 ; AC1 = nh − 3

MOV AC1, *SP(0)

MOV *SP(0), CSR ; Set inner loop count to nh − 3

||MOV #0, H_ptr ; Initialize index of filter to 0

MOV #0, DB_ptr ; Initialize index of delay buffer to 0

; (in parallel with RPTB below)

; Begin outer loop on nx samples

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

||RPTB Loop1−1

; Move next input sample into delay buffer

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MOV *X_ptr+, *DB_ptr

Note: The algebraic instructions code example for Code in Example 4−10 Modified to Take Advantage of Parallelism is shown
in Example B−24 on page B-25.

4-40
 Using Parallel Execution Features

Example 4−11. Code in Example 4−10 Modified to Take Advantage of Parallelism

(Continued)

; Sum h * x for next y value

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MPYM *H_ptr+, *DB_ptr+, AC0

||RPT CSR

MACM *H_ptr+, *DB_ptr+, AC0

MACMR *H_ptr+, *DB_ptr, AC0 ; Round result

; Store result
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

MOV HI(AC0), *R_ptr+

Loop1:

; Clear FRCT bit to restore normal C operating environment

; Return overflow condition of AC0 (shown in ACOV0) in T0
; Restore stack to previous value, FRAME, etc.
; Update current index of delay buffer pointer
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

END_FUNCTION:

MOV *SP(1), AR0 ; AR0 = pointer to delay buffer pointer

||AADD #FRAME_SZ, SP ; Remove local stack frame

MOV DB_ptr, *AR0 ; Update delay buffer pointer with current

; index
||BCLR FRCT ; Clear fractional mode bit

MOV #0, T0 ; Make T0 = 0 for no overflow (return value)

||XCC Label, overflow(AC0)
MOV #1, T0 ; Make T0 = 1 for overflow (return value)
Label:

||RET
; **********************************************************************

Note: The algebraic instructions code example for Code in Example 4−10 Modified to Take Advantage of Parallelism is shown
in Example B−24 on page B-25.

Optimizing Assembly Code 4-41

Implementing Efficient Loops

4.3 Implementing Efficient Loops

There are four common methods to implement instruction looping in the C55x
DSP:
- Single repeat: repeat(CSR/k8/k16),[CSR += TA/k4]
- Local block repeat: localrepeat{}
- Block repeat: blockrepeat{}
- Branch on auxiliary register not zero: if (ARn_mod != 0) goto #loop_start

The selection of the looping method to use depends basically on the number
of instructions that need to be repeated and in the way you need to control the
loop counter parameter. The first three methods in the preceding list offer zero-
overhead looping and the fourth one offers a 5-cycle loop overhead.

Overall, the most efficient looping mechanisms are the repeat() and the
localrepeat{} mechanisms. The repeat() mechanism provides a way to repeat
a single instruction or a parallel pair of instructions in an interruptible way. re-
peat(CSR), in particular, allows you to compute the loop counter at runtime.
Refer to section 4.3.2, Efficient Use of repeat(CSR) Looping.

Note:
If you are migrating code from a TMS320C54x DSP, be aware that a single-
repeat instruction is interruptible on a TMS320C55x DSP. On a
TMS320C54x DSP, a single-repeat instruction cannot be interrupted.

The localrepeat{} mechanism provides a way to repeat a block from the in-
struction buffer queue. Reusing code that has already been fetched and
placed in the queue brings the following advantages:

- Fewer program-memory access pipeline conflicts

- Overall lower power consumption

- No repetition of wait-state and access penalties when executing loop code

from external RAM

4.3.1 Nesting of Loops

You can create up to two levels of block-repeat loops without any cycle penalty.
You can have one block-repeat loop nested inside another, creating an inner
(level 1) loop and an outer (level0) loop. In addition, you can put any number
of single-repeat loops inside each block-repeat loop. Example 4−12 shows a
multi-level loop structure with two block-repeat loops and two single-repeat
loops. (The examples in this section use instructions from the algebraic in-
struction set, but the concepts apply equally for the mnemonic instruction set.)

4-42
 Implementing Efficient Loops

Example 4−12 shows one block-repeat loop nested inside another block-re-
peat loop. If you need more levels of multiple-instruction loops, use branch on
auxiliary register not zero constructs to create the remaining outer loops. In
Example 4−13 (page 4-44), a branch on auxiliary register not zero construct
(see the last instruction in the example) forms the outermost loop of a Fast
Fourier Transform algorithm. Inside that loop are two localrepeat{} loops. No-
tice that if you want the outermost loop to execute n times, you must initialize
AR0 to (n − 1) outside the loop.

Example 4−12. Nested Loops

MOV #(n0−1), BRC0

MOV #(n1−1), BRC1
; ...
RPTBLOCAL Loop1−1 ; Level 0 looping (could instead be blockrepeat):
; Loops n0 times
; ...
RPT #(n2−1)
; ...
RPTBLOCAL Loop2−1 ; Level 1 looping (could instead be blockrepeat):
; Loops n1 times
; ...
RPT #(n3−1)
; ...
Loop2:
; ...
Loop1:

Note: The algebraic instructions code example for Nested Loops is shown in Example B−25 on page B-28.

Optimizing Assembly Code 4-43

Implementing Efficient Loops

Example 4−13. Branch-On-Auxiliary-Register-Not-Zero Construct

(Shown in Complex FFT Loop Code)

_cfft:

radix_2_stages:
; ...

outer_loop:
; ...
MOV T1, BRC0
; ...
MOV T1, BRC1
; ...
SFTS AR4, #1 ; outer loop counter
||BCC no_scale, AR5 == #0 ; determine if scaling required
; ...

no_scale:

RPTBLOCAL Loop1−1

MOV dbl(*AR3), AC0 ; Load ar,ai

SUB AC0, dual(*AR2), AC2 ; tr = ar − br

; ti = ai − bi

RPTBLOCAL Loop2−1

ADD dual(*AR2), AC0, AC1 ; ar’ = ar + br

; ai’ = ai + bi
||MOV AC2, dbl(*AR6) ; Store tr, ti

MPY *AR6, *CDP+, AC2 ; c*tr

::MPY *AR7, *CDP+, AC3 ; c*ti

MOV AC1, dbl(*AR2+) ; Store ar, ai

||MOV dbl(*AR3(T0)), AC0 ; * load ar,ai

MASR *AR6, *CDP−, AC3 ; bi’ = c*ti − s*tr

::MACR *AR7, *CDP−, AC2 ; br’ = c*tr + s*ti

MOV pair(HI(AC2)), dbl(*AR3+) ; Store br’, bi’

||SUB AC0, dual(*AR2), AC2 ; * tr = ar − br
; * ti = ai − bi

Note: The algebraic instructions code example for Branch-On-Auxiliary-Register-Not-Zero Construct (Shown in Complex FFT
Loop Code) is shown in Example B−26 on page B-28.

4-44
 Implementing Efficient Loops

Example 4−13. Branch-On-Auxiliary-Register-Not-Zero Construct

(Shown in Complex FFT Loop Code) (Continued)

Loop2:
ADD dual(*AR2), AC0, AC1 ; ar’ = ar + br
; ai’ = ai + bi
||MOV AC2, dbl(*AR6) ; Store tr, ti

MPY *AR6, *CDP+, AC2 ; c*tr

::MPY *AR7, *CDP+, AC3 ; c*ti

MOV AC1, dbl(*(AR2+T1)) ; Store ar, ai

MASR *AR6, *CDP+, AC3 ; bi’ = c*ti − s*tr

::MACR *AR7, *CDP+, AC3 ; br’ = c*tr + s*ti

MOV pair(HI(AC2)), dbl(*(AR3+T1)) ; Store br’, bi’

Loop1:

SFTS AR3, #1
||MOV #0, CDP ; rewind coefficient pointer
SFTS T3, #1
||BCC outer_loop, AR4 != #0

Note: The algebraic instructions code example for Branch-On-Auxiliary-Register-Not-Zero Construct (Shown in Complex FFT
Loop Code) is shown in Example B−26 on page B-28.

To achieve an efficient nesting of loops, apply the following guidelines:

- Use a single-repeat instruction for the innermost loop if the loop contains
only a single instruction (or a pair of instructions that have been placed in
parallel).

- Use a local block-repeat instruction (localrepeat in the algebraic syntax)

for a loop containing more than a single instruction or instruction pair—
provided the loop contains no more than 56 bytes of code. If there are
more than 56 bytes but you would still like to use the local block-repeat in-
struction, consider the following possibilities:

J Split the existing loop into two smaller loops.

J Reduce the number of bytes in the loop. For example, you can reduce
the number of instructions that use embedded constants.

- Use a standard block-repeat instruction (blockrepeat in the algebraic syn-

tax) in cases where a local block-repeat instruction cannot be used. The
standard block-repeat mechanism always refetches the loop code from
memory.

Optimizing Assembly Code 4-45

Implementing Efficient Loops

- When you nest a block-repeat loop inside another block-repeat loop, ini-
tialize the block-repeat counters (BRC0 and BRC1) in the code outside of
both loops. This technique is shown in Example 4−12 (page 4-43).
Neither counter needs to be re-initialized inside its loop; placing such init-
ializations inside the loops only adds extra cycles to the loops. The CPU
uses BRC0 for the outer (level 0) loop and BRC1 for the inner (level 1) loop.
BRC1 has a shadow register, BRS1, that preserves the initial value of
BRC1. Each time the level 1 loop must begin again, the CPU automatically
re-initializes BRC1 from BRS1.

4.3.2 Efficient Use of repeat(CSR) Looping

The single-repeat instruction syntaxes allow you to specify the repeat count
as a constant (embedded in the repeat instruction) or as the content of the
computed single-repeat register (CSR). When CSR is used, it is not decrem-
ented during each iteration of the single-repeat loop. Before the first execution
of the instruction or instruction pair to be repeated, the content of CSR is cop-
ied into the single-repeat counter (RPTC). RPTC holds the active loop count
and is decremented during each iteration. Therefore, CSR needs to be initial-
ized only once. Initializing CSR outside the outer loop, rather than during every
iteration of the outer loop, saves cycles. There are advantages to using CSR
for the repeat count:

- The repeat count can be dynamically computed during runtime and stored
to CSR. For example, CSR can be used when the number of times an in-
struction must be repeated depends on the iteration number of a higher
loop structure.

- Using CSR saves outer loop cycles when the single-repeat loop is an inner
loop.

- An optional syntax extension enables the repeat instruction to modify the

CSR after copying the content of CSR to RPTC. When the single-repeat
loop is repeated in an outer loop, CSR contains a new count.

Example 4−14 (page 4-47) uses CSR for a single-repeat loop that is nested
inside a block-repeat loop. In the example, CSR is assigned the name
inner_cnt.

4-46
 Implementing Efficient Loops

Example 4−14. Use of CSR (Shown in Real Block FIR Loop Code)

; ...
.asg CSR, inner_cnt ; inner loop count
.asg BRC0, outer_cnt ; outer loop count
; ...
.asg AR0, x_ptr ; linear pointer
.asg AR1, db_ptr1 ; circular pointer
.asg AR2, r_ptr ; linear pointer
.asg AR3, db_ptr2 ; circular pointer
.asg CDP, h_ptr ; circular pointer

; ...

_fir2:

; ...
AMAR *db_ptr2− ; index of 2nd oldest db entry
;
; Setup loop counts
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
||SFTS T0, #−1 ; T0 = nx/2

SUB #1, T0 ; T0 = (nx/2 − 1)

MOV T0, outer_cnt ; outer loop executes nx/2 times
SUB #3, T1, T0 ; T0 = nh−3
MOV T0, inner_cnt ; inner loop executes nh−2 times
ADD #1, T1 ; T1 = nh+1, adjustment for db_ptr1, bd_ptr2
;
; Start of outer loop
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
||RPTBLOCAL OuterLoop−1 ; start the outer loop

MOV *x_ptr+, *db_ptr1 ; get 1st new input value

MOV *x_ptr+, *db_ptr2 ; get 2nd new input value

; 1st interation
MPY *db_ptr1+, *h_ptr+, AC0 ; part 1 of dual−MPY
::MPY *db_ptr2+, *h_ptr+, AC1 ; part 2 of dual−MPY

; inner loop
||RPT inner_cnt
MAC *db_ptr1+, *h_ptr+, AC0 ; part 1 of dual−MAC
::MAC *db_ptr2+, *h_ptr+, AC1 ; part 2 of dual−MAC

Note: This example shows portions of the file fir2.asm in the TI C55x DSPLIB (introduced in Chapter 8).

Optimizing Assembly Code 4-47

Implementing Efficient Loops

Example 4−14. Use of CSR (Shown in Real Block FIR Loop Code) (Continued)

; last iteration has different pointer adjustment and rounding

MACR *(db_ptr1−T1), *h_ptr+, AC0 ; part 1 of dual−MAC
::MACR *(db_ptr2−T1), *h_ptr+, AC1 ; part 2 of dual−MAC

; store result to memory

MOV HI(AC0), *r_ptr+ ; store 1st Q15 result to memory
MOV HI(AC1), *r_ptr+ ; store 2nd Q15 result to memory

OuterLoop: ; end of outer loop

; ...

Note: This example shows portions of the file fir2.asm in the TI C55x DSPLIB (introduced in Chapter 8).

4.3.3 Avoiding Pipeline Delays When Accessing Loop-Control Registers

Accesses to loop-control registers like CSR, BRC0, and BRC1 can cause de-
lays in the instruction pipeline if nearby instructions make competing ac-
cesses. For recommendations on avoiding this type of pipeline delay, see the
“Loop control” section of Table 4−6 (page 4-54) .

4-48
 Minimizing Pipeline and IBQ Delays

4.4 Minimizing Pipeline and IBQ Delays

The C55x instruction pipeline is a protected pipeline that has two, decoupled
segments:

- The first segment, referred to as the fetch pipeline, fetches 32-bit instruc-
tion packets from memory into the instruction buffer queue (IBQ), and then
feeds the second pipeline segment with 48-bit instruction packets. The
fetch pipeline is illustrated in Figure 4−8.

- The second segment, referred to as the execution pipeline, decodes in-

structions and performs data accesses and computations. The execution
pipeline is illustrated and described in Figure 4−9. Table 4−5 provides ex-
amples to help you understand the activity in the key phases of the execu-
tion pipeline.

Figure 4−8. First Segment of the Pipeline (Fetch Pipeline)

Time

Prefetch 1 Prefetch 2 Fetch Predecode

(PF1) (PF2) (F) (PD)

Pipeline
Phase Description
PF1 Present the program fetch address to memory.

PF2 Wait for memory to respond.

F Fetch an instruction packet from memory and place it in the IBQ.

PD Predecode instructions in the IBQ (identify where instructions be-

gin and end; identify parallel instructions).

Optimizing Assembly Code 4-49

Minimizing Pipeline and IBQ Delays

ÉÉÉÉ
Figure 4−9. Second Segment of the Pipeline (Execution Pipeline)

ÉÉÉÉ
Time

ÉÉ ÉÉÉÉ
Decode Address Access 1 Access 2 Read Execute Write Write+
(D) (AD) (AC1) (AC2) (R) (X) (W) (W+)

Note:
ÉÉ
Pipeline
Only for memory write operations.

Phase Description
Decode (D) - Read six bytes from the instruction buffer queue.
- Decode an instruction pair or a single instruction.
- Dispatch instructions to the appropriate CPU functional units.
- Read STx_55 bits associated with data address generation:
ST1_55(CPL) ST2_55(ARnLC)
ST2_55(ARMS) ST2_55(CDPLC)
- Read STx bits related to address generation (ARxLC, CPL,
ARMS) for address generation in the AD phase.
Address - Read/modify registers when involved in data address
(AD) generation. For example:
ARx and Tx in *ARx+(T0)
BK03 if AR2LC=1
SP during pushes and pops
SSP, same as for SP if in 32-bit stack mode
- Read A-unit register in the case of a R/W (in AD-phase)
conflict.
- Perform operations that use the A-unit ALU. For example:
Arithmetic using AADD instruction
Swapping A-unit registers with a SWAP instruction
Writing constants to A-unit registers (BKxx, BSAxx, BRCx, CSR,
etc.)
- Decrement ARx for the conditional branch instruction that
branches on ARx not zero.
- (Exception) Evaluate the condition of the XCC instruction
(execute(AD-unit) attribute in the algebraic syntax).
Access 1 Refer to section 4.4.3.1 for description.
(AC1)
Access 2 Refer to section 4.4.3.1 for description.
(AC2)

4-50
 Minimizing Pipeline and IBQ Delays

Figure 4−9. Second Segment of the Pipeline (Execution Pipeline) (Continued)

Pipeline
Phase Description
Read (R) - Read data from memory, I/O space, and MMR-addressed
registers.
- Read A-unit registers when executing specific D-unit instructions
that “prefetch” A-unit registers (listed in Appendix A) in the R
phase rather than reading them in the X phase.
- Evaluate the conditions of conditional instructions. Most but not
all condition evaluation is performed in the R phase. Exceptions
are marked with (Exception) in this table.
Execute (X) - Read/modify registers that are not MMR-addressed.
- Read/modify individual register bits.
- Set conditions.
- (Exception) Evaluate the condition of the XCCPART instruction
(execute(D-unit) attribute in the algebraic syntax), unless the
instruction is conditioning a write to memory (in this case, the
condition is evaluated in the R phase).
- (Exception) Evaluate the condition of the RPTCC instruction.
Write (W) - Write data to MMR-addressed registers or to I/O space
(peripheral registers).
- Write data to memory. From the perspective of the CPU, the write
operation is finished in this pipeline phase.
Write+ (W+) Write data to memory. From the perspective of the memory, the
write operation is finished in this pipeline phase.

Table 4−5. Pipeline Activity Examples

Example Pipeline Explanation

AMOV #k23, XARx XARx is initialized with a constant in the AD phase.

MOV #k, ARx ARx is not MMR-addressed. ARx is initialized with a

constant in the X phase.

MOV #k, mmap(ARx) ARx is MMR-addressed. ARx is initialized with a

constant in the W phase.

AADD #k, ARx With this special instruction, ARx is initialized with a
constant in the AD phase.

MOV #k, *ARx+ The memory write happens in the W+ phase. See
special write pending and memory bypass cases in
section 4.4.3.

Optimizing Assembly Code 4-51

Minimizing Pipeline and IBQ Delays

Table 4−5. Pipeline Activity Examples (Continued)

Example Pipeline Explanation
MOV *ARx+, AC0 ARx is read and updated in the AD phase. AC0 is
loaded in the X phase.

ADD #k, ARx ARx is read at the beginning of the X phase and is
modified at the end of the X phase.

ADD ACy, ACx ACx and ACy read and write activity occurs in the X
phase.

MOV mmap(ARx), ACx ARx is MMR-addressed and so is read in the R

phase. ACx is modified in the X phase.

MOV ARx, ACx ARx is not MMR-addressed and so is read in the X

phase. ACx is modified in the X phase.

BSET CPL The CPL bit is set in the X phase.

PUSH, POP, RET or SP is read and modified in the AD phase. SSP is also
AADD #K8, SP affected if the 32-bit stack mode is selected.

XCCPART overflow(ACx) The condition is evaluated in the X phase.

|| MOV *AR1+, AC1 Note: AR1 is incremented regardless of whether the
condition is true.

XCCPART overflow(ACx) The condition is evaluated in the R phase because it

|| MOV AC1, *AR1+ conditions a write to memory.
Note: AR1 is incremented regardless of whether the
condition is true.

XCC overflow(ACx) The condition is evaluated in the AD phase.

|| MOV *AR1+, AC1 Note: AR1 is incremented only if the condition is true.

Multiple instructions are executed simultaneously in the pipeline, and different

instructions perform modifications to memory, I/O space, and register values
during different phases of the pipeline. In an unprotected pipeline, this could
lead to data-access errors—reads and writes at the same location happening
out of the intended order. The pipeline-protection unit of the C55x DSP inserts
extra cycles to prevent these errors. If an instruction (say, instruction 3) must
access a location but a previous instruction (say, instruction 1) is not done with
the location, instruction 3 is halted in the pipeline until instruction 1 is done. To
minimize delays, you can take steps to prevent many of these pipeline-protec-
tion cycles.

The instruction set reference guides (see Related Documentation From Texas
Instruments in the preface) show how many cycles an instruction takes to exe-
cute when the pipeline is full and experiencing no delays. Pipeline-protection

4-52
 Minimizing Pipeline and IBQ Delays

cycles add to that best-case execution time. As it will be shown, most cases
of pipeline conflict can be solved with instruction rescheduling.

This section provides examples to help you to better understand the impact
of the pipeline structure on the way your code performs. It also provides you
with recommendations for coding style and instruction usage to minimize con-
flicts or pipeline stalls. This section does not cover all of the pipeline potential
conflicts, but some of the most common pipeline delays and IBQ delays found
when writing C55x code.

4.4.1 Process to Resolve Pipeline and IBQ Conflicts

A pipeline conflict occurs when one instruction attempts to access a resource
before a previous instruction is done with that resource. The pipeline-protec-
tion unit adds extra cycles to delay the later instruction.The following process
is recommended for resolving pipeline conflicts that are causing delays. Try
to focus your pipeline optimization effort on your key, inner code kernels first,
to achieve a greater payback.

Step 1: Make your code functional.

Write your code first, without pipeline optimization in mind. In the
C55x DSP, the pipeline is protected. Code is executed in the order
in which it is written, and stalls are automatically inserted by the hard-
ware to prevent incorrect operation. This makes programming the
DSP easier and makes the code easier to debug than an open-pipe-
line device, in which the sequence of the code might not be the se-
quence of operation.

Step 2: Determine where the pipeline conflicts exist.

If you are using the C55x simulator, take advantage of its pipeline-
conflict detection capabilities. Watch the clock variable when step-
ping through your code in the C55x simulator to view the intervention
of the pipeline protection unit. If the clock increments by more than
1, there might be a pipeline or memory conflict in the instruction you
just single stepped.
The C55x emulator/debugger does not support the clock variable,
and setting breakpoints before and after may not give you accurate
results for a single instruction due to the initial pipeline fill and the fi-
nal pipeline flush during single stepping. In this case, you should try
to benchmark small pieces of looped code.
Also, C55x simulator version 2.1 or higher offers a pipeline visibility
plug-in to help you identify and resolve pipeline conflicts.

Optimizing Assembly Code 4-53

Minimizing Pipeline and IBQ Delays

Step 3: Apply the pipeline optimization coding recommendations sum-

marized in Table 4−6.
After step 2 or if you are not using the simulator but the emulator, you
should try to apply the recommendations directly.
Tip: When suspecting a pipeline conflict between two instructions,
try to add NOP instructions in between. If the entire code cycle count
does not increase by adding NOPs, then you can try to rearrange
your code to replace those NOPs with useful instructions.

Software solutions to apply for pipeline and memory conflicts include:

- Reschedule instructions.

- Reduce memory accesses by using CPU registers to hold data.

- Reduce memory accesses by using a local repeat instruction, an instruc-

tion that enables the CPU to repeatedly execute a block of code from the
instruction buffer queue.

- Relocate variables and data arrays in memory, or consider temporarily

copying arrays to other nonconflicting memory banks at run time.

4.4.2 Recommendations for Preventing Pipeline Delays

Table 4−6 lists recommendations for avoiding common causes for pipeline de-
lays. The rightmost column of the table directs you to the section that contains
the details behind each recommendation. (The examples in this section use
instructions from the algebraic instruction set, but the concepts apply equally
for the mnemonic instruction set.)

Table 4−6. Recommendations for Preventing Pipeline Delays

Recommendation
Category Recommendation See ...

General - In the case of a conflict, the front runner wins. Section 4.4.2.1, page 4-56

Register access - Avoid consecutive accesses to the same Section 4.4.2.2, page 4-56
register.

4-54
 Minimizing Pipeline and IBQ Delays

Table 4−6. Recommendations for Preventing Pipeline Delays (Continued)

Recommendation
Category Recommendation See ...

- Use MAR type of instructions, when pos- Section 4.4.2.3, page 4-60
sible, to modify ARx and Tx registers, but
avoid read/write register sequences, and
pay attention to instruction size.

Pipeline-protection - Pay attention to pipeline-protection granular- Section 4.4.2.4, page 4-62

granularity ity when updating status register bit fields.

- Pay attention to pipeline-protection granular- Section 4.4.2.5, page 4-63

ity when accessing different registers in con-
secutive instructions.

Loop control - Understand when the loop-control registers Section 4.4.2.6, page 4-64
are accessed in the pipeline

- Avoid writing the BRC register in the last few Section 4.4.2.7, page 4-65
instructions of a block-repeat or local block-
repeat structure to prevent an unprotected
pipeline situation.

- Initialize the BRCx or CSR register at least 4 Section 4.4.2.8, page 4-66
cycles before the repeat instruction, or initial-
ize the register with an immediate value.

Condition evaluation - Try to set conditions well in advance of the Section 4.4.2.9, page 4-67
time that the condition is tested.

- When making an instruction execute condi- Section 4.4.2.10, page 4-68

tionally, use XCC instead of XCCPART to
create fully protected pipeline conditions, but
be aware of potential pipeline delays.

- Understand the XCCPART condition evalua- Section 4.4.2.11, page 4-70

tion exception.

Optimizing Assembly Code 4-55

Minimizing Pipeline and IBQ Delays

Table 4−6. Recommendations for Preventing Pipeline Delays (Continued)

Recommendation
Category Recommendation See ...

Memory usage - When working with dual-MAC and FIR in- Section 4.4.3.2, page 4-78
structions, put the Cmem operand in a differ-
ent memory bank.

- Map program code to a dedicated SARAM Section 4.4.3.3, page 4-79

memory block to avoid conflicts with data ac-
cesses.

- For 32-bit accesses (using an Lmem oper- Section 4.4.3.4, page 4-79
and), no performance hit is incurred if you
use SARAM (there is no need to use
DARAM).

IBQ usage - Align PC discontinuities in 32-bit memory Section 4.4.4, page 4-79
boundaries.
- Use short instructions as the first instructions
after a PC discontinuity.
- Use LOCALREPEAT when possible

4.4.2.1 In the case of a conflict, the front runner wins.

A pipeline conflict arises when two instructions in different phases in the pipe-
line compete for the use of the same resource. The resource is granted to the
instruction that is ahead in terms of pipeline execution, to increase overall in-
struction throughput.

4.4.2.2 Avoid consecutive accesses to the same register.

As shown in Figure 4−9, registers are not accessed in the same pipeline
phase. Therefore, pipeline conflicts can occur, especially in write/read or read/
write sequences to the same register. Following are three common register
pipeline conflict cases and how to resolve them.

Case 1: ARx write followed by an indirect addressing ARx read/update

Example 4−15 shows an AR write followed by an indirect-addressing AR read

and update. In the example, I2 has a 4-cycle latency due to pipeline protec-
tion. I2 must wait in the AD phase until I1 finishes its X phase. I2 requires
5 cycles (minimum 1 cycle + 4 pipeline-protection cycles).

4-56
 Minimizing Pipeline and IBQ Delays

Example 4−15. A-Unit Register (Write in X Phase/Read in AD Phase) Sequence

I1: MOV #k, AR1 ; Load AR1 with constant
; (AR1 modified in X phase of I1)
I2: MOV *AR1+, AC0 ; Load AC0 with value pointed to by AR1
; (AR1 read in AD phase of I2)
; Results: AC0 = content of memory at
; location #y16, AR1 = #y16 + 1

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3 I2 delayed 4 cycles
I2 I1 4
I2 I1 5
I2 I1 6 I1 writes to AR1
I2 I1 7 I2 reads AR1
I2 8
I2 9
I2 10
I2 11
I2 12

One solution is instruction rescheduling. Between I1 and I2 you can place 4

cycles worth of instructions from elsewhere.
I1: MOV #k, AR1
nop ; Replace NOPs with useful instructions
nop
nop
nop
I2: MOV *AR1+, AC0

Another solution is to use a MAR instruction to write to AR1:

I1: AMOV #k, AR1
I2: MOV *AR1+, AC0

A AMOV instruction modifies a register in the AD phase, preventing a pipeline

conflict. This solution is covered in more detail in section 4.4.2.3 (page 4-60).

Optimizing Assembly Code 4-57

Minimizing Pipeline and IBQ Delays

Case 2: ARx read followed by an indirect addressing ARx read/update

Example 4−16 shows an AR read followed by an indirect addressing AR up-

date. In the example, I2 is delayed 2 cycles due to pipeline protection. I2 must
wait in the AD phase until I1 finishes its R phase. I2 executes in 3 cycles
(minimum 1 cycle + 2 pipeline-protection cycles). Notice that AR1 is read by
I1 and is incremented by I2 in the same cycle (cycle 5). This is due to a special
A-unit register prefetch mechanism by a D-unit instruction that reads the A-unit
register in the R phase instead of the X phase. Appendix A lists the D-unit in-
structions that offer this feature.

Example 4−16. A-Unit Register Read/(Write in AD Phase) Sequence

I1: MOV AR1, AC1 ; AR1 read in R phase (A-unit register
prefetch)
I2: MOV *AR1+, AC0 ; AR1 updated in AD phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3 I2 delayed 2 cycles
I2 I1 4
I1 reads AR1 at the beginning of
cycle 5 (A-unit register pre-
I2 I1 5
fetch); I2 increments AR1 at the
end of cycle 5
I2 I1 6
I2 I1 7
I2 8
I2 9
I2 10

To prevent 1 to 3 of the pipeline-protection cycles, you can reschedule instruc-

tions. If possible, take up to 3 cycles worth of instructions from elsewhere in
the program and place them between I1 and I2.
MOV AR1, AC1
nop ; Replace NOPs with useful instructions
nop
MOV *AR1+, AC0
One could consider that using a MAR for the ARx register update could also
be solution. However as Example 4−16 shows, using MAR can cause unnec-
essary pipeline conflicts. This is covered in detail in section 4.4.2.3.

4-58
 Minimizing Pipeline and IBQ Delays

Case 3: ACx write followed by an ACx read

Accumulators and registers not associated with address generation are read
and written in the X phase, if MMR addressing is not used. Otherwise, the read
and write happen in the R and W phases, respectively.

The (write in X phase)/(read in R phase) sequence shown in Example 4−17

costs 1 cycle for pipeline protection. AC0 is updated in the X phase of I1. AC0
must be read in the R phase of I2, but I2 must wait until I1 has written to AC0.
The 1 cycle can be regained if you move a 1-cycle instruction between I1 and
I2:
I1:ADD #1, AC0
nop ; Replace NOP with useful instruction
I2:MOV mmap(AC0_L), AC2

Notice that
ADD #1, AC0
MOV AC0, AC2

will not cause pipeline conflicts because AC0 is read by (AC2=AC0) in the X
phase. When AC0_L is accessed via the memory map (@AC0_L || mmap()),
it is treated as a memory access and read in the read phase.

Example 4−17. Register (Write in X Phase)/(Read in R Phase) Sequence

I1: ADD #1, AC0 ; AC0 updated in X phase
I2: MOV mmap(AC0_L), AC2 ; AC0_L read in R phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3
I2 I1 4
I2 I1 5
I2 I1 6 I1 updates AC0; I2 delayed
I2 I1 7 I2 reads AC0
I2 8
I2 9

Optimizing Assembly Code 4-59

Minimizing Pipeline and IBQ Delays

4.4.2.3 Use MAR type of instructions, when possible, to modify ARx and Tx registers, but
avoid read/write register sequences, and pay attention to instruction size.

The MAR type of instructions (AMOV, AMAR, AADD, ASUB) use independent
hardware in the data-address generation unit (DAGEN) to update ARx and Tx
registers in the AD phase of the pipeline. You can take advantage of this fact
to avoid pipeline conflicts, as shown in Example 4−18. Because AR1 is up-
dated by the MAR instruction prior to being used by I2 for addressing genera-
tion, no cycle penalty is incurred.

However, using a MAR instruction could increase instruction size. For exam-
ple, AADD T1, AR1 requires 3 bytes, while ADD T1, AR1 requires 2 bytes. You
must consider the tradeoff between code size and speed.

Example 4−18. Good Use of MAR Instruction (Write/Read Sequence)

I1: AMOV #k, AR1 ; AR1 updated in AD phase
I2: MOV *AR1+, AC0 ; AR1 read in AD phase
; (No cycle penalty)

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2 I1 updates AR1
I2 I1 3 I2 reads AR1
I2 I1 4
I2 I1 5
I2 I1 6
I2 I1 7
I2 8

Example 4−19 shows that sometimes, using a MAR instruction can cause
pipeline conflicts. The MAR instruction (I2) attempts to write to AR1 in the AD
phase, but due to pipeline protection, I2 must wait for AR1 to be read in the
R phase by I1. This causes a 2-cycle latency. Notice that AR1 is read by I1
and is updated by I2 in the same cycle (cycle 5). This is made possible by the
A-unit register prefetch mechanism activated in the R phase of the C55x DSP
(see page 4-58). I1 is one of the D-unit instructions listed in Appendix A.

One way to avoid the latency in Example 4−19 is to use the code in
Example 4−20:
I1: ADD mmap(AR1), T2, AC1 ; AR1 read in R phase and
I2: ADD T1, AR1 ; AR1 updated in X phase
; (No cycle penalty)

4-60
 Minimizing Pipeline and IBQ Delays

Example 4−19. Bad Use of MAR Instruction (Read/Write Sequence)

I1: ADD mmap(AR1), T2, AC1 ; AR1 read in R phase
; (A-unit register prefetch)
I2: AADD T1, AR1 ; AR1 updated in AD phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3 I2 delayed 2 cycles
I2 I1 4
I1 reads AR1 at the beginning
of cycle 5 (A-unit register
I2 I1 5
prefetch); I2 updates AR1 at
the end of cycle 5
I2 I1 6
I2 I1 7
I2 8
I2 9
I2 10

Example 4−20. Solution for Bad Use of MAR Instruction (Read/Write Sequence)
I1: ADD mmap(AR1), T2, AC1 ; AR1 read in R phase
; (A-unit register prefetch)
I2: ADD T1, AR1 ; AR1 updated in X phase
; (No cycle penalty)

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3
I2 I1 4
I2 I1 5 I1 reads AR1
I2 I1 6
I2 I1 7 I2 updates AR1
I2 8

Optimizing Assembly Code 4-61

Minimizing Pipeline and IBQ Delays

4.4.2.4 Pay attention to pipeline-protection granularity when updating status register

bit fields.

A write to STx can generate the following not-so-obvious pipeline stalls:

- Case 1: Stall during the decode phase when followed by an instruc-

tion doing a memory access.
This is illustrated in Example 4−21. I2 needs to wait in the D phase until
the value of CPL is set by I1. This causes an 5-cycle pipeline delay.

Example 4−21. Stall During Decode Phase

I1: OR #k, mmap(ST1)
I2: MOV *SP(1), AC2

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3
I2 I1 4
I2 I1 5
I2 I1 6
ST1 writes by I1 and I2 read
I2 I1 7
CPL in same cycle
I2 8
I2 9
I2 10
I2 11
I2 12
I2 13

4-62
 Minimizing Pipeline and IBQ Delays

- Case 2: Stall due to coarse pipeline-protection granularity on STx

register.
STx register bits can be viewed by the pipeline-protection unit as individual
bits or as part of the same pipeline bit group (listed in Table 4−7).

Table 4−7. Bit Groups for STx Registers

Registers Bit Group Members

ST0 AC0V0, AC0V1, AC0V2, AC0V3

ST1 CPL, C54CM, ARMS

ST2 AR0LC to AR7LC, CDPLC

ST1 and ST3 C54CM, C16, SXMD, SATD, M40, SST

An instruction accessing an STx register bit can have a pipeline conflict

with another instruction accessing a different bit of the same STx register,
if:
J One of instructions makes a direct reference to the STx register name.
J The bits being accessed belong to the same group as listed in
Table 4−7.
Bits within the same group have coarse granularity: you cannot ac-
cess one bit without delaying the access to other bits within the same
group.
In the example below, I2 gets delayed 4 cycles until AR1LC gets set
because AR1LC and AR2LC belong to the same group.
I1: BSET AR7LC ; AR7LC is written in the
; X phase
I2: MOV *AR1++%, AC1 ; AR1 circular update in the D
; phase conflicts with the
; AR7LC update because AR1LC
; and AR7LC belong to the same
; STx group.

4.4.2.5 Pay attention to pipeline-protection granularity when accessing different registers

in consecutive instructions.

Registers can be viewed by the pipeline-protection unit as individual registers

or as part of the same pipeline register group (listed in Table 4−8).

A pipeline delay can occur if two different registers belonging to the same
group are accessed at the same time.

Optimizing Assembly Code 4-63

Minimizing Pipeline and IBQ Delays

Table 4−8. Pipeline Register Groups

Register Group Group Members

Accumulator 0 AC0L, AC0H, AC0G

Accumulator 1 AC1L, AC1H, AC1G

Accumulator 2 AC2L, AC2H, AC2G

Accumulator 3 AC3L, AC3H, AC3G

Block repeat registers RSA0, REA0, RSA1, REA1

Transition registers TRN0, TRN1

4.4.2.6 Understand when the loop-control registers are accessed in the pipeline.

As in any register, the h/w loop controller registers (CSR, RPTC, RSA0, REA0,
RSA1, REA1, BRC0, BRC1, and BRS1) can be read and written in:

- The AD phase (example: MOV #K12, BRC0)

- The R or write phase when accessed using MMR addressing (example:

@BRC0)

- The X phase when not in MMR addressing mode.

Loop-control registers can also be modified by the repeat instruction. For ex-
ample:

- During RPT loop execution:

J CSR or an instruction constant is loaded into RPTC in the AD phase of

the single RPT instruction.
J RPTC is tested and decremented in the Decode (D) phase of each re-
peated instruction.

- During RPTB and RPTBLOCAL loop execution:

J BRS1 is loaded into BRC1 in the AD phase of the blockrepeat or local-

repeat instructions.
J RSAx and REAx are loaded in the AD phase of the blockrepeat or lo-
calrepeat instructions.
J BRCx is decremented in the decode phase of the last instruction of the
loop.
J RSAx and REAx are constantly read throughout the loop operation.

4-64
 Minimizing Pipeline and IBQ Delays

4.4.2.7 Avoid writing the BRC register in the last few instructions of a block-repeat
or local block-repeat structure to prevent an unprotected pipeline situation.

Writing to BRC in one of the last few instructions (exact number depends on
the specific instruction) of a block-repeat structure could cause an unprotected
pipeline situation. On the other hand, reading BRC is pipeline protected and
does not insert an extra pipeline stall cycle.

BRC write accesses may not be protected in the last cycles of a block-repeat
structure. Do not write to BRC0 or BRC1 within those cycles. This can be seen
in Example 4−22. BRC0 is to be written to by I1 in the W phase (cycle 7),while
BRC0 is decremented in the D phase of I2. The pipeline-protection unit can-
not guarantee the proper sequence of these operations (write to BRC0 and
then decrement BRC0). BRC0 is decremented by I2 before BRC0 changed
by I1. On the other hand, certain instructions are protected (for example,
BRC0 = #k is pipeline protected).

Example 4−22. Unprotected BRC Write

I1: ADD #1, mmap(BRC0) ;BRC0 written in W phase
I2: Inst2 ; Last instruction of a
; block−repeat loop BRC0
; decremented in D phase of last
; instruction

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2 CPU decrements BRC0
I2 I1 3
I2 I1 4
I2 I1 5
I2 I1 6
I2 I1 7 I1 changes BRC0 late
I2 8

Optimizing Assembly Code 4-65

Minimizing Pipeline and IBQ Delays

4.4.2.8 Initialize the BRCx or CSR register at least 4 cycles before the repeat instruction,
or initialize the register with an immediate value.

Whenever BRC1 is loaded, BRS1 is loaded with the same value. In

Example 4−23, BRC1 and BRS1 are to be loaded by I1 in X phase of the pipe-
line (cycle 6), while BRS1 is to be read by I2 in the AD phase (cycle 3) to initial-
ize BRC1. The pipeline-protection unit keeps the proper sequence of these op-
erations (write to BRS1 and then read BRS1) by delaying the completion of I2
by 4 cycles. Example 4−24 shows a similar situation with CSR.

Instruction rescheduling or initialization of BRC1/CSR with an immediate val-

ue will remove the pipeline conflict in Example 4−23. An instruction that loads
BRC1/CSR with an immediate value will do so in the AD phase of the
instruction.

Example 4−23. BRC Initialization

I1: MOV *AR1, BRC1 ; BRC1 and BRS1 loaded in X phase
I2: RPTB label ; BRS1 is copied to BRC1 in AD phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3 I2 delayed 4 cycles
I2 I1 4
I2 I1 5
I2 I1 6 I1 loads BRS1
I2 I1 7 I2 copies BRS1 into BRC1
I2 8
I2 9
I2 10
I2 11
I2 12

4-66
 Minimizing Pipeline and IBQ Delays

Example 4−24. CSR Initialization

I1: MOV *AR1, CSR ; CSR written in X phase
I2: RPT CSR ; CSR read in AD phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3 I2 delayed 4 cycles
I2 I1 4
I2 I1 5
I2 I1 6 I1 loads CSR
I2 I1 7 I2 reads CSR
I2 8
I2 9
I2 10
I2 11
I2 12

4.4.2.9 Try to set conditions well in advance of the time that the condition is tested.

Conditions are typically evaluated in the R phase of the pipeline with the follow-
ing exceptions:

- For XCC instruction (execute(AD_unit) keyword used in algebraic syntax),

the condition is evaluated in the AD phase; for example:
XCC cond

- For XCCPART instruction (execute(D_unit) keyword used in algebraic

syntax), the condition is evaluated in the X phase; for example:
XCCPART cond
The exception is when the XCCPART instruction conditions a write to
memory (see section 4.4.2.11). In this case the condition is evaluated in
the R phase; for example:
XCCPART cond
|| MOV AC1, *AR1+

- In an RPTCC, the condition is evaluated in the X phase.

Optimizing Assembly Code 4-67

Minimizing Pipeline and IBQ Delays

Example 4−25 involves a condition evaluation preceded too closely by a write

to the register affecting the condition. AC1 is updated by I1 in the X phase,
while I2 must read AC1 in the R phase to evaluate the condition (AC1 == #0).
The pipeline-protection unit ensures the proper sequence of these operations
(write to AC1 and then test AC1) by delaying the completion of I2 by 1 cycle.
The solution is to move I1 within the program, such that it updates AC1 at least
1 cycle sooner.

Example 4−25. Condition Evaluation Preceded by a X-Phase Write to the Register

Affecting the Condition
I1: ADD #1, AC1 ; AC1 update in X phase
I2: BCC #label, AC1==#0 ; AC1 test in R phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I2 I1 3
I2 I1 4
I2 I1 5
I2 I1 6 I1 updates AC1; I2 delayed
I2 I1 7 I2 tests AC1
I2 8
I2 9

4.4.2.10 When making an instruction execute conditionally, use XCC instead of

XCCPART to create fully protected pipeline conditions, but be aware of
potential pipeline delays.

- Use of XCC: Example 4−26 shows a case where a register load operation
(MOV *AR3+, AC1) is made conditional with the XCC instruction. When
XCC is used, the condition is evaluated in the AD phase, and if the
condition is true, the conditional instruction performs its operations
in the AD through W phases. In Example 4−26, the AR3+ update in I3
depends on whether AC0 > 0; therefore, the update is postponed until I1
updates AC0. As a result, I3 is delayed by 4 cycles.

- Use of XCCPART: One solution to the latency problem in Example 4−26

is to move I1 such that AC0 is updated 4 cycles earlier. Another solution
is to use XCCPART, as shown in Example 4−27. When XCCPART is
used, the condition is evaluated in the X phase, and if the condition
is true, the conditional instruction performs its X phase operation.

4-68
 Minimizing Pipeline and IBQ Delays

Operations in the AD through R phases will happen unconditionally.

In Example 4−27, AR3 is updated in the AD phase regardless of the condi-
tion. Also, the memory read operation in I2 will always happen. However,
the memory value will be written AC1 only if the condition is true. Other-
wise, the memory value is discarded. Overall, a zero-latency execution is
achieved.
Notice that the advantage of using XCC is that it conditions the entire effect
of the instruction, not just part. However, this could come at the expense of
added latency cycles.

Example 4−26. Making an Operation Conditional With XCC

I1: ADD #1, AC0 ; AC0 updated in X phase
I2: XCC AC0>#0 ; AC0 tested in AD phase
I3: MOV *AR3+, AC1 ; If AC0 > 0, increment AR3 in AD
; phase and load AC1 in X phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
I2 I1 2
I3 I2 I1 3 I2 and I3 delayed 4 cycles
I3 I2 I1 4
I3 I2 I1 5
I3 I2 I1 6 I1 updates AC0
I3 I2 I1 7 I2 tests AC0
I3 I2 8 If AC0 > 0, I3 updates AR3
I3 I2 9
I3 I2 10
I3 I2 11
I3 I2 12 If AC0 > 0, I3 updates AC1
I3 13

Optimizing Assembly Code 4-69

Minimizing Pipeline and IBQ Delays

Example 4−27. Making an Operation Conditional With execute(D_unit)

I1: ADD #1, AC0 ; AC0 updated in X phase
I2: XCCPART AC0>#0 ; AC0 tested in X phase
I3: MOV *AR3+, AC1 ; If AC0 > 0, load AC1 in X phase.
; Update AR3 in AD phase regardless
of the condition.

D AD AC1 AC2 R X W Cycle Comment

I1
I2 I1
I3 I2 I1
I3 I2 I1 I3 updates AR3 unconditionally
I3 I2 I1
I3 I2 I1 I1 updates AC0
I3 I2 I1 I2 tests AC0
I3 I2 If AC0 > 0, I3 updates AC1
I3

4.4.2.11 Understand the XCCPART condition evaluation exception.

Typically, XCCPART causes the CPU to evaluate the condition in the execute
(X) phase. The exception is when you make a memory write operation depen-
dent on a condition, in which case the condition is evaluated in the read (R)
phase.

In Example 4−28, AR3 is to be read by I1 in R phase to evaluate the condition,

while AR3 is to be modified by I3 in the AD phase. The pipeline-protection unit
keeps the proper sequence of these operations (read AR3 and then write to
AR3) by delaying the completion of I3 by 2 cycles. Notice that AR3 is tested
by I1 and is modified by I3 in the same cycle (cycle 5). This is enabled by an
A-unit register prefetch mechanism activated in the R phase of the C55x DSP.

To prevent the 2-cycle delay in Example 4−28, you can insert two other, non-
conflicting instructions between I2 and I3.

4-70
 Minimizing Pipeline and IBQ Delays

Example 4−28. Conditional Parallel Write Operation Followed by an

AD-Phase Write to the Register Affecting the Condition
I1: XCCPART AR3==#0 ; AR3 tested in R phase
I2: || MOV AC1, *AR1+ ; If AR3 contains 0, write to
; memory in W phase. Update AR1
; regardless of condition.
I3: MOV *(AR3+T0), AC1 ; AR3 updated in AD phase

D AD AC1 AC2 R X W Cycle Comment

I1 1
||I2
I3 I1 2 I2 updates AR1 unconditionally
||I2
I3 I1 3 I3 delayed 2 cycles
||I2
I3 I1 4
||I2
I3 I1 5 I1 tests AR3; I3 modifies AR3
||I2
I3 I1 6
||I2
I3 I1 7
||I2
I3 8

I3 9

I3 10

Optimizing Assembly Code 4-71

Minimizing Pipeline and IBQ Delays

4.4.3 Memory Accesses and the Pipeline

This section discusses the impact of the memory mapping of variables/array
on the execution of C55x instructions. The two factors to consider are the type
of memory (DARAM vs SARAM) and the memory block allocation.
- DARAM supports 2 accesses/cycle to the same memory block by per-
forming one access in the first half-cycle and the other in the next half-
cycle. Table 4−9 lists the accesses performed in each half-cycle, and
Table 4−11 shows how the different memory accesses are performed. Ad-
dress bus loads are omitted from Table 4−11 for simplicity.
- SARAM supports 1 access/cycle to one memory block.

Table 4−9. Memory Accesses

Bus Used Half-Cycle Used
Type of Access (see Table 4−10) for DARAM
Instruction fetch P First and second halves
Data operand read (smem) D Second half
First data operand read (xmem) D First half
Second data operand read (ymem) C Second half
Third data operand (cmem) B Second half
Long (32-bit) data operand read C and D (Note 1) Second half
(lmem)
Data operand write (smem) E First half
Long (32-bit) data operand write E and F (Note 2) First and second halves
(lmem)
First data operand write (xmem) F Second half
Second data operand write (ymem) E First half
Coefficient read (cmem) B Second half
Notes: 1) Address is driven by D bus. Data is driven in C and D buses.
2) Address is driven by E bus. Data is driven in E and F buses.

4-72
 Minimizing Pipeline and IBQ Delays

Table 4−10. C55x Data and Program Buses

Bus Description
B BB. This data-read data bus carries a 16-bit coefficient data value (Cmem)
from data space to the CPU.
C, D CB, DB. Each of these data-read data buses carries a 16-bit data value
from data space to the CPU. DB carries a value from data space or from
I/O-space. In the case of an Lmem (32-bit) read, DB presents the data
address but each half of the 32-bit data comes in DB and CB.
E, F EB, FB. Each of these data-write data buses carries a 16-bit data value to
data space from the CPU. EB carries a value to data space or to I/O-
space. In the case of an Lmem (32-bit) write, EB presents the data ad-
dress but each half of the 32-bit data comes in EB and FB.
P PB. This program read bus carries 32-bit instructions from program space
to the IBQ.

Optimizing Assembly Code 4-73

Minimizing Pipeline and IBQ Delays

Table 4−11. Half-Cycle Accesses to Dual-Access Memory (DARAM) and the Pipeline
(Note 1)

Access Type D AD AC1 AC2 R X W W+

Smem read D

Smem write E

Lmem read C D

Lmem write E F

Xmem read
C D
|| Ymem read

Xmem write
E F
|| Ymem write

Xmem read
|| Ymem read D
C
|| Cmem read B
(Note 2)

Xmem read
D E
|| Ymem write

Lmem read
C D E F
|| Lmem write

Notes: 1) For detailed memory access description, refer to section 4.4.3.1.

2) B (Cmem) should be in a different DARAM block than the D (Xmem) operand to avoid one delay cycle.

4.4.3.1 Memory Access and the Pipeline

A CPU memory access is pipelined across three clock cycles as described in

Table 4−12.

- For a read operation: In the first cycle (request cycle), the request and
address are placed on the bus. In the second cycle (memory read cycle),
the read access is done to the memory. In the third cycle (data read cycle),
the data is delivered to the buses.

- For a write operation: In the first cycle (request cycle), the request and
address are placed on the bus. In the second cycle (data write cycle), the
data is written to the buses. In the third cycle (memory write), the write ac-
cesses are done to the memory.

4-74
 Minimizing Pipeline and IBQ Delays

Table 4−12. Memory Accesses and the Pipeline

Pipeline Phase AC1 AC2 R X W W+
Read operation Request Memory Data
read read

Write operation Request Data Memory

written written

The memory access happens in the AC2 phase for reads and in the W+
phase for writes.
As seen in Table 4−9, two simultaneous acceses can occur to the same DA-
RAM block, and only one access to a SARAM block.
Ideally, we should allocate all data into DARAM due to its higher memory band-
width (2 accesses/cycle). However, DARAM is a limited resource and should
be used only when it is advantageous. Following are recommendations to
guide your memory mapping decisions.
- Reschedule instructions.

- Reduce memory accesses by using CPU registers to hold data.

- Reduce memory accesses by using a local repeat instruction, an instruc-

tion that enables the CPU to repeatedly execute a block of code from the
instruction buffer queue.
- Relocate variables and data arrays in memory, or consider temporarily
copying arrays to other nonconflicting memory banks at run time.
It is important to note that in case of a memory access contention between
CPU and DMA, the C55x has the following bus priority:
- CPU data accesses (highest priority)

- CPU program access

- DMA access (lowest priority)

Following are some typical memory conflicts:

Case 1. A write-pending case: Write conflicting with a Dual-Operand
Read
Example 4−29 shows a conflict between a write and a dual-operand read. In
this case, there is a conflict between the operand write access (E bus) in W+
and the second data read access (C bus) in AC2. This conflict is known as write
pending and is resolved automatically by delaying the write access by one
cycle (write pends).

Optimizing Assembly Code 4-75

Minimizing Pipeline and IBQ Delays

The actual execution time of these instructions does not increase, because the
delayed (pending) write memory access (I1) is performed while the read in-
struction (I5) is in the R phase.

Example 4−29. A Write Pending Case

I1: MOV AC0, *AR3+ ; Write happens in the W+ phase
I2: NOP
I3: NOP
I4: NOP
I5: ADD *AR4+, *AR5+, A ; Dual read happens in the R
; phase

D AD AC1 AC2 R X W W+ Cycle Comment

I1 1
I2 I1 2
I3 I2 I1 3
I4 I3 I2 I1 4
I5 I4 I3 I2 I1 5
I5 I4 I3 I2 I1 6
I5 I4 I3 I2 I1 7
I1 write conflicts with I5
reads. I1 write pends in the
I5 I4 I3 I2 I1 8
memory buffers and reads
complete.
I5 I4 I3 I2 9 I1 write completes.
I5 I4 I3 10
I5 I4 11
I5 12

4-76
 Minimizing Pipeline and IBQ Delays

Example 4−30. A Memory Bypass Case

I1: MOV AC0, *AR3+ ; Initially, AR3 and AR4 have the
; same values. Write happens in
; the W+ phase.
I2: ADD *AR4+, *AR5+, A ; Dual read happens in the R
; phase

D AD AC1 AC2 R X W W+ Cycle Comment

I1 1
I2 I1 2
I2 I1 3
I2 I1 4
I2 tries to read from
memory. Pipeline detects I1
I2 I1 5 instruction processing and
waits for “memory bypass” in
R phase.
I2 tries to read from inter-
I2 I1 6 nal buses, but has to wait
for I2 W phase.
I1 writes data to internal
I2 I1 7 buses in W phase. I2 delayed
one cycle.
I2 I1 8 I2 can now proceed.
I2 9
I2 10

Case 2. A memory-bypass case:

If in Example 4−30, any read access (via D or C buses) is from the same
memory location in memory where the write access should occur, the CPU by-
passes reading the actual memory location; instead, it reads the data directly
from the internal buses. This allows the pipeline to perform a memory write ac-
cess in a later pipeline phase than that in which the next instruction reads from
the same memory location. Without this memory bypass feature, the delay
would have been 3 cycles.

Optimizing Assembly Code 4-77

Minimizing Pipeline and IBQ Delays

4.4.3.2 When working with dual-MAC and FIR instructions, put the Cmem operand
in a different memory bank.
Provided code is not executed from the same memory block in which you have
the data being accessed by that code, the only memory access type which can
generate a conflict in a DARAM is the execution of instructions requiring three
data operands in 1 cycle: Xmem, Ymem, and Cmem (coefficient operand). Ex-
amples of two commonly used instructions that use three data operands are:
- Dual multiply-and-accumulate (MAC) instruction:
MAC Xmem, Cmem, ACx
:: MAC Ymem, Cmem, ACy
- Finite impulse response filter instructions:
FIRSADD Xmem, Ymem, Cmem, ACx, ACy
FIRSSUB X mem, Ymem, Cmem, ACx, ACy
This memory conflict can be solved by maintaining the Ymem and Xmem oper-
ands in the same DARAM memory bank but putting the Cmem operand into
a different memory bank (SARAM or DARAM).
When cycle intensive DSP kernels are developed, it is extremely important to
identify and document software integration recommendations stating which
variables/arrays must not be mapped in the same DARAM memory block.
The software developer should also document the associated cycle cost when
the proposed optimized mapping is not performed. That information will pro-
vide the software integrator enough insight to make trade-offs.
When cycle intensive DSP kernels are developed, it is extremely important to
identify and document software integration recommendations stating which
variables/arrays must not be mapped in the same dual access memory. The
software developer should also document the associated cycle cost when the
proposed optimized mapping is not performed. That information will provide
the software integrator enough insight to make trade-offs. Table 4−13 provides
an example of such table: if the 3 arrays named “input.” “output,” and “coeffi-
cient” are in the same DARAM, the subroutine named “filter” will have 200
cycle overhead.

Table 4−13. Cross-Reference Table Documented By Software Developers to Help

Software Integrators Generate an Optional Application Mapping

Array 1 Array 2 Array 3 Cycle

Cycle Name Name Name Cost per
Routine Weight (Xmem) Size (Ymem) Size (Cmem) Size Routine
filter 10% input 40 output 40 coefficient 10 10*20

… … … … … … … … …

4-78
 Minimizing Pipeline and IBQ Delays

4.4.3.3 Map program code to a dedicated SARAM memory block to avoid conflicts with
data accesses.

If a DARAM block maps both program and data spaces of the same routine,
a program code fetch will conflict, for example, with a dual (or triple) data oper-
and read (or write) access if they are performed in the same memory block.
C55x DSP resolves the conflict by delaying the program code fetch by one
cycle. It is therefore recommended to map the program code in a dedicated
program memory bank: generally a SARAM memory bank is preferred. This
avoids conflicts with data variables mapped in the high bandwidth DARAM
banks.

Another way to avoid memory conflicts is to use the 56-byte IBQ to execute
blocks of instructions without refetching code after the 1st iteration (see local-
repeat{} instruction). Conflict will only occur in the first loop iteration.

4.4.3.4 For 32-bit accesses (using an Lmem operand), no performance hit is

incurred if you use SARAM (there is no need to use DARAM).

When a 32-bit memory access is performed with Lmem, only one address bus
(DAB or EAB) is used to specify the most and least significant words of the
32-bit value. Therefore, reading from or writing to a 32-bit memory location in
an SARAM bank occurs in 1 cycle.

4.4.4 Recommendations for Preventing IBQ Delays

The IBQ is filled by the fetch pipeline in 32-bit instruction packets each aligned
in a 32-bit memory boundary. Instructions stored in the IBQ are dispatched to
the execution pipeline in 48-bit packets and consumed by the decode (D)
phase of the pipeline at a rate that depends on the length of the instruction be-
ing decoded. The number of bytes waiting in the IBQ to be decoded is known
as IBQ fetch advance and should be greater than the instruction length to be
decoded in order to have enough data to feed the execution pipeline. Other-
wise an IBQ stall in the fetch pipeline will occur. IBQ fetch advance can be be-
tween 0 and 24 bytes.

4.4.4.1 IBQ Performance on PC Discontinuities

When a PC discontinuity occurs (a subroutine call, a branch, a block repeat

loop,...), the IBQ content is flushed and the fetch pipeline starts filling the IBQ
with the instructions at the target address. One delay cycle will occur if the first
32-bit aligned memory fetch packet does not contain a complete parallel in-
struction pair (for example, branching to a 4-byte-aligned nop_16||nop_16
does not cause an extra stall) or a complete single instruction and the first byte
of the following instruction.

Optimizing Assembly Code 4-79

Minimizing Pipeline and IBQ Delays

To avoid IBQ delay cycles and increase the chances of the first 32-bit packet
to contain the target address instruction and the first byte of the next instruc-
tion, following are recommendations:

- Align PC discontinuities in a 32-bit memory boundary

This means for example, to align

J the starting address of a subroutine
J the first instruction inside a block repeat

- Use short instructions at the target address of PC discontinuities. Howev-

er, if the size of the first instruction after the PC discontinuity is 4 bytes or
larger there will be a delay of 1 cycle.

To better understand the effect of the IBQ in code cycles, we present 3 typical
PC discontinuitites cases:

Case 1: IBQ Behavior During Branches

Examples 1a, 1b, and 1c show branch examples where the IBQ behavior is
illustrated.

Example 1a: No Delay

B label1
.....
....
.align 4 ; force 32−bit alignment
label1: XOR ac1 << #1, ac0 ; 3 bytes
MOV t0, ac2 ; 2 bytes
By using the .align 4 assembler directive, we are able to avoid an IBQ stall. The
first 32-bit fetch packet at the PC discontinuity target address, label1, contains
the first instruction and the first byte of the next instruction.

Example 1b: One Delay

B label2
.....
....
.align 4 ; force 32−bit alignment
label2: MOV AC0, *AR3−
|| MOV #15, BRC1 ; 5 bytes
MOV t0, ac2 ; 2 bytes
In this case, even though we align the target address in a 32-bit boundary, the
first packet cannot contain the first instruction and the first byte of the next in-
struction. Using a shorter instruction at the target address (less than 3 bytes)
can prevent this delay cycle. For example, we can rearrange the parallelism
as shown in Example 1c.

4-80
 Minimizing Pipeline and IBQ Delays

Example 1c: No Delay

B label2
.....
.align 4
label2: MOV AC0, *AR3− ; 2 bytes
MOV #15, BRC1
|| MOV t0, ac2 ; 5 bytes

Case 2: IBQ Behavior During Repeat Block

IBQ delays can also occur when the PC discontinuity is caused by a block re-
peat loop. During a block repeat processing the IBQ fetches sequentially the
instructions until it reaches the last instruction. At this point, since the IBQ does
not know what the size of the last instruction is (it only knows the address of
the last instruction), it fetches 7 more bytes. Thus, the IBQ may be fetching
more program bytes than it actually needs. This overhead can cause some de-
lays. The following example will show this:

Example 2
.align 4
4000 nop
4001 rptb label3
4004 amov #0x8000 , xar0
400a ;;nop_16 ;← commented nop_16 instruction
400a nop_16 || nop_16 ;#1
400e nop_16 || nop_16 ;#2
4012 nop_16 || nop_16 ;#3
4016 nop_16 || nop_16 ;#4
401a nop_16 || nop_16 ;#5
401e nop_16 || nop_16 ;#6
4022 nop_16 || nop_16 ;#7
4026 nop_16 || nop_16 ;#8
402a nop_16 || nop_16 ;#9
label3:
402e nop_16 || nop_16 ;#10
4032

The block repeat loop has 11 instructions. Since the size of the first instruction
is 6 bytes, the total cycle count for this loop should be 12 cycles since it requires
two 32-bit fetch packets to have the entire first instruction (AMOV) and the first
byte of the next instruction. However, this code actually takes 13 cycles, as
illustrated below.

Optimizing Assembly Code 4-81

Minimizing Pipeline and IBQ Delays

Cycle 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Program
Address 04 08 0c 10 14 18 1c 20 24 28 2c 30 34 04 08 0c 10 14 18
Bus
Decode 04 0a 0e 12 16 1a 1e 22 26 2a 2e X X 04
PC
Note: S The Cycle row corresponds to the cycle timeline.
S The Program Address Bus row corresponds to the address of the 32-bit block requested from memory (fetch packet).
(For simplifying the notation, the 0x4004 is written as 04.)
S The Decode PC row corresponds to the address of the instruction that is being decoded.

There are two X slots that are used to fetch packets that are never decoded.
Therefore, we have 11 instructions for 13 cycles. One of the X slots is unavoid-
able because of the first 6-byte instruction (AMOV). The second, however,
could be used by another instruction. For example, if we uncomment the
nop_16 instruction, the cycle count will be the same—13 cycles—indicating
the opportunity to rearrange your code and insert a useful instruction instead
of the nop_16. This kind of delay can occur in block repeat loops but it does
not occur in local repeat loops. For this reason, the use of RPTBLOCAL is rec-
ommended. To avoid this kind of delay, use the following recommendation:

- Use RPTBLOCAL when possible. It will also save power.

IBQ delays can also occur when there are many 5- or 6-byte instructions con-
secutively. This will reduce the fetch advance and eventually produce delays.
To avoid this kind of delay, use the following recommendation:

- Avoid using too many 5- or 6-byte instructions (or 5- or 6-byte parallel in-
structions) consecutively. Doing so will eventually starve the IBQ. Instead,
mix in short instructions to balance the longer instructions, keeping in mind
that the average sustainable fetch, without incurring a stall, is 4 bytes.

Case 3: IBQ Behavior During Local Repeat Block

The RPTBLOCAL loop differs from a RPTB loop in that the size of the content
of the loop is able to fit completely in the IBQ. The maximum size of a RPTBLO-
CAL is 55 bytes between the address of the first instruction and the last instruc-
tion of the loop. If the last instruction of the loop is a 6-byte instruction then the
size of the loop is 61 bytes. During the processing of the loop the IBQ keeps
fetching instructions until it reaches the maximum fetch advance which is 24
bytes.

Since all the instructions of the local loop are in the IBQ, there are no IBQ de-
lays inside a RPTBLOCAL loop.

4-82
 Minimizing Pipeline and IBQ Delays

Special Case: IBQ Delays When the size of the Local Loop is Close to 61
Bytes

There is not much space left in the IBQ to fetch additional instructions. In this
case, after the loop completes, IBQ delays in the first instruction outside
the loop may occur. This delay will occur when the IBQ did not have enough
space to fetch the first instruction following the loop plus the first byte of the
second instruction (see Example 3). This type of IBQ delay could range from
1 to 6, depending on the size of the local loop and the length of the instruction
following the local loop.

Example 3: Delays After the Completion of a RPTBLOCAL Loop

RPTBLOCAL label4

; loop is 61 bytes

0x4000 NOP_16|| NOP_16 ; 4 bytes

0x4004 NOP_16|| NOP_16 ; 4 bytes
0x4008 NOP_16|| NOP_16 ; 4 bytes
0x400c NOP_16|| NOP_16 ; 4 bytes
0x4010 NOP_16|| NOP_16 ; 4 bytes
0x4014 NOP_16|| NOP_16 ; 4 bytes
0x4018 NOP_16|| NOP_16 ; 4 bytes
0x401c NOP_16|| NOP_16 ; 4 bytes
0x4020 NOP_16|| NOP_16 ; 4 bytes
0x4024 NOP_16|| NOP_16 ; 4 bytes
0x4028 NOP_16|| NOP_16 ; 4 bytes
0x402c NOP_16|| NOP_16 ; 4 bytes
0x4030 NOP_16 ; 2 bytes
0x4032 MAXDIFF AC0, AC1, AC2, AC1 ; 5 bytes
||MOV #0, AC3
label4:
0x4037 SUBADD T3,*mold_ptr+,AC0 ; 6 bytes
||AADD #1, mnew_ptr
0x403d NOP_16|| NOP_16 ; 4 bytes

The size of the RPTBLOCAL loop in the previous example is 61 bytes. Since
the size of the IBQ is 64 bytes, and since, in this case, the content of the loop
is aligned on a 32-bit boundary, the IBQ fetches also the first three bytes of the
first instruction after the loop. However, that is not sufficient to avoid delays as
the first instruction outside the loop (at 0x403d) address is a 4-byte instruction.
Therefore, in this case, five IBQ delays will occur after the completion of the
loop.

Optimizing Assembly Code 4-83

Minimizing Pipeline and IBQ Delays

4.4.4.2 The Speculative Pre-Fetch Feature

The C55x offers a speculative pre-fetch feature that can save several execu-
tion cycles of conditional control flow instructions when the condition detected
is “true.” For example, in the case of a conditional branch (BCC) with an imme-
diate value, the branch target address is known in the decode (D) phase (in
the case of an immediate value) or in the address (AD) phase (in the case of
a relative offset) of the pipeline but the condition is evaluated later in the Read
(R) phase. To avoid the extra cycle delays that could imply to wait for the condi-
tion to be known before fetching the target address, the C55x fetches the
branch target address speculatively and the fetch packet is stored in the IBQ.
If the condition is evaluated as “true,” then the instruction decoder can get the
branch target instruction from the IBQ with minimum latency.

4-84
 Chapter 5

 

The TMS320C55x (C55x) DSP is a 16-bit, fixed-point processor. This chapter

explains important considerations for performing standard- and extended-pre-
cision fixed-point arithmetic with the DSP. Assembly-language code examples
are provided to demonstrate the concepts.

Topic Page

5.1 Fixed-Point Arithmetic − a Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2

5.2 Extended-Precision Addition and Subtraction . . . . . . . . . . . . . . . . . . . 5-9
5.3 Extended-Precision Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
5.4 Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-17
5.5 Methods of Handling Overflows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-24

5-1
Fixed-Point Arithmetic − a Tutorial

5.1 Fixed-Point Arithmetic − a Tutorial

Digital signal processors (DSPs) have been developed to process complex al-
gorithms that require heavy computations. DSPs can be divided into two
groups: floating-point DSPs and fixed-point DSPs.

Typically, floating-point DSPs use 32-bit words composed of a 24-bit mantissa

and an 8-bit exponent, which together provide a dynamic range from 2−127 to
2128(1 − 2−23). This vast dynamic range in floating-point devices means that
dynamic range limitations may be virtually ignored in a design. Floating-point
devices are usually more expensive and consume more power than fixed-
point devices.

Fixed-point DSPs, like the TMS320C55x DSP, typically use 16-bit words. They
use less silicon area than their floating-point counterparts, which translates
into cheaper prices and less power consumption. Due to the limited dynamic
range and the rules of fixed-point arithmetic, a designer must play a more ac-
tive role in the development of a fixed-point DSP system. The designer has to
decide whether the 16-bit words will be interpreted as integers or fractions, ap-
ply scale factors if required, and protect against possible register overflows.

5.1.1 2s-Complement Numbers

In binary form, a number can be represented as a signed magnitude, where
the left-most bit represents the sign and the remaining bits represent the mag-
nitude.

+52 = 0 011 0100b

−52 = 1 011 0100b

This representation is not used in a DSP architecture because the addition al-
gorithm would be different for numbers that have the same signs and for num-
bers that have different signs. The DSP uses the 2s-complement format, in
which a positive number is represented as a simple binary value and a nega-
tive value is represented by inverting all the bits of the corresponding positive
value and then adding 1.

Example 5−1 shows the decimal number 353 as a 16-bit signed binary num-
ber. Each bit position represents a power of 2, with 20 at the position of the least
significant bit and 215 at the position of the most significant bit. The 0s and 1s
of the binary number determine how these powers of 2 are weighted (times 0
or times 1) when summed to form 353. Because the number is signed, 215 is
given a negative sign. Example 5−2 shows how to compute the negative of a
2s-complement number.

5-2
 Fixed-Point Arithmetic − a Tutorial

Example 5−1. Signed 2s-Complement Binary Number Expanded to Decimal Equivalent

215 214 213 212 211 210 29 28 27 26 25 24 23 22 21 20

0 0 0 0 0 0 0 1 0 1 1 0 0 0 0 1

= (0 x (−215)) + (0 x 214) + (0 x 213) + (0 x 212) + (0 x 211) + (0 x 210) + (0 x 29) + (1 x 28) + (0 x 27)

+ (1 x 26) + (1 x 25) + (0 x 24) + (0 x 23) + (0 x 22) + (0 x 21) + (1 x 20)
= (1 x 28) + (1 x 26) + (1 x 25) + (1 x 20)
= 256 + 64 + 32 + 1 = 353

Example 5−2. Computing the Negative of a 2s-Complement Number

Begin with a positive binary number (353 decimal): 0000 0001 0110 0001

Invert all bits to get the 1s complement: 1111 1110 1001 1110
Add 1 to get the 2s complement: + 1
Result: negative binary number (−353 decimal): 1111 1110 1001 1111

5.1.2 Integers Versus Fractions

The most common formats used in DSP programming are integers and frac-
tions. In signal processing, fractional representation is more common. A frac-
tion is defined as a ratio of two integers such that the absolute value of the ratio
is less than or equal to 1. When two fractions are multiplied together, the result
is also a fraction. Multiplicative overflow, therefore, never occurs. Note, how-
ever, that additive overflow can occur when fractions are added. Overflows are
discussed in section 5.5, beginning on page 5-24.

Figure 5−1 shows how you can interpret 2s-complement numbers as integers.
The most significant bit (MSB) is given a negative weight, and the integer is
the sum of all the applicable bit weights. If a bit is 1, its weight is included in
the sum; if the bit is 0, its weight is not applicable (the effective weight is 0). For
simplicity, the figure shows 4-bit binary values; however, the concept is easily
extended for larger binary values. Compare the 4-bit format in Figure 5−1 with
the 8-bit format in Figure 5−2. The LSB of a binary integer always has a bit
weight of 1, and the absolute values of the bit weights increase toward the
MSB. Adding bits to the left of a binary integer does not change the absolute
bit weights of the original bits.

Fixed-Point Arithmetic 5-3

Fixed-Point Arithmetic − a Tutorial

Figure 5−1. 4-Bit 2s-Complement Integer Representation

4-bit 2s-complement integer MSB LSB

Bit weights −23 = −8 22 = 4 21 = 2 20 = 1

Least positive value 0 0 0 1 =0+0+0+1=1

Most positive value 0 1 1 1 =0+4+2+1=7

Least negative value 1 1 1 1 = −8 + 4 + 2 + 1 = −1

Most negative value 1 0 0 0 = −8 + 0 + 0 + 0 = −8

Other examples: 0 1 0 1 =0+4+0+1=5

1 1 0 1 = −8 + 4 + 0 + 1 = −3

Figure 5−2. 8-Bit 2s-Complement Integer Representation

MSB LSB

−27 = −128 26 = 64 25 = 32 24 = 16 23 = 8 22 = 4 21 = 2 20 = 1

Figure 5−3 shows how 2s-complement numbers can be interpreted as a frac-

tions. The concept is much the same as that in Figure 5−1, but the bit weights
are fractional, meaning that the number cannot have an absolute value larger
than 1. Compare the 4-bit format in Figure 5−3 with the 8-bit format in
Figure 5−4. The MSB of a binary fraction always has a bit weight of −1, and
the absolute values of the bit weights decrease toward the LSB. Unlike adding
bits to the left of a binary integer, adding bits to the left of a binary fraction
changes the bit weights of the original bits.

5-4
 Fixed-Point Arithmetic − a Tutorial

Figure 5−3. 4-Bit 2s-Complement Fractional Representation

4-bit binary
MSB LSB
fraction

Bit weights −20 = −1 2−1 = 1/2 2−2 = 1/4 2−3 = 1/8

Least positive = 0 + 0 + 0 +1/8

0 0 0 1
value = 1/8

Most positive = 0 + 1/2+ 1/4+ 1/8

0 1 1 1
value = 7/8

Least negative = −1 + 1/2 + 1/4 + 1/8

1 1 1 1
value = −1/8

Most negative = −1 + 0 + 0 + 0
1 0 0 0
value = −1

= 0 + 1/2 + 0 + 1/8
Other examples: 0 1 0 1
= 5/8

= −1 + 1/2 + 0 + 1/8
1 1 0 1
= −3/8

Figure 5−4. 8-Bit 2s-Complement Fractional Representation

MSB LSB

−20 = −1 2−1 = 1/2 2−2 = 1/4 2−3 = 1/8 2−4 = 1/16 2−5 = 1/32 2−6 = 1/64 2−7 = 1/128

5.1.3 2s-Complement Arithmetic

An important advantage of the 2s-complement format is that addition is per-
formed with the same algorithm for all numbers. To become more familiar with
2s-complement binary arithmetic, refer to the examples in this section. You
may want to try a few examples yourself. It is important to understand how 2s-
complement arithmetic is performed by the DSP instructions, in order to effi-
ciently debug your program code.

Example 5−3 shows two 2s-complement additions. These binary operations

are completely independent of the convention the programmer uses to convert
them into decimal numbers. To highlight this fact, an integer interpretation and

Fixed-Point Arithmetic 5-5

Fixed-Point Arithmetic − a Tutorial

a fractional interpretation are shown for each addition. For simplicity, the ex-
amples use 8-bit binary values; however, the concept is easily extended for
larger binary values. For a better understanding of how the integer and frac-
tional interpretations were derived for the 8-bit binary numbers, see
Figure 5−2 (page 5-4) and Figure 5−4 (page 5-5), respectively.

Example 5−3. Addition With 2s-Complement Binary Numbers

2s-Complement Addition Integer Interpretation Fractional Interpretation

1 (carry)
0000 0101 5 5/128
+ 0000 0100 + 4 + 4/128
−−−−−−−−−−−−−− −−−− −−−−−−−
0000 1001 9 9/128

1 1 1 (carries)
0000 0101 5 5/128
+ 0000 1101 + 13 + 13/128
−−−−−−−−−−−−−− −−−−− −−−−−−−−
0001 0010 18 18/128

Example 5−4 shows subtraction. As with the additions in Example 5−3, an in-
teger interpretation and a fractional interpretation are shown for each com-
putation. It is important to notice that 2s-complement subtraction is the same
as the addition of a positive number and a negative number. The first step is
to find the 2s-complement of the number to be subtracted. The second step
is to perform an addition using this negative number.

5-6
 Fixed-Point Arithmetic − a Tutorial

Example 5−4. Subtraction With 2s-Complement Binary Numbers

2s-Complement Subtraction Integer Interpretation Fractional Interpretation

Original form:
0000 0101 5 5/128
− 0000 0100 − 4 − 4/128
−−−−−−−−−−−−−− −−−− −−−−−−−
2s complement of subtracted term:
1 1 (carries)
1111 1011
+ 1
−−−−−−−−−−−−−−
1111 1100
Addition form:
11111 1 (carries)
0000 0101 5 5/128
+ 1111 1100 + (−4) + (−4/128)
−−−−−−−−−−−−−− −−−−−− −−−−−−−−−
0000 0001 1 1/128
(final carry ignored)

Original form:
0000 0101 5 5/128
− 0000 1101 − 13 − 13/128
−−−−−−−−−−−−−− −−−−− −−−−−−−−
2s complement of subtracted term:
1111 0010
+ 1
−−−−−−−−−−−−−−
1111 0011
Addition form:
1 1 1 (carries)
0000 0101
+ 1111 0011 5 5/128
−−−−−−−−−−−−−− + (−13) + (−13/128)
1111 1000 −−−−−−− −−−−−−−−−−
−8 −8/128

Fixed-Point Arithmetic 5-7

Fixed-Point Arithmetic − a Tutorial

Example 5−5 shows 2s-complement multiplication. For simplicity, the exam-

ple uses a 4-bit by 4-bit multiplication and assumes an 8-bit accumulator for
the result. Notice that the 7-bit mathematical result is sign extended to fill the
accumulator. The C55x DSP sign extends multiplication results in this way, ex-
tending the result to either 32 bits or 40 bits. The effects of this type of sign ex-
tension can be seen in the integer and fractional interpretations of
Example 5−5. The integer is not changed by sign extension, but the fraction
can be misinterpreted. Sign extension adds an extra sign bit. If your program
assumes that the MSB of the result is the only sign bit, you must shift the result
left by 1 bit to remove the extra sign bit. In the C55x DSP, there is a control bit
called FRCT to automate this shift operation. When FRCT = 1, the DSP auto-
matically performs a left shift by 1 bit after a multiplication. You can clear and
set FRCT with the following instructions:
BCLR FRCT ; Clear FRCT
BSET FRCT ; Set FRCT

Example 5−5. Multiplication With 2s-Complement Binary Numbers

2s-Complement Multiplication Integer Interpretation Fractional Interpretation

0100 Multiplicand 4 4/8

x 1101 Multiplier x (−3) x (−3/8)
−−−−−−− −−−−−− −−−−−−−−−
0000100
000000
00100
1100 (see Note)
−−−−−−−−−
1110100 7-bit mathematical result −12 −12/64 (The MSB of the re-
sult is the only sign bit)

11110100 8-bit sign-extended result −12 −12/64 if properly inter-

in accumulator preted; −12/128 if incorrectly
interpreted. To remove extra
Note: Because the MSB is a sign bit, the
sign bit in MSB position, shift
final partial product is the 2s complement
result left by 1 bit.
negative of the multiplicand.

5-8
 Extended-Precision Addition and Subtraction

5.2 Extended-Precision Addition and Subtraction

Numerical analysis, floating-point computations, and other operations may re-
quire arithmetic operations with more than 32 bits of precision. Because the
C55x device is a 16-/32-bit fixed-point processor, software is required for arith-
metic operations with extended precision. These arithmetic functions are per-
formed in parts, similar to the way in which longhand arithmetic is done.

The C55x DSP has several features that help make extended-precision cal-
culations more efficient.

- CARRY bit: One of the features is the CARRY status bit, which is affected
by most arithmetic D-unit ALU instructions, as well as the rotate and shift
operations. CARRY depends on the M40 status bit. When M40 = 0, the
carry/borrow is detected at bit position 31. When M40 = 1, the carry/borrow
reflected in CARRY is detected at bit position 39. Your code can also expli-
citly modify CARRY by loading ST0_55 or by using a status bit clear/set
instruction. For proper extended-precision arithmetic, the saturation mode
bit should be cleared (SATD = 0) to prevent the accumulator from saturat-
ing during the computations.

- 32-bit addition, subtraction, and loads: Two C55x data buses, CB and
DB, allow some instructions to handle 32-bit operands in a single cycle.
The long-word load and double-precision add/subtract instructions use
32-bit operands and can efficiently implement extended-precision arith-
metic.

- 16-bit signed/unsigned multiplication: The hardware multiplier can

multiply 16-bit signed/unsigned numbers, as well as multiply two signed
numbers and two unsigned numbers. This makes 32-bit multiplication op-
erations efficient.

5.2.1 A 64-Bit Addition Example

The code in Example 5−6 adds two 64-bit numbers to obtain a 64-bit result.
The following are some code highlights:

- The partial sum of the 64-bit addition is efficiently performed by the follow-
ing instructions, which handle 32-bit operands in a single cycle.
Mnemonic instructions: MOV40 dbl(Lmem), ACx
ADD dbl(Lmem), ACx
Algebraic instructions: ACx = dbl(Lmem)
ACx = ACx + dbl(Lmem)

- For the upper half of a partial sum, the instruction that follows this para-
graph uses the carry bit generated in the lower 32-bit partial sum. Each

Fixed-Point Arithmetic 5-9

Extended-Precision Addition and Subtraction

partial sum is stored in two memory locations using MOV ACx, dbl(Lmem)
or dbl(Lmem) = ACx.
Mnemonic instruction: ADD uns(Smem), CARRY, ACx
Algebraic instruction: ACx = ACx + uns(Smem) + CARRY

- Typically, if a computation does not generate a carry, the CARRY bit is

cleared.
As shown in Figure 5−5, the ADD instruction with a 16-bit shift (shown fol-
lowing this paragraph) is an exception because it can only set the CARRY
bit. If this instruction does not generate a carry, the CARRY bit is left un-
changed. This allows the D-unit ALU to generate the appropriate carry
when adding to the lower or upper half of the accumulator causes a carry.
Mnemonic instruction: ADD Smem << #16, ACx, ACy
Algebraic instruction: ACy = ACx + (Smem << #16)
For illustration purposes, Figure 5−5 shows the normal effect of several
32-bit additions on the CARRY bit (referred to as C in the figure).

Example 5−6. 64-Bit Addition

;*********************************************************************
; 64−Bit Addition Pointer assignments:
;
; X3 X2 X1 X0 AR1 −> X3 (even address)
; + Y3 Y2 Y1 Y0 X2
; −−−−−−−−−−−−−− X1
; W3 W2 W1 W0 X0
; AR2 −> Y3 (even address)
; Y2
; Y1
; Y0
; AR3 −> W3 (even address)
; W2
; W1
; W0
;
;*********************************************************************

MOV40 dbl(*AR1(#2)), AC0 ; AC0 = X1 X0

ADD dbl(*AR2(#2)), AC0 ; AC0 = X1 X0 + Y1 Y0
MOV AC0,dbl(*AR3(#2)) ; Store W1 W0.
MOV40 dbl(*AR1), AC0 ; AC0 = X3 X2
ADD uns(*AR2(#1)),CARRY,AC0 ; AC0 = X3 X2 + 00 Y2 + CARRY
ADD *AR2<< #16, AC0 ; AC0 = X3 X2 + Y3 Y2 + CARRY
MOV AC0, dbl(*AR3) ; Store W3 W2.

Note: The algebraic instructions code example for 64-Bit Addition is shown in Example B−27 on page B-30.

5-10
 Extended-Precision Addition and Subtraction

Figure 5−5. Effect on CARRY of Addition Operations

C MSB LSB C MSB LSB
X F F F F F F F F F F ACx X F F F F F F F F F F ACx
_ + 1 +F F F F F F F F F F
1 0 0 0 0 0 0 0 0 0 0 1 F F F F F F F F F E

C MSB LSB C MSB LSB

X 0 0 7 F F F F F F F ACx X 0 0 7 F F F F F F F ACx
_ + 1 +F F F F F F F F F F
0 0 0 8 0 0 0 0 0 0 0 1 0 0 7 F F F F F F E

C MSB LSB C MSB LSB

X F F 8 0 0 0 0 0 0 0 ACx X F F 8 0 0 0 0 0 0 0 ACx
+ 1 _+F F F F F F F F F F
0 F F 8 0 0 0 0 0 0 1 1 F F 7 F F F F F F F

ACy = ACx + Smem + CARRY

C MSB LSB C MSB LSB
1 0 0 0 0 0 0 0 0 0 0 ACx 1 F F F F F F F F F F ACx
+ 0 _ + 0
0 0 0 0 0 0 0 0 0 0 1 ACy 1 0 0 0 0 0 0 0 0 0 0

Special Case: ACy = ACx + (Smem <<16)

C MSB LSB C MSB LSB
1 F F 8 0 0 0 F F F F ACx 1 F F 8 0 0 0 F F F F ACx
+0 0 0 0 0 1 0 0 0 0 +0 0 7 F F F 0 0 0 0
1 F F 8 0 0 1 F F F F ACy 1 F F F F F F F F F F

C keeps a values of 1, even when a CARRY is not generated.

5.2.2 A 64-Bit Subtraction Example

Example 5−7 subtracts two 64-bit numbers to obtain a 64-bit result. The fol-
lowing are some code highlights:

- The partial remainder of the 64-bit subtraction is efficiently performed by

the following instructions, which handle 32-bit operands in a single cycle.
Mnemonic instructions: MOV40 dbl(Lmem), ACx
SUB dbl(Lmem), ACx
Algebraic instructions: ACx = dbl(Lmem)
ACx = ACx − dbl(Lmem)

- For the upper half of the partial remainder, the instruction that follows this
paragraph uses the borrow generated in the lower 32-bit partial remainder.
The borrow is not a physical bit in a status register; it is the logical inverse
of CARRY. Each partial sum is stored in two memory locations using
MOV ACx, dbl(Lmem) or dbl(Lmem) = ACx.

Fixed-Point Arithmetic 5-11

Extended-Precision Addition and Subtraction

Mnemonic instruction: SUB uns(Smem), BORROW, ACx

Algebraic instruction: ACx = ACx − uns(Smem) − BORROW

- Typically, if a borrow is not generated, the CARRY is set.

As shown in Figure 5−6, the SUB instruction with a 16-bit shift (shown fol-
lowing this paragraph) is an exception because it only resets the carry bit.
This allows the D-unit ALU to generate the appropriate carry when sub-
tracting to the lower or upper half of the accumulator causes a borrow.
Mnemonic instruction: SUB Smem << #16, ACx, ACy
Algebraic instruction: ACy = ACx − (Smem << #16)
Figure 5−6 shows the effect of subtractions on the CARRY bit.

Example 5−7. 64-Bit Subtraction

;**********************************************************************
; 64−Bit Subtraction Pointer assignments:
;
; X3 X2 X1 X0 AR1 −> X3 (even address)
; − Y3 Y2 Y1 Y0 X2
; −−−−−−−−−−−−−− X1
; W3 W2 W1 W0 X0
; AR2 −> Y3 (even address)
; Y2
; Y1
; Y0
; AR3 −> W3 (even address)
; W2
; W1
; W0
;
;**********************************************************************

MOV40 dbl(*AR1(#2)), AC0 ; AC0 = X1 X0

SUB dbl(*AR2(#2)), AC0 ; AC0 = X1 X0 − Y1 Y0
MOV AC0, dbl(*AR3(#2)) ; Store W1 W0.
MOV40 dbl (*AR1), AC0 ; AC0 = X3 X2
SUB uns(*AR2(#1)), BORROW, AC0 ; AC0 = X3 X2 − 00 Y2 − BORROW
SUB *AR2 << #16, AC0 ; AC0 = X3 X2 − Y3 Y2 − BORROW
MOV AC0, dbl(*AR3) ; Store W3 W2.

Note: The algebraic instructions code example for 64-Bit Subtraction is shown in Example B−28 on page B-31.

5-12
 Extended-Precision Addition and Subtraction

Figure 5−6. Effect on CARRY of Subtraction Operations

C MSB LSB C MSB LSB
X 0 0 0 0 0 0 0 0 0 0 ACx X F F 0 0 0 0 0 0 0 0 ACx
− 1 −F F F F F F F F F F
0 F F F F F F F F F F 0 0 0 0 0 0 0 0 0 0 1

C MSB LSB C MSB LSB

X 0 0 7 F F F F F F F ACx X 0 0 7 F F F F F F F ACx
− 1 −F F F F F F F F F F
1 0 0 7 F F F F F F E C F F 8 0 0 0 0 0 0 0

C MSB LSB C MSB LSB

X F F 8 0 0 0 0 0 0 0 ACx X F F 8 0 0 0 0 0 0 0 ACx
− 1 −F F F F F F F F F F
1 F F 7 F F F F F F F 0 F F 8 0 0 0 0 0 0 1

ACy = ACx − Smem − BORROW

C MSB LSB C MSB LSB
0 0 0 0 0 0 0 0 0 0 0 ACx 0 F F F F F F F F F F ACx
− 0 − 0
0 F F F F F F F F F F ACy 1 F F F F F F F F F E

ACy = ACx − (Smem << 16)

C MSB LSB C MSB LSB
1 F F 8 0 0 0 F F F F ACx 0 F F 8 0 0 0 F F F F ACx
−0 0 0 0 0 1 0 0 0 0 −F F F F F F 0 0 0 0
0 0 0 7 F F F F F F F ACy 0 F F 8 0 0 1 F F F F
C keeps a zero value, even
when a borrow is generated.

Fixed-Point Arithmetic 5-13

Extended-Precision Multiplication

5.3 Extended-Precision Multiplication

Extended precision multiplication (operands larger than 16 bit) can be per-

formed using basic C55x instructions. The C55x instruction set provides the
user with a very flexible set of 16-bit multiply instructions that accept signed
and unsigned operands and with a very efficient set of multiply-and-accumu-
late instructions that shift the value of the accumulator before adding it to the
multiplication result. Figure 5−5 shows how two 32-bit numbers yield a 64-bit
product.

Figure 5−7. 32-Bit Multiplication

X1X1 X0
X0

Y1Y1 Y0
Y0
×

Unsigned multiplication X0 x Y0

Signed/unsigned multiplication X1 x Y0

Signed/unsigned multiplication X0 x Y1

Signed multiplication X1 x Y1

Final 64-bit result W3 W2 W1 W0

Example 5−8 shows that a multiplication of two 32-bit integer numbers re-
quires one multiplication, two multiply/accumulate/shift operations, and a mul-
tiply/accumulate operation. The product is a 64-bit integer number.
Example 5−9 shows a fractional multiplication. The operands are in Q31 for-
mat, while the product is in Q31 format.

5-14
 Extended-Precision Multiplication

Example 5−8. 32-Bit Integer Multiplication

;****************************************************************
; This routine multiplies two 32−bit signed integers, giving a
; 64−bit result. The operands are fetched from data memory and the
; result is written back to data memory.
;
; Data Storage: Pointer Assignments:
; X1 X0 32−bit operand AR0 −> X1
; Y1 Y0 32−bit operand X0
; W3 W2 W1 W0 64−bit product AR1 −> Y1
; Y0
; Entry Conditions: AR2 −> W0
; SXMD = 1 (sign extension on) W1
; SATD = 0 (no saturation) W2
; FRCT = 0 (fractional mode off) W3
;
; RESTRICTION: The delay chain and input array must be
; long-word aligned.
;***************************************************************

AMAR *AR0+ ; AR0 points to X0

|| AMAR *AR1+ ; AR1 points to Y0
MPYM uns(*AR0), uns(*AR1), AC0 ; ACO = X0*Y0
MOV AC0,*AR2+ ; Save W0
MACM *AR0+, uns(*AR1−), AC0 >> #16, AC0 ; AC0 = X0*Y0>>16 + X1*Y0
MACM uns(*AR0−), *AR1, AC0 ; AC0 = X0*Y0>>16 + X1*Y0 + X0*Y1
MOV AC0, *AR2+ ; Save W1
MACM *AR0, *AR1, AC0 >> #16, AC0 ; AC0 = AC0>>16 + X1*Y1
MOV AC0, *AR2+ ; Save W2
MOV HI(AC0), *AR2 ; Save W3

Note: The algebraic instructions code example for 32-Bit Integer Multiplication is shown in Example B−29 on page B-32.

Fixed-Point Arithmetic 5-15

Extended-Precision Multiplication

Example 5−9. 32-Bit Fractional Multiplication

;**************************************************************************
; This routine multiplies two Q31 signed integers, resulting in a
; Q31 result. The operands are fetched from data memory and the
; result is written back to data memory.
;
; Data Storage: Pointer Assignments:
; X1 X0 Q31 operand AR0 −> X1
; Y1 Y0 Q31 operand X0
; W1 W0 Q31 product AR1 −> Y1
; Y0
; Entry Conditions: AR2 −> W1 (even address)
; SXMD = 1 (sign extension on) W0
; SATD = 0 (no saturation)
; FRCT = 1 (shift result left by 1 bit)
;
; RESTRICTION: W1 W0 is aligned such that W1 is at an even address.
;***************************************************************************
AMAR *AR0+ ; AR0 points to X0
MPYM uns(*AR0−), *AR1+, AC0 ; AC0 = X0*Y1
MACM *AR0, uns(*AR1−), AC0 ; AC0 =X0*Y1 + X1*Y0
MACM *AR0, *AR1, AC0 >> #16, AC0 ; AC0 = AC0>>16 + X1*Y1
MOV AC0, dbl(*AR2) ; Save W1 W0

Note: The algebraic instructions code example for 32-Bit Fractional Multiplication is shown in Example B−30 on page B-33.

5-16
 Division

5.4 Division

Binary division is the inverse of multiplication. Multiplication consists of a se-

ries of shift and add operations, while division can be broken into a series of
subtract and shift operations. On the C55x DSP you can implement this kind
of division by repeating a form of the conditional subtract (SUBC) instruction.

In a fixed-point processor, the range of the numbers we can use is limited by

the number of bits and the convention we use to represent these numbers. For
example, with a 16-bit unsigned representation, it is not possible to represent
a number larger than 216 − 1 (that is, 65 535). There can be problems with divi-
sion operations that require computing the inverse of a very small number.
Some digital signal processing algorithms may require integer or fractional di-
vision operations that support a large range of numbers. This kind of division
can be implemented with the conditional subtract (SUBC) instruction.

The difference between integers and fractional numbers is so great in a fixed-

point architecture that it requires different algorithms to perform the division
operation. Section 5.4.1 shows how to implement signed and unsigned integer
division. Section 5.4.2 (page 5-23) describes fractional division.

5.4.1 Integer Division

To prepare for a SUBC integer division, place a 16-bit positive dividend in an

accumulator. Place a 16-bit positive divisor in memory. When you write the
SUBC instruction, make sure that the result will be in the same accumulator
that supplies the dividend; this creates a cumulative result in the accumulator
when the SUBC instruction is repeated. Repeating the SUBC instruction 16
times produces a 16-bit quotient in the low part of the accumulator (bits 15−0)
and a remainder in the high part of the accumulator (bits 31−16). During each
execution of the conditional subtract instruction:

1) The 16-bit divisor is shifted left by 15 bits and is subtracted from the value
in the accumulator.

2) If the result of the subtraction is greater than or equal to 0, the result is

shifted left by 1 bit, added to 1, and stored in the accumulator. If the result
of the subtraction is less than 0, the result is discarded and the value in the
accumulator is shifted left by 1 bit.

The following examples show the implementation of the signed/unsigned inte-

ger division using the SUBC instruction.

Fixed-Point Arithmetic 5-17

Division

5.4.1.1 Examples of Unsigned Integer Division

Example 5−10 shows how to use the SUBC instruction to implement unsigned
division with a 16-bit dividend and a 16-bit divisor.

Example 5−10. Unsigned, 16-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend Divisor ) Dividend
; AR1 −> Divisor
; AR2 −> Quotient
; AR3 −> Remainder
;
; Algorithm notes:
; − Unsigned division, 16−bit dividend, 16−bit divisor
; − Sign extension turned off. Dividend & divisor are positive numbers.
; − After division, quotient in AC0(15−0), remainder in AC0(31−16)
;***************************************************************************

BCLR SXMD ; Clear SXMD (sign extension off)

MOV *AR0, AC0 ; Put Dividend into AC0
RPT #(16 − 1) ; Execute subc 16 times
SUBC *AR1, AC0, AC0 ; AR1 points to Divisor
MOV AC0, *AR2 ; Store Quotient
MOV HI(AC0), *AR3 ; Store Remainder

Note: The algebraic instructions code example for Unsigned, 16-Bit Integer Division is shown in Example B−31 on page B-33.

Example 5−11 shows how to implement unsigned division with a 32-bit divi-
dend and a 16-bit divisor. The code uses two phases of 16-bit by 16-bit integer
division. The first phase takes as inputs the high 16 bits of the 32-bit dividend
and the 16-bit divisor. The result in the low half of the accumulator is the high
16 bits of the quotient. The result in the high half of the accumulator is shifted
left by 16 bits and added to the lower 16 bits of the dividend. This sum and the
16-bit divisor are the inputs to the second phase of the division. The lower 16
bits of the resulting quotient is the final quotient and the resulting remainder
is the final remainder.

5-18
 Division

Example 5−11. Unsigned, 32-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend high half Divisor ) Dividend
; Dividend low half
; ...
; AR1 −> Divisor
; ...
; AR2 −> Quotient high half
; Quotient low half
; ...
; AR3 −> Remainder
;
; Algorithm notes:
; − Unsigned division, 32−bit dividend, 16−bit divisor
; − Sign extension turned off. Dividend & divisor are positive numbers.
; − Before 1st division: Put high half of dividend in AC0
; − After 1st division: High half of quotient in AC0(15−0)
; − Before 2nd division: Put low part of dividend in AC0
; − After 2nd division: Low half of quotient in AC0(15−0) and
; Remainder in AC0(31−16)
;***************************************************************************

BCLR SXMD ; Clear SXMD (sign extension off)

MOV *AR0+, AC0 ; Put high half of Dividend in AC0
|| RPT #(15 − 1) ; Execute subc 15 times
SUBC *AR1, AC0, AC0 ; AR1 points to Divisor
SUBC *AR1, AC0, AC0 ; Execute subc final time
|| MOV #8, AR4 ; Load AR4 with AC0_L memory address
MOV AC0, *AR2+ ; Store high half of Quotient
MOV *AR0+, *AR4 ; Put low half of Dividend in AC0_L
RPT #(16 − 1) ; Execute subc 16 times
SUBC *AR1, AC0, AC0
MOV AC0, *AR2+ ; Store low half of Quotient
MOV HI(AC0), *AR3) ; Store Remainder
BSET SXMD ; Set SXMD (sign extension on)

Note: The algebraic instructions code example for Unsigned, 32-Bit by 16-Bit Integer Division is shown in Example B−32 on
page B-34.

Fixed-Point Arithmetic 5-19

Division

5.4.1.2 Examples of Signed Integer Division

Some applications might require doing division with signed numbers instead
of unsigned numbers. The conditional subtract instruction works only with pos-
itive integers. The signed integer division algorithm computes the quotient as
follows:

1) The sign of the quotient is determined and preserved in AC0.

2) The quotient of the absolute values of the dividend and the divisor is deter-
mined using repeated conditional subtract instructions.

3) The negative of the result is computed if required, according to the sign

of AC0.

Example 5−12 shows the implementation of division with a signed 16-bit divi-
dend and a 16-bit signed divisor, and Example 5−13 extends this algorithm to
handle a 32-bit dividend.

5-20
 Division

Example 5−12. Signed, 16-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend Divisor ) Dividend
; AR1 −> Divisor
; AR2 −> Quotient
; AR3 −> Remainder
;
; Algorithm notes:
; − Signed division, 16−bit dividend, 16−bit divisor
; − Sign extension turned on. Dividend and divisor can be negative.
; − Expected quotient sign saved in AC0 before division
; − After division, quotient in AC1(15−0), remainder in AC1(31−16)
;***************************************************************************

BSET SXMD ; Set SXMD (sign extension on)

MPYM *AR0, *AR1, AC0 ; Sign of (Dividend x Divisor) should be
; sign of Quotient
MOV *AR1, AC1 ; Put Divisor in AC1
ABS AC1, AC1 ; Find absolute value, |Divisor|
MOV AC1, *AR2 ; Store |Divisor| temporarily
MOV *AR0, AC1 ; Put Dividend in AC1
ABS AC1, AC1 ; Find absolute value, |Dividend|
RPT #(16 − 1) ; Execute subc 16 times
SUBC *AR2, AC1, AC1 ; AR2 −> |Divisor|
MOV HI(AC1), *AR3 ; Save Remainder
MOV AC1, *AR2 ; Save Quotient
SFTS AC1, #16 ; Shift quotient: Put MSB in sign position
NEG AC1, AC1 ; Negate quotient
XCCPART label, AC0 < #0 ; If sign of Quotient should be negative,
MOV HI(AC1), *AR2 ; replace Quotient with negative version
label:

Note: The algebraic instructions code example for Signed, 16-Bit by 16-Bit Integer Division is shown in Example B−33 on
page B-35.

Fixed-Point Arithmetic 5-21

Division

Example 5−13. Signed, 32-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: (Dividend and Quotient are long−word aligned)
; AR0 −> Dividend high half (NumH) (even address)
; Dividend low half (NumL)
; AR1 −> Divisor (Den)
; AR2 −> Quotient high half (QuotH) (even address)
; Quotient low half (QuotL)
; AR3 −> Remainder (Rem)
;
; Algorithm notes:
; − Signed division, 32−bit dividend, 16−bit divisor
; − Sign extension turned on. Dividend and divisor can be negative.
; − Expected quotient sign saved in AC0 before division
; − Before 1st division: Put high half of dividend in AC1
; − After 1st division: High half of quotient in AC1(15−0)
; − Before 2nd division: Put low part of dividend in AC1
; − After 2nd division: Low half of quotient in AC1(15−0) and
; Remainder in AC1(31−16)
;***************************************************************************

BSET SXMD ; Set SXMD (sign extension on)

MPYM *AR0, *AR1, AC0 ; Sign( NumH x Den ) is sign of actual result
MOV *AR1, AC1 ; AC1 = Den
ABS AC1, AC1 ; AC1 = abs(Den)
MOV AC1, *AR3 ; Rem = abs(Den) temporarily
MOV40 dbl(*AR0), AC1 ; AC1 = NumH NumL
ABS AC1, AC1 ; AC1 = abs(Num)
MOV AC1, dbl(*AR2) ; QuotH = abs(NumH) temporarily
; QuotL = abs(NumL) temporarily

MOV *AR2, AC1 ; AC1 = QuotH

RPT #(15 − 1) ; Execute subc 15 times
SUBC *AR3, AC1, AC1
SUBC *AR3, AC1, AC1 ; Execute subc final time
|| MOV #11, AR4 ; Load AR4 with AC1_L memory address
MOV AC1, *AR2+ ; Save QuotH
MOV *AR2, *AR4 ; AC1_L = QuotH
RPT #(16 − 1) ; Execute subc 16 times
SUBC *AR3, AC1, AC1
MOV AC1, *AR2− ; Save QuotL
MOV HI(AC1), *AR3 ; Save Rem

BCC skip, AC0 >= #0 ; If actual result should be positive, goto skip.
MOV40 dbl(*AR2), AC1 ; Otherwise, negate Quotient.
NEG AC1, AC1
MOV AC1, dbl(*AR2)

skip:
RET

Note: The algebraic instructions code example for Signed, 32-Bit by 16-Bit Integer Division is shown in Example B−34 on
page B-36.

5-22
 Division

5.4.2 Fractional Division

The algorithms that implement fractional division compute first an approxima-
tion of the inverse of the divisor (denominator) using different techniques such
as Taylor Series expansion, line of best fit, and successive approximation. The
result is then multiplied by the dividend (numerator). The C55x DSP function
library (see Chapter 8) implements this function under the name ldiv16.

To calculate the value of Ym this algorithm uses the successive approximation

method. The approximations are performed using the following equation:
2
Y m)1 + 2ĂY m * Y m ĂX norm

If we start with an initial estimate of Ym, then the equation will converge to a
solution very rapidly (typically in three iterations for 16-bit resolution). The ini-
tial estimate can either be obtained from a look-up table, from choosing a mid-
point, or simply from linear interpolation. The ldiv16 algorithm uses linear inter-
polation. This is accomplished by taking the complement of the least signifi-
cant bits of the Xnorm value.

Fixed-Point Arithmetic 5-23

Methods of Handling Overflows

5.5 Methods of Handling Overflows

An overflow occurs when the result of an arithmetical operation is larger than
the largest number that can be represented in the register that must hold the
result. Due to the 16-bit format, fixed-point DSPs provide a limited dynamic
range. You must manage the dynamic range of your application to avoid pos-
sible overflows. The overflow depends on the nature of the input signal and of
the algorithm in question.

5.5.1 Hardware Features for Overflow Handling

The C55x DSP offers several hardware features for overflow handling:

- Guard bits:

Each of the C55x accumulators (AC0, AC1, AC2, and AC3) has eight
guard bits (bits 39−32), which allow up to 256 consecutive multiply-and-
accumulate operations before an accumulator overflow.

- Overflow flags:

Each C55x accumulator has an associated overflow flag (see the following
table). When an operation on an accumulator results in an overflow, the
corresponding overflow flag is set.

- Saturation mode bits:

The DSP has two saturation mode bits: SATD for operations in the D unit of
the CPU and SATA for operations in the A unit of the CPU. When the SATD
bit is set and an overflow occurs in the D unit, the CPU saturates the result.
Regardless of the value of SATD, the appropriate accumulator overflow
flag is set. Although no flags track overflows in the A unit, overflowing re-
sults in the A unit are saturated when the SATA bit is set.
Saturation replaces the overflowing result with the nearest range bound-
ary. Consider a 16-bit register which has range boundaries of 8000h (larg-
est negative number) and 7FFFh (largest positive number). If an operation
generates a result greater than 7FFFh, saturation can replace the result
with 7FFFh. If a result is less than 8000h, saturation can replace the result
with 8000h.

5-24
 Methods of Handling Overflows

5.5.1.1 Overview of Overflow Handling Techniques

There are a number of general methodologies to handle overflows. Among the

methodologies are saturation, input scaling, fixed scaling, and dynamic scal-
ing. We will give an overview of these methodologies and will see some exam-
ples illustrating their application.

- Saturation:

One possible way to handle overflows is to use the hardware saturation

modes mentioned in section 5.5.1. However, saturation has the effect of
clipping the output signal, potentially causing data distortion and non-lin-
ear behavior in the system.

- Input scaling:

You can analyze the system that you want to implement and scale the in-
put signal, assuming worst conditions, to avoid overflow. However, this ap-
proach can greatly reduce the precision of the output.

- Fixed scaling:

You can scale the intermediate results, assuming worst conditions. This
method prevents overflow but also increases the system’s signal-to-noise
ratio.

- Dynamic scaling:

The intermediate results can be scaled only when needed. You can ac-
complish this by monitoring the range of the intermediate results. This
method prevents overflow but increases the computational requirements.

The next sections demonstrate these methodologies applied to FIR (finite im-
pulse response) filters, IIR (infinite impulse response) filters and FFTs (fast
Fourier transforms).

5.5.1.2 Scaling Methods for FIR Filters

The best way to handle overflow problems in FIR (finite impulse response) fil-
ters is to design the filters with a gain less than 1 to avoid having to scale the
input data. This method, combined with the guard bits available in each of the
accumulators, provides a robust way to handle overflows in the filters.

Fixed scaling and input scaling are not used due to their negative impact on
signal resolution (basically one bit per multiply-and-accumulate operation).
Dynamic scaling can be used for an FIR filter if the resulting increase in cycles
is not a concern. Saturation is also a common option for certain types of audio
signals.

Fixed-Point Arithmetic 5-25

Methods of Handling Overflows

5.5.1.3 Scaling Methods for IIR Filters

Fixed-point realization of an IIR (infinite impulse response) filter in cascaded
second-order phases is recommended to minimize the frequency response
sensitivity of high-order filters. In addition to round-off error due to filter coeffi-
cient quantization, overflow avoidance is critical due to the recursive nature of
the IIR filter.
Overflow between phases can be avoided by maintaining the intermediate val-
ue in the processor accumulator. However, overflow can happen at the internal
filter state (delay buffer) inside each phase. To prevent overflow at phase k, the
filter unit-pulse response f(n) must be scaled (feed forward path) by a gain fac-
tor Gk given by

Option1 : G k + ȍ absǒf (n)Ǔ

n

or

ȍǒabsǒf (n)Ǔ Ǔ
1ń2
2
OptionĂ2 : G k +
n

Option 1 prevents overflows, but at the expense of precision. Option 2 allows

occasional overflows but offers an improved precision. In general, these tech-
niques work well if the input signal does not have a large dynamic range.
Another method to handle overflow in IIR filters is to use dynamic scaling. In
this approach, the internal filter states are scaled down by half only if an over-
flow is detected at each phase. The result is a higher precision but at the ex-
pense of increased MIPS.

5.5.1.4 Scaling Methods for FFTs

In FFT (Fast Fourier Transform) operations, data will grow an average of one
bit on the output of each butterfly. Input scaling will require shifting the data in-
put by log n (n = size of FFT) that will cause a 6(log n) dB loss even before com-
puting the FFT. In fixed scaling, the output of the butterfly will be scaled by 2
at each phase. This is probably the most common scaling approach for FFTs
because it is simple and has a better sound-to-noise ratio (SNR). However, for
larger FFTs this scaling may cause information loss.
Another option is to implement a dynamic scaling approach in which scaling
by 2 at each phase occurs only when bit growth occurs. In this case, an expo-
nent is assigned to the entire phase block (block floating-point method). When
scaling by 2 happens, the exponent is incremented by 1. At the end of the FFT,
the data are scaled up by the resulting exponent. Dynamic scaling provides
the best SNR but increases the FFT cycle count because you have to detect
bit growth and update the exponent accordingly.

5-26
 Chapter 6

!
"  

This chapter introduces the concept and the syntax of bit-reverse addressing.
It then explains how bit-reverse addressing can help to speed up a Fast Fourier
Transform (FFT) algorithm. To find code that performs complex and real FFTs
(forward and reverse) and bit-reversing of FFT vectors, see Chapter 8, TI C55x
DSPLIB.

Topic Page

6.1 Introduction to Bit-Reverse Addressing . . . . . . . . . . . . . . . . . . . . . . . . . 6-2

6.2 Using Bit-Reverse Addressing in FFT Algorithms . . . . . . . . . . . . . . . . 6-4
6.3 In-Place Versus Off-Place Bit-Reversing . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6.4 Using the C55x DSPLIB for FFTs and Bit-Reversing . . . . . . . . . . . . . . 6-7

6-1
Introduction to Bit-Reverse Addressing

6.1 Introduction to Bit-Reverse Addressing

Bit-reverse addressing is a special type of indirect addressing. It uses one of

the auxiliary registers (AR0−AR7) as a base pointer of an array and uses tem-
porary register 0 (T0) as an index register. When you add T0 to the auxiliary
register using bit-reverse addressing, the address is generated in a bit-re-
versed fashion, with the carry propagating from left to right instead of from right
to left.

Table 6−1 shows the syntaxes for each of the two bit-reversed addressing
modes supported by the TMS320C55x (C55x) DSP.

Table 6−1. Syntaxes for Bit-Reverse Addressing Modes

Operand
Syntax Function Description
*(ARx −T0B) address = ARx After access, T0 is subtracted from ARx with
ARx = (ARx − T0) reverse carry (rc) propagation.

*(ARx + T0B) address = ARx After access, T0 is added to ARx with reverse
ARx = (ARx + T0) carry (rc) propagation.

Assume that the auxiliary registers are 8 bits long, that AR2 represents the
base address of the data in memory (0110 0000b), and that T0 contains the
value 0000 1000b (decimal 8). Example 6−1 shows a sequence of modifica-
tions of AR2 and the resulting values of AR2.

Table 6−2 shows the relationship between a standard bit pattern that is repeat-
edly incremented by 1 and a bit-reversed pattern that is repeatedly increm-
ented by 1000b with reverse carry propagation. Compare the bit-reversed pat-
tern to the 4 LSBs of AR2 in Example 6−1.

Example 6−1. Sequence of Auxiliary Registers Modifications in Bit-Reversed Addressing

*(AR2+T0B) ;AR2 = 0110 0000 (0th value)

*(AR2+T0B) ,AR2 = 0110 1000 (1st value)
*(AR2+T0B) ;AR2 = 0110 0100 (2nd value)
*(AR2+T0B) ;AR2 = 0110 1100 (3rd value)
*(AR2+T0B) ;AR2 = 0110 0010 (4th value)
*(AR2+T0B) ;AR2 = 0110 1010 (5th value)
*(AR2+T0B) ;AR2 = 0110 0110 (6th value)
*(AR2+T0B) ;AR2 = 0110 1110 (7th value)

6-2
 Introduction to Bit-Reverse Addressing

Table 6−2. Bit-Reversed Addresses

Step Bit Pattern Bit-Reversed Pattern Bit-Reversed Step

0 0000 0000 0
1 0001 1000 8
2 0010 0100 4
3 0011 1100 12
4 0100 0010 2
5 0101 1010 10
6 0110 0110 6
7 0111 1110 14
8 1000 0001 1
9 1001 1001 9
10 1010 0101 5
11 1011 1101 13
12 1100 0011 3
13 1101 1011 11
14 1110 0111 7
15 1111 1111 15

Bit-Reversed Addressing 6-3

Using Bit-Reverse Addressing In FFT Algorithms

6.2 Using Bit-Reverse Addressing In FFT Algorithms

Bit-reversed addressing enhances execution speed for Fast-Fourier Trans-
form (FFT) algorithms. Typical FFT algorithms either take an in-order vector
input and produce a bit-reversed vector output or take a bit-reversed vector
input and produce an in-order vector output. In either case, bit-reverse ad-
dressing can be used to resequence the vectors. Figure 6−1 shows a flow
graph of an 8-point decimation-in-frequency FFT algorithm with a bit-reversed
input and an in-order output.

Figure 6−1. FFT Flow Graph Showing Bit-Reversed Input and In-Order Output
x(0) X(0)
W 0
N
x(4) X(1)
−1
W 0
N
x(2) X(2)
−1
W 2 W 0
x(6) N N X(3)
−1 −1
W 0
x(1) N X(4)
−1
W 1 W 0
N N
x(5) X(5)
−1 −1
W2 W0
N N
x(3) X(6)
−1 −1 0
W3 W2 W
N
N N
x(7) X(7)
−1 −1 −1

Consider a complex FFT of size N (that is, an FFT with an input vector that con-
tains N complex numbers). You can bit-reverse either the input or the output
vectors by executing the following steps:

1) Write 0 to the ARMS bit of status register 2 to select the DSP mode for AR
indirect addressing. (Bit-reverse addressing is not available in the control
mode of AR indirect addressing.) Then use the .arms_off directive to notify
the assembler of this selection.

2) Use Table 6−3 to determine how the base pointer of the input array must
be aligned to match the given vector format. Then load an auxiliary register
with the proper base address.

3) Consult Table 6−3 to properly load the index register, T0.

4) Ensure that the entire array fits within a 64K boundary (the largest possible
array addressable by the 16-bit auxiliary register).

6-4
 Using Bit-Reverse Addressing In FFT Algorithms

As an example of how to use Table 6−3, suppose you need to bit-reverse a

vector with N = 64 complex elements in the Re-Im-Re-Im format. The (n + 1)
least significant bits (LSBs) of the base address must be 0s, where

(n + 1) = (log2 N + 1) = (log2 64 + 1) = (6 + 1) = 7 LSBs

Therefore, AR0 must be loaded with a base address of this form (Xs are don’t
cares):

AR0 = XXXX XXXX X000 0000b

The index loaded into T0 is equal to the number of elements:

T0 = 2n = 26 = 64

Table 6−3. Typical Bit-Reverse Initialization Requirements

Vector Format in T0 Initialization Alignment of

Memory Value Vector Base Address

Re-Im-Re-Im. The real and 2n (n+1) LSBs must be 0s,

imaginary parts of each
where n = log2 N where n = log2 N
complex element are
stored at consecutive
memory locations. For ex-
ample:
Real
Imaginary
...
Real
Imaginary

Re-Re...Im-Im. The real 2(n−1) n LSBs must be 0s,

and imaginary data are
where n = log2 N where n = log2 N
stored in separate arrays.
For example:
Real
Real
...
Imaginary
Imaginary

Bit-Reversed Addressing 6-5

In-Place Versus Off-Place Bit-Reversing

6.3 In-Place Versus Off-Place Bit-Reversing

You can have a bit-reversed array write over the original array (in-place
bit-reversing) or you can place the bit-reversed array in a separate place in
memory (off-place bit-reversing). Example 6−2 shows assembly language
code for an off-place bit-reverse operation. An input array of N complex ele-
ments pointed to by AR0 is bit-reversed into an output array pointed to by AR1.
The vector format for input and output vectors is assumed to be Re-Im-Re-Im.
Each element (Re-Im) of the input array is loaded into AC0. Then the element
is transferred to the output array using bit-reverse addressing. Each time the
index in T0 is added to the address in AR1, the addition is done with carries
propagating from left to right, instead of from right to left.

Although it requires twice the memory, off-place bit-reversing is faster than

in-place bit-reversing. Off-place bit-reversing requires 2 cycles per complex
data point, while in-place bit-reversing requires approximately 4 cycles per
complex data point.

Example 6−2. Off-Place Bit Reversing of a Vector Array (in Assembly)

;...
BCLR ARMS ; reset ARMS bit to allow bit−reverse addressing
.arms_off ; notify the assembler of ARMS bit = 0
;...
off_place:
RPTBLOCAL LOOP
MOV dbl(*AR0+), AC0 ; AR0 points to input array
LOOP: MOV AC0, dbl(*(AR1+T0B)) ; AR1 points to output array
; T0 = NX = number of complex elements in
; array pointed to by AR0

Note: The algebraic instructions code example for Off-Place Bit Reversing of a Vector Array (in Assembly) is shown in
Example B−35 on page B-37.

6-6
 Using the C55x DSPLIB for FFTs and Bit-Reversing

6.4 Using the C55x DSPLIB for FFTs and Bit-Reversing

The C55x DSP function library (DSPLIB) offers C-callable DSP assembly-opti-
mized routines. Among these are a bit-reversing routine (cbrev()) and complex
and real FFT routines.

Example 6−3 shows how you can invoke the cbrev() DSPLIB function from C
to do in-place bit-reversing. The function bit-reverses the position of in-order
elements in a complex vector x and then computes a complex FFT of the bit-re-
versed vector. The function uses in-place bit-reversing. See Chapter 8 for an
introduction to the C55x DSPLIB.

Example 6−3. Using DSPLIB cbrev() Routine to Bit Reverse a Vector Array (in C)

#define NX 64
short x[2*NX] ;
short scale = 1 ;

void main(void)
{
;...
cbrev(x,x,NX) // in−place bit−reversing on input data (Re−Im format)
cfft(x,NX,scale) // 64−point complex FFT on bit−reversed input data with
// scaling by 2 at each phase enabled
;...
}

Note: This example shows portions of the file cfft_t.c in the TI C55x DSPLIB (introduced in Chapter 8).

Bit-Reversed Addressing 6-7

6-8
 Chapter 7

#$ 

This chapter presents examples of efficient implementations of some common

signal processing and telecommunications functions. These examples illus-
trate the use of some application-specific instructions on the TMS320C55x
(C55x) DSP. (Most of the examples in this chapter use instructions from the
mnemonic instruction set, corresponding algebraic instruction set examples
are shown in Appendix A.)

Topic Page

7.1 Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN) . . . . . . . . . . 7-2

7.2 Adaptive Filtering (LMS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6
7.3 Convolutional Encoding (BFXPA, BFXTR) . . . . . . . . . . . . . . . . . . . . . . 7-10
7.4 Viterbi Algorithm for Channel Decoding
(ADDSUB, SUBADD, MAXDIFF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-16

7-1
Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN)

7.1 Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN)

FIR (finite impulse response) filters are often used in telecommunications ap-
plications because they are unconditionally stable and they may be designed
to preserve important phase information in the processed signal. A linear
phase FIR provides a phase shift that varies in proportion to the input frequen-
cy and requires that the impulse response be symmetric: h(n) = h(N−n).

Another class of FIR filter is the antisymmetric FIR: h(n) = −h(N−n). A com-
mon example is the Hilbert transformer, which shifts positive frequencies by
+90 degrees and negative frequencies by −90 degrees. Hilbert transformers
may be used in applications, such as modems, in which it is desired to cancel
lower sidebands of modulated signals.

Figure 7−1 gives examples of symmetric and antisymmetric filters, each with
eight coefficients (a0 through a7). Both symmetric and antisymmetric filters
may be of even or odd length. However, even-length symmetric filters lend
themselves to computational shortcuts which will be described in this section.
It is sometimes possible to reformulate an odd-length filter as a filter with one
more tap, to take advantage of these constructs.

Because (anti)symmetric filters have only N/2 distinct coefficients, they may
be folded and performed with N/2 additions (subtractions) and N/2 multiply-
and-accumulate operations. Folding means that pairs of elements in the delay
buffer which correspond to the same coefficient are pre-added(subtracted)
prior to multiplying and accumulating.

The C55x DSP offers two different ways to implement symmetric and asym-
metric filters. This section shows how to implement these filters using specific
instructions, FIRS and FIRSN. To see how to implement symmetric and asym-
metric filters using the dual-MAC hardware, see section 4.1.1, Implicit Algo-
rithm Symmetry, which begins on page 4-4. The firs/firsn implementation and
the dual-MAC implementation are equivalent from a throughput standpoint.

7-2
 Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN)

Figure 7−1. Symmetric and Antisymmetric FIR Filters

Symmetric Antisymmetric

a0 a1 a2 a3

a0 a1 a2 a3 a4 a5 a6 a7 a4 a5 a6 a7

Symmetric FIR Filter a[n]: Antisymmetric FIR Filter a[n]:

aā [ā nā ] + aā [ā N * n] where 0 t n t N aā [ā nā ] + * aā [ā N * n] where 0 t n t N
2 2
Symmetric FIR filter output:
Y(n) + aā [0]ā ǒā xā [n * 7] ) xā [0]ā Ǔ ) aā [1]ā ǒā xā [n * 6] ) xā [1]ā Ǔ ) aā [2]ā ǒā xā [n * 5] ) xā [2]ā Ǔ ) aā [3]ā ǒā xā [n * 4] ) xā [3]ā Ǔ

Antisymmetric FIR filter output:

Y(n) + aā [0]ā ǒā xā [n * 7] * xā [0]ā Ǔ ) aā [1]ā ǒā xā [n * 6] * xā [1]ā Ǔ ) aā [1]ā ǒā xā [n * 5] * xā [2]ā Ǔ ) aā [3]ā ǒā xā [n * 4] * xā [3]ā Ǔ

Definitions:
a = Filter coefficient Y = Filter output
n = Sample index x = Filter input data value
N = Number of filter taps

7.1.1 Symmetric FIR Filtering With the firs Instruction

The C55x instruction for symmetric FIR filtering is:

firs(Xmem,Ymem,Cmem,ACx,ACy)

This instruction performs two parallel operations: a multiply-and-accumulate

(MAC) operation, and an addition. The firs() instruction performs the following
parallel operations:

ACy = ACy + (ACx * Cmem),

ACx = (Xmem << #16) + (Ymem << #16)

The first operation performs a multiplication and an accumulation in a MAC unit

of the CPU. The input operands of the multiplier are the content of ACx(32−16)
and a data memory operand, which is addressed using the coefficient ad-
dressing mode and is sign extended to 17 bits. Table 7−1 explains the oper-
ands necessary for the operation.

Application-Specific Instructions 7-3

Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN)

Table 7−1. Operands to the firs or firsn Instruction

Operand(s) Description

Xmem and Ymem One of these operands points to the newest value in the
delay buffer. The other points to the oldest value in the
delay buffer.

Cmem This operand points to the filter coefficient.

ACx ACx is one of the four accumulators (AC0−AC3). It holds

the sum of the two delayed input values referenced by
Xmem and Ymem.

ACy ACy is one of the four accumulators (AC0−AC3) but is not

the same accumulator as ACx. ACy holds the output of
each filter tap. After all the filter taps have been performed,
ACy holds the final result.

7.1.2 Antisymmetric FIR Filtering With the firsn Instruction

The antisymmetric FIR is the same as the symmetric FIR except that the pre-
addition of sample pairs is replaced with a pre-subtraction. The C55x instruc-
tion for antisymmetric FIR filtering is:

firsn(Xmem,Ymem,Cmem,ACx,ACy)

This instruction performs two parallel operations: a multiply-and-accumulate

(MAC) operation, and a subtraction. The firsn() instruction performs the follow-
ing parallel operations:

ACy = ACy + (ACx * Cmem),

ACx = (Xmem << #16) − (Ymem << #16)

The first operation performs a multiplication and an accumulation in a MAC unit

of the CPU. The input operands of the multiplier are the content of ACx(32−16)
and a data memory operand, which is addressed using the coefficient ad-
dressing mode and is sign extended to 17 bits. Table 7−1 (page 7-4) explains
the operands necessary for the operation.

7.1.3 Implementation of a Symmetric FIR Filter on the TMS320C55x DSP

The C55x DSPLIB features an efficient implementation of the Symmetric FIR

on the C55x device. Example 7−1 presents the kernel of that implementation
to illustrate the usage of the firs instruction.

7-4
 Symmetric and Asymmetric FIR Filtering (FIRS, FIRSN)

Example 7−1. Symmetric FIR Filter

;
; Start of outer loop
;−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
localrepeat { ; Start the outer loop

; Get next input value

*db_ptr1 = *x_ptr+ ; x_ptr: pointer to input data buffer

; db_ptr1: pointer to newest input value

; Clear AC0 and pre−load AC1 with the sum of the 1st and last inputs
||AC0 = #0;

; 1st and last inputs

AC1 = (*db_ptr1+ << #16) + (*db_ptr2− << #16)
; Inner loop
||repeat(inner_cnt)
firs(*db_ptr1+, *db_ptr2−, *h_ptr+, AC1, AC0)

; 2nd to last iteration has different pointer adjustment

firs(*(db_ptr1−T0), *(db_ptr2+T1), coef(*h_ptr+), AC1, AC0)

; Last iteration is a MAC with rounding

AC0 = rnd(AC0 + (*h_ptr+ * AC1))

; Store result to memory

*r_ptr+ = HI(AC0) ;store Q15 value to memory

} ;end of outer loop

Note: This example shows portions of the file firs.asm in the TI C55x DSPLIB (introduced in Chapter 8).

Application-Specific Instructions 7-5

Adaptive Filtering (LMS)

7.2 Adaptive Filtering (LMS)

Some applications for adaptive FIR (finite impulse response) and IIR (infinite
impulse response) filtering include echo and acoustic noise cancellation. In
these applications, an adaptive filter tracks changing conditions in the environ-
ment. Although in theory, both FIR and IIR structures can be used as adaptive
filters, stability problems and the local optimum points of IIR filters makes them
less attractive for this use. Therefore, FIR filters are typically used for practical
adaptive filter applications. The least mean square (LMS), local block-repeat,
and parallel instructions on the C55x DSP can be used to efficiently implement
adaptive filters. The block diagram of an adaptive FIR filter is shown in
Figure 7−2.

Figure 7−2. Adaptive FIR Filter Implemented With the

Least-Mean-Squares (LMS) Algorithm

x(n)
z−1 z−1 z−1

b0 b1 bN−1

Desired
response
d(n)

+ + +
y(n)

+
ȍ −
ȍ

LMS

Two common algorithms employed for least mean squares adaptation are the
non-delayed LMS and the delayed LMS algorithm. When compared to non-
delayed LMS, the more widely used delayed LMS algorithm has the advan-
tage of greater computational efficiency at the expense of slightly relaxed con-
vergence properties. Therefore, section 7.2.1 describes only the delayed LMS
algorithm.

7-6
 Adaptive Filtering (LMS)

7.2.1 Delayed LMS Algorithm

In the delayed LMS, the convolution is performed to compute the output of the
adaptive filter:

ȍ b Ăxā (n * k)
N*1
y(n) + k
k+0

where

y = Filter output
n = Sample index
k = Delay index
N = Number of filter taps
bk = Adaptive coefficient
x = Filter input data value

The value of the error is computed and stored to be used in the next invocation:

e(n) + d(n) * y(n)

where

e = Error
d = Desired response
y = Actual response (filter output)

The coefficients are updated based on an error value computed in the previous
invocation of the algorithm (β is the conversion constant):

b kā (n ) 1) + b kā (n) ) 2ĂbĂeā (n * 1)Ăx(n * k * 1)

The delayed LMS algorithm can be implemented with the LMS instruction—
lms(Xmem, Ymem, ACx, ACy)—which performs a multiply-and-accumulate
(MAC) operation and, in parallel, an addition with rounding:

ACy = ACy + (Xmem * Ymem),

ACx = rnd(ACx + (Xmem << #16)

Application-Specific Instructions 7-7

Adaptive Filtering (LMS)

The input operands of the multiplier are the content of data memory operand
Xmem, sign extended to 17 bits, and the content of data memory operand
Ymem, sign extended to 17 bits. One possible implementation would assign
the following roles to the operands of the LMS instruction:

Operand(s) Description

Xmem This operand points to the coefficient array.

Ymem This operand points to the data array.

ACx ACx is one of the four accumulators (AC0−AC3). ACx is used to

update the coefficients.

ACy ACy is one of the four accumulators (AC0−AC3) but is not the
same accumulator as ACx. ACy holds the output of the FIR filter.

An efficient implementation of the delayed LMS algorithm is available in the

C55x DSP function library (see Chapter 8). NO TAG shows the kernel of this
implementation.

7-8
 Adaptive Filtering (LMS)

Example 7−2. Delayed LMS Implementation of an Adaptive Filter

; ar_data: index in the delay buffer

; ar_input: pointer to input vector
; ar_coef: pointer to coefficient vector

StartSample:

; Clear AC0 for initial error term

MOV #0, AC1
||RPTBLOCAL OuterLoop−1
MOV *ar_input+, *ar_data+ ;copy input −> state(0)

; Place error term in T3

MOV HI(AC1), T3

; Place first update term in AC0

;...while clearing FIR value
MPYM *ar_data+, T3, AC0
||MOV #0, AC1

;AC0 = update coef

;AC1 = start of FIR output
LMS *ar_coef, *ar_data, AC0, AC1
||RPTLOCAL InnerLoop−1
MOV HI(AC0), *ar_coef+
||MPYM *ar_data+, T3, AC0

;AC0 = update coef

;AC1 = update of FIR output
LMS *ar_coef, *ar_data, AC0, AC1
InnerLoop:

; Store Calculated Output

MOV HI(AC0), *ar_coef+
||MOV rnd(HI(AC1)), *ar_output+

; AC2 is error amount

; Point to oldest data sample
SUB AC1, *ar_des+ << #16, AC2
||AMAR *ar_data+

; Place updated mu_error term in AC1

AMPYMR T_step, AC2, AC1
OuterLoop:

Note: The algebraic instructions code example for Delayed LMS Implementation of an Adaptive Filter is shown in
Example B−36 on page B-38.

Application-Specific Instructions 7-9

Convolutional Encoding (BFXPA, BFXTR)

7.3 Convolutional Encoding (BFXPA, BFXTR)

The goal in every telecommunication system is to achieve maximum data
transfer, using a minimum bandwidth, while maintaining an acceptable quality
of transmission. Convolutional codes are a forward error control (FEC) tech-
nique in which extra binary digits are added to the original information binary
digits prior to transmission to create a code structure which is resistant to er-
rors that may occur within the channel. A decoder at the receiver exploits the
code structure to correct any errors that may have occurred. The redundant
bits are formed by XORing the current bit with time-delayed bits within the past
K input sample history. This is effectively a 1-bit convolution sum; hence the
term convolutional encoder. The coefficients of the convolution sum are de-
scribed using polynomial notation. A convolutional code is defined by the fol-
lowing parameters:

n = Number of function generators

G0, G1, ... , Gn = Polynomials that define the convolutions of bit streams
K = Constraint length (number of delays plus 1)

The rate of the convolutional encoder is defined as R = 1/n. Figure 7−3 gives
an example of a convolutional encoder with K=5 and R = 1/2.

Figure 7−3. Example of a Convolutional Encoder

XOR
G0
bits

Input k0 k1 k2 k3
bits Z−1 Z−1 Z−1 Z−1 k4

XOR
G1
bits

The C55x DSP creates the output streams (G0 and G1) by XORing the shifted
input stream (see Figure 7−4).

7-10
 Convolutional Encoding (BFXPA, BFXTR)

Figure 7−4. Generation of an Output Stream G0

15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
0 1 0 1 1 0 0 1 0 1 1 1 0 0 1 0 x

0 1 0 1 1 0 0 1 0 1 1 1 0 0 1 0 + x >> 3

0 1 0 1 1 0 0 1 0 1 1 1 0 0 1 0 + x >> 4

12 bits of G0
x(11) + x(14) + x(15) x(0) + x(3) + x(4)

Example 7−3 shows an implementation of the output bit streams for the con-
volutional encoder of Figure 7−3.

Example 7−3. Generation of Output Streams G0 and G1

MOV #09H, AR3 ; AC0_H

MOV #in_bit_stream, AR1

; Load 32 bits into the accumulators

MOV *AR1+, AC0
ADD *AR1 << #16, AC0
MOV AC0, AC1

; Generate G0
XOR AC1 << #−1, AC0 ; A = A XOR B>>1
XOR AC1 << #−3, AC0 ; A = A XOR B>>3
MOV AC0, T0 ; Save G0

; Generate G1
XOR AC1 << #−1, AC0 ; A = A XOR B>>1
XOR AC1 << #−3, AC0 ; A = A XOR B>>3
XOR AC1 << #−4, AC0 ; A = A XOR B>>4 −−> AC0_L = G1

MOV T0, *AR3 ; AC0_H = G0 −−−−−−> AC0 = G0G1

Note: The algebraic instructions code example for Generation of Output Streams G0 and G1 is shown in Example B−37 on
page B-39.

Application-Specific Instructions 7-11

Convolutional Encoding (BFXPA, BFXTR)

7.3.1 Bit-Stream Multiplexing and Demultiplexing

After a bit stream is convolved by the various generating polynomials, the re-
dundant bits are typically multiplexed back into a single higher-rate bit stream.
The C55x DSP has dedicated instructions that allow the extraction or insertion
of a group of bits anywhere within a 32-bit accumulator. These instructions can
be used to greatly expedite the bit-stream multiplexing operation on convolu-
tional encoders.

Figure 7−5 illustrates the concept of bit multiplexing.

Figure 7−5. Bit Stream Multiplexing Concept

G0
Multiplexor G0G1
G1

The C55x DSP has a dedicated instruction to perform the multiplexing of the
bit streams:

dst = field_expand(ACx,k16)

This instruction executes in 1 cycle according to the following algorithm:

1) Clear the destination register.

2) Reset to 0 the bit index pointing within the destination register:

index_in_dst.

3) Reset to 0 the bit index pointing within the source accumulator:

index_in_ACx.

4) Scan the bit field mask k16 from bit 0 to bit 15, testing each bit. For each
tested mask bit:
If the tested bit is 1:
a) Copy the bit pointed to by index_in_ACx to the bit pointed to by
index_in_dst.
b) Increment index_in_ACx.
c) Increment index_in_dst, and test the next mask bit.
If the tested bit is 0:
Increment index_in_dst, and test the next mask bit.

Example 7−4 demonstrates the use of the field_expand() instruction.

7-12
 Convolutional Encoding (BFXPA, BFXTR)

Example 7−4. Multiplexing Two Bit Streams With the Field Expand Instruction

.asg AC0, G0 ; Assign G0 to register AC0.

.asg AC1, G1 ; Assign G1 to register AC1.
.asg AC2, Temp ; Assign Temp to register AC2.
.asg AC3, G0G1 ; Assign G0G1 to register AC3.

G0 = *get_G0 ; Load G0 stream.

G1 = *get_G1 ; Load G1 stream.
G0G1 = field_expand(G0, #5555h) ; Expand G0 stream.
Temp = G0G1 ; Temporarily store expanded G0 stream.
G0G1 = field_expand(G1, #AAAAh) ; Expand G1 stream
G0G1 = G0G1 | Temp ; Interleave expanded streams

G0G1 = field_expand(G0,#5555h)

5555h

ÉÉÉÉÉ
ÉÉÉÉÉÉÉÉÉÉÉÉÉ
ÉÉÉ
0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

G0(15−0) X X X X X X X
ÉÉÉÉÉ
X
ÉÉÉÉÉÉÉÉÉÉÉÉÉ
1 1
ÉÉÉ
0 0 1 1 1 1

ÉÉÉÉ ÉÉÉÉÉÉÉÉÉÉÉ ÉÉ
ÉÉÉÉ ÉÉÉÉÉÉÉÉÉÉÉ ÉÉ
0 1 0 1 0 0 0 0 0 1 0 0 0 1 0 1

G0G1 = field_expand(G1,#AAAAh)

AAAAh 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0

G1(15−0) X X X X X X X X 0 1 1 1 0 0 1 0

0 0 1 0 1 0 1 0 0 0 0 0 1 0 0 0

ÉÉÉÉÉÉÉÉÉ ÉÉÉÉÉÉÉÉÉ
G0G1 = G0G1 | Temp

0
ÉÉÉÉÉÉÉÉÉ ÉÉÉÉÉÉÉÉÉ
1 1 1 1 0 1 0 0 1 0 0 1 1 0 1

Application-Specific Instructions 7-13

Convolutional Encoding (BFXPA, BFXTR)

At the receiver G0 and G1 must be extracted by de-interleaving the G0 G1

stream. The DSP has dedicated instructions to perform the de-interleaving of
the bit streams:

dst = field_extract(ACx,k16)

This instruction executes in 1 cycle according to the following algorithm:

1) Clear the destination register.

2) Reset to 0 the bit index pointing within the destination register:

index_in_dst.

3) Reset to 0 the bit index pointing within the source accumulator:

index_in_ACx.

4) Scan the bit field mask k16 from bit 0 to bit 15, testing each bit. For each
tested mask bit:
If the tested bit is 1:
a) Copy the bit pointed to by index_in_ACx to the bit pointed to by
index_in_dst.
b) Increment index_in_dst.
c) Increment index_in_ACx, and test the next mask bit.
If the tested bit is 0:
Increment index_in_ACx, and test the next mask bit.

Example 7−5 demonstrates the use of the field_extract() instruction. The ex-
ample shows how to de-multiplex the signal created in Example 7−4.

7-14
 Convolutional Encoding (BFXPA, BFXTR)

Example 7−5. Demultiplexing a Bit Stream With the Field Extract Instruction

.asg T2, G0 ; Assign G0 to register T2.

.asg T3, G1 ; Assign G1 to register T3.
.asg AC0, G0G1 ; Assign G0G1 to register AC0.
.asg AR1, receive ; Assign receive to register AR1.

G0G1 = *receive ; Get bit stream.

G0 = field_extract(G0G1, #05555h) ; Extract G0 from bit stream.
G1 = field_extract(G0G1, #0AAAAh) ; Extract G1 from bit stream.

G0 = field_extract(G0G1,#5555h)

5555h 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

G0G1(15−0) 0 1 1 1 1 0 1 0 0 1 0 0 1 1 0 1

G0 X X X X X X X X 1 1 0 0 1 0 1 1

G1 = field_extract(G0G1,#AAAAh)

AAAAh 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0

G0G1(15−0) 0 1 1 1 1 0 1 0 0 1 0 0 1 1 0 1

G1 X X X X X X X X 0 1 1 1 0 0 1 0

Application-Specific Instructions 7-15

Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

7.4 Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

The Viterbi algorithm is widely used in communications systems for decoding
information that has been convolutionally encoded. The most computationally
intensive part of the routine is comprised of many add-compare-select (ACS)
iterations. Minimizing the time for each ACS calculation is important. For a giv-
en system, the number of ACS calculations depends on the constraint length
K and is equal to 2(K − 2). Thus, as K increases, the number of ACS calculations
increases exponentially. The C55x DSP can perform the ACS operation in 1
cycle, due to dedicated instructions that support the Viterbi algorithm.

The convolutional encoder depicted in Figure 7−3 (page 7-10) is used in the
global system for mobile communications (GSM) and is described by the fol-
lowing polynomials (K=5):

G 0Ă(x) + 1 ) x 3 ) x 4 G 1Ă(x) + 1 ) x ) x 3 ) x 4

The convolutionally encoded outputs are dependent on past data inputs.

Moreover, the contents of the encoder can be viewed as a finite state machine.
A trellis diagram can represent the allowable state transitions, along with their
corresponding path states. Decoding the data involves finding the optimal path
through the trellis, by iteratively selecting possible paths to each delay state,
for a given number of symbol time intervals. Two path metrics are calculated
by adding a local distance to two old metrics. A comparison is made and a new
path metric is selected from the two.

In the case of the GSM encoder, there are 16 possible states for every symbol
time interval. For rate 1/n systems, there is some inherent symmetry in the trel-
lis structure, which simplifies the calculations. The path states leading to a
delay state are complementary. That is, if one path has G0G1 = 00, the other
path has G0G1 = 11. This symmetry is based on the encoder polynomials and
is true for most systems. Two starting and ending complementary states can
be paired together, including all the paths between them, to form a butterfly
structure (see Figure 7−6). Hence, only one local distance is needed for each
butterfly; it is added and subtracted for each new state. Additionally, the old
metric values are the same for both updates, so address manipulation is mini-
mized.

7-16
 Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

Figure 7−6. Butterfly Structure for K = 5, Rate 1/2 GSM Convolutional Encoder

0000 00 0000
Old_Metric(0) New_Metric(0)
11

Old_Metric(1)
11
0001

00 New_Metric(8)
1000

Time
Symbol Symbol
Time 0 Time 1

Delay States Path States G0 G1 Input = 0 Input = 1

The following equation defines a local distance for the rate 1/2 GSM system:

LD + SD 0ĂG 0Ăǒ j Ǔ ) SD 1ĂG 1Ăǒ j Ǔ

where

SDx = Soft-decision input into the decoder

Gx(j) = Expected encoder output for the symbol interval j

Usually, the Gx(j)s are coded as signed antipodal numbers, meaning that “0”
corresponds to +1 and “1” corresponds to −1. This coding reduces the local
distance calculation to simple addition and subtraction.

As shown in Example 7−6, the DSP can calculate a butterfly quickly by using
its accumulators in a dual 16-bit computation mode. To determine the new path
metric j, two possible path metrics, 2j and 2j+1, are calculated in parallel with
local distances (LD and −LD) using the add-subtract (ADDSUB) instruction
and an accumulator. To determine the new path metric (j+2(K − 2)), the subtract-
add (SUBADD) instruction is also used, using the old path metrics plus local
distances stored in a separate accumulator. The MAXDIFF instruction is then
used on both accumulators to determine the new path metrics. The MAXDIFF
instruction compares the upper and lower 16-bit values for two given accumu-
lators, and stores the larger values in a third accumulator.

Application-Specific Instructions 7-17

Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

Example 7−6 shows two macros for Viterbi butterfly calculations. The add-
subtract (ADDSUB) and subtract-add (SUBADD) computations in the two
macros are performed in alternating order, which is based on the expected en-
coder state. A local distance is stored in the register T3 beforehand. The
MAXDIFF instruction performs the add-compare-select function in 1 cycle.
The updated path metrics are saved to memory by the next two lines of code.

Two 16-bit transition registers (TRN0 and TRN1) are updated with every com-
parison done by the MAXDIFF instruction, so that the selected path metric can
be tracked. TRN0 tracks the results from the high part data path, and TRN1
tracks the low part data path. These bits are later used in traceback, to deter-
mine the original uncoded data. Using separate transition registers allows for
storing the selection bits linearly, which simplifies traceback. In contrast, the
TMS320C54x (C54x) DSP has only one transition register, storing the
selection bits as 0, 8, 1, 9, etc. As a result, on the C54x DSP, additional lines
of code are needed to process these bits during traceback.

You can make the Viterbi butterfly calculations faster by implementing user-
defined instruction parallelism (see section 4.2, page 4-16) and software pipe-
lining. Example 7−7 (page 7-20) shows the inner loop of a Viterbi butterfly al-
gorithm. The algorithm places some instructions in parallel (||) in the CPU, and
the algorithm implements software pipelining by saving previous results at the
same time it performs new calculations. Other operations, such as loading the
appropriate local distances, are coded with the butterfly algorithm.

7-18
 Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

Example 7−6. Viterbi Butterflies for Channel Coding

BFLY_DIR_MNEM .MACRO
;new_metric(j)&(j+2^(K−2))
;
ADDSUB T3, *AR5+, AC0 ; AC0(39−16) = Old_Met(2*j)+LD
; AC0(15−0) = Old_met(2*j+1)−LD

SUBADD T3, *AR5+, AC1 ; AC1(39−16) = Old_Met(2*j)−LD

; AC1(15−0) = Old_met(2*j+1)+LD

MAXDIFF AC0, AC1, AC2, AC3 ; Compare AC0 and AC1

MOV AC2, *AR3+, *AR4+ ; Store the lower maxima (with AR3)
; and upper maxima (with AR4)
.ENDM

BFLY_REV_MNEM .MACRO
;new_metric(j)&(j+2^(K−2))
SUBADD T3, *AR5+, AC0 ; AC0(39−16) = Old_Met(2*j)−LD
; AC0(15−0) = Old_met(2*j+1)+LD

ADDSUB T3, *AR5+, AC1 ; AC1(39−16) = Old_Met(2*j)+LD

; AC1(15−0) = Old_met(2*j+1)−LD

MAXDIFF AC0, AC1, AC2, AC3 ; Compare AC0 and AC1

MOV AC2, *AR3+, *AR4+ ; Store the lower maxima (with AR3)
; and upper maxima (with AR4)
.ENDM

Note: The algebraic instructions code example for Viterbi Butterflies for Channel Coding is shown in Example B−38 on
page B-39.

Application-Specific Instructions 7-19

Viterbi Algorithm for Channel Decoding (ADDSUB, SUBADD, MAXDIFF)

Example 7−7. Viterbi Butterflies Using Instruction Parallelism

RPTBLOCAL end
butterfly:
ADDSUB T3, *AR0+, AC0 ; AC0(39−16) = Old_Met(2*j)+LD
; AC0(15−0) = Old_met(2*j+1)−LD
|| MOV *AR5+, AR7

SUBADD T3, *AR0+, AC1 ; AC1(39−16) = Old_Met(2*j)−LD

; AC1(15−0) = Old_met(2*j+1)+LD
|| MOV *AR6, T3 ; Load new local distance

MOV AC2, *AR2+, *AR2(T0) ; Store lower and upper maxima

; from previous MAXDIFF operation
|| MAXDIFF AC0, AC1, AC2, AC3 ; Compare AC0 and AC1

ADDSUB T3, *AR0+, AC0 ; AC0(39−16) = Old_Met(2*j)+LD

; AC0(15−0) = Old_met(2*j+1)−LD
|| MOV *AR5+, AR6

SUBADD T3, *AR0+, AC1 ; AC1(39−16) = Old_Met(2*j)−LD

; AC1(15−0) = Old_met(2*j+1)+LD
|| MOV *AR7, T3 ; Load new local distance

end MOV AC2, *AR2(T0), *AR2+ ; Store lower and upper maxima
; from previous MAXDIFF operation
|| MAXDIFF AC0, AC1, AC2, AC3 ; Compare AC0 and AC1

Note: The algebraic instructions code example for Viterbi Butterflies Using Instruction Parallelism is shown in Example B−39
on page B-40.

7-20
 Chapter 8

 %%&#'!

The TI C55x DSPLIB is an optimized DSP function library for C programmers

on TMS320C55x (C55x) DSP devices. It includes over 50 C-callable assem-
bly-optimized general-purpose signal processing routines. These routines are
typically used in computationally intensive real-time applications where opti-
mal execution speed is critical. By using these routines you can achieve exe-
cution speeds considerably faster than equivalent code written in standard
ANSI C language. In addition, by providing ready-to-use DSP functions, TI
DSPLIB can shorten significantly your DSP application development time.

The TI DSPLIB includes commonly used DSP routines. Source code is pro-
vided to allow you to modify the functions to match your specific needs
and is shipped as part of the C55x Code Composer Studio product under
the c:\ti\C5500\dsplib\55x_src directory.

Full documentation on C55x DSPLIB can be found in the TMS320C55x DSP

Library Programmer’s Reference (SPRU422).

Topic Page

8.1 Features and Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2

8.2 DSPLIB Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.3 DSPLIB Arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.4 Calling a DSPLIB Function from C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-3
8.5 Calling a DSPLIB Function from Assembly Language
Source Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
8.6 Where to Find Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-4
8.7 DSPLIB Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5

8-1
Features and Benefits / DSPLIB Data Types / DSPLIB Arguments

8.1 Features and Benefits

- Hand-coded assembly optimized routines

- C-callable routines fully compatible with the C55x DSP compiler

- Fractional Q15-format operands supported

- Complete set of examples on usage provided

- Benchmarks (cycles and code size) provided

- Tested against Matlab scripts

8.2 DSPLIB Data Types

DSPLIB functions generally operate on Q15-fractional data type elements:

- Q.15 (DATA): A Q.15 operand is represented by a short data type (16 bit)
that is predefined as DATA, in the dsplib.h header file.

Certain DSPLIB functions use the following data type elements:

- Q.31 (LDATA): A Q.31 operand is represented by a long data type (32 bit)
that is predefined as LDATA, in the dsplib.h header file.

- Q.3.12: Contains 3 integer bits and 12 fractional bits.

8.3 DSPLIB Arguments

DSPLIB functions typically operate over vector operands for greater efficiency.
Though these routines can be used to process short arrays or scalars (unless
a minimum size requirement is noted), the execution times will be longer in
those cases.

- Vector stride is always equal 1: vector operands are composed of vector

elements held in consecutive memory locations (vector stride equal to 1).

- Complex elements are assumed to be stored in a Real-Imaginary (Re-

Im) format.

- In-place computation is allowed (unless specifically noted): Source

operand can be equal to destination operand to conserve memory.

8-2
 Calling a DSPLIB Function from C

8.4 Calling a DSPLIB Function from C

In addition to installing the DSPLIB software, to include a DSPLIB function in
your code you have to:

- Include the dsplib.h include file

- Link your code with the DSPLIB object code library, 55xdsp.lib.

- Use a correct linker command file describing the memory configuration

available in your C55x DSP board.

For example, the following code contains a call to the recip16 and q15tofl rou-
tines in DSPLIB:

#include ”dsplib.h”

DATA x[3] = { 12398 , 23167, 564};

DATA r[NX];
DATA rexp[NX];
float rf1[NX];
float rf2[NX];

void main()
{
short i;

for (i=0;i<NX;i++)
{
r[i] =0;
rexp[i] = 0;
}

recip16(x, r, rexp, NX);

q15tofl(r, rf1, NX);

for (i=0; i<NX; i++)

{
rf2[i] = (float)rexp[i] * rf1[i];
}

return;
}

In this example, the q15tofl DSPLIB function is used to convert Q15 fractional
values to floating-point fractional values. However, in many applications, your
data is always maintained in Q15 format so that the conversion between float-
ing point and Q15 is not required.

TI C55x DSPLIB 8-3

Calling
Calling aa DSPLIB
DSPLIB Function
Function from
from C
Assembly Language Source Code / Where to FInd Sample Code

8.5 Calling a DSPLIB Function from Assembly Language Source Code

The DSPLIB functions were written to be used from C. Calling the functions
from assembly language source code is possible as long as the calling-func-
tion conforms with the C55x DSP C compiler calling conventions. Refer to the
TMS320C55x Optimizing C Compiler User’s Guide (SPRU281), if a more in-
depth explanation is required.

Realize that the DSPLIB is not an optimal solution for assembly-only program-
mers. Even though DSPLIB functions can be invoked from an assembly pro-
gram, the resulting execution times and code size may not be optimal due to
unnecessary C-calling overhead.

8.6 Where to Find Sample Code

You can find examples on how to use every single function in DSPLIB, in the
examples subdirectory. This subdirectory contains one subdirectory for each
function. For example, the c:\ti\cstools\dsplib\examples directory contains the
following files:

- araw_t.c: main driver for testing the DSPLIB acorr (raw) function.

- test.h: contains input data(a) and expected output data(yraw) for the acorr
(raw) function as. This test.h file is generated by using Matlab scripts.

- test.c: contains function used to compare the output of araw function with
the expected output data.

- ftest.c: contains function used to compare two arrays of float data types.

- ltest.c: contains function used to compare two arrays of long data types.

- 55x.cmd: an example of a linker command you can use for this function.

8-4
 DSPLIB Functions

8.7 DSPLIB Functions

DSPLIB provides functions in the following 8 functional catagories:

- Fast-Fourier Transforms (FFT)

- Filtering and convolution
- Adaptive filtering
- Correlation
- Math
- Trigonometric
- Miscellaneous
- Matrix

For specific DSPLIB function API descriptions, refer to the TMS320C55x DSP
Library Programmer’s Reference (SPRU422).

TI C55x DSPLIB 8-5

8-6
 Appendix
AppendixAA

# &(  

This appendix is a list of D-unit instructions where A-unit registers are read
(“pre-fetched”) in the Read phase of the execution pipeline.

Instruction Description
ABS_RR dst = |src|

ADD_RM dst = src + Smem

ADD_RM_IS ACy = ACx + (Smem << Tx)

ADD_RR dst = dst + src

ADD_RR_IS ACy = ACy + (ACx << Tx)

ADD_RWK dst = src + K16

ADD_SUB_RM_H HI(ACx) = Smem + Tx , LO(ACx) = Smem − Tx

ADSC_RM_4 ACy = ads2c(Smem,ACx,Tx,TC1,TC2)

AND_RBK dst = src & k8

AND_RM dst = src & Smem

AND_RR dst = dst & src

AND_RWK dst = src & k16

CMPL_RR dst = ~src

CMPR_RR_10 TCx = uns(src RELOP dst) {==,<,>=,!=}

CMPR_RR_11 TCx = TCy & uns(src RELOP dst) {==,<,>=,!=}

CMPR_RR_12 TCx = !TCy & uns(src RELOP dst) {==,<,>=,!=}

CMPR_RR_13 TCx = TCy | uns(src RELOP dst) {==,<,>=,!=}

CMPR_RR_14 TCx = !TCy | uns(src RELOP dst) {==,<,>=,!=}

DADD_RLM_D HI(ACx) = HI(Lmem) + Tx , LO(ACx) = LO(Lmem) + Tx

DADS_RLM HI(ACx) = HI(Lmem) + Tx , LO(ACx) = LO(Lmem) − Tx

A-1
 Instruction Description
DRSUB_RLM_D HI(ACx) = HI(Lmem) − Tx , LO(ACx) = LO(Lmem) − Tx

DSAD_RLM HI(ACx) = HI(Lmem) − Tx , LO(ACx) = LO(Lmem) + Tx

DSUB_RLM_D HI(ACx) = Tx − HI(Lmem) , LO(ACx) = Tx − LO(Lmem)

LD_MAC_R ACx = rnd(ACx + (Tx * Xmem)) ,ACy = Ymem << #16[,T3 =

Xmem]

LD_MAS_R ACx = rnd(ACx − (Tx * Xmem)) ,ACy = Ymem << #16[,T3 =

Xmem]

LD_RM_IS ACx = rnd(Smem << Tx)

MAC_R_RBK ACy = rnd(ACx + (Tx * K8))

MAC_R_RM ACy = rnd(ACx + (Tx * Smem)) [,T3 = Smem]

MAC_R_RR ACy = rnd(ACy + (ACx * Tx))

MAC_R_RWK_S ACy = rnd(ACx + (Tx * K16))

MAS_R_RM ACy = rnd(ACx − (Tx * Smem)) [,T3 = Smem]

MAS_R_RR ACy = rnd(ACy − (ACx * Tx))

MAX_RR dst = max(src,dst)

MIN_RR dst = min(src,dst)

MPOLY_R_RR ACy = rnd((ACy * Tx) + ACx)

MPYU_R_RM ACy = rnd(uns(Tx * Smem)) [,T3 = Smem]

MPY_R_RR_DR ACy = rnd(ACx * Tx)

MV_RR dst = src

MV_R_ACH HI(ACx) = DAx

MV_XRR dst = src

NEG_RR dst = −src

OR_RBK dst = src | k8

OR_RM dst = src | Smem

OR_RR dst = dst | src

OR_RWK dst = src | k16

ROL_RR_1 dst = {TC2,Carry} \\ src \\ {TC2,Carry}

A-2
Instruction Description
ROR_RR_1 dst = {TC2,Carry} // src // {TC2,Carry}

RSUB_RM dst = Smem − src

SFTA_RR_IS ACy = ACx << Tx

SFTA_RR_ISC ACy = ACx <<C Tx

SFTL_RR_IS ACy = ACx <<< Tx

STH_RS_RM_ASM Smem = HI(saturate(uns(rnd(ACx << Tx))))

STH_R_RM_ASM Smem = HI(rnd(ACx << Tx))

ST_ADD ACy = ACx + (Xmem << #16) ,Ymem = HI(ACy << T2)

ST_LD ACy = Xmem << #16 ,Ymem = HI(ACx << T2)

ST_MAC_R ACy = rnd(ACy + (Tx * Xmem)) ,Ymem = HI(ACx <<

T2)[,T3 = Xmem]

ST_MAS_R ACy = rnd(ACy − (Tx * Xmem)) ,Ymem = HI(ACx <<

T2)[,T3 = Xmem]

ST_MPY_R ACy = rnd(Tx * Xmem) ,Ymem = HI(ACx << T2)[,T3 =

Xmem]

ST_RM_ASM Smem = LO(ACx << Tx)

ST_SUB ACy = (Xmem << #16) − ACx ,Ymem = HI(ACy << T2)

SUB_ADD_RM_H HI(ACx) = Smem − Tx , LO(ACx) = Smem + Tx

SUB_RM dst = src − Smem

SUB_RM_IS ACy = ACx − (Smem << Tx)

SUB_RR dst = dst − src

SUB_RR_IS ACy = ACy − (ACx << Tx)

SUB_RWK dst = src − K16

XOR_RBK dst = src ^ k8

XOR_RM dst = src ^ Smem

XOR_RR dst = dst ^ src

XOR_RWK dst = src ^ k16

Special D-Unit Instructions A-3

 Appendix
AppendixBA

    

This appendix shows the algebraic instructions code examples that corre-
spond to the mnemonic instructions code examples shown in Chapters 2
through 7.

B-1
Example B−1. Partial Assembly Code of test.asm (Step 1)

* Step 1: Section allocation

* −−−−−−
.def x,y,init
x .usect “vars”,4 ; reserve 4 uninitalized locations for var x
y .usect “vars”,1 ; reserve 1 uninitialized location for result y

.sect “table” ; create initialized section “table” to

init .int 1,2,3,4 ; contain initialization values for x

.text ; create code section (default is .text)

.def start ; make the start label global
start ; define label to the start of the code

Example B−2. Partial Assembly Code of test.asm (Step 2)

* Step 2: Processor mode initialization

* −−−−−−
bit(ST1, #ST1_C54CM) = #0 ; set processor to C55x native mode instead of
C54x compatibility mode (reset value)
bit(ST2, #ST2_AR0LC) = #0 ; set AR0 register in linear mode (reset value)
bit(ST2, #ST2_AR6LC) = #0 ; set AR6 register in linear mode (reset value)

B-2
Example B−3. Partial Assembly Code of test.asm (Part3)

* Step 3a: Copy initialization values to vector x using indirect addressing

* −−−−−−−
copy
XAR0 = #x ; XAR0 pointing to startn of x array
XAR6 = #init ; XAR6 pointing to start of init array

*AR0+ = *AR6+ ; copy from source ”init” to destination ”x”

*AR0+ = *AR6+
*AR0+ = *AR6+
*AR0 = *AR6

* Step 3b: Add values of vector x elements using direct addressing

* −−−−−−−
add
XDP = #x ; XDP pointing to variable x
.dp x ; and the assembler is notified

AC0 = @x
AC0 += @(x+3)
AC0 += @(x+1)
AC0 += @(x+2)

* Step 3c. Write the result to y using absolute addressing

* −−−−−−−
*(#y) = AC0

end
nop
goto end

Algebraic Instructions Code Examples B-3

Example B−4. Assembly Code Generated With −o3 and −pm Options

_sum:
;** Parameter deleted n == 9u
T0 = #0 ; |3|
repeat(#9)
T0 = T0 + *AR0+

return ; |11|
_main:
SP = SP + #−1
XAR0 = #_a ; |9|
call #_sum ; |9|
; call occurs [#_sum] ; |9|
*(#_sum1) = T0 ; |9|
XAR0 = #_b ; |10|
call #_sum ; |10|
; call occurs [#_sum] ; |10|
*(#_sum2) = T0 ; |10|
S = SP + #1P
return
; return occurs

B-4
Example B−5. Assembly Generated Using −o3, −pm, and −oi50

_sum:
T0 = #0 ; |3|
repeat(#9)
T0 = T0 + *AR0+
return ; |11|
_main:
XAR3 = #_a ; |9|
repeat(#9)
|| AR1 = #0 ; |3|
AR1 = AR1 + *AR3+

*(#_sum1) = AR1 ; |11|

AR1 = #0 ; |3|
XAR3 = #_b ; |10|
repeat(#9)
AR1 = AR1 + *AR3+

*(#_sum2) = AR1 ; |11|

return

Example B−6. Assembly Code for localrepeat Generated by the Compiler

_vecsum:
AR3 = T0 − #1
BRC0 = AR3
localrepeat {
AC0 = (*AR0+ << #16) + (*AR1+ << #16) ; |7|
*AR2+ = HI(AC0) ; |7|
} ; loop ends ; |8|
L2:
return

Algebraic Instructions Code Examples B-5

Example B−7. Inefficient Loop Code for Variable and Constraints (Assembly)

_sum:
AR1 = #0 ; |3|
if (T0 <= #0) goto L2 ; |6|
; branch occurs ; |6|
AR2 = T0 − #1
CSR = AR2
repeat(CSR)
AR1 = AR1 + *AR0+

T0 = AR1 ; |11|
return ; |11|

Example B−8. Assembly Code Generated With the MUST_ITERATE Pragma

_sum:
AR2 = T0 − #1
CSR = AR2
AR1 = #0 ; |3|
repeat(CSR)
AR1 = AR1 + *AR0+

T0 = AR1 ; |12|
return ; |12|

B-6
Example B−9. Generated Assembly for FIR Filter Showing Dual-MAC

_fir:
AR3 = T0 + #1
AR3 = AR3 >> #1
AR3 = AR3 − #1
BRC0 = AR3
push(T3,T2)
T3 = #0 ; |6|
|| XCDP = XAR0
SP = SP + #−1
localrepeat {

T2 = T1 − #1
XAR3 = XAR1
CSR = T2
AR3 = AR3 + T3
XAR4 = XAR3
AR4 = AR4 + #1
AC0 = #0 ; |8|
repeat(CSR)
|| AC1 = AC0 ; |8|

AC0 = AC0 + (*AR4+ * coef(*CDP+)), AC1 = AC1 + (*AR3+ *

coef(*CDP+))

XAR0 = XCDP
T3 = T3 + #2
AR0 = AR0 − T1
|| *AR2(short(#1)) = HI(AC0)
AR2 = AR2 + #2
|| *AR2 = HI(AC1)
XCDP = XAR0
}

SP = SP + #1
T3,T2 = pop()
return

Algebraic Instructions Code Examples B-7

Example B−10. Inefficient Assembly Code Generated by C Version of Saturated Addition

_sadd:
AR1 = T1 ; |5|
T1 = T1 ^ T0 ; |9|
TC1 = bit(T1,@#15) ; |9|
AR1 = AR1 + T0
if (TC1) goto L2 ; |9|
; branch occurs ; |9|
AR2 = T0 ; |9|
AR2 = AR2 ^ AR1 ; |9|
TC1 = bit(AR2,@#15) ; |9|
if (!TC1) goto L2 ; |9|
; branch occurs ; |9|
if (T0 < #0) goto L1 ; |22|
; branch occurs ; |22|
T0 = #32767 ; |22|
goto L3 ; |22|
; branch occurs ; |22|
L1:
AR1 = #−32768 ; |22|
L2:
T0 = AR1 ; |25|
L3:
return ; |25|
; return occurs ; |25|

B-8
Example B−11. Assembly Code Generated When Using Compiler Intrinsic for
Saturated Add

_sadd:
bit(ST3, #ST3_SATA) = #1
T0 = T0 + T1 ; |3|
bit(ST3, #ST3_SATA) = #0
return ; |3|
; return occurs ; |3|

Example B−12. Assembly Output for Circular Addressing C Code

_circ:
AC0 = #0 ; |7|
if (T1 <= #0) goto L2 ; |9|
; branch occurs ; |9|
AR3 = T1 − #1
BRC0 = AR3
AR2 = #0 ; |6|
localrepeat {
; loop starts
L1:
AC0 = AC0 + (*AR1+ * *AR0+) ; |11|
|| AR2 = AR2 + #1
TC1 = (AR2 < T0) ; |12|
if (!TC1) execute (D_Unit) ||
AR1 = AR1 − T0
if (!TC1) execute (D_Unit) ||
AR2 = AR2 − T0
} ; loop ends ; |13|
L2:
return ; |14|
; return occurs ; |14|

Algebraic Instructions Code Examples B-9

Example B−13. Assembly Output for Circular Addressing Using Modulo

_circ:
push(T3,T2)
SP = SP + #−7
dbl(*SP(#0)) = XAR1
|| AC0 = #0 ; |4|
dbl(*SP(#2)) = AC0 ; |4|
|| T2 = T0 ; |2|
if (T1 <= #0) goto L2 ; |6|
; branch occurs ; |6|
T0 = #0 ; |3|
T3 = T1
|| dbl(*SP(#4)) = XAR0
L1:
XAR3 = dbl(*SP(#0))
T1 = *AR3(T0) ; |8|
XAR3 = dbl(*SP(#4))
AC0 = dbl(*SP(#2)) ; |8|
T0 = T0 + #1
AC0 = AC0 + (T1 * *AR3+) ; |8|
dbl(*SP(#2)) = AC0 ; |8|
dbl(*SP(#4)) = XAR3
call #I$$MOD ; |9|
|| T1 = T2 ; |9|
; call occurs [#I$$MOD] ; |9|
T3 = T3 − #1
if (T3 != #0) goto L1 ; |10|
; branch occurs ; |10|
L2:
AC0 = dbl(*SP(#2))
SP = SP + #7 ; |11|
T3,T2 = pop()
return ; |11|
; return occurs ; |11|

B-10
 Algebraic Instructions Code Examples

Example B−14. Complex Vector Multiplication Code

N .set 3 ; Length of each complex vector

.data
A .int 1,2,3,4,5,6 ; Complex input vector #1
B .int 7,8,9,10,11,12 ; Complex input vector #2

;Results are: 0xfff7, 0x0016, 0xfff3, 0x0042, 0xffef, 0x007e

.bss C, 2*N, ,1 ; Results vector, long−word aligned

.text
bit(ST2,#ST2_ARMS) = #0 ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

cplxmul:
XAR0 = #A ; Pointer to A vector
XCDP = #B ; Pointer to B vector
XAR1 = #C ; Pointer to C vector
BRC0 = #(N−1) ; Load loop counter
T0 = #1 ; Pointer offset
T1 = #2 ; Pointer increment

localrepeat { ; Start the loop

AC0 = *AR0 * coef(*CDP+),

AC1 = *AR0(T0) * coef(*CDP+)

AC0 = AC0 − (*AR0(T0) * coef(*CDP+)),

AC1 = AC1 + (*(AR0+T1) * coef(*CDP+))

*AR1+ = pair(LO(AC0)) ; Store complex result

} ; End of loop

Algebraic Instructions Code Examples B-11

Algebraic Instructions Code Examples

Example B−15. Block FIR Filter Code (Not Optimized)

N_TAPS .set 4 ; Number of filter taps

N_DATA .set 11 ; Number of input values

.data
COEFFS .int 1,2,3,4 ; Coefficients
IN_DATA .int 1,2,3,4,5,6,7,8,9,10,11 ; Input vector

;Results are: 0x0014, 0x001E, 0x0028, 0x0032,

; 0x003C, 0x0046, 0x0050, 0x005A

.bss OUT_DATA, N_DATA − N_TAPS + 1 ; Output vector

.text
bit(ST2,#ST2_ARMS) = #0 ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

bfir:
XCDP = #COEFFS ; Pointer to coefficient array
XAR0 = #(IN_DATA + N_TAPS − 1) ; Pointer to input vector
XAR1 = #(IN_DATA + N_TAPS) ; 2nd pointer to input vector
XAR2 = #OUT_DATA ; Pointer to output vector
BRC0 = #((N_DATA − N_TAPS + 1)/2 − 1) ; Load outer loop counter
CSR = #(N_TAPS − 1) ; Load inner loop counter

localrepeat { ; Start the outer loop

AC0 = #0 ; Clear AC0

AC1 = #0 ; Clear AC1

repeat(CSR) ; Start the inner loop

AC0 = AC0 + ( *AR0− * coef(*CDP+) ), ; All taps
AC1 = AC1 + ( *AR1− * coef(*CDP+) )

*AR2+ = AC0 ; Write 1st result

*AR2+ = AC1 ; Write 2nd result

CDP = #COEFFS ; Rewind coefficient pointer

AR0 = AR0 + #(N_TAPS + 2) ; Adjust 1st input vector
; pointer
AR1 = AR1 + #(N_TAPS + 2) ; Adjust 2nd input vector
; pointer

} ; End of outer loop

B-12
 Algebraic Instructions Code Examples

Example B−16. Block FIR Filter Code (Optimized)

N_TAPS .set 4 ; Number of filter taps

N_DATA .set 11 ; Number of input values

.data
COEFFS .int 1,2,3,4 ; Coefficients
IN_DATA .int 1,2,3,4,5,6,7,8,9,10,11 ; Input vector

;Results are: 0x0014, 0x001E, 0x0028, 0x0032,

; 0x003C, 0x0046, 0x0050, 0x005A

.bss OUT_DATA, N_DATA − N_TAPS + 1, ,1

; Output vector, long−word
; aligned

.text
bit(ST2,#ST2_ARMS) = #0 ; Clear ARMS bit (select DSP mode)
.arms_off ; Tell assembler ARMS = 0

bfir:
XCDP =
#COEFFS ; Pointer to coefficient array
XAR0 =
#(IN_DATA + N_TAPS − 1) ; Pointer to input vector
XAR1 =
#(IN_DATA + N_TAPS) ; 2nd pointer to input vector
XAR2 =
#OUT_DATA ; Pointer to output vector
BRC0 =
#((N_DATA − N_TAPS + 1)/2 − 1)
; Load outer loop counter
CSR = #(N_TAPS − 3) ; Load inner loop counter
T0 = #(−(N_TAPS − 1)) ; CDP rewind increment

T1 = #(N_TAPS + 1) ; ARx rewind increment

||localrepeat { ; Start the outer loop

AC0 = *AR0− * coef(*CDP+), ; 1st tap

AC1 = *AR1− * coef(*CDP+)

repeat(CSR) ; Start the inner loop

AC0 = AC0 + ( *AR0− * coef(*CDP+) ), ; Inner taps
AC1 = AC1 + ( *AR1− * coef(*CDP+) )

AC0 = AC0 + ( *(AR0+T1) * coef(*(CDP+T0)) ), ; Last tap

AC1 = AC1 + ( *(AR1+T1) * coef(*(CDP+T0)) )

*AR2+ = pair(LO(AC0)) ; Store both results

} ; End of outer loop

Algebraic Instructions Code Examples B-13

Algebraic Instructions Code Examples

Example B−17. A-Unit Code With No User-Defined Parallelism

; Variables
.data

COEFF1 .word 0x0123 ; First set of coefficients

.word 0x1234
.word 0x2345
.word 0x3456
.word 0x4567

COEFF2 .word 0x7654 ; Second set of coefficients

.word 0x6543
.word 0x5432
.word 0x4321
.word 0x3210

HST_FLAG .set 0x2000 ; Host flag address

HST_DATA .set 0x2001 ; Host data address

CHANGE .set 0x0000 ; “Change coefficients” command from host

READY .set 0x0000 ; “READY” Flag from Host
BUSY .set 0x1111 ; “BUSY” Flag set by DSP

.global start_a1

.text

start_a1:
AR0 = #HST_FLAG ; AR0 points to Host Flag
AR2 = #HST_DATA ; AR2 points to Host Data
AR1 = #COEFF1 ; AR1 points to COEFF1 buffer initially
AR3 = #COEFF2 ; AR3 points to COEFF2 buffer initially
CSR = #4 ; Set CSR = 4 for repeat in COMPUTE
BIT(ST1, #ST1_FRCT) = #1 ; Set fractional mode bit
BIT(ST1, #ST1_SXMD) = #1 ; Set sign−extension mode bit

B-14
 Algebraic Instructions Code Examples

Example B−17. A-Unit Code With No User-Defined Parallelism (Continued)

LOOP:
T0 = *AR0 ; T0 = Host Flag
if (T0 == #READY) GOTO PROCESS ; If Host Flag is ”READY”, continue
GOTO LOOP ; process − else poll Host Flag again

PROCESS:
T0 = *AR2 ; T0 = Host Data

if (T0 == #CHANGE) EXECUTE(AD_UNIT)

; The choice of either set of
; coefficients is based on the value
; of T0. COMPUTE uses AR3 for
; computation, so we need to
; load AR3 correctly here.

SWAP(AR1, AR3) ; Host message was ”CHANGE”, so we

; need to swap the two coefficient
; pointers.

CALL COMPUTE ; Compute subroutine

*AR2 = AR4 ; Write result to Host Data

*AR0 = #BUSY ; Set Host Flag to Busy
GOTO LOOP ; Infinite loop continues
END

COMPUTE:
AC1 = #0 ; Initialize AC1 to 0
REPEAT(CSR) ; CSR has a value of 4
AC1 = AC1 + (*AR2 * *AR3+) ; This MAC operation is performed
; 5 times
AR4 = AC1 ; Result is in AR4

RETURN

HALT:
GOTO HALT

Algebraic Instructions Code Examples B-15

Algebraic Instructions Code Examples

Example B−18. A-Unit Code in Example B−17 Modified to Take Advantage of Parallelism

; Variables
.data

COEFF1 .word 0x0123 ; First set of coefficients

.word 0x1234
.word 0x2345
.word 0x3456
.word 0x4567

COEFF2 .word 0x7654 ; Second set of coefficients

.word 0x6543
.word 0x5432
.word 0x4321
.word 0x3210

HST_FLAG .set 0x2000 ; Host flag address

HST_DATA .set 0x2001 ; Host data address

CHANGE .set 0x0000 ; “Change coefficients” command from host

READY .set 0x0000 ; “READY” Flag from Host
BUSY .set 0x1111 ; “BUSY” Flag set by DSP

.global start_a2

.text

start_a2:

AR0 = #HST_FLAG ; AR0 points to Host Flag

AR2 = #HST_DATA ; AR2 points to Host Data
AR1 = #COEFF1 ; AR1 points to COEFF1 buffer initially
AR3 = #COEFF2 ; AR3 points to COEFF2 buffer initially

CSR = #4 ; Set CSR = 4 for repeat in COMPUTE

|| BIT(ST1, #ST1_FRCT) = #1 ; Set fractional mode bit

BIT(ST1, #ST1_SXMD) = #1 ; Set sign−extension mode bit

LOOP:
T0 = *AR0 ; T0 = Host Flag
if (T0 == #READY) GOTO PROCESS ; If Host Flag is “READY”, continue
GOTO LOOP ; process − else poll Host Flag again

B-16
 Algebraic Instructions Code Examples

Example B−18. A-Unit Code in Example B−17 Modified to Take Advantage of Parallelism
(Continued)

PROCESS:
T0 = *AR2 ; T0 = Host Data

if (T0 == #CHANGE) EXECUTE(AD_UNIT)

; The choice of either set of
; coefficients is based on the value
; of T0. COMPUTE uses AR3 for
; computation, so we need to
; load AR3 correctly here.

|| SWAP(AR1, AR3) ; Host message was ”CHANGE”, so we

; need to swap the two coefficient
; pointers.

CALL COMPUTE ; Compute subroutine

*AR2 = AR4 ; Write result to Host Data

|| *AR0 = #BUSY ; Set Host Flag to Busy

GOTO LOOP ; Infinite loop continues

END

COMPUTE:
AC1 = #0 ; Initialize AC1 to 0
|| REPEAT(CSR) ; CSR has a value of 4
AC1 = AC1 + (*AR2 * *AR3+) ; This MAC operation is performed
; 5 times
AR4 = AC1 ; Result is in AR4
|| RETURN

HALT:
GOTO HALT

Algebraic Instructions Code Examples B-17

Algebraic Instructions Code Examples

Example B−19. P-Unit Code With No User-Defined Parallelism

.CPL_off ; Tell assembler that CPL bit is 0

; (Direct addressing is done with DP)

; Variables
.data

var1 .word 0x0004

var2 .word 0x0000

.global start_p1

.text

start_p1:
XDP = #var1
AR3 = #var2

BRC0 = #0007h ; BRC0 loaded using KPB

BRC1 = *AR3 ; BRC1 loaded using DB

AC2 = #0006h

BLOCKREPEAT {
AC1 = AC2
AR1 = #8000h
LOCALREPEAT {
AC1 = AC1 − #1
*AR1+ = AC1
}
AC2 = AC2 + #1
}
@(AC0_L) = BRC0 ; AC0_L loaded using EB
@(AC1_L) = BRC1 ; AC1_L loaded using EB

if (AC0 >= #0) goto start_p1:

if (AC1 >= #0) goto start_p1:
end_p1

B-18
 Algebraic Instructions Code Examples

Example B−20. P-Unit Code in Example B−19 Modified to Take Advantage of Parallelism

.CPL_off ; Tell assembler that CPL bit is 0

; (Direct addressing is done with DP)

; Variables
.data

var1 .word 0x0004

var2 .word 0x0000

.global start_p2

.text

start_p2:
XDP = #var1
AR3 = #var2

BRC0 = #0007h ; BRC0 loaded using KPB

|| BRC1 = *AR3 ; BRC1 loaded using DB

AC2 = #0006h
|| BLOCKREPEAT {
AC1 = AC2
AR1 = #8000h
|| LOCALREPEAT {
AC1 = AC1 − #1
*AR1+ = AC1
}
AC2 = AC2 + #1
}
@(AC0_L) = BRC0 ; AC0_L loaded using EB
@(AC1_L) = AR1 ; AC1_L loaded using EB

if (AC0 >= #0) goto start_p2

if (AC1 >= #0) goto start_p2
end_p2

Algebraic Instructions Code Examples B-19

Algebraic Instructions Code Examples

Example B−21. D-Unit Code With No User-Defined Parallelism

; Variables
.data

var1 .word 0x8000

var2 .word 0x0004

.global start_d1

.text

start_d1:
AR3 = #var1
AR4 = #var2

AC0 = #0004h ; AC0 loaded using KDB

AC2 = *AR3 ; AC2 loaded using DB

T0 = #5A5Ah ; T0 loaded with constant, 0x5A5A

AC2 = AC2 + (AC0 * T0) ; MAC

AC1 = AC1 + (AC2 * T0) ; MAC
SWAP(AC0, AC2) ; SWAP

*AR3 = HI(AC1) ; Store result in AC1

*AR4 = HI(AC0) ; Store result in AC0
end_d1

B-20
 Algebraic Instructions Code Examples

Example B−22. D-Unit Code in Example B−21 Modified to Take Advantage of Parallelism

; Variables
.data

var1 .word 0x8000

var2 .word 0x0004

.global start_d2

.text

start_d2:
AR3 = #var1
AR4 = #var2

AC0 = #0004h ; AC0 loaded using KDB

|| AC2 = *AR3 ; AC2 loaded using DB

T0 = #5A5Ah ; T0 loaded with constant, 0x5A5A

AC2 = AC2 + (AC0 * T0) ; MAC

AC1 = AC1 + (AC2 * T0) ; MAC
||SWAP(AC0, AC2) ; SWAP

*AR3 = HI(AC1) ; Store result in AC1

|| *AR4 = HI(AC0) ; Store result in AC0
end_d2

Algebraic Instructions Code Examples B-21

Algebraic Instructions Code Examples

Example B−23. Code That Uses Multiple CPU Units But No User-Defined Parallelism

.CPL_ON ; Tell assembler that CPL bit is 1

; (SP direct addressing like *SP(0) is enabled)

; Register usage
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

.asg AR0, X_ptr ; AR0 is pointer to input buffer − X_ptr

.asg AR1, H_ptr ; AR1 is pointer to coefficients − H_ptr
.asg AR2, R_ptr ; AR2 is pointer to result buffer − R_ptr
.asg AR3, DB_ptr ; AR3 is pointer to delay buffer − DB_ptr

FRAME_SZ .set 2

.global _fir

.text

;**********************************************************************

_fir

; Create local frame for temp values

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

SP = SP − #FRAME_SZ

; Turn on fractional mode

; Turn on sign−extension mode
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BIT(ST1, #ST1_FRCT) = #1 ; Set fractional mode bit

BIT(ST1, #ST1_SXMD) = #1 ; Set sign−extension mode bit

; Set outer loop count by subtracting 1 from nx and storing into

; block−repeat counter
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AC1 = T1 − #1 ; AC1 = number of samples (nx) − 1

*SP(0) = AC1 ; Top of stack = nx − 1
BRC0 = *SP(0) ; BRC0 = nx − 1 (outer loop counter)

B-22
 Algebraic Instructions Code Examples

Example B−23. Code That Uses Multiple CPU Units But No User-Defined Parallelism
(Continued)

; Store length of coefficient vector/delay buffer in BK register

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BIT(ST2, #ST2_AR1LC) = #1 ; Enable AR1 circular configuration

BSA01 = *(#0011h) ; Set buffer (filter) start address
; AR1 used as filter pointer

BIT(ST2, #ST2_AR3LC) = #1 ; Enable AR3 circular configuration

*SP(1) = DB_ptr ; Save pointer to delay buffer pointer
AC1 = *DB_ptr ; AC1 = delay buffer pointer
DB_ptr = AC1 ; AR3 (DB_ptr) = delay buffer pointer
BSA23 = *(#0013h) ; Set buffer (delay buffer) start address
; AR3 used as filter pointer

*SP(0) = T0 ; Save filter length, nh − used as buffer

; size
BK03 = *SP(0) ; Set circular buffer size − size passed
; in T0

AC1 = T0 − #3 ; AC1 = nh − 3
*SP(0) = AC1
CSR = *SP(0) ; Set inner loop count to nh − 3

H_ptr = #0 ; Initialize index of filter to 0

DB_ptr = #0 ; Initialize index of delay buffer to 0

; Begin outer loop on nx samples

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BLOCKREPEAT {

; Move next input sample into delay buffer

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

*DB_ptr = *X_ptr+

; Sum h * x for next y value

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AC0 = *H_ptr+ * *DB_ptr+

REPEAT (CSR)
AC0 = AC0 + (*H_ptr+ * *DB_ptr+)

AC0 = rnd(AC0 + (*H_ptr+ * *DB_ptr)) ; Round result

Algebraic Instructions Code Examples B-23

Algebraic Instructions Code Examples

Example B−23. Code That Uses Multiple CPU Units But No User-Defined Parallelism
(Continued)

; Store result
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

*R_ptr+ = HI(AC0)
}

; Clear FRCT bit to restore normal C operating environment

; Return overflow condition of AC0 (shown in ACOV0) in T0
; Restore stack to previous value, FRAME, etc..
; Update current index of delay buffer pointer
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

END_FUNCTION:

AR0 = *SP(1) ; AR0 = pointer to delay buffer pointer

SP = SP + #FRAME_SZ ; Remove local stack frame
*AR0 = DB_ptr ; Update delay buffer pointer with current
; index

BIT(ST1, #ST1_FRCT) = #0 ; Clear fractional mode bit

T0 = #0 ; Make T0 = 0 for no overflow (return value)

if(overflow(AC0)) execute(AD_unit)
T0 = #1 ; Make T0 = 1 for overflow (return value)

RETURN
;********************************************************************

B-24
 Algebraic Instructions Code Examples

Example B−24. Code in Example B−23 Modified to Take Advantage of Parallelism

.CPL_ON ; Tell assembler that CPL bit is 1

; (SP direct addressing like *SP(0) is enabled)

; Register usage
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

.asg AR0, X_ptr ; AR0 is pointer to input buffer − X_ptr

.asg AR1, H_ptr ; AR1 is pointer to coefficients − H_ptr
.asg AR2, R_ptr ; AR2 is pointer to result buffer − R_ptr
.asg AR3, DB_ptr ; AR3 is pointer to delay buffer − DB_ptr

FRAME_SZ .set 2

.global _fir

.text

;**************************************************************************

_fir

; Create local frame for temp values

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

SP = SP − #FRAME_SZ ; (Attempt to put this in parallel with

; the following AC1 modification failed)

; Set outer loop count by subtracting 1 from nx and storing into

; block−repeat counter
; Turn on fractional mode
; Turn on sign−extension mode
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AC1 = T1 − #1 ; AC1 = number of samples (nx) − 1

BIT(ST1, #ST1_FRCT) = #1 ; Set fractional mode bit

|| *SP(0) = AC1 ; Top of stack = nx − 1

BRC0 = *SP(0) ; BRC0 = nx − 1 (outer loop counter)

|| BIT(ST1, #ST1_SXMD) = #1 ; Set sign−extension mode bit

Algebraic Instructions Code Examples B-25

Algebraic Instructions Code Examples

Example B−24. Code in Example B−23 Modified to Take Advantage of Parallelism

(Continued)

; Store length of coefficient vector/delay buffer in BK register

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

BIT(ST2, #ST2_AR1LC) = #1 ; Enable AR1 circular configuration

BSA01 = *(#0011h) ; Set buffer (filter) start address
; AR1 used as filter pointer

BIT(ST2, #ST2_AR3LC) = #1 ; Enable AR3 circular configuration

|| *SP(1) = DB_ptr ; Save pointer to delay buffer pointer

AC1 = *DB_ptr ; AC1 = delay buffer pointer

DB_ptr = AC1 ; AR3 (DB_ptr) = delay buffer pointer
|| *SP(0) = T0 ; Save filter length, nh − used as buffer
; size
BSA23 = *(#0013h) ; Set buffer (delay buffer) start address
; AR3 used as filter pointer

BK03 = *SP(0) ; Set circular buffer size − size passed

; in T0

AC1 = T0 − #3 ; AC1 = nh − 3
*SP(0) = AC1

CSR = *SP(0) ; Set inner loop count to nh − 3

|| H_ptr = #0 ; Initialize index of filter to 0

DB_ptr = #0 ; Initialize index of delay buffer to 0

; (in parallel with BLOCKREPEAT below)

; Begin outer loop on nx samples

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

||BLOCKREPEAT {

; Move next input sample into delay buffer

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

*DB_ptr = *X_ptr+

B-26
 Algebraic Instructions Code Examples

Example B−24. Code in Example B−23 Modified to Take Advantage of Parallelism

(Continued)

; Sum h * x for next y value

; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

AC0 = *H_ptr+ * *DB_ptr+

|| REPEAT (CSR)

AC0 = AC0 + (*H_ptr+ * *DB_ptr+)

AC0 = rnd(AC0 + (*H_ptr+ * *DB_ptr)) ; Round result

; Store result
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

*R_ptr+ = HI(AC0)
}

; Clear FRCT bit to restore normal C operating environment

; Return overflow condition of AC0 (shown in ACOV0) in T0
; Restore stack to previous value, FRAME, etc..
; Update current index of delay buffer pointer
; −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−

END_FUNCTION:

AR0 = *SP(1) ; AR0 = pointer to delay buffer pointer

|| SP = SP + #FRAME_SZ ; Remove local stack frame

*AR0 = DB_ptr ; Update delay buffer pointer with current

; index

|| BIT(ST1, #ST1_FRCT) = #0 ; Clear fractional mode bit

T0 = #0 ; Make T0 = 0 for no overflow (return value)

|| if(overflow(AC0)) execute(AD_unit)
T0 = #1 ; Make T0 = 1 for overflow (return value)

|| RETURN
;********************************************************************

Algebraic Instructions Code Examples B-27

Algebraic Instructions Code Examples

Example B−25. Nested Loops

BRC0 = #(n0−1)
BRC1 = #(n1−1)
; ...
localrepeat{ ; Level 0 looping (could instead be blockrepeat):
; Loops n0 times
; ...
repeat(#(n2−1)
; ...
localrepeat{ ; Level 1 looping (could instead be blockrepeat):
; Loops n1 times
; ...
repeat(#(n3−1))
; ...
}
; ...
}

Example B−26. Branch-On-Auxiliary-Register-Not-Zero Construct

(Shown in Complex FFT Loop Code)

_cfft:

radix_2_stages:
; ...

outer_loop:
; ...
BRC0 = T1
; ...
BRC1 = T1
; ...
AR4 = AR4 >> #1 ; outer loop counter
|| if (AR5 == #0) goto no_scale ; determine if scaling required
; ...

no_scale:

localrepeat{

AC0 = dbl(*AR3) ; Load ar,ai

HI(AC2) = HI(*AR2) − HI(AC0), ; tr = ar − br

LO(AC2) = LO(*AR2) − LO(AC0) ; ti = ai − bi

localrepeat {

Note: This example shows portions of the file cfft.asm in the TI C55x DSPLIB (introduced in Chapter 8).

B-28
 Algebraic Instructions Code Examples

Example B−26. Branch-On-Auxiliary-Register-Not-Zero Construct

(Shown in Complex FFT Loop Code) (Continued)

HI(AC1) = HI(*AR2) + HI(AC0), ; ar’ = ar + br

LO(AC1) = LO(*AR2) + LO(AC0) ; ai’ = ai + bi
|| dbl(*AR6)=AC2 ; Store tr, ti

AC2 = *AR6 * coef(*CDP+), ; c*tr

AC3 = *AR7 * coef(*CDP+) ; c*ti

dbl(*AR2+) = AC1 ; Store ar, ai

|| AC0 = dbl(*AR3(T0)) ; * load ar,ai

AC3 = rnd(AC3 − (*AR6 * coef(*CDP−))), ; bi’ = c*ti − s*tr

AC2 = rnd(AC2 + (*AR7 * coef(*CDP−))) ; br’ = c*tr + s*ti

*AR3+ = pair(HI(AC2)) ; Store br’, bi’

|| HI(AC2) = HI(*AR2) − HI(AC0), ; * tr = ar − br
LO(AC2) = LO(*AR2) − LO(AC0) ; * ti = ai − bi

HI(AC1) = HI(*AR2) + HI(AC0), ; ar’ = ar + br

LO(AC1) = LO(*AR2) + LO(AC0) ; ai’ = ai + bi
|| dbl(*AR6)=AC2 ; Store tr, ti

AC2 = *AR6 * coef(*CDP+), ; c*tr

AC3 = *AR7 * coef(*CDP+) ; c*ti

dbl(*(AR2+T1)) = AC1 ; Store ar, ai

AC3 = rnd(AC3 − (*AR6 * coef(*CDP+))), ; bi’ = c*ti − s*tr

AC2 = rnd(AC2 + (*AR7 * coef(*CDP+))) ; br’ = c*tr + s*ti

*(AR3+T1) = pair(HI(AC2)) ; Store br’, bi’

AR3 = AR3 << #1

|| CDP = #0 ; rewind coefficient pointer
T3 = T3 >> #1
|| if (AR4 != #0) goto outer_loop

Note: This example shows portions of the file cfft.asm in the TI C55x DSPLIB (introduced in Chapter 8).

Algebraic Instructions Code Examples B-29

Algebraic Instructions Code Examples

Example B−27. 64-Bit Addition

;*********************************************************************
; 64−Bit Addition Pointer assignments:
;
; X3 X2 X1 X0 AR1 −> X3 (even address)
; + Y3 Y2 Y1 Y0 X2
; −−−−−−−−−−−−−− X1
; W3 W2 W1 W0 X0
; AR2 −> Y3 (even address)
; Y2
; Y1
; Y0
; AR3 −> W3 (even address)
; W2
; W1
; W0
;
;*********************************************************************

AC0 = dbl(*AR1(#2)) ; AC0 = X1 X0

AC0 = AC0 + dbl(*AR2(#2)) ; AC0 = X1 X0 + Y1 Y0
dbl(*AR3(#2)) = AC0 ; Store W1 W0.
AC0 = dbl (*AR1) ; AC0 = X3 X2
AC0 = AC0 + uns(*AR2(#1))+ CARRY ; AC0 = X3 X2 + 00 Y2 + CARRY
AC0 = AC0 + (*AR2<< #16) ; AC0 = X3 X2 + Y3 Y2 + CARRY
dbl(*AR3) = AC0 ; Store W3 W2.

B-30
 Algebraic Instructions Code Examples

Example B−28. 64-Bit Subtraction

;**********************************************************************
; 64−Bit Subtraction Pointer assignments:
;
; X3 X2 X1 X0 AR1 −> X3 (even address)
; − Y3 Y2 Y1 Y0 X2
; −−−−−−−−−−−−−− X1
; W3 W2 W1 W0 X0
; AR2 −> Y3 (even address)
; Y2
; Y1
; Y0
; AR3 −> W3 (even address)
; W2
; W1
; W0
;
;**********************************************************************

AC0 = dbl(*AR1(#2)) ; AC0 = X1 X0

AC0 = AC0 − dbl(*AR2(#2)) ; AC0 = X1 X0 − Y1 Y0
dbl(*AR3(#2)) = AC0 ; Store W1 W0.
AC0 = dbl (*AR1) ; AC0 = X3 X2
AC0 = AC0 − uns(*AR2(#1)) − BORROW ; AC0 = X3 X2 − 00 Y2 − BORROW
AC0 = AC0 − (*AR2<< #16) ; AC0 = X3 X2 − Y3 Y2 − BORROW
dbl(*AR3) = AC0 ; Store W3 W2.

Algebraic Instructions Code Examples B-31

Algebraic Instructions Code Examples

Example B−29. 32-Bit Integer Multiplication

;****************************************************************
; This routine multiplies two 32−bit signed integers, giving a
; 64−bit result. The operands are fetched from data memory and the
; result is written back to data memory.
;
; Data Storage: Pointer Assignments:
; X1 X0 32−bit operand AR0 −> X1
; Y1 Y0 32−bit operand X0
; W3 W2 W1 W0 64−bit product AR1 −> Y1
; Y0
; Entry Conditions: AR2 −> W0
; SXMD = 1 (sign extension on) W1
; SATD = 0 (no saturation) W2
; FRCT = 0 (fractional mode off) W3
;
; RESTRICTION: The delay chain and input array must be
; long-word aligned.
;***************************************************************

mar(*AR0+) ; AR0 points to X0

|| mar(*AR1+) ; AR1 points to Y0
AC0 = uns(*AR0−)*uns(*AR1) ; ACO = X0*Y0
*AR2+ = AC0 ; Save W0
AC0 = (AC0 >> #16) + ((*AR0+)*uns(*AR1−)) ; AC0 = X0*Y0>>16 + X1*Y0
AC0 = AC0 + (uns(*AR0−)* (*AR1)) ; AC0 = X0*Y0>>16 + X1*Y0 + X0*Y1
*AR2+ = AC0 ; Save W1
AC0 = (AC0 >> #16) + ((*AR0)*(*AR1)) ; AC0 = AC0>>16 + X1*Y1
*AR2+ = AC0 ; Save W2
*AR2 = HI(AC0) ; Save W3

B-32
 Algebraic Instructions Code Examples

Example B−30. 32-Bit Fractional Multiplication

;**************************************************************************
; This routine multiplies two Q31 signed integers, resulting in a
; Q31 result. The operands are fetched from data memory and the
; result is written back to data memory.
;
; Data Storage: Pointer Assignments:
; X1 X0 Q31 operand AR0 −> X1
; Y1 Y0 Q31 operand X0
; W1 W0 Q31 product AR1 −> Y1
; Y0
; Entry Conditions: AR2 −> W1 (even address)
; SXMD = 1 (sign extension on) W0
; SATD = 0 (no saturation)
; FRCT = 1 (shift result left by 1 bit)
;
; RESTRICTION: W1 W0 is aligned such that W1 is at an even address.
;***************************************************************************
mar(*AR0+) ; AR0 points to X0
AC0 = uns(*AR0−)*(*AR1+) ; AC0 = X0*Y1
AC0 = AC0 + ((*AR0)* uns(*AR1−)) ; AC0 =X0*Y1 + X1*Y0
AC0 = (AC0 >> #16) + ((*AR0)*(*AR1)) ; AC0 = AC0>>16 + X1*Y1
dbl(*AR2) = AC0 ; Save W1 W0

Example B−31. Unsigned, 16-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend Divisor ) Dividend
; AR1 −> Divisor
; AR2 −> Quotient
; AR3 −> Remainder
;
; Algorithm notes:
; − Unsigned division, 16−bit dividend, 16−bit divisor
; − Sign extension turned off. Dividend & divisor are positive numbers.
; − After division, quotient in AC0(15−0), remainder in AC0(31−16)
;***************************************************************************

bit(ST1,#ST1_SXMD) = #0 ; Clear SXMD (sign extension off)

AC0 = *AR0 ; Put Dividend into AC0
repeat( #(16 − 1) ) ; Execute subc 16 times
subc( *AR1, AC0, AC0 ) ; AR1 points to Divisor
*AR2 = AC0 ; Store Quotient
*AR3 = HI(AC0) ; Store Remainder

Algebraic Instructions Code Examples B-33

Algebraic Instructions Code Examples

Example B−32. Unsigned, 32-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend high half Divisor ) Dividend
; Dividend low half
; ...
; AR1 −> Divisor
; ...
; AR2 −> Quotient high half
; Quotient low half
; ...
; AR3 −> Remainder
;
; Algorithm notes:
; − Unsigned division, 32−bit dividend, 16−bit divisor
; − Sign extension turned off. Dividend & divisor are positive numbers.
; − Before 1st division: Put high half of dividend in AC0
; − After 1st division: High half of quotient in AC0(15−0)
; − Before 2nd division: Put low part of dividend in AC0
; − After 2nd division: Low half of quotient in AC0(15−0) and
; Remainder in AC0(31−16)
;***************************************************************************

bit(ST1,#ST1_SXMD) = #0 ; Clear SXMD (sign extension off)

AC0 = *AR0+ ; Put high half of Dividend in AC0
|| repeat( #(15 − 1) ) ; Execute subc 15 times
subc( *AR1, AC0, AC0) ; AR1 points to Divisor
subc( *AR1, AC0, AC0) ; Execute subc final time
|| AR4 = #8 ; Load AR4 with AC0_L memory address
*AR2+ = AC0 ; Store high half of Quotient
*AR4 = *AR0+ ; Put low half of Dividend in AC0_L
repeat( #(16 − 1) ) ; Execute subc 16 times
subc( *AR1, AC0, AC0)
*AR2+ = AC0 ; Store low half of Quotient
*AR3 = HI(AC0) ; Store Remainder
bit(ST1,#ST1_SXMD) = #1 ; Set SXMD (sign extension on)

B-34
 Algebraic Instructions Code Examples

Example B−33. Signed, 16-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: ___________
; AR0 −> Dividend Divisor ) Dividend
; AR1 −> Divisor
; AR2 −> Quotient
; AR3 −> Remainder
;
; Algorithm notes:
; − Signed division, 16−bit dividend, 16−bit divisor
; − Sign extension turned on. Dividend and divisor can be negative.
; − Expected quotient sign saved in AC0 before division
; − After division, quotient in AC1(15−0), remainder in AC1(31−16)
;***************************************************************************

bit(ST1,#ST1_SXMD) = #1 ; Set SXMD (sign extension on)

AC0 = (*AR0) * (*AR1) ; Sign of (Dividend x Divisor) should be
; sign of Quotient
AC1 = *AR1 ; Put Divisor in AC1
AC1 = |AC1| ; Find absolute value, |Divisor|
*AR2 = AC1 ; Store |Divisor| temporarily
AC1 = *AR0 ; Put Dividend in AC1
AC1 = |AC1| ; Find absolute value, |Dividend|
repeat( #(16 − 1) ) ; Execute subc 16 times
subc( *AR2, AC1, AC1) ; AR2 −> |Divisor|
*AR3 = HI(AC1) ; Save Remainder
*AR2 = AC1 ; Save Quotient
AC1 = AC1 << #16 ; Shift quotient: Put MSB in sign position
AC1 = − AC1 ; Negate quotient
if(AC0 < #0) execute (D_unit) ; If sign of Quotient should be negative,
*AR2 = HI(AC1) ; replace Quotient with negative version

Algebraic Instructions Code Examples B-35

Algebraic Instructions Code Examples

Example B−34. Signed, 32-Bit By 16-Bit Integer Division

;***************************************************************************
; Pointer assignments: (Dividend and Quotient are long−word aligned)
; AR0 −> Dividend high half (NumH) (even address)
; Dividend low half (NumL)
; AR1 −> Divisor (Den)
; AR2 −> Quotient high half (QuotH) (even address)
; Quotient low half (QuotL)
; AR3 −> Remainder (Rem)
;
; Algorithm notes:
; − Signed division, 32−bit dividend, 16−bit divisor
; − Sign extension turned on. Dividend and divisor can be negative.
; − Expected quotient sign saved in AC0 before division
; − Before 1st division: Put high half of dividend in AC1
; − After 1st division: High half of quotient in AC1(15−0)
; − Before 2nd division: Put low part of dividend in AC1
; − After 2nd division: Low half of quotient in AC1(15−0) and
; Remainder in AC1(31−16)
;***************************************************************************

bit(ST1,#ST1_SXMD) = #1 ; Set SXMD (sign extension on)

AC0 = (*AR0)* (*AR1) ; Sign( NumH x Den ) is sign of actual result
AC1 = *AR1 ; AC1 = Den
AC1 = |AC1| ; AC1 = abs(Den)
*AR3 = AC1 ; Rem = abs(Den) temporarily
AC1 = dbl(*AR0) ; AC1 = NumH NumL
AC1 = |AC1| ; AC1 = abs(Num)
dbl(*AR2) = AC1 ; QuotH = abs(NumH) temporarily
; QuotL = abs(NumL) temporarily

AC1 = *AR2 ; AC1 = QuotH

repeat( #(15 − 1) ) ; Execute subc 15 times
subc( *AR3, AC1, AC1)
subc( *AR3, AC1, AC1) ; Execute subc final time
||AR4 = #11 ; Load AR4 with AC1_L memory address
*AR2+ = AC1 ; Save QuotH
*AR4 = *AR2 ; AC1_L = QuotH
repeat( #(16 − 1) ) ; Execute subc 16 times
subc( *AR3, AC1, AC1)
*AR2− = AC1 ; Save QuotL
*AR3 = HI(AC1) ; Save Rem

if (AC0 >= #0) goto skip ; If actual result should be positive, goto skip.
AC1 = dbl(*AR2) ; Otherwise, negate Quotient.
AC1 = − AC1
dbl(*AR2) = AC1

skip:
return

B-36
 Algebraic Instructions Code Examples

Example B−35. Off-Place Bit Reversing of a Vector Array (in Assembly)

;...
bit (ST2, #ST2_ARMS) = #0 ; reset ARMS bit to allow bit−reverse addressing
.arms_off ; notify the assembler of ARMS bit = 0
;...
off_place:
localrepeat{
AC0 = dbl(*AR0+) ; AR0 points to input array
dbl(*(AR1+T0B)) = AC0 ; AR1 points to output array
; T0 = NX = number of complex elements in
; array pointed to by AR0
}

Note: This example shows portions of the file cbrev.asm in the TI C55x DSPLIB (introduced in Chapter 8).

Algebraic Instructions Code Examples B-37

Algebraic Instructions Code Examples

Example B−36. Delayed LMS Implementation of an Adaptive Filter

; ar_data: index in the delay buffer

; ar_input: pointer to input vector
; ar_coef: pointer to coefficient vector

StartSample:

; Clear AC0 for initial error term

AC1 = #0
|| localrepeat {
*ar_data+ = *ar_input+ ;copy input −> state(0)

; Place error term in T3

T3 = HI(AC1)

; Place first update term in AC0

;...while clearing FIR value
AC0 = T3 * *ar_data+
|| AC1 = #0

;AC0 = update coef

;AC1 = start of FIR output
LMS(*ar_coef, *ar_data, AC0, AC1)
|| localrepeat {
*ar_coef+ = HI(AC0)
|| AC0 = T3 * *ar_data+

;AC0 = update coef

;AC1 = update of FIR output
LMS(*ar_coef, *ar_data, AC0, AC1)
}

; Store Calculated Output

*ar_coef+ = HI(AC0)
|| *ar_output+ = HI(rnd(AC1))

; AC2 is error amount

; Point to oldest data sample
AC2 = (*ar_des+ << #16) − AC1
|| mar(*ar_data+)

; Place updated mu_error term in AC1

AC1 = rnd(T_step*AC2)
}

Note: This example shows portions of the file dlms.asm in the TI C55x DSPLIB (introduced in Chapter 8).

B-38
 Algebraic Instructions Code Examples

Example B−37. Generation of Output Streams G0 and G1

AR3 = #09H ; AC0_H

AR1 = #in_bit_stream

; Load 32 bits into the accumulators

AC0 = *AR1+
AC0 = AC0 + (*AR1 << #16)
AC1 = AC0

; Generate G0
AC0 = AC0 ^ (AC1 << #−1) ; A = A XOR B>>1
AC0 = AC0 ^ (AC1 << #−3) ; A = A XOR B>>2
T0 = AC0 ; Save G0

; Generate G1
AC0 = AC0 ^ (AC1 << #−1) ; A = A XOR B>>1
AC0 = AC0 ^ (AC1 << #−3) ; A = A XOR B>>3
AC0 = AC0 ^ (AC1 << #−4) ; A = A XOR B>>4 −−> AC0_L = G1

*AR3 = T0 ; AC0_H = G0 −−−−−−> AC0 = G0G1

Example B−38. Viterbi Butterflies for Channel Coding

BFLY_DIR_ALG .MACRO
;new_metric(j)&(j+2^(K−2))
hi(AC0) = *AR5+ + T3, ; AC0(39−16) = Old_Met(2*j)+LD
lo(AC0) = *AR5+ − T3 ; AC0(15−0) = Old_met(2*j+1)−LD

hi(AC1) = *AR5+ − T3, ; AC1(39−16) = Old_Met(2*j)−LD

lo(AC1) = *AR5+ + T3 ; AC1(15−0) = Old_met(2*j+1)+LD

max_diff(AC0, AC1, AC2, AC3) ; Compare AC0 and AC1

*AR3+ = lo(AC2), *AR4+ = hi(AC2) ; Store the lower maxima (with AR3)
; and upper maxima (with AR4)
.ENDM

BFLY_REV_ALG .MACRO
;new_metric(j)&(j+2^(K−2))
hi(AC0) = *AR5+ − T3, ; AC0(39−16) = Old_Met(2*j)−LD
lo(AC0) = *AR5+ + T3 ; AC0(15−0) = Old_met(2*j+1)+LD

hi(AC1) = *AR5+ + T3, ; AC1(39−16) = Old_Met(2*j)+LD

lo(AC1) = *AR5+ − T3 ; AC1(15−0) = Old_met(2*j+1)−LD

max_diff(AC0, AC1, AC2, AC3) ; Compare AC0 and AC1

*AR3+ = lo(AC2), *AR4+ = hi(AC2) ; Store the lower maxima (with AR3)
; and upper maxima (with AR4)
.ENDM

Algebraic Instructions Code Examples B-39

Algebraic Instructions Code Examples

Example B−39. Viterbi Butterflies Using Instruction Parallelism

localrepeat {
butterfly:
hi(AC0) = *AR0+ + T3, ; AC0(39−16) = Old_Met(2*j)+LD
lo(AC0) = *AR0+ − T3 ; AC0(15−0) = Old_met(2*j+1)−LD
|| AR7 = *AR5+

hi(AC1) = *AR0+ − T3, ; AC1(39−16) = Old_Met(2*j)−LD

lo(AC1) = *AR0+ + T3 ; AC1(15−0) = Old_met(2*j+1)+LD
|| T3 = *AR6 ; Load new local distance

*AR2+ = lo(AC2), *AR2(T0) = hi(AC2) ; Store lower and upper maxima

; from previous max_diff operation
|| max_diff( AC0, AC1, AC2, AC3) ; Compare AC0 and AC1

hi(AC0) = *AR0+ + T3, ; AC0(39−16) = Old_Met(2*j)+LD

lo(AC0) = *AR0+ − T3 ; AC0(15−0) = Old_met(2*j+1)−LD
|| AR6 = *AR5+

hi(AC1) = *AR0+ − T3, ; AC1(39−16) = Old_Met(2*j)−LD

lo(AC1) = *AR0+ + T3 ; AC1(15−0) = Old_met(2*j+1)+LD
|| T3 = *AR7 ; Load new local distance

*AR2(T0) = lo(AC2), *AR2+ = hi(AC2) ; Store lower and upper maxima

; from previous max_diff operation
|| max_diff( AC0, AC1, AC2, AC3) ; Compare AC0 and AC1
}

B-40
 Index



2s-complement addition address buses available 1-2

concept 5-5 address generation units 1-2
extended-precision 5-9 addressing modes
2s-complement arithmetic ARn 2-8
concepts 5-5 bit addressing by instruction 2-8
extended-precision (details) 5-9, 5-14 DP direct 2-8
2s-complement division 5-17 k23 2-8
2s-complement fractional format 5-5 ADDSUB instruction used in Viterbi code 7-19
2s-complement integer format 5-4 advanced parallelism rules 4-22
2s-complement multiplication aliasing 3-5
concept 5-5 analyzing C performance 3-6
extended-precision 5-14 application-specific instructions 7-1
2s-complement numbers architecture of TMS320C55x DSPs
concept 5-2 features supporting parallelism 4-17
decimal equivalent (example) 5-3 overview 1-2
negative of (example) 5-3
arguments to functions in TMS320C55x DSP
2s-complement subtraction library 8-2
concept 5-5
arithmetic
extended-precision 5-9
2s-complement 5-5
extended-precision 2s-complement
A (details) 5-9, 5-14
assembly code optimization 4-1
A unit, parallel optimization within 4-25 asymmetric FIR filtering 7-2
AC1 (access 1) phase of pipeline 4-50 with FIRSN instruction 7-4
AC2 (access 2) phase of pipeline 4-50
access 1 (AC1) phase of pipeline 4-50
access 2 (AC2) phase of pipeline 4-50
B
accesses to dual-access RAM (DARAM) 4-73 B bus (BB), use in dual-MAC operations 4-2
AD (address) phase of pipeline 4-50 basic parallelism rules 4-21
adaptive FIR filtering BB bus, use in dual-MAC operations 4-2
concept 7-6 bit field expand instruction
example 7-9 concept 7-12
addition example 7-13
2s-complement 5-5 bit field extract instruction
extended-precision 2s-complement 5-9 concept 7-14
address (AD) phase of pipeline 4-50 example 7-15

Index-1
Index

bit-reverse addressing 6-1 comments 2-3

in FFT algorithms 6-4 compatibility with TMS320C54x DSPs 1-2
in TMS320C55x DSP function library 6-7
compiler options
in-place versus off-place bit-reversing 6-6
for information 3-9
introduction 6-2
recommended to improve performance 3-8
syntaxes 6-2
that may improve code size 3-8
typical initialization requirements 6-5
to avoid on performance-critical code 3-7
bit-stream (de)multiplexing 7-12
complex vector multiplication using dual-MAC
block-repeat instructions, identifying which type to hardware 4-4
use 4-45
computation blocks 1-2
built-in parallelism 4-16
computed single-repeat register (CSR), using with a
buses available 1-2 single-repeat instruction 4-46
buses supporting parallelism 4-19
constants 2-5
butterfly structure (figure) 7-17
control code tips 3-39
byte extensions (parallelism rule) 4-22
convolutional encoding 7-10
CSR, using with a single-repeat instruction 4-46
C
C bus (CB), use in dual-MAC operations 4-2 D
C code, optimizing 3-1
C code performance analysis 3-6 D (decode) phase of pipeline 4-50
C function library for TMS320C55x DSPs 8-1 D bus (DB), use in dual-MAC operations 4-2
calling a function from assembly source 8-4 D unit, parallel optimization within 4-33
calling a function from C 8-3 DARAM accesses (buses and pipeline
data types used 8-2 phases) 4-73
function arguments 8-2
data and code allocation tips 3-44, 3-48
list of functions 8-5
where to find sample code 8-4 data buses available 1-2
C/C++ code optimization techniques 3-15 data type tips 3-2
calling a function in TMS320C55x DSP library data types used in TMS320C55x DSP function
from assembly source 8-4 library 8-2
from C 8-3 DATA_ALIGN pragma 3-33
CARRY bit 5-9 DATA_SECTION pragma 3-46
effected by addition 5-11 DB bus, use in dual-MAC operations 4-2
effected by subtraction 5-13
decode (D) phase of pipeline 4-50
CB bus, use in dual-MAC operations 4-2
delayed LMS algorithm 7-7
char size in C55x compiler 3-2
dependences between instructions 3-4
circular addressing 3-35
simulating in C 3-35 directives
.def 2-5
clock() function for profiling 3-13
.init 2-5
code and data allocation tips 3-44, 3-48 MEMORY 2-10
code development flow 1-3 .sect 2-5
code size compiler options 3-8 SECTIONS 2-10
CODE_SECTION pragma 3-48 .usect 2-5
coding idioms in C for generating efficient division, 2s-complement 5-17
assembly 3-40 double size in C55x compiler 3-2

Index-2
 Index

DSP function library for TMS320C55x DSPs 8-1 extended auxiliary register usage in dual-MAC
calling a function from assembly source 8-4 operations 4-3
calling a function from C 8-3 extended coefficient data pointer usage in
data types used 8-2 dual-MAC operations 4-3
function arguments 8-2
extended-precision 2s-complement arithmetic,
list of functions 8-5
details 5-9, 5-14
where to find sample code 8-4
dual-MAC operations, helping the C compiler
generate them 3-22 F
dual-access RAM (DARAM) accesses (buses and
pipeline phases) 4-73 FFT flow graph with bit-reversed input 6-4
dual-MAC hardware/operations field expand instruction
efficient use of 4-2 concept 7-12
matrix multiplication 4-13 example 7-13
pointer usage in 4-3 field extract instruction
taking advantage of implicit algorithm concept 7-14
symmetry 4-4 example 7-15
dynamic scaling for overflow handling 5-25 filtering
adaptive FIR (concept) 7-6
adaptive FIR (example) 7-9
symmetric/asymmetric FIR 7-2
E with dual-MAC hardware 4-4, 4-6, 4-14
FIR filtering
encoding, convolutional 7-10 adaptive (concept) 7-6
examples adaptive (example) 7-9
adaptive FIR filter (delayed LMS) 7-9 symmetric/asymmetric 7-2
block FIR filter (dual-MAC implementation) 4-9, FIRS instruction 7-3, 7-4
4-10
FIRSN instruction 7-4
branch-on-auxiliary-register-not-zero loop
construct 4-44 fixed scaling for overflow handling 5-25
complex vector multiplication (dual-MAC fixed-point arithmetic 5-1
implementation) 4-6 float size in C55x compiler 3-2
demultiplexing bit streams (convolutional fractional representation (2s-complement) 5-5
encoder) 7-15
fractions versus integers 5-3
multiplexing bit streams (convolutional
encoder) 7-13 function inlining 3-11
nested loops 4-43 function library for TMS320C55x DSPs 8-1
output streams for convolutional encoder 7-11 calling a function from assembly source 8-4
parallel optimization across CPU functional calling a function from C 8-3
units 4-35 data types used 8-2
parallel optimization within A unit 4-25 function arguments 8-2
parallel optimization within CPU functional list of functions 8-5
units 4-25 where to find sample code 8-4
parallel optimization within D unit 4-33
parallel optimization within P unit 4-30
symmetric FIR filter 7-5 G
use of CSR for single-repeat loop 4-47
Viterbi butterfly 7-19 global symbols 3-43
execute (X) phase of pipeline 4-51 guard bits used for overflow handling 5-24

Index-3
Index

intrinsics (continued)
H long_lssub 3-32
long_smac 3-32
hardware resource conflicts (parallelism rule) 4-21 long_smas 3-32
holes caused by data alignment 3-42

K
I keyword
inline 3-11
I/O space available 1-2 onchip 3-23
IBQ, minimizing delays 4-49 restrict 3-4, 3-5, 3-26
idioms in C for generating efficient assembly 3-40
if-then-else and switch/case constructs 3-39 L
implicit algorithm symmetry 4-4 labels 2-3
inline keyword 3-11 least mean square (LMS) calculation for adaptive
input scaling for overflow handling 5-25 filtering 7-6
input/output space available 1-2 least mean square (LMS) instruction 7-6
instruction buffer 1-2 linker command file sample 3-47
linking 2-10
instruction length limitation (parallelism rule) 4-21
local block-repeat instruction, when to use 4-45
instruction pipeline, introduced 1-2
local symbols 3-43
instructions executed in parallel 4-16
localrepeat/RPTBLOCAL instruction 3-16
instructions for specific applications 7-1 long data accesses for 16-bit data 3-33
int _norm 3-33 long long size in C55x compiler 3-2
int size in C55x compiler 3-2 long size in C55x compiler 3-2
integer representation (2s-complement) 5-4 loop code tips 3-16
integers versus fractions 5-3 loop unrolling 4-6
interrupt-control logic 1-2 loop-control registers
avoiding pipeline delays when accessing 4-48
intrinsics 3-29 to 3-33
when they are accessed in the pipeline 4-64
int_abss 3-32
int_lnorm 3-33 loops
int_lshrs 3-33 implementing efficient 4-42
int_md 3-33 nesting of 4-42
int_norm 3-33
int_sadd 3-32
int_shrs 3-33
M
int_smacr 3-33 MAC hardware tips 3-21
int_smasr 3-33 MAC units 1-2
int_smpy 3-32, 3-33 map file, example 2-11
int_sneg 3-32 matrix mathematics using dual-MAC
int_ssubb 3-32 hardware 4-13
long_labss 3-32
MAXDIFF instruction used in Viterbi code 7-19
long_lsadd 3-32
long_lsmpy 3-32 maximum instruction length (parallelism rule) 4-21
long_lsneg 3-32 memory accesses and the pipeline 4-72
long_lssh 3-33 memory available 1-2
long_lsshl 3-33 memory dependences 3-4

Index-4
 Index

memory management 3-42

allocating data 3-42 P
data alignment 3-42 P unit, parallel optimization within 4-30
stack configuration 3-42
parallel enable bit 4-21
symbol declaration 3-42
parallel execution features 4-16
mmap() qualifier (parallelism rule) 4-22
parallel optimization, examples 4-25, 4-35
multi-algorithm applications enhanced with
parallelism
dual-MAC hardware 4-15
architectural features 4-17
multichannel applications implemented with built-in 4-16
dual-MAC hardware 4-14 user-defined 4-16
multiplexing bit streams (convolutional parallelism rules for user-defined parallelism 4-20
encoder) 7-12 parallelism tips 4-24
multiplication performance tips for C code 3-15
2s-complement 5-5
pipeline
extended-precision 2s-complement 5-14
introduced 1-2
writing multiplication expressions in C 3-3
minimizing delays 4-49
multiply-and-accumulate (MAC) 3-21 register groups 4-64
hardware tips 3-21 segments and phases 4-49
multiply-and-accumulate (MAC) units 1-2 pipeline conflicts, process to resolve 4-53
MUST_ITERATE pragma 3-16 to 3-20, 3-25 to pipeline delays
3-27 recommendations for preventing 4-54
when accessing loop-control registers 4-48
−pm compiler option 3-9
N pointer usage in dual-MAC operations 4-3
negative of 2s-complement number (example) 5-3 port() qualifier (parallelism rule) 4-22
nesting of loops 4-42 pragma
CODE_SECTION 3-48
DATA_ALIGN 3-33
O DATA_SECTION 3-46
MUST_ITERATE 3-16 to 3-20, 3-25 to 3-27
onchip keyword 3-23 procedure for efficient code 1-3
onchip type qualifier 3-23 process for user-defined parallelism 4-23
operands 3-46 process to resolve pipeline conflicts 4-53
global 3-46 processor initialization 2-3, 2-7
local 3-46 profiling C code 3-13
operators supporting parallelism 4-17 program-level optimization 3-9
optimization in compiler, program-level 3-9
optimizing assembly code 4-1
optimizing C code 3-1
R
overflow flags used for overflow handling 5-24 R (read) phase of pipeline 4-51
Overflow handling 5-24 read (R) phase of pipeline 4-51
overflow handling recommendations for preventing pipeline
FFTs 5-26 delays 4-54
FIR filters 5-25 refining C/C++ code 3-15
hardware features for 5-24 registers, when they are accessed in the
IIR filters 5-26 pipeline 4-56, 4-64
techniques for 5-25 resource conflicts (parallelism rule) 4-21

Index-5
Index

restrict keyword 3-4, 3-5, 3-26 switch/case and if-then-else constructs 3-39
RPTBLOCAL/localrepeat instruction 3-16 symbol declarations
rules for user-defined parallelism 4-20 global vs. local 3-43
local vs. global 3-43
symmetric FIR filtering 7-2
S with FIRS instruction (concept) 7-3
with FIRS instruction (example) 7-4
saturation for overflow handling 5-25
saturation mode bits used for overflow
handling 5-24
scaling
T
dynamic 5-25
fixed 5-25 table to help generate optional application
input 5-25 mapping 4-78
methods for FFTs 5-26 TI C55x DSPLIB 8-1
methods for FIR filters 5-25 calling a function in assembly source 8-4
methods for IIR filters 5-26 calling a function in C 8-3
scaling methods (overflow handling) data types used 8-2
FFTs 5-26 function arguments 8-2
FIR filters 5-25 list of functions 8-5
IIR filters 5-26 where to find sample code 8-4
section allocation 2-5 tips
example 2-6 applying parallelism 4-24
code and data allocation 3-44, 3-48
sections 3-45
control code 3-39
.bss 3-45
data and code allocation 3-44, 3-48
.cinit 3-45
data types 3-2
.const 3-45
loop code 3-16
.do 3-45
MAC hardware 3-21
.ioport 3-45
nesting loops 4-45
.stack 3-45
performance for C code 3-15
.switch 3-45
preventing pipeline delays 4-54
.sysmem 3-45
producing efficient code 1-3
.sysstack 3-45
resolving pipeline conflicts 4-53
.text 3-45
short size in C55x compiler 3-2 TMS320C54x-compatible mode 1-2
single-repeat instruction, when to use 4-45 TMS320C55x DSP function library 8-1
calling a function from assembly source 8-4
size of C data types 3-2
calling a function from C 8-3
soft dual encoding 4-21 data types used 8-2
stack configuration 3-43 function arguments 8-2
stacks available 1-2 list of functions 8-5
where to find sample code 8-4
standard block-repeat instruction, when to
use 4-45 transition registers (TRN0, TRN1) used in Viterbi
algorithm 7-18
SUBADD instruction used in Viterbi code 7-19
subtraction trip count 3-17
2s-complement 5-5 unsigned integer types 3-17
extended-precision 2s-complement 5-9 tutorial 2-1

Index-6
 Index

U W
unroll-and-jam transformation 3-25 W (write) phase of pipeline 4-51
user-defined parallelism 4-16 W+ (write+) phase of pipeline 4-51
process 4-23 write (W) phase of pipeline 4-51
rules 4-20
write+ (W+) phase of pipeline 4-51
writing assembly code 2-3
V
variables 2-5 X
vector multiplication using dual-MAC hardware 4-4
Viterbi algorithm for channel decoding 7-16 X (execute) phase of pipeline 4-51
Viterbi butterfly XARn, example 2-9
examples 7-19 XARn usage in dual-MAC operations 4-3
figure 7-17 XCDP usage in dual-MAC operations 4-3

Index-7
Index-8