TMS320C64x DSP

Two-Level Internal Memory

Reference Guide

Literature Number: SPRU610C

February 2006
 IMPORTANT NOTICE

Texas Instruments Incorporated and its subsidiaries (TI) reserve the right to make corrections, modifications,
enhancements, improvements, and other changes to its products and services at any time and to discontinue
any product or service without notice. Customers should obtain the latest relevant information before placing
orders and should verify that such information is current and complete. All products are sold subject to TI’s terms
and conditions of sale supplied at the time of order acknowledgment.

TI warrants performance of its hardware products to the specifications applicable at the time of sale in
accordance with TI’s standard warranty. Testing and other quality control techniques are used to the extent TI
deems necessary to support this warranty. Except where mandated by government requirements, testing of all
parameters of each product is not necessarily performed.

TI assumes no liability for applications assistance or customer product design. Customers are responsible for
their products and applications using TI components. To minimize the risks associated with customer products
and applications, customers should provide adequate design and operating safeguards.

TI does not warrant or represent that any license, either express or implied, is granted under any TI patent right,
copyright, mask work right, or other TI intellectual property right relating to any combination, machine, or process
in which TI products or services are used. Information published by TI regarding third-party products or services
does not constitute a license from TI to use such products or services or a warranty or endorsement thereof.
Use of such information may require a license from a third party under the patents or other intellectual property
of the third party, or a license from TI under the patents or other intellectual property of TI.

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. Reproduction
of this information with alteration is an unfair and deceptive business practice. TI is not responsible or liable for
such altered documentation.

Resale of TI products or services with statements different from or beyond the parameters stated by TI for that
product or service voids all express and any implied warranties for the associated TI product or service and
is an unfair and deceptive business practice. TI is not responsible or liable for any such statements.

Following are URLs where you can obtain information on other Texas Instruments products and application
solutions:

Products Applications
Amplifiers amplifier.ti.com Audio www.ti.com/audio
Data Converters dataconverter.ti.com Automotive www.ti.com/automotive
DSP dsp.ti.com Broadband www.ti.com/broadband
Interface interface.ti.com Digital Control www.ti.com/digitalcontrol
Logic logic.ti.com Military www.ti.com/military
Power Mgmt power.ti.com Optical Networking www.ti.com/opticalnetwork
Microcontrollers microcontroller.ti.com Security www.ti.com/security
Telephony www.ti.com/telephony
Video & Imaging www.ti.com/video
Wireless www.ti.com/wireless

Mailing Address: Texas Instruments

Post Office Box 655303 Dallas, Texas 75265

Copyright  2006, Texas Instruments Incorporated

 Preface

Read This First

About This Manual

The TMS320C64x digital signal processors (DSPs) of the
TMS320C6000 DSP family have a two-level memory architecture for
program and data. The first-level program cache is designated L1P, and the
first-level data cache is designated L1D. Both the program and data memory
share the second-level memory, designated L2. L2 is configurable, allowing
for various amounts of cache and SRAM. This document discusses the C64x
two-level internal memory.

Notational Conventions
This document uses the following conventions.
- Hexadecimal numbers are shown with the suffix h. For example, the
following number is 40 hexadecimal (decimal 64): 40h.
- Registers in this document are shown in figures and described in tables.
J Each register figure shows a rectangle divided into fields that represent
the fields of the register. Each field is labeled with its bit name, its
beginning and ending bit numbers above, and its read/write properties
below. A legend explains the notation used for the properties.
J Reserved bits in a register figure designate a bit that is used for future
device expansion.

Related Documentation From Texas Instruments

The following documents describe the C6000 devices and related support
tools. Copies of these documents are available on the Internet at www.ti.com.
Tip: Enter the literature number in the search box provided at www.ti.com.
TMS320C6000 CPU and Instruction Set Reference Guide (literature
number SPRU189) describes the TMS320C6000 CPU architecture,
instruction set, pipeline, and interrupts for these digital signal processors.
TMS320C6000 DSP Peripherals Overview Reference Guide (literature
number SPRU190) describes the peripherals available on the
TMS320C6000 DSPs.

SPRU610C TMS320C64x Two-Level Internal Memory 3

Trademarks
Related Documentation From Texas Instruments / Trademarks

TMS320C64x Technical Overview (SPRU395) The TMS320C64x technical

overview gives an introduction to the TMS320C64x digital signal
processor, and discusses the application areas that are enhanced by the
TMS320C64x VelociTI.

TMS320C6000 DSP Cache User’s Guide (literature number SPRU656)

explains the fundamentals of memory caches and describes how to
efficiently utilize the two-level cache-based memory architecture in the
digital signal processors (DSPs) of the TMS320C6000 DSP family. It
shows how to maintain coherence with external memory, how to use
DMA to reduce memory latencies, and how to optimize your code to
improve cache efficiency.

TMS320C6000 Programmer’s Guide (literature number SPRU198)

describes ways to optimize C and assembly code for the
TMS320C6000 DSPs and includes application program examples.

TMS320C6000 Code Composer Studio Tutorial (literature number

SPRU301) introduces the Code Composer Studio integrated develop-
ment environment and software tools.

Code Composer Studio Application Programming Interface Reference

Guide (literature number SPRU321) describes the Code Composer
Studio application programming interface (API), which allows you to
program custom plug-ins for Code Composer.

TMS320C6x Peripheral Support Library Programmer’s Reference

(literature number SPRU273) describes the contents of the
TMS320C6000 peripheral support library of functions and macros. It
lists functions and macros both by header file and alphabetically,
provides a complete description of each, and gives code examples to
show how they are used.

TMS320C6000 Chip Support Library API Reference Guide (literature

number SPRU401) describes a set of application programming interfaces
(APIs) used to configure and control the on-chip peripherals.

Trademarks

Code Composer Studio, C6000, C62x, C64x, C67x, TMS320C6000,

TMS320C62x, TMS320C64x, TMS320C67x, and VelociTI are trademarks of
Texas Instruments.

4 TMS320C64x Two-Level Internal Memory SPRU610C

 Contents

Contents

1 Memory Hierarchy Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Cache Terms and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Level 1 Data Cache (L1D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.1 L1D Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 L1D Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.1 L1D Memory Banking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.2 L1D Miss Penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.3 L1D Write Buffer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.4 L1D Miss Pipelining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4 Level 1 Program Cache (L1P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.1 L1P Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 L1P Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 L1P Miss Penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.2 L1P Miss Pipelining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Level 2 Unified Memory (L2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.1 L2 Cache and L2 SRAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.2 L2 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.3 L2 Bank Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.4 L2 Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.4.1 L1D/L1P-to-L2 Request Servicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.4.2 EDMA-to-L2 Request Servicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.4.3 L2 Request Servicing Using EDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.4.4 EDMA Access to Cache Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.4.5 HPI and PCI Access to Memory Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6 Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.1 Cache Configuration Register (CCFG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.2 L2 EDMA Access Control Register (EDMAWEIGHT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.3 L2 Allocation Registers (L2ALLOC0−L2ALLOC03) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4 L2 Writeback Base Address Register (L2WBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5 L2 Writeback Word Count Register (L2WWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.6 L2 Writeback−Invalidate Base Address Register (L2WIBAR) . . . . . . . . . . . . . . . . . . . . . . 44
6.7 L2 Writeback−Invalidate Word Count Register (L2WIWC) . . . . . . . . . . . . . . . . . . . . . . . . . 44

SPRU610C TMS320C64x Two-Level Internal Memory 5

Contents

6.8 L2 Invalidate Base Address Register (L2IBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6.9 L2 Invalidate Word Count Register (L2IWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.10 L1P Invalidate Base Address Register (L1PIBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.11 L1P Invalidate Word Count Register (L1PIWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.12 L1D Writeback−Invalidate Base Address Register (L1DWIBAR) . . . . . . . . . . . . . . . . . . . 47
6.13 L1D Writeback−Invalidate Word Count Register (L1DWIWC) . . . . . . . . . . . . . . . . . . . . . . 47
6.14 L1D Invalidate Base Address Register (L1DIBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.15 L1D Invalidate Word Count Register (L1DIWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.16 L2 Writeback All Register (L2WB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.17 L2 Writeback−Invalidate All Register (L2WBINV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.18 L2 Memory Attribute Registers (MAR0−MAR255) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

7 Memory System Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

7.1 Cache Mode Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.1.1 L1D Mode Selection Using DCC Field in CSR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.1.2 L1P Mode Selection Using PCC Field in CSR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.1.3 L2 Mode Selection Using L2MODE Field in CCFG . . . . . . . . . . . . . . . . . . . . . . . . 53
7.2 Cacheability Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7.3 Program-Initiated Cache Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.3.1 Global Cache Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.3.2 Block Cache Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3.3 Effect of L2 Commands on L1 Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.4 L2-to-EDMA Request Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
7.5 EDMA Access Into L2 Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

8 Memory System Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

8.1 Memory System Coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.2 EDMA Coherence in L2 SRAM Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
8.3 Memory Access Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
8.3.1 Program Order of Memory Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
8.3.2 Strong and Relaxed Memory Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Revision History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6 TMS320C64x Two-Level Internal Memory SPRU610C

 Figures

Figures
1 TMS320C64x DSP Block Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 TMS320C64x Two-Level Internal Memory Block Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 L1D Address Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4 Address to Bank Number Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5 Potentially Conflicting Memory Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6 L1P Address Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
7 L2 Address Allocation, 256K Cache (L2MODE = 111b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
8 L2 Address Allocation, 128K Cache (L2MODE = 011b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
9 L2 Address Allocation, 64K Cache (L2MODE = 010b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
10 L2 Address Allocation, 32K Cache (L2MODE = 001b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
11 Cache Configuration Register (CCFG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
12 L2 EDMA Access Control Register (EDMAWEIGHT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
13 L2 Allocation Registers (L2ALLOC0−L2ALLOC3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
14 L2 Writeback Base Address Register (L2WBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
15 L2 Writeback Word Count Register (L2WWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
16 L2 Writeback−Invalidate Base Address Register (L2WIBAR) . . . . . . . . . . . . . . . . . . . . . . . . . 44
17 L2 Writeback−Invalidate Word Count Register (L2WIWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
18 L2 Invalidate Base Address Register (L2IBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
19 L2 Invalidate Word Count Register (L2IWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
20 L1P Invalidate Base Address Register (L1PIBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
21 L1P Invalidate Word Count Register (L1PIWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
22 L1D Writeback−Invalidate Base Address Register (L1DWIBAR) . . . . . . . . . . . . . . . . . . . . . . 47
23 L1D Writeback−Invalidate Word Count Register (L1DWIWC) . . . . . . . . . . . . . . . . . . . . . . . . 47
24 L1D Invalidate Base Address Register (L1DIBAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
25 L1D Invalidate Word Count Register (L1DIWC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
26 L2 Writeback All Register (L2WB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
27 L2 Writeback-Invalidate All Register (L2WBINV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
28 L2 Memory Attribute Register (MAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
29 CPU Control and Status Register (CSR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
30 Block Cache Operation Base Address Register (BAR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
31 Block Cache Operation Word Count Register (WC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
32 Streaming Data Pseudo-Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
33 Double Buffering Pseudo-Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
34 Double-Buffering Time Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
35 Double Buffering as a Pipelined Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

SPRU610C TMS320C64x Two-Level Internal Memory 7

Tables

Tables
1 TMS320C621x/C671x/C64x Internal Memory Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Terms and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Cycles Per Miss for Different Numbers of L1D Misses That Hit L2 Cache . . . . . . . . . . . . . . 26
4 Cycles Per Miss for Different Numbers of L1D Misses that Hit L2 SRAM . . . . . . . . . . . . . . . 26
5 Average Miss Penalties for Large Numbers of Sequential Execute Packets . . . . . . . . . . . . 29
6 Internal Memory Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7 Cache Configuration Register (CCFG) Field Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
8 L2 EDMA Access Control Register (EDMAWEIGHT) Field Descriptions . . . . . . . . . . . . . . . 41
9 L2 Allocation Registers (L2ALLOC0−L2ALLOC3) Field Descriptions . . . . . . . . . . . . . . . . . . 42
10 L2 Writeback Base Address Register (L2WBAR) Field Descriptions . . . . . . . . . . . . . . . . . . . 43
11 L2 Writeback Word Count Register (L2WWC) Field Descriptions . . . . . . . . . . . . . . . . . . . . . 43
12 L2 Writeback−Invalidate Base Address Register (L2WIBAR) Field Descriptions . . . . . . . . 44
13 L2 Writeback−Invalidate Word Count Register (L2WIWC) Field Descriptions . . . . . . . . . . . 44
14 L2 Invalidate Base Address Register (L2IBAR) Field Descriptions . . . . . . . . . . . . . . . . . . . . 45
15 L2 Invalidate Word Count Register (L2IWC) Field Descriptions . . . . . . . . . . . . . . . . . . . . . . . 45
16 L1P Invalidate Base Address Register (L1PIBAR) Field Descriptions . . . . . . . . . . . . . . . . . 46
17 L1P Invalidate Word Count Register (L1PIWC) Field Descriptions . . . . . . . . . . . . . . . . . . . . 46
18 L1D Writeback−Invalidate Base Address Register (L1DWIBAR) Field Descriptions . . . . . 47
19 L1D Writeback−Invalidate Word Count Register (L1DWIWC) Field Descriptions . . . . . . . . 47
20 L1D Invalidate Base Address Register (L1DIBAR) Field Descriptions . . . . . . . . . . . . . . . . . 48
21 L1D Invalidate Word Count Register (L1DIWC) Field Descriptions . . . . . . . . . . . . . . . . . . . . 48
22 L2 Writeback All Register (L2WB) Field Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
23 L2 Writeback−Invalidate All Register (L2WBINV) Field Descriptions . . . . . . . . . . . . . . . . . . . 50
24 Memory Attribute Register (MAR) Field Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
25 L1D Mode Setting Using DCC Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
26 L1P Mode Setting Using PCC Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
27 L2 Mode Switch Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
28 Memory Attribute Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
29 Summary of Program-Initiated Cache Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
30 L2ALLOC Default Queue Allocations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
31 Coherence Assurances in the Two-Level Memory System . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
32 Program Order for Memory Operations Issued From a Single Execute Packet . . . . . . . . . 76
33 Document Revision History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

8 TMS320C64x Two-Level Internal Memory SPRU610C

 TMS320C64x TwoĆLevel Internal Memory

The TMS320C621x, TMS320C671x, and TMS320C64x digital signal

processors (DSPs) of the TMS320C6000 DSP family have a two-level
memory architecture for program and data. The first-level program cache is
designated L1P, and the first-level data cache is designated L1D. Both the
program and data memory share the second-level memory, designated L2. L2
is configurable, allowing for various amounts of cache and SRAM. This
document discusses the C64x two-level internal memory. For a discussion
of the C621x/C671x two-level internal memory, see TMS320C621x/C671x DSP
Two-Level Internal Memory Reference Guide (SPRU609).

1 Memory Hierarchy Overview

Figure 1 shows the block diagram of the C64x DSP. Table 1 summarizes the
differences between the C621x/C671x and C64x internal memory. Figure 2
illustrates the bus connections between the CPU, internal memories, and the
enhanced DMA (EDMA) of the C6000 DSP.

Figure 1. TMS320C64x DSP Block Diagram

EMIFA
L1P cache

EMIFB

C6000 DSP core

Instruction fetch Control
Other Instruction dispatch registers
peripherals
Enhanced Instruction decode Control
L2 logic
DMA memory
controller Data path A Data path B
Test
A register file B register file
Interrupt In-circuit
selector emulation
L1 S1 M1 D1 D2 M2 S2 L2 Interrupt
control
Power down
logic

Boot L1D cache

configuration PLL

Note: EMIFB is available only on certain C64x devices. Refer to the device-specific data sheet for the available peripheral set.

SPRU610C TMS320C64x Two-Level Internal Memory 9

Memory Hierarchy Overview

Table 1. TMS320C621x/C671x/C64x Internal Memory Comparison

TMS320C621x/C671x DSP TMS320C64x DSP

Internal memory structure Two Level

L1P size 4 Kbytes 16 Kbytes

L1P organization Direct mapped

L1P CPU access time 1 cycle

L1P line size 64 bytes 32 bytes

L1P read miss action 1 line allocated in L1P

L1P read hit action Data read from L1P

L1P write miss action L1P writes not supported

L1P write hit action L1P writes not supported

L1P → L2 request size 2 fetches/L1P line 1 fetch/L1P line

L1P protocol Read Allocate Read Allocate; Pipelined Misses

L1P memory Single-cycle RAM

L1P → L2 single request stall 5 cycles for L2 hit 8 cycles for L2 hit

L1P → L2 minimum cycles between Pipelined misses not supported 1 cycle

pipelined misses

L1D size 4 Kbytes 16 Kbytes

L1D organization 2-way set associative

L1D CPU access time 1 cycle

L1D line size 32 bytes 64 bytes

L1D replacement strategy 2-way Least Recently Used

L1D banking 64-bit-wide dual-ported RAM 8 × 32 bit banks

L1D read miss action 1 line allocated in L1D

L1D read hit action Data read from L1D

L1D write miss action No allocation in L1D, data sent to L2

L1D write hit action Data updated in L1D; line marked dirty

L1D protocol Read Allocate Read allocate; Pipelined Misses

L1D → L2 request size 2 fetches/L1D line

† Some C64x devices may not support the 256K cache mode. Refer to the device-specific datasheet.

10 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory Hierarchy Overview

Table 1. TMS320C621x/C671x/C64x Internal Memory Comparison (Continued)

TMS320C621x/C671x DSP TMS320C64x DSP
L1D → L2 single request stall 4 cycles for L2 hit 6 cycles/L2 SRAM hit

L1D → L2 minimum cycles between Pipelined misses not supported 2 cycles

pipelined misses

L2 total size Varies by part number. Refer to the datasheet for the specific device.

L2 SRAM size Varies by part number. Refer to the datasheet for the specific device.

L2 cache size† 0/16/32/48/64 Kbytes 0/32/64/128/256 Kbytes

L2 organization 1/2/3/4-way set associative 4-way set associative cache

L2 line size 128 bytes

L2 replacement strategy 1/2/3/4-way Least Recently Used 4-way Least Recently Used

L2 banking 4 × 64 bit banks 8 × 64 bit banks

L2-L1P protocol Coherency invalidates

L2-L1D protocol Coherency snoop-invalidates Coherency snoops and

snoop-invalidates

L2 protocol Read and Write Allocate

L2 read miss action Data is read via EDMA into newly allocated line in L2; requested
data is passed to the requesting L1

L2 read hit action Data read from L2

L2 write miss action Data is read via EDMA into newly allocated line in L2; write data is
then written to the newly allocated line.

L2 write hit action Data is written into hit L2 location

L2 → L1P read path width 256 bit

L2 → L1D read path width 128 bit 256 bit

L1D → L2 write path width 32 bit 64 bit

L1D → L2 victim path width 128 bit 256 bit

L2 → EDMA read path width 64 bit

L2 → EDMA write path width 64 bit

† Some C64x devices may not support the 256K cache mode. Refer to the device-specific datasheet.

SPRU610C TMS320C64x Two-Level Internal Memory 11

Memory Hierarchy Overview

Figure 2. TMS320C64x Two-Level Internal Memory Block Diagram

Snoop address
Cache L1 program cache Data
RAM controller Address

Program Program
address data

C6000 CPU
RAM
Program fetch L2 cache
controller

Data path A Data path B EDMA

ST1 store data

ST2 store data

LD1 load data

LD2 load data

DA1 address

DA2 address

Address
Data
Cache L1 data cache
RAM controller Data
Snoop address

12 TMS320C64x Two-Level Internal Memory SPRU610C

 Cache Terms and Definitions

2 Cache Terms and Definitions

Table 2 lists the terms used throughout this document that relate to the
operation of the C64x two-level memory hierarchy.

Table 2. Terms and Definitions

Term Definition
Allocation The process of finding a location in the cache to store newly cached data. This
process can include evicting data that is presently in the cache to make room for the
new data.
Associativity The number of line frames in each set. This is specified as the number of ways in the
cache.

Capacity miss A cache miss that occurs because the cache does not have sufficient room to hold the
entire working set for a program. Compare with compulsory miss and conflict miss.

Clean A cache line that is valid and that has not been written to by upper levels of memory
or the CPU. The opposite state for a valid cache line is dirty.

Coherence Informally, a memory system is coherent if any read of a data item returns the most
recently written value of that data item. This includes accesses by the CPU and the
EDMA. Cache coherence is covered in more detail in section 8.1.
Compulsory miss Sometimes referred to as a first-reference miss. A compulsory miss is a cache miss
that must occur because the data has had no prior opportunity to be allocated in the
cache. Typically, compulsory misses for particular pieces of data occur on the first
access of that data. However, some cases can be considered compulsory even if
they are not the first reference to the data. Such cases include repeated write misses
on the same location in a cache that does not write allocate, and cache misses to
noncacheable locations. Compare with capacity miss and conflict miss.
Conflict miss A cache miss that occurs due to the limited associativity of a cache, rather than due
to capacity constraints. A fully-associative cache is able to allocate a newly cached
line of data anywhere in the cache. Most caches have much more limited
associativity (see set-associative cache), and so are restricted in where they may
place data. This results in additional cache misses that a more flexible cache would
not experience.
Direct-mapped cache A direct-mapped cache maps each address in the lower-level memory to a single
location in the cache. Multiple locations may map to the same location in the cache.
This is in contrast to a multi-way set-associative cache, which selects a place for the
data from a set of locations in the cache. A direct-mapped cache can be considered
a single-way set-associative cache.
Dirty In a writeback cache, writes that reach a given level in the memory hierarchy may
update that level, but not the levels below it. Thus, when a cache line is valid and
contains updates that have not been sent to the next lower level, that line is said to
be dirty. The opposite state for a valid cache line is clean.

SPRU610C TMS320C64x Two-Level Internal Memory 13

Cache Terms and Definitions

Table 2. Terms and Definitions (Continued)

Term Definition
DMA Direct Memory Access. Typically, a DMA operation copies a block of memory from
one range of addresses to another, or transfers data between a peripheral and
memory. On the C64x DSP, DMA transfers are performed by the enhanced DMA
(EDMA) engine. These DMA transfers occur in parallel to program execution. From a
cache coherence standpoint, EDMA accesses can be considered accesses by a
parallel processor.
Eviction The process of removing a line from the cache to make room for newly cached data.
Eviction can also occur under user control by requesting a writeback-invalidate for an
address or range of addresses from the cache. The evicted line is referred to as the
victim. When a victim line is dirty (that is, it contains updated data), the data must be
written out to the next level memory to maintain coherency.
Execute packet A block of instructions that begin execution in parallel in a single cycle. An execute
packet may contain between 1 and 8 instructions.

Fetch packet A block of 8 instructions that are fetched in a single cycle. One fetch packet may
contain multiple execute packets, and thus may be consumed over multiple cycles.

First-reference miss A cache miss that occurs on the first reference to a piece of data. First-reference
misses are a form of compulsory miss.

Fully-associative A cache that allows any memory address to be stored at any location within the
cache cache. Such caches are very flexible, but usually not practical to build in hardware.
They contrast sharply with direct-mapped caches and set-associative caches, both of
which have much more restrictive allocation policies. Conceptually, fully-associative
caches are useful for distinguishing between conflict misses and capacity misses
when analyzing the performance of a direct-mapped or set-associative cache. In
terms of set-associative caches, a fully-associative cache is equivalent to a
set-associative cache that has as many ways as it does line frames, and that has
only one set.

Higher-level memory In a hierarchical memory system, higher-level memories are memories that are
closer to the CPU. The highest level in the memory hierarchy is usually the Level 1
caches. The memories at this level exist directly next to the CPU. Higher-level
memories typically act as caches for data from lower-level memory.

Hit A cache hit occurs when the data for a requested memory location is present in the
cache. The opposite of a hit is a miss. A cache hit minimizes stalling, since the data
can be fetched from the cache much faster than from the source memory. The
determination of hit versus miss is made on each level of the memory hierarchy
separately—a miss in one level may hit in a lower level.

14 TMS320C64x Two-Level Internal Memory SPRU610C

 Cache Terms and Definitions

Table 2. Terms and Definitions (Continued)

Term Definition
Invalidate The process of marking valid cache lines as invalid in a particular cache. Alone, this
action discards the contents of the affected cache lines, and does not write back any
updated data. When combined with a writeback, this effectively updates the next
lower level of memory that holds the data, while completely removing the cached
data from the given level of memory. Invalidates combined with writebacks are
referred to as writeback-invalidates, and are commonly used for retaining coherence
between caches.
Least Recently Used For set-associative and fully-associative caches, least-recently used allocation refers
(LRU) allocation to the method used to choose among line frames in a set when allocating space in
the cache. When all of the line frames in the set that the address maps to contain
valid data, the line frame in the set that was read or written the least recently (furthest
back in time) is selected to hold the newly cached data. The selected line frame is
then evicted to make room for the new data.
Line A cache line is the smallest block of data that the cache operates on. The cache line
is typically much larger than the size of data accesses from the CPU or the next
higher level of memory. For instance, although the CPU may request single bytes
from memory, on a read miss the cache reads an entire line’s worth of data to satisfy
the request.
Line frame A location in a cache that holds cached data (one line), an associated tag address,
and status information for the line. The status information can include whether the
line is valid, dirty, and the current state of that line’s LRU.
Line size The size of a single cache line, in bytes.
Load through When a CPU request misses both the first-level and second-level caches, the data is
fetched from the external memory and stored to both the first-level and second-level
cache simultaneously. A cache that stores data and sends that data to the
upper-level cache at the same time is a load-through cache. Using a load-through
cache reduces the stall time compared to a cache that first stores the data in a lower
level and then sends it to the higher-level cache as a second step.
Long-distance access Accesses made by the CPU to a noncacheable memory. Long-distance accesses
are used when accessing external memory that is not marked as cacheable.

Lower-level memory In a hierarchical memory system, lower-level memories are memories that are further
from the CPU. In a C64x system, the lowest level in the hierarchy includes the
system memory below L2 and any memory-mapped peripherals.
LRU Least Recently Used. See least recently used allocation for a description of the LRU
replacement policy. When used alone, LRU usually refers to the status information
that the cache maintains for identifying the least-recently used line in a set. For
example, consider the phrase “accessing a cache line updates the LRU for that line.”

SPRU610C TMS320C64x Two-Level Internal Memory 15

Cache Terms and Definitions

Table 2. Terms and Definitions (Continued)

Term Definition
Memory ordering Defines what order the effects of memory operations are made visible in memory.
(This is sometimes referred to as consistency.) Strong memory ordering at a given
level in the memory hierarchy indicates it is not possible to observe the effects of
memory accesses in that level of memory in an order different than program order.
Relaxed memory ordering allows the memory hierarchy to make the effects of
memory operations visible in a different order. Note that strong ordering does not
require that the memory system execute memory operations in program order, only
that it makes their effects visible to other requestors in an order consistent with
program order. Section 8.3 covers the memory ordering assurances that the C64x
memory hierarchy provides.
Miss A cache miss occurs when the data for a requested memory location is not in the
cache. A miss may stall the requestor while the line frame is allocated and data is
fetched from the next lower level of memory. In some cases, such as a CPU write
miss from L1D, it is not strictly necessary to stall the CPU. Cache misses are often
divided into three categories: compulsory misses, conflict misses, and capacity
misses.
Miss pipelining The process of servicing a single cache miss is pipelined over several cycles. By
pipelining the miss, it is possible to overlap the processing of several misses, should
many occur back-to-back. The net result is that much of the overhead for the
subsequent misses is hidden, and the incremental stall penalty for the additional
misses is much smaller than that for a single miss taken in isolation.
Read allocate A read-allocate cache only allocates space in the cache on a read miss. A write miss
does not cause an allocation to occur unless the cache is also a write-allocate cache.
For caches that do not write allocate, the write data would be passed on to the next
lower-level cache.

Set A collection of line frames in a cache that a single address can potentially reside. A
direct-mapped cache contains one line frame per set, and an N-way set-associative
cache contains N line frames per set. A fully-associative cache has only one set that
contains all of the line frames in the cache.

Set-associative A set-associative cache contains multiple line frames that each lower-level memory
cache location can be held in. When allocating room for a new line of data, the selection is
made based on the allocation policy for the cache. The C64x devices employ a least
recently used allocation policy for its set-associative caches.

Snoop A method by which a lower-level memory queries a higher-level memory to

determine if the higher-level memory contains data for a given address. The primary
purpose of snoops is to retain coherency, by allowing a lower-level memory to
request updates from a higher-level memory. A snoop operation may trigger a
writeback, or more commonly, a writeback-invalidate. Snoops that trigger
writeback-invalidates are sometimes called snoop-invalidates.

16 TMS320C64x Two-Level Internal Memory SPRU610C

 Cache Terms and Definitions

Table 2. Terms and Definitions (Continued)

Term Definition
Tag A storage element containing the most-significant bits of the address stored in a
particular line. Tag addresses are stored in special tag memories that are not directly
visible to the CPU. The cache queries the tag memories on each access to
determine if the access is a hit or a miss.

Thrash An algorithm is said to thrash the cache when its access pattern causes the
performance of the cache to suffer dramatically. Thrashing can occur for multiple
reasons. One possible situation is that the algorithm is accessing too much data or
program code in a short time frame with little or no reuse. That is, its working set is
too large, and thus the algorithm is causing a significant number of capacity misses.
Another situation is that the algorithm is repeatedly accessing a small group of
different addresses that all map to the same set in the cache, thus causing an
artificially high number of conflict misses.
Touch A memory operation on a given address is said to touch that address. Touch can also
refer to reading array elements or other ranges of memory addresses for the sole
purpose of allocating them in a particular level of the cache. A CPU-centric loop used
for touching a range of memory in order to allocate it into the cache is often referred
to as a touch loop. Touching an array is a form of software-controlled prefetch for data.
Valid When a cache line holds data that has been fetched from the next level memory, that
line frame is valid. The invalid state occurs when the line frame holds no data, either
because nothing has been cached yet, or because previously cached data has been
invalidated for whatever reason (coherence protocol, program request, etc.). The
valid state makes no implications as to whether the data has been modified since it
was fetched from the lower-level memory; rather, this is indicated by the dirty or
clean state of the line.

Victim When space is allocated in a set for a new line, and all of the line frames in the set
that the address maps to contain valid data, the cache controller must select one of
the valid lines to evict in order to make room for the new data. Typically, the
least-recently used (LRU) line is selected. The line that is evicted is known as the
victim line. If the victim line is dirty, its contents are written to the next lower level of
memory using a victim writeback.
Victim Buffer A special buffer that holds victims until they are written back. Victim lines are moved
to the victim buffer to make room in the cache for incoming data.

Victim Writeback When a dirty line is evicted (that is, a line with updated data is evicted), the updated
data is written to the lower levels of memory. This process is referred to as a victim
writeback.
Way In a set-associative cache, each set in the cache contains multiple line frames. The
number of line frames in each set is referred to as the number of ways in the cache.
The collection of corresponding line frames across all sets in the cache is called a
way in the cache. For instance, a 4-way set-associative cache has 4 ways, and each
set in the cache has 4 line frames associated with it, one associated with each of the
4 ways. As a result, any given cacheable address in the memory map has 4 possible
locations it can map to in a 4-way set-associative cache.

SPRU610C TMS320C64x Two-Level Internal Memory 17

Cache Terms and Definitions

Table 2. Terms and Definitions (Continued)

Term Definition
Working set The working set for a program or algorithm is the total set of data and program code
that is referenced within a particular period of time. It is often useful to consider the
working set on an algorithm-by-algorithm basis when analyzing upper levels of
memory, and on a whole-program basis when analyzing lower levels of memory.

Write allocate A write-allocate cache allocates space in the cache when a write miss occurs. Space
is allocated according to the cache’s allocation policy (LRU, for example), and the
data for the line is read into the cache from the next lower level of memory. Once the
data is present in the cache, the write is processed. For a writeback cache, only the
current level of memory is updated—the write data is not immediately passed to the
next level of memory.
Writeback The process of writing updated data from a valid but dirty cache line to a lower-level
memory. After the writeback occurs, the cache line is considered clean. Unless
paired with an invalidate (as in writeback-invalidate), the line remains valid after a
writeback.

Writeback cache A writeback cache will only modify its own data on a write hit. It will not immediately
send the update to the next lower-level of memory. The data will be written back at
some future point, such as when the cache line is evicted, or when the lower-level
memory snoops the address from the higher-level memory. It is also possible to
directly initiate a writeback for a range of addresses using cache control registers. A
write hit to a writeback cache causes the corresponding line to be marked as
dirty—that is, the line contains updates that have yet to be sent to the lower levels of
memory.
Writeback-invalidate A writeback operation followed by an invalidation. See writeback and invalidate. On
the C64x devices, a writeback-invalidate on a group of cache lines only writes out
data for dirty cache lines, but invalidates the contents of all of the affected cache lines.
Write merging Write merging combines multiple independent writes into a single, larger write. This
improves the performance of the memory system by reducing the number of
individual memory accesses it needs to process. For instance, on the C64x device,
the L1D write buffer can merge multiple writes under some circumstances if they are
to the same double-word address. In this example, the result is a larger effective
write-buffer capacity and a lower bandwidth impact on L2.
Write-through cache A write-through cache passes all writes to the lower-level memory. It never contains
updated data that it has not passed on to the lower-level memory. As a result, cache
lines can never be dirty in a write-through cache. The C64x devices do not utilize
write-through caches.

18 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Data Cache (L1D)

3 Level 1 Data Cache (L1D)

The level 1 data cache (L1D) services data accesses from the CPU. The
following sections describe the parameters and operation of the L1D. The
operation of L1D is controlled by various registers, as described in section 7,
Memory System Controls.

3.1 L1D Parameters

The L1D is a 16K-byte cache. It is a two-way set associative cache with a
64-byte line size and 128 sets. It also features a 64-bit by 4-entry write buffer
between L1D and the L2 memory.
Physical addresses map onto the cache in a straightforward manner. The
physical address divides into three fields as shown in Figure 3. Bits 5−0 of the
address specify an offset within the line. Bits 12−6 of the address select one
of the 128 sets within the cache. Bits 31−13 of the address serve as the tag
for the line.

Figure 3. L1D Address Allocation

31 13 12 6 5 0
Tag Set Index Offset

Because L1D is a two-way cache, each set contains two cache lines, one for
each way. On each access, the L1D compares the tag portion of the address
for the access to the tag information for both lines in the appropriate set. If the
tag matches one of the lines and that line is marked valid, the access is a hit.
If these conditions are not met, the access is a miss. Miss penalties are
discussed in detail under section 3.2.
The L1D is a read-allocate-only cache. This means that new lines are allocated
in L1D for read misses, but not for write misses. For this reason, a 4-entry write
buffer exists between the L1D and L2 caches that captures data from write
misses. The write buffer is enhanced in comparison to the write buffer on the
C621x/C671x devices. The write buffer is described in section 3.2.3.
The L1D implements a least-recently used (LRU) line allocation policy. This
means that on an L1D read miss, the L1D evicts the least-recently read or
written line within a set in order to make room for the incoming data. Note that
invalid lines are always considered least-recently used.
If the selected line is dirty, that is, its contents are updated, then the victim line’s
data is prepared for writeback to L2 as a victim writeback. The actual victim
writeback occurs after the new data is fetched, and then only if the newly
fetched data is considered cacheable. If the newly fetched data is
noncacheable, the victim writeback is cancelled and the victim line remains in
the L1D cache.

SPRU610C TMS320C64x Two-Level Internal Memory 19

Level 1 Data Cache (L1D)

3.2 L1D Performance

3.2.1 L1D Memory Banking

The C64x DSP has a least-significant bit (LSB) based memory banking
structure that is similar to the structure employed by the C620x/C670x
families. The L1D on C64x devices divides memory into eight 32-bit-wide
banks. These banks are single-ported, allowing only one access per cycle.
This is in contrast to the C621x/C671x devices, which use a single bank of
dual-ported memory rather than multiple banks of single-ported memory. In
Figure 4, bits 4−2 of the address select the bank and bits 1−0 select the byte
within the bank.

Figure 4. Address to Bank Number Mapping

31 5 4 2 1 0
Bank
Upper Address Bits Offset
Number

The shaded areas in Figure 5 show combinations of parallel accesses that

may result in bank-conflict stalls according to the LSBs of addresses for the
two accesses. Two simultaneous accesses to the same bank incur a one-cycle
stall penalty, except under the following special cases:

- The memory accesses are both writes to nonoverlapping bytes within the
same word. That is, bits 31−2 of the address are the same.

- The memory accesses are both reads that access all or part of the same
word. That is, bits 31−2 of the address are the same. In this case, the two
accesses may overlap.

- One or both of the memory accesses is a write that misses L1D and is
serviced by the write buffer instead. (See section 3.2.3 for information on
the write buffer.)

- The memory accesses form a single nonaligned access. Nonaligned

accesses do not cause bank-conflict stalls, even though the memory
system may subdivide them into multiple accesses.

Notice that a read access and a write access in parallel to the same bank will
always cause a stall. Two reads or two writes to the same bank may not stall
as long as the above conditions are met.

20 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Data Cache (L1D)

Figure 5. Potentially Conflicting Memory Accesses

DA1 Byte Halfword Word DW
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 1 1 1 1 0 0 1 1
0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0 1 1 0 0 1 1 0 1 0 1
DA2

Bits
0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 1 0 1 0 1 0 1 0 0 0 0
4−0 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0
0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
00000
00001
00010
00011
00100
00101
00110
00111
01000
01001
01010
01011
01100
01101
Byte

01110
01111
10000
10001
10010
10011
10100
10101
10110
10111
11000
11001
11010
11011
11100
11101
11110
11111
00000
00010
00100
00110
01000
01010
01100
Halfword

01110
10000
10010
10100
10110
11000
11010
11100
11110
00000
00100
01000
Word

01100
10000
10100
11000
11100
00000
01000
DW

10000
11000

Note: Conflicts shown in shaded areas.

SPRU610C TMS320C64x Two-Level Internal Memory 21

Level 1 Data Cache (L1D)

While similar to C620x/C670x devices memory banking, the C64x device

banking scheme differs in two ways:

- Banks are 16-bits wide on the C620x/C670x devices, and 32-bits wide on
the C64x device.

- The C620x/C670x devices always stall on parallel accesses to the same

bank, regardless of whether those accesses are to the same word. The
C64x device allows many special cases to proceed without stalls, as
previously described.

3.2.2 L1D Miss Penalty

The L1D can service up to two data accesses from the CPU every cycle.
Accesses that hit L1D complete without stalls, unless a bank conflict occurs
as described in section 3.2.1.

Reads that miss L1D stall the CPU while the requested data is fetched. The
L1D is a read-allocate cache, and so it will allocate a new line for the requested
data, as described in section 3.1. An isolated L1D read miss that hits L2 SRAM
stalls the CPU for 6 cycles, and an isolated L1D read miss that hits L2 cache
stalls the CPU for 8 cycles. This assumes there is no other memory traffic in
L2 that delays the processing of requests from L1D. Section 5.4 discusses
interactions between the various requestors that access L2.

An L1D read miss that also misses L2 stalls the CPU while the L2 retrieves the
data from external memory. Once the data is retrieved, it is stored in L2 and
transferred to the L1D. The external miss penalty varies depending on the type
and width of external memory used to hold external data, as well as other
aspects of system loading. Section 5.2 describes how L2 handles cache
misses on behalf of L1D.

If there are two read misses to the same line in the same cycle, only one miss
penalty is incurred. Similarly, if there are two accesses in succession to the
same line and the first one is a miss, the second access will not incur any
additional miss penalty.

The process of allocating a line in L1D can result in a victim writeback. Victim
writebacks move updated data out of L1D to the lower levels of memory. When
updated data is evicted from L1D, the cache moves the data to the victim
buffer. Once the data is moved to the victim buffer, the L1D resumes
processing of the current read miss. Further processing of the victim writeback
occurs in the background. Subsequent read and write misses, however, must
wait for the victim writeback to be processed. As a result, victim writebacks can
noticeably lengthen the time for servicing cache misses.

22 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Data Cache (L1D)

The L1D pipelines read misses. Consecutive read misses to different lines
may be overlapped, reducing the overall stall penalty. The incremental stall
penalty can be as small as 2 cycles per miss. Section 3.2.4 discusses miss
pipelining.

Write misses do not stall the CPU directly. Rather, write misses are queued in
the write buffer that exists between L1D and L2. Although the CPU does not
always stall for write misses, the write buffer can stall the CPU under various
circumstances. Section 3.2.3 describes the effects of the write buffer.

3.2.3 L1D Write Buffer

The L1D does not write allocate. Rather, write misses are passed directly to
L2 without allocating a line in L1D. A write buffer exists between the L1D cache
and the L2 memory to capture these write misses. The write buffer provides
a 64-bit path for writes from L1D to L2 with room for four outstanding write
requests.

Writes that miss L1D do not stall the CPU unless the write buffer is full. If the
write buffer is full, a write miss will stall the CPU until there is room in the buffer
for the write. The write buffer can also indirectly stall the CPU by extending the
time for a read miss. Reads that miss L1D will not be processed as long as the
write buffer is not empty. Once the write buffer has emptied, the read miss will
be processed. This is necessary as a read miss may overlap an address for
which a write is pending in the write buffer.

The L2 can process a new request from the write buffer every cycle, provided
that the requested L2 bank is not busy. Section 5.3 describes the L2 banking
structure and its impact on performance.

The C64x write buffer allows merging of write requests. It merges two write
misses into a single transaction providing all of the following rules are obeyed:

- The double-word addresses (that is, the upper 29 bits) for the two
accesses are the same.

- The two writes are to locations in L2 SRAM (not locations that may be held
in L2 cache).

- The first write has just been placed in the write buffer queue.

- The second write is presently being placed in the buffer queue.

- The first write has not yet been presented to the L2 controller.

SPRU610C TMS320C64x Two-Level Internal Memory 23

Level 1 Data Cache (L1D)

The previous conditions occur in a number of situations, such as when a

program makes a large series of sequential writes or when it makes a burst
of small writes to a structure in memory. Write merging increases the effective
capacity of the write buffer in these cases by reducing the number of
independent stores that are present in the write buffer. This reduces the stall
penalty for programs with a large number of write misses.

As a secondary benefit, write merging reduces the number of memory

operations executed in L2. This improves the overall performance of the L2
memory by reducing the total number of individual write operations L2 must
process. Adjacent accesses are combined into a single access to an L2 bank,
rather than multiple accesses to that bank. This allows other requestors to
access that bank more quickly, and it allows the CPU to move on to the next
bank immediately in the next cycle.

3.2.4 L1D Miss Pipelining

The L1D cache pipelines read misses. A single L1D read miss takes 6 cycles
when serviced from L2 SRAM, and 8 cycles when serviced from L2 cache.
Miss pipelining can hide much of this overhead by overlapping the processing
of several cache misses.

For L1D miss pipelining to be effective, there must be multiple outstanding L1D
read misses. Load instructions on the C64x DSP have a 5-cycle-deep pipeline,
and the C64x DSP may issue up to two accesses per cycle. In this pipeline,
the L1D performs tag comparisons in one pipeline stage (E2), and services
cache hits and misses on the following stage (E3). Cache read misses result
in a CPU stall.

L1D processes single read misses only when there are no outstanding victim
writebacks and when the write buffer is empty. When two cache misses occur
in parallel, the L1D processes the misses in program order. (The program
order is described in section 8.3.1.) In the case of two write misses, the misses
are inserted in the write buffer and the CPU does not stall unless the write
buffer is full. (Section 3.2.3 describes the write buffer.) In the case of two read
misses or a read and a write miss, the misses are overlapped as long as they
are to different sets, that is, their addresses differ in bits 13−6.

24 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Data Cache (L1D)

Cache misses are processed in the E3 pipeline stage. Once L1D has issued
commands to L2 for all of the cache misses in E3, the L1D may decide to
advance its state internally by one pipeline stage to consider cache misses due
to accesses that were in the E2 pipeline stage. This allows L1D to aggressively
overlap requests for cache misses that occur in parallel and cache misses that
occur on consecutive cycles. L1D considers the accesses in E2 only if the write
buffer and victim writeback buffer are empty. Although the L1D internal state
advances, the CPU stall is not released until the data returns for accesses that
were in the E3 stage.

Once the CPU stall is released, memory accesses that were in the E2 stage
advance to the E3 pipeline stage. This may bring one or two new accesses into
the E2 pipeline stage. It also potentially brings one or two unprocessed cache
misses from E2 into E3. The L1D first issues commands for any cache misses
that are now in E3 but that have not yet been processed. Once the accesses
in E3 are processed, the L1D may consider accesses in E2 as previously
described. In any case, the L1D stalls the CPU when there are accesses in E3
that have not yet completed.

The net result is that the L1D can generate a continuous stream of requests
to L2. Code that issues pairs of memory reads to different cache lines every
cycle will maximize this effect. As noted above, this pipelining can result in
improved performance, especially in the presence of sustained read misses.

The incremental miss penalty can be as small as 2 cycles per miss when the
L1D is able to overlap the processing for a new cache miss with that of prior
misses. Therefore, the average miss penalty for a sustained sequence of
back-to-back misses approaches 2 cycles per miss in the ideal case. Table 3
and Table 4 illustrate the performance for various numbers of consecutive L1D
read misses that hit in L2 cache and L2 SRAM, assuming all misses are able
to overlap. These further assume that there is no other memory traffic in L2 that
may lengthen the time required for an L1D cache miss, and that all misses are
within the same half of the affected L1D cache lines.

SPRU610C TMS320C64x Two-Level Internal Memory 25

Level 1 Data Cache (L1D)

Table 3. Cycles Per Miss for Different Numbers of L1D Misses That Hit L2 Cache

Number of Misses Total Stall Cycles Mean Cycles Per Miss

1 8 8

2 10 5

3 12 4

4 14 3.5

> 4, even 6 + (2 * M) 2 + (6 / M)

Note: M = Number of total misses.

Table 4. Cycles Per Miss for Different Numbers of L1D Misses that Hit L2 SRAM

Number of Misses Total Stall Cycles Mean Cycles Per Miss

1 6 6

2 8 4

3 10 3.33

4 12 3

> 4, even 4 + (2 * M) 2 + (4 / M)

Note: M = Number of total misses.

26 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Program Cache (L1P)

4 Level 1 Program Cache (L1P)

The level 1 program cache (L1P) services program fetches from the CPU. The
following sections describe the parameters and operation of the L1P. The
operation of L1P is controlled by various registers, as described in section 7,
Memory System Controls.

4.1 L1P Parameters

The L1P is a 16K-byte cache. It is a direct-mapped cache with a 32-byte line
size and 512 sets.

Physical addresses map onto the cache in a fixed manner. The physical
address divides into three fields as shown in Figure 6. Bits 4−0 of the address
specify an instruction within a set. Bits 13−5 of the address select one of the
512 sets within the cache. Bits 31−14 of the address serve as the tag for the
line.

Figure 6. L1P Address Allocation

31 14 13 5 4 0
Tag Set Index Offset

Because L1P is direct-mapped cache, each address maps to a fixed location

in the cache. That is, each set contains exactly one line frame. On a cache
miss, the cache allocates the corresponding line for the incoming data.
Because L1P does not support writes from the CPU, the previous contents of
the line are discarded.

4.2 L1P Performance

4.2.1 L1P Miss Penalty
A program fetch which hits L1P completes in a single cycle without stalling the
CPU. An L1P miss that hits in L2 may stall the CPU for up to 8 cycles,
depending on the parallelism of the execute packets in the vicinity of the miss.
Section 4.2.2 describes this in more detail.

An L1P miss that misses in L2 cache stalls the CPU until the L2 retrieves the
data from external memory and transfers the data to the L1P, which then
returns the data to the CPU. This delay depends upon the type of external
memory used to hold the program, as well as other aspects of system loading.

The C64x DSP allows an execute packet to span two fetch packets. This
spanning does not change the penalty for a single miss. However, if both fetch
packets are not present in L1P, two cache misses occur.

SPRU610C TMS320C64x Two-Level Internal Memory 27

Level 1 Program Cache (L1P)

4.2.2 L1P Miss Pipelining

The L1P cache pipelines cache misses. A single L1P cache miss requires
8 cycles to retrieve data from L2. Miss pipelining can hide much of this
overhead by overlapping the processing for several cache misses.
Additionally, some amount of the cache miss overhead can be overlapped with
dispatch stalls that occur in the fetch pipeline.

For L1P miss pipelining to be effective, there must be multiple outstanding

cache misses. The C64x DSP fetch pipeline accomplishes this by attempting
to fetch one new fetch packet every cycle, so long as there is room in the fetch
pipeline. To understand how this works, it is necessary to understand the
nature of the fetch pipeline itself.

The fetch and decode pipeline is divided into 6 stages leading up to but not
including the first execution stage, E1. The stages are:

- PG − Program Generate
- PS − Program Send
- PW − Program Wait
- PR − Program Read
- DP − Dispatch
- DC − Decode

C6000 DSP instructions are grouped into two groupings: fetch packets and
execute packets. The CPU fetches instructions from memory in fixed bundles
of 8 instructions, known as fetch packets. The instructions are decoded and
separated into bundles of parallel-issue instructions known as execute
packets. A single execute packet may contain between 1 and 8 instructions.
Thus, a single fetch packet may contain multiple execute packets. On the
C64x DSP, an execute packet may also span two fetch packets. The Program
Read (PR) stage of the pipeline is responsible for identifying a sequence of
execute packets within a sequence of fetch packets. The Dispatch (DP) stage
is responsible for extracting and dispatching them to functional units.

As a result of the disparity between fetch packets and execute packets, the
entire fetch pipeline need not advance every cycle. Rather, the PR pipeline
stage only allows the Program Wait (PW) stage to advance its contents into
the PR stage when the DP stage has consumed the complete fetch packet
held in PR. The stages before PR advance as needed to fill in gaps. Thus,
when there are no cache misses, the early stages of the fetch pipeline are
stalled while the DP stage pulls the individual execute packets from the current
fetch packet. These stalls are referred to as dispatch stalls.

28 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 1 Program Cache (L1P)

The C64x DSP takes advantage of these dispatch stalls by allowing the earlier
stages of the pipeline to advance toward DP while cache misses for those
stages are still pending. Cache misses may be pending for the PR, PW, and
PS pipeline stages. Because the DP stage stalls the PR stage with a dispatch
stall while it consumes the fetch packets in the PR stage of the pipeline, it is
not necessary to expose these cache stalls to the CPU. When a fetch packet
is consumed completely, however, the contents of the PW stage must advance
into the PR stage. At this point, the CPU is stalled if DP requests an execute
packet from PR for which there is still an outstanding cache miss.

When a branch is taken, the fetch packet containing the branch target
advances through the fetch pipeline every cycle until the branch target
reaches the E1 pipeline stage. Branch targets override the dispatch stall
described above. As a result, they do not gain as much benefit from miss
pipelining as other instructions. The fetch packets that immediately follow a
branch target do benefit, however. Although the code in the fetch packets that
follows the branch target may not execute immediately, the branch triggers
several consecutive fetches for this code, and thus pipelines any misses for
that code. In addition, no stalls are registered for fetch packets that were
requested prior to the branch being taken, but that never made it to the DP
pipeline stage.

The miss penalty for a single L1P miss is 8 cycles. The second miss in a pair
of back-to-back misses will see an incremental stall penalty of up to 2 cycles.
Sustained back-to-back misses in straight-line (nonbranching) code incurs an
average miss penalty based on the average parallelism of the code. The
average miss penalty for a long sequence of sustained misses in straight-line
code is summarized in Table 5.

Table 5. Average Miss Penalties for Large Numbers of Sequential Execute Packets
Instructions Per Average Stalls Per
Execute Packet Execute Packet
1 0.125

2 0.125

3 0.688

4 1.500

5 1.813

6 2.375

7 2.938

8 4.000

SPRU610C TMS320C64x Two-Level Internal Memory 29

Level 2 Unified Memory (L2)

5 Level 2 Unified Memory (L2)

The level 2 unified memory (L2) can operate as SRAM, cache, or both. It
services cache misses from both L1P and L1D, as well as DMA accesses
using the EDMA controller. The following sections describe the parameters
and operation of the L2. The operation of L2 is controlled by various registers,
as described in section 7, Memory System Controls.

5.1 L2 Cache and L2 SRAM

The L2 can operate as SRAM, as cache, or as both, depending on its mode.
L2 SRAM refers to the portion of L2 mapped as SRAM and L2 cache refers
to the portion of L2 that acts as cache. For a given device, the total capacity
of L2 SRAM and L2 cache together is fixed, regardless of the L2 mode.

The total size of the L2 depends on the specific C64x device. The C6414,
C6415, and C6416 devices provide 1024K bytes of L2 memory. The C6411
and C6412 devices provide 256K bytes of L2 memory. For other C64x devices,
consult the data sheet to determine the L2 size for the device.

After reset, the entire L2 memory is mapped as SRAM. The L2 SRAM is

contiguous and always starts at address 0000 0000h in the memory map. The
C6414, C6415, and C6416 devices map L2 SRAM over the address range
0000 0000h to 000F FFFFh. The C6411 and C6412 devices map L2 SRAM
over the address range 0000 0000h to 0003 FFFFh.

The L2 cache is a 4-way set associative cache whose capacity varies between
32K bytes and 256K bytes depending on its mode. The L2 cache is enabled
by the L2MODE field in the cache configuration register (CCFG). Enabling L2
cache reduces the amount of available L2 SRAM. Section 7.1.3 discusses how
the L2 memory map varies according to cache mode and the specific device
being used.

External physical memory addresses map onto the L2 cache differently

depending on the cache mode. The physical address divides into three fields
as shown in Figure 7, Figure 8, Figure 9, and Figure 10. Bits 6−0 of the
address specify an offset within a line. The next 6 to 9 bits (depending on
mode) of the address specify the set within the cache. The remaining bits of
the address serve as the tag for the line.

30 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 2 Unified Memory (L2)

Figure 7. L2 Address Allocation, 256K Cache (L2MODE = 111b)

31 16 15 7 6 0
Tag Set Index Offset

Figure 8. L2 Address Allocation, 128K Cache (L2MODE = 011b)

31 15 14 7 6 0
Tag Set Index Offset

Figure 9. L2 Address Allocation, 64K Cache (L2MODE = 010b)

31 14 13 7 6 0
Tag Set Index Offset

Figure 10. L2 Address Allocation, 32K Cache (L2MODE = 001b)

31 13 12 7 6 0
Tag Set Index Offset

SPRU610C TMS320C64x Two-Level Internal Memory 31

Level 2 Unified Memory (L2)

5.2 L2 Operation

When L2 cache is enabled, it services requests from L1P and L1D to external
addresses. The operation of the L2 cache is similar to the operation of both
L1P and L1D. When a request is made, the L2 first determines if the address
requested is present in the cache. If the address is present, the access is
considered a cache hit and the L2 services the request directly within the
cache. If the address is not present, the access results in a cache miss. On a
miss, the L2 processes the request according to the cacheability of the
affected address.

On a cache hit, the L2 updates the LRU status for the corresponding set in L2
cache. If the access is a read, the L2 returns the requested data. If the access
is a write, the L2 updates the contents of the cache line and marks the line as
dirty. L2 is a writeback cache, and so write hits in L2 are not immediately
forwarded to external memory. The external memory will be updated when this
line is later evicted, or is written back using the block-writeback control
registers described in section 7.3.2.

The L2 allocates a new line within the L2 cache on a cache miss to a cacheable
external memory location. Note that unlike L1D, the L2 allocates a line for both
read and write misses. The L2 cache implements a least-recently used policy
to select a line within the set to allocate on a cache miss. If this line being
replaced is valid, it is evicted as described below. Once space allocated for the
new data, an entire L2 line’s worth of data is fetched via the EDMA into the
allocated line.

Evicting a line from L2 requires several steps, regardless of whether the victim
is clean or dirty. For each line in L2, L2 tracks whether the given line is also
cached in L1D. If it detects that the victim line is present in L1D, it sends L1D
snoop-invalidate requests to remove the affected L1D lines. L1D responds by
invalidating the corresponding line. If the line in L1D was dirty, the updated
data is passed to L2 and merged with the L2 line that is being evicted. The
combined result is written to external memory. If the victim line was not dirty
in either L1D or L2, its contents are discarded. These actions ensure that the
most recent writes to the affected cache line are written to external memory,
but that clean lines are not needlessly written to external memory.

Note that L1P is not consulted when a line is evicted from L2. This allows
program code to remain in L1P despite having been evicted from L2. This
presumes that program code is never written to. In those rare situations where
this is not the case, programs may use the cache controls described in
section 7.3.2 to remove cached program code from L1P.

32 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 2 Unified Memory (L2)

If the cache miss was a read, the data is stored in L2 cache when it arrives.
It is also forwarded directly to the requestor, thereby reducing the overall stall
time. The portion of the L2 cache line requested by L1 is sent directly to that
L1, and is allocated within that cache. The entire L2 line is also stored within
the L2 cache as it arrives.

If the cache miss was a write, the incoming data is merged with the write data
from L1D, and the merged result is stored in L2. In the case of a write, the line
is not immediately allocated in L1D, as L1D does not write allocate.

A cache-miss to a non-cacheable location results in a long-distance access.

A long distance read causes the L2 to issue a read transfer for the requested
data to the EDMA controller. When the requested data is returned, the L2
forwards the data to the requestor. It does not allocate space for this data in
the L2 cache. A long distance read from L1D reads exactly the amount of data
requested by the CPU. A long distance read from L1P reads exactly 32 bytes.
The data read is not cached by L1D, L1P, or L2.

A long distance write causes the L2 to store a temporary copy of the written
data. It then issues a write transfer for the write miss to the EDMA controller.
Long distance writes can only originate from L1D using the L1D write buffer.
Because the written data is stored in a special holding buffer, it is not necessary
to stall the CPU while the long-distance write is being processed. Also, further
writes to the L2 SRAM address space or on-chip peripherals may be
processed while the long-distance access is being executed.

The L2 cache allows only one long-distance access to be in progress at one

time. A long distance write will potentially stall the CPU if a previous long
distance write is still in progress. Likewise, a long distance read will not be
processed until all previous long distance writes are complete.

The ordering constraints placed on long-distance accesses effectively make

accesses to non-cacheable regions of memory strongly ordered. Program
order for these accesses is retained, thus making these accesses useful for
interfacing to peripherals or synchronizing with other CPUs.

SPRU610C TMS320C64x Two-Level Internal Memory 33

Level 2 Unified Memory (L2)

5.3 L2 Bank Structure

The L2 memory is divided into 8 64-bit banks that operate at the CPU’s clock
rate, but pipeline accesses over two cycles. The L2 begins processing a new
request on each cycle, although each bank can process new requests no
faster than one every two cycles.

Because each bank requires two cycles to process a request, an access to a

bank on one cycle blocks all accesses to that bank on the following cycle.
Thus, bank conflicts can occur between accesses on adjacent cycles.
Repeated accesses to the same bank are processed at a rate of one every two
cycles, regardless of the requestor. This highlights the importance of merging
write misses from L1D.

The L2 must service requests from the following sources:

- L1P read miss that hits in L2.

- L1D read or write miss that hits in L2.
- EDMA reads and writes.
- Internal cache operations (victim writebacks, line fills, snoops).

The internal cache machinery is the fourth source of requests to the L2

memory. These requests represent data movement triggered by normal cache
operation. Most of these requests are triggered by cache misses, or by
user-initiated cache control operations.

Only one requestor may access the L2 in a given cycle. In the case of an L2
conflict, the L2 prioritizes requests in the above order. That is, L1P read hits
have highest priority, followed by L1D, and so on.

L2 contention due simultaneous access or bank conflicts can lengthen the

time required for servicing L1P and L1D cache misses. The priority scheme
above minimizes this effect by servicing CPU-initiated accesses before EDMA
accesses and the background activities performed by the cache.

34 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 2 Unified Memory (L2)

5.4 L2 Interfaces
The L2 services requests from the L1D cache, L1P cache, and the EDMA. The
L2 provides access to its own memory, the peripheral configuration bus
(CFGBUS), and the various cache control registers described in section 6,
Registers. The following sections describe the interaction between L2 and the
various requestors that it interfaces.

5.4.1 L1D/L1P-to-L2 Request Servicing

The L2 controller allows the L1P and L1D to access both L2 SRAM and
L2 cache. The L2 also acts as intermediary for long-distance accesses to
addresses in external memory, and for accesses to on-chip peripherals. For
each access, the address and the L2 mode determine the behavior.
Memory accesses to addresses that are mapped as L2 SRAM are serviced
directly by L2 and are treated as cacheable by L1P and L1D. Memory
accesses to addresses that are outside L2 SRAM but are on-chip are treated
as non-cacheable. These accesses include accesses to on-chip peripherals
and cache control registers.
L2 SRAM accesses by L1D do not trigger cache operations in either L1P or
L2 cache. Likewise, L1P accesses to L2 SRAM do not trigger cache operations
in either L1D or L2 cache. This is consistent with the coherence policy outlined
in section 8.1. This contrasts with EDMA accesses to L2 SRAM, described in
section 5.4.2.
L1P and L1D memory accesses to external addresses may be serviced by the
L2 cache, when the cache is enabled and caching is permitted on the given
range of addresses. The cacheability of a given external address range is
determined by the cache enable bit (CE) in the corresponding memory
attribute register (MAR). Section 7.2 describes MAR operation.
When the L2 is in all SRAM mode (L2MODE = 000b), it does not attempt to
cache memory accesses directly. However, when servicing read requests for
L1P and L1D, the L2 returns the cacheability status of the fetched data, thus
allowing L1P and L1D to directly retain copies of memory that is marked as
cacheable. Writes and writebacks from L1D for external addresses are sent
to external memory directly when in this mode.
Accesses to noncacheable external locations are handled as long-distance
accesses. Long distance accesses are described in section 5.2.

Note:
The L2 only supports 32-bit accesses to on-chip peripheral addresses,
including its own control registers. Byte, half-word, and double-word
accesses may not work as expected.

SPRU610C TMS320C64x Two-Level Internal Memory 35

Level 2 Unified Memory (L2)

5.4.2 EDMA-to-L2 Request Servicing

The L2 controller handles EDMA access to L2 SRAM addresses. EDMA read
and write requests to L2 SRAM result in specific actions to enforce the memory
coherence policy described in section 8.1.

The L2 controller tracks information on what portions of L2 SRAM are held in

L1D cache. When the EDMA reads from a location in L2 SRAM, the L2
controller snoops L1D for updated data, if necessary, before servicing the read
for the EDMA. If L1D contains updated data, the update is written to L2 SRAM
before allowing the EDMA read to proceed. The affected line in L1D is left valid
but clean. No action is made relative to L1P.

When the EDMA writes to a location in L2 SRAM, the L2 controller sends an

invalidate command to L1P and a snoop-invalidate command to L1D, as
necessary. For L1P, the invalidate command is always sent since L2 does not
track the state of L1P. If the address is present in L1P, the affected line is
invalidated. As L1P cannot contain updated data, L2 does not interact further
with L1P. For L1D, the snoop-invalidate is sent only if L2 detects that L1D holds
a copy of the address being written. If L1D contains updated data, the update
is merged with the EDMA write and stored to L2, with the incoming data taking
precedence over the data returned by L1D. The affected line in L1D is then
marked invalid.

The end result of this system of snoops and invalidates is that coherence is
retained between EDMA and CPU accesses to L2 SRAM. The example in
section 8.1 illustrates this protocol in action.

5.4.3 L2 Request Servicing Using EDMA

The L2 controller relies on the EDMA to service requests on its behalf. Cache
misses, victim writebacks, and long-distance accesses to external addresses
cause the L2 controller to issue DMA transfer requests to the EDMA controller.

L2 requests are queued along other with other EDMA requests, and are
serviced according to the policies set by the EDMA. The L2 requests may be
placed on any of the EDMA priority queues. The priority level for requests and
the number of outstanding requests permitted can be controlled as described
in section 7.4. The EDMA is described in TMS320C6000 DSP Enhanced
Direct Memory Access (EDMA) Peripheral Reference Guide (SPRU234).

36 TMS320C64x Two-Level Internal Memory SPRU610C

 Level 2 Unified Memory (L2)

5.4.4 EDMA Access to Cache Controls

The L2 controller manages access to the memory mapped cache control
registers. While the L2 controller permits CPU access to these registers, it
does not permit EDMA access to any of these registers. EDMA accesses to
the cache control registers are dropped. EDMA writes are ignored and EDMA
reads return undefined values.

As a result of these restrictions, cache control operations may only be initiated

from the CPU. External devices that wish to trigger a cache control operation
should do so using an interrupt handler or some other mechanism that triggers
CPU writes to the appropriate registers on behalf of the requestor.

5.4.5 HPI and PCI Access to Memory Subsystem

The host port interface (HPI) and peripheral component interconnect (PCI)
peripherals are not directly connected to the two-level memory subsystem.
HPI and PCI requests are serviced indirectly by the EDMA. Therefore, HPI and
PCI accesses follow the same coherence policies and are subject to the same
restrictions as EDMA accesses.

Sections 5.4.2 through 5.4.4 describe the interaction between the EDMA and
the two-level memory system. Section 8, Memory System Policies, describes
the coherence and ordering policies for the memory system.

SPRU610C TMS320C64x Two-Level Internal Memory 37

Registers

6 Registers
The two-level memory hierarchy is controlled by several memory-mapped
control registers listed in Table 6. It is also controlled by the data cache control
(DCC) and program cache control (PCC) fields in the CPU control and status
register (CSR), see section 7.1. See the device-specific datasheet for the
memory address of these registers.

Table 6. Internal Memory Control Registers

Acronym Register Name Section
CCFG Cache configuration register 6.1

EDMAWEIGHT L2 EDMA access control register 6.2

L2ALLOC0−3 L2 allocation registers 0−3 6.3

L2WBAR L2 writeback base address register 6.4

L2WWC L2 writeback word count register 6.5

L2WIBAR L2 writeback-invalidate base address register 6.6

L2WIWC L2 writeback-invalidate word count register 6.7

L2IBAR L2 invalidate base address register 6.8

L2IWC L2 invalidate word count register 6.9

L1PIBAR L1P invalidate base address register 6.10

L1PIWC L1P invalidate word count register 6.11

L1DWIBAR L1D writeback-invalidate base address register 6.12

L1DWIWC L1D writeback-invalidate word count register 6.13

L1DIBAR L1D invalidate base address register 6.14

L1DIWC L1D invalidate word count register 6.15

L2WB L2 writeback all register 6.16

L2WBINV L2 writeback-invalidate all register 6.17

MAR0−MAR95 Reserved −

MAR96−MAR111† L2 memory attribute register 96−111: controls EMIFB CE0−CE3 6.18

MAR112−MAR127 Reserved −

MAR128−MAR191 L2 memory attribute register 128−191: controls EMIFA CE0−CE3 6.18

MAR192−MAR255 Reserved −
† MAR96−MAR111 only available on the C6414/C6415/C6416 devices; on all other devices, these registers are reserved.

38 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.1 Cache Configuration Register (CCFG)

The cache configuration register (CCFG) is shown in Figure 11 and described
in Table 7.

Figure 11. Cache Configuration Register (CCFG)

31 29 28 16
P Reserved
R/W-0 R-0

15 10 9 8 7 3 2 0
Reserved IP ID Reserved L2MODE
R-0 W-0 W-0 R-0 R/W-0
Legend: R = Read only; W = Write only; R/W = Read/write; -n = value after reset

Table 7. Cache Configuration Register (CCFG) Field Descriptions

Bit field† symval† Value Description
31−29 P OF(value) 0−7h L2 requestor priority bits.

DEFAULT 0 L2 controller requests are placed on urgent priority level.

URGENT

HIGH 1h L2 controller requests are placed on high priority level.

MEDIUM 2h L2 controller requests are placed on medium priority level.

LOW 3h L2 controller requests are placed on low priority level.

− 4h−7h Reserved

28−10 Reserved − 0 Reserved. The reserved bit location is always read as 0. Bits in
this field should always be written with 0.

9 IP OF(value) Invalidate L1P bit.

DEFAULT 0 Normal L1P operation.

NORMAL

INVALIDATE 1 All L1P lines are invalidated.

† For CSL implementation, use the notation CACHE_CCFG_field_symval
‡ C64x devices with less than 256K of L2 memory may not support all L2 cache modes. See the device-specific datasheet to
determine the supported cache sizes.

SPRU610C TMS320C64x Two-Level Internal Memory 39

Registers

Table 7. Cache Configuration Register (CCFG) Field Descriptions (Continued)

Bit field† symval† Value Description
8 ID OF(value) Invalidate L1D bit.

DEFAULT 0 Normal L1D operation.

NORMAL

INVALIDATE 1 All L1D lines are invalidated.

7−3 Reserved − 0 Reserved. The reserved bit location is always read as 0. Bits in
this field should always be written with 0.

2−0 L2MODE‡ OF(value) 0−7h L2 operation mode bits.See the device-specific datasheet to
determine the available cache options for a device and how the
various L2 cache modes affect the memory map for different
sizes of L2.

DEFAULT 0 L2 cache disabled (all SRAM mode)

0KC

32KC 1h 4-way cache (32K L2 cache)

64KC 2h 4-way cache (64K L2 cache)

128KC 3h 4-way cache (128K L2 cache)

− 4h-6h Reserved

256KC 7h 4-way cache (256K L2 cache). Not supported on devices with

L2 cache sizes of less than 256-Kbytes memory, such as the
C6410 device.
† For CSL implementation, use the notation CACHE_CCFG_field_symval
‡ C64x devices with less than 256K of L2 memory may not support all L2 cache modes. See the device-specific datasheet to
determine the supported cache sizes.

40 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.2 L2 EDMA Access Control Register (EDMAWEIGHT)

The L2 EDMA access control register (EDMAWEIGHT) is shown in Figure 12
and described in Table 8.

Figure 12. L2 EDMA Access Control Register (EDMAWEIGHT)

31 2 1 0
Reserved EDMAWEIGHT
R-0 R/W-1
Legend: R = Read only; R/W = Read/Write; -n = value after reset

Table 8. L2 EDMA Access Control Register (EDMAWEIGHT) Field Descriptions

Bit Field Value Description

31−2 Reserved 0 Reserved. The reserved bit location is always read as 0. A value written
to this field has no effect.

1−0 EDMAWEIGHT 0−3h EDMA weight limits the amount of time L1D blocks EDMA access to L2.

0 L1D access always a higher priority than EDMA access to L2.

EDMA never receives priority.

1h EDMA receives priority after 16 L1D priority cycles.

2h EDMA receives priority after 4 L1D priority cycles.

3h EDMA receives priority after 1 L1D priority cycle.

SPRU610C TMS320C64x Two-Level Internal Memory 41

Registers

6.3 L2 Allocation Registers (L2ALLOC0−L2ALLOC03)

The L2 allocation registers (L2ALLOCn) are shown in Figure 13 and described
in Table 9.

Figure 13. L2 Allocation Registers (L2ALLOC0−L2ALLOC3)

(a) L2ALLOC0
31 3 2 0
Reserved Q0CNT
R-0 R/W-110
(b) L2ALLOC1
31 3 2 0
Reserved Q1CNT
R-0 R/W-010
(c) L2ALLOC2
31 3 2 0
Reserved Q2CNT
R-0 R/W-010
(d) L2ALLOC3
31 3 2 0
Reserved Q3CNT
R-0 R/W-010
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 9. L2 Allocation Registers (L2ALLOC0−L2ALLOC3) Field Descriptions

Bit field† symval† Value Description

31−3 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

2−0 QnCNT OF(value) 0−7h The the total number of outstanding L2 and QDMA requests
permitted on the corresponding EDMA priority level. Further
requests on that priority level are stalled until the number of
outstanding requests falls below the QnCNT setting.

DEFAULT 2h For L2ALLOC1, L2ALLOC2, and L2ALLOC3.

DEFAULT 6h For L2ALLOC0.

† For CSL implementation, use the notation CACHE_L2ALLOCn_field_symval

42 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.4 L2 Writeback Base Address Register (L2WBAR)

The L2 writeback base address register (L2WBAR) is shown in Figure 14 and
described in Table 10.

Figure 14. L2 Writeback Base Address Register (L2WBAR)

31 0

L2 Writeback Base Address (L2WBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 10. L2 Writeback Base Address Register (L2WBAR) Field Descriptions

Bit Field symval† Value Description
31−0 L2WBAR OF(value) 0−FFFF FFFFh L2 writeback base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2WBAR_L2WBAR_symval

6.5 L2 Writeback Word Count Register (L2WWC)

The L2 writeback word count register (L2WWC) is shown in Figure 15 and
described in Table 11.

Figure 15. L2 Writeback Word Count Register (L2WWC)

31 16 15 0

Reserved L2 Writeback Word Count (L2WWC)

R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 11. L2 Writeback Word Count Register (L2WWC) Field Descriptions

Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L2WWC OF(value) 0−FFFFh L2 writeback word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2WWC_L2WWC_symval

SPRU610C TMS320C64x Two-Level Internal Memory 43

Registers

6.6 L2 Writeback−Invalidate Base Address Register (L2WIBAR)

The L2 writeback−invalidate base address register (L2WIBAR) is shown in
Figure 16 and described in Table 12.

Figure 16. L2 Writeback−Invalidate Base Address Register (L2WIBAR)

31 0

L2 Writeback−Invalidate Base Address (L2WIBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 12. L2 Writeback−Invalidate Base Address Register (L2WIBAR) Field Descriptions

Bit Field symval† Value Description
31−0 L2WIBAR OF(value) 0−FFFF FFFFh L2 writeback−invalidate base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2WIBAR_L2WIBAR_symval

6.7 L2 Writeback−Invalidate Word Count Register (L2WIWC)

The L2 writeback−invalidate word count register (L2WIWC) is shown in
Figure 17 and described in Table 13.

Figure 17. L2 Writeback−Invalidate Word Count Register (L2WIWC)

31 16 15 0

Reserved L2 Writeback−Invalidate Word Count (L2WIWC)

R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 13. L2 Writeback−Invalidate Word Count Register (L2WIWC) Field Descriptions

Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L2WIWC OF(value) 0−FFFFh L2 writeback−invalidate word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2WIWC_L2WIWC_symval

44 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.8 L2 Invalidate Base Address Register (L2IBAR)

The L2 invalidate base address register (L2IBAR) is shown in Figure 18 and
described in Table 12.

Figure 18. L2 Invalidate Base Address Register (L2IBAR)

31 0

L2 Invalidate Base Address (L2IBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 14. L2 Invalidate Base Address Register (L2IBAR) Field Descriptions

Bit Field symval† Value Description
31−0 L2IBAR OF(value) 0−FFFF FFFFh L2 invalidate base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2IBAR_L2IBAR_symval

6.9 L2 Invalidate Word Count Register (L2IWC)

The L2 invalidate word count register (L2IWC) is shown in Figure 19 and
described in Table 15.

Figure 19. L2 Invalidate Word Count Register (L2IWC)

31 16 15 0

Reserved L2 Invalidate Clean Word Count (L2IWC)

R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 15. L2 Invalidate Word Count Register (L2IWC) Field Descriptions

Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L2IWC OF(value) 0−FFFFh L2 invalidate word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L2IWC_L2IWC_symval

SPRU610C TMS320C64x Two-Level Internal Memory 45

Registers

6.10 L1P Invalidate Base Address Register (L1PIBAR)

The L1P invalidate base address register (L1PIBAR) is shown in Figure 20
and described in Table 16.

Figure 20. L1P Invalidate Base Address Register (L1PIBAR)

31 0

L1P Invalidate Base Address (L1PIBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 16. L1P Invalidate Base Address Register (L1PIBAR) Field Descriptions
Bit Field symval† Value Description
31−0 L1PIBAR OF(value) 0−FFFF FFFFh L1P invalidate base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1PIBAR_L1PIBAR_symval

6.11 L1P Invalidate Word Count Register (L1PIWC)

The L1P invalidate word count register (L1PIWC) is shown in Figure 21 and
described in Table 17.

Figure 21. L1P Invalidate Word Count Register (L1PIWC)

31 16 15 0

Reserved L1P Invalidate Word Count (L1PIWC)

R-0 R/W-x
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 17. L1P Invalidate Word Count Register (L1PIWC) Field Descriptions
Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L1PIWC OF(value) 0−FFFFh L1P invalidate word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1PIWC_L1PIWC_symval

46 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.12 L1D Writeback−Invalidate Base Address Register (L1DWIBAR)

The L1D writeback−invalidate base address register (L1DWIBAR) is shown in
Figure 22 and described in Table 18.

Figure 22. L1D Writeback−Invalidate Base Address Register (L1DWIBAR)

31 0

L1D Writeback−Invalidate Base Address (L1DWIBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 18. L1D Writeback−Invalidate Base Address Register (L1DWIBAR) Field Descriptions
Bit Field symval† Value Description
31−0 L1DWIBAR OF(value) 0−FFFF FFFFh L1D writeback−invalidate base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1DWIBAR_L1DWIBAR_symval

6.13 L1D Writeback−Invalidate Word Count Register (L1DWIWC)

The L1D writeback−invalidate word count register (L1DWIWC) is shown in
Figure 23 and described in Table 19.

Figure 23. L1D Writeback−Invalidate Word Count Register (L1DWIWC)

31 16 15 0

Reserved L1D Writeback−Invalidate Word Count (L1DWIWC)

R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 19. L1D Writeback−Invalidate Word Count Register (L1DWIWC) Field Descriptions
Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L1DWIWC OF(value) 0−FFFFh L1D writeback−invalidate word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1DWIWC_L1DWIWC_symval

SPRU610C TMS320C64x Two-Level Internal Memory 47

Registers

6.14 L1D Invalidate Base Address Register (L1DIBAR)

The L1D invalidate base address register (L1DIBAR) is shown in Figure 24
and described in Table 20.

Figure 24. L1D Invalidate Base Address Register (L1DIBAR)

31 0

L1D Invalidate Base Address (L1DIBAR)

R/W-0
Legend: R/W = Read/write; -n = value after reset

Table 20. L1D Invalidate Base Address Register (L1DIBAR) Field Descriptions
Bit Field symval† Value Description
31−0 L1DIBAR OF(value) 0−FFFF FFFFh L1D invalidate base address.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1DIBAR_L1DIBAR_symval

6.15 L1D Invalidate Word Count Register (L1DIWC)

The L1D invalidate word count register (L1DIWC) is shown in Figure 25 and
described in Table 21.

Figure 25. L1D Invalidate Word Count Register (L1DIWC)

31 16 15 0

Reserved L1D Invalidate Word Count (L1DIWC)

R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 21. L1D Invalidate Word Count Register (L1DIWC) Field Descriptions
Bit Field symval† Value Description
31−16 Reserved − 0 Reserved. The reserved bit location is always read as 0.
Bits in this field should always be written with 0.

15−0 L1DIWC OF(value) 0−FFFFh L1D invalidate word count.

DEFAULT 0
† For CSL implementation, use the notation CACHE_L1DIWC_L1DIWC_symval

48 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.16 L2 Writeback All Register (L2WB)

The L2 writeback all register (L2WB) is shown in Figure 26 and described in
Table 22.

Figure 26. L2 Writeback All Register (L2WB)

31 1 0
Reserved C
R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 22. L2 Writeback All Register (L2WB) Field Descriptions

Bit Field symval† Value Description

31−1 Reserved − 0 Reserved. The reserved bit location is always read as 0. Bits in
this field should always be written with 0.

0 C OF(value) L2 writeback all.

DEFAULT 0 Normal L2 operation

NORMAL

FLUSH 1 All L2 lines writeback all

† For CSL implementation, use the notation CACHE_L2WB_C_symval

SPRU610C TMS320C64x Two-Level Internal Memory 49

Registers

6.17 L2 Writeback−Invalidate All Register (L2WBINV)

The L2 writeback−invalidate all register (L2WBINV) is shown in Figure 27 and
described in Table 23.

Figure 27. L2 Writeback-Invalidate All Register (L2WBINV)

31 1 0
Reserved C
R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 23. L2 Writeback−Invalidate All Register (L2WBINV) Field Descriptions

Bit Field symval† Value Description

31−1 Reserved − 0 Reserved. The reserved bit location is always read as 0. Bits in
this field should always be written with 0.

0 C OF(value) L2 writeback−invalidate all.

DEFAULT 0 Normal L2 operation

NORMAL

CLEAN 1 All L2 lines writeback−invalidate all

† For CSL implementation, use the notation CACHE_L2WBINV_C_symval

50 TMS320C64x Two-Level Internal Memory SPRU610C

 Registers

6.18 L2 Memory Attribute Registers (MAR0−MAR255)

The L2 memory attribute register (MAR) is shown in Figure 28 and described
in Table 24. Each MAR controls the cacheability of a 16-Mbyte address range,
see section 7.2.

Figure 28. L2 Memory Attribute Register (MAR)

31 1 0
Reserved CE
R-0 R/W-0
Legend: R = Read only; R/W = Read/write; -n = value after reset

Table 24. Memory Attribute Register (MAR) Field Descriptions

Bit Field symval† Value Description

31−1 Reserved − 0 Reserved. The reserved bit location is always read as 0. Bits in
this field should always be written with 0.

0 CE OF(value) Cache enable bit determines whether the L1D, L1P, and L2 are
allowed to cache the corresponding address range.

DEFAULT 0 Memory range is not cacheable.

DISABLE

ENABLE 1 Memory range is cacheable.

† For CSL implementation, use the notation CACHE_MAR_CE_symval

SPRU610C TMS320C64x Two-Level Internal Memory 51

Memory System Control

7 Memory System Control

7.1 Cache Mode Selection

The cache mode for the two-level memory hierarchy is determined by the
cache configuration register (CCFG), and the data cache control (DCC) and
program cache control (PCC) fields of the CPU control and status register
(CSR). The CSR is shown in Figure 29.

Figure 29. CPU Control and Status Register (CSR)

31 24 23 16
CPU ID Revision ID

15 10 9 8 7 5 4 2 1 0
PWRD SAT EN PCC DCC PGIE GIE

7.1.1 L1D Mode Selection Using DCC Field in CSR

The L1D only operates as a cache and cannot be memory mapped. The L1D
does not support freeze or bypass modes. The only values allowed for the data
cache control (DCC) field are 000b and 010b. All other values for DCC are
reserved, as shown in Table 25.

Table 25. L1D Mode Setting Using DCC Field

Bit Field Value Description

4−2 DCC Data cache control bit.

000 2-way cache enabled

001 Reserved

010 2-way cache enabled

011−111 Reserved

52 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

7.1.2 L1P Mode Selection Using PCC Field in CSR

The L1P only operates as a cache and cannot be memory mapped. The L1P
does not support freeze or bypass modes. The only values allowed for the
program cache control (PCC) field are 000b and 010b. All other values for PCC
are reserved, as shown in Table 26.

Table 26. L1P Mode Setting Using PCC Field

Bit Field Value Description
7−5 PCC Program cache control bit.

000 Direct-mapped cache enabled

001 Reserved

010 Direct-mapped cache enabled

011−111 Reserved

7.1.3 L2 Mode Selection Using L2MODE Field in CCFG

The L2 memory can function as mapped SRAM, as cache, or as some
combination of both. The L2MODE field in the cache configuration register
(CCFG) determines what portion of L2 is mapped as SRAM and what portion
acts as cache. For some settings of L2MODE, either L2 SRAM or L2 cache
may not be present. The mode with no L2 cache is referred to as all SRAM
mode. The CCFG is shown in Figure 11 (page 39) and described in Table 7.
The reset value of the L2MODE field is 000b, so the L2 is configured in all
SRAM mode after reset. That is, there is no L2 cache enabled and the entire
L2 SRAM address range is available. The L2 still services external memory
requests for the L1P and L1D in this mode. It will not, however, retain copies
of the data itself.
By default, external memory is not marked as cacheable. Cacheability is
controlled separately of the cache mode. Section 7.2 discusses the memory
attribute registers (MAR) that control the cacheability of external memory.
As various amounts of L2 cache are enabled, the amount of L2 SRAM
available decreases. Addresses at the top of the L2 SRAM address space are
converted to cache. For example, consider the C6414 DSP, which offers
1024K bytes of L2 memory at addresses from 0000 0000h−000F FFFFh.
When the L2 is in 256K cache mode (L2MODE = 111b), the uppermost
256K-byte range of addresses, 000C 0000h−000F FFFFh, is not available to
programs.

SPRU610C TMS320C64x Two-Level Internal Memory 53

Memory System Control

Note:
Reads or writes to L2 address ranges that are configured as L2 cache may
result in undesired operation of the cache hierarchy. Programs must confine
L2 accesses to L2 addresses that are mapped as L2 SRAM to ensure correct
program operation.
The L2 controller processes reads to the address range
0000 0000h−0010 0FFFh without undesired cache operation, even if some
of these addresses are configured as L2 cache. This address range
represents the entire allotted L2 address range, plus some additional space
to allow for certain program optimizations. Therefore, the restriction above
does not apply to reads; however, programs should not interpret values
returned by reads nor should programs perform writes to L2 addresses that
are not configured as L2 SRAM.

To ensure correct operation when switching L2 cache modes, you must

perform a series of operations. Table 27 specifies the required operations
when either adding or removing L2 memory as mapped SRAM. Failure to
follow these guidelines may result in data loss and undefined L2 operation.

54 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

Table 27. L2 Mode Switch Procedure

To Switch From... To... Perform the Following Steps

All SRAM Mode Mode with L2 cache 1) Use EDMA to transfer any data needed out of the L2 SRAM
space to be converted into cache.
2) Perform a block writeback-invalidate in L1D of L2 SRAM
addresses that are about to become L2 cache.
3) Wait for block writeback-invalidate to complete.
4) Perform block writeback-invalidates for any external address
ranges that may be cached in L1D. (This step is unnecessary
if no CE bit has been set to 1 in any MAR.)
5) Wait for the block writeback-invalidates from step 4 to
complete.
6) Write to CCFG to change mode.
7) Force CPU to wait for CCFG modification by reading CCFG.
8) Execute 8 cycles of NOP.

Mode with mixed Mode with less L2 1) Use EDMA to transfer any data needed out of the L2 SRAM
L2 SRAM and mapped SRAM space to be converted into cache.
L2 cache
2) Perform a block writeback-invalidate in L1D of L2 SRAM
addresses that are about to become L2 cache.
3) Wait for block writeback-invalidate to complete.
4) Perform global writeback-invalidate of L2 (L2WBINV).
5) Wait for L2WBINV to complete.
6) Write to CCFG to change mode.
7) Force CPU to wait for CCFG modification by reading CCFG.
8) Execute 8 cycles of NOP.

Any L2 mode Mode with more L2 1) Perform global writeback-invalidate of L2 (L2WBINV).

mapped SRAM
2) Wait for L2WBINV to complete.
3) Write to CCFG to change mode.
4) Force CPU to wait for CCFG modification by reading CCFG.
5) Execute 8 cycles of NOP.

SPRU610C TMS320C64x Two-Level Internal Memory 55

Memory System Control

7.2 Cacheability Controls

The cacheability of external address ranges is controlled by the memory
attribute registers (MAR0−MAR255). Each MAR controls the cacheability of
a 16-Mbyte address range as listed in Table 28. MAR is shown in Figure 28
(page 51) and described in Table 24. See the device-specific datasheet for the
memory address of these registers.

The cache enable (CE) bit in each MAR determines whether the L1D, L1P, and
L2 are allowed to cache the corresponding address range. After reset, the CE
bit in each MAR is cleared to 0, thereby disabling caching of external memory
by default. This is in contrast to L2 SRAM, which is always considered cacheable.

To enable caching on a particular external address range, an application

should set the CE bit in the appropriate MAR to 1. No special procedure is
necessary. Subsequent accesses to the affected address range are cached
by the two-level memory system.

To disable caching for a given address range, programs should follow the
following sequence to ensure that all future accesses to the particular address
range are not cached.

1) Ensure that all addresses within the affected range are removed from the
L1 and L2 caches. This is accomplished in one of the following ways. Any
one of the following operations should be sufficient.
a) If L2 cache is enabled, invoke a global writeback-invalidate using
L2WBINV. Wait for the C bit in L2WBINV to read as 0. Alternately,
invoke a block writeback-invalidate of the affected range using
L2WIBAR/L2WIWC. Wait for L2WIWC to read as 0.
b) If L2 is in all SRAM mode, invoke a block writeback-invalidate of the
affected range using L1DWIBAR/L1DWIWC. Wait for L1DWIWC to
read as 0.
Note that the block-oriented cache controls can only operate on a
256K-byte address range at a time, so multiple block writeback-invalidate
operations may be necessary to remove the entire affected address range
from the cache. These cache controls are discussed in section 7.3.

2) Clear the CE bit in the appropriate MAR to 0.

56 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

Table 28. Memory Attribute Registers

Acronym Register Name

MAR0−MAR95 Reserved
MAR96† L2 memory attribute register 96: controls EMIFB CE0 range 6000 0000h−60FF FFFFh
MAR97† L2 memory attribute register 97: controls EMIFB CE0 range 6100 0000h−61FF FFFFh
MAR98† L2 memory attribute register 98: controls EMIFB CE0 range 6200 0000h−62FF FFFFh
MAR99† L2 memory attribute register 99: controls EMIFB CE0 range 6300 0000h−63FF FFFFh
MAR100† L2 memory attribute register 100: controls EMIFB CE1 range 6400 0000h−64FF FFFFh
MAR101† L2 memory attribute register 101: controls EMIFB CE1 range 6500 0000h−65FF FFFFh
MAR102† L2 memory attribute register 102: controls EMIFB CE1 range 6600 0000h−66FF FFFFh
MAR103† L2 memory attribute register 103: controls EMIFB CE1 range 6700 0000h−67FF FFFFh
MAR104† L2 memory attribute register 104: controls EMIFB CE2 range 6800 0000h−68FF FFFFh
MAR105† L2 memory attribute register 105: controls EMIFB CE2 range 6900 0000h−69FF FFFFh
MAR106† L2 memory attribute register 106: controls EMIFB CE2 range 6A00 0000h−6AFF FFFFh
MAR107† L2 memory attribute register 107: controls EMIFB CE2 range 6B00 0000h−6BFF FFFFh
MAR108† L2 memory attribute register 108: controls EMIFB CE3 range 6C00 0000h−6CFF FFFFh
MAR109† L2 memory attribute register 109: controls EMIFB CE3 range 6D00 0000h−6DFF FFFFh
MAR110† L2 memory attribute register 110: controls EMIFB CE3 range 6E00 0000h−6EFF FFFFh
MAR111† L2 memory attribute register 111: controls EMIFB CE3 range 6F00 0000h−6FFF FFFFh
MAR112−MAR127 Reserved
MAR128 L2 memory attribute register 128: controls EMIFA CE0 range 8000 0000h−80FF FFFFh
MAR129 L2 memory attribute register 129: controls EMIFA CE0 range 8100 0000h−81FF FFFFh
MAR130 L2 memory attribute register 130: controls EMIFA CE0 range 8200 0000h−82FF FFFFh
MAR131 L2 memory attribute register 131: controls EMIFA CE0 range 8300 0000h−83FF FFFFh
MAR132 L2 memory attribute register 132: controls EMIFA CE0 range 8400 0000h−84FF FFFFh
MAR133 L2 memory attribute register 133: controls EMIFA CE0 range 8500 0000h−85FF FFFFh
MAR134 L2 memory attribute register 134: controls EMIFA CE0 range 8600 0000h−86FF FFFFh
MAR135 L2 memory attribute register 135: controls EMIFA CE0 range 8700 0000h−87FF FFFFh
MAR136 L2 memory attribute register 136: controls EMIFA CE0 range 8800 0000h−88FF FFFFh
MAR137 L2 memory attribute register 137: controls EMIFA CE0 range 8900 0000h−89FF FFFFh
† MAR96−MAR111 only available on the C6414/C6415/C6416 devices; on all other devices, these registers are reserved.

SPRU610C TMS320C64x Two-Level Internal Memory 57

Memory System Control

Table 28. Memory Attribute Registers (Continued)

Acronym Register Name
MAR138 L2 memory attribute register 138: controls EMIFA CE0 range 8A00 0000h−8AFF FFFFh
MAR139 L2 memory attribute register 139: controls EMIFA CE0 range 8B00 0000h−8BFF FFFFh
MAR140 L2 memory attribute register 140: controls EMIFA CE0 range 8C00 0000h−8CFF FFFFh
MAR141 L2 memory attribute register 141: controls EMIFA CE0 range 8D00 0000h−8DFF FFFFh
MAR142 L2 memory attribute register 142: controls EMIFA CE0 range 8E00 0000h−8EFF FFFFh
MAR143 L2 memory attribute register 143: controls EMIFA CE0 range 8F00 0000h−8FFF FFFFh
MAR144 L2 memory attribute register 144: controls EMIFA CE1 range 9000 0000h−90FF FFFFh
MAR145 L2 memory attribute register 145: controls EMIFA CE1 range 9100 0000h−91FF FFFFh
MAR146 L2 memory attribute register 146: controls EMIFA CE1 range 9200 0000h−92FF FFFFh
MAR147 L2 memory attribute register 147: controls EMIFA CE1 range 9300 0000h−93FF FFFFh
MAR148 L2 memory attribute register 148: controls EMIFA CE1 range 9400 0000h−94FF FFFFh
MAR149 L2 memory attribute register 149: controls EMIFA CE1 range 9500 0000h−95FF FFFFh
MAR150 L2 memory attribute register 150: controls EMIFA CE1 range 9600 0000h−96FF FFFFh
MAR151 L2 memory attribute register 151: controls EMIFA CE1 range 9700 0000h−97FF FFFFh
MAR152 L2 memory attribute register 152: controls EMIFA CE1 range 9800 0000h−98FF FFFFh
MAR153 L2 memory attribute register 153: controls EMIFA CE1 range 9900 0000h−99FF FFFFh
MAR154 L2 memory attribute register 154: controls EMIFA CE1 range 9A00 0000h−9AFF FFFFh
MAR155 L2 memory attribute register 155: controls EMIFA CE1 range 9B00 0000h−9BFF FFFFh
MAR156 L2 memory attribute register 156: controls EMIFA CE1 range 9C00 0000h−9CFF FFFFh
MAR157 L2 memory attribute register 157: controls EMIFA CE1 range 9D00 0000h−9DFF FFFFh
MAR158 L2 memory attribute register 158: controls EMIFA CE1 range 9E00 0000h−9EFF FFFFh
MAR159 L2 memory attribute register 159: controls EMIFA CE1 range 9F00 0000h−9FFF FFFFh
MAR160 L2 memory attribute register 160: controls EMIFA CE2 range A000 0000h−A0FF FFFFh
MAR161 L2 memory attribute register 161: controls EMIFA CE2 range A100 0000h−A1FF FFFFh
MAR162 L2 memory attribute register 162: controls EMIFA CE2 range A200 0000h−A2FF FFFFh
MAR163 L2 memory attribute register 163: controls EMIFA CE2 range A300 0000h−A3FF FFFFh
MAR164 L2 memory attribute register 164: controls EMIFA CE2 range A400 0000h−A4FF FFFFh
MAR165 L2 memory attribute register 165: controls EMIFA CE2 range A500 0000h−A5FF FFFFh
† MAR96−MAR111 only available on the C6414/C6415/C6416 devices; on all other devices, these registers are reserved.

58 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

Table 28. Memory Attribute Registers (Continued)

Acronym Register Name
MAR166 L2 memory attribute register 166: controls EMIFA CE2 range A600 0000h−A6FF FFFFh
MAR167 L2 memory attribute register 167: controls EMIFA CE2 range A700 0000h−A7FF FFFFh
MAR168 L2 memory attribute register 168: controls EMIFA CE2 range A800 0000h−A8FF FFFFh
MAR169 L2 memory attribute register 169: controls EMIFA CE2 range A900 0000h−A9FF FFFFh
MAR170 L2 memory attribute register 170: controls EMIFA CE2 range AA00 0000h−AAFF FFFFh
MAR171 L2 memory attribute register 171: controls EMIFA CE2 range AB00 0000h−ABFF FFFFh
MAR172 L2 memory attribute register 172: controls EMIFA CE2 range AC00 0000h−ACFF FFFFh
MAR173 L2 memory attribute register 173: controls EMIFA CE2 range AD00 0000h−ADFF FFFFh
MAR174 L2 memory attribute register 174: controls EMIFA CE2 range AE00 0000h−AEFF FFFFh
MAR175 L2 memory attribute register 175: controls EMIFA CE2 range AF00 0000h−AFFF FFFFh
MAR176 L2 memory attribute register 176: controls EMIFA CE3 range B000 0000h−B0FF FFFFh
MAR177 L2 memory attribute register 177: controls EMIFA CE3 range B100 0000h−B1FF FFFFh
MAR178 L2 memory attribute register 178: controls EMIFA CE3 range B200 0000h−B2FF FFFFh
MAR179 L2 memory attribute register 179: controls EMIFA CE3 range B300 0000h−B3FF FFFFh
MAR180 L2 memory attribute register 180: controls EMIFA CE3 range B400 0000h−B4FF FFFFh
MAR181 L2 memory attribute register 181: controls EMIFA CE3 range B500 0000h−B5FF FFFFh
MAR182 L2 memory attribute register 182: controls EMIFA CE3 range B600 0000h−B6FF FFFFh
MAR183 L2 memory attribute register 183: controls EMIFA CE3 range B700 0000h−B7FF FFFFh
MAR184 L2 memory attribute register 184: controls EMIFA CE3 range B800 0000h−B8FF FFFFh
MAR185 L2 memory attribute register 185: controls EMIFA CE3 range B900 0000h−B9FF FFFFh
MAR186 L2 memory attribute register 186: controls EMIFA CE3 range BA00 0000h−BAFF FFFFh
MAR187 L2 memory attribute register 187: controls EMIFA CE3 range BB00 0000h−BBFF FFFFh
MAR188 L2 memory attribute register 188: controls EMIFA CE3 range BC00 0000h−BCFF FFFFh
MAR189 L2 memory attribute register 189: controls EMIFA CE3 range BD00 0000h−BDFF FFFFh
MAR190 L2 memory attribute register 190: controls EMIFA CE3 range BE00 0000h−BEFF FFFFh
MAR191 L2 memory attribute register 191: controls EMIFA CE3 range BF00 0000h−BFFF FFFFh
MAR192−MAR255 Reserved
† MAR96−MAR111 only available on the C6414/C6415/C6416 devices; on all other devices, these registers are reserved.

SPRU610C TMS320C64x Two-Level Internal Memory 59

Memory System Control

7.3 Program-Initiated Cache Operations

The memory system provides a set of cache control operations. These allow
programs to specifically request that certain data be written back or
invalidated. The cache operations fall into two categories, global operations
that operate on the entire cache, and block operations that operate on a
specific range of addresses. The global operations are discussed in
section 7.3.1 and the block operations are discussed in section 7.3.2.

The memory system can only perform one program-initiated cache operation
at a time. This includes global operations, block operations, and mode
changes. For this reason, the memory system may stall accesses to cache
control registers while a cache control operation is in progress. Table 29 gives
a summary of the available operations and their impact on the memory
system.

Table 29. Summary of Program-Initiated Cache Operations

Type of Cache Register L1P Cache
Operation Operation Usage Effect L1D Cache Effect L2 Cache Effect
Global L2 L2WB No effect. Updated lines holding Updated data is
Operation Writeback addresses also held in written back. All lines
All L2 cache are written kept valid.
back. All lines
corresponding to
addresses held in L2
cache are invalidated.†

L2 L2WBINV Entire contents Updated lines holding Updated lines are

Writeback are discarded.† addresses also held in written back. All lines
and L2 cache are written are invalidated.
Invalidate back. All lines
All corresponding to
addresses held in L2
cache are invalidated.†

L1P IP bit in Entire contents No effect. No effect.

Invalidate CCFG are discarded.
All

L1D ID bit in No effect. Entire contents are No effect.

Invalidate CCFG discarded. No updated
All data is written back.
† As described in section 7.3.3, these operations only affect L1D and L1P when L2 cache is enabled. When L2 cache is enabled,
these operations only affect those portions of L1P and L1D that are also held in L2 cache. Future C6000 devices may cause
these operands to act on the entire contents of L1P and L1D, regardless of whether the same addresses are cached in L2, or
whether L2 cache is enabled.
‡ In contrast to its behavior on C621x/C671x devices, L1DWIBAR/L1DWIWC on C64x devices do not cause the corresponding
block to be invalidated in L1P.

60 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

Table 29. Summary of Program-Initiated Cache Operations (Continued)

Type of Cache Register L1P Cache
Operation Operation Usage Effect L1D Cache Effect L2 Cache Effect
Block L2 Block L2WBAR, No effect Updated lines in block Updated lines in
Operation Writeback L2WWC are written to L2. All block are written to
lines in block are external memory.
invalidated in L1D.† Lines in block are
kept valid in L2.

L2 Block L2WIBAR, All lines in block Updated lines in block Updated lines in
Writeback L2WIWC are invalidated.† are written to L2. All block are written to
with lines in block are external memory. All
Invalidate invalidated in L1D.† lines in block are
invalidated in L2.

L2 Block L2IBAR, All lines in block All lines in block are All lines in block are
Invalidate L2IWC are invalidated.† invalidated in L1D. invalidated in L2.
Updated data in block is Updated data in
discarded.† block is discarded.

L1P Block L1PIBAR, All lines in block No effect. No effect.

Invalidate L1PIWC are invalidated.

L1D Block L1DWIBAR, No effect.‡ Updated lines in block No effect.

Writeback L1DWIWC are written to L2. All
with lines in block are
Invalidate invalidated in L1D.

L1D Block L1DIBAR, No effect. All lines in block are No effect.

Invalidate L1DIWC invalidated in L1D.
Updated data in block is
discarded.
† As described in section 7.3.3, these operations only affect L1D and L1P when L2 cache is enabled. When L2 cache is enabled,
these operations only affect those portions of L1P and L1D that are also held in L2 cache. Future C6000 devices may cause
these operands to act on the entire contents of L1P and L1D, regardless of whether the same addresses are cached in L2, or
whether L2 cache is enabled.
‡ In contrast to its behavior on C621x/C671x devices, L1DWIBAR/L1DWIWC on C64x devices do not cause the corresponding
block to be invalidated in L1P.

SPRU610C TMS320C64x Two-Level Internal Memory 61

Memory System Control

7.3.1 Global Cache Operations

The global cache operations execute on the entire contents of a particular
cache. Global operations take precedence over program accesses to the
cache. Program accesses (either data or program fetches) to a particular
cache are stalled while a global cache operation is active on that cache.

Global operations in L1D and L1P are initiated by the ID and IP bits in CCFG
(Figure 11). The L1D and L1P only offer global invalidation. By writing a 1 to
the ID or IP bit in CCFG, a program can invalidate the entire contents of the
corresponding L1 cache. Upon initiation of the global invalidate, the entire
contents of the corresponding cache is discarded — no updated data is written
back. Reading CCFG after the write will stall the program until the invalidate
operation is complete.

The L1D global-invalidate causes all updated data in L1D to be

discarded, rather than written back to the lower levels of memory.
This can cause incorrect operation in programs that expect the
updates to be written to the lower levels of memory. Therefore,
most programs use either the L1D block writeback-invalidate
(described in section 7.3.2) or the global L2 operations, rather than
the L1D global-invalidate.

The L2 offers both global writeback and global writeback-invalidate

operations. Global cache operations in L2 are initiated by writing a 1 to the C
bit in either L2WB shown in Figure 26 (page 49) and described in Table 22, or
L2WBINV shown in Figure 27 (page 50) and described in Table 23. Writing a
1 to the C bit of L2WB initiates a global writeback of L2; writing a 1 to the C bit
of L2WBINV initiates a global writeback-invalidate of L2. The C bit continues
to read as 1 until the cache operation is complete. Programs can poll to
determine when a cache operation is complete.

Global operations on L2 have indirect effects on the L1 caches, as discussed

in section 7.3.3.

62 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

7.3.2 Block Cache Operations

The block cache operations execute on a range of addresses that may be held
in the cache. Block operations execute in the background, allowing other
program accesses to interleave with the block cache operation.

Programs initiate block cache operations with two writes. The program first
writes the starting address to one of the base address registers (BAR), shown
in Figure 30. Next, the program writes the total number of words to operate on
to the corresponding word count register (WC), shown in Figure 31. The cache
operation begins as soon as the word count register is written with a non-zero
value. The cache provides a set of BAR/WC pseudo-register pairs, one for
each block operation the cache supports. The complete list of supported
operations is shown in Table 29.

Notice that the word count field in WC is only 16-bits wide. This limits the block
size to 65535 words (approximately 256K bytes). Larger ranges require
multiple block commands to be issued to cover the entire range.

Although block operations specify the block in terms of a word address and
word count, the block operations always operate on whole cache lines. Whole
lines are always be written back and/or invalidated in each affected cache. For
this reason, programs should be careful to align arrays on cache-line
boundaries, and to pad arrays to be a multiple of the cache line size. This is
especially true when invoking the invalidate-only commands with respect to
these arrays.

Figure 30. Block Cache Operation Base Address Register (BAR)

31 0
Base Address
R/W-0

Legend: R/W = Read/write; -n = value after reset

Figure 31. Block Cache Operation Word Count Register (WC)

31 16 15 0
Reserved Word Count
R-0 R/W-0

Legend: R = Read only; R/W = Read/write; -n = value after reset

SPRU610C TMS320C64x Two-Level Internal Memory 63

Memory System Control

In current device implementations, the block command registers (BAR and

WC) are implemented as a single register pair in the underlying hardware.
They appear as pseudo-registers at multiple addresses in the address map.
Therefore, programs must not interleave writes to different command register
pairs. Writes to one BAR/WC pair should be performed atomically with respect
to other BAR/WC writes. Depending on the nature of the program, it may be
necessary to disable interrupts while writing to BAR and WC.

Despite this restriction, programs may attempt to initiate a new block

command while a previous command is in progress. The cache controller will
stall the write to BAR until the previous cache operation completes, thus
preserving correct operation. To avoid stalling, programs may determine if a
block cache operation has completed by polling WC. WC returns a non-zero
value while the cache operation is in progress, and zero when the operation
has completed. The non-zero value returned may vary between device
implementations, so programs should only rely on the register being non-zero.

Programs should not assume that the value of BAR is retained between block
cache operations. Programs should always write an address to BAR and a
word count to WC for each cache operation. Also, programs should not
assume that the various BAR and WC map to the same physical register. For
each cache operation, programs should write to BAR and WC for that
operation.

The L1P block invalidate can be used in conjunction with the L1D block
writeback-invalidate to provide software controlled coherence between L1D
and L1P. (Section 8.1 discusses the memory system coherence policies.) To
execute code that was previously written to the CPU, the program should use
L1PIBAR/L1PIWC to invalidate the block in L1P, and L1DWIBAR/L1DWIWC
to writeback-invalidate the block in L1D. These operations can be performed
in either order. The specific timing of these operations relative to program
fetches is not well defined. Therefore, programs should wait for L1DWIWC and
L1PIWC to read as zero prior to branching to an address range that has been
invalidated in this manner. (Note that the behavior of L1DIBAR/L1DIWC differs
on C621x/C671x devices. See TMS320C621x/C671x DSP Two-Level
Internal Memory Reference Guide, SPRU609, for details.)

Block cache operations in L2 can indirectly affect L1D and L1P, as noted in
Table 29. Section 7.3.3 discusses these interactions in detail.

64 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

Note:
Reads or writes to the addresses within the block being operated on while
a block cache operation is in progress may cause those addresses to not be
written back or invalidated as requested. To avoid this, programs should not
access addresses within the range of cache lines affected by a block cache
operation while the operation is in progress. Programs may consult the
appropriate WC to determine when the block operation is complete.

7.3.3 Effect of L2 Commands on L1 Caches

Note:
The behaviors described in this section may change on future C6000
devices that implement the two-level memory subsystem. Future devices
may remove the inclusivity of L1D within L2, and may cause the L2 cache
operations to act on L1, regardless of the current L2 cache mode. Forward-
compatible programs should not rely on these specific memory system
behaviors.

Cache operations in L2 indirectly operate on the L1D and L1P cache. As a

result, the L2 cache operations have no effect on any of the caches when L2
is in all SRAM mode. This is true for both the global and block commands.
Otherwise, when L2 cache is enabled, program-initiated cache operations in
L2 may operate on the corresponding contents of the L1D and L1P.

Under normal circumstances, the L1D cache is inclusive in L2, and L1P is not.
Inclusive implies that the entire contents of L1D are also held either in
L2 SRAM or L2 cache. The L2 cache operations are designed with these
properties in mind.

Because L1P is not inclusive in L2, the L2 cache operations that invalidate
lines in L2 send explicit invalidation commands to L1P. A global
writeback-invalidate of L2 (L2WBINV) triggers a complete invalidation of L1P.
Block invalidate and writeback-invalidate operations in L2 blindly send
invalidate commands to L1P for the corresponding L1P cache lines. This
ensures that L1P always fetches the most recent contents of memory after the
cache operation is complete.

SPRU610C TMS320C64x Two-Level Internal Memory 65

Memory System Control

Because L1D is normally inclusive in L2, the L2 relies on normal cache

protocol operation to operate on L1D indirectly. Writebacks and invalidates
triggered for lines in L2 result in snoop-invalidate commands sent to L1D when
L2 detects that L1D is also holding a copy of the corresponding addresses.
Therefore, the L2 global writeback (L2WB) and L2 global writeback-invalidate
(L2WBINV) causes all external addresses held in L1D to be written back and
invalidated from L1D. Likewise, block operations in L2 cause the
corresponding range in L1D to be written to L2 and invalidated from L1D using
the indirect snoop-invalidates.

One result of this is that L2 SRAM addresses cached in L1D are not affected
by program-initiated cache operations in L2, as L2 cache never holds copies
of L2 SRAM. To remove L2 SRAM addresses from L1D, programs must use
the L1D block cache operations directly. Ordinarily, direct removal of L2 SRAM
addresses from L1D is required only when changing L2 cache modes. The
coherence policy described in section 8.1 makes unnecessary most of the
need for programs to manually write back portions of L1D to L2 SRAM.

Another result is nonintuitive behavior when L1D is not inclusive in L2. L1D is
inclusive in L2 under normal circumstances, and so most programs do not
need to be concerned about this situation. Indeed, the recommended L2
cache mode-change procedure in section 7.1.3 ensures that the memory
system is never in this state. When not following the procedure precisely, it is
possible for L1D to hold copies of external memory that are not held in L2. This
noninclusive state is achieved in the following rare sequence:

- The program enabled caching for an external address range while L2 was
in all SRAM mode.

- The program read directly from this external address range.

- The program then enabled L2 cache without first removing the external
address range from L1D with a block cache operation.

To prevent this situation, programs should ensure no external addresses are

cached in L1D when enabling L2 cache. This is accomplished by not enabling
caching for external addresses prior to enabling L2 cache, or by removing
external addresses from L1D using the block cache operations described in
section 7.3.2.

66 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Control

7.4 L2-to-EDMA Request Control

As described in section 5.4.3, the L2 controller relies on the EDMA to service

requests on its behalf. L2 requests may be placed on any of the four EDMA
priority queues. The priority queue selection for cache-service requests is
made using the P field in CCFG (Figure 11, page 39).

Programs should take care when changing the priority level for L2 requests in
order to ensure proper operation of the cache. The following sequence should
be followed:

1) Poll the EDMA priority queue status register (PQSR) and wait for the PQ
bit that corresponds to the current priority level to read as 1. PQSR is
described in TMS320C6000 DSP Enhanced Direct Memory Access
(EDMA) Peripheral Reference Guide (SPRU234).

This step may require that other transfers using this same priority queue,
such as externally-triggered EDMA transfers, be disabled. Otherwise, in a
heavily loaded system, the PQ bit in PQSR may not read as 1 for an
arbitrarily long period of time.

2) Write a new value to the P field in CCFG.

3) Read CCFG back to ensure the write has completed.

To ensure fairness between L2 requests and other EDMA requests in the

system, the L2 provides four priority queue allocation registers, L2ALLOC0,
L2ALLOC1, L2ALLOC2, and L2ALLOC3 (Figure 13, page 42). These
registers control the rate at which L2 requests are issued to the EDMA.
Because the L2 also contains the QDMA control registers, the limits set by
L2ALLOC0−3 also apply to QDMA requests. QDMAs are described in
TMS320C6000 DSP Enhanced Direct Memory Access (EDMA) Peripheral
Reference Guide (SPRU234).

The setting of a given L2ALLOC determines the the total number of

outstanding L2 and QDMA requests permitted on the corresponding EDMA
priority level. Further requests on that priority level are stalled until the number
of outstanding requests falls below the L2ALLOC setting. An outstanding
transfer in this context refers to a transfer that has been submitted to a given
priority queue, but which has not been extracted from the queue into the EDMA
channel registers. L2ALLOC0−3 behave similarly to the priority queue
allocation registers (PQAR) described in TMS320C6000 DSP Enhanced
Direct Memory Access (EDMA) Peripheral Reference Guide (SPRU234).

SPRU610C TMS320C64x Two-Level Internal Memory 67

Memory System Control

Table 30 lists the default queue allocation. The L2ALLOC settings must also
take into account the current settings for PQAR so that no transfers are lost.
The correct procedure for modifying the L2ALLOC and PQAR settings is
described in TMS320C6000 DSP Enhanced Direct Memory Access (EDMA)
Peripheral Reference Guide (SPRU234).

Table 30. L2ALLOC Default Queue Allocations

Default L2/QDMA
Priority Level Allocation Register Allocation
Urgent L2ALLOC0 6

High L2ALLOC1 2

Medium L2ALLOC2 2

Low L2ALLOC3 2

7.5 EDMA Access Into L2 Control

The C64x DSP incorporates an L2 EDMA access control register
(EDMAWEIGHT), located in the L2 cache register memory map, that controls
the relative priority weighting of EDMA versus L1D access into L2.
EDMAWEIGHT gives EDMA accesses a temporary boost in priority by limiting
the amount of time L1D blocks EDMA access to L2. This priority boost only
applies when competing with write data from the CPU that misses in L1D, but
hits in L2 cache or L2 SRAM. Normal line allocate and eviction operation result
in gaps in the L1D-to-L2 access pattern that can be used by EDMA accesses.
EDMAWEIGHT lets you control how often this priority boost is given. When the
EDMA priority is raised, it is allowed to complete one access before priority is
returned to the CPU data. The EDMAWEIGHT is shown in Figure 12, page 41,
and described in Table 8.

If L1D blocks L2 for n consecutive cycles, then EDMA is given priority for a
single cycle.

68 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

8 Memory System Policies

This section discusses the various policies of the memory system, such as
coherence between CPU and EDMA or host accesses and the order in which
memory updates are made.

8.1 Memory System Coherence

Cache memories work by retaining copies of data from lower levels of the
memory hierarchy in the hierarchy’s higher levels. This provides a
performance benefit as higher levels of the memory hierarchy may be
accessed more quickly than the lower levels, and so the lower levels need not
be consulted on every access. Because many accesses to memory are
captured at higher levels of the hierarchy, the opportunity exists for the CPU
and other devices to see a different picture of what is in memory.

A memory system is coherent if all requestors into that memory see updates
to individual memory locations occur in the same order. A requestor is a device
such as a CPU or EDMA. A coherent memory system ensures that all writes
to a given memory location are visible to future reads of that location from any
requestor, so long as no intervening write to that location overwrites the value.
If the same requestor is writing and reading, the results of a write are
immediately visible. If one requestor writes and different requestor reads, the
reader may not see the updated value immediately, but it will be able to see
updates after a sufficient period of time. Coherence also implies that all writes
to a given memory location appear to be serialized. That is, that all requestors
see the same order of writes to a memory location, even when multiple
requestors are writing to one location.

The two-level cache memory system is coherent with various qualifications.

The C64x devices ensure coherence in a limited manner for the programmer.
This limited coherence model is designed to simplify and improve the
performance of the hardware, while still providing a reasonable programming
model.

Memory system coherence only applies to cacheable data. For noncacheable

data, the memory system merely passes the access on to its final destination
without retaining any intermediate copies, and thus coherence is not an issue.
Therefore, this discussion focuses only on cacheable memory, which is
generally confined to various forms of RAM and ROM.

SPRU610C TMS320C64x Two-Level Internal Memory 69

Memory System Policies

The coherence model of the C64x memory system is expressed in terms of

requestors and where the memory is located. The two-level memory system
must support requests from three sources: CPU program fetches, CPU data
accesses, and EDMA accesses. (HPI and PCI accesses to the two-level
memory system are handled using the EDMA.) Cacheable areas of interest
include both internal and external RAM. Since ROM is typically not writeable,
we consider only RAM. Table 31 shows where the memory system ensures
coherency.

Table 31. Coherence Assurances in the Two-Level Memory System

Requestor CPU Program Fetch CPU Data Access EDMA Access

CPU program fetch Coherent Software-managed, On-chip SRAM only
L2 level only

CPU data access Software-managed, Coherent On-chip SRAM only

L2 level only

EDMA access On-chip SRAM only On-chip SRAM only Coherent

Notice that the hardware ensures that accesses by the CPU and EDMA to
internal SRAM addresses are coherent, but external addresses are not.
Software must ensure external addresses accessed by the EDMA are not held
in cache when the EDMA accesses them. Failure to do so can result in data
corruption on the affected range of addresses. See section 7.3 for the steps
required to ensure particular ranges of addresses are not held in cache.

Also notice that CPU data and program accesses are not coherent. The
reason that these are not considered coherent is that the L1P does not query
the L1D when it makes a program fetch, and thus CPU writes captured in L1D
may not be visible to L1P for an arbitrarily long period of time. Therefore,
programs that write to locations that are subsequently executed from must
ensure that the updated data is written from L1D to at least L2 before
execution. The L1DWIBAR/L1DWIWC and L1PIBAR/L1PIWC described in
section 7.3 can be used for this purpose.

Although the memory system is coherent, it does not ensure that all requestors
see updates to cacheable memory occurring in the same order. This is of
primary importance when a given buffer of memory is accessed by the CPU
and an EDMA or host port access. Unless care is taken, it is possible for a
EDMA or host port access to see a mixture of old and new data if the access
occurs while the CPU updates are being processed by the memory system.
Section 8.3 discusses the order in which CPU accesses are made visible to
the memory system.

70 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

8.2 EDMA Coherence in L2 SRAM Example

As indicated in section 8.1, the memory system ensures coherence between
CPU and EDMA accesses to L2 SRAM. It does not provide coherence
between CPU and EDMA accesses to external memory. This section
illustrates how the L2 SRAM coherence assurances are enforced by the
hardware within the context of an example.

A common DSP programming technique is to process streams of data with a

double-buffering scheme. In this setup, a large set of data external to the
device is processed by transferring in small chunks and operating on them.
The results are then transferred to external memory. These transfers are
generally executed by the EDMA controller. In pseudo-code, a typical
processing sequence might as shown in Figure 32.

Because the EDMA operates independently of the CPU, it is possible to

overlap the EDMA transfers with data processing. Double-buffering allows this
by having two sets of input and output buffers—one for processing and one for
transfers-in-progress. The CPU and EDMA alternate sets of buffers,
“ping-ponging” between them. (For this reason, double-buffering is sometimes
referred to as ping-pong buffering.) In pseudo-code, a double-buffered
processing sequence might be as shown in Figure 33. Here, PING and PONG
are pairs pointers to the double-buffers. PING points to the buffers that the
CPU will process, and PONG points to the buffers the EDMA is filling.

Graphically, the time sequence for the internal input and output buffers would
look as shown in Figure 34. EDMA reads are one step ahead of the processing
and EDMA writes are one step behind. In Figure 34, step 2 operates in parallel
with steps 3 and 4; steps 5 and 6 overlap with steps 7 and 8.

Figure 35 shows the read/process/write pipeline in software, with the EDMA

transfers overlapping processing.

SPRU610C TMS320C64x Two-Level Internal Memory 71

Memory System Policies

Figure 32. Streaming Data Pseudo-Code

while (data left to process)

{
Read next block to internal input buffer

Process internal input buffer, placing result in internal output buffer

Write the internal output buffer to the external output

}

Figure 33. Double Buffering Pseudo-Code

Read first block to internal PING input buffer.

while (data left to process)
{
Start reading next block to internal ”PONG” input buffer.

Wait for PING input and output buffers to be ready.

Process internal PING input buffer, placing result in internal

PING output buffer.

Swap PING and PONG buffer pointers.

Start writing out PONG output buffer.

}

72 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

Figure 34. Double-Buffering Time Sequence

Before while loop: initial setup

1. Initial read InBuff 0

(PING)
EDMA writes to PING
InBuff 1
(PONG)

OutBuff 0
(PING)

OutBuff 1
(PONG)

First while loop iteration: first half

InBuff 0 3. First process

(PING)
CPU reads from PING
2. Second read InBuff 1
(PONG)
EDMA writes to PONG

OutBuff 0 4. First process

(PING)
CPU writes to PING
OutBuff 1
(PONG)

First while loop iteration: second half, PING and PONG switched

InBuff 0
(PONG)

InBuff 1
(PING)

5. First write OutBuff 0

(PONG)
EDMA reads from PONG
OutBuff 1
(PING)

SPRU610C TMS320C64x Two-Level Internal Memory 73

Memory System Policies

Figure 34. Double-Buffering Time Sequence (Continued)

Second while loop iteration

6. Third read InBuff 0

EDMA writes to PONG (PONG)

InBuff 1 7. Second process

(PING) CPU reads from PING

OutBuff 0
(PONG)

9. Second write† OutBuff 1 8. Second process

EDMA reads from PONG (PING) CPU writes to PING

† PING and PONG labels switch before step 9.

Figure 35. Double Buffering as a Pipelined Process

Iteration Time

Iter 1 RD PROC WR

Iter 2 RD PROC WR

Iter 3 RD PROC WR

Iter 4 RD PROC WR

Iter 5 RD PROC WR

Iter 6 RD PROC WR

74 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

The memory system’s coherence directly supports this programming

technique. A system of snoops and invalidates (as discussed in section 5.4.2)
keeps L2 SRAM and L1D in sync without programmer intervention. In the
context of this example, the memory system performs the following operations
in response to each of the numbered steps in Figure 34.

1) The EDMA writes to InBuff 0. As the EDMA writes to InBuff 0, the L2

controller sends snoop-invalidates for the corresponding lines in L1D, if it
detects that L1D may be caching these addresses. Updated data written
back from L1D is merged with the data written by the EDMA, so that L2
contains the most up-to-date data. The corresponding lines are
invalidated in L1D.

2) This step is similar to step 1, EDMA writes to InBuff 1.

3) The CPU reads InBuff 0. These lines were snoop-invalidated from L1D in
step 1. Therefore, these accesses miss L1D, forcing L1D to read the new
data from L2 SRAM.

4) The CPU writes to OutBuff 0. In practice, this step may overlap or

interleave with step 3. Suppose initially OutBuff 0 is present in L1D. In this
case, the writes hit L1D and mark the lines as dirty. If OutBuff 0 were not
present in L1D, the writes would go directly to L2 SRAM.

5) The EDMA reads OutBuff 0. The EDMA reads in L2 trigger a snoop for
each cache line held in L1D. Any dirty data for the line is written to L2
before the EDMA’s read is processed. Thus, the EDMA sees the most
up-to-date data. The line is marked clean and is left valid. (On
C621x/C671x devices, the line is subsequently invalidated in L1D.)

6) The EDMA writes to InBuff 0. This step proceeds identically to step 1. That
is, InBuff 0 is snoop-invalidated from L1D as needed. The EDMA writes
are processed after any dirty data is written back to L2 SRAM.

7) The CPU reads InBuff 1. These lines were snoop-invalidated in step 6.

Therefore, these accesses miss L1D, forcing L1D to read the new data
from L2 SRAM.

8) The CPU writes to OutBuff 1. This step proceeds identically to step 4.

9) The EDMA reads OutBuff 1. This step proceeds identically to step 5.

Notice that the system of snoops and snoop-invalidates automatically keeps

the EDMA and CPU synchronized for the input and output buffers. Further,
double-buffering allows EDMA and CPU accesses to occur in parallel.

SPRU610C TMS320C64x Two-Level Internal Memory 75

Memory System Policies

8.3 Memory Access Ordering

8.3.1 Program Order of Memory Accesses

To optimize throughput, power, and ease of programming, the C6000 DSP

architecture supports a combination of strongly ordered and relaxed ordered
memory models. These terms are defined relative to the order of memory
operations implied by a particular program sequence. This ordering is referred
to as the program order of memory accesses.

The C6000 DSP cores may initiate up to two parallel memory operations per
cycle. The program order of memory accesses defines the outcome of
memory accesses in terms of a hypothetical serial implementation of the
architecture. That is, it describes the order that the parallel memory operations
are processed such that time-sequence terms such as earlier and later are
used precisely with respect to a particular sequence of operations.

Program order is defined with respect to instructions within an execute packet

and with respect to a sequence of execute packets. Memory operations
(including those that are issued in parallel) are described as being earlier or
later with respect to each other. The terms earlier and later are strictly
opposites: if X is not earlier than Y, then X is later than Y.

Memory accesses initiated from different execute packets have the same
temporal ordering as the execute packets themselves. That is, in the defined
program order, memory operations issued on cycle i are always earlier than
memory accesses issued on cycle i + 1, and are always later than those issued
on cycle i - 1.

For accesses issued in parallel, the type of operations (reads or writes), and
the data address ports that execute the operations determine the ordering.
Table 32 describes the ordering rules.

Table 32. Program Order for Memory Operations Issued From a Single Execute Packet

Data Address 1 (DA1) Data Address 2 (DA2)

Operation Operation Program Order of Accesses
Load Load DA1 is earlier than DA2

Load Store DA1 is earlier than DA2

Store Load DA2 is earlier than DA1

Store Store DA1 is earlier than DA2

76 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

The data address port for a load or store instruction is determined by the
datapath that provides the data (as opposed to the address) for the memory
operation. Load and store instructions that operate on data in the A datapath
use DA1. Load and store instructions that operate on data in the B datapath
use DA2. Note that the datapath that provides the data to be operated on
determines whether DA1 or DA2 is used. The datapath that provides the
address for the access is irrelevant.

The C64x DSP supports nonaligned memory accesses to memory using the
LDNW, STNW, LDNDW, and STNDW instructions. The memory system does
not assure that these memory accesses will be atomic. Rather, it may divide
the accesses for these instructions into multiple operations. The program
order of memory accesses does not define the order of the individual memory
operations that comprise a single nonaligned access. The program order only
defines how the entire nonaligned access is ordered relative to earlier and later
accesses. So, although the complete nonaligned access does follow the
program order defined above with respect to the CPU itself, other requestors
may see the nonaligned memory access occur in pieces.

The previous definition describes the memory system semantics. The memory
system assures that the semantics of the program order of memory accesses
will be retained for CPU accesses relative to themselves. The memory system
may, however, relax the ordering of operations as they are executed within the
memory hierarchy so long as the correct semantics are retained. It may also
allow other requestors to the memory system to see the accesses occur in an
order other than the original program order. Section 8.3.2 describes this in
detail.

8.3.2 Strong and Relaxed Memory Ordering

The program order of memory accesses (as described in section 8.3.1)

describes the desired semantics of a particular sequence of memory
accesses. In most circumstances, it is not necessary for the memory system
to execute the memory accesses in this exact order at all levels of memory
hierarchy in order to retain these semantics. Indeed, for performance and
power reasons, it is extremely advantageous for the memory hierarchy to relax
the order of accesses.

When communicating with peripherals and when coordinating with other

devices (the EDMA, other CPUs) that are accessing the same memory,
retaining program order potentially becomes more important. Given these
diverse needs, the memory hierarchy supports both strongly ordered and
relaxed orderings for memory operations.

SPRU610C TMS320C64x Two-Level Internal Memory 77

Memory System Policies

Memory system coherence implies that writes to a single memory location are
serialized for all requestors, and that all requestors see the same sequence
of writes to that location. Coherence does not make any implications about the
ordering of accesses to different locations, or the ordering of reads with
respect to other reads of the same location. Rather, the memory system
ordering rules (strong or relaxed) describe the ordering assurances applied to
accesses to different locations.

A sequence of memory operations are said to be strongly ordered if it is not

possible to observe a different sequence of memory operations from some
place in the system. For instance, if one requestor writes to location X earlier
than it writes to a different location Y, a second requestor must not see Y
updated before it sees X updated for the accesses to be considered strongly
ordered. In this example, regardless of the order in which the memory system
processes writes from the first requestor and reads from the second requestor,
the second requestor must not be able to observe the write to Y occurring
before the write to X. Strong ordering does not require the writes to be
serialized, however. The writes may occur in parallel or even in opposite order
so long as it is not possible for the second CPU to observe Y being updated
before X.

Nonaligned accesses (those issued by the LDNW, STNW, LDNDW, and

STNDW instructions) may be strongly ordered relative to other accesses.
Because nonaligned accesses are not assured to be atomic, strong ordering
assurances apply only to the nonaligned access as a whole. No ordering
assurances are made between the individual operations that comprise a
single nonaligned access, should the memory system divide the nonaligned
access into multiple memory operations.

The memory hierarchy provides strong ordering for all noncacheable

long-distance accesses. Such accesses are typically to peripherals and
external memory for which the corresponding MAR cache enable bit is not set.

For cacheable locations in external memory, the memory hierarchy provides

only a relaxed ordering. Thus, it is possible for other requestors to observe
updates occurring in memory in a different order than the original program
order of memory accesses. Despite this relaxed ordering, the CPU still sees
the desired memory system semantics as described in section 8.3.1.

For cacheable locations in L2 SRAM, the C64x device provides strong

ordering for CPU accesses to locations that are within the same L1D cache
line, and for locations that are not present in L1D. (On the C621x/C671x
devices, this means that data accesses to addresses whose upper 27 bits are
equal are strongly ordered.) On the C64x device, this means that data
accesses to addresses whose upper 26 bits are equal are strongly ordered.

78 TMS320C64x Two-Level Internal Memory SPRU610C

 Memory System Policies

For locations in L2 SRAM that are not within the same cache line, strong
ordering is provided only on writes and only as long as addresses involved are
not present in L1D. This can be ensured by using the L1DWIBAR and
L1DWIWC control registers described in section 7.3.2. In all other cases, a
relaxed ordering is provided for CPU accesses to L2 SRAM.

The L2 provides limited ordering assurances for EDMA access to L2 SRAM.

EDMA reads to L2 are not reordered relative to other EDMA reads. EDMA
writes to L2 are not reordered relative to other EDMA writes. Reads and writes,
however, may be reordered relative to each other.

SPRU610C TMS320C64x Two-Level Internal Memory 79

 This page is intentionally left blank.

80 TMS320C64x Two-Level Internal Memory SPRU610C

 Revision History

Table 33 lists the changes made since the previous version of this document.

Table 33. Document Revision History

Page Additions/Modifications/Deletions
35 Changed the text, section 7, Memory System Controls, to section 6, Registers, in section 5.4.

72 Changed Figure 33.

SPRU610C TMS320C64x Two-Level Internal Memory 81

 This page is intentionally left blank.

82 TMS320C64x Two-Level Internal Memory SPRU610C

 Index

Index

B G
block cache operation base address register global cache operations 62
(BAR) 63
block cache operation word count register (WC) 63 H
block cache operations 63
HPI and PCI access to memory subsystem 37
block diagram
C64x DSP 9
two-level internal memory 12 I
ID bit 39
IP bit 39
C
C bit L
in L2WB 49
in L2WBINV 50 L1D invalidate base address register (L1DIBAR) 48
L1D invalidate word count register (L1DIWC) 48
cache configuration register (CCFG) 39
L1D memory banking 20
cacheability controls 56
L1D miss penalty 22
CCFG 39
L1D miss pipelining 24
CE bit 51 L1D parameters 19
L1D performance 20
L1D write buffer 23
D L1D writeback−invalidate base address register
(L1DWIBAR) 47
DCC field in CSR 52
L1D writeback−invalidate word count register
(L1DWIWC) 47
L1D/L1P-to-L2 request servicing 35
E L1DIBAR 48
L1DIWC 48
EDMA access into L2 control 68
L1DWIBAR 47
EDMA access to cache controls 37
L1DWIWC 47
EDMA coherence in L2 SRAM example 71 L1P invalidate base address register (L1PIBAR) 46
EDMA-to-L2 request servicing 36 L1P invalidate word count register (L1PIWC) 46
EDMAWEIGHT 41 L1P miss penalty 27
effect of L2 commands on L1 caches 65 L1P miss pipelining 28

SPRU610C TMS320C64x Two-Level Internal Memory 83

Index

L1P parameters 27 level 1 program cache (L1P) 27

L1P performance 27 mode selection using PCC field in CSR 53
L1PIBAR 46 parameters 27
performance 27
L1PIWC 46
level 2 unified memory (L2) 30
L2 allocation registers (L2ALLOC0−L2ALLOC3) 42
bank structure 34
L2 bank structure 34 cache and SRAM 30
L2 cache and L2 SRAM 30 interfaces 35
L2 EDMA access control register mode selection using L2MODE field in
(EDMAWEIGHT) 41 CCFG 53
L2 interfaces 35 operation 32
L2 invalidate base address register (L2IBAR) 45
L2 invalidate word count register (L2IWC) 45
L2 memory attribute registers
M
(MAR0−MAR255) 51 MAR 51
address ranges 57 memory access ordering 76
L2 operation 32 program order of memory accesses 76
L2 request servicing using EDMA 36 strong and relaxed memory ordering 77
L2 writeback all register (L2WB) 49 memory hierarchy overview 9
L2 writeback base address register (L2WBAR) 43 memory system coherence 69
L2 writeback word count register (L2WWC) 43 memory system control 52
L2 writeback−invalidate all register (L2WBINV) 50 cache mode selection 52
L2 writeback−invalidate base address register cacheability controls 56
(L2WIBAR) 44 EDMA access into L2 control 68
L2 writeback−invalidate word count register L2-to-EDMA request control 67
(L2WIWC) 44 program-initiated cache operations 60
L2-to-EDMA request control 67 memory system policies 69
coherence 69
L2ALLOC0−L2ALLOC3 42
EDMA coherence in L2 SRAM example 71
L2IBAR 45 memory access ordering 76
L2IWC 45
L2MODE bits 39
L2WB 49 N
L2WBAR 43
notational conventions 3
L2WBINV 50
L2WIBAR 44
L2WIWC 44 P
L2WWC 43
P bits 39
level 1 data cache (L1D) 19
mode selection using DCC field in CSR 52 PCC field in CSR 53
parameters 19 program order of memory accesses 76
performance 20 program-initiated cache operations 60

84 TMS320C64x Two-Level Internal Memory SPRU610C

 Index

L2 writeback all register (L2WB) 49

R L2 writeback base address register
registers 38 (L2WBAR) 43
cache configuration register (CCFG) 39 L2 writeback word count register (L2WWC) 43
L1D invalidate base address register L2 writeback−invalidate all register
(L1DIBAR) 48 (L2WBINV) 50
L1D invalidate word count register L2 writeback−invalidate base address register
(L1DIWC) 48 (L2WIBAR) 44
L1D writeback−invalidate base address register L2 writeback−invalidate word count register
(L1DWIBAR) 47 (L2WIWC) 44
L1D writeback−invalidate word count register related documentation from Texas Instruments 3
(L1DWIWC) 47 revision history 81
L1P invalidate base address register
(L1PIBAR) 46
L1P invalidate word count register (L1PIWC) 46
L2 allocation registers
S
(L2ALLOC0−L2ALLOC3) 42
strong and relaxed memory ordering 77
L2 EDMA access control register
(EDMAWEIGHT) 41
L2 invalidate base address register
(L2IBAR) 45 T
L2 invalidate word count register (L2IWC) 45
L2 memory attribute registers terms and definitions 13
(MAR0−MAR255) 51 trademarks 4

SPRU610C TMS320C64x Two-Level Internal Memory 85