DSP

DSP Solutions Advanced Design Guide
Edition 1.0
DSP: Designing for Optimal Results

High-Performance DSP Using Virtex-4 FPGAs
DSP PRODUCTS: ADVANCED DESIGN GUIDE
EDITION 1.0

Xilinx i
DSP: DESIGNING FOR OPTIMAL RESULTS
2005 Xilinx, Inc. All rights reserved. XILINX, the Xilinx Logo, and other designated brands included herein are trademarks of Xilinx, Inc. PowerPC is a trademark of IBM, Inc. All other trademarks are the property of their respective owners. NOTICE OF DISCLAIMER: The information stated in this book is not to be used for design purposes. Xilinx is providing this design, code, or information "as is." By providing the design, code, or information as one possible implementation of this feature, application, or standard, Xilinx makes no representation that this implementation is free from any claims of infringement. You are responsible for obtaining any rights you may require for your implementation. Xilinx expressly disclaims any warranty whatsoever with respect to the adequacy of the implementation, including but not limited to any warranties or representations that this implementation is free from claims of infringement and any implied warranties of merchantability or fitness for a particular purpose. All terms mentioned in this book are known to be trademarks or service marks and are the property of their respective owners. Use of a term in this book should not be regarded as affecting the validity of any trademark or service mark. All rights reserved. No part of this book may be reproduced, in any form or by any means, without written permission from the publisher.
Edition 1.0 March 2005
ii Xilinx
Acknowledgements
Chapter 2 through Appendix A have been sourced from the Xilinx Advanced Product Divisions (APD) "XtremeDSP Design Considerations User Guide". For up-to-date information, download the online version located here: http://www.xilinx.com/bvdocs/userguides/ug073.pdf For more information, contact: Gregg C. Hawkes Senior Staff Applications Manager, Xilinx, Inc.
Xilinx iii
iv Xilinx
TABLE OF CONTENTS
Chapter 1: Digital Signal Processing Design Challenges

The Performance Gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 The Ideal Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
XtremeDSP Slice Delivers Maximum Performance, Minimum Power, and Best Economy . 2 Simplicity and Efficiency of the Cascade Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Extremely Low Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Increased Flexibility for Cost Effectiveness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Easy to Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Virtex-4 FPGAs A Platform for Every Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Reduce Time-to-Market with World-Class Xilinx Support . . . . . . . . . . . . . . . . . 3 A Must-Read . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Chapter 2: XtremeDSP Design Considerations

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Architecture Highlights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Number of DSP48 Slices Per Virtex-4 Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
DSP48 Slice Primitive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 DSP48 Slice Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Attributes in VHDL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Attributes in Verilog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 DSP48 Tile and Interconnect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Simplified DSP48 Slice Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 Timing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A, B, C, and P Port Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 OPMODE, SUBTRACT, and CARRYINSEL Port Logic . . . . . . . . . . . . . . . . . . . . . . . . 20 Twos Complement Multiplier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 X, Y, and Z Multiplexer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Three-Input Adder/Subtracter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Carry Input Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Symmetric Rounding Supported by Carry Logic . . . . . . . . . . . . . . . . . . . . . . . . . . 26 Forming Larger Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Basic FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Multi-Channel FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Creating FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Adder Cascade vs. Adder Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 DSP48 Slice Functional Use Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Single Slice, Multi-Cycle, Functional Use Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Single Slice, 35 x 18 Multiplier Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Single Slice, 35 x 35 Multiplier Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Fully Pipelined Functional Use Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Fully Pipelined, 35 x 18 Multiplier Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Fully Pipelined, 35 x 35 Multiplier Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 Xilinx v

Fully Pipelined, Complex, 18 x 18 Multiplier Use Model . . . . . . . . . . . . . . . . . . . . . . . . 41 Fully Pipelined, Complex, 18 x 18 MAC Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Fully Pipelined, Complex, 35 x 18 Multiplier Usage Model . . . . . . . . . . . . . . . . . . . . . . . 46
Miscellaneous Functional Use Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic, 18-bit Circular Barrel Shifter Use Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . VHDL and Verilog Instantiation Templates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . VHDL Instantiation Template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Verilog Instantiation Template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47 48 50 50 51
Chapter 3: DSP48 Slice Math Functions

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Basic Math Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Add/Subtract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Accumulate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multiply Accumulate (MAC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multiplexer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Barrel Shifter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Counter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multiply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Divide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dividing with Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dividing with Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Square Root . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Square Root of the Sum of Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 54 55 55 55 55 56 56 56 58 59 61 62
Chapter 4: MAC FIR Filters

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 Single-Multiplier MAC FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Bit Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Generic Saturation Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Coefficient Specific Saturation Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Control Logic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Embedding the Control Logic into the Block RAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rounding without an Extra Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Using Distributed RAM for Data and Coefficient Buffers . . . . . . . . . . . . . . . . . . . . . . . Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Symmetric MAC FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dual-Multiplier MAC FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 65 65 65 67 69 70 70 72 72 73 74
Chapter 5: Parallel FIR Filters

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
vi Xilinx
TABLE OF CONTENTS
Parallel FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Transposed FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Advantages and Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 Resource Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 Systolic FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Advantages and Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Resource Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Symmetric Systolic FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Resource Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Rounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Chapter 6: Semi-Parallel FIR Filters

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Semi-Parallel FIR Filter Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Four-Multiplier, Distributed-RAM-Based, Semi-Parallel FIR Filter . . . . . . . . 87
Data Memory Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 Coefficient Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 Control Logic and Address Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Resource Utilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 Three-Multiplier, Block RAM-Based, Semi-Parallel FIR Filter . . . . . . . . . . . . . 92 Other Semi-Parallel FIR Filter Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 Semi-Parallel, Transposed, Four-Multiplier FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . 94 Advantages and Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 Rounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Chapter 7: Multi-Channel FIR Filters

Multi-Channel FIR Implementation Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Top Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 DSP48 Tile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 Combining Separate Input Streams into an Interleaved Stream . . . . . . . . . . . 101 Coefficient RAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 Control Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 Implementation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Appendix A: References
Xilinx vii
viii Xilinx
Chapter 1
Digital Signal Processing Design Challenges
Our insatiable hunger for electronic gadgets that provide high-quality audio, video, data or all three, is spiraling up the processing power that is needed to process these signals. Digital signal processing (DSP) systems, within both infrastructure and customer premise equipment must provide increasing levels of performance and flexibility to handle the new requirements yet provide greater scalability for achieving higher economies of scale.
The Performance Gap

Algorithmic complexity increases as application demands increase. Figure 1-1 illustrates performance demands over time. In order to process these new algorithms, higher-performance signal processing engines are required. Typical fixed architecture DSP processors cannot keep pace on their own. A DSP co-processor is often needed to handle the highest performance portions of these ever increasingly complex algorithms. The performance gap in Figure 1-1 illustrates this expanding co-processing requirement.
4G
Performance (Algorithm and Processor Forecast)
Performance Gap
Alg
th ori
Co
it lex mp
SDR 3G Imaging Radar SD/HD Video
DSP/GPP*
Traditional Processor Architectures
1960
1970
1980
1990
2000
2010
*Source: Jan Rabaey, BWRC
Figure 1-1:
The Performance Gap Xilinx 1
DSP: DESIGNING FOR OPTIMAL RESULTS Field Programmable Gate Arrays (FPGAs) are very well suited to fill the performance gap for a variety of reasons: They offer extremely high-performance signal processing capability through parallelism. They provide very low risk due to the flexible architecture. They allow design migration to handle changing standards. Developers can use them to create a customized and differentiated solution. They are quickly coming down in price. In fact, it is possible to find FPGAs for less then $2 per device. They provide very low power per function.
The Ideal Solution

With the revolutionary XtremeDSP Slices, Xilinx Virtex-4 FPGAs deliver the ideal solution for high-performance digital signal processing. They satisfy high-performance signal processing tasks traditionally serviced by an ASIC or ASSP. They allow you to create high-performance DSP engines that can boost the signal processing performance of your system for a host of applications including digital communications and video/imaging. And they are the ideal choice to increase system level performance by complementing a programmable DSP system as either a pre- or co-processor.
XtremeDSP Slice Delivers Maximum Performance, Minimum Power, and Best Economy
The XtremeDSP Sliceoperating at a blazing 500 MHzlies at the heart of Virtex-4 FPGAs XtremeDSP performance. As the most powerful addition to the Xilinx XtremeDSP took kit, it is a unique piece of hard coded IP embedded in each Virtex-4 device. It provides industry-leading DSP processing performance, unrivalled economy, and the lowest power consumption of any device in this performance range.
Simplicity and Efficiency of the Cascade Logic

The built-in cascade logic of the XtremeDSP Slice allows multiple slices to be connected together to implement complex filters and multi-precision functions while operating at full speed. And the cascade logic provides tremendous cost advantage. Other solutions require additional FPGA resources to build costly and inefficient adder trees to implement this common function. They require a much larger FPGA to implement the same level of functionality that can be attained in an XtremeDSPenabled Virtex-4 FPGA. The result is a tremendous performance and cost advantage of the Virtex-4 device over other FPGA DSP solutions.
Extremely Low Power Consumption

Each XtremeDSP Slice consumes only 2.3 mW/100 MHz in a typical system implementation. This extremely low power is enabled by the optimized hard implementation of the XtremeDSP Slice. Also, the programmable logic fabric of the Virtex-4 family has a significant power advantage. For example, power-per-CLB has been cut in half, with static power reduced by 40% and dynamic power reduced by 50%. In addition, certain hard-logic silicon functions in the Virtex-4 FPGA reduce consumption by approximately 90%. This results in a lower power budget and all its associated benefitshigher reliability, smaller power devices, smaller fans, and so on.
2 Xilinx
DIGITAL SIGNAL PROCESSING DESIGN CHALLENGES
Increased Flexibility for Cost Effectiveness

The Virtex-4 FPGA flexibility boosts cost effectiveness for all application designs. For example, Virtex-4 FPGAs enable you to buy a customer device that supports two applicationsone for diagnostic testing and one for the application. Here, Virtex-4 FPGAs can be tested for two designs or two variations of the same design. Savings are realized right down the line, from inventory costs, to design costs, to system costs, to consumer costs.
Easy to Use
Xilinx and its partners provide the easiest-to-use design solutions for FPGA-based DSP solutions with features such as: System Generator for DSP reduces design time. A rich DSP IP library implements fast, highly optimized algorithms. Award-winning technical support and DSP services enable you to bring products to market much faster. Whether you are working with spread-spectrum, multi-carrier, or narrowband communication systems, Virtex-4 FPGAs are the ideal choice for ease of use.
Virtex-4 FPGAs A Platform for Every Application

All Virtex-4 platforms offer XtremeDSP capabilities. Choose the device that provides the optimal DSP performance for your application: Virtex-4 SX devices offer the most cost-effective implementation of ultra-high performance DSP functionality for high-end DSP applications. They provide the highest ratio of XtremeDSP slicesup to 512and deliver up to 256 GMACS (18x18-bit multiply, 48-bit addition/accumulation) performance. Virtex-4 LX devices offer ample XtremeDSP slices and include more logic, memory, and I/O resources for logic applications. Virtex-4 FX devices include embedded PowerPC processors and RocketIO multigigabit transceivers for embedded processing and high-speed serial applications. XtremeDSP platform solutions accelerate your products time-to-market through superior design, design tools, intellectual property cores, and design services. They provide the fastest means of designing, verifying, and deploying your DSP algorithms and systems in FPGAs.
Reduce Time-to-Market with World-Class Xilinx Support

Xilinx supplies a host of support functions to designers including DSP training courses, award winning technical support, technical data, implementation data, and design consulting.
A Must-Read
This book is a must-read for DSP designers who want to tap the power of the Virtex-4 XtremeDSP Slice. It provides a detailed description of the multiple features of the slice as well as providing multiple examples that show you how to harness the power and flexibility of this powerful IP block. Tap into the XtremeDSP Slice and reap the rewards of highest performance, lowest power at the lowest cost.
Xilinx 3
4 Xilinx
Chapter 2
XtremeDSP Design Considerations
This chapter provides technical details for the XtremeDSP Digital Signal Processing (DSP) element, the DSP48 slice. The DSP48 slice is a new element in the Xilinx development model referred to as Application Specific Modular Blocks (ASMBL). The purpose of this model is to deliver off-the-shelf programmable devices with the best mix of logic, memory, I/O, processors, clock management, and digital signal processing. ASMBL is an efficient FPGA development model for delivering off-theshelf, flexible solutions ideally suited to different application domains. Each XtremeDSP tile contains two DSP48 slices to form the basis of a versatile coarse-grain DSP architecture. Many DSP designs follow a multiply with addition. In Virtex-4 devices these elements are supported in dedicated circuits. The DSP48 slices support many independent functions, including multiplier, multiplieraccumulator (MAC), multiplier followed by an adder, three-input adder, barrel shifter, wide bus multiplexers, magnitude comparator, or wide counter. The architecture also supports connecting multiple DSP48 slices to form wide math functions, DSP filters, and complex arithmetic without the use of general FPGA fabric. The DSP48 slices available in all Virtex-4 family members support new DSP algorithms and higher levels of DSP integration than previously available in FPGAs. Minimal use of general FPGA fabric leads to low power, very high performance, and efficient silicon utilization.
Introduction
The DSP48 slices facilitate higher levels of DSP integration than previously possible in FPGAs. Many DSP algorithms are supported with minimal use of the general-purpose FPGA fabric, resulting in low power, high performance, and efficient device utilization. At first look, the DSP48 slice is an 18 x 18 bit twos complement multiplier followed by a 48-bit sign-extended adder/subtracter/accumulator, a function that is widely used in digital signal processing (DSP). A second look reveals many subtle features that enhance the usefulness, versatility, and speed of this arithmetic building block. Programmable pipelining of input operands, intermediate products, and accumulator outputs enhances throughput. The 48-bit internal bus allows for practically unlimited aggregation of DSP slices. Xilinx 5
DSP: DESIGNING FOR OPTIMAL RESULTS One of the most important features is the ability to cascade a result from one XtremeDSP Slice to the next without the use of general fabric routing. This path provides high-performance and lowpower post addition for many DSP filter functions of any tap length. For multi-precision arithmetic this path supports a right-wire-shift. Thus a partial product from one XtremeDSP Slice can be right-justified and added to the next partial product computed in an adjacent such slice. Using this technique, the XtremeDSP Slices can be configured to support any size operands. Another key feature for filter composition is the ability to cascade an input stream from slice to slice. The C input port, allows the formation of many 3-input mathematical functions, such as 3-input addition, 2-input multiplication with a single addition. One subset of this function is the very valuable support of rounding a multiplication away from zero.
Architecture Highlights
The Virtex-4 DSP slices are organized as vertical DSP columns. Within the DSP column, two vertical DSP slices are combined with extra logic and routing to form a DSP tile. The DSP tile is four CLBs tall. Each DSP48 slice has a two-input multiplier followed by multiplexers and a three-input adder/subtracter. The multiplier accepts two 18-bit, two's complement operands producing a 36-bit, two's complement result. The result is sign extended to 48 bits and can optionally be fed to the adder/subtracter. The adder/subtracter accepts three 48-bit, two's complement operands, and produces a 48-bit two's complement result. Higher level DSP functions are supported by cascading individual DSP48 slices in a DSP48 column. One input (cascade B input bus) and the DSP48 slice output (cascade P output bus) provide the cascade capability. For example, a Finite Impulse Response (FIR) filter design can use the cascading input to arrange a series of input data samples and the cascading output to arrange a series of partial output results. For details on this technique, refer to the section titled Adder Cascade vs. Adder Tree, page 31. Architecture highlights of the DSP48 slices are: 18-bit by 18-bit, two's-complement multiplier with a full-precision 36-bit result, sign extended to 48 bits Three-input, flexible 48-bit adder/subtracter with optional registered accumulation feedback Dynamic user-controlled operating modes to adapt DSP48 slice functions from clock cycle to clock cycle Cascading 18-bit B bus, supporting input sample propagation Cascading 48-bit P bus, supporting output propagation of partial results Multi-precision multiplier and arithmetic support with 17-bit operand right shift to align wide multiplier partial products (parallel or sequential multiplication) Symmetric intelligent rounding support for greater computational accuracy Performance enhancing pipeline options for control and data signals are selectable by configuration bits Input port C typically used for multiply-add operation, large three-operand addition, or flexible rounding mode Separate reset and clock enable for control and data registers 6 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS I/O registers, ensuring maximum clock performance and highest possible sample rates with no area cost OPMODE multiplexers A number of software tools support the DSP48 slice. The Xilinx ISE software supports DSP48 slice instantiations. The Architecture Wizard is a GUI for creating instantiation VHDL and/or Verilog code. It also helps generate code for designs using a single DSP48 slice (i.e., Multiplier, Adder, Multiply-Accumulate or MAC, and Dynamic Control modes). Using the Architecture Wizard, CORE Generator tool, or System Generator, a designer can quickly generate math or other functions using Virtex-4 DSP48 slices.
Number of DSP48 Slices Per Virtex-4 Device

Table 2-1 shows the number of DSP48 slices for each device in the Virtex-4 families. The Virtex-4 SX family offers the highest ratio of DSP48 slices to logic, making it ideal for math-intensive applications. Table 2-1: Number of DSP48 Slices per Family Member DSP48 32 48 64 64 80 96 96 96 Device XC4VSX25 XC4VSX35 XC4VSX55 DSP48 128 192 512 Device XC4VFX12 XC4VFX20 XC4VFX40 XC4VFX60 XC4VFX100 XC4VFX140 DSP48 32 32 48 128 160 192
Device XC4VLX15 XC4VLX25 XC4VLX40 XC4VLX60 XC4VLX80 XC4VLX100 XC4VLX160 XC4VLX200
Xilinx 7
DSP48 Slice Primitive

Figure 2-1 shows the DSP48 slice primitive.
A[17:0] B[17:0] C[47:0] OPMODE[6:0] SUBTRACT CARRYIN 2 CARRYINSEL[1:0] CEA CEB CEC CEM CEP CECTRL CECINSUB CECARRYIN RSTA RSTB RSTC RSTM RSTP RSTCTRL RSTCARRYIN CLK 18 48 BCIN[17:0] PCIN[47:0]
ug073_c1_01_060304
18 18 48 7
BCOUT[17:0] PCOUT[47:0] P[47:0]
18 48 48
Figure 2-1:
DSP48 Slice Primitive
Table 2-2 lists the available ports in the DSP48 slice primitive. Table 2-2: A B C 8 Xilinx DSP48 Slice Port List and Definitions Direction Size I 18 I I 18 48 Function The multiplier's A input. This signal can also be used as the adder's Most Significant Word (MSW) input The multiplier's B input. This signal can also be used as the adder's Least Significant Word (LSW) input The adder's C input
Signal Name
XTREMEDSP DESIGN CONSIDERATIONS Table 2-2: DSP48 Slice Port List and Definitions (Continued) Direction Size I 7 I I I I I I I I I I I I I I I I I I I I I O O O 1 1 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 18 48 18 48 48 Function Controls the input to the X, Y, and Z multiplexers in the DSP48 slices (see OPMODE, Table 2-7) 0 = add, 1 = subtract The carry input to the carry select logic Selects carry source (see CARRYINSEL, Table 2-8) Clock enable: 0 = hold, 1 = enable AREG Clock enable: 0 = hold, 1 = enable BREG Clock enable: 0 = hold, 1 = enable CREG Clock enable: 0 = hold, 1 = enable MREG Clock enable: 0 = hold, 1 = enable PREG Clock enable: 0 = hold, 1 = enable OPMODEREG, CARRYINSELREG Clock enable: 0 = hold, 1 = enable SUBTRACTREG and general interconnect carry input Clock enable: 0 = hold, 1 = enable (carry input from internal paths) Reset: 0 = no reset, 1 = reset AREG Reset: 0 = no reset, 1 = reset BREG Reset: 0 = no reset, 1 = reset CREG Reset: 0 = no reset, 1 = reset MREG Reset: 0 = no reset, 1 = reset PREG Reset: 0 = no reset, 1 = reset SUBTRACTREG, OPMODEREG, CARRYINSELREG Reset: 0 = no reset, 1 = reset (carry input from general interconnect and internal paths) The DSP48 clock The multiplier's cascaded B input. This signal can also be used as the adder's LSW input Cascaded adder's Z input from the previous DSP slice The B cascade output The P cascade output The product output
Signal Name OPMODE SUBTRACT CARRYIN CARRYINSEL CEA CEB CEC CEM CEP CECTRL CECINSUB CECARRYIN RSTA RSTB RSTC RSTM RSTP RSTCTRL RSTCARRYIN CLK BCIN PCIN BCOUT PCOUT P
Xilinx 9
DSP48 Slice Attributes

The synthesis attributes for the DSP48 slice are described in detail throughout this section. With the exception of the B_INPUT and LEGACY_MODE attributes, all other attributes call out pipeline registers in the control and datapaths. The value of the attribute sets the number of pipeline registers. The attribute settings are as follows: The AREG and BREG attributes can take a value of 0, 1, or 2. The values define the number of pipeline registers in the A and B input paths. See the A, B, C, and P Port Logic section for more information. The CREG, MREG, and PREG attributes can take a value of 0 or 1. The value defines the number of pipeline registers at the output of the multiplier (MREG) (shown in Figure 2-11) and at the output of the adder (PREG) (shown in Figure 2-9). The CREG attribute is used to select the pipeline register at the 'C' input (shown in Figure 2-8). The CARRYINREG, CARRYINSELREG, OPMODEREG, and SUBTRACTREG attributes take a value of 0 if there is no pipelining register on these paths, and take a value of 1 if there is one pipeline register in their path. The CARRYINSELREG, OPMODEREG, and SUBTRACTREG paths are shown in Figure 2-10, and the CARRYINREG path is shown in Figure 2-12. The B_INPUT attribute defines whether the input to the B port is routed from the parallel input (attribute: DIRECT) or the cascaded input from the previous slice (attribute: CASCADE). The LEGACY_MODE attribute serves two purposes. The first purpose is similar in nature to the MREG attribute. It defines whether or not the multiplier is "flow through" in nature (i.e., LEGACY_MODE value equal to MULT18x18) or contains a single pipeline register in the middle of the multiplier (i.e., LEGACY_MODE value equal to MULT18x18S is the same as MREG value equal to one). While this is redundant to the MREG attribute, it was deemed useful for customers used to the Virtex-II and Virtex-II Pro multipliers since the DSP48 setup and hold timing most closely matches those of the Virtex-II and Virtex-II Pro MULT18x18S when the MREG is used. Any disagreement between the MREG attribute and LEGACY_MODE attribute settings are flagged as a software Design Rule Check (DRC) error. The second purpose for the attribute is to convey to the timing tools whether the A and B port through the combinatorial multiplier path (slower timing) or faster X multiplexer bypass path for A:B should be used in the timing calculations. Since the OPMODE can change dynamically, the timing tools cannot determine this without an attribute. To summarize the timing tools behavior: If (attribute: NONE), then timing analysis/simulation bypasses the multiplier for the highest performance. The lowest power dissipation is achieved by setting MREG to one while CEM input is grounded. If (attribute: MULT18x18), then timing analysis/simulation uses the combinatorial path through the multiplier. In this case, MREG must be set to zero or a DRC error occurs. If (attribute: MULT18x18S), then timing analysis/simulation uses a pipelined multiplier. In this case MREG must be set to one or a DRC error occurs.
10 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS
Attributes in VHDL
DSP48 generic map ( AREG => 1,-- Number of pipeline registers on the A input, 0, 1 or 2 BREG => 1,-- Number of pipeline registers on the B input, 0, 1 or 2 B_INPUT => DIRECT, -- B input DIRECT from fabric or CASCADE from -- another DSP48 CARRYINREG => 1, -- Number of pipeline registers for the CARRYIN -- input, 0 or 1 CARRYINSELREG => 1, -- Number of pipeline registers for the -- CARRYINSEL, 0 or 1 CREG => 1, -- Number of pipeline registers on the C input, 0 or 1 LEGACY_MODE => MULT18X18S, -- Backward compatibility, NONE, -- MULT18X18 or MULT18X18S MREG => 1, -- Number of multiplier pipeline registers, 0 or 1 OPMODEREG => 1,-- Number of pipeline registers on OPMODE input, -- 0 or 1 PREG => 1, -- Number of pipeline registers on the P output, 0 or 1 SIM_X_INPUT => GENERATE_X_ONLY, -- Simulation parameter for behavior for X on input. -- Possible values: GENERATE_X, NONE or WARNING SUBTRACTREG => 1)-- Number of pipeline registers on the SUBTRACT -- input, 0 or 1
Attributes in Verilog
defparam DSP48_inst.AREG = 1; // Number of pipeline registers on the A input, 0, 1 or 2 defparam DSP48_inst.BREG = 1; // Number of pipeline registers on the B input, 0, 1 or 2 defparam DSP48_inst.B_INPUT = DIRECT; // B input DIRECT from fabric or CASCADE from another DSP48 defparam DSP48_inst.CARRYINREG = 1; // Number of pipeline registers for the CARRYIN input, 0 or 1 defparam DSP48_inst.CARRYINSELREG = 1; // Number of pipeline registers for the CARRYINSEL, 0 or 1 defparam DSP48_inst.CREG = 1; // Number of pipeline registers on the C input, 0 or 1 defparam DSP48_inst.LEGACY_MODE = MULT18X18S; // Backward compatibility, NONE, MULT18X18 or MULT18X18S defparam DSP48_inst.MREG = 1; // Number of multiplier pipeline registers, 0 or 1 defparam DSP48_inst.OPMODEREG = 1; // Number of pipeline registers on OPMODE input, 0 or 1 defparam DSP48_inst.PREG = 1; // Number of pipeline registers on the P output, 0 or 1 defparam DSP48_inst.SIM_X_INPUT = GENERATE_X_ONLY; // Simulation parameter for behavior for X on input. // Possible values: GENERATE_X, NONE or WARNING defparam DSP48_inst.SUBTRACTREG = 1; // Number of pipeline registers on the SUBTRACT input, 0 or 1
Xilinx 11
DSP48 Tile and Interconnect

Two DSP48 slices, a shared 48-bit C bus, and dedicated interconnect form a DSP48 tile. The DSP48 tiles stack vertically in a DSP48 column. The height of a DSP48 tile is the same as four CLBs and also matches the height of one block RAM. This regularity enhances the routing of wide datapaths. Smaller Virtex-4 family members have one DSP48 column while the larger Virtex-4 family members have two, four, or eight DSP48 columns. As shown in Figure 2-2, the multipliers and block RAM share interconnect resources in the Virtex-II and Virtex-II Pro architectures. Virtex-4 devices, however, have independent routing for the DSP48 tiles and block RAM, effectively doubling the available data bandwidth between the elements.
Virtex-II and Virtex-II Pro Devices
Interconnect
Multiplier
BRAM
Virtex-4 Devices
Interconnect
Interconnect
BRAM
DSP48 DSP48 Slice Slice
ug073_c1_02_060304
Figure 2-2: DSP48 Interconnect and Relative Dedicated Element Sizes Figure 2-3 shows two DSP48 slices and their associated datapaths stacked vertically in a DSP48 column. The inputs to the shaded multiplexers are selected by configuration control signals. These are set by attributes in the HDL source code or by the User Constraint File (UCF).
12 Xilinx
BCOUT 18 18 A 18 B 18 18 Note 1 36 X 48 36 Note 2 72 Note 3 36 48 48 C 48 Note 7 48 18 Zero Z 48 SUBTRACT Note 8 Y 48 CIN Note 4
PCOUT
48
48
Notes 4, 5 48 Note 5 BCIN BCOUT 18 18 Note 4 Note 1 36 X 48 36 Note 2 72 Note 3 36 48 48 Zero Z 48 SUBTRACT Note 8 Y 48 CIN 48 48 48 Wire Shift Right by 17 bits PCIN PCOUT
A 18 B 18
18
18 Notes 4, 5 48 Note 5 BCIN Wire Shift Right by 17 bits 48 PCIN

ug073_c1_03_020405
Figure 2-3: A DSP48 Tile Consisting of Two DSP48 Slices Notes: 1. The 18-bit A bus and B bus are concatenated, with the A bus being the most significant. 2. The X,Y, and Z multiplexers are 48-bit designs. Selecting any of the 36-bit inputs provides a 48-bit sign-extended output. 3. The multiplier outputs two 36-bit partial products, sign extended to 48 bits. The partial products feed the X and Y multiplexers. When OPMODE selects the multiplier, both X and Y multiplexers are utilized and the adder/subtracter combines the partial products into a valid multiplier result. 4. The multiply-accumulate path for P is through the Z multiplexer. The P feedback through the X multiplexer enables accumulation of P cascade when the multiplier is not used. 5. The Right Wire Shift by 17 bits path truncates the lower 17 bits and sign extends the upper 17 bits. 6. The grey-colored multiplexers are programmed at configuration time. 7. The shared C register supports multiply-add, wide addition, or rounding. 8. Enabling SUBTRACT implements Z (X+Y+CIN) at the output of the adder/subtracter.
Xilinx 13
Simplified DSP48 Slice Operation

The math portion of the DSP48 slice consists of an 18-bit by 18-bit, twos complement multiplier followed by three 48-bit datapath multiplexers (with outputs X, Y, and Z) followed by a three-input, 48-bit adder/subtracter. The data and control inputs to the DSP48 slice feed the arithmetic portions directly, or are optionally registered one or two times to assist the construction of different, highly pipelined, DSP application solutions. The data inputs A and B can be registered once or twice. The other data inputs and the control inputs can be registered once. Full speed operation is 500 MHz when using the pipeline registers. More detailed timing information is available in the Timing Section. In its most basic form the output of the adder/subtracter is a function of its inputs. The inputs are driven by the upstream multiplexers, carry select logic, and multiplier array. Equation 2-1 summarizes the combination of X, Y, Z, and CIN by the adder/subtracter. The CIN, X multiplexer output, and Y multiplexer output are always added together. This combined result can be selectively added to or subtracted from the Z multiplexer output. A d d e r O u t = (Z ( X + Y + C I N ) ) Equation 2-1
Equation 2-2 describes a typical use where A and B are multiplied and the result is added to or subtracted from the C register. More detailed operations based on control and data inputs are described in later sections. Selecting the multiplier function consumes both X and Y multiplexer outputs to feed the adder. The two 36-bit partial products from the multiplier are sign extended to 48 bits before being sent to the adder/subtracter. A d d e r O u t = C (A B + CI N) Equation 2-2
Figure 2-4 shows the DSP48 slice in a very simplified form. The seven OPMODE bits control the selection of the 48-bit datapaths by the three multiplexers feeding each of the three inputs to the adder/subtracter. In all cases, the 36-bit input data to the multiplexers is sign extended, forming 48bit input datapaths to the adder/subtracter. Based on 36-bit operands and a 48-bit accumulator output, the number of guard bits (i.e., bits available to guard against overflow) is 12. Therefore, the number of multiply accumulations possible before overflow occurs is 4096. Combinations of OPMODE, SUBTRACT, CARRYINSEL, and CIN control the function of the adder/subtracter.
OPMODE Controls Behavior P A:B A B C Zero PCIN P
UG073_c1_04_070904
OPMODE, CARRYINSEL, CIN, and SUBTRACT Control Behavior
Figure 2-4: 14 Xilinx
Simplified DSP48 Slice Model
Timing Model
Table 2-3 lists the XtremeDSP switching characteristics. Table 2-3: XtremeDSP Switching Characteristics Symbol Setup and Hold of CE Pins TDSPCCK_CE/TDSPCKC_CE TDSPCCK_RST/TDSPCKC_RST Setup/Hold of all CE inputs of the DSP48 slice Setup/Hold of all RST inputs of the DSP48 slice Setup/Hold of {A, B, C} input to {A, B, C} register Setup/Hold of {A, B} input to M register Setup/Hold of {A, B} input to P register (LEGACY_MODE = MULT18X18) Setup/Hold of {A, B, C} input to P register (LEGACY_MODE = NONE for A and B) Setup/Hold of {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT} input to {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT} register Setup/Hold of {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT, PCIN} input to P register Clock Enable Reset CE RST Description Function Control Signal
Setup and Hold Times of Data/Control Pins TDSPDCK_{AA, BB, CC}/ TDSPCKD_{AA, BB, CC} TDSPDCK_{AM, BM}/ TDSPCKD_{AM, BM} TDSPDCK_{AP, BP}_L/ TDSPCKD_{AP, BP}_L TDSPDCK_{AP_NL, BP_NL, CP}/ TDSPCKD_{AP_NL, BP_NL, CP} TDSPDCK_{CRYINC, CRYINSC, OPO, SUBS}/ TDSPCKD_{CRYINC, CRYINSC,
OPO, SUBS}
Data In Data In Data In Data In
A, B, C A, B A, B A, B, C
Control In
Various
TDSPDCK_{CRYINP, CRYINSP, OPP, SUBPPCINP}/ TDSPCKD_{CRYINP, CRYINSP,

OPP, SUBP, PCINP}
Control In
Various
Clock to Out TDSPCKO_PP TDSPCKO_{PA, PB}_L Clock to out from P register to P output Clock to out from {A, B} register to P output (LEGACY_MODE = MULT18X18) Clock to out from {A, B, C} register to P output (LEGACY_MODE = NONE for A and B) Clock to out from {M, CARRYIN, CARRYINSEL, OPMODE, SUBTRACT} register to P output Data Out Data Out P Output P Output
TDSPCKO_{PA_NL, PB_NL, PC}
Data Out
P Output
TDSPCKO_{PM, PCRYIN,
PCRYINS, POP, PSUB}
Data Out
P Output
Xilinx 15
DSP: DESIGNING FOR OPTIMAL RESULTS Table 2-3: XtremeDSP Switching Characteristics (Continued) Symbol TDSPCKO_PCOUTP Description Clock to out from P register to PCOUT output Function Data Out Data Out Control Signal P Output P Output
TDSPCKO_{PCOUTA, PCOUTB}_L Clock to out from {A, B} register to PCOUT output (LEGACY_MODE = MULT18X18) TDSPCKO_{PCOUTA_NL,
PCOUTB_NL, PCOUTC}
Clock to out from {A, B, C} register to PCOUT output (LEGACY_MODE = NONE for A and B)
Data Out
P Output
TDSPCKO_{PCOUTM,
PCOUTCRYIN, PCOUTCRYINS, PCOUTOP, PCOUTSUB}
Data Out Clock to out from {M, CARRYIN, CARRYINSEL, OPMODE, SUBTRACT} register to PCOUT output {A, B} input to P output (LEGACY_MODE = MULT18X18) {A, B, C} input to P output (LEGACY_MODE = NONE for A and B) {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT, PCIN} input to P output {A, B} input to PCOUT output (LEGACY_MODE = MULT18X18) {A, B, C} input to PCOUT output (LEGACY_MODE = NONE for A and B) {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT, PCIN} input to PCOUT output From {A, B} register to P register (LEGACY_MODE = MULT18X18) From {A, B, C, P} register to P register (LEGACY_MODE = NONE for A and B) Data In to Out Data In to Out Control to Data Out Data In to PC Out
P Output
Combinatorial TDSPDO_{AP, BP}_L TDSPDO_{AP_NL, BP_NL, CP} A, B to P A, B, C to P Various
TDSPDO_{CRYINP, CRYINSP,
OPMODEP, SUBTRACTP, PCINP}
TDSPDO_{APCOUT, BPCOUT}_L TDSPDO_{APCOUT_NL,

BPCOUT_NL, CPCOUT}
A, B to PC Out
Data In A, B, C to to PC Out PC Out Control to PC Out Various
TDSPDO_{CRYINPCOUT,
CRYINSPCOUT, OPMODEPCOUT, SUBTRACTPCOUT, PCINPCOUT}
Sequential TDSPCKCK_{AP, BP}_L TDSPCKCK_{AP_NL, BP_NL, CP,

PP}
Register to register Register to register
16 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS Table 2-3: XtremeDSP Switching Characteristics (Continued) Symbol TDSPCKCK_{CRYINP, CRYINSP,
OPMODEP, SUBTRACTP}
Description From {CARRYIN, CARRYINSEL, OPMODE, SUBTRACT} register to P register From {A, B} register to M register
Function Register to register Register to register
Control Signal
TDSPCKCK__{AM, BM}
The timing diagram in Figure 2-5 uses OPMODE equal to 0x05 with all pipeline registers turned on. For other applications, the clock latencies and the parameter names must be adjusted.
CLK Event 1 CLK TDSPCCK_CE CE TDSPCCK_RST RST TDSPDCK_AA A Don't Care Data A1 Data A2 Data A3 Data A4 CLK Event 4 CLK Event 5
TDSPDCK_BB B Don't Care Data B1 Data B2 Data B3 Data B4
TDSPDCK_CC C Don't Care Data C1 Data C2 TDSPCKO_CC P Don't Care Result 1 0

UG073_c1_27_071204
Data C3
Data C4 TDSPCKO_CC
Figure 2-5:
XtremeDSP Timing Diagram
The following events occur in Figure 2-5: 1. At time TDSPCCK_CE before CLK event 1, CE becomes valid High to allow all DSP registers to sample incoming data. 2. At time TDSPDCK_{AA,BB,CC} before CLK event 1, data inputs A, B, C have remained stable for sampling into the DSP slice. 3. At time TDSPCKO_PP after CLK event 4, the P output switches into the results of the data captured at CLK event 1. This occurs three clock cycles after CLK event 1. Xilinx 17
DSP: DESIGNING FOR OPTIMAL RESULTS 4. 5. At time TDSPCCK_RST before CLK event 5, the RST signal becomes valid High to allow a synchronous reset at CLK event 5. At time TDSPCKO_PP after CLK event 5, the output P becomes a logic 0.
A, B, C, and P Port Logic

The DSP48 slice input and output data ports support many common DSP and math algorithms. The DSP48 slice has two direct 18-bit input data ports labeled A and B. Two DSP48 slices within a DSP48 tile share a direct 48-bit input data port labeled C. Each DSP48 slice has one direct 48-bit output port labeled P, a cascaded input datapath (B cascade), and a cascaded output datapath (P cascade), providing a cascaded input and output stream between adjacent DSP48 slices. Applications benefiting from this feature include FIR filters, complex multiplication, multi-precision multiplication, complex MACs, adder cascade, and adder tree (the final summation of several multiplier outputs) support. The 18-bit A and B port can supply input data to the 18-bit by 18-bit, two's complement multiplier. A and B concatenated can bypass the multiplier and feed the X multiplexer input. The 48bit C port is used as a general input to the Y and Z multiplexer to perform multiply, add, subtract, three-input add/subtract functions, or rounding. Multiplexers controlled by configuration bits select flow through paths, optional registers, or cascaded inputs. The data port registers allow users to typically trade off increased clock frequency (i.e., higher performance) vs. data latency. There is also a configuration controlled pipeline register between the multiplier and adder/subtracter known as the M register. The registers have independent clock enables and resets as described in Table 2-2 and shown in Figure 2-1. The configuration bit enables the C register to select between two potentially different clock domains as shown in Figure 2-8, page 19. The selection of the clock multiplexer is not set by user attributes. If the C register is used, the DSP48 slices packed in the same DSP48 tile must either be in the same clock domain or meet multicycle clock constraints. The shared C input within the DSP tile can be used by the two slices within a tile in any one of the following modes: 1. Neither DSP48 slice uses the C port. The C inputs in both the slices are connected to GND, 0 in the HDL code. The place and route software maps the two slices in one tile. 2. Both DSP48 slices use the same C port inputs. The C inputs in both the slices are connected to C in the HDL code. The place and route software maps the two slices in one tile. 3. Only one DSP48 slice uses the C port. In this case, the C input on slice 1 is connected to C, and the C input on slice 2 is connected to 0 in the HDL code. A C port connected to 0 is taken as an unused C port in the software. The software can map the two slices in one tile. The simulation shows the C input connected to 0 for slice 2 in the code. However, in the hardware, the C port on slice 2 is connected to the C port on slice 1, causing a potential simulation mismatch for the C port on slice 2. To avoid this potential mismatch, the C port must not be selected on the Y and Z multiplexers of slice 2. To get a 0 at the output of multiplexers Y and Z, choose the 0 input of these multiplexers using OPMODE. Do not use the C input to get a zero at the output of Y and Z multiplexers on slice 2.
18 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS The A, B, C, and P port logics are shown in Figure 2-6, Figure 2-7, Figure 2-8, and Figure 2-9, respectively.
A
18 18
18 D EN RST Q D EN RST Q 18
18
A input to Multiplier
CEA RSTA
UG073_c1_05_061304
Figure 2-6:
A Input Logic
B
18 18
BCIN
18 18 D EN RST Q D EN RST Q 18 18
B input to Multiplier
CEB RSTB
UG073_c1_06_061304
Figure 2-7:
C
B Input Logic
48 48 D Q 48
To Both DSP48 Slices
CLK_0
CEC
EN RST
CLK_1 RSTC
UG073_c1_07_061304
Figure 2-8:
C Input Logic
Xilinx 19
P
48 48 D Q 48
DSP48 Slice Output
CEP
EN RST
RSTP
UG073_c1_08_061304
Figure 2-9:
P Output Logic
OPMODE, SUBTRACT, and CARRYINSEL Port Logic

The OPMODE, SUBTRACT, and CARRYINSEL port logic supports flowthrough or registered input control signals. Similar to the datapaths, multiplexers controlled by configuration bits select flowthrough or optional registers. The control port registers allow users to trade off increased clock frequency (i.e., higher performance) vs. data latency. The registers have independent clock enables and resets as described in Table 2-2 and shown in Figure 2-1. The OPMODE, SUBTRACT, and CARRYINSEL registers are reset by RSTCTRL. The SUBTRACT register has a separate enable labeled CECINSUB from OPMODE and CARRYINSEL. This enable signal is also used to enable the carry input from the general interconnect described in the Carry Input Logic subsection. Figure 2-10 shows the OPMODE, SUBTRACT, and CARRYINSEL port logic.
OPMODE 7 D CECTRL EN RST RSTCTRL SUBTRACT To Adder/Subtracter D CECINSUB EN RST Q Q 7 To the X, Y, Z Multiplexers and Carry Input Select Logic
CARRYINSEL
2 D EN RST Q 2
To Carry Input Select Logic
ug073_c1_09_070904
OPMODE, SUBTRACT, and CARRYINSEL Port Logic
Twos Complement Multiplier

The two's complement multiplier inside the DSP48 slice accepts two 18-bit x 18-bit two's complement inputs and produces a 36-bit two's complement result. Cascading of multipliers to achieve larger products is supported with a 17-bit right-shifted cascaded bus input to the adder/subtracter to right justify partial products by the correct number of bits. MAC functions can also right justify intermediate results for multi-precision. The multiplier can emulate unsigned math by setting the MSB of an 18-bit operand to zero. The output of the multiplier consists of two 36-bit partial products. The 36-bit partial products are sign extended to 48 bits prior to being input to the adder/subtracter. Selecting the output of the multiplier consumes both X and Y multiplexers whereby the adder/subtracter combines the partial products to form the final result. Figure 2-11 shows an optional pipeline register (MREG) for the output of the multiplier. Using the register provides increased performance with a single clock cycle of increased latency. The grey multiplexer indicates selected at configuration time by configuration bits.
A B
36 72
Partial Product 1 Partial Product 2

ug073_c1_10_070904
Optional MREG
36
Figure 2-11:
Twos Complement Multiplier Followed by Optional MREG
X, Y, and Z Multiplexer
The Operating Mode (OPMODE) inputs provide a way for the design to change its functionality from clock cycle to clock cycle (e.g., when altering the initial or final state of the DSP48 relative to the middle part of a given calculation). The OPMODE bits can be optionally registered under the control of the configuration memory cells (as denoted by the grey colored MUX symbol in Figure 2-10). Table 2-4, Table 2-5, and Table 2-6 list the possible values of OPMODE and the resulting function at the outputs of the three multiplexers (X, Y, and Z multiplexers). The multiplexer outputs supply three operands to the following adder/subtracter. Not all possible combinations for the multiplexer select bits are allowed. Some are marked in the tables as illegal selection and give undefined results. If the multiplier output is selected, then both the X and Y multiplexers are consumed supplying the multiplier output to the adder/subtracter. Table 2-4: Z XXX XXX XXX XXX OPMODE Control Bits Select X, Y, and Z Multiplexer Outputs OPMODE Binary Y XX 01 XX XX X 00 01 10 11 X Multiplexer Output Fed to Add/Subtract ZERO (Default) Multiplier Output (Partial Product 1) P A concatenate B
Xilinx 21
DSP: DESIGNING FOR OPTIMAL RESULTS Table 2-5: Z XXX XXX XXX XXX OPMODE Control Bits Select X, Y, and Z Multiplexer Outputs OPMODE Binary Y 00 01 10 11 X XX 01 XX XX Y Multiplexer Output Fed to Add/Subtract ZERO (Default) Multiplier Output (Partial Product 2) Illegal selection C
Table 2-6: Z
000 001 010 011 100 101 110 111
OPMODE Control Bits Select X, Y, and Z Multiplexer Outputs OPMODE Binary Y

XX XX XX XX XX XX XX XX
X
XX XX XX XX XX XX XX XX
Z Multiplexer Output Fed to Add/Subtract ZERO (Default) PCIN P C Illegal selection Shift (PCIN) Shift (P) Illegal selection
There are seven possible non-zero operands for the three-input adder as selected by the three multiplexers, and the 36-bit operands are sign extended to 48 bits at the multiplexer outputs: 1. Multiplier output, supplied as two 36-bit partial products 2. Multiplier bypass bus consisting of A concatenated with B 3. C bus, 48 bits, shared by two slices 4. Cascaded P bus, 48 bits, from a neighbor DSP48 slice 5. Registered P bus output, 48 bits, for accumulator functions 6. Cascaded P bus, 48 bits, right shifted by 17 bits from a neighbor DSP48 slice 7. Registered P bus output, 48 bits, right shifted by 17 bits, for accumulator functions
Three-Input Adder/Subtracter
The adder/subtracter output is a function of control and data inputs. OPMODE, as shown in the previous section, selects the inputs to the X, Y, Z multiplexer directed to the associated three adder/subtracter inputs. It also describes how selecting the multiplier output consumes both X and Y multiplexers. As with the input multiplexers, the OPMODE bits specify a portion of this function. Table 2-7 shows OPMODE combinations and the resulting functions. The symbol in the table means either add or subtract and is specified by the state of the SUBTRACT control signal (SUBTRACT = 1 is defined as subtraction). The symbol : in the table means concatenation. The outputs of the X and Y
22 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS multiplexer and CIN are always added together. This result is then added to or subtracted from the output of the Z multiplexer. Table 2-7: Hex OPMODE [6:0] 0x00 0x02 0x03 0x05 0x0c 0x0e 0x0f 0x10 0x12 0x13 0x15 0x1c 0x1e 0x1f 0x20 0x22 0x23 0x25 0x2c 0x2e 0x2f 0x30 0x32 0x33 0x35 0x3c 0x3e 0x3f 0x50 0x52 OPMODE Control Bits Adder/Subtracter Function Binary OPMODE Z Y X 000 00 00 000 00 10 000 00 11 000 01 01 000 11 00 000 11 10 000 11 11 001 00 00 001 00 10 001 00 11 001 01 01 001 11 00 001 11 10 001 11 11 010 00 00 010 00 10 010 00 11 010 01 01 010 11 00 010 11 10 010 11 11 011 00 00 011 00 10 011 00 11 011 01 01 011 11 00 011 11 10 011 11 11 101 00 00 101 00 10 Z 0 0 0 0 0 0 0 PCIN PCIN PCIN PCIN PCIN PCIN PCIN P P P P P P P C C C C C C C Shift (PCIN) Shift (PCIN) XYZ Multiplexer Outputs and Adder/Subtracter Output Y
0 0 0
X
0
Adder/Subtracter Output CIN (P + CIN)
A:B (A:B + CIN) Note 1 (A B + CIN)

0
C C C
0 0 0
(C + CIN) (C + P + CIN) PCIN CIN
P
0
A:B (A:B + C + CIN) P
PCIN (P + CIN) A:B PCIN (A:B + CIN) Note 1 PCIN (A B + CIN)

0
C C C
0 0 0
PCIN (C + CIN)
PCIN (C + P + CIN) A:B PCIN (A:B + C + CIN)

0
P CIN
P P (P + CIN) A:B P (A:B + CIN) Note 1 P (A B + CIN)

0
C C C
0 0 0
P (C + CIN)
P P (C + P + CIN) A:B P (A:B + C + CIN)

0
C CIN C (P + CIN) C (A B + CIN) C (C + CIN) C (C + P + CIN) Shift(PCIN) CIN Shift(PCIN) (P + CIN) Xilinx 23
A:B C (A:B + CIN)

0
Note 1 C C C
0 0
P
0
A:B C (A:B + C + CIN) P
DSP: DESIGNING FOR OPTIMAL RESULTS Table 2-7: Hex OPMODE [6:0] 0x53 0x55 0x5c 0x5e 0x5f 0x60 0x62 0x63 0x65 0x6c 0x6e 0x6f OPMODE Control Bits Adder/Subtracter Function (Continued) Binary OPMODE Z Y X 101 00 11 101 01 01 101 11 00 101 11 10 101 11 11 110 00 00 110 00 10 110 00 11 110 01 01 110 11 00 110 11 10 110 11 11 XYZ Multiplexer Outputs and Adder/Subtracter Output Z Shift (PCIN) Shift (PCIN) Shift (PCIN) Shift (PCIN) Shift (PCIN) Shift (P) Shift (P) Shift (P) Shift (P) Shift (P) Shift (P) Shift (P) Y
0
X Adder/Subtracter Output A:B Shift(PCIN) (A:B + CIN) Note 1 Shift(PCIN) (A B + CIN)

0
C C C
0 0 0
Shift(PCIN) (C + CIN) Shift(PCIN) (C + P + CIN) Shift(P) CIN
P
0
A:B Shift(PCIN) (A:B + C + CIN) P
Shift(P) (P + CIN) A:B Shift(P) (A:B + CIN) Note 1 Shift(P) (A B + CIN)

0
C C C
Shift(P) (C + CIN)
Shift(P) (C + P + CIN) A:B Shift(P) (A:B + C + CIN)
Notes: 1. When the multiplier output is selected, both X and Y multiplexers are used to feed the multiplier partial products to the adder input.
Carry Input Logic

The carry input logic result is a function of the OPMODE control bits and CARRYINSEL. The inputs to the carry input logic appear in Figure 2-12. Carry inputs used to form results for adders and subtracters are always in the critical path. High performance is achieved by implementing this logic in the diffused silicon. The possible carry inputs to the carry logic are gathered prior to the outputs of the X, Y, and Z multiplexers. In a sense, the X, Y, and Z multiplexer function is duplicated for the carry inputs to the carry logic. Both OPMODE and CARRYINSEL must be in the correct state to ensure the correct carry input (CIN) is selected.
24 Xilinx

CARRYINSEL 2 Carry input from general fabric (to cause counter increment, etc.) CARRYIN 00 D CECINSUB OPMODE RST Round a previous P result Round a previous PCIN result ~P[47] RSTCARRYIN 01 ~PCIN[47] OPMODE Round an external value input via A:B Round the output of the multiplier 10 ~A[17] Carry Input (CIN) to Adder/Subtracter EN Q
A[17] XNOR B[17] D CECARRYIN EN RST RSTCARRYIN

UG073_c1_11_070904
11
Similar Function as X, Y, Z Data MUX
Figure 2-12: Carry Input Logic Feeding the Adder/Subtracter Figure 2-12 shows four inputs, selected by the 2-bit CARRYINSEL control with the OPMODE bits providing additional control. The first input CARRYIN (CARRYINSEL is equal to binary 00) is driven from general logic. This option allows implementation of a carry function based on user logic. It can be optionally registered to match the pipeline delay of the MREG when used. This register delay is controlled by configuration. The next input (CARRYINSEL is equal to binary 01) is the inverted MSB of either the output P or the cascaded output, PCIN (from an adjacent DSP48 slice). The final selection between P or PCIN is dictated by OPMODE[4] and OPMODE[6]. The third input (CARRYINSEL is equal to binary 10) is the inverted MSB of A, for rounding A concatenated with B values, or A[17] XNOR B[17] for rounding multiplier outputs. Again, the state of OPMODE determines the final selection. The fourth and final input is merely a registered version of the third input to adjust the carry input delay when using the multiplier output register or MREG. Table 2-8 lists the possible values of the two carry input select bits (CARRYINSEL), the operation mode bus (OPMODE), and the resulting carry inputs or sources.
Xilinx 25
DSP: DESIGNING FOR OPTIMAL RESULTS Table 2-8: OPMODE and CARRYINSEL Control Carry Source OPMODE XXX XX XX Z MUX output = P or Shift(P) Z MUX output = PCIN or Shift(PCIN) X and Y MUX output = multiplier partial products X and Y MUX output = multiplier partial products X MUX output = A:B X MUX output = A:B Carry Source CARRYIN ~P[47] ~PCIN[47] Comments General fabric carry source (registered or not) Rounding P or Shift(P) Rounding the cascaded PCIN or Shift(PCIN) from adjacent slice
CARRYINSEL[1:0] 00 01 01
10
11
10 11
A[17] xnor B[17] Rounding multiplier (MREG pipeline register disabled) A[17] xnor B[17] Rounding multiplier (MREG pipeline register enabled) ~A[17] Rounding A:B (not pipelined) ~A[17] Rounding A:B (pipelined)
Symmetric Rounding Supported by Carry Logic

Arithmetic rounding is a process where a result is quantized in an intelligent manner. The bit position placement where rounding occurs is up to the designer and is determined solely by a constant loaded in the C register. While the binary point placement and bit position where rounding occurs are independent of each other, the following discussion assumes one wants to round off the fractional bits. One form of rounding is simple truncation or just dropping undesired LSBs from a large result to obtain a reduced number of result bits. The problem with truncation happens after the bits are dropped and the new reduced result is biased in the wrong direction. For example, if a number has the decimal value 2.8 and the fractional part of the number is truncated, then the result is two. In this example, the original number is closer to 3 than to 2, and a rounded result of 3 is more desirable than the simple truncated result of 2. Another method of quantization is known as symmetric rounding. Symmetric rounding accomplishes the more desirable effect of quantizing numbers to keep them from becoming biased in the wrong direction. For example, the number 2.8 rounds to 3.0 and the number 2.2 rounds to 2.0. Negative numbers, such as 2.8 and 2.2, round to 3.0 and 2.0 respectively. The midpoint number 2.5 rounds to 3.0 and 2.5 rounds to 3. Another way to describe this type of quantization (for fractional rounding) is to round to the nearest integer and at the midpoint round away from zero. For positive numbers this effect is achieved by adding 0.1000 binary and truncating the fraction of the result. For negative numbers this effect is achieved by adding 0.0111 and truncating the fraction of the result. The implementation of the symmetric rounding in the DSP48 slice allows the user to load a single constant. If the design calls for eight bits (out of 48 total bits) to be rounded, then load 0x00000000007F into the C register. The number of bits to be rounded off is one more than the 26 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS number of ones present in the C register. Table 2-9 has examples for rounding off the fractional bits from a value (binary point placement and rounded bits placement coincide). Table 2-9: Symmetric Rounding Examples C Value 0000.0111 0000.0111 0000.0111 0000.0111 0000.0111 0000.0111 Internally Generated CIN 1 1 1 0 0 0 Multiplier Plus C Plus CIN 0010.1111 0011.0000 0011.0001 1110.0000 1101.1111 1101.1110 After After Truncation Truncation (Decimal) (Binary) 0010 2 0011 3 0011 3 1110 -2 1101 -3 1101 -3 Multiplier Multiplier Output Output (Binary) (Decimal) 2.4375 0010.0111 2.5 0010.1000 2.5625 0010.1001 2.4375 1101.1001 2.5 1101.1000 2.5625 1101.0111
Forming Larger Multipliers

Figure 2-13 illustrates the formation of a 35 x 35-bit multiplication from smaller 18 x 18-bit multipliers. The notation 0,B[16:0] denotes B has a leading zero followed by 17 bits, forming a positive two's complement number.
x AU = A[34:17] BU = B[34:17] AL = 0,A[16:0] BL = 0,B[16:0]
Sign Extend 36 bits of '0'
BL * AL = 34 bits [33:17] [16:0]

17-bit Offset
Sign Extend 18 bits of A[34] Sign Extend 18 bits of B[34]
BL * AU = 35 bits [34:17] [16:0] BU * AL = 35 bits [34:17] [16:0]
[35:18] P[69:52]
BU * AU = 36 bits [17:0] P[51:34] P[33:17]
34-bit Offset
P[16:0]
ug073_c1_12_070904
Figure 2-13:
35x35-bit Multiplication from 18x18-bit Multipliers
When separating two's complement numbers into two parts, only the most-significant part carries the original sign. The least-significant part must have a forced zero in the sign position meaning they are positive operands. While it seems logical to separate a positive number into the sum of two positive numbers, it can be counter intuitive to separate a negative number into a negative most-significant part and a positive least-significant part. However, after separation, the mostsignificant part becomes more negative by the amount the least-significant part becomes more
Xilinx 27
DSP: DESIGNING FOR OPTIMAL RESULTS positive. The forced zero sign bit in the least-significant part is why only 35-bit operands are found instead of 36-bit operands. The DSP48 slices, with 18 x 18 multipliers and post adder, can now be used to implement the sum of the four partial products shown in Figure 2-13. The lessor significant partial products must be right-shifted by 17 bit positions before being summed with the next most-significant partial products. This is accomplished with a built in wire shift applied to PCIN supplied as one selectable Z multiplexer input. The entire process of multiplication, shifting, and addition using adder cascade to form the 70-bit result can remain in the dedicated silicon of the DSP48 slice, resulting in maximum performance with minimal power consumption. Figure 2-21, page 41 illustrates the implementation of a 35 x 35 multiplier using the DSP48 slices.
FIR Filters
Basic FIR Filters
FIR filters are used extensively in video broadcasting and wireless communications. DSP filter applications include, but are not limited to the following: Wireless Communications Image Processing Video Filtering Multimedia Applications Portable Electrocardiogram (ECG) Displays Global Positioning Systems (GPS) Equation 2-3 shows the basic equation for a single-channel FIR filter. k = N1 y(n) =
k=0
h ( k )x ( n k ) Equation 2-3
The terms in the equation can be described as input samples, output samples, and coefficients. Imagine x as a continuous stream of input samples and y as a resulting stream (i.e., a filtered stream) of output samples. The n and k in the equation correspond to a particular instant in time, so to compute the output sample y(n) at time n, a group of input samples at N different points in time, or x(n), x(n-1), x(n-2), x(n-N+1) is required. The group of N input samples are multiplied by N coefficients and summed together to form the final result y. The main components used to implement a digital filter algorithm include adders, multipliers, storage, and delay elements. The DSP48 slice includes all of the above elements, which makes it ideal to implement digital filter functions. All of the input samples from the set of n samples are present at the input of each DSP48 slice. Each slice multiplies the samples with the corresponding coefficients within the DSP48 slice. The outputs of the multipliers are combined in the cascaded adders.
28 Xilinx
x(n)
Z-1
Z-1
Z-1
Z-1
Z-1
h(0)
h(1)
h(2)
h(3)
h(4)
h(N-1)
y(n)
+
UG073_c6_01_070904
Figure 2-14:
Conventional Tapped Delay Line FIR Filter
In Figure 2-14, the sample delay logic is denoted by Z-1, whereas the 1 represents a single clock delay. The delayed input samples are supplied to one input of the multiplier. The coefficients (denoted by h0 to h(N-1)) are supplied to the other input of the multiplier through individual ROMs, RAMs, registers, or constants. Y(n) is merely the summation of a set of input samples, and in time, multiplied by their respective coefficients.
Multi-Channel FIR Filters

Multi-channel filters are used to filter multiple data streams of input signals using the same set of coefficients for all the channels, or using different coefficient sets for different channels. A common example of a multi-channel filter is a radio receivers digital down converter. Equation 2-4 shows the equation, and Figure 2-15 shows the block diagram. A digitized baseband signal is applied to a matched low-pass filter M(z) to reduce the data rate from the input sample rate to the bit rate. The resulting in-phase and quadrature components are each processed by the same filter and, therefore, could be processed by a single, multi-channel filter running at twice the sample rate. x (n ) = x I ( n) + j x Q ( n ) Equation 2-4
Xilinx 29
x ( n ) = x ( n ) + jx Q ( n ) I
v(n)
xl(n)
M(z)
xQ(n)
M(z)
DDS Direct Digital Synthesizer
UG073_c6_02_070904
Figure 2-15: Software-Defined Radio Digital Down Converter Some video applications use multi-channel implementations for multiple components of a video stream. Typical video components are red, green, and blue (RGB) or luma, chroma red, and chroma blue (YCrCb). The different video components can have the same coefficient sets or different coefficient sets for each channel by simply changing the coefficient ROM structure.
Creating FIR Filters

Referring to Figure 2-4, Table 2-4, Table 2-5, and Table 2-6, an inner product MAC operation starts by loading the first operand into the P register. The output of the multiplier is passed through the X and Y multiplexer, added to zero, and loaded into the P register. Note the load operation OPMODE with value 0000101 selects zero to be output on the Z multiplexer supplying one of the adder inputs. A previous MAC inner product can exit via the P bus during this clock cycle. In subsequent clock cycles, the MAC operation requires the X and Y multiplexers to supply the multiplier output and the Z multiplexer to supply the output of the P register to the adder. The OPMODE for this operation, which differs from the load cycle by a single bit, has a value of 0100101. The description above allows for continuous operation with the previous resulting output and initial load occurring in the same clock cycle. Refer to Chapter 4, MAC FIR Filters, for detailed information on using DSP48 slices to create MAC FIR filters. To create a simple multiply-add processing element using the DSP48 slice shown in Figure 2-4, set the X and Y multiplexers to multiply and select the cascaded input from another DSP48 output (PCIN) as the Z MUX input to the arithmetic unit. For a normal multiply-add operation, the OPMODE value is set to 0010101. Refer to Chapter 5, Parallel FIR Filters, for detailed information on using DSP48 slices to create Parallel FIR filters.
30 Xilinx
Adder Cascade vs. Adder Tree

In typical direct form FIR filters, an input stream of samples is presented to one data input of separate multipliers where coefficients supply the other input to the multipliers. An adder tree follows the multipliers where the outputs from many multipliers are combined as shown in Figure 2-16.
h7(n) 18 18 48 h6(n) 18 X(n-4) Z-2 18
48
h5(n) 18 18
48
+
48
h4(n) 18 X(n-2) Z-2 18
+
48 y(n-6)
h3(n) 18 18
+
48
h2(n) X(n) Z-2 18 18
The final stages of the post addition in logic are the performance bottleneck that consume more power.
h1(n) 18 18
48
+
48
h0(n) X(n) 18 18
ug073_c1_13_070904
Figure 2-16: FIR Filter Adder Tree Using DSP48 Slices One difficulty of the adder tree concept is defining the size. Filters come in various lengths and consume a variable number of adders forming an adder tree. Placing a fixed number of adder tree components in silicon displaces other elements or requires a larger FPGA, thereby increasing the cost of the design. In addition, the adder tree structure with a fixed number of additions forces the designer to use logic resources when the fixed number of additions is exceeded. Using logic resources dramatically reduces performance and increases power consumption. The key to maximizing
Xilinx 31
DSP: DESIGNING FOR OPTIMAL RESULTS performance and lowering power for DSP math is to remain inside the DSP48 column consisting entirely of dedicated silicon. The Virtex-4 solution accomplishes the post-addition process while guaranteeing no wasted silicon resources. It involves computing the additive result incrementally utilizing a cascaded approach as illustrated in Figure 2-17. Figure 2-17 is a systolic version of a direct form FIR with a latency of 10 clocks versus an adder tree latency of six clocks.
32 Xilinx

Slice 8
h7(n-7) 18 18
48
+
No wire shift 48
48
Y(n10)
Slice 7
h6(n-6) 18 18
48
+
No wire shift 48
Slice 6
h5(n-5) 18 18
48
+
No wire shift 48
Slice 5
h4(n-4) 18 18
48
+
No wire shift 48
Slice 4
h3(n-3) 18 18
48
+
No wire shift 48
The post adders are contained wholly in dedicated silicon for highest performance and lowest power
Slice 3
h2(n-2) 18 18
48
+
No wire shift 48
Slice 2
h1(n-1) 18 18
48
+
No wire shift 48
Slice 1
h0(n) X(n) 18 18 Zero Sign extended from 36 bits to 48 bits
48
+
ug073_c1_14_070904
Figure 2-17: Systolic FIR with Adder Cascade Care should be taken to balance the input sample delay and the coefficients with the cascaded adder. The adaptive coefficients are staggered in time (wave coefficients).
Xilinx 33
DSP48 Slice Functional Use Models

The use models in this section explain how the DSP48 slices are used in various DSP applications. Starting with simple multiplication and then growing in complexity, DSP48 slices can be connected in a variety of ways, trading performance and slice utilization. The tables and use models illustrate a sampling of different connections. In some designs full performance is desired and several slices with pipelined registers are used. In designs with lower sample rates, a single slice is used with multiple clock cycles creating partial results to be combined at the very end of the computation. Performance choices (i.e., using multiple clock cycles) can produce efficient slice counts. In either case, the use of pipeline registers allows the DSP48 slice to run at a very fast, full performance clock rate. Block diagrams showing the basic connections are also included. The VHDL and Verilog Instantiation Templates section shows how to instantiate and connect the DSP48 slice. In many cases, starting or ending states are different than the middle states of operation.
Single Slice, Multi-Cycle, Functional Use Models

Table 2-10 lists and summarizes four single slice use models. These examples use the high speed of the DSP48 slice to accomplish a complicated multi-cycle function by changing the OPMODE bits from cycle to cycle. Entries in the table name the function with suggestions for DSP48 slice function during different clock cycles. Further details are in the following subsections. DSP48 designs support extra pipeline stages to increase overall performance, however, the function remains the same with increased clock-cycle latency. Table 2-10: Single Slice DSP48 Implementation Inputs B B[17:0] B[17:0] C X X Function and OPMODE[6:0] Multiply 17-Bit Shift Feedback Multiply Add Multiply 17-Bit Shift Feedback Multiply Add MultiplyAccumulate 17-Bit Shift Feedback Multiply Add 0x05 0x65 Output P[16:0] P[52:17]
Single Slice Slice Cycle Mode Number A 35 x 18 1 1 0,A[16:0 Multiply ] 2 A[34:17 ] 35 x 35 Multiply 1 1 2 0,A[16:0 ] A[34:17 ] 0,A[16:0 ] A[34:17 ]
0,B[16:0] 0,B[16:0]
X X
0x05 0x65
P[16:0]
3 4
B[34:17] B[34:17]
X X
0x25 0x65
P[33:17] P[69:34]
34 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS Table 2-10: Single Slice DSP48 Implementation (Continued) Inputs B BRe[17:0] BIm[17:0] BIm[17:0] BRe[17:0] C X X X X Function and OPMODE[6:0] Multiply MultiplyAccumulate Multiply MultiplyAccumulate 0x05 0x25 0x05 0x25 P (Imaginary) P (Real) Output
Single Slice Slice Cycle Mode Number A Complex 1 1 ARe[17: Multiply 0] 2 AIm[17: 0] 3 ARe[17: 0] 4 AIm[17: 0]
Single Slice, 35 x 18 Multiplier Use Model

The first entry in Table 2-10 indicates how the signed 35 x 18 multiply is designed using a single DSP48 slice. The 35-bit A and 18-bit B operands are assumed to be signed, two's complement numbers with results also expressed as a signed, two's complement, 53-bit output. Operand A can only be 35 bits because when separating an operand into two 18-bit parts, the least-significant part must have the MSB forced to zero, thereby reducing the available operand bits from 36 to 35. The multiply function uses one slice (labeled slice 0 in Table 2-10) and computes the final result in two clocks. The 36-bit, least-significant partial product output formed during the first clock cycle is computed by multiplying the least-significant 17 bits of Operand A, which are forced positive (sign bit = 0), with the 18 bits of Operand B (including the original sign). 0,A[16:0] x B[17:0] The first product is loaded into the output register on this cycle. The lower 17 bits of the first partial product are the lower 17 bits of the final result. During the second clock cycle, the first partial product is shifted right by 17 bits, leaving the remaining bits to be fed back and added to the next partial product. This partial product is formed by multiplying the signed 18-bit Operand B with the signed upper 18 bits of Operand A. The lower 36 bits of the second partial product are the upper 36 bits of the final result. A[34:17] x B[17:0] Figure 2-18 shows the function during both clock cycles for a single DSP48 slice used as a 35bit x 18-bit, signed, two's complement multiplier. Increased performance is obtained by using the pipeline registers before and after the multiplier, however, the clock latency is increased.
Xilinx 35

Clock Cycle 2
A[34:17] 18 B[17:0] 18 A[34:17] B[17:0] P = right shifted PREG + (A[34:0] B[17:0])
48
P[52:17] 48 48
right wire shift by 17 bits
Clock Cycle 1
0,A[16:0] 18 B[17:0] 18
0,A[16:0] B[17:0]
PREG = 0,A[16:0] B[17:0]
+
48 Zero 48 48
P[16:0]
Sign extended from 36 bits to 48 bits
ug073_c1_15_071004
Figure 2-18:
Single Slice, 35 x 18-bit Multiplier
Single Slice, 35 x 35 Multiplier Use Model

The next entry in Table 2-10 indicates how the signed 35 x 35 multiply is designed using a single DSP48 slice. The 35-bit A and B operands are assumed to be signed, two's complement numbers with results expressed as a signed, two's complement, 70-bit output. Operands can only be 35 bits because when separating an operand into two 18-bit parts. The least-significant 18-bit part must have the MSB forced to zero, thereby reducing the available operand bits from 36 to 35. The flow is similar to the 35 x 18 multiply, but instead of two partial products, there are four: a lower partial product, two middle partial products, and an upper partial product. The multiply function uses one slice (labeled slice 1 in Table 2-10) and computes the final result in four clocks. The 36-bit lower partial product formed during the first clock cycle is computed by multiplying the least-significant 17 bits of Operand A, which are forced positive (sign bit = 0), with the least-significant 17 bits of Operand B, also forced positive. 0,A[16:0] x 0,B[16:0] The first product is loaded into the output register on this cycle. All 36-bit products from the multiplier are sign extended to 48 bits. During the second and third clock cycles, the two middle products are computed. In clock cycle two, the first or lower partial product in the P register is shifted right by 17 bits and fed back to the adder/subtracter. The output of the multiplier is the first middle product, expressed as: A[34:17] x 0,B[16:0] The adder/subtracter is set to add and the two partial products are added. In the third clock cycle, the previous result is fed back to the adder/subtracter, but not right shifted since its bits align with the next computed middle product expressed as: B[34:17] x 0,A[16:0] The adder/subtracter is again set to add, and the P register receives the sum of the three partial products. Finally, in the fourth clock cycle, the accumulated sum of partial products is again shifted right by 17 bits, and sign extended leaving the remaining bits to be fed back and added to the next partial 36 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS product. The upper partial product is formed by multiplying the signed upper 18 bits of B with the signed upper 18 bits of A. A[34:17] x B[34:17] The 70-bit result is output sequentially in 17-bit, 17-bit, and 36-bit segments as shown in Figure 2-19. Figure 2-19 shows the function during all four clock cycles for a single DSP48 slice used as a 35bit x 35-bit, signed, two's complement multiplier. Increased performance can be obtained by using the pipeline registers before and after the multiplier, however, the clock latency is increased.
Clock Cycle 4
A[34:17] 18 B[34:17] 18 A[34:17] x B[34:17] P = right shifted PREG + (A[34:0] x B[34:17])
48
+
48
P[69:34]
Clock Cycle 3
0,A[16:0] 18 B[34:17] 18
0,A[16:0] x B[34:17]
PREG = PREG + (0,A[16:0] x B[34:17])
48
+
48 48
P[33:17]
Clock Cycle 2
A[34:17] 18 0,B[16:0] 18
A[34:17] x 0,B[16:0]
PREG = right shifted PREG + (A[34:17] x 0,B[16:0])
48
+
48 right wire shift by 17 bits
Clock Cycle 1
0,A[16:0] 18 0,B[16:0] 18
0,A[16:0] x 0,B[16:0]
PREG = 0,A[16:0] x 0,B[16:0]
48 Zero
+
48
48
P[16:0]
ug073_c1_16_071004
Figure 2-19:
Single Slice, 35 x 35-bit Multiplier
Xilinx 37
Fully Pipelined Functional Use Models

Table 2-11 summarizes six fully pipelined functional use models. The table provides the function name and suggests what each DSP48 slice is doing. More details are provided in the following subsections. The designs are fully pipelined and run at the maximum DSP48 slice clock rate. Table 2-11: Multiple Slice Mode 35 x 18 Multiply Figure 2-20 35 x 35 Multiply Figure 2-21 Fully Pipelined DSP48 Implementations Slice 1 2 Inputs A 0,A[16:0] A[34:17] B B[17:0] B[17:0] C X X Function and OPMODE[6:0] Multiply 17-Bit Shifted Feedback Multiply Add Multiply 17-Bit Shifted Feedback Multiply Add Multiply Accumulate 17-Bit Shifted Feedback Multiply Add Multiply Multiply Accumulate Multiply Multiply Accumulate 0x05 0x65 Output P[16:0] P[52:17]
1 2
0,A[16:0] A[34:17]
0,B[16:0] 0,B[16:0]
X X
0x05 0x65
P[16:0]
3 4
0,A[16:0] A[34:17]
B[34:17] B[34:17]
X X
0x25 0x65
P[33:17] P[69:34]
18 x 18 Complex Multiply Figure 2-22 18 x 18 Complex MAC Figure 2-23 35 x 18 Complex Multiply Real Part Figure 2-26 35 x 18 Complex Multiply Imaginary Part Figure 2-27
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0] ARe[17:0] AIm[17:0]
BRe[17:0] BIm[17:0] BIm[17:0] BRe[17:0] BRe[17:0] BIm[17:0] BIm[17:0] BRe[17:0] BRe[17:0] BIm[17:0] BIm[17:0] BRe[17:0] BRe[17:0] BIm[17:0] BIm[17:0] BRe[17:0]
X X X X
0x05 0x25 0x05 0x25 0x05 0x25 0x05 0x25 0x05 0x25 0x05 0x25 0x05 0x25 0x05 0x25
P (Real) P (Imaginary) P (Real) P (Imaginary) P (Real) P (Imaginary) P (Real) P (Imaginary)
X Multiply X Multiply Accumulate X Multiply X Multiply Accumulate X Multiply X Multiply Accumulate X Multiply X Multiply Accumulate X Multiply X Multiply Accumulate X Multiply X Multiply Accumulate
38 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS Table 2-12 summarizes utilization of more complex digital filters possible using the DSP48. The small n in the Silicon Utilization column indicates the number of DSP48 filter taps. The construction and operation of complex filters is discussed in Chapter 4, MAC FIR Filters, Chapter 5, Parallel FIR Filters, and Chapter 6, Semi-Parallel FIR Filters. Table 2-12: Composite Digital Filters Silicon Utilization n DSP slices, n RAM n DSP slices n DSP slices n DSP slices n DSP slices, n RAM n DSP slices, n RAM 1 DSP slice per stage OPMODE Static Static Static Static Static Dynamic Static
Digital Filter Multi-Channel FIR Direct Form FIR Transposed Form FIR Systolic Form FIR Polyphase Interpolator Polyphase Decimator CIC Decimation/Interpolation Filters
Fully Pipelined, 35 x 18 Multiplier Use Model

The previous single slice use models show how performance and power consumption can be traded for a very small implementation (i.e., single slice). However, many DSP solutions require very high sample rates. When sample rates approach the maximum inherent clock rate for the math elements in the FPGA, it becomes necessary to design using parallel, fully pipelined math elements. With fully pipelined designs, inputs can be presented and an output computed every single clock cycle. In addition, the DSP48 slice circuits and interconnect are very carefully matched, ensuring no path becomes the timing bottleneck. Keeping math implementations mostly inside the DSP48 maximizes performance and minimizes power consumption. Of course, pipelining does have increased clock latency, but this is usually not a problem in DPS algorithms. In the single slice versions of this algorithm, partial products are computed sequentially and summed in the adder. For the fully pipelined version of the algorithm, the same partial products are computed in parallel and summed in the last slice producing a result and consuming new input operands every clock cycle. The single slice version of the 35 x 18 multiply uses two clock cycles. In each clock cycle the slice is presented with different operands, and switching the OPMODE bits modifies the behavior. The fully pipelined versions connect separate slices with fixed behavior. In the 35 x 18-bit multiply block diagram (Figure 2-20), the most-significant input data part for the 35-bit A is delayed with an extra input register in the second slice. This allows the cascading B input to be available to the second slice multiply at the same time as the most-significant data part for A. The extra register delay for the cascading B input and most-significant part of A also guarantee the output of the multiply in the second slice arrive at the same time as the partial product result from the first slice multiply.
Xilinx 39

Slice 2
A[34:17] 18 18 A[34:17] B[17:0]
+
48 48
P[52:17]
PREG2 = right shifted PREG1+ (A[34:17] B[17:0]) right wire shift by 17 bits 48 0,A[16:0] B[17:0] 48 P[16:0]
Slice 1
0,A[16:0] 18 B[17:0] 18
+
48 Zero
PREG1[16:0] = 0,A[16:0] B[17:0]
UG073_c1_17_071004
Figure 2-20:
Fully Pipelined, 35 x 18 Multiplier
Fully Pipelined, 35 x 35 Multiplier Use Model

Similar to the 35 x 18-bit example, this fully pipelined design can present inputs every clock cycle. An output is also computed every single clock cycle. Once again, no particular path becomes the timing bottleneck. The single slice version of the 35 x 35-bit multiply uses four clock cycles. In each clock cycle the slice is presented with different operands and switching the OPMODE bits modifies the behavior. The fully pipelined version connects separate slices with fixed behavior. In the single slice versions of this algorithm, partial products are computed sequentially and summed in the adder. For the fully pipelined version of the algorithm, the same partial products are computed in parallel and summed in the last slice, producing a result and consuming new input operands every clock cycle. As in the 35 x 18-bit example, there are additional register stages placed in the input paths to delay input data until the needed cascading results arrive. In Figure 2-21, the block diagram for the fully pipelined, 35 x 35 multiply shows where additional input register stages are placed. The 35 x 35-bit multiplier has additional output registers outside of the slice to align the output data. The notation Z-3 is in the external register to signify the data must be delayed by three clock cycles. If the delay is only one cycle, then registers are typically used. When the delay is larger than one, an SRL16 followed by the associated CLB flip-flop achieves maximum design performance.
40 Xilinx

Slice 4
Z-3 A[34:17] 18 18 A[34:17] B[34:17]
+
48
P[69:34] 48 PREG4 = right shifted PREG3 + (A[34:0] B[34:17])
Slice 3
0,A[16:0] B[34:17] 0,A[16:0] 18 B[34:17] 18
48 P[33:17] 48
48
PREG3 = PREG2 + (0,A[16:0] B[34:17])
Slice 2
A[34:17] 18 18
48 A[34:17] 0,B[16:0] 48
+
48 PREG2 = right shifted PREG1 + (A[34:17] 0,B[16:0]) right wire shift by 17 bits
Slice 1
0,A[16:0] 18 0,B[16:0] 18
48 0,A[16:0] 0,B[16:0] 48
Z-3 P[16:0]
+
48 PREG1 = 0,A[16:0] 0,B[16:0]
Zero Sign extended from 36 bits to 48 bits
ug073_c1_18_071004
Figure 2-21:
Fully Pipelined, 35 x 35 Multiplier
Fully Pipelined, Complex, 18 x 18 Multiplier Use Model

Complex multiplication used in many DSP applications combines operands having both real and imaginary parts into results with real and imaginary parts. Two operands A and B, each having real and imaginary parts, are combined as shown in the following equations: (A_real B_real) (A_imaginary B_imaginary) = P_real (A_real B_imaginary) + (A_imaginary B_real) = P_imaginary The real and imaginary results use the same slice configuration with the exception of the adder/subtracter. The adder/subtracter performs subtraction for the real result and addition for the imaginary result. Figure 2-22 shows several DSP48 slices used as a complex, 18-bit x 18-bit multiplier.
Xilinx 41

The two input registers to the left align operands with the first output register below and avoid fabric. The benefit is increased performance and lower power.
Slice 4
A_imag 18 B_imag 18
_
48 48
P_real
Slice 3
A_real 18 B_real 18
48
+
48
Slice 2
A_imag 18 B_real 18
+
48
48
P_imag
Slice 1
A_real 18 B_imag 18 Zero Sign extended from 36 bits to 48 bits
48
48
UG073_c1_19_071004
Figure 2-22: Pipelined, Complex, 18 x 18 Multiply Note: The real and the imaginary computations are functionally similar using different input data. The real output subtracts the multiplied terms, and the imaginary output adds the multiplied terms.
Fully Pipelined, Complex, 18 x 18 MAC Use Model

The differences between complex multiply and complex MAC implementations using several DSP48 slices is illustrated with the next set of equations. As shown, the addition and subtraction of the terms only occur after the desired number of MAC operations. For N Cycles: Slice 1 = (A_real B_imaginary) accumulation Slice 2 = (A_imaginary B_real) accumulation Slice 3 = (A_real B_real) accumulation Slice 4 = (A_imaginary B_imaginary) accumulation
42 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS Last Cycle: Slice 1 + Slice 2 = P_imaginary Slice 3 Slice 4 = P_real During the last cycle, the input data must stall while the final terms are added. To avoid having to stall the data, instead of using the complex multiply implementation shown in Figure 2-23 and Figure 2-24, use the complex multiply implementation shown in Figure 2-25.
Slice 4
A_imag 18 B_imag 18 48
48
P_real
Slice 3
A_real 18 B_real 18
48
+
48
Slice 2
A_imag 18 B_real 18
48
+
48
P_imag
Slice 1
A_real 18 B_imag 18 Sign extended from 36 bits to 48 bits 48
48
+
ug073_c1_20_071004
Figure 2-23:
Fully Pipelined, Complex, 18 x 18 MAC (N Cycles)
Xilinx 43
DSP: DESIGNING FOR OPTIMAL RESULTS In Figure 2-24, the N+1 cycle adds the accumulated products, and the input data stalls one cycle.
Slice 4
48
48
P_real
Slice 3
A_real 18 B_real 18 Zero
48
48
48
Slice 2
48
P_imag
Slice 1
48
+
48
ug073_c1_21_070904
Figure 2-24:
Fully Pipelined, Complex, 18 x 18 MAC (Last or N+1 Cycle)
44 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS An additional slice used for the accumulation is shown in Figure 2-25. The extra slice prevents the input data from stalling on the last cycle. The capability of accumulating the P cascade through the X mux feedback eliminates the pipeline stall.
Slice 6
+
48
P_real
Slice 5
A_imag 18 B_imag 18
48
_
48
Slice 4
A_real 18 B_real 18 Zero Sign extended from 36 bits to 48 bits
48
48
Slice 3
+
48
P_imag
Slice 2
A_imag 18 B_real 18
48
+
48
Slice 1
48
+
48
ug073_c1_22_070904
Figure 2-25:
Fully Pipelined, Complex, 18 x 18 MAC with Extra Slice Xilinx 45
Fully Pipelined, Complex, 35 x 18 Multiplier Usage Model

Many complex multiply algorithms require higher precision in one of the operands. The equations for combining the real and imaginary parts in complex multiplication are the same, but the larger operands must be separated into two parts and combined using partial product techniques. As shown in the other examples, the real and imaginary results use the same slice configuration with the exception of the adder/subtracter. The adder/subtracter performs subtraction for the real result and addition for the imaginary result. The following equations describe the math used to form the real and imaginary parts for the fully pipelined, complex, 35-bit x 18-bit multiplication. (A_real B_real) (A_imaginary B_imaginary) = P_real (A_real B_imaginary) + (A_imaginary B_real) = P_imaginary Figure 2-26 shows the real part of a fully pipelined, complex, 35-bit x 18-bit multiplier.
Slice 4
Z 3 18 B_imag[17:0] 18 Z 3 48 A_imag[34:17] B_imag[17:0]
A_imag[34:17]
_
48 48
P[52:17] PREG4 = PREG3 (A_imag[34:17] B_imag[17:0])
Slice 3
A_real[34:17] 18 18
A_real[34:17] B_real[17:0]
+
48 PREG3 = right shifted PREG2 + (A_real[34:17] B_real[17:0]) Right wire shift by 17 bits 48 Z 2 48 P[16:0]
Slice 2
0,A_real[16:0] 18 B_real[17:0] 18
0,A_real[16:0] B_real[17:0]
+
48 PREG2 = PREG1 + (0,A_real[16:0] B_real[17:0])
48
Slice 1
0,A_imag[16:0] 18 B_imag[17:0] 18
0,A_imag[16:0] B_imag[17:0]
_
48 Zero PREG1 = (0,A_imag[16:0] B_imag[17:0])
ug073_c1_23_071004
Figure 2-26:
Real Part of a Pipelined, Complex, 35 x 18 Multiply
46 Xilinx
XTREMEDSP DESIGN CONSIDERATIONS Figure 2-27 shows the imaginary part of a fully pipelined, complex, 35-bit x 18-bit multiplier.
Slice 4
A_imag[34:17] 18 B_real[17:0] Z 3 18 Z 3 A_imag[34:17] B_real[17:0]
48
48
P[52:17]
PREG4 = PREG3 + (A_imag[34:17] B_real[17:0])
48
Slice 3
A_real[34:17] 18
A_real[34:17] B_imag[17:0]
+
48 PREG3 = right shifted PREG2 + (A_real[34:17] B_imag[17:0]) Right wire shift by 17 bits 48 Z 2 48 P[16:0]
18
Slice 2
0,A_real[16:0] 18 B_imag[17:0] 18
0,A_real[16:0] B_imag[17:0]
+
48 PREG2 = PREG1+ 0,(A_real[16:0] B_imag[17:0])
48
Slice 1
0,A_imag[16:0] 18 B_real[17:0] 18
0,A_imag[16:0] B_real[17:0]
+
48 Zero PREG1 = (0,A_imag[16:0] B_real[17:0])
ug073_c1_24_071004
Figure 2-27:
Imaginary Part of a Pipelined, Complex, 35 x 18 Multiply
Miscellaneous Functional Use Models

Table 2-13 summarizes a few miscellaneous functional use models. Table 2-13: Miscellaneous Functional Use Models Silicon Utilization 2 DSP slices 2 DSP slices n DSP slices 2n DSP slices n DSP slices 1 DSP slice 1 DSP slice, logic 1 DSP slice, logic OPMODE Static Static Static Dynamic Dynamic Static Static Static Xilinx 47
Miscellaneous 18-bit Barrel Shifter 48-bit Add Subtract 36-bit Add Subtract Cascade n word MUX, 48 bit words n word MUX, 36 bit words 48-bit Counter Magnitude Compare Equal to Zero Compare
DSP: DESIGNING FOR OPTIMAL RESULTS Table 2-13: Miscellaneous Functional Use Models (Continued) Silicon Utilization 1 DSP slice 1 DSP slice 1 DSP slice OPMODE Static Static Static
Miscellaneous 24 2-input ANDs 24 2-input XORs Up to 48-bit AND
Dynamic, 18-bit Circular Barrel Shifter Use Model

The barrel shift function is very useful when trying to realign data very quickly. Using two DSP48 slices, an 18-bit circular barrel shifter can be implemented. This implementation shifts 18 bits of data left by the number of bit positions represented by n. The bits shifted out of the most-significant part reappear in the lower significant part of the answer completing the circular shift. The equations in Figure 2-28 describe what value is carried out of the first slice, what this value looks like after shifting right 17 bits, and finally what is visible as a result.
48 18 N zeros N zeros
Slice 1, P Carry Out
PCOUT[000..., A[17:1], ...000]

17 bits of A 48 18 N + 17 zeros
Slice 2, P Carry In after 17 bit right shift
A[000..., A[17:17 N 1]]

N MSBs of A 48 18 N zeros N MSBs of A
Slice 2, P Result
A[000..., A[17 N:0], A[17:17 N 1]]

18 N LSBs of A
ug073_c1_25_061304
Figure 2-28:
Circular Barrel Shifter Equations
Figure 2-29 shows the DSP48 used an 18-bit circular barrel shifter. The P register for slice 1 contains leading zeros in the MSBs, followed by the most-significant 17 bits of A, followed by n trailing zeros. If n equals zero, then are no trailing zeros and the P register contains leading zeros followed by 17 bits of A.
48 Xilinx

Slice 2
A[17:0] 18
+
48 48 Right wire shift by 17 bits 48
AR[17:0] PREG2 = right shifted by 17 PREG1+ n bit left shifted A[17:0]
Slice 1
A[0,17:1] 18 2n 18 Zero
+
48 PREG1 = [000..., 000..., 12 36 18 n + 1 zeros zeros A17:1, ...000] 17-bit n A zeros
ug073_c1_26_071004
Figure 2-29: Dynamic 18-Bit Barrel Shifter In the case of n equal to zero (i.e., no shift), the P register of slice 1 is passed to slice 2 with 17 bits of right shift. This leaves all 48 bits of the P carry input effectively equal to zero since A[17:1] were shifted toward the least-significant direction. If there is a positive shift amount, then P carry out of slice 1 contains A[17:1] padded in front by 48 17 n zeros and in back by n zeros. After the right shift by 17, only the n most-significant bits of A remain in the lower 48 bits of the P carry input. This n-bit guaranteed positive number is added to the A[17:0], left shifted by n bits. In the n least-significant bits there are zeros. The end result contained in A[17:0] of the second slice P register is A[17 n:n, 17:17 n + 1] or a barrel shifted A[17:0]. The design is fully pipelined and can generate a new result every clock cycle at the maximum DSP48 clock rate. A single slice version of the dynamic 18-bit barrel shifter can be implemented. For this implementation, Table 2-14 describes the DSP48 slice function and OPMODE settings for each clock cycle. Table 2-14: Miscellaneous DSP48 Implementations A A[17:0] A[17:0] A[17:0] A[17:0] Inputs B B[17:0] B[17:0] B[17:0] B[17:0] C X X X X Function and OPMODE[6:0] Multiply Multiply Accumulate Multiply Multiply Accumulate 0x05 0x25 0x05 0x25 Output P P
Single Slice Mode Cycle 18-bit Barrel Shifter 0 1 2 3
Xilinx 49
VHDL and Verilog Instantiation Templates

This section describes the VHDL and Verilog instantiation templates. In VHDL, each template has a component declaration section and an architecture section. Insert each part of the template within the VHDL design file. The port map of the architecture section should include the design signal names.
VHDL Instantiation Template

-- DSP48 design, -- VHDL placed -- instance -- declaration -- code reference and ---- Library use -- declaration -- for library -- Xilinx -- primitives used -: In order to incorporate this function into the : the following instance declaration needs to be : in the body of the design code. The instance name : (DSP48_inst) and/or the port declarations after the : => declaration maybe changed to properly : connect this function to the design. All inputs : and outputs must be connected. : In addition to adding the instance declaration, a : statement for the UNISIM.vcomponents library needs : to be added before the entity declaration. This : contains the component declarations for all Xilinx : primitives and points to the models that will be : for simulation.
-- Copy the following two statements and paste them before the -- Entity declaration, unless they already exists. Library UNISIM; use UNISIM.vcomponents.all; -- <-----Cut code below this line and paste into the architecture body----> -- DSP48: DSP Function Block -Virtex-4 -- Xilinx HDL Language Template version 6.1i DSP48_inst: DSP48 generic map ( AREG => 1, -- Number of pipeline registers on the A input, 0, 1 or 2 BREG => 1, -- Number of pipeline registers on the B input, 0, 1 or 2 B_INPUT => DIRECT, -- B input DIRECT from fabric or CASCADE from another DSP48 CARRYINREG => 1, -- Number of pipeline registers for the CARRYIN input, 0 or 1 CARRYINSELREG => 1, -- Number of pipeline registers for the CARRYINSEL, 0 or 1 CREG => 1, -- Number of pipeline registers on the C input, 0 or 1 LEGACY_MODE => MULT18X18S, -- Backward compatibility, NONE, -- MULT18X18 or MULT18X18S MREG => 1, -- Number of multiplier pipeline registers, 0 or 1 OPMODEREG => 1, -- Number of pipeline registers on OPMODE input, 0 or 1 PREG => 1, -- Number of pipeline registers on the P output, 0 or 1 SIM_X_INPUT => GENERATE_X_ONLY, -- Simulation parameter for behavior for X on input. -- Possible values: GENERATE_X, NONE or WARNING SUBTRACTREG => 1) -- Number of pipeline registers on the SUBTRACT input, 0 or 1 port map ( BCOUT => BCOUT, -- 18-bit B cascade output
50 Xilinx

P => P, -- 48-bit product output PCOUT => PCOUT, -- 38-bit cascade output A => A, -- 18-bit A data input B => B, -- 18-bit B data input BCIN => BCIN, -- 18-bit B cascade input C => C, -- 48-bit cascade input CARRYIN => CARRYIN, -- Carry input signal CARRYINSEL => CARRYINSEL, -- 2-bit carry input select CEA => CEA, -- A data clock enable input CEB => CEB, -- B data clock enable input CEC => CEC, -- C data clock enable input CECARRYIN => CECARRYIN, -- CARRYIN clock enable input CECINSUB => CECINSUB, -- CINSUB clock enable input CECTRL => CECTRL, -- Clock Enable input for CTRL registers CEM => CEM, -- Clock Enable input for multiplier registers CEP => CEP, -- Clock Enable input for P registers CLK => CLK, -- Clock input OPMODE => OPMODE, -- 7-bit operation mode input PCIN => PCIN, -- 48-bit PCIN input RSTA => RSTA, -- Reset input for A pipeline registers RSTB => RSTB, -- Reset input for B pipeline registers RSTC => RSTC, -- Reset input for C pipeline registers RSTCARRYIN => RSTCARRYIN, -- Reset input for CARRYIN registers RSTCTRL => RSTCTRL, -- Reset input for CTRL registers RSTM => RSTM, -- Reset input for multiplier registers RSTP => RSTP, -- Reset input for P pipeline registers SUBTRACT => SUBTRACT -- SUBTRACT input ); -- End of DSP48_inst instantiation
Verilog Instantiation Template

The following is a synthesis instantiation template for the DSP48 slice in Verilog. After the port list are synthesis attributes with syntax written for the Xilinx Synthesis Tool (XST). If using a different synthesis tool, consult the tools user guide and change the attributes appropriately. The section after the synthesis attributes consists of defparam statements that are ignored by the synthesis process, but are used to initialize the simulation model to match the synthesis attributes during simulation.
// // // // // // // DSP48 Verilog instance declaration code : : : : : : : In order to incorporate this function into the design, the following instance declaration needs to be placed in the body of the design code. The instance name (DSP48_inst) and/or the port declarations within the parenthesis maybe changed to properly reference and connect this function to the design. All inputs and outputs must be connected.
// <-----Cut code below this line----> // DSP48: DSP Function Block // Virtex-4 // Xilinx HDL Language Template version 7.1i DSP48 DSP48_inst ( .BCOUT(BCOUT), // 18-bit B cascade output .P(P), // 48-bit product output .PCOUT(PCOUT), // 38-bit cascade output .A(A), // 18-bit A data input .B(B), // 18-bit B data input .BCIN(BCIN), // 18-bit B cascade input .C(C), // 48-bit cascade input .CARRYIN(CARRYIN), // Carry input signal .CARRYINSEL(CARRYINSEL), // 2-bit carry input select .CEA(CEA), // A data clock enable input
Xilinx 51

.CEB(CEB), .CEC(CEC), .CECARRYIN(CECARRYIN), .CECINSUB(CECINSUB), .CECTRL(CECTRL), .CEM(CEM), registers .CEP(CEP), .CLK(CLK), .OPMODE(OPMODE), .PCIN(PCIN), .RSTA(RSTA), .RSTB(RSTB), .RSTC(RSTC), .RSTCARRYIN(RSTCARRYIN), .RSTCTRL(RSTCTRL), .RSTM(RSTM), .RSTP(RSTP), .SUBTRACT(SUBTRACT) ); // // // // // // // // // // // // // // // // // // B data clock enable input C data clock enable input CARRYIN clock enable input CINSUB clock enable input Clock Enable input for CTRL registers Clock Enable input for multiplier Clock Enable input for P registers Clock input 7-bit operation mode input 48-bit PCIN input Reset input for A pipeline registers Reset input for B pipeline registers Reset input for C pipeline registers Reset input for CARRYIN registers Reset input for CTRL registers Reset input for multiplier registers Reset input for P pipeline registers SUBTRACT input
// The following defparams specify the behavior of the DSP48 slice. // If the instance name to the DSP48 is changed, that change needs to // be reflected in the defparam statements. defparam input, 0, 1 defparam input, 0, 1 defparam DSP48_inst.AREG = 1; // Number of pipeline registers on the A or 2 DSP48_inst.BREG = 1; // Number of pipeline registers on the B or 2 DSP48_inst.B_INPUT = DIRECT; // B input DIRECT from fabric // or CASCADE from another // Number of pipeline // for the CARRYIN input, 0 or 1 defparam DSP48_inst.CARRYINSELREG = 1; registers for the // Number of pipeline
DSP48 defparam DSP48_inst.CARRYINREG = 1; registers
// CARRYINSEL, 0 or 1 defparam DSP48_inst.CREG = 1; // Number of pipeline registers on the C input, 0 or 1 defparam DSP48_inst.LEGACY_MODE = MULT18X18S; // Backward compatibility, // NONE, MULT18X18 or MULT18X18S defparam DSP48_inst.MREG = 1; // Number of multiplier pipeline registers, 0 or 1 defparam DSP48_inst.OPMODEREG = 1; // Number of pipeline registers on // OPMODE input, 0 or 1 defparam DSP48_inst.PREG = 1; // Number of pipeline registers on the P output, 0 or 1 defparam DSP48_inst.SIM_X_INPUT = GENERATE_X_ONLY; // Simulation parameter for behavior // for X on input. Possible values: // GENERATE_X, NONE or WARNING defparam DSP48_inst.SUBTRACTREG = 1; // Number of pipeline registers // on the SUBTRACT input, 0 or 1 // End of DSP48_inst instantiation
52 Xilinx
Chapter 3
DSP48 Slice Math Functions
The DSP48 slice efficiently performs a wide range of basic math functions, including adders, subtracters, accumulators, MACs, multiply multiplexers, counters, dividers, square-root functions, and shifters. The optional pipeline stage within the DSP48 tile ensures high performance arithmetic functions. The DSP48 column structure and associated routing provides fast routing between DSP48 tiles with less routing congestion to the FPGA fabric. This chapter describes how to use the DSP48 slice to perform some basic arithmetic functions.
Overview
The DSP48 slice is shown in Figure 3-1. Refer to Figure 2-3, page 13 for a diagram showing two slices cascaded together.
Cascade Out to Next Slice BCOUT 18 18 A 18 B 18 18 36 X 36 48 CIN 48 48 PCOUT
72 36 48 Y 48
SUBTRACT
ZERO 48
48 Z 48
48 18 Wire Shift Right by 17 Bits Cascade In from Previous Slice
48 48 PCIN
ug073_c2_01_061304
BCIN
Figure 3-1:
DSP Slice Architecture
Xilinx 53
Basic Math Functions

Add/Subtract
The DSP48 slice contains an adder/subtracter unit allowing different combinations of add/subtract logic to be implemented in a single DSP slice. The output of the DSP48 slice in adder/subtracter mode is: Output = Z (X + Y +CIN) The X, Y, and Z terms in this equation refer to the X, Y, and Z multiplexers shown in Figure 3-1. The inputs to the X, Y, and Z multiplexers are routed to the outputs using OPMODE settings as shown in Table 3-1. The CIN term is the Carry Input to the Adder/subtracter unit. Determining whether an addition or a subtraction () takes place is controlled by the SUBTRACT input to the adder/subtracter unit. The SUBTRACT input must be set to 0 to add, and 1 to subtract. Table 3-1: Z 0 PCin P C ShiftPCin ShiftP OPMODE Settings for the Z, Y, and X Multiplexers OPMODE[6:4] 000 001 010 011 101 110 Y 0 AxB Illegal C OPMODE[3:2] 00 01 10 11 X 0 AxB P A:B OPMODE[1:0] 00 01 10 11
Notes: 1. If one of X or Y is set to 01, the other one must also be set to 01.
2. For Carryin Select (CIN) see Carry Input Logic in Chapter 2.
The Verilog code for this 48-bit adder is in the reference design file: ADDSUB48.v, and the VHDL code is in the reference design file: ADDSUB48.vdh. This code can be used to implement any data combination for this equation by using the different OPMODEs found in Table 3-1.
Accumulate
A DSP48 slice can implement add and accumulate functions with up to 36-bit inputs. The output equation of the accumulator is: Output = Output + A:B + C Concatenate (:) the A and B inputs to provide a 36-bit input from Multiplexer X using the setting OPMODE[1:0] = 0b11. Select the C input to Multiplexer Y using the setting OPMODE[3:2] = 0b11. To add (accumulate) the output of the slice, select the feedback path (P) through the Z multiplexer using the setting OPMODE[6:4] = 0b010. Other accumulate functions can be implemented by changing the OPMODE selection for the Z input multiplexer. To get an output of: Output = Shift(P) (A:B + C) 54 Xilinx
DSP48 SLICE MATH FUNCTIONS use the setting OPMODE[6:4] = 0b110 to select the Shift(P) input to the Z multiplexer. To get an output of: Output = 0 (A:B +C) (no accumulation) use the setting OPMODE [6:4] = 0b0000 to select the ZERO input to the Z multiplexer. The Verilog code for the accumulator is in the reference design file ACCUM48.v, and the VHDL code is in the reference design file ACCUM48.vhd.
Multiply Accumulate (MAC)

The DSP48 slice allows two 18-bit numbers to be multiplied together, and the product to be added to or subtracted from a previous product, a 0, or a shifted product. In addition, rounding of any of the add, subtract, previous product, 0, or shifted product options is also possible. The input added or subtracted from the product is from the output of the Z multiplexer. This output is set using the corresponding OPMODE setting as shown in Table 3-1. Cascade the MAC tree by selecting the PCIN signal from the previous slice as the output from the Z multiplexer. The Verilog code for the multiply-accumulate function is in the reference design file MAC.v, and the VHDL code is in the reference design file MAC.vhd.
Multiplexer
There are three multiplexers in a DSP48 slice: the 3:1 Y multiplexer, the 4:1 X multiplexer, and the 6:1 Z multiplexer. Only one multiplexer is active to use the slice as a pure multiplexer. Make the other two multiplexers inactive by choosing the OPMODE selecting the ZERO inputs. The two DSP48 tiles in a slice can be combined to make wider input multiplexers.
Barrel Shifter
An 18-bit barrel shifter can be implemented using the two DSP48 tiles in the DSP slice. To barrel shift the 18-bit number A[17:0] two positions to the left, the output from the barrel shifter is A[15:0], A[17], and A[16]. This operation is implemented as follows. The first DSP48 is used to multiply {0,A[17:1]} by 22. The output of this DSP48 tile is now {0,A[17:1],0,0}. The output from the first tile is fed into the second DSP48 tile over the PCIN/PCOUT signals, and is passed through the 17-bit right-shifted input. The input to the Z multiplexer becomes {0,A[17],A[16]}, or {0,A[17:0],0,0} shifted right by 17 bits. The multiplier inputs to the second DSP48 tile are A = A[17:0] and B = 22. The output of this multiplier is {A[17:0], 0,0}. This output is added to the 17-bit right-shifted value of {0,A[17],A[16]} coming from the previous slice. The 18-bit output of the adder is {A[15:0],A[17],A[16]}. This is the initial A input shifted by two to the left. The Verilog code is in the reference design file barrelshifter_18bit.v, and the VHDL code is in the reference design file barrelshifter_18bit.vhd).
Counter
The DSP48 slice can be used as a counter to count up by one on each clock cycle. Setting the SUBTRACT input to 0, the carry-in input (CIN) to 1, and OPMODE [6:0] = 0b0100000 gives an output of P + CIN. After the first clock, the output P is 0 + 1 = 1. Subsequent outputs are P + 1. Xilinx 55
DSP: DESIGNING FOR OPTIMAL RESULTS This method is equivalent to counting up by one. The counter can be used as a down counter by setting the SUBTRACT input to a 1 at the start. The counter can also be preloaded using the C input to provide the preload value. Setting the Carry In input (CIN) to 1 and OPMODE [6:4] = 0b0110000 gives an output of P = C+1 in the first cycle. For subsequent clocks, set the OPMODE to select P = P+1 by changing OPMODE [6:4] from 0b0110000 to 0b0100000. The Verilog code for a loadable counter is in the reference design file CNTR_LOAD.v, and the VHDL code for a loadable counter is in the reference design file CNTR_LOAD.vhd.
Multiply
A single DSP48 slice can implement an 18x18 signed or unsigned multiplier. Larger multipliers can be implemented in a single DSP48 slice by sequentially shifting the appropriate number of bits in each clock cycle. The Verilog implementation of an 18x18 multiplier is in the reference design file MULT18X18_PARALLEL.v, and the VHDL implementation is in the reference design file MULT18X18_PARALLEL.vhd. The Verilog implementation of a 35x35 multiplier and a sequential 35x35 multiplier are in the reference design files MULT35X35_PIPE.v and MULT35X35_SEQUENTIAL_PIPE.v respectively. The VHDL implementation of a 35x35 multiplier and a sequential 35x35 multiplier are in the reference design files MULT35X35_PIPE.vhd and MULT35x35_SEQUENTIAL_PIPE.vhd, respectively.
Divide
Binary division can be implemented in the DSP48 slice by performing a shift and subtract or a multiply and subtract. The DSP48 slice includes a shifter, a multiplier, and adder/subtracter unit to implement binary division. The division by subtraction and division by multiplication algorithms are shown below. These algorithms assume: 1. N > D 2. N and D are both positive If either N or D is negative, use the same algorithms by taking the absolute positive values for N and D and making the appropriate sign change in the result. The terms N and D in the algorithms refer to the number to be divided (N) and the divisor (D). The terms Q and R in the algorithms refer to the quotient and remainder, respectively.
Dividing with Subtraction

The shift and subtract algorithm can be explained as follows: If N is an 8-bit integer and D is not more than 8 bits wide, N/D = Q + R 1. Assign the 8-bit register R the value 00000000. 2. Shift the R register one bit to the left and fill in the LSB with N[8-n]. 3. Calculate R-D. 4. Set R and set Q: a. If R-D is positive, set Q[8-n] to 1 and R = R-D b. 5. If R-D is negative, set Q[0] to 0 and R = R Repeat Steps 2 to 4, filling in R[n] each time with N[8-n], where n is the number of the iteration. Q[8-n] is filled each time in Step 4.
56 Xilinx
DSP48 SLICE MATH FUNCTIONS After the eighth iteration, Q[7:0] contains the quotient, and R[7:0] contains the remainder. For example: N 8 0000, 1000 --- = --- = ------------------------ = Q(10) + R(10) D 3 011 Step 1 2 3 4 2 3 4 2 3 4 2 3 4 2 3 4 2 3 4 2 3 4 2 3 Iteration (n) 1 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 6 6 6 7 7 7 8 8 8 R = 0000,0000 R <-- N[7] = 0000,0000 R-D = Negative Q[7] = 0 R <-- N[6] = 0000,0000 R-D = Negative Q[6] = 0 R <-- N[5] = 0000,0000 R-D = Negative Q[5] = 0 R <-- N[4] = 0000,0000 R-D = Negative Q[4] = 0 R <-- N[3] = 0000,0001 R-D = Negative Q[3] = 0 R <-- N[2] = 0000,0010 R-D = Negative Q[2] = 0 R <-- N[1] = 0000,0100 R-D = Positive Q[1] = 1, R = 0000,0001 R <-- N[0] = 0000,0010 R-D = Negative Q[2] = 0 Action After Action Q xxxx,xxxx xxxx,xxxx xxxx,xxxx 0xxx,xxxx 0xxx,xxxx 0xxx,xxxx 00xx,xxxx 00xx,xxxx 00xx,xxxx 000x,xxxx 000x,xxxx 000x,xxxx 0000,xxxx 0000,xxxx 0000,xxxx 0000,0xxx 0000,0xxx 0000,0xxx 0000,00xx 0000,00xx 0000,00xx 0000,001x 0000,001x 0000,001x 0000,0010 R 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0000 0000,0001 0000,0001 0000,0001 0000,0010 0000,0010 0000,0010 0000,0100 0000,0100 0000,0001 0000,0010 0000,0010 0000,0010
Xilinx 57
Dividing with Multiplication

The multiply and subtract method consists of rewriting N/D = Q + R as N= D * (Q + R). The answer is calculated using the following steps for an 8-bit N/D: 1. Set the initial value of Q[8-n] =1 and the bits right of Q[8-n] to 0. 2. Calculate D*Q. 3. Calculate N- (D*Q). a. If step 2 is positive, N > (D*Q), set Q[8-n] to a 1. b. 4. If step 2 is negative, N < (D*Q), set Q[8-n] to a 0. Repeat steps 1 to 3. After the eighth iteration, Q[7:0] contains the quotient and N - (D*Q) contains the remainder. Using the same example: 8 0000, 1000 --- = ------------------------ = Q(10) + R(10) 3 011 Step 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 Iteration (n) 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 6 6 6 7 7 7 Action Q[8-1] = 1, Set the bits right of Q[8-1] to 0 D*Q = 3 * 128 = 384 N - (D*Q) = 8 - 384 = Negative Q[8-1] = 0 Q[8-2] = 1, Set the bits right of Q[8-2] to 0 D*Q = 3 * 64 = 192 N - (D*Q) = 8 - 192 = Negative Q[8-2] = 0 Q[8-3] = 1, Set the bits right of Q[8-3] to 0 D*Q = 3 * 32 = 96 N - (D*Q) = 8 - 96 = Negative Q[8-3] = 0 Q[8-4] = 1, Set the bits right of Q[8-4] to 0 D*Q = 3 * 16 = 48 N - (D*Q) = 8 - 48 = Negative Q[8-4] = 0 Q[8-5] = 1, Set the bits right of Q[8-5] to 0 D*Q = 3 * 8 = 24 N - (D*Q) = 8 - 24 = Negative Q[8-5] = 0 Q[8-6] = 1, Set the bits right of Q[8-6] to 0 D*Q = 3 * 4 = 12 N - (D*Q) = 8 - 12 = Negative Q[8-6] = 0 Q[8-7] = 1, Set the bits right of Q[8-7] to 0 D*Q = 3 * 2 = 6 N - (D*Q) = 8 - 6 = Positive Q[8-7] = 1 After Action Q 1000,0000 1000,0000 0000,0000 0100,0000 0100,0000 0000,0000 0010,0000 0010,0000 0000,0000 0001,0000 0001,0000 0000,0000 0000,1000 0000,1000 0000,0000 0000,0100 0000,0100 0000,0000 0000,0010 0000,0010 0000,0010
58 Xilinx
DSP48 SLICE MATH FUNCTIONS Step 1 2 3 Iteration (n) 8 8 8 Action Q[8-8] = 1 D*Q = 3 * 3 = 9 N - (D*Q) = 8 9 = Negative Q[8-8] = 0 Remainder = N-(D*Q) = 8-(3*2) = 2 After Action Q 0000, 0011 0000,0011 0000,0010
Both of the division implementations are possible in one DSP48 slice. The slice usage for 8-bit division is one DSP48, and the latency is eight clock cycles. The Verilog code for the Divide by Subtraction implementation is in the reference design file DIV_SUB.v, and the VHDL code is in the reference design file DIV_SUB.vhd. The Verilog code for the Divide by Multiplication implementation is in DIV_MULT.v and the VHDL code for the second implementation is in DIV_MULT.vhd.
Square Root
The square root of an integer number can be calculated by successive multiplication and subtraction. This is similar to the subtraction method used to divide two numbers. The square root of an N-bit number will have N/2 (rounded up) bits. If the square root is a fractional number, N/2 clocks are needed for the integer part of the answer, and every following clock gives one bit of the fraction part. The logic needed to compute this is shown in Figure 3-2.
1'b1 Register A Multiplier 1'b0
Register A Input
Subtractor
Input = Reg C
UG073_c2_02_061304
Figure 3-2:
Square Root Logic
The square root for an 8-bit number can be calculated as follows: X = Y.Z Y is the integer part of the root, and Z is the fraction part. Register A refers to the registers found on the A input to the DSP48 slice, and Register C refers to the registers found on the C input to the DSP48 slice 1. Read the number into Register C. Set Register A to 8b10000000. 2. Calculate Register C (Register A * Register A). 3. If step 2 is positive, set Register A[(8-clock)] = 1, Register A[(8-clock)-1] = 1 Xilinx 59
DSP: DESIGNING FOR OPTIMAL RESULTS If step 2 is negative, set 4. Register A[(8-clock)] = 0, Register A[(8-clock)-1] = 1
Repeat steps 1 to 3. Four clocks are required to calculate the integer part of the value (Y). The number of clocks required for the fraction part (Z) depends on the precision required. For an 8-bit input value, the value in Reg_A after eight clocks includes the integer part given by the four MSBs and the fractional part given by the four LSBs. For example, find the square root of 11 decimal = 3.3166. Because 11 decimal is a 4-bit binary number, the integer part is two bits wide and is obtained in two clock cycles. The bit width of the fractional part depends on the precision required. In this example, four bits of precision are used requiring four clock cycles. The binary of value of 11 decimal is 1011. Expressed as an 8-bit number, it becomes 0000,1011. Store this value as 0000,1011,0000,0000. The last eight bits are necessary because the result is an 8-bit number, and 8 bits * 8 bits gives a 16-bit multiplication result. Clock 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5 6 6 6 7 7 7 8 60 Xilinx Step 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 Register A = 1000,0000 0000,1011,0000,0000 (1000,0000 * 1000,0000) Step 2 is negative. Set Register A to 0100,0000 Register A = 0100,0000 0000,1011,0000,0000 (0100,0000 * 0100,0000) Step 2 is negative. Set Register A to 0010,0000 Register A = 0010,0000 0000,1011,0000,0000 (0010,0000 * 0010,0000) Step 2 is positive. Set Register A to 0011,0000 Register A = 0011,0000 0000,1011,0000,0000 (0011,0000* 0011,0000) Step 2 is positive. Set Register A to 0011,1000 Register A = 0011,1000 0000,1011,0000,0000 (0011,1000* 0011,1000) Step 2 is negative. Set Register A to 0011,0100 Register A = 0011,0100 0000,1011,0000,0000 (0011,0100* 0011,0100) Step 2 is positive. Set Register A to 0011,0110 Register A = 0011,0110 0000,1011,0000,0000 (0011,0110* 0011,0110) Step 2 is negative. Set Register A to 0011,0101 Register A = 0011,0101 Action
DSP48 SLICE MATH FUNCTIONS Clock 8 8 Step 2 3 Step 2 is positive. Action 0000,1011,0000,0000 (0011,0101* 0011,0101)
The output is in Register A and is 0011,0101. The final answer is 11.0101. The Verilog code for this implementation (8-bit input, 8 clocks) is in SQRT.v, and the VHDL code is in SQRT.vhd.
Square Root of the Sum of Squares

The sum of squares is a widely used DSP function. The sum of squares can be either of the forms listed in Equation 3-1 or Equation 3-2.
I = n1
SoS = A + B
or
SoS =
i=0
Ai
Equation 3-1
I = n1
SoS = A + B
or
SoS =
i=0
Ai
Equation 3-2 These functions are basic multiply-accumulate operations easily implemented on the DSP48 slice as described in Multiply Accumulate (MAC), page 55. A variation of this function is when the square root of either of the above equations is needed. In this case, the OPMODE does the MAC function for n cycles and then switches to do the square root function for the next n cycles. The Subtract input is dynamic and does an add for the MAC cycles and a subtract for the square root cycles. With the SUBTRACT input equal to 0, the OPMODE for the function is 0110101. A square root function is implemented by changing the SUBTRACT input to a 1. Xilinx 61
Conclusion
The DSP48 slice has a variety of features for fast and easy implementation of many basic math functions. The dedicated routing region around the DSP48 slice and the feedback paths provided in each slice result routing improvements. The high-speed multiplier and adder/subtracter unit in the slice delivers high-speed math functions.
62 Xilinx
Chapter 4
MAC FIR Filters
This chapter describes the implementation of a Multiply-Accumulate (MAC) Finite Impulse Response (FIR) filter using the DSP48 slice in a Virtex-4 device. Because the Virtex-4 architecture is flexible, constructing FIR filters for specific application requirements is practical. Creating optimized filter structures of a sequential nature saves resources and potential clock cycles. This chapter demonstrates two sequential filter architectures: the single-multiplier and the dualmultiplier MAC FIR filter. Reference design files are available for the System Generator in DSP, VHDL, and Verilog. These reference designs permit filter parameter changes including coefficients and the number of taps.
Overview
A large array of filtering techniques is available to signal processing engineers. A common filter implementation uses the single multiplier MAC FIR filter. In the past, this structure used the Virtex-II embedded multipliers and 18K block RAMs. The Virtex- 4 DSP48 slice contains higher performance multiplication and arithmetic capabilities specifically designed to enhance the use of MAC FIR filters in FPGA-based Digital Signal Processing (DSP).
Single-Multiplier MAC FIR Filter

The single-multiplier MAC FIR is one of the simplest DSP filter structures. The MAC structure uses a single multiplier with an accumulator to implement a FIR filter sequentially versus a full parallel FIR filter. This trade-off reduces hardware by a factor of N, but also reduces filter throughput by the same factor. The general FIR filter equation is a summation of products (also know as an inner product), defined as:
N1
yn =
xn i hi
Equation 4-1
i=0 In this equation, a set of N coefficients is multiplied by N respective data samples, and the inner products are summed together to form an individual result. The values of the coefficients determine the characteristics of the filter (e.g., low-pass filter, band-pass filter, high-pass filter). The equation can be mapped to many different implementations (e.g., sequential, semi-parallel, or parallel) in the different available architectures. Xilinx 63
DSP: DESIGNING FOR OPTIMAL RESULTS For slow sample rate requirements and a large number of coefficients, the single MAC FIR filter is well suited and dual-port block RAM is the optimal choice for the memory buffer. This structure is illustrated in Figure 4-1. If the number of coefficients is small, distributed memory and the SRL16E can be used as the data and coefficient buffers. For more information on using distributed memory, refer to Using Distributed RAM for Data and Coefficient Buffers, page 70.
Dual-Port Block RAM 18 Data Samples 96 x 18 Data Addr A P B load Optional Output Register Used Z-4
UG073_c3_02_081804
DSP48 Slice OPMODE = 0100101
18
Control
WE
Coefficients 96 x 18
Coef Addr
Figure 4-1:
Single-Multiplier MAC FIR Filter
The input data buffer is implemented in dual-port block RAM. The read address port is clocked N times faster than the input samples are written into the data port, where N is the number of filter taps. The filter coefficients are also stored in the same dual-port block RAM, and are output at port B. Hence, the RAM is used in a mixed-mode configuration. The data is written and read from port A (RAM mode), and the coefficients are read only from port B (ROM mode). The control logic provides the necessary address logic for the dual-port block RAM and creates a cyclic RAM buffer for port A (data buffer) to create the FIR filter delay line. An optional output capture register maybe required for streaming operation, if the accumulation result can not be immediately used in downstream processing. The multiplier followed by the accumulator sums the products over the same number of cycles as there are coefficients. With this relationship, the performance of the MAC FIR filter is calculated by the following equation: M a x i m u m I n p u t S a m pl e R a t e = C l o c k S p ee d / N u mb e r o f Ta ps Equation 4-2
If the coefficients possess a symmetric shape, a slightly costlier structure is available (see Symmetric MAC FIR Filter, page 72), however, the maximum sampled rate is doubled. The sample rate of the costlier structure is defined as follows: S am p l e R at e = C l o c k S p e e d / ( 1 /2 x nu m b e r o f t a p s ) Equation 4-3
64 Xilinx
MAC FIR FILTERS
Bit Growth
The nature of the FIR filter, with numerous multiplies and adds, outputs a larger number of bits from the filter than are present on the filters input. This effect is the "bit growth" or the "gain" of a filter. These larger results cannot be maintained throughout a system due to cost implications. Therefore, the full precision result is typically rounded and quantized (refer to Rounding, page 69) back to a desired level. However, it is important to calculate the full precision output in order to select the correct bits from the output of the MAC. A simple explanation for implementation purposes involves considering the maximum value expected at the output (saturation level). A greater understanding of the specific filter enhances the accuracy of the output bit width. The following two techniques help determine the full precision output bit width.
Generic Saturation Level

This technique assumes every value in the filter could be the worst possible for the size of the twos complement numbers specified. Using the generic saturation level is a good starting point when the coefficients are unknown, but the number of bits required to represent them is known. For example, if the coefficients are reloadable, as in adaptive filters. O u t pu t W i dt h = c e i l ( l o g 2 ( 2 ( b - 1 ) x 2 (c -1) x N ) + 1 where: ceil: Rounds up to the nearest integer b: Number of bits in the data samples c: Number of bits in the coefficients Equation 4-4
Coefficient Specific Saturation Level

This technique uses the magnitude-only sum of actual coefficient values and applies the worst-case data samples to the filter. More accurate calculations could be required if a bit maximum is reached. With actual coefficients, the output for the worst possible inputs can be determined. O u t p u t W i d th = c e i l ( l og 2 ( 2 ( b - 1 ) x a b s ( s u m ( c oe f )) x N ) + 1 Equation 4-5
where: ceil: Rounds up to the nearest integer abs: Makes the absolute value of a number (not negative) sum: Sums all the values in an array B: Number of bits in the data samples C: Number of bits in the coefficients If the output width exceeds 48 bits, there are notable effects on the size (in terms of the number of DSP48 slices used to implement the filter), because the DSP48 slice is limited to a 48-bit result. The output width can be extended by using more DSP48 slices, however, reconsidering the specification is more practical.
Control Logic
The control logic is very straightforward when using an SRL16E for the data buffer. For dual-port block RAM implementations the cyclic RAM buffer is required. This can complicate the control logic, and there are two different ways this control can be implemented. Both techniques produce the same results, but one way uses all slice-based logic to produce the results, while the other way embeds Xilinx 65
DSP: DESIGNING FOR OPTIMAL RESULTS the control in the available space in the Block RAM. The basic architecture of the control logic for the slice based approach is outlined in Figure 4-2.
en
addr Data Counter
Data Address WE A = 2N 2 Load ACC
en
addr Coefficient Counter
Coefficient Address
UG073_c3_03_090204
Figure 4-2:
Dual-Port Block RAM MAC FIR Filter Control Logic Using Slices
The control logic consists of two counters. One counter drives the address of the coefficient section of the dual-port block RAM, while the other controls the address for the data buffer. A comparator controls an enable to the data buffer counter to disable the count for one cycle every output sample, and writes a new sample into the data buffer every N cycles. A simplified diagram of the control logic and the memory is shown in Figure 4-3.
Dual-Port RAM Counter 0 (N1) en addr DIN A WE data addr DOUT A
RAM must be: Read after Write A = 2N2
Counter N (2N1) addr
DIN B WE coef addr
Coefficient ROM
DOUT B
DIN Data Addr Coeff Addr
D1 X 0 1
X 2
X 3
X D2 X X 1
X 2
X D3 X
X 1
94 95 95 0
93 94 94 95 0
96 97 98 99
190 191 96 97 98 99
190 191 96 97 98 99
WE
WE
UG073_c3_04_090204
Figure 4-3:
Control Logic and Memory
66 Xilinx
MAC FIR FILTERS The cyclic data RAM buffer is required to emulate the delay line shift register of the FIR filter while using a static RAM. The RAM is addressed sequentially every clock cycle. The counter rolls over to have the last coefficient (N1) read out. At this point, the data buffer is stalled by the controlling clock enable and the newest sample is read into the buffer AFTER the oldest data sample is read out. This newest data sample is now multiplied by the first coefficient (as the coefficient address counter is never disabled) and the cycle is repeated. The effect is of data shifting over time as the FIR filter equation requires. The ability to perform a simultaneous read and write requires the RAM buffer to have a read port and a write port (called read before write mode). The inverted WE signal is also used to drive the load input (OPMODE[5]) on the DSP48 slice. This signal must be delayed with a simple SRL16E to make sure the latency on the signal matches the latency through the MAC engine. This delay is typically four clocks, but depends upon the number of pipelining registers used in the DSP48 slice and block RAM. The number of required pipelining stages is a function of the desired achievable clock frequency. The number of resources used for the control logic is easily calculated. The counters are always two bits per slice plus the additional logic required to count limit the counter (unless the counter is a power of two limit). The count limiter circuit size is determined by the number of bits needed to represent the count limit value divided by four. Therefore, n/2 + n/4 slices are required for each counter, but the coefficient counter is larger due to the higher count value. The other control logic typically yields about N/4 slices due to the comparator required for the enable circuitry and the inverter to disable the data counter. The total number of slices for the control logic for an 18 x 18 MAC FIR filter with 96 coefficients is listed in Table 4-1. Table 4-1: Control Logic Using Slices Resource Utilization Elements Coefficient Counter Data Counter Relational Operator Other Logic Total Slices 5 4 1 1 11
Embedding the Control Logic into the Block RAM

The total number of slices for the control logic can be reduced if required by embedding the coefficient address sequence, CE, WE, and capture signal into the leftover block RAM space. This clever trick is enabled by the separation of the DSP48 from the Virtex-4 block RAM. It is different from Virtex-II Pro FPGAs where the embedded multiplier and block RAM share routing and prevent this kind of
Xilinx 67
DSP: DESIGNING FOR OPTIMAL RESULTS trick. Figure 4-4 illustrates the control logic and memory layout for this embedded control logic implementation.
coef addr
coef addr(0) 0 DIN A WE_A
29 28 27 26 ... 18 17 ... 0 WE CE load coef addr coef (0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . coef (N) .
Load ACC WE_B DOUT A CE coef addr 18 Coefficients RAM must be: - Read before Write - Output Register On
18
DIN B WE DOUT B 18
en
addr
data addr
Counter N -> (2N1)

DIN D1 X Coeff Addr 0 Data Addr 1 X 2 X 3 X
Dual-Port RAM
X D2 X X 0 1 2 X 3 X X D3 X 0 1 X 2 X 3 94 95 94 95
96 97 98 99
190 191 191 96 97 98
189 190 190 191 96 97
WE
WE
UG073_c3_05_090204
Figure 4-4:
Control Logic Using Embedded Control Technique
Figure 4-4 demonstrates how the predictable and repeatable control sequence for the coefficient side of the memory can be embedded into the remaining space of the memory. The coefficient address value, accumulator Load signal, CE, and WE for the data buffer are precalculated and concatenated on to the coefficient values. The memory must be used in 512 x 36 mode, instead of 1024 x 18 mode. The individual signals are split up correctly on the output of the memory. This costs nothing in logic utilization apart from routing. Due to the feedback nature of the address line, it is important to set the initial state of the dualport block RAMs output register to effectively kick- start the MAC process. The initial values need to be different from each other to start the correct addressing, however, the silicon forces them to be the same. This changes the 1-bit masking of the LSB of the coefficient address such that the first value is 0 instead of the initialized value of 1. The initial value of the output latch is on the address bus the next cycle and, by unmasking the LSB, the count is successfully kick-started. Because the coefficients are placed in the upper half of the memory, only a single LSB must be masked, not the complete address bus. The masking signal can take the form of a reset signal or a registered permanent value to get the required single cycle mask. Each address concatenated onto its respective coefficient is the next required address (ahead by two cycles due to the output latch and register) to keep cycling through the coefficients. This technique enables a reduction in the control logic required for the MAC FIR filter, but it can only be exploited when the number of coefficients is smaller than 256 for greater than 9-bit data (256 68 Xilinx
MAC FIR FILTERS data and 256 coefficient elements are required to be stored). Table 4-2 highlights the smaller resource utilization. Table 4-2: Control Logic Using Embedded Block RAMs Resource Utilization Element Control Counter Total Slices 5 5
Rounding
As noted earlier, the number of bits on the output of the filter is much larger than the number of bits on the input, and must be reduced to a manageable width. The output can be truncated by simply selecting the MSBs required from the filter. However, truncation introduces an undesirable DC data shift due to the nature of twos complement numbers. Negative numbers become more negative, and positive numbers also become more negative. The DC shift can be improved with the use of symmetric rounding, where positive numbers are rounded up and negative numbers are rounded down. The rounding capability built into the DSP48 slice maintains performance and minimizes the use of the FPGA fabric. This is implemented in the DSP48 slice using the C input port and the Carry-In port. The rounding is achieved in the following manner: For positive numbers: Binary Data Value + 0.10000 and then truncate For negative numbers: Binary Data Value + 0.01111... and then truncate The actual implementation always adds 0.0111 to the data value using the C input port, as in the negative case, and then adds the extra carry in required to adjust for positive numbers. Table 4-3 illustrates some examples of symmetric rounding. Table 4-3: Symmetric Rounding Examples Binary Value
0010.0111 0010.1000 0010.1001 1101.1001 1101.1000 1101.0111
Decimal Value 2.4375 2.5 2.5625 -2.4375 -2.5 -2.5625
Add Round
0010.1111 0011.0000 0011.0001 1110.0000 1101.1111 1101.1110
Truncate: Finish
0010 0011 0011 1110 1101 1101
Rounded Value 2 3 3 -2 -3 -3
In the instance of the MAC FIR filter, the C input is available for continued use because the Z multiplexer is used for the feedback from the P output. Therefore, for rounding to be performed, either an extra cycle or another DSP48 slice is required. Typically, an extra cycle is used to save on DSP48 slices. On the extra cycle, OPMODE is changed for the X and Y multiplexers, setting the X multiplexer to zero and the Y multiplexer to use the C input to add the user-specified requirements for a negative rounding scenario.
Xilinx 69
DSP: DESIGNING FOR OPTIMAL RESULTS The Z multiplexer remains unchanged, as the feedback loop is still required, leading to the opcode being 0101100. The simplified diagram in Figure 4-5 shows how the DSP48 slice functions during this extra cycle.
18 Data Samples 96 x 18 Data Addr A P B CIN C Dual-Port Block RAM Z-4
UG073_c3_06_081804
DSP48 Slice
18
Control
WE
Coef Addr
Rounding Constant
OPMODE Translation
Figure 4-5:
MAC FIR Filter in Rounding Mode
Rounding without an Extra Cycle

A trade-off can be made to avoid using the extra cycle required for true symmetric rounding. In this instance, the rounding constant is added to first inner product when the load of the first inner product occurs, leading to an OPMODE value of 0110101 instead of 0100101. The carry-in value is applied on the final cycle to complete the rounding. The trade-off is using the penultimate accumulated inner product as the basis for rounding, which is fine unless the penultimate inner product value is very close to zero. In this case, if the value is positive and the final inner product makes the result negative, leading to a rounding down, an incorrect result occurs due to the rounding function assuming a positive number instead of a negative. The last coefficient in typical FIR filters is very small, hence, this case rarely occurs. This form of not quite perfect rounding does save a cycle if absolutely necessary and also gives a significant improvement over truncation.
Using Distributed RAM for Data and Coefficient Buffers

For smaller-sized MAC FIR filters (typically those under 32 taps), it can be considered wasteful to use block RAM as a means to store filter input samples and coefficients. Using block RAM for a 16-tap, 18-bit filter, for example, only uses up to 3% of the memory block. Block RAMs are not as abundant as the smaller distributed RAMs found inside the slice, making them an excellent option for smaller
70 Xilinx
MAC FIR FILTERS FIR filters. Figure 4-6 illustrates the MAC FIR filter implementation using distributed RAM for the coefficient bank and an SRL16E for the data buffer.
SRL16E 18 A P WE Addr B 18 DSP48 Slice OPMODE = 0100101
Control
Coefficients 16 x 18 Single-Port Distributed Memory
load
Z-3
UG073_c3_07_081804
Figure 4-6:
Tap-Distributed RAM MAC FIR Filter
The resource utilization is still small for these small memories. For a 16-tap (or less), n-bit memory bank, the cost is n/2 slices. Therefore, for this example, the cost is nine slices per memory bank (18 slices in total). The added benefit of using SRL16Es is the embedded shifting capabilities leading to a reduction in control logic. Only a single count value is required to address both the coefficient buffer and the data buffer. The terminal count signal is used to write the slower input samples into the data buffer and capture the results and to load the accumulator with the new set of inner products. The size of the control logic and memory buffer for a 16-tap, 18-bit data and coefficient FIR is detailed in Table 4-4. Table 4-4: Data Buffer Coefficient Memory Control Counter Relational Operator Capture/Load Delay Total Control Logic Resource Utilization Element Slices 9 9 2 1 1 22
All aspects of the DSP48 and capture register approach to the MAC FIR filter using distributed RAM are identical to the block RAM based MAC FIR.
Xilinx 71
Performance
Table 4-5 compares the performance of a Virtex-4 MAC FIR filter with a Virtex-II Pro solution. Overall, the Virtex-4 DSP48 slice greatly reduces the logic fabric resource requirement, improves the speed of the design, and reduces filter power consumption. Table 4-5: 18 x 18 MAC FIR Filter (96 Tap) Comparison 18 x 18 MAC FIR Filter (96 Tap) Virtex-II Pro FPGA Virtex-4 FPGA 99 slices, 1 Embedded Multiplier, 24 slices, 1 DSP48 Slice, 1 block RAM 1 block RAM 3.125 MSPS 4.69 MSPS 250 MHz 450 MHz 170 mW 57 mW
Parameter Size Performance (Clock Speed) Power
Symmetric MAC FIR Filter

The HDL code provided in the reference design is for a single multiplier MAC FIR filter. other techniques can also be explored. This section describes how the symmetric nature of FIR filter coefficients can double the capable sample rate performance of the filter (assuming the same clock speed). By rearranging the FIR filter equation, the coefficients are exploited as follows: ( X0 x C 0 ) + ( Xn x C n ) ( X 0 + Xn ) x C 0 ( if C 0 = C n ) Equation 4-6
Figure 4-7 shows the architecture for a symmetric MAC FIR filter.
17 Data Samples 96 x 18 Data1 Addr A P Dual Read Access B
18
Control
WE
Data2 Addr Coef Addr Dual-Port Block RAM DSP48 Slice OPMODE = 010010 Z-4 load
Single-Port Block RAM
UG073_c3_08_020405
Figure 4-7:
Symmetric MAC FIR Filter
72 Xilinx
MAC FIR FILTERS There are limitations to using the symmetric MAC FIR filter. Due to the 1-bit growth from the pre-adder shown in Figure 4-5, the data input to the filter must be less than 18 bits to fit into one DSP48 slice. If necessary, the pre-adder can be implemented in slices or in another DSP48 slice. The performance of this fabric-based adder represents the critical path through the filter and limits the maximum clock speed. There are extra resources required for the filter to support symmetry. Three memory ports are needed along with the pre-adder. The control portion increases in resource utilization since the data is read out of one port in a forward direction and in reverse on the second port. This technique should only be utilized when extra sample rate performance is required.
Dual-Multiplier MAC FIR Filter

Another technique used to improve the data throughput of an MAC FIR filter is to increase the number of multipliers used to process the data. This introduces parallelism into the DSP design, and can be extrapolated into completely parallel techniques supporting the highest of sample rates. Figure 4-8 and Figure 4-9 illustrate how a dual-multiplier MAC FIR filter can be implemented using two DSP slices. Figure 4-8 shows the accumulation of the coefficients of each of the two MAC engines. These partial results must be combined together and then rounded to achieve the final result. This process uses an extra cycle and the OPMODE switching of the DSP48 slice. This extra cycle is illustrated in Figure 4-9.
18 Data Samples 43 x 18 Data Addr A DSP48 Slice
P Coefficients 43 x 18 B
18
Control
WE
Coef Addr OPMODE Translation Dual-Port Block RAM OPMODE = 010010
18 Data Samples 43 x 18 A
DSP48 Slice
Dual-Port Block RAM
OPMODE = 010010
UG073_c3_09_081804
Figure 4-8:
Dual-Multiplier MAC FIR Filter
Xilinx 73

18 Data Samples 43 x 18 Data Addr A P B 18 DSP48 Slice
Control
WE
Cin OPMODE = 0011110
Coef Addr
OPMODE Translation
Dual-Port Block RAM
Rounding Constant DSP48 Slice
18 Data Samples 43 x 18 A
OPMODE = 0100101 Dual-Port Block RAM

UG073_c3_10_081804
Figure 4-9:
Dual-Multiplier MAC FIR Filter with Extra Cycle
Conclusion
MAC FIR filters are commonly used in DSP applications. With the introduction of the Virtex-4 DSP48 slice, this function can be achieved in a smaller area, while at the same time producing higher performance with less power resources. Designers have tremendous flexibility in determining the desired implementation as well as the ability to change the implementation parameters. Each specification and design scenario brings a different set of restrictions for the design. Several more techniques are discussed in the next chapters. The ability to "tune" a filter in an existing system or to have multiple filter settings is a distinct advantage. The HDL and System Generator for DSP reference designs are easily modified to achieve specific requirements, such as different coefficients, smaller data and coefficient bit widths, and coefficient values.
74 Xilinx
Chapter 5
Parallel FIR Filters
This chapter describes the implementation of high-performance, parallel, full-precision FIR filters using the DSP48 slice in a Virtex-4 device. Because the Virtex-4 architecture is flexible, it is practical to construct custom FIR filters to meet the requirements of a specific application. Creating optimized, parallel filters saves either resources and potential clock cycles. This chapter demonstrates two parallel filter architectures: the Transposed and Systolic Parallel FIR filters. The reference design files in VHDL and Verilog permit filter parameter changes including coefficients and the number of taps.
Overview
There are many filtering techniques available to signal processing engineers. A common filter implementation for high-performance applications is the fully parallel FIR filter. Implementing this structure in the Virtex-II architecture uses the embedded multipliers and slice based arithmetic logic. The Virtex- 4 DSP48 slice introduces higher performance multiplication and arithmetic capabilities specifically designed to enhance the use of parallel FIR filters in FPGA-based DSP.
Parallel FIR Filters

A wide variety of filter architectures are available to FPGA designers due to the liquid hardware nature of FPGAs. The type of architecture chosen is typically determined by the amount of processing required in the available number of clock cycles. The two most important factors are: Sample Rate (Fs) Number of Coefficients (N) In Figure 5-1, as the sample rate increases and the number of coefficients increase, the architecture selected for a desired FIR filter becomes a more parallel structure involving more multiply and add elements. Chapter 4, MAC FIR Filters addresses the details of the sequential processing FIR filters including the single and dual MAC FIR filter. This chapter investigates the other extreme of the fully parallel FIR filter as required to filter the fastest data streams.
Xilinx 75
500 400 300 200
Parallel FIR Filters (Chapter 4)
100
Sample Rate (MHz)
50 Log Scale
Semi-Parallel FIR Filters (Chapter 5)
10 5
Sequential FIR Filters (Chapter 3)
1 0.5 5 1 10 20 50 100 200 500 1000
Number of Coefficients (N)

Log Scale
UG073_c4_02_082404
Figure 5-1:
Selecting Filter Architectures
The basic parallel architecture, shown in Figure 5-2, is referred to as the Direct Form Type 1.
18 h3 h2 h1 h0
18
38
UG073_c4_03_060404
Figure 5-2:
Direct Form Type 1 FIR Filter
76 Xilinx
PARALLEL FIR FILTERS This structure implements the general FIR filter equation of a summation of products as defined in Equation 5-1.
N1
yn =
i=0
xn i hi
Equation 5-1
In Equation 5-1, a set of N coefficients is multiplied by N respective data samples. The results are summed together to form an individual result. The values of the coefficients determine the characteristics of the filter (e.g., a low-pass filter). The history of data is stored in the individual registers chained together across the top of the architecture. Each clock cycle yields a new complete result and all multiplication and arithmetic required occurs simultaneously. In sequential FIR filter architectures, the data buffer is created using Virtex-4 dedicated block RAMs or distributed RAMs. This demonstrates a trend; as algorithms become faster, the memory requirement is reduced. However, the memory bandwidth increases dramatically since all N coefficients must be processed at the same time. The performance of the Parallel FIR filter is calculated in Equation 5-2. M a x i m u m I n p u t S a m p l e R a t e = C l o c k S p e ed Equation 5-2
The bit growth through the filter is the same for all FIR filters and is explained in the section Bit Growth in Chapter 4.
Transposed FIR Filter

The DSP48 arithmetic units are designed to be easily and efficiently chained together using dedicated routing between slices. The Direct Form Type I uses an adder tree structure. This makes it difficult to chain the blocks together. The Transposed FIR filter structure (Figure 5-3) is more optimal for use with the DSP48 Slice.
18 B B B B
h3
h2
h1
h0
0 DSP48 Slice OPMODE = 0000101 DSP48 Slice OPMODE = 0010101 DSP48 Slice OPMODE = 0010101 DSP48 Slice OPMODE = 0010101
20 P
UG073_c4_04_081104
Figure 5-3:
Transposed FIR Filter
The input data is broadcast across all the multipliers simultaneously, and the coefficients are ordered from right to left with the first coefficient, h0, on the right. These results are fed into the pipelined adder chain acting as a data buffer to store previously calculated inner products in the adder chain. The rearranged structure yields identical results to the Direct Form structure, but gains from the use of an adder chain. This different structure is easily mapped to the DSP48 slice without Xilinx 77
DSP: DESIGNING FOR OPTIMAL RESULTS additional external logic. If more coefficients are required, then more DSP48 slices are required to be added to the chain. The configuration of the DSP48 slice for each segment of the Transposed FIR filter is shown in Figure 5-4. Apart from the very first segment, all processing elements are to be configured as in Figure 5-4. OPMODE is set to multiply mode with the adder combining the results from the multiplier and from the previous DSP48 slice through the dedicated cascade input (PCIN). OPMODE is set to binary 0010101.
B
PCIN DSP48 Slice OPMODE = 0010101
PCOUT
UG073_c5_05_081104
Figure 5-4:
Transpose Multiply-Add Processing Element
Advantages and Disadvantages

The advantages to using the Transposed FIR filter are: Low Latency: The maximum latency never exceeds the pipelining time through the slice containing the first coefficient. Typically, this is three clock cycles between the data input and the result appearing. Efficient Mapping to the DSP48 Slice: Mapping is enabled by the adder chain structure of the Transposed FIR filter. This extendable structure supports both large and small FIR filters. No External Logic: No external FPGA fabric is required, enabling the highest possible performance to be achieved. The disadvantage to using the Transposed FIR filter is: Limited Performance: Performance may be limited by a high fanout input signal if there are a large number of taps.
Resource Utilization
An N coefficient filter uses N DSP48 slices. A design cannot use symmetry to reduce the number of DSP48 slices when using the Transposed FIR filter structure.
78 Xilinx
PARALLEL FIR FILTERS
Systolic FIR Filter

The systolic FIR filter is considered an optimal solution for parallel filter architectures. The systolic FIR filter also uses adder chains to fully utilize the DSP48 slice architecture (Figure 5-5).
B 18 h0 18 h1 h2 h3
20 P
UG073_c6_06_081104
Figure 5-5:
Systolic FIR Filter
The input data is fed into a cascade of registers acting as a data buffer. Each register delivers a sample to a multiplier where it is multiplied by the respective coefficient. In contrast to the Transposed FIR filter, the coefficients are aligned from left to right with the first coefficients on the left side of the structure. The adder chain stores the gradually combined inner products to form the final result. As with the Transposed FIR filter, no external logic is required to support the filter and the structure is extendable to support any number of coefficients. The configuration of the DSP48 slice for each segment of the Systolic FIR filter is shown in Figure 5-6. Apart from the very first segment, all processing elements are to be configured as shown in Figure 5-6. OPMODE is set to multiply mode with the adder combining the results from the multiplier and from the previous DSP48 slice through the dedicated cascade input (PCIN). OPMODE is set to binary 0010101. The dedicated cascade input (BCIN) and dedicated cascade output (BCOUT) are used to create the necessary input data buffer cascade.
BCIN A
BCOUT
PCOUT
ug073_c4_07_081104
Figure 5-6:
Systolic Multiply-Add Processing Element
Xilinx 79

The advantages to using the Systolic FIR filter are: Highest Performance: Maximum performance can be achieved with this structure because there is no high fanout input signal. Larger filters can be routing-limited if the number of coefficients exceeds the number of DSP slices in a column on a device. Efficient Mapping to the DSP48 Slice: Mapping is enabled by the adder chain structure of the Systolic FIR Filter. This extendable structure supports large and small FIR filters. No External Logic: No external FPGA fabric is required, enabling the highest possible performance. The disadvantage to using the Systolic FIR filter is: Higher Latency: The latency of the filter is a function of how many coefficients are in the filter. The larger the filter, the higher the latency.
An N coefficient filter uses N DSP48 slices.
Symmetric Systolic FIR Filter

In Chapter 4, MAC FIR Filters, symmetry was examined and an implementation was illustrated to exploit this symmetric nature of the coefficients. Exploiting symmetry is extremely powerful in Parallel FIR filters because it halves the required number of multipliers, which is advantageous due to the finite number of DSP48 slices. Equation 5-3 demonstrates how the data is pre-added before being multiplied by the single coefficient. (X 0 x C 0 ) + (X n x C n ) (X0 + Xn) x C0 (if C0 = Cn) Equation 5-3
Figure 5-7 shows the implementation of this type of Systolic FIR Filter structure.
17
h0
h1
h2
h3
38
UG073_c6_10_082404
Symmetric Systolic FIR
PARALLEL FIR FILTERS In this structure, DSP48 slices have been traded off for Fabric slices. From a performance viewpoint, to achieve the full speed of the DSP48 slice, the fabric 18-bit adder has to run at the same speed. To achieve this, register duplication can be performed on the output from the last tap that feeds all the other multipliers. The two register delay in the input buffer time series is implemented as an SRL16E and a register output to save on logic area. A further benefit of the symmetric implementation is the reduction in latency, due to the adder chain being half the length. Figure 5-8 shows the configuration of the DSP48 slice for each segment of the Symmetric Systolic FIR filter. Apart from the very first segment, all processing elements are to be configured as in Figure 5-8. OPMODE is set to multiply mode with the adder combining results from the multiplier and from the previous DSP48 slice via the dedicated cascade input (PCIN). OPMODE is set to binary 0010101.
B
PCOUT
UG073_c5_11_082404
Figure 5-8:
Symmetric Systolic Multiply-Add (MADD) Processing Element
An N symmetric coefficient filter uses N DSP48 slices. The slice count for the pre-adder and input buffer time series is a factor of the input bit width (n) and N. The equation for the size in slices is: (( n +1 ) * ( N / 2 ) ) + (n / 2 ) Equation 5-4
For the example illustrated in Figure 5-7, the size is (17+1) * 8/2 + 17/2 = 81 slices.
Rounding
The number of bits on the output of the filter is much larger than the input and must be reduced to a manageable width. The output can be truncated by simply selecting the MSBs required from the filter. However, truncation introduces an undesirable DC data shift. Due to the nature of two's complement numbers, negative numbers become more negative and positive numbers also become more negative. The DC shift can be improved with the use of symmetric rounding, where positive numbers are rounded up and negative numbers are rounded down. The rounding capability in the DSP48 slice maintains performance and minimizes the use of the FPGA fabric. This is implemented in the DSP48 slice using the C input port and the Carry In port. Rounding is achieved by: For positive numbers: Binary Data Value + 0.10000 and then truncate For negative numbers: Binary Data Value + 0.01111... and then truncate Xilinx 81
DSP: DESIGNING FOR OPTIMAL RESULTS The actual implementation always adds 0.0111 to the data value through the C port input as in the negative case, and then adds the extra carry in required to adjust for positive numbers. Table 5-1 illustrates some examples of symmetric rounding. Table 5-1: 2.4375 2.5 2.5625 2.4375 2.5 2.5625 Symmetric Rounding Examples Binary Value
0010.0111 0010.1000 0010.1001 1101.1001 1101.1000 1101.0111
Decimal Value
Add Round
0010.1111 0011.0000 0011.0001 1110.0000 1101.1111 1101.1110
Truncate: Finish
0010 0011 0011 1110 1101 1101
Rounded Value 2 3 3 2 3 3
For both the Transposed and Systolic Parallel FIR filters, the C input is used at the beginning of the adder chain to drive the carry value into the accumulated result. The final segment uses the MSB of the PCIN as the carry-in value to determine if the accumulated product is positive or negative. CARRYINSEL is used to select the appropriate carry-in value. If positive, the carry-in value is used, and if negative, the result is kept the same (see Figure 5-9).
B 18 h0 18 h1 h2 h3
0.49999 C DSP48 Slice OPMODE = 0'b0000101 DSP48 Slice OPMODE = 0'b0010101 DSP48 Slice OPMODE = 0'b0010101 DSP48 Slice OPMODE = 0'b0010101 Carryinsel = 01
18 P
UG073_c4_08_060704
Figure 5-9:
Systolic FIR Filter with Rounding
The one problem with this solution occurs when the final accumulated inner product input to the final DSP48 slice is very close to zero. If the value is positive and the final inner product makes the result negative (leading to a rounding down), then an incorrect result occurs because the rounding function assumes a positive number instead of a negative. The last coefficient in typical FIR filters is very small, so this situation rarely occurs. However, if absolute certainty is required, an extra DSP48
82 Xilinx
PARALLEL FIR FILTERS slice can perform the rounding function (see Figure 5-10). A Transposed FIR filter can have exactly the same problem as the Systolic FIR filter.
0.4999 C B 18 h0 18 h1 h2 h3
P C 0
DSP48 Slice OPMODE = 0'b0000101 DSP48 Slice OPMODE = 0'b0010101 DSP48 Slice OPMODE = 0'b0010101 DSP48 Slice DSP48 Slice OPMODE = 0'b0010101 OPMODE = 0'b0011101 Carryinsel = 01
UG073_c4_09_061304
18
Figure 5-10:
Systolic FIR Filter with Separate Rounding Function
Performance
When examining the performance of a Virtex-4 Parallel FIR filter, a Virtex-II Pro design is a valuable reference. Table 5-2 illustrates the ability of the Virtex-4 DSP48 slice to greatly reduce logic fabric resources requirements while improving the speed of the design and reducing the power utilization of the filter. Table 5-2: Performance Analysis Device Family Virtex-II Pro FPGA Virtex-II Pro FPGA Virtex-4 FPGA Size 1860 Slices 26 Embedded Multipliers 2958 Slices 26 Embedded Multipliers 0 Slices 51 DSP48 Slices Performance 300 MHz Clock Speed 300 MSPS 300 MHz Clock Speed 300 MSPS 400 MHz Clock Speed 400 MSPS Power (Watts) TBD
Filter Type 18 x 18 Parallel Transposed FIR Filter (51 Tap Symmetric) 18 x 18 Parallel Systolic FIR Filter (51 Tap Symmetric) 18 x 18 Parallel Transposed FIR Filter (51 Tap Symmetric)
TBD
TBD
Xilinx 83
DSP: DESIGNING FOR OPTIMAL RESULTS Table 5-2: Performance Analysis (Continued) Device Family Virtex-4 FPGA Size 0 Slices 51 DSP48 Slices Performance 450 MHz Clock Speed 450 MSPS 400 MHz Clock Speed 400 MSPS Power (Watts) TBD
Filter Type 17 x 18 Systolic FIR Filter (51 Tap Nonsymmetric) 17 x 18 Systolic FIR Filter (51 Tap Symmetric)
Virtex-4 FPGA
477 Slices 26 DSP48 Slices
TBD
Conclusion
Parallel FIR filters are commonly used in high-performance DSP applications. With the introduction of the Virtex-4 DSP48 slice, DSPs can be achieved in a smaller area, thereby producing higher performance with less power penalty. Designers have tremendous flexibility in determining the desired implementation, and also have the ability to change the implementation parameters. The ability to tune a filter in an existing system or to have multiple filter settings is a distinct advantage. By making the necessary coefficient changes in the synthesizable HDL code, the reconfigurable nature of the FPGA is fully exploited. The coefficients can be either hardwired to the A input of the DSP48 slices or stored in small memories and selected to change the filter characteristics. The HDL and System Generator for DSP reference designs are easily modified to achieve specific requirements.
84 Xilinx
Chapter 6
Semi-Parallel FIR Filters
This chapter describes the implementation of semi-parallel or hardware-folded, full-precision FIR filters using the Virtex-4 DSP48 slice. Because the Virtex-4 architecture is flexible, constructing FIR filters for specific application requirements is practical. Creating optimum filter structures of a semi-parallel nature saves resources and potential clock cycles. Therefore, optimum filter structures of a semi-parallel nature can be created without draining resources or clock cycles. This chapter demonstrates two semi-parallel filter architectures: the four- multiplier FIR filter using distributed RAM and the three-multiplier FIR filter using block RAM. These filters illustrate how resources are saved by using available clock cycles and hardware-folding techniques. Reference design files are available for system generator in DSP, VHDL, and Verilog. The reference designs permit filter parameter changes including coefficients and the number of taps.
Overview
A large array of filtering techniques are available to signal processing engineers. A common filter implementation to exploit available clock cycles, while still achieving moderate to high sample rates, is the semi-parallel (also known as folded-hardware) FIR filter. In the past, this structure used the Virtex-II embedded multipliers and slice-based arithmetic logic. However, the Virtex-4 DSP48 slice introduces higher performance multiplication and arithmetic capabilities to enhance the use of semiparallel FIR filters in FPGA-based DSP designs.
Semi-Parallel FIR Filter Structure

A wide variety of filter architectures are available to FPGA designers due to the liquid hardware nature of FPGAs. The type of architecture is typically determined by the amount of processing required in the number of available clock cycles. The two most important factors are: Sample Rate (Fs) Number of Coefficients (N) As illustrated in Figure 6-1, as the sample rate increases and the number of coefficients increase, the architecture selected for a desired FIR filter becomes a more parallel structure involving more multiply-add elements. Chapter 4, MAC FIR Filters addresses the details of sequential processing FIR filters including the single and dual MAC FIR Filter. Chapter 5, Parallel FIR Filters investigates the polar extreme of the fully-parallel FIR filter required for the highest sample rate filters. This chapter examines the common scenario requiring multiple processing elements working Xilinx 85
DSP: DESIGNING FOR OPTIMAL RESULTS over numerous clock cycles to achieve the result. These techniques are often referred to as semi-parallel and are used to maximize efficiency of the filter (see Figure 6-1).
500 400 300 200
Parallel FIR Filters (Chapter 4)
100
Sample Rate (MHz)
50 Log Scale
Semi-Parallel FIR Filters (Chapter 5)
10 5
Sequential FIR Filters (Chapter 3)
1 0.5 5 1 10 20 50 100 200 500 1000
Number of Coefficients (N)

Log Scale
UG073_c5_01_082404
Figure 6-1:
Selecting Filter Architectures
The semi-parallel FIR structure implements the general FIR filter equation of a summation of products defined as shown in Equation 6-1. N1 yn =
xn i hi
i=0
Equation 6-1
Here a set of N coefficients is multiplied by N respective time series data samples, and the results are summed together to form an individual result. The values of the coefficients determine the characteristics of the filter (for example, a low-pass filter). Along with achievable clock speed and the number of coefficients (N), the number of multipliers (M) is also a factor in calculating semi-parallel FIR filter performance. The following equation demonstrates how the more multipliers used, the greater the achievable performance of the filter. 86 Xilinx
SEMI-PARALLEL FIR FILTERS Maximum Input Sample rate = (Clock speed / Number of Coefficients) x Number of Multipliers The above equation is rearranged to determine how many multipliers to use for a particular semiparallel architecture: Number of Multipliers = (Maximum Input Sample rate x Number of Coefficients) / Clock speed The number of clock cycles between each result of the FIR filter is determined by the following equation: Number of Clock cycles per result = Number of Coefficients / Number of Multipliers The bit growth on the output of the filter is the same as for all FIR filters and is explained in Bit Growth in Chapter 4. The large 48-bit internal precision of the DSP48 slice means that little concern needs to be paid to the internal bit growth of the filter.
Four-Multiplier, Distributed-RAM-Based, Semi-Parallel FIR Filter

Once the required number of multipliers is determined, there is an extendable architecture using the DSP48 slice for use as the basis of the filter. This section assumes the specifications in Table 6-1 describe the filter implementation and its functions. Table 6-1: Four-Multiplier, Semi-Parallel FIR Filter Specifications 112.5 MSPS 16 450 MHz 18 Bits 18 Bits 4 4
Sampling Rate Number of Coefficients Assumed Clock Speed Input Data Width Output Data Width Number of Multipliers Number of Clock Cycles Between Each Result
Figure 6-2 illustrates the main structure for the four-multiplier, semi-parallel FIR filter.
CE 18 SRL16E CE B CE1 SRL16E CE1 B CE2 SRL16E CE2 B CE3 SRL16E B
Coefficients 4 x 18
18
Coefficients 4 x 18
18
Coefficients 4 x 18
18
Coefficients 4 x 18
18 PCIN 40 P CE4
0 DSP48 Slice OPMODE = 0110101 DSP48 Slice OPMODE = 0010101 DSP48 Slice OPMODE = 0010101
DSP48 Slice DSP48 Slice OPMODE = 0010101 OPMODE = 0010010

UG073_c5_02_081104
Figure 6-2:
Four-Multiplier, Semi-Parallel FIR Filter in Accumulation Mode
The DSP48 slice arithmetic units are designed to be chained together easily and efficiently due to dedicated routing between slices. Figure 6-2 shows how the four DSP48 slice multiply-add elements are cascaded together to form the main part of the filter structure. Figure 6-3 provides a detailed view Xilinx 87
DSP: DESIGNING FOR OPTIMAL RESULTS of the main multiply-add elements. The two pipeline registers are used on the B input to compensate for the register on the output of the coefficient memory.
B
PCIN
PCOUT

UG073_c5_03_081104
Figure 6-3:
Detailed Diagram of a Single Multiply-Add Element
An extra DSP48 slice is required on the end to perform the accumulation of the partial results, thus creating the final result. A new result is created every four cycles. Every four cycles, the accumulation must be reset to the first partial value of the next result. As in the MAC FIR Filter, this reset (or load) is achieved by changing the OPMODE value of the DSP48 slice for a single cycle. OPMODE is changed from binary 0010010 to binary 0010000 (just a single bit change). At the same time, the capture register is also enabled, and the final result is stored on the output (see Figure 6-4).
CE 18 SRL16E CE B CE1 SRL16E CE1 B CE2 SRL16E CE2 B CE3 SRL16E B
Coefficients 4 x 18
18
Coefficients 4 x 18
18
Coefficients 4 x 18
18
Coefficients 4 x 18
0 18 PCIN 40 P CE4
0 DSP48 Slice OPMODE = 0110101 DSP48 Slice OPMODE = 0010101 DSP48 Slice OPMODE = 0010101
DSP48 Slice DSP48 Slice OPMODE = 0010101 OPMODE = 0010000

UG073_c5_02_081104
Figure 6-4:
Four-Multiplier, Semi-Parallel FIR Filter at the Start of a New Result Cycle
Control logic is required to make this dynamic change occur. The specifics are detailed in Control Logic and Address Sequencing, page 90.
Data Memory Buffers

This example uses eight memories. Four SRL16Es are used as data buffers. Each SRL16E holds the four samples needed for the result. They are written to once every four cycles (the input data rate is 4x slower than the internal rate), and the shifting characteristic of the SRL16E is exploited to pass old 88 Xilinx
SEMI-PARALLEL FIR FILTERS samples along the time series buffer. The extra register on the output of each data buffer is required to match up the data buffer pipeline with the extra delay caused by the adder chain. The extra register should not cost extra resources, because it is already present in the slice containing the SRL16E (see Figure 6-5).
1/2 SliceM SRL16E Register DIN CE ADDR[3:0]
CE CE
DOUT
UG073_c5_05_081904
Figure 6-5:
Single Bit of One Input Memory Buffer
As long as the depth does not exceed 16, the resources required for each of these input memory buffers is determined by the bit width of the input data (n). Therefore, n/2 SliceM is required for each memory buffer, leading to nine slices per buffer in this filter example. For depths up to 32, resources are a little more than doubled because two SRL16Es are needed, as well as an extra output multiplexer. For more information on SliceM, refer to the CLB section in the Virtex-4 User Guide.
Coefficient Memory
The coefficients are divided up into four groups of four. This arrangement is determined by dividing the total number of coefficients by the number of multipliers used in the implementation. In this example, if the total number of coefficients is 16, and the number of multipliers is four, four coefficients per memory are needed. Note that filters with a total number of coefficients that are integer-divisible by the required number of multipliers are very desirable. System designers should take this into account when designing their filters to get the optimal filter specification for the implementation used. Otherwise, the coefficients will have to be padded with zeros to achieve a number of coefficients that are integerdivisible by the number of multipliers. The coefficients are simply split into groups according to their order. The first four in the first memory, the second four in the second memory, and so on (see Figure 6-6).
Xilinx 89

Coefficients Driving First DSP48 Slice h0 h1 h2 h3 Coefficients Driving Second DSP48 Slice h4 h5 h6 h7 Coefficients Driving Third DSP48 Slice h8 h9 h10 h11 Coefficients Driving Fourth DSP48 Slice h12 h13 h14 h15
UG073_c5_06_060804
Figure 6-6:
Coefficient Memory Arrangement
The adder chain architecture of the DSP48 slice means that each Multiply-Add cascade multiplication must be delayed by a single cycle so that the results synchronize appropriately when added together. This delay is achieved by addressing of the memories and is explained in Control Logic and Address Sequencing. Distributed RAM (refer to Chapter 2, XtremeDSP Design Considerations, for detailed information on distributed RAMs) are used for the coefficient memories. The reason for their use is that it would be an extremely inefficient usage of the larger block RAMs, especially given their scarcity versus the smaller abundant distributed RAMs. The larger block RAM comes into play when the number of coefficients per memory starts to increase to the point where the cost in slice resources becomes significant (for example, greater than 64). The total cost of the current example is 36 slices. The coefficient width is 18 bits, and distributed RAMs cost n/2 slices (that is, nine slices per memory and four memories). For larger distributed RAMs (larger than 16 elements), the size begins to increase as Write Enable (WE) control logic and an output multiplexer is needed. The distributed memory v7.0 in the CORE Generator system can be easily used to create these little distributed RAMs and get accurate size estimates.
Control Logic and Address Sequencing

The Control Logic and Address Sequencing is the most important and complicated aspect of semiparallel FIR filters, and getting it right is crucial to the operation of the filter. The control logic is discussed in two separate sections: Memory Addressing Clock Enable Sequencing Memory addressing must provide the necessary delay for each multiply-add element mentioned in Coefficient Memory, page 89, caused by the adder chain. This is not the case when using an adder tree; the DSP48 slice is most efficiently used in adder chains. Figure 6-7 illustrates the control logic required to create the necessary memory addressing. The counter creates the fundamental zero through three count. This is then delayed by one cycle by the use of a register in the control path. Each successive delay is used to address both the coefficient memory and the data buffer of their respective multiply-add elements. A single delay for the second multiplyadd element, two delays for the third multiply-add element, etc. Note that this is extensible control logic for M number of multipliers.
90 Xilinx
SEMI-PARALLEL FIR FILTERS

Coefficient Memory 0 & Data Buffer 0 Address Coefficient Memory 1 & Data Buffer 1 Address Coefficient Memory 2 & Data Buffer 2 Address Coefficient Memory 3 & Data Buffer 3 Address
addr Counter 0 -> N/M - 1
N/M - 1
CE
CE1
CE2
CE3
CE4
UG073_c5_07_081104
Figure 6-7:
Control Logic for the Four-Multiplier, Semi-Parallel FIR Filter
Figure 6-7 also shows clock enable sequencing. A relational operator is required to determine when the count limited counter resets its count. This signal is High for one clock cycle every four cycles, to represent the input and output data rates. The Clock Enable signal is delayed by a single register just like the coefficient address, and each delayed version of the signal is tied to the respective section of the filter. Refer to Figure 6-2 to see the signal connections to the element. Figure 6-8 illustrates the control logic waveforms changing over time.
Clock Input Data x(n) 0 0 1 2 3 0 1 2 3 0 1 2 3 0 x(n-1) x(n-2) y(n) 1 2 3 x(n-3) y(n-1) 0 1 2 3 x(n-4) x(n-5) y(n-2)
Output Data Address for First DSP48 Slice MADD Element Control CE Address for Second DSP48 Slice MADD Element Control CE1 Address for Third DSP48 Slice Design Element Control CE2 Address for Fourth DSP48 Slice Design Element Control CE3 2 3 0 1 2
UG073_c5_08_082504
Figure 6-8:
Control Waveforms for Semi-Parallel FIR Filters Xilinx 91
Table 6-2 shows the resources used by a 16-tap, four-multiplier, distributed-RAM-based, semiparallel FIR filter. Table 6-2: Resource Utilization Elements Multiply-Add Input Data Buffers Coefficient Memories Capture Register Main Control Counter Relational Operator Multiply-Add Element Control Total 36 36 20 2 1 9 (3 per extra element) 104 5 Slices DSP48 Slices 5
Three-Multiplier, Block RAM-Based, Semi-Parallel FIR Filter

This section investigates a different filter structure, the three-multiplier, block RAM-based, semiparallel FIR filter (see Figure 6-9).
WE 18 Cyclic Data Buffer 100 x 18 B WE1 Cyclic Data Buffer 100 x 18 B WE2 Cyclic Data Buffer 100 x 18 B
Coefficients 100 x 18 18 0 DSP48 Slice OPMODE = 0110101
Coefficients 100 x 18 18
18 PCIN P 40 CE4

UG073_c5_09_081104
Figure 6-9:
Three-Multiplier, Block-RAM-Based, Semi-Parallel FIR Filter
The decision to use this implementation is based on the filter specification. The filter specifications are described in Table 6-3. Table 6-3: Three-Multiplier, Block RAM-Based, Semi-Parallel FIR Filter Specifications Parameter Sampling Rate Number of Coefficients Assumed Clock Speed Input Data Width Output Data Width 92 Xilinx Value 4.5 MSPS 300 450 MHz 18 Bits 18 Bits
SEMI-PARALLEL FIR FILTERS Table 6-3: Three-Multiplier, Block RAM-Based, Semi-Parallel FIR Filter Specifications Parameter Number of Mulipliers Number of Clock Cycles Between Each Result Value 3 100
The structure is similar to the four-multiplier filter studied earlier. In this instance, the lower sample rate of the filter specification and the larger number of taps indicates that only three multipliers are required, each servicing 100 coefficients, leading to a new result yielded every 100 clock cycles. Each memory buffer is required to hold 100 coefficients and also 100 input data history values. The dedicated Virtex-4 block RAM can be used in dual-port mode with a cyclic data buffer established in the first half of the memory to serve the shifting input data series. Chapter 4, MAC FIR Filters, describes using these memories to store the input data series, the coefficients, and also the control logic required to make the cyclic RAM buffer operate. The rest of the control logic and data flow is identical to the first filter investigated except that only three multipliers are serviced, therefore, the control logic can be scaled back by one element. Also note that the WE signals are the inversion of their respective CE pair. Table 6-4 shows the resource utilization for the 300-tap, three-multiplier, semi-parallel FIR filter. Table 6-4: Resource Utilization Elements Multiply-Add Input Data Buffers and Coefficient Memories Capture Register Main Control Counter Relational Operator Multiply-Add Element Control Total 20 5 1 12 (6 per extra element) 38 Slices DSP48 Slices 4 Block RAMs 3
Other Semi-Parallel FIR Filter Structures

As with many DSP functions there are many different ways to implement a function. There is never one solution fitting all requirements for all specifications. For example, should distributed or block RAM be used for data storage? Should a systolic or a transposed implementation be used for a given filter? This chapter describes in detail the different techniques using single-rate FIR filters to get the maximum performance and low resource utilization using the Virtex-4 architecture. This section introduces other possible semi-parallel FIR filter implementations and discusses the advantages and disadvantages of their structures.
Xilinx 93
Semi-Parallel, Transposed, Four-Multiplier FIR Filter

This structure is very different in nature to the main architecture discussed in this chapter (see Figure 6-10).
Cyclic Data Buffer 400 x 18 A 43 CE 18 B
18

Coefficients 100 x 36 36 A
18

A
18

Coefficients 100 x 36 36 A
18
UG073_c5_10_061404
Figure 6-10:
Semi-Parallel, Transposed FIR Filter
Only one data storage buffer is required, typically a block RAM. The data buffer output is also broadcast to all DSP48 slices. Each DSP48 slice works in accumulator mode until the last cycle of the calculation, when OPMODE changes to form an adder chain, and then passes the results to the next DSP48 slice. Actually, four results are being calculated at one time, and the completed result is output 94 Xilinx
SEMI-PARALLEL FIR FILTERS from the last DSP48 slice. The previous elements are working on their respective parts of the next results. Figure 6-11 shows the filter structure every time the DSP48 slice OPMODE is changed, which occurs once every result cycle.
18 Cycle Data Buffer 400 x 18 A 43 CE 18 B

Coefficients 100 x 36 36 A Coefficients 100 x 36 18 B
18

36 A
18
UG073_c5_11_060804
Figure 6-11:
Semi-Parallel, Transposed FIR Filter (Combination of the Results)

The advantages to using the Semi-Parallel, Transposed FIR filter are: Lower resource utilization due to one less DSP48 slice required and a single input memory buffer. Xilinx 95
DSP: DESIGNING FOR OPTIMAL RESULTS Low latency due to the transpose nature of the filter implementation is lower than the Systolic approach. The latency is equal to the size of one coefficient bank. The disadvantages to using the Semi-Parallel, Transposed FIR filter are: Lower performance due to the broadcast nature of the data buffer output can limit performance of the filter. Control logic is more difficult to understand, but is still of a compact nature.
Rounding
The number of bits on the output of the filter is much larger than the input and must be reduced to a manageable width. The output can be truncated by simply selecting the MSBs required from the filter. However, truncation introduces an undesirable DC shift on the data set. Due to the nature of twos complement numbers, negative numbers become more negative and positive numbers also become more negative. The DC shift can be improved with the use of symmetric rounding, where positive numbers are rounded up and negative numbers are rounded down. The rounding capability built into the DSP48 slice maintains performance and minimizes the use of FPGA fabric. This is ingrained in the DSP48 slice via the C input port and also the Carry-In port. Rounding is achieved in the following manner: For positive numbers: Binary Data Value + 0.10000 and then truncate For negative numbers: Binary Data Value + 0.01111... and then truncate The actual implementation always adds 0.0111 to the data value using the C port input as in the negative case, and then adds the extra carry in required to adjust for positive numbers. Table 6-5 illustrates some examples of symmetric rounding. Table 6-5: 2.4375 2.5 2.5625 -2.4375 -2.5 -2.5625 Symmetric Rounding Examples Binary Value
0010.0111 0010.1000 0010.1001 1101.1001 1101.1000 1101.0111
Decimal Value
Add Round
0010.1111 0011.0000 0011.0001 1110.0000 1101.1111 1101.1110
Truncate: Finish
0010 0011 0011 1110 1101 1101
Rounded Value 2 3 3 -2 -3 -3
In the instance of the semi-parallel FIR filter, an extra DSP48 slice is required to perform the rounding functionality. It cannot be ingrained into the final accumulator because the rounding cannot be done on the final result. If the C input is used and the accumulator is put into three-input add mode, then rounding is performed on the partial result. The more multipliers in the filter, the worse the rounding performance because even fewer inner products are included in the result. An extra DSP48 slice is required to perform the rounding. Due to the finite nature of the DSP48 slices, it is recommended that the symmetric rounder be actually implemented in the fabric outside of the slices. The function is small and does not have to run at a high frequency because the results are running at the much slower input data rate.
96 Xilinx
SEMI-PARALLEL FIR FILTERS
Performance
It does not make sense to compare the performance of the semi-parallel FIR filter in a Virtex-4 device with Virtex-II Pro results because completely different techniques are used to build the filters. As a general statement though, Virtex-4 devices improve the speed of the design, shrink the area, and reduce power drawn by the filter. All designs assume 18-bit data and 18-bit coefficient widths. Table 6-6 through Table 6-8 compare the specifications for three filters. Table 6-6: 4-Multiplier, Memory-Based, Semi-Parallel FIR Filter Specifications (16-Tap Symmetric) Parameter Size Performance Power Specification 94 slices, 5 DSP48 slices 458 MHz clock speed, 114.5 MSPS TBD Watt
Table 6-7: 3-Multiplier, Block-RAM-Based, Semi-Parallel FIR Filter Specifications (300-Tap Symmetric) Parameter Size Performance Power Specification 38 slices, 4 DSP48 slices, 4 block RAMs 450 MHz clock speed, 4.5 MSPS TBD Watt
Table 6-8: 4-Multiplier, Block-RAM-Based, Semi-Parallel Transposed FIR Filter Specifications (400-Tap Symmetric) Parameter Size Performance Power Specification 46 slices, 4 DSP48 slices, 2 block RAMs 450 MHz clock speed, 4.5 MSPS TBD Watt
Xilinx 97
Conclusion
Semi-parallel FIR filters probably are the most frequently used filter techniques in Virtex-4 highperformance DSP applications. Figure 6-12 shows the necessary implementation decisions and provides guidelines for choosing the required structure based on the filter specifications.
Transposed FIR 500 400 300 200 Symmetric MACC FIR
Systolic FIR (symmetric & non-symmetric) Increasing Number of Multipliers Semi-Parallel Distributed Memory FIR 10-Multiplier Semi-Parallel FIR
100
Sample Rate (MHz) Log Scale
50
10 5 Distributed Memory MACC FIR 1 0.5 1 5 10 20 50 100 200 Embedded Control MACC FIR
Block RAM
Semi-Parallel BRAM FIR
Distributed Memory
Normal Control MACC FIR 500
1000
Number of Coefficients (N) Log Scale
UG072_c5_12_082304
Figure 6-12:
Selecting the Correct Filter Architecture for Semi-Parallel FIR Filters
The major lines indicate the guideline thresholds between given implementation techniques. For instance, the shift to using block RAM is desirable when the number of taps needed to be stored in a given memory exceeds 32. This correlates to two SRL16Es for the data buffers. If more than two SRL16Es are used in a data buffer, it will be difficult to reach the high clock rate indicated in Chapter 4, MAC FIR Filters, Chapter 5, Parallel FIR Filters, and this chapter. However, this is only a guideline. A great deal depends upon how many slices or block RAMs are remaining in the device, the power requirements, and the available clock frequencies. A given filter implementation is subjective because a different set of restrictions is provided by every application and design. In general, the guidelines provided in the past three chapters should enable designers to make sensible and efficient decisions when designing filters. These chapters also complete the foundations required for filter construction in Virtex-4 devices so that more complex, multi-channel and interpolation or decimation multi-rate filters can be constructed. The supplied referenced designs further aid in helping to understand and utilize these filters. 98 Xilinx
Chapter 7
Multi-Channel FIR Filters
This chapter illustrates the use of the advanced Virtex-4 DSP features when implementing a widely used DSP function known as multi-channel FIR filtering. Multi-channel filters are used to filter multiple input sample streams in a variety of applications, including communications and multimedia. The main advantage of using a multi-channel filter is leveraging very fast math elements across multiple input streams (i.e., channels) with much lower sample rates. This technique increases silicon efficiency by a factor almost equal to the number of channels. The Virtex-4 DSP48 slice is one of the new and highly innovative diffused elements that form the basis of the Application Specific Modular BLock or ASMBL architecture. This modular architecture enables Xilinx to rapidly and cost-effectively build FPGA platforms by combining different elements, such as logic, memory, processors, I/O, and of course, DSP functionality targeting specific applications such as wireless or video DSP. The Virtex-4 DSP48 slice contains the basic elements of classic FIR filters: a multiplier followed by an adder, delay or pipeline registers, plus the ability to cascade an input stream (B bus) and an output stream (P bus) without exiting to a general slice fabric. The resulting DSP designs can have optional pipelining that permits aggregate multi-channel sample rates of up to 500 million samples per second, while minimizing power consumption and external slice logic. In the implementation described in this chapter, multi-channel filtering can be looked at as time-multiplexed, single-channel filters. In a typical multi-channel filtering scenario, multiple input channels are filtered using a separate digital filter for each channel. Due to the high performance of the DSP48 block within the Virtex-4 device, a single digital filter can be used to filter all eight input channels by clocking the single filter with an 8x clock. This implementation uses 1/8th of the total FPGA resource as compared to implementing each channel separately.
Multi-Channel FIR Implementation Overview

Top Level
The implementation of a six-channel, eight-tap FIR filter using DSP48 elements is depicted in Figure 7-1. The design elements used in the implementation include the following:
Xilinx #
DSP: DESIGNING FOR OPTIMAL RESULTS Six-to-one multiplexer that is implemented in slice logic as described in Combining Separate Input Streams into an Interleaved Stream, page 101 Coefficient ROMs using SRL16Es connected in head-to-tail fashion Input sample delay-by-seven SRL16Es to hold the interleaved streams DSP48 slices for multiplication and additions
SRL16 Coefficient ROM xo(n) x1(n) x2(n) x3(n) x4(n) x5(n) Z-7 Z-7 Z-7 SRL16 Coefficient ROM SRL16 Coefficient ROM SRL16 Coefficient ROM
UG073_c6_03_081804
Figure 7-1:
Block Diagram of a 6-Channel, 8-Tap FIR Filter
All datapaths and coefficient paths for this example are 8 bits wide. The coefficient ROMs and input sample delay elements are designed using SRL16Es. The SRL16E is a very compact and efficient memory element, running at the very high 6x clock rate. For adaptive filtering, where coefficients can be different depending upon their input signals, coefficient RAMs can be used to update the coefficient values. The DSP48 slices and interconnects also run at the 6x clock rate, providing unparalleled performance for multiplication and additions in todays FPGAs.
DSP48 Tile
The multi-channel filter block is a cascade implementation of the DSP48 tile. Each tile is implemented as shown in Figure 7-2. An SRL16E is used to shift the input from the six channels. The
# Xilinx
SEMI-PARALLEL FIR FILTERS product cascade path between two DSP48 slices within the tile can be used to bring the product output from one tap into the cascading input of the next tap for the final addition.
Input: 6 Channels SRL16 Taps SRL16 Taps
8 bits
C1
C2
Add
UG073_c6_04_081804
Figure 7-2:
DSP48 Tile Cascading Diagram
Combining Separate Input Streams into an Interleaved Stream

As shown in Figure 7-3, six separate video input sample streams must be combined into one interleaved sample stream for this multi-channel FIR filter example. Conceptually, a high-speed, sixto-one multiplexer feeds a seven deep SRL16E shift register to accomplish this task. The SRL16E depth is the number of channels plus one.
Counter
Shift Register
X6(n-1) X5(n-1) X4(n-1) X3(n-1) X2(n-1) X1(n-1) X0(n-1)
18
ug073_c6_05_060904
Figure 7-3:
Converting Eight Input Streams to One Interleaved Input Stream
For each clock tick, the counter selects a different input stream (in order), and then supplies this value to the SRL16E shift register. After six clock ticks, the six input samples for a given time period are loaded sequentially, or interleaved into a single stream. A six-to-one multiplexer must be designed carefully, as it is constructed with slice logic that must run at the 6x clock rate. At 446 MHz, good design practices dictate connections point-to-point, a maximum of one Look-Up Table (LUT) between flip-flops and RLOC techniques.
Xilinx 15
DSP: DESIGNING FOR OPTIMAL RESULTS To reduce the high fanouts on the selected lines of the multiplexer, the conceptual multiplexer in Figure 7-3 is implemented as shown in Figure 7-4. This circuit is repeated for all eight bits of the input sample width.
Shift Register
'1' X0(n) X1(n) LUT
'1'
LUT X2(n) X3(n)
LUT LUT
X4(n) X5(n)
LUT '0'
LUT
UG073_c6_06_060904
Figure 7-4:
High-Speed 8-to-1 Multiplexer Used in the Filter
Coefficient RAM
The six coefficient sets are stored in the SRL16 memories. If the same coefficient set is used for all channels, then only a single set is stored in the SRL16. If the different channels use different coefficients, then six sets of SRL16s are used for each tap. (Six RAMs can be used instead, one for each channel.) Each RAM is 8 bits wide and six deep, corresponding to the six taps. The optional Load input is used to change or load a new coefficient set. Six clock cycles are needed to load all six RAMs. Input C1 is used to load the eight locations of RAM1 which are used for Channel1. C8 is used to load the eight locations of RAM8 which are used for Channel8. At the eighth clock, all eight locations of the eight RAMs are loaded; the filter then becomes an adaptive filter. The speed of the overall filter will be reduced when the coefficients are stored in the RAM.
Control Logic
The control logic is used to ensure proper functioning of the different blocks. If the coefficient RAM block is used, the control logic ensures that the load signal is High for six clocks. Different tapenabled signals are used to make sure that RAM values are read into the DSP48 correctly. For instance, clock1 reads in the first location from RAM1, but the first location of RAM2 is read only at the clock number equal to shift register length. The design assumes a clock is running at 6x that of the input # Xilinx
SEMI-PARALLEL FIR FILTERS signals. The DCM can also be used to multiply the clock if the only available clock is running at the input channel frequency. The control logic also takes care of the initial latency such that the final output is enabled only after the initial latency period is complete.
Implementation Results
The initial latency of the design is equal to the [(number of channels + 1) * number of taps] plus three pipe stages within the DSP48. After placement and routing, the design uses 216 slices and eight DSP48 blocks. The design has a speed of 454 MHz.
Conclusion
The available arithmetic functions within the DSP48 block, combined with fine granularity and high speed, makes the Virtex-4 FPGA an ideal device to implement high-speed, multi-channel filter functions. The design shows the efficient implementation of a six-channel, eight-tap filter. Due to the high-performance capability within the DSP48 block, a single channel, eight-tap filter can be used to implement the six-channel, eight-tap filter, reducing the area utilization by 1/6th.
Xilinx 15
# Xilinx
Appendix A
References
1. 2. 3. 4.
A Digital Signal Processing Primer by Ken Steiglitz, ISBN: 0-8053-1684-1 Digital Video and HDTV Algorithms and Interfaces by Charles Poynton, ISBN: 1-55860-792-7 DSP Primer by C.Britton Rorabaugh, ISBN:0-07-054004-7 Xilinx, Inc., Virtex-4 User Guide
Xilinx #
# Xilinx
DSP Solutions Advanced Design Guide

How Do You Get 256 GigaMAC DSP Performance?
For high-performance DSP design there is no better solution than XtremeDSP from Xilinx. This book discusses the many advantages of the Xilinx DSP solution and how to implement working DSP designs, including:
I
The Virtex-4 FPGA family The ideal platform for creating high-performance DSP systems and for boosting the performance of a DSP processor-based system. The Virtex-4 technology Innovative architectural features and design techniques that dramatically reduce power consumption while increasing DSP performance. The XtremeDSP slice Optimal performance IP, fully integrated into the on-chip DSP-specific architecture. (up to 512 XtremeDSP slices operating at 500 MHz speed.) Design tools and support A broad pallet of easy-to-use design tools and libraries, in addition to in-depth design support, from both Xilinx and its partners. Education and design services Get up and running as quickly as possible.
Xcell Publications help you solve design challenges, bringing you the awareness of the latest tools, devices, and technologies; knowledge on how to design most effectively; and the next steps for implementing working solutions. See all of our books, magazines, technical journals, solutions guides, and brochures at: www.xilinx.com/xcell
Edition 1.0 April, 2005

DSP

Uploaded by

Copyright:

Available Formats

DSP

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

DSP

Uploaded by

Copyright:

Available Formats

DSP Solutions Advanced Design Guide

DSP: Designing for Optimal Results

DSP PRODUCTS: ADVANCED DESIGN GUIDE

DSP: Designing for Optimal Results

DSP: DESIGNING FOR OPTIMAL RESULTS

Edition 1.0 March 2005

DSP: DESIGNING FOR OPTIMAL RESULTS

Chapter 1: Digital Signal Processing Design Challenges

Chapter 2: XtremeDSP Design Considerations

DSP: DESIGNING FOR OPTIMAL RESULTS

Chapter 3: DSP48 Slice Math Functions

Chapter 4: MAC FIR Filters

Chapter 5: Parallel FIR Filters

Parallel FIR Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Transposed FIR Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Chapter 6: Semi-Parallel FIR Filters

Chapter 7: Multi-Channel FIR Filters

DSP: DESIGNING FOR OPTIMAL RESULTS

Digital Signal Processing Design Challenges

The Performance Gap

Performance (Algorithm and Processor Forecast)

SDR 3G Imaging Radar SD/HD Video

Traditional Processor Architectures

The Performance Gap Xilinx 1

The Ideal Solution

Simplicity and Efficiency of the Cascade Logic

Extremely Low Power Consumption

DIGITAL SIGNAL PROCESSING DESIGN CHALLENGES

Increased Flexibility for Cost Effectiveness

Virtex-4 FPGAs A Platform for Every Application

Reduce Time-to-Market with World-Class Xilinx Support

DSP: DESIGNING FOR OPTIMAL RESULTS

XtremeDSP Design Considerations

Number of DSP48 Slices Per Virtex-4 Device

Device XC4VLX15 XC4VLX25 XC4VLX40 XC4VLX60 XC4VLX80 XC4VLX100 XC4VLX160 XC4VLX200

DSP: DESIGNING FOR OPTIMAL RESULTS

DSP48 Slice Primitive

BCOUT[17:0] PCOUT[47:0] P[47:0]

DSP48 Slice Primitive

DSP: DESIGNING FOR OPTIMAL RESULTS

DSP48 Slice Attributes

XTREMEDSP DESIGN CONSIDERATIONS

DSP: DESIGNING FOR OPTIMAL RESULTS

DSP48 Tile and Interconnect

Virtex-II and Virtex-II Pro Devices

DSP48 DSP48 Slice Slice

XTREMEDSP DESIGN CONSIDERATIONS

18 Notes 4, 5 48 Note 5 BCIN Wire Shift Right by 17 bits 48 PCIN

DSP: DESIGNING FOR OPTIMAL RESULTS

Simplified DSP48 Slice Operation

OPMODE, CARRYINSEL, CIN, and SUBTRACT Control Behavior

Figure 2-4: 14 Xilinx

Simplified DSP48 Slice Model

XTREMEDSP DESIGN CONSIDERATIONS

Data In Data In Data In Data In

TDSPDCK_{CRYINP, CRYINSP, OPP, SUBPPCINP}/ TDSPCKD_{CRYINP, CRYINSP,

TDSPCKO_{PA_NL, PB_NL, PC}

Combinatorial TDSPDO_{AP, BP}_L TDSPDO_{AP_NL, BP_NL, CP} A, B to P A, B, C to P Various

TDSPDO_{APCOUT, BPCOUT}_L TDSPDO_{APCOUT_NL,

Data In A, B, C to to PC Out PC Out Control to PC Out Various