Introduction to Digital Signal ProcessORS

01/03/2007 Bryan Stenquist - University of Utah

1

Overview of topics

• DSP in the world

• Motivations of DSP Architectures

• Overview DSP specific hardware

• Types of DSP Architectures

• State of DSP


2

Where DSPs found?

• Consumer products– DVD, MP3, HDTV

• Communications– Radio, Cellular Phones

• Medical– Pacemakers, MRI, Spectrometers

• Industrial– Motor controllers, environmental controls


3

Technologies used for Digital Signal ProcessING

• General Purpose Processors (GPP)– Intel IPX400, IBM PowerPC

• Field Programmable Gate Array (FPGA)– Xilinx, Altera

• Application Specific Integrated Circuit (ASIC)

• DSP Processors (or DSPs)


4

Advantages of DSPs

• Reprogrammable– High level to assembly programming with well

documented tools

• Cost effective in low volume applications– Modestly priced development hardware

• Desirable speed / cost / efficiency for evaluation of design trade-offs


5

Motivations for DSP Architecture

• Most features of DSPs have the structures of a DSP algorithm– Fast multipliers– Multiple Execution Units– Efficient Memory access– Data formats– Streamlines I/O– Special instruction sets


6

• Case study: The dot product (FIR filter)

– Requires fast Multiplication and addition of product terms referred to as accumulation.

– MAC (Multiply - ACcumulate)

The Fast Multiplier

0 1 N

0 1 N

where

{ , , , }

{ , , , }

and is the number of terms

y h x

h h h h

x x x x

N


7

The Fast Multiplier(cont)

• Other MAC intensive algorithms include:

– convolution,– correlation,– IIR filtering, – Fourier transform

• All DSPs provide single cycle MAC


8

The Multiple Execution Unit

• High computation complexity requires multiple types of arithmetic and logic to be preformed in parallel to increase speed.

• Parallel execution units will have:– MAC– ALU– Shifter


9

The Multiple Execution Unit (cont.)

• Example: Texas Instruments C6713 DSP– It has 8 execution units ( two sections

called .L, .M, .S, .D )– All blocks are multi-function and can add

Integers– Only the .M blocks can multiply– .D is mostly used to save and store data– .L mostly does logical operations – .S mostly does shifting


10

Efficient Memory Access

• Performing a MAC every cycle would require:– (1) instruction (Perform MAC)– (1) retained accumulated value– (2) data ( Two new terms to multiply)

(Lots of Bandwidth for data is required)


11

Efficient Memory Access(cont)

• Technique used to provide greater bandwidth– Use of separate Instruction bus and multiple data

buses– Caching of instruction so they don’t consumer

bandwidth– Providing L1 and L2 caching for single cycle access– Automatic address generation that is incremented

without hardware intervention.• FIFO buffers• Auto increment for arrays access• Delay lines


12

DSP Data formats

• Most DSPs use Fixed Point Math which is a form of Integer math– Smaller area used on chip– Less power– Decrease cost– Lower precision

• For application which require high precision floating point math is used


13

DSP Data Formats (cont.)

• DSP processors use the shortest data word width that can provide the required accuracy– 16-bit fixed point is common– 24 or 32 bit hardware is for better accuracy– Accumulation registers may have more bits to

avoid overflow / saturation


14

Zero overhead looping

• Special loop instructions or hardware is used to implement “for loops” such that it will automatically– Store and update an increment – Evaluate a conditional– Repeat only a specific number of times


15

Streamlined I/O

• Built-in types of Input/Output (I/O) units are provide on the DSP chip such as– Serial and/or parallel ports– Low overhead interrupt handling– Direct Memory Access (DMA) to provide data

to on chip memory without processor intervention


16

The Special Instruction Set

• Special instruction sets are provided for a DSP to maximize hardware use. – Perform parallel operations with single intructions– Include data operations, pointer updates and

arithmetic simultaneously

• They also can minimize memory space used– Use of smaller instruction word length– Change of Hardware configuration to decrease

instructions that are used

• Specialized instruction may require coded in assembly due to added complexity


17

Conventional DSP Processors

• Low-cost / Good performance– Operated at 20 to 50 Mhz– Include only 1 MAC an ALU and few additional

execution units• Target applications include

– Disk drives– Answering machines

• Examples: – ADI ADSP-21xx, – TI TMS320C2xxx, – Freescale (motorola) DSP 560xx


18

Conventional DSP Processors(cont)

• Mid-range higher performance– Increaseed speed– Extra hardware such as shifter or instruction

cache– Deeper pipeline

• Example include– Freescale 563xx– TI TMS320C5xxx


19

Enhanced Conv. DSP Processors

• Maintain cost/performance trade-off for more intensive applications

• Provide Multiple MACs with ALU and shifter with parallel execution.

• Extended instruction sets• Wider data/instruction• Examples include:

– Lucent DSP16xxx– ADI Blackfin BF533


20

Multi-issue Architecture

• Motivation: – Enhance-conv DSP use such specialized

hardware ad compound instructions, they require assembly programming

– Provide friendly structures for C compiler optimization

– Use simpler instructions in parallel groups


21

Types of Multi-issue Architecture

• VLIW( Very Long Instruction Word) or its hybrid VelocTI (both are TI inventions)– Many execution units each having its own instruction.

Four to eight executions to cycle– Instructions and executions are group at the time the

program is compile/assembled• Superscalar

– Uses special hardware to group instruction and executions in parallel

– Extremely complex design and not deterministic (same code many execute in different orders each time its call changing the time taken to perform)


22

State of DSPs

• They have diversified• They are a key part in designs to

– increasing speed– decreasing energy consumption– decrease memory usagewhile balancing cost.

• There is a need for more efficient compilers because assembly is too time intensive

Introduction to Digital Signal ProcessORS

Documents

parallel execution units

multiple execution unit

store data

multiple types of arithmetic

accumulated value2 data

fourier transformall

addition of product

new terms