2000/03/051 This presentation will probably involve audience discussion, which will create action items. Use PowerPoint to keep track of these action items during your presentation In Slide Show, click on the right mouse button Select “Meeting Minder” Select the “Action Items” tab Type in action items as they come up Click OK to dismiss this box This will automatically create an Action Item slide at the end of your presentation with your points entered. Processor Requirements needed to optimize DSP performance M. R. Smith, Electrical and Computer Engineering, University of Calgary, Alberta, Canada ucalgary.ca

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 2 / 48 To be tackled today Characteristics of DSP algorithms Specialized handling of Multiplication Division (21K has no division instruction) ENCM515 Reference Material How RISCy Is DSP, IEEE Micro (Jan-10) Simply Signal Processing (Jan-40) Fast Scaling, CCI (Apr-10) Saturation Arithmetic (Apr-20)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 3 / 48 DSP Algorithms DSP algorithms require specialized features on processors Processors are a compromise speed, cost, silicon When have you as a designer found a compromise that meets your requirements? As a consultant may have to add DSP characteristics to an existing system or add DSP coprocessor to an existing system

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 4 / 48 FIR Multiply/Addition intensive Sum operation with high precision -- overflow considerations Long simple loop Online operation -- “infinite” amount of data Store coefficients on-chip for fast access Complex domain arithmetic

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 5 / 48 IIR-1 Interrelated and order dependent multiplications and additions Small number of delays via register moves? short loop -- low number of instructions in loop which makes it difficult to optimize Precision -- very important because of feedback Multiple stages -- I.e. IIR follows IIR etc

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 6 / 48 IIR-2 LDI Short complicated loop Many intermediate values Pipeline issues because of interdependence

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 7 / 48 FFT Complex variables (A and B) and fixed coefficients (W) Address calculations complex Memory accesses numerable Multiplication and additions Need for fast access to many registers, address pointers, constants, variables

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 8 / 48 Fast instruction cycle -- needed DSP chips -- two cycle instructions (on top of FETCH/DECODE) during which the processor performs many parallel operations More recent technology -- 1 clock cycle Many processors takes 6 to 32 cycles to handle MULT, FMULT, FDIV or even FADD Make processor highly pipelined -- pipeline must be started and then kept full FIR (easy to pipeline) IIR (hard to pipeline) FFT (challenging to pipeline)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 9 / 48 Loop Overhead -- must be minimized Use specialized hardware specialized decrement and branch instructions occurring in a single cycle instruction cached with counter superscalar operations delayed branches hardware loop control Use specialized software techniques loop unrolling down counting loops

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 10 / 48 Memory operations -- Many of them Data/instruction and data/data conflicts Data caches Will also have external data memory banks Harvard architecture branch target caches multi-ported memory register pre-forwarding -- avoid stalls while trying to write back result of ALU operation only to re-- access the same register large register banks -- avoid memory ops associated with just calculated values

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 11 / 48 Precision -- high but without speed loss FIR -- accumulated value can grow big IIR -- recursive use of a value External Memory bus width Internal Memory bus width Data width of registers and ALU Saturation arithmetic

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 12 / 48 Saturation Arithmetic For full discussion see 21K SHARC user manual and also “Being Assertive with your processor” (APR-20) Internal register 80 bits but external busses only 32 wide 0xFFFF F stored as F xFFFF stored as (normal math) stored as (saturation) Can be good solution (FIR) or bad solution (IIR) to the problem of overflow

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 13 / 48 Complex arithmetic -- frequency domain operations Need to fetch real and imaginary parts in at different times during the algorithm Need fast access to adjacent memory locations -- burst memory Need for many internal registers to temporarily store real/imaginary components (FFT butterfly and last years exams) Duplication of resources -- was custom, but consider now 21160

2000/03/0514 DAG 2 8 x 4 x 32 DAG 1 8 x 4 x 32 CACHE MEMORY 32 x 48 PROGRAM SEQUENCER PMD BUS DMD BUS 32 PMA BUS PMD DMD PMA 32DMA BUS DMA 64 JTAG TEST & EMULATION FLAGS TIMER TigerSHARC ADSP Core Architecture BUS CONNECT FLOATING & FIXED- POINT MULTIPLIER, FIXED-POINT ACCUMULATOR REGISTER FILE 16 x BIT BARREL SHIFTER FLOATING-POINT &FIXED-POINT ALU FLOATING & FIXED- POINT MULTIPLIER, FIXED-POINT ACCUMULATOR REGISTER FILE 16 x BIT BARREL SHIFTER FLOATING-POINT &FIXED-POINT ALU

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 15 / 48

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 16 / 48 Address calculations -- frequent Complex addressing modes -- take many clock cycles Use pointers and autoincrement rather than calculating pointer + offset need many address-related registers address calculations compete with ALU calculations group instructions within program e.g. read and store often use same or similar addresses so don’t recalculate the addresses.

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 17 / 48 Specialized addressing modes standard memory access premodify postmodify circular buffers (modulo arithmetic on the address registers) bit-reverse addressing structure handling auto-increment with size accounted for

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 18 / 48 Key issue -- ease of development Microcontrollers -- onboard peripherals Host communication Multiprocessor communications Simulators Multi-processor operations Application notes Good working environment Compatibility to previous processor versions -- legacy code (advantage and a disadvantage)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 19 / 48 Multiplication Extensive algorithms Off-chip multipliers have big bottlenecks Get and then give instruction to multiplier Get and then give first, second data to multiplier Wait till cooked, and then get value Newer chips have on-board multiplication or intelligent co-processors (F-LINE exceptions) Many chips do multiplication using specialized techniques introduced by optimizing compiler

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 20 / 48 Smart Multiplication through optimizing compiler techniques 29K RISC FMULT execution takes 6 cycles + fetch 16bit x 16bit INTEGER multiplication on 68K CISC takes 70 cycles regardless of operations Use adds and shift instead since these take less time -- easy with integer, but floats? What are equivalent operations on 21K. Discussed in early lecture on Quirks and SHARCs

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 21 / 48 Smart Integer 68k Multiplication Multiplication by 2, 4, 8, 16 Achieved by shifting 1, 2, 3 or 4 times (done in 6 + 2n operations on 68K) D2 = D0 * 19 MOVE.W D0, D2 ASL.W #4, D2D2 = D0 * 16 ADD.W D0, D2D2 = D0 * 17 ASL.W #1, D0D0 = D0 *2 ADD.W D0, D2 D2 = D0 * 19 (29 cycles compared to 70) Watch out for overflow, may need conversion to 32 bits (SSI, SSF on some processssors -- not only 21k) Waste of time if have single cycle multipliers (21k?). Careful because multiplication results may end in special register.

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 22 / 48 Multiplication Extensive algorithms Highly pipelined, therefore complex instruction interdependence R0 = R1 * R2BUTR0 = R1 * R2 R3 = R4 * R5R3 = R0 * R5 <- delay dependency Need automated tools to schedule instructions Need multiple destinations (registers) for multiplier result Multiple and Accumulate (MAC) instruction Super-scalar operations even on a simpler processor Cause problems in short loops Many types of MACs needed Not all processors have the single cycle multiplication operation See “In the AM29050 a FIR-bearing animal” (FEB-80 in class notes))

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 23 / 48 Typically need “Normalization” of result N point DFT Result = DFT (Input) ; 0 <= n < N N point inverse DFT Result = IDFT (Input) / N ; 0 <= n < N Division is typically done by the equivalent of repeated subtraction cycles on 68K result = 0; do { Numerator = Numerator - Denom; result++; } while (Numerator > 0); result--; Special shift-subtract tricks speed operations

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 24 / 48 Smart Integer Division Division by 2, 4, 8, 16 unsigned signed LSL #1, D0ASL #1, D0 Need to propagate (or not propagate) the sign bit Unsigned original = 0x80 (128) final = 0x40 (64) Signed original = 0x80 ( - 128) final = 0xC0 ( - 64)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 25 / 48 Floating Point Division The FDIV on 29K takes 15 cycles There is not a FDIV on the 21K -- use recursion!!

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 26 / 48 Why is floating point so difficult? NumberInternal representation 1.00x3F x x41 FF D9 9A x44 7F D9 9A = / 32 = / 2^5

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 27 / 48 Why is floating point so difficult? “Fast scaling Routine for Floating-point RISC and DSP processors” (APR-10) Floating Point Format S bexp frac

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 28 / 48 Floating point number K s (bexp -127) (-1) x 1.frac x = 0x1.0 x 2 0 ( ) (-1) x 0x x 2

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 29 / 48 Floating point number K s (bexp -127) (-1) x 1.frac x = 0x10.0 = % = % x 2 (0x1.4 x 2 ) 0 ( ) (-1) x 0x x 2

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 30 / 48 IEEE Std. 754, 1985 NumberInternal s bexp frac representation 1.00x3F x7F 0x x x84 0x x41 FF D9 9A 0 0x83 0x7F D9 9A x44 7F D9 9A 0 0x88 0x7F D9 9A 1.frac -- only fractional part is stored Remember JAMES BOND helped by M (Smith) “The ONE is remembered and not stored”

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 31 / 48 Fast floating pt division possible NumberInternal s bexp frac representation 1.00x3F x7F 0x x x84 0x BEXP DIFF = x41 FF D9 9A 0 0x83 0x7F D9 9A x44 7F D9 9A 0 0x88 0x7F D9 9A BEXP DIFF = 5 K = K / flip the sign bit with XOR instruction p K = K / N where N = 2 -- decrease bexp = bexp -5

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 32 / 48 Fast Floating Point Division by 32 Doing it 29K -- FP# K is in gr96 Setting up the power CONST BEXPchange, 5 Setting up the bexp-diff SLL BEXPchange, BEXPchange, 23 result = K / 32 SUB result, K, BEXPchange <- REPEATED Note -- when processing a large array -- only the last step needed for every number (inside the loop)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 33 / 48 Fast Floating Point Division by FP M when M is known to be 2^p F0 = 1.0 R0 = R8 - R0 // NOTE integer operation Setting up the bexp-diff R0 = ASHIFT R0 BY 23 result = K / 32 R4 = R4 - R0 Works because F8 = 32.0 (0x ) F0 = 1.0 (0x3F )

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 34 / 48 PROBLEMS? Try to do 0 / 32 Get a large negative number Number s bexp frac x00 0x subtract 0 0x05 0x * 10^37 1 0xFB 0x If dividing by 2^p -- problems if number is smaller than 2^(p-127) Must be overcome on many processors Non-issue on 21k which has single cycle multiplication and division. Calculate reciprocal and then multiply

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 35 / 48 Must guarantee result 68K, 29K, MIPS and 21k problems ADD.W R0, R1ADD gr96, gr97, gr98 Every addition (subtraction) result has the possibility of being out of range -- overflow. Must be tested. 68K solution ADD.W R0, R1 BVS Somewhere <- Test takes cycles 29K and MIPS solution Special instructions -- ADDU and ADDS 21k solution is what?

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 36 / 48 Specialized coding techniques e.g. 29k has the ability of “throwing” SWI as part of compare (ASSERT) Test for FP number too small from previous special Division operation CMP.L #toosmall, D0 68K code BGE okay<- EXTRA cycles always executed MOVE.L #0, D0 BRA continue okay: SUB.L #b_exp, D0 continue: ASGE TRAP#, temp, BEXPchange <- Only “compare” for 29k SUB gr96, gr96, BEXPchange <- Not in a delay slot? where TOOSMALL:CONST gr96, 0 RTI Extra code only executed in the special case that it is needed

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 37 / 48 Specialized conditional instructions on 21k 21K -- F4 contains the FP value -- need F4/32 R0 = 5 R0 = ASHIFT R0 BY 23 F1 = minimum value ( 2^(5-127) ) F2 = ABS F4 COMP (F2, F1) IF GE R4 = R4 - R0 ELSE R4 = R4 - R4 <- NO DELAY Can’t use ELSE R4 = 0 As this not a compute operation but uses 32-bit constant.

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 38 / 48 LIES -- ALL LIES IF GE R4 = R4 - R0 ELSE R4 = R4 - R4 This is not a legal instruction either!! COMPUTE instructions take 22 bits to describe IF JUMP/CALL ELSE R4 = R4 - R4 is allowed Useless approach anyway since there are better ways on 21k to do repeated division by a constant.

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 39 / 48 Processors compared IEEE Micro Magazine Special Feature 1992 DSP TMS320C25, 030 DSP56000/1, DSP96002 (Motorola) RISC i860 (Intel) MC88100 (Motorola) SPARC (Sparc Consortium NOT Sun) Am29050 Ideal -- SMITH CRISP

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 40 / 48 CRISP -- triple pun as well Comprehensive RISC -- Predicted 1992 Harvard architecture MAC (rather than Super -- Scalar instructions) Ability to do X = R+S, Y = R-S operations many registers for address/values FP as well as integer capability Bit-reverse addressing Peripherals with DMA Low power standby High precision -- double precision Efficient pipeline with parallel completion of many operations (dual-ported memory and register banks)

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 41 / 48 Comparisons -- 1

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 42 / 48 FIR/IIR

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 43 / 48 FFT -- Radix 2 and Radix 4

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 44 / 48 Requirements for “perfect” DSP Fast instruction cycle -- different from high clock speed Cycle time adjustable according to instruction type Fast hardware multiplier Floating point for easier algorithm design High precision, implying wide data buses for memory, internal processor transfers, registers and on-board processing units

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 45 / 48 Requirements for “perfect” DSP Several data buses available to reduce bus conflict transfer overhead Harvard architecture and/or instruction cache to avoid instruction and data-fetch clashes Duplicate resources for parallel computation of real and imaginary components of complex numbers Dedicated hardware required for address calculations to avoid APU clash with main algorithm

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 46 / 48 Requirements for “perfect” DSP Extensive temporary registers to reduce unwanted fetches of continually used data Or single cycle, highly parallel, memory operations Fast and reliable, easily programmed, developed and upgraded Inexpensive and easy to develop peripherals High level of customer support Inexpensive to purchase Lower power consumption with a standby mode

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 47 / 48 Requirements for “perfect” DSP Several data buses available to reduce bus conflict transfer overhead Harvard architecture and/or instruction cache to avoid instruction and data-fetch clashes Duplicate resources for parallel computation of real and imaginary components of complex numbers Dedicated hardware required for address calculations to avoid APU

2000/03/05 ENCM Characteristics needed in DSP processors Copyright 48 / 48 Tackled today Characteristics of DSP algorithms Specialized handling of Multiplication Division (21K has no division instruction) ENCM515 Reference Material How RISCy Is DSP, IEEE Micro (Jan-10) Simply Signal Processing (Jan-40) Fast Scaling, CCI (Apr-10) Saturation Arithmetic (Apr-20)