What are the characteristics of DSP algorithms? M. Smith and S. Daeninck
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Tackled today What are the basic characteristics of a DSP algorithm? Information on the TigerSHARC arithmetic, multiplier and shifter units Practice examples of C++ to assembly code conversion DSP Introduction, M. Smith, ECE, University of Calgary, Canada
IEEE Micro Magazine Article How RISCy is DSP? Smith, M.R.; IEEE Micro, Volume: 12, Issue: 6, Dec. 1992, Pages:10 - 23 Available on line via the library “Electronic web links” Copy placed on ENCM515 Web site. Make sure you read it before midterm DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Characteristics of an FIR algorithm Involves one of the three basic types of DSP algorithms FIR (Type 1), IIR (Type 2) and FFT (Type 3) Representative of DSP equations found in filtering, convolution, correlation (Lab) and modeling Multiplication / addition intensive Simple format within a (long) loop Many memory fetches of fixed and changing data Handle “infinite amount of input data” – need FIFO buffer when handling ON-LINE data All calculations “MUST” be completed in the time interval between samples DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada FIR Input Value must be stored in circular buffer Filter operation must be performed on circular buffer For operational efficiency – Note that latest value is the “last in the array” Xarray = {Xm-1, Xm-2, Xm-3, … X1, X0 } Harray = {Hm-1, Hm-2, Hm-3, … H1, H0 } DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada FIR COMMON MISTAKE = MUCH WASTED LAB. TIME For operational efficiency – Note that latest value is the “last in the array” Xarray = {Xm-1, Xm-2, Xm-3, … X1, X0 } Harray = {Hm-1, Hm-2, Hm-3, … H1, H0 } Can work with latest value “first in the array” when doing C++, but does not work for assembly code optimization DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada FIR X[n – 1] = NewInputValue Into last place of Input Buffer Sum = 0; For (count = 0 to N – 1) -- N of size 100+ Xvalue = X[count]; Hvalue = H[count]; Product = Xvalue * Hvalue; Sum = Sum + Product; Multiply and Accumulate -- MAC NewOutputValue = Sum; Update Buffer – The T-operation in the picture For (count = 1 to N – 1) -- Discard oldest X[0]; X[count – 1] = X[count]; DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Comparing IIR and FIR filters Infinite Impulse Response filters – few operations to produce output from input for each IIR stage 3 – 7 stages Finite Impulse Response filters – many operations to produce output from input. Long FIFO buffer which may require as many operations As FIR calculation itself. Easy to optimize DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada IIR -- Biquad For (Stages = 0 to 3) Do S0 = Xin * H5 + S2 * H3 + S1 * H4 Yout = S0 * H0 + S1 * H1 + S2 * H2 S2 = S1 S1 = S0 This second solution gives DIFFERENT result. Order of calculation is different. The actual output difference depends on how frequently samples are taken relative to how rapidly the signal changes CALCULATION SPEED IS DIFFERENT DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada We need to know how the processor architecture affects speed of calculation Register File and Compute Block Volatile registers Data Summation Multiply and Accumulate (MAC) DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Register File and COMPUTE Units Key Points DAB – Data Alignment Buffer (special for quad fetches NOT writes) Each block can load/store 4x32bit registers in a cycle. 4 inputs to Compute block, but only 3 Outputs to Register Block. Highly parallel operations UNDER THE RIGHT CONDITIONS DSP Introduction, M. Smith, ECE, University of Calgary, Canada
NOTE – DATA PATH ISSUES OF THE X-REGISTER FILE 1 output path (128 bit) TO memory 2 input paths FROM memory 4 output (64-bit) paths TO ALU, multiplier, shifter 3 input paths (64-bit) FROM ALU, multiplier, shifter NUMBER OF PATHS HAS IMPLICATIONS ON WHAT THINGS CAN HAPPEN IN PARALLEL DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Register File - Syntax Key Points Each Block has 32x32 bit Data registers Each register can store 4x8 bit, 2x16 bit or 1x32 bit words. Registers can be combined into dual or quad groups. These groups can store 8, 16, 32, 40 or 64 bit words. XSR3:2 -> 4x16 bit words XFR1:0 -> 1x40 bit float XR7 -> 1x32 bit word XBR3:0 -> 16x8 bit words Multiple of 4 Multiple of 2 XLR7:6 -> 1x64 bit word Register Syntax DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Register File – BIT STORAGE Both 32 bit and 64 bit registers 128 bit examples are not shown but they are the same. DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Volatile Data Registers Non-preserved during a function call Volatile registers – no need to save 24 Volatile DATA registers in each block XR0 – XR23 YR0 – YR23 2 ALU SUMMATION registers in each block XPR0, XPR1, YPR0, YPR1 5 MAC ACCUMULATE registers in each block XMR0 – XMR3, YMR0 – YMR3 XMR4, YMR4 – Overflow registers PR stands for parallel results register MR stands for Multiplier results register DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Arithmetic Logic Unit (ALU) 2x64 bit input paths 2x64 bit output paths 8, 16, 32, or 64 bit addition/subtraction - Fixed-point 32 or 64 bit logical operations - fixed-point 32 or 40 bit floating-point operations Can do the same on Y ALU AT THE SAME TIME DAB – Data Alignment Buffer(2x128 bit FIFO)-> used to align misaligned quad or dual 32 bit data loads DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Sample ALU Instruction Example of 16 bit addition XYSR1:0 = R31:30 + R25:24 Performs “short” addition in X and Y Compute Blocks XR1.HH = XR31.HH + XR25.HH XR1.HL = XR31.HL + XR25.HL XR0.LH = XR30.LH + XR24.LH XR0.LL = XR30.LL + XR24.LL YR1.HH = YR31.HH + YR25.HH 8 additions at the same time YR1.HL = YR31.HL + YR25.HL YR0.LH = YR30.LH + YR24.LH .LH, .HH is my notation YR0.LL = YR30.LL + YR24.LL Other additions/subtractions look the same, but use 32 or 8 bits DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Sample ALU Instructions A neat instruction is the sideways addition sum (SUM) Fixed-Point long word, word, short word, byte (char) Floating-Point Single, double precision DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Pass is an interesting instruction XR4 = R5 Assignment statement -- makes XR4 XR5 XR4 = PASS R5 Still makes XR4 XR5 BUT USES A DIFFERENT PATH THROUGH THE PROCESSOR Sets the ALU flags (so that they can be used for conditional tests) PASS instructions can be put in parallel with different instructions than assignments DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Example code – parallel operations occurring int x_two = 64, y_two = 16; int x_three = 128, y_three = 8; int x_four = 128, y_four = 8; int x_five = 64, y_five = 16; int x_odd = 0, y_odd = 0; int x_even = 0, y_even = 0; x_odd = x_five + x_three; x_even = x_four + x_two; y_odd = y_five + y_three; y_even = y_four + y_two; XR2 = 64;; XR3 = 128;; XR4 = 128;; XR5 = 64;; YR2 = 16;; YR3 = 8;; YR4 = 8;; YR5 = 16;; XYR1:0 = R5:4 + R3:2;; //XR1 = x_odd, XR0 = x_even //YR1 = y_odd, YR1 = y_even WRONG SYNTAX nice example of the tigerSharc, it accomplishes thecode in less lines than C DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Multiplier Operates on fixed, floating and complex numbers. Fixed-Point numbers 32x32 bit with 32 or 64 bit results 4 (16x16 bit) with 4x16 or 4x32 bit results Data compaction inputs – 16, 32, 64 bits, outputs 16, 32 bit results Floating-Point numbers 32x32 bit with 32 bit result 40x40 bit with 40 bit result COMPLEX Numbers 32x32 bit with results stored in MR register FIXED-POINT ONLY Complex – imaginary part is in the MSB part of the 32 bit word DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Multiplier XR0 = R1*R2;; XR1:0 = R3*R5;; XMR1:0 = R3*R5;; //uses XMR4 overflow XR2 = MR3:2, XMR3:2 = R3*R5;; XR3:2 = MR1:0, XMR1:0 = R3*R5;; XFR0 = R1*R2;; // 32 bit mult – 24 bit mantissa XFR1:0 = R3:2*R5:4;; //40 bit MULTIPLY //32 bit mantissa // high precision float XR2 = MR3:2, XMR3:2 = R3*R5;; if integer multiply, R2 gets MR2, if Fractional gets MR3 MR stands for Multiplier results register DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Multiplier --- with 32 or 16 bit results Note minor changes in syntax XR5:4 = R1:0*R3:2;;(16 bit results) XR7:4 = R3:2*R5:4;; (32 bit results) XMR1:0 += R3:2*R5:4;;(16 bit results) XMR3:0 += R3:2*R5:4;; (32 bit results) XR3:2 = MR3:2, XMR3:2 = R1:0*R5:4;; (16 bit results) one instruction XR3:0 = MR3:0, XMR3:0 = R1:0*R5:4;; (32 bit results) 16 bit multiplies results can be 16 bit or 32 bit RED for 16 bit, Blue for 32 bit MR4 contains four overflow bits for every MR register No need to tell the instruction if it is a short or normal word DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Practice Examples Convert from “C” into assembly code – use volatile registers long int value = 6; long int number = 7; long int temp = 8; value = number * temp; BAD DESIGN OF FLOATING PT CODE WILL INTRODUCE MANY ERRORS RE-WRITE CODE TO FIX float value = 6; float number = 7; long int temp = 8; value = number * temp; DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Avoiding common design errors XR12 = 6.0;; //valueF12 // Sets XFR12 6.0 XR13 = 7.0;;//numberF13 XR18 = 8;; //tempR18 //(float) tempR18 XFR18 = FLOAT R18;; //valueF12 = numberF13 * tempF18 XFR12 = R13 * R18;; Convert from “C” into assembly code – use volatile registers float value = 6.0; (XFR12) float number = 7.0; (XFR13) long int temp = 8; (XR18) value = number * temp; // Treat as value = number * (float) temp; XFR23:22 = R21*R22;; not allowed DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Shifter Instructions FEXT – bit field extraction, FDEP – bit field deposit 2x64 bit input paths and 2x64 bit output paths 32, or 64 bit shifting operations 32 or 64 bit manipulation operations DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Examples --- shift only integers There is a FSCALE for floats (not shifter) long int value = 128; long int high, low; low = value >> 2; high = value << 2; POSITIVE VALUE – LEFT SHIFT NEGATIVE VALUE – RIGHT SHIFT XR0 = 2;; XR1 = -XR2;; XR2 = 128;; //low = value >> 2; XR23 = ASHIFT XR2 BY –2;; Or XR23 = ASHIFT XR2 BY XR1;; //high = value << 2; XR22 = ASHIFT XR2 BY 2;; XR22 = ASHIFT XR2 BY XR0;; DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada ALU instructions Under the RIGHT conditions can do multiple operations in a single instruction. Instruction line has 4x32 bit instruction slots. Can do 2 Compute and 2 memory operations. This is actually 4 Compute operations counting both compute blocks. One instruction per unit of a compute block, ie. ALU. Since there are only 3 result buses, only one unit (ALU or Multiplier) can use 2 result buses. Not all instructions can be used in parallel. DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Dual Operation Examples FRm = Rx + Ry, FRn = Rx – Ry;; Note that uses 4(8) different registers and not 6(12) FR4 = R2 + R1, FR5 = R2 - R1;; The source registers used around the + and – must be the same. Very useful in FFT code Can be floating(single or extended precision) or fixed(32 or 64 bit) add/subtract. Rm = MRa, MRa += Rx * Ry;; MRa must be the same register(s) (MR1:0 or MR 3:2) Can be used on fixed(32 or 64 bit results) COMPLEX numbers (on 16 bit values) Rm = MRa, MRa += Rx ** Ry;; DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Practice Examples Convert to assembly code Convert from “C” into assembly code – use volatile registers long int value = 6; long int number = 7; long int temp = 8; value = number * temp; #define value_XR12 XR12 Assignment operation value_XR12 = 6;; Multiply operations value_XR12 = R5 * R6; DSP Introduction, M. Smith, ECE, University of Calgary, Canada
Avoiding common design errors Convert to assembly code float value = 6.0; float number = 7.0; long int temp = 8; value = value + 1; number = number + 2; temp = value + number; Questionable if XFR12 = 1.0;; is allowed, assembler complains XR23:22 = 10.0;; not allowed, there may not be an immediate load for 40 bit floats DSP Introduction, M. Smith, ECE, University of Calgary, Canada
DSP Introduction, M. Smith, ECE, University of Calgary, Canada Tackled today What are the basic characteristics of a DSP algorithm? Information on the TigerSHARC arithmetic, multiplier and shifter units Practice examples of C++ to assembly code conversion DSP Introduction, M. Smith, ECE, University of Calgary, Canada