Lecture 11 Oct 12 Floating-point numbers Circuits for floating-point operations Floating-point Representation - Basics Floating-point numbers - provide a dynamic range of representable real numbers Want to represent very large numbers and very small numbers More importantly, we want to represent moderate size numbers (like p or e) with great precision. Representation - similar to scientific notation Floating Point Standard Defined by IEEE Std 754-1985 Developed in response to divergence of representations Portability issues for scientific code Now almost universally adopted Two representations Single precision (32-bit) Double precision (64-bit) Two parts - significand (or mantissa) M and exponent (or characteristic) E The floating-point number F represented by the pair (M,E) has the value F=ME ( - base of exponent) Base - common to all numbers in a given system - implied - not included in the representation of a floating point number Example Floating Point Format • M = largest possible exponent, m = smallest possible exponent • Smallest number is 1/8, Largest number is 7/4 • Smallest gap is 1/32 • Largest gap is 1/4 IEEE 754 floating-point standard Leading “1” bit of significand is implicit Exponent is “biased” to make sorting easier all 0s is smallest exponent all 1s is largest bias of 127 for single precision and 1023 for double precision sign x (1+significand) x 2exponent – bias summary: (–1) Example: decimal: -.75 = - ( ½ + ¼ ) binary: -.11 = -1.1 x 2-1 floating point: exponent = 126 = 01111110 IEEE single precision: 10111111010000000000000000000000 IEEE Floating-Point Format single: 8 bits double: 11 bits S Exponent single: 23 bits double: 52 bits Fraction x (1)S (1 Fraction) 2(Exponent Bias) S: sign bit (0 non-negative, 1 negative) Normalize significand: 1.0 ≤ |significand| < 2.0 Always has a leading pre-binary-point 1 bit, so no need to represent it explicitly (hidden bit) Significand is Fraction with the “1.” restored Exponent: excess representation: actual exponent + Bias Ensures exponent is unsigned Single: Bias = 127; Double: Bias = 1023 Precision n bits partitioned into two parts - significand M and exponent E n bits – 2n different values Range between smallest and largest representable values increases distance between any two consecutive values increases Floating-point numbers sparser than fixed-point numbers - lower precision Real number between two consecutive floating-point numbers is mapped onto one of the two A larger distance between two consecutive numbers results in a lower precision of representation Example: What decimal representation of real number is represented by the following 32 bit string (in IEEE format)? 11000000101000…0 Answer: Exponent is 210 (=12910 – 12710) Sign is negative. Significand value is (1.01000 …)2 = 1.2510 Value of the number is –(22 x 1.25)10 = - 510 Design choices for FP format • Recall that significand (mantissa) uses sign- magnitude format, exponent uses excess notation. • Choices mainly motivated by need to compare two FP numbers very quickly. • Order relation is determined by comparing the exponents first. Normalized and denormalized numbers If the exponent is between 1 and 254, a normal real number is represented. If the exponent is 0: • if significand is 0, then value = 0. • if significand is not zero, it represents a denormalized number. b1 b2 … b23 represents 0. b1 b2 … b23 rather than 1.b1b2 … b23 Why? To reduce the chance of underflow. Nonstandard numbers When the exponent is 255, the number represented is • + or – infinity (if the significand is 0) • NaN (otherwise) Example: What do the following numbers represent? 0 0 0 0 0 0 0 0 0 1 0 0 0 …0 1 0 0 0 0 0 0 0 0 0 1 0 0 …0 Answer: The first one represents 0.510 The second one represents -0.2510 Set of representable numbers using FP format The number of distinct normal numbers that can be represented using a single-precision floating point format is 232 – 224. Reason: There are no ambiguous representations, i.e., two distinct 32 bits representing normal numbers represent two distinct real numbers. Main advantage of double precision FP format is that it can represent “typical” real numbers more precisely. The expanded range of representable numbers (upto 2.0 x 10308) is less important. Absolute and relative error in FP representation A real number r in the range -2127 x(2- 2-23) <= r <= 2127 x(2- 2-23) is either representable (exactly), or there exist two representable real numbers r1 and r2 such that r1 < r < r2. r should be represented as r1 (if |r – r1| < |r – r2|) or r2 (otherwise). The (absolute) error is |r – r’| where r is the real number being represented, and r’ is the chosen representation. Relative error is |r – r’|/|r’| Problem: Show the representation of the following decimal numbers in the ANSI/IEEE short and long floating-point formats. When the number is not exactly representable, use the round to nearest even rule. –6.25 Problem: Show the representation of the following decimal numbers in the ANSI/ISSS short and long floatingpoint formats. When the number is not exactly representable, use the round to nearest even rule. –6.25 Answer: sign bit is 1 6.25 = 22 x 1.5625 = 1 10000001 10010000 … 0 This is exact representation. Problem: What are the absolute and relative errors in the representation of the real number 0.110 ? (Use the standard rounding rule.) FP representation of r = 0.1 is r1 = 0 00000000 0001100110011001100 (in base 2) r1 = .099999904632568359375 (in base 10) Absolute error is ~ 0.00000009 Relative error is ~ 0.0000009 = 9 x 10-7 Relative error is more or less independent of the actual value of the number. Relative error for a large number Example: Consider a large number, e.g. 1035. What is its FP representation? What is the relative error in its representation? Relative error associated with a large number Example: Consider a large number, e.g. 1035. What is its FP representation? What is the relative error in its representation? Solution: Step 1: Write 1035 as 1035 = 2E y where y is in [1, 2). Step 2: Write E in excess 127 notation. Step 3: Write y – 1 in base ½. Step 1: Write 1035 as 1035 = 2E y where y is in [1, 2). We can compute E using the c++ code: E = 1; val = 2.0; while (val < 1.0E35) { val *= 2; E++; } E--; y = 1.0E35/val; The values of E and y are computed as follows: E = 116 y = 1.203706215242022408 Step 2: Write E = 116 in excess 127 notation. Step 3: Write y – 1 in base ½. 11110011 y – 1 = 0.203706215242022408 Converting y – 1 in binary: 0.203706215242022408 x 2 = 0. 407412430484044816 < 1 so first bit is 0 0.407412430484044816 x 2 = 0.814824860968089632 < 1 so second bit is 0 0.814824860968089632 x 2 = 1.629649721936179264 > 1 so 3rd bit is 1 1.629649721936179264 – 1 = 0.629649721936179264 0.629649721936179264 x 2 = 1.259299443872358528 > 1 so 4th bit is 1 etc. The final result (correct to 23 bits is): 001101000010011000010111 The FP representation of x = 1035 is y = 0 11110011 001101000010011000010111 Relative error in the FP representation of 1035 Let x = 1035 and y be the number whose FP representation is 0 11110011 001101000010011000010111 What is (the value of) y? y ~ 0.999999991400877188214542197 x 1035 Relative error is |x – y|/ x ~ 0.859912281 x 10-9 Why is the integer 33554431 not exactly representable? Slide source: UIUC, CS 232 class lecture Effect of Loss of Precision According to the General Accounting Office of the U.S. Government, a loss of precision in converting 24bit integers into 24-bit floating point numbers was responsible for the failure of a Patriot anti- missile battery. Effect of Loss of Precision Slide source: UIUC, CS 232 class lecture Floating-Point Addition Consider a 4-digit decimal example 1. Align decimal points 9.999 × 101 + 0.016 × 101 = 10.015 × 101 3. Normalize result & check for over/underflow Shift number with smaller exponent 9.999 × 101 + 0.016 × 101 2. Add significands 9.999 × 101 + 1.610 × 10–1 1.0015 × 102 4. Round and renormalize if necessary 1.002 × 102 Floating-Point Addition Now consider a 4-digit binary example 1. Align binary points 1.0002 × 2–1 + –0.1112 × 2–1 = 0.0012 × 2–1 3. Normalize result & check for over/underflow Shift number with smaller exponent 1.0002 × 2–1 + –0.1112 × 2–1 2. Add significands 1.0002 × 2–1 + –1.1102 × 2–2 (0.5 + –0.4375) 1.0002 × 2–4, with no over/underflow 4. Round and renormalize if necessary 1.0002 × 2–4 (no change) = 0.0625 FP Adder Hardware Much more complex than integer adder Doing it in one clock cycle would take too long Much longer than integer operations Slower clock would penalize all instructions FP adder usually takes several cycles Can be pipelined Floating point addition Start 1. Compare the exponents of the two numbers. Shift the smaller number to the right until its exponent would match the larger exponent 2. Add the significands 3. Normalize the sum, either shifting right and incrementing the exponent or shifting left and decrementing the exponent Overflow or underflow? Yes No 4. Round the significand to the appropriate number of bits No Still normalized? Yes Done Exception FP Adder Hardware Step 1 Step 2 Step 3 Step 4 Addition and Subtraction Exponents of both operands must be equal before adding or subtracting significands Significands aligned by shifting the significand of the smaller operand |E1-E2| base- positions to the right, increasing its exponent, until exponents are equal E1E2 - Exponent of larger number not decreased - this will result in a significand larger than 1 - a larger significand adder required Addition/Subtraction – post-normalization Addition - resultant significand M (sum of two aligned significands) is in range 1/ M < 2 If M>1 - a postnormalization step - shifting significand to the right to yield M3 and increasing exponent by one - is required (an exponent overflow may occur) Subtraction - Resultant significand M is in range 0 |M|< 1 - postnormalization step - shifting significand to left and decreasing exponent - is required if M<1/ (an exponent underflow may occur) In extreme cases, the postnormalization step may require a shift left operation over all bits in significand, yielding a zero result Example F1=(0.100000)16 163 ; F2=(0.FFFFFF)16 162 Short IBM format ; calculate F1 – F2 Significand of smaller number (F2) is shifted to the right - least-significant digit lost Shift is time consuming - result is wrong Example - Continued Correct result (with “unlimited" number of significand digits) Error (also called loss of significance) is 0.1 16-2 - 0.1 16-3 = 0.F 16-3 Solution to problem - guard digits - additional digits to the right of the significand to hold shiftedout digits In example - a single (hexadecimal) guard digit is sufficient Steps in Addition/Subtraction of Floating-Point Numbers Step 1: Calculate difference d of the two exponents - d=|E1 - E2| Step 2: Shift significand of smaller number by d base positions to the right Step 3: Add aligned significands and set exponent of result to exponent of larger operand Step 4: Normalize resultant significand and adjust exponent if necessary Step 5: Round resultant significand and adjust exponent if necessary Floating Point Complexities Operations are somewhat more complicated (see text) In addition to overflow we can have “underflow” Accuracy can be a big problem IEEE 754 keeps two extra bits, guard and round four rounding modes positive divided by zero yields “infinity” zero divide by zero yields “not a number” other complexities Implementing the standard can be tricky Not using the standard can be even worse see text for description of 80x86 and Pentium bug! FP Arithmetic Hardware FP multiplier is of similar complexity to FP adder FP arithmetic hardware usually does But uses a multiplier for significands instead of an adder Addition, subtraction, multiplication, division, reciprocal, square-root FP integer conversion Operations usually takes several cycles Can be pipelined FP Instructions in MIPS FP hardware is coprocessor 1 Adjunct processor that extends the ISA Separate FP registers 32 single-precision: $f0, $f1, … $f31 Paired for double-precision: $f0/$f1, $f2/$f3, … Release 2 of MIPs ISA supports 32 × 64-bit FP reg’s FP instructions operate only on FP registers Programs generally don’t do integer ops on FP data, or vice versa More registers with minimal code-size impact FP load and store instructions lwc1, ldc1, swc1, sdc1 e.g., ldc1 $f8, 32($sp) FP Instructions in MIPS Single-precision arithmetic add.s, sub.s, mul.s, div.s Double-precision arithmetic add.d, sub.d, mul.d, div.d e.g., mul.d $f4, $f4, $f6 Single- and double-precision comparison c.xx.s, c.xx.d (xx is eq, lt, le, …) Sets or clears FP condition-code bit e.g., add.s $f0, $f1, $f6 e.g. c.lt.s $f3, $f4 Branch on FP condition code true or false bc1t, bc1f e.g., bc1t TargetLabel FP Example: °F to °C C code: float f2c (float fahr) { return ((5.0/9.0)*(fahr - 32.0)); } fahr in $f12, result in $f0, literals in global memory space Compiled MIPS code: f2c: lwc1 lwc2 div.s lwc1 sub.s mul.s jr $f16, $f18, $f16, $f18, $f18, $f0, $ra const5($gp) const9($gp) $f16, $f18 const32($gp) $f12, $f18 $f16, $f18 FP Example: Array Multiplication X=X+Y×Z All 32 × 32 matrices, 64-bit double-precision elements C code: void mm (double x[][], double y[][], double z[][]) { int i, j, k; for (i = 0; i! = 32; i = i + 1) for (j = 0; j! = 32; j = j + 1) for (k = 0; k! = 32; k = k + 1) x[i][j] = x[i][j] + y[i][k] * z[k][j]; } Addresses of x, y, z in $a0, $a1, $a2, and i, j, k in $s0, $s1, $s2 FP Example: Array Multiplication MIPS code: li li L1: li L2: li sll addu sll addu l.d L3: sll addu sll addu l.d … $t1, 32 $s0, 0 $s1, 0 $s2, 0 $t2, $s0, 5 $t2, $t2, $s1 $t2, $t2, 3 $t2, $a0, $t2 $f4, 0($t2) $t0, $s2, 5 $t0, $t0, $s1 $t0, $t0, 3 $t0, $a2, $t0 $f16, 0($t0) # # # # # # # # # # # # # # $t1 = 32 (row size/loop end) i = 0; initialize 1st for loop j = 0; restart 2nd for loop k = 0; restart 3rd for loop $t2 = i * 32 (size of row of x) $t2 = i * size(row) + j $t2 = byte offset of [i][j] $t2 = byte address of x[i][j] $f4 = 8 bytes of x[i][j] $t0 = k * 32 (size of row of z) $t0 = k * size(row) + j $t0 = byte offset of [k][j] $t0 = byte address of z[k][j] $f16 = 8 bytes of z[k][j] FP Example: Array Multiplication … sll $t0, $s0, 5 addu $t0, $t0, $s2 sll $t0, $t0, 3 addu $t0, $a1, $t0 l.d $f18, 0($t0) mul.d $f16, $f18, $f16 add.d $f4, $f4, $f16 addiu $s2, $s2, 1 bne $s2, $t1, L3 s.d $f4, 0($t2) addiu $s1, $s1, 1 bne $s1, $t1, L2 addiu $s0, $s0, 1 bne $s0, $t1, L1 # # # # # # # # # # # # # # $t0 = i*32 (size of row of y) $t0 = i*size(row) + k $t0 = byte offset of [i][k] $t0 = byte address of y[i][k] $f18 = 8 bytes of y[i][k] $f16 = y[i][k] * z[k][j] f4=x[i][j] + y[i][k]*z[k][j] $k k + 1 if (k != 32) go to L3 x[i][j] = $f4 $j = j + 1 if (j != 32) go to L2 $i = i + 1 if (i != 32) go to L1 Interpretation of Data The BIG Picture Bits have no inherent meaning Interpretation depends on the instructions applied Computer representations of numbers Finite range and precision Need to account for this in programs Parallel programs may interleave operations in unexpected orders Assumptions of associativity may fail (x+y)+z x+(y+z) -1.50E+38 x -1.50E+38 y 1.50E+38 0.00E+00 z 1.0 1.0 1.50E+38 1.00E+00 0.00E+00 Need to validate parallel programs under varying degrees of parallelism §3.6 Parallelism and Computer Arithmetic: Associativity Associativity Originally based on 8087 FP coprocessor FP values are 32-bit or 64 in memory 8 × 80-bit extended-precision registers Used as a push-down stack Registers indexed from TOS: ST(0), ST(1), … Converted on load/store of memory operand Integer operands can also be converted on load/store Very difficult to generate and optimize code Result: poor FP performance §3.7 Real Stuff: Floating Point in the x86 x86 FP Architecture x86 FP Instructions Data transfer Arithmetic Compare Transcendent al FILD mem/ST(i) FISTP mem/ST(i) FLDPI FLD1 FLDZ FIADDP FISUBRP FIMULP FIDIVRP FSQRT FABS FRNDINT FICOMP FIUCOMP FSTSW AX/mem FPATAN F2XMI FCOS FPTAN FPREM FPSIN FYL2X mem/ST(i) mem/ST(i) mem/ST(i) mem/ST(i) Optional variations I: integer operand P: pop operand from stack R: reverse operand order But not all combinations allowed Streaming SIMD Extension 2 (SSE2) Adds 4 × 128-bit registers Extended to 8 registers in AMD64/EM64T Can be used for multiple FP operands 2 × 64-bit double precision 4 × 32-bit double precision Instructions operate on them simultaneously Single-Instruction Multiple-Data Left shift by i places multiplies an integer by 2i Right shift divides by 2i? Only for unsigned integers For signed integers Arithmetic right shift: replicate the sign bit e.g., –5 / 4 111110112 >> 2 = 111111102 = –2 Rounds toward –∞ c.f. 111110112 >>> 2 = 001111102 = +62 §3.8 Fallacies and Pitfalls Right Shift and Division Who Cares About FP Accuracy? Important for scientific code But for everyday consumer use? “My bank balance is out by 0.0002¢!” The Intel Pentium FDIV bug The market expects accuracy See Colwell, The Pentium Chronicles ISAs support arithmetic Bounded range and precision Signed and unsigned integers Floating-point approximation to reals Operations can overflow and underflow MIPS ISA Core instructions: 54 most frequently used 100% of SPECINT, 97% of SPECFP Other instructions: less frequent §3.9 Concluding Remarks Concluding Remarks Floating-point numbers Denormals allow graceful underflow – Negative numbers max – FLP – min – Sparser Overflow region 0 Denser Positive numbers min + FLP + Denser Underflow example + Sparser Underflow regions Midway example max + Overflow region Typical example Overflow example Distribution of floating-point numbers on the real line. Representation for non-integral numbers Like scientific notation –2.34 × 1056 +0.002 × 10–4 +987.02 × 109 normalized not normalized In binary Including very small and very large numbers ±1.xxxxxxx2 × 2yyyy Types float and double in C §3.5 Floating Point Floating Point Single-Precision Range Exponents 00000000 and 11111111 reserved Smallest value Exponent: 00000001 actual exponent = 1 – 127 = –126 Fraction: 000…00 significand = 1.0 ±1.0 × 2–126 ≈ ±1.2 × 10–38 Largest value exponent: 11111110 actual exponent = 254 – 127 = +127 Fraction: 111…11 significand ≈ 2.0 ±2.0 × 2+127 ≈ ±3.4 × 10+38 Double-Precision Range Exponents 0000…00 and 1111…11 reserved Smallest value Exponent: 00000000001 actual exponent = 1 – 1023 = –1022 Fraction: 000…00 significand = 1.0 ±1.0 × 2–1022 ≈ ±2.2 × 10–308 Largest value Exponent: 11111111110 actual exponent = 2046 – 1023 = +1023 Fraction: 111…11 significand ≈ 2.0 ±2.0 × 2+1023 ≈ ±1.8 × 10+308 Floating-Point Precision Relative precision all fraction bits are significant Single: approx 2–23 Equivalent to 23 × log102 ≈ 23 × 0.3 ≈ 6 decimal digits of precision Double: approx 2–52 Equivalent to 52 × log102 ≈ 52 × 0.3 ≈ 16 decimal digits of precision Floating-Point Example Represent –0.75 –0.75 = (–1)1 × 1.12 × 2–1 S=1 Fraction = 1000…002 Exponent = –1 + Bias Single: –1 + 127 = 126 = 011111102 Double: –1 + 1023 = 1022 = 011111111102 Single: 1011111101000…00 Double: 1011111111101000…00 Floating-Point Example What number is represented by the single-precision float 11000000101000…00 S=1 Fraction = 01000…002 Fxponent = 100000012 = 129 x = (–1)1 × (1 + 012) × 2(129 – 127) = (–1) × 1.25 × 22 = –5.0