am p a r
play

AM P A R CudA Multiple Precision ARithmetic librarY Floating - PowerPoint PPT Presentation

Dec 19, 2022 •282 likes •662 views

Tight and rigorous error bounds for basic building blocks of double-word arithmetic Mioara Joldes, Jean-Michel Muller, Valentina Popescu JNCF - January 2017 AM P A R CudA Multiple Precision ARithmetic librarY Floating point arithmetics

  1. Tight and rigorous error bounds for basic building blocks of double-word arithmetic Mioara Joldes, Jean-Michel Muller, Valentina Popescu JNCF - January 2017 AM P A R CudA Multiple Precision ARithmetic librarY

  2. Floating point arithmetics A real number X is approximated in a machine by a rational x = M x · 2 e x − p +1 , – M x is the significand , a signed integer number of p digits in radix 2 s.t. 2 p − 1 ≤ | M x | ≤ 2 p − 1 ; – e x is the exponent , a signed integer ( e min ≤ e x ≤ e max ). 1 / 18

  3. Floating point arithmetics A real number X is approximated in a machine by a rational x = M x · 2 e x − p +1 , – M x is the significand , a signed integer number of p digits in radix 2 s.t. 2 p − 1 ≤ | M x | ≤ 2 p − 1 ; – e x is the exponent , a signed integer ( e min ≤ e x ≤ e max ). IEEE 754-2008: Most common formats Single ( binary 32 ) precision format: 1 8 23 s e m Double ( binary 64 ) precision format: 1 11 52 s e m 1 / 18

  4. Floating point arithmetics A real number X is approximated in a machine by a rational x = M x · 2 e x − p +1 , – M x is the significand , a signed integer number of p digits in radix 2 s.t. 2 p − 1 ≤ | M x | ≤ 2 p − 1 ; – e x is the exponent , a signed integer ( e min ≤ e x ≤ e max ). IEEE 754-2008: Most common formats Single ( binary 32 ) precision format: 1 8 23 s e m Double ( binary 64 ) precision format: 1 11 52 s e m Rounding modes 4 rounding modes: RD, RU, RZ, RN; Correct rounding for: + , − , × , ÷ , √ ; Portability, determinism. 1 / 18

  5. Applications → Computations with increased (multiple) precision in numerical applications. 0.4 Chaotic dynamical systems: 0.2 x2 0 bifurcation analysis; -0.2 compute periodic orbits (e.g., H´ enon map, Lorenz -0.4 -1.5 -1 -0.5 0 0.5 1 1.5 attractor); x1 celestial mechanics (e.g., stability of the solar system). Experimental mathematics: computational geometry (e.g., kissing numbers); polynomial optimization etc. 2 / 18

  6. What is a double-word number? Definition: A double-word number x is the unevaluated sum x h + x ℓ of two floating-point numbers x h and x ℓ such that x h = RN ( x ) . 3 / 18

  7. What is a double-word number? Definition: A double-word number x is the unevaluated sum x h + x ℓ of two floating-point numbers x h and x ℓ such that x h = RN ( x ) . → Called double-double when using the binary64 standard format. Example: π in double-double p h = 11 . 001001000011111101101010100010001000010110100011000 2 , and p ℓ = 1 . 0001101001100010011000110011000101000101110000000111 2 × 2 − 53 ; p h + p ℓ ↔ 107 bit FP approx. 3 / 18

  8. Remark − → Not the same as IEEE 754-2008 standard’s binary128/quadruple -precision. Double-word (using binary64/double -precision): Binary128/quadruple -precision: 4 / 18

  9. Remark − → Not the same as IEEE 754-2008 standard’s binary128/quadruple -precision. Double-word (using binary64/double -precision): Binary128/quadruple -precision: “wobbling precision” ≥ 107 bits of precision; 113 bits of precision; 4 / 18

  10. Remark − → Not the same as IEEE 754-2008 standard’s binary128/quadruple -precision. Double-word (using binary64/double -precision): Binary128/quadruple -precision: “wobbling precision” ≥ 107 bits of precision; 113 bits of precision; exponent range limited by larger exponent range ( 15 bits): binary64 ( 11 bits) i.e. − 1022 to − 16382 to 16383 ; 1023 ; 4 / 18

  11. Remark − → Not the same as IEEE 754-2008 standard’s binary128/quadruple -precision. Double-word (using binary64/double -precision): Binary128/quadruple -precision: “wobbling precision” ≥ 107 bits of precision; 113 bits of precision; exponent range limited by larger exponent range ( 15 bits): binary64 ( 11 bits) i.e. − 1022 to − 16382 to 16383 ; 1023 ; defined with all rounding modes lack of clearly defined rounding modes; 4 / 18

  12. Remark − → Not the same as IEEE 754-2008 standard’s binary128/quadruple -precision. Double-word (using binary64/double -precision): Binary128/quadruple -precision: “wobbling precision” ≥ 107 bits of precision; 113 bits of precision; exponent range limited by larger exponent range ( 15 bits): binary64 ( 11 bits) i.e. − 1022 to − 16382 to 16383 ; 1023 ; defined with all rounding modes lack of clearly defined rounding not implemented in hardware on modes; widely available processors. manipulated using error-free transforms (next slide). 4 / 18

  13. Error-Free Transforms Theorem 1 ( 2Sum algorithm) Let a and b be FP numbers. Algorithm 2Sum computes two FP numbers s and e that satisfy the following: s + e = a + b exactly; s = RN ( a + b ) . ( RN stands for performing the operation in rounding to nearest rounding mode.) Algorithm 1 ( 2Sum ( a, b ) ) s ← RN ( a + b ) t ← RN ( s − b ) e ← RN ( RN ( a − t ) + RN ( b − RN ( s − t ))) return ( s, e ) − → 6 FP operations (proved to be optimal unless we have information on the ordering of | a | and | b | ) 5 / 18

  14. Error-Free Transforms Theorem 2 ( Fast2Sum algorithm) Let a and b be FP numbers that satisfy e a ≥ e b ( | a | ≥ | b | ) . Algorithm Fast2Sum computes two FP numbers s and e that satisfy the following: s + e = a + b exactly; s = RN ( a + b ) . Algorithm 2 ( Fast2Sum ( a, b ) ) s ← RN ( a + b ) z ← RN ( s − a ) e ← RN ( b − z ) return ( s, e ) − → 3 FP operations 6 / 18

  15. Error-Free Transforms Theorem 3 ( 2ProdFMA algorithm) Let a and b be FP numbers, e a + e b ≥ e min + p − 1 . Algorithm 2ProdFMA computes two FP numbers p and e that satisfy the following: p + e = a · b exactly; p = RN ( a · b ) . Algorithm 3 ( 2ProdFMA ( a, b ) ) p ← RN ( a · b ) e ← fma ( a, b, − p ) return ( p, e ) − → 2 FP operations − → hardware-implemented FMA available in latest processors. 7 / 18

  16. Previous work: concept introduced by Dekker [DEK71] together with some algorithms for basic operations; Linnainmaa [LIN81] suggested similar algorithms assuming an underlying wider format; library written by Briggs [BRI98] - that is no longer maintained; QD library written by Bailey [Li.et.al02]. 8 / 18

  17. Previous work: concept introduced by Dekker [DEK71] together with some algorithms for basic operations; Linnainmaa [LIN81] suggested similar algorithms assuming an underlying wider format; library written by Briggs [BRI98] - that is no longer maintained; QD library written by Bailey [Li.et.al02]. Problems: 1 . most algorithms come without correctness proof and error bound; 2 . some error bounds published without a proof; 3 . differences between published and implemented algorithms. 8 / 18

  18. Previous work: concept introduced by Dekker [DEK71] together with some algorithms for basic operations; Linnainmaa [LIN81] suggested similar algorithms assuming an underlying wider format; library written by Briggs [BRI98] - that is no longer maintained; QD library written by Bailey [Li.et.al02]. Problems: 1 . most algorithms come without correctness proof and error bound; 2 . some error bounds published without a proof; 3 . differences between published and implemented algorithms. − → Strong need to “clean up” the existing literature. 8 / 18

  19. Previous work: concept introduced by Dekker [DEK71] together with some algorithms for basic operations; Linnainmaa [LIN81] suggested similar algorithms assuming an underlying wider format; library written by Briggs [BRI98] - that is no longer maintained; QD library written by Bailey [Li.et.al02]. Problems: 1 . most algorithms come without correctness proof and error bound; 2 . some error bounds published without a proof; 3 . differences between published and implemented algorithms. − → Strong need to “clean up” the existing literature. Notation: p represents the precision of the underlying FP format; ulp ( x ) = 2 ⌊ log 2 | x |⌋− p +1 , for x � = 0 ; u = 2 − p = 1 2 ulp (1) denotes the roundoff error unit. 8 / 18

  20. Addition: DWPlusFP ( x h , x ℓ , y ) Algorithm 4 1: ( s h , s ℓ ) ← 2Sum ( x h , y ) 2: v ← RN ( x ℓ + s ℓ ) 3: ( z h , z ℓ ) ← Fast2Sum ( s h , v ) 4: return ( z h , z ℓ ) – implemented in the QD library; – no previous error bound published; – relative error bounded by 2 · 2 − 2 p 1 − 2 · 2 − p = 2 · 2 − 2 p + 4 · 2 − 3 p + 8 · 2 − 4 p + · · · , which is less than 2 · 2 − 2 p + 5 · 2 − 3 p as soon as p ≥ 4 ; 9 / 18

  21. Addition: DWPlusFP ( x h , x ℓ , y ) Algorithm 4 1: ( s h , s ℓ ) ← 2Sum ( x h , y ) 2: v ← RN ( x ℓ + s ℓ ) 3: ( z h , z ℓ ) ← Fast2Sum ( s h , v ) 4: return ( z h , z ℓ ) – implemented in the QD library; – no previous error bound published; – relative error bounded by 2 · 2 − 2 p 1 − 2 · 2 − p = 2 · 2 − 2 p + 4 · 2 − 3 p + 8 · 2 − 4 p + · · · , which is less than 2 · 2 − 2 p + 5 · 2 − 3 p as soon as p ≥ 4 ; Asymptotically optimal bound Let x h = 1 , x ℓ = (2 p − 1) · 2 − 2 p , and y = − 1 2 (1 − 2 − p ) . Then: 2 + 3 · 2 − p − 1 and; – z h + z ℓ = 1 2 + 3 · 2 − p − 1 − 2 − 2 p ; – x + y = 1 – relative error 2 · 2 − 2 p 1 + 3 · 2 − p − 2 · 2 − 2 p ≈ 2 · 2 − 2 p − 6 · 2 − 3 p . 9 / 18

  22. Sketch of the proof Lemma 4 (Sterbenz Lemma) Let a and b be two positive FP numbers. If a 2 ≤ b ≤ 2 a, then a − b is a floating-point number, so that RN ( a − b ) = a − b . Lemma 5 Let a and b be FP numbers, and let s = RN ( a + b ) . If s � = 0 then � 1 2 ulp ( a ) , 1 � s ≥ max 2 ulp ( b ) . 10 / 18

Recommend

More recommend