1. 3/02/2010CS267 Lecture 131 CS 267 Dense Linear Algebra: History and Structure, Parallel Matrix Multiplication James Demmel www.cs.berkeley.edu/~demmel/cs267_Spr10

2. 3/02/2010CS267 Lecture 132 Outline History and motivation Structure of the Dense Linear Algebra motif Parallel Matrix-matrix multiplication Parallel Gaussian Elimination (next time)

3. 3/02/2010CS267 Lecture 133 Outline History and motivation Structure of the Dense Linear Algebra motif Parallel Matrix-matrix multiplication Parallel Gaussian Elimination (next time)

4. 4 Motifs The Motifs (formerly “Dwarfs”) from “The Berkeley View” (Asanovic et al.) Motifs form key computational patterns

5. A brief history of (Dense) Linear Algebra software (1/6) Libraries like EISPACK (for eigenvalue problems) Then the BLAS (1) were invented (1973-1977) Standard library of 15 operations (mostly) on vectors “AXPY” ( y = α·x + y ), dot product, scale (x = α·x ), etc Up to 4 versions of each (S/D/C/Z), 46 routines, 3300 LOC Goals Common “pattern” to ease programming, readability, self- documentation Robustness, via careful coding (avoiding over/underflow) Portability + Efficiency via machine specific implementations Why BLAS 1 ? They do O(n 1 ) ops on O(n 1 ) data Used in libraries like LINPACK (for linear systems) Source of the name “LINPACK Benchmark” (not the code!) 3/02/2010CS267 Lecture 135 In the beginning was the do-loop…

6. 3/02/2010CS267 Lecture 136 Current Records for Solving Dense Systems (11/2009) Gigaflops Machine n=100 n=1000 Any n Peak “Jaguar” (ORNL) 1.76M 2.3M (1.76 Petaflops, 224K cores, 7 MW, n=5.5M) www.top500.org, www.netlib.orgwww.top500.org, www.netlib.org/performance Palm Pilot III.00000169 (1.69 Kiloflops) NEC SX 8 (data from 2005) (8 proc, 2 GHz) 75.1 128 (1 proc, 2 GHz) 2.2 15.0 16 …

7. A brief history of (Dense) Linear Algebra software (2/6) But the BLAS-1 weren’t enough Consider AXPY ( y = α·x + y ): 2n flops on 3n read/writes Computational intensity = (2n)/(3n) = 2/3 Too low to run near peak speed (read/write dominates) Hard to vectorize (“SIMD’ize”) on supercomputers of the day (1980s) So the BLAS-2 were invented (1984-1986) Standard library of 25 operations (mostly) on matrix/vector pairs “GEMV”: y = α·A·x + β·x, “GER”: A = A + α·x·y T, x = T -1 ·x Up to 4 versions of each (S/D/C/Z), 66 routines, 18K LOC Why BLAS 2 ? They do O(n 2 ) ops on O(n 2 ) data So computational intensity still just ~(2n 2 )/(n 2 ) = 2 OK for vector machines, but not for machine with caches 3/02/2010CS267 Lecture 137

8. A brief history of (Dense) Linear Algebra software (3/6) The next step: BLAS-3 (1987-1988) Standard library of 9 operations (mostly) on matrix/matrix pairs “GEMM”: C = α·A·B + β·C, C = α·A·A T + β·C, B = T -1 ·B Up to 4 versions of each (S/D/C/Z), 30 routines, 10K LOC Why BLAS 3 ? They do O(n 3 ) ops on O(n 2 ) data So computational intensity (2n 3 )/(4n 2 ) = n/2 – big at last! Good for machines with caches, other mem. hierarchy levels How much BLAS1/2/3 code so far (all at www.netlib.org/blas) Source: 142 routines, 31K LOC, Testing: 28K LOC Reference (unoptimized) implementation only Ex: 3 nested loops for GEMM Lots more optimized code (eg Homework 1) Motivates “automatic tuning” of the BLAS (details later) 3/02/2010CS267 Lecture 138

9. 02/25/2009CS267 Lecture 89

10. A brief history of (Dense) Linear Algebra software (4/6) LAPACK – “Linear Algebra PACKage” - uses BLAS-3 (1989 – now) Ex: Obvious way to express Gaussian Elimination (GE) is adding multiples of one row to other rows – BLAS-1 How do we reorganize GE to use BLAS-3 ? (details later) Contents of LAPACK (summary) Algorithms we can turn into (nearly) 100% BLAS 3 –Linear Systems: solve Ax=b for x –Least Squares: choose x to minimize ||r|| 2   r i 2 where r=Ax-b Algorithms that are only 50% BLAS 3 (so far) –“Eigenproblems”: Find and x where Ax = x –Singular Value Decomposition (SVD): (A T A)x=  2 x Generalized problems (eg Ax = Bx) Error bounds for everything Lots of variants depending on A’s structure (banded, A=A T, etc) How much code? (Release 3.2, Nov 2008) (www.netlib.org/lapack) Source: 1582 routines, 490K LOC, Testing: 352K LOC Ongoing development (at UCB and elsewhere) (class projects!) 3/02/2010CS267 Lecture 1310

11. A brief history of (Dense) Linear Algebra software (5/6) Is LAPACK parallel? Only if the BLAS are parallel (possible in shared memory) ScaLAPACK – “Scalable LAPACK” (1995 – now) For distributed memory – uses MPI More complex data structures, algorithms than LAPACK Only (small) subset of LAPACK’s functionality available Details later (class projects!) All at www.netlib.org/scalapack 3/02/2010CS267 Lecture 1311

12. 3/02/2010CS267 Lecture 1312 Success Stories for Sca/LAPACK (6/6) Cosmic Microwave Background Analysis, BOOMERanG collaboration, MADCAP code (Apr. 27, 2000). ScaLAPACK Widely used Adopted by Mathworks, Cray, Fujitsu, HP, IBM, IMSL, Intel, NAG, NEC, SGI, … 5.5M webhits/year @ Netlib (incl. CLAPACK, LAPACK95) New Science discovered through the solution of dense matrix systems Nature article on the flat universe used ScaLAPACK Other articles in Physics Review B that also use it 1998 Gordon Bell Prize www.nersc.gov/news/reports/ newNERSCresults050703.pdf

13. 13 Back to basics: Why avoiding communication is important (1/2) Algorithms have two costs: 1.Arithmetic (FLOPS) 2.Communication: moving data between levels of a memory hierarchy (sequential case) processors over a network (parallel case). CPU Cache DRAM CPU DRAM CPU DRAM CPU DRAM CPU DRAM 3/02/2010CS267 Lecture 13

14. Why avoiding communication is important (2/2) Running time of an algorithm is sum of 3 terms: # flops * time_per_flop # words moved / bandwidth # messages * latency 14 communication Time_per_flop << 1/ bandwidth << latency Gaps growing exponentially with time Goal : organize linear algebra to avoid communication Between all memory hierarchy levels L1 L2 DRAM network, etc Not just hiding communication (overlap with arith) (speedup  2x ) Arbitrary speedups possible Annual improvements Time_per_flopBandwidthLatency Network26%15% DRAM23%5% 59% 3/02/2010

15. for i = 1 to n {read row i of A into fast memory, n 2 reads} for j = 1 to n {read C(i,j) into fast memory, n 2 reads} {read column j of B into fast memory, n 3 reads} for k = 1 to n C(i,j) = C(i,j) + A(i,k) * B(k,j) {write C(i,j) back to slow memory, n 2 writes} 15 Review: Naïve Sequential MatMul: C = C + A*B =+* C(i,j) A(i,:) B(:,j) C(i,j) n 3 + O(n 2 ) reads/writes altogether 3/02/2010CS267 Lecture 13

16. Less Communication with Blocked Matrix Multiply Blocked Matmul C = A·B explicitly refers to subblocks of A, B and C of dimensions that depend on cache size 16 … Break A nxn, B nxn, C nxn into bxb blocks labeled A(i,j), etc … b chosen so 3 bxb blocks fit in cache for i = 1 to n/b, for j=1 to n/b, for k=1 to n/b C(i,j) = C(i,j) + A(i,k)·B(k,j) … b x b matmul, 4b 2 reads/writes (n/b) 3 · 4b 2 = 4n 3 /b reads/writes altogether Minimized when 3b 2 = cache size = M, yielding O(n 3 /M 1/2 ) reads/writes What if we had more levels of memory? (L1, L2, cache etc)? Would need 3 more nested loops per level 3/02/2010CS267 Lecture 13

17. Blocked vs Cache-Oblivious Algorithms Blocked Matmul C = A·B explicitly refers to subblocks of A, B and C of dimensions that depend on cache size 17 … Break A nxn, B nxn, C nxn into bxb blocks labeled A(i,j), etc … b chosen so 3 bxb blocks fit in cache for i = 1 to n/b, for j=1 to n/b, for k=1 to n/b C(i,j) = C(i,j) + A(i,k)·B(k,j) … b x b matmul … another level of memory would need 3 more loops Cache-oblivious Matmul C = A·B is independent of cache Function C = RMM(A,B) … R for recursive If A and B are 1x1 C = A · B else … Break A nxn, B nxn, C nxn into (n/2)x(n/2) blocks labeled A(i,j), etc for i = 1 to 2, for j = 1 to 2, for k = 1 to 2 C(i,j) = C(i,j) + RMM( A(i,k), B(k,j) ) … n/2 x n/2 matmul 3/02/2010 CS267 Lecture 13

18. Communication Lower Bounds: Prior Work on Matmul Assume n 3 algorithm (i.e. not Strassen-like) Sequential case, with fast memory of size M Lower bound on #words moved to/from slow memory =  (n 3 / M 1/2 ) [Hong, Kung, 81] Attained using blocked or cache-oblivious algorithms 18 Parallel case on P processors: Let NNZ be total memory needed; assume load balanced Lower bound on #words communicated =  (n 3 /(P· NNZ ) 1/2 ) [Irony, Tiskin, Toledo, 04] NNZ (name of alg)Lower bound on #words Attained by 3n 2 (“2D alg”)  (n 2 / P 1/2 ) [Cannon, 69] 3n 2 P 1/3 (“3D alg”)  (n 2 / P 2/3 ) [Johnson,93] 3/02/2010

19. New lower bound for all “direct” linear algebra Holds for BLAS, LU, QR, eig, SVD, … Some whole programs (sequences of these operations, no matter how they are interleaved, eg computing A k ) Dense and sparse matrices (where #flops << n 3 ) Sequential and parallel algorithms Some graph-theoretic algorithms (eg Floyd-Warshall) See (partial) proof sketch from Lecture 2 19 Let M = “fast” memory size per processor = cache size (sequential case) or n 2 /p (parallel case) #words_moved by at least one processor =  (#flops / M 1/2 ) #messages_sent by at least one processor =  (#flops / M 3/2 ) 3/02/2010CS267 Lecture 13

20. Can we attain these lower bounds? Do conventional dense algorithms as implemented in LAPACK and ScaLAPACK attain these bounds? Mostly not If not, are there other algorithms that do? Yes Goals for algorithms: Minimize #words =  (#flops/ M 1/2 ) Minimize #messages =  (#flops/ M 3/2 ) Need new data structures Minimize for multiple memory hierarchy levels Cache-oblivious algorithms would be simplest Fewest flops when matrix fits in fastest memory Cache-oblivious algorithms don’t always attain this Attainable for nearly all dense linear algebra Just a few prototype implementations so far (class projects!) Only a few sparse algorithms so far (eg Cholesky) 203/02/2010CS267 Lecture 13

21. A brief future look at (Dense) Linear Algebra software PLASMA and MAGMA (now) Planned extensions to Multicore/GPU/Heterogeneous Can one software infrastructure accommodate all algorithms and platforms of current (future) interest? How much code generation and tuning can we automate? Details later (Class projects!) Other related projects BLAST Forum (www.netlib.org/blas/blast-forum) Attempt to extend BLAS to other languages, add some new functions, sparse matrices, extra-precision, interval arithmetic Only partly successful (extra-precise BLAS used in latest LAPACK) FLAME (www.cs.utexas.edu/users/flame/)www.cs.utexas.edu/users/flame/ Formal Linear Algebra Method Environment Attempt to automate code generation across multiple platforms 3/02/2010CS267 Lecture 1321

22. 3/02/2010CS267 Lecture 1322 Outline History and motivation Structure of the Dense Linear Algebra motif Parallel Matrix-matrix multiplication Parallel Gaussian Elimination (next time)

23. What could go into the linear algebra motif(s)? For all linear algebra problems For all matrix/problem structures For all data types For all programming interfaces Produce best algorithm(s) w.r.t. performance and accuracy (including error bounds, etc) For all architectures and networks Need to prioritize, automate! CS267 Lecture 13233/02/2010

24. For all linear algebra problems: Ex: LAPACK Table of Contents Linear Systems Least Squares Overdetermined, underdetermined Unconstrained, constrained, weighted Eigenvalues and vectors of Symmetric Matrices Standard (Ax = λx), Generalized (Ax=λxB) Eigenvalues and vectors of Unsymmetric matrices Eigenvalues, Schur form, eigenvectors, invariant subspaces Standard, Generalized Singular Values and vectors (SVD) Standard, Generalized Level of detail Simple Driver Expert Drivers with error bounds, extra-precision, other options Lower level routines (“apply certain kind of orthogonal transformation”) CS267 Lecture 13243/02/2010

25. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13253/02/2010

26. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13263/02/2010

27. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13273/02/2010

28. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13283/02/2010

29. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13293/02/2010

30. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13303/02/2010

31. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13313/02/2010

32. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed CS267 Lecture 13323/02/2010

33. Organizing Linear Algebra – in books www.netlib.org/lapackwww.netlib.org/scalapack www.cs.utk.edu/~dongarra/etemplateswww.netlib.org/templates gams.nist.gov

34. 3/02/2010CS267 Lecture 1334 Outline History and motivation Structure of the Dense Linear Algebra motif Parallel Matrix-matrix multiplication Parallel Gaussian Elimination (next time)

35. 3/02/2010CS267 Lecture 1335 Different Parallel Data Layouts for Matrices (not all!) 0123012301230123 01230123 1) 1D Column Blocked Layout2) 1D Column Cyclic Layout 3) 1D Column Block Cyclic Layout 4) Row versions of the previous layouts Generalizes others 01010101 23232323 01010101 23232323 01010101 23232323 01010101 23232323 6) 2D Row and Column Block Cyclic Layout 0123 01 23 5) 2D Row and Column Blocked Layout b

36. 3/02/2010CS267 Lecture 1336 Parallel Matrix-Vector Product Compute y = y + A*x, where A is a dense matrix Layout: 1D row blocked A(i) refers to the n by n/p block row that processor i owns, x(i) and y(i) similarly refer to segments of x,y owned by i Algorithm: Foreach processor i Broadcast x(i) Compute y(i) = A(i)*x Algorithm uses the formula y(i) = y(i) + A(i)*x = y(i) +  j A(i,j)*x(j) x y P0 P1 P2 P3 P0 P1 P2 P3 A(0) A(1) A(2) A(3)

37. 3/02/2010CS267 Lecture 1337 Matrix-Vector Product y = y + A*x A column layout of the matrix eliminates the broadcast of x But adds a reduction to update the destination y A 2D blocked layout uses a broadcast and reduction, both on a subset of processors sqrt(p) for square processor grid P0 P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 P15

38. 3/02/2010CS267 Lecture 1338 Parallel Matrix Multiply Computing C=C+A*B Using basic algorithm: 2*n 3 Flops Variables are: Data layout Topology of machine Scheduling communication Use of performance models for algorithm design Message Time = “latency” + #words * time-per-word =  + n*  Efficiency (in any model): serial time / (p * parallel time) perfect (linear) speedup  efficiency = 1

39. 3/02/2010CS267 Lecture 1339 Matrix Multiply with 1D Column Layout Assume matrices are n x n and n is divisible by p A(i) refers to the n by n/p block column that processor i owns (similiarly for B(i) and C(i)) B(i,j) is the n/p by n/p sublock of B(i) in rows j*n/p through (j+1)*n/p - 1 Algorithm uses the formula C(i) = C(i) + A*B(i) = C(i) +  j A(j)*B(j,i) p0p1p2p3p5p4p6p7 May be a reasonable assumption for analysis, not for code

40. 3/02/2010CS267 Lecture 1340 Matrix Multiply: 1D Layout on Bus or Ring Algorithm uses the formula C(i) = C(i) + A*B(i) = C(i) +  j A(j)*B(j,i) First consider a bus-connected machine without broadcast: only one pair of processors can communicate at a time (ethernet) Second consider a machine with processors on a ring: all processors may communicate with nearest neighbors simultaneously

41. 3/02/2010CS267 Lecture 1341 MatMul: 1D layout on Bus without Broadcast Naïve algorithm: C(myproc) = C(myproc) + A(myproc)*B(myproc,myproc) for i = 0 to p-1 for j = 0 to p-1 except i if (myproc == i) send A(i) to processor j if (myproc == j) receive A(i) from processor i C(myproc) = C(myproc) + A(i)*B(i,myproc) barrier Cost of inner loop: computation: 2*n*(n/p) 2 = 2*n 3 /p 2 communication:  +  *n 2 /p

42. 3/02/2010CS267 Lecture 1342 Naïve MatMul (continued) Cost of inner loop: computation: 2*n*(n/p) 2 = 2*n 3 /p 2 communication:  +  *n 2 /p … approximately Only 1 pair of processors (i and j) are active on any iteration, and of those, only i is doing computation => the algorithm is almost entirely serial Running time: = (p*(p-1) + 1)*computation + p*(p-1)*communication  2*n 3 + p 2 *  + p*n 2 *  This is worse than the serial time and grows with p. When might you still want to do this?

43. 3/02/2010CS267 Lecture 1343 Matmul for 1D layout on a Processor Ring Pairs of adjacent processors can communicate simultaneously Copy A(myproc) into Tmp C(myproc) = C(myproc) + Tmp*B(myproc, myproc) for j = 1 to p-1 Send Tmp to processor myproc+1 mod p Receive Tmp from processor myproc-1 mod p C(myproc) = C(myproc) + Tmp*B( myproc-j mod p, myproc) Same idea as for gravity in simple sharks and fish algorithm May want double buffering in practice for overlap Ignoring deadlock details in code Time of inner loop = 2*(  +  *n 2 /p) + 2*n*(n/p) 2

44. 3/02/2010 CS267 Lecture 13 44 Matmul for 1D layout on a Processor Ring Time of inner loop = 2*(  +  *n 2 /p) + 2*n*(n/p) 2 Total Time = 2*n* (n/p) 2 + (p-1) * Time of inner loop  2*n 3 /p + 2*p*  + 2*  *n 2 (Nearly) Optimal for 1D layout on Ring or Bus, even with Broadcast: Perfect speedup for arithmetic A(myproc) must move to each other processor, costs at least (p-1)*cost of sending n*(n/p) words Parallel Efficiency = 2*n 3 / (p * Total Time) = 1/(1 +  * p 2 /(2*n 3 ) +  * p/(2*n) ) = 1/ (1 + O(p/n)) Grows to 1 as n/p increases (or  and  shrink)

45. 3/02/2010CS267 Lecture 1345 MatMul with 2D Layout Consider processors in 2D grid (physical or logical) Processors can communicate with 4 nearest neighbors Broadcast along rows and columns Assume p processors form square s x s grid, s = p 1/2 p(0,0) p(0,1) p(0,2) p(1,0) p(1,1) p(1,2) p(2,0) p(2,1) p(2,2) p(0,0) p(0,1) p(0,2) p(1,0) p(1,1) p(1,2) p(2,0) p(2,1) p(2,2) p(0,0) p(0,1) p(0,2) p(1,0) p(1,1) p(1,2) p(2,0) p(2,1) p(2,2) =*

46. 3/02/2010CS267 Lecture 1346 Cannon’s Algorithm … C(i,j) = C(i,j) +  A(i,k)*B(k,j) … assume s = sqrt(p) is an integer forall i=0 to s-1 … “skew” A left-circular-shift row i of A by i … so that A(i,j) overwritten by A(i,(j+i)mod s) forall i=0 to s-1 … “skew” B up-circular-shift column i of B by i … so that B(i,j) overwritten by B((i+j)mod s), j) for k=0 to s-1 … sequential forall i=0 to s-1 and j=0 to s-1 … all processors in parallel C(i,j) = C(i,j) + A(i,j)*B(i,j) left-circular-shift each row of A by 1 up-circular-shift each column of B by 1 k

47. 3/02/2010CS267 Lecture 1347 C(1,2) = A(1,0) * B(0,2) + A(1,1) * B(1,2) + A(1,2) * B(2,2) Cannon’s Matrix Multiplication

48. 3/02/2010CS267 Lecture 1348 Initial Step to Skew Matrices in Cannon Initial blocked input After skewing before initial block multiplies A(1,0) A(2,0) A(0,1)A(0,2) A(1,1) A(2,1) A(1,2) A(2,2) A(0,0) B(0,1)B(0,2) B(1,0) B(2,0) B(1,1)B(1,2) B(2,1)B(2,2) B(0,0) A(1,0) A(2,0) A(0,1)A(0,2) A(1,1) A(2,1) A(1,2) A(2,2) A(0,0) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0)

49. 3/02/2010CS267 Lecture 1349 Skewing Steps in Cannon All blocks of A must multiply all like-colored blocks of B First step Second Third A(1,0) A(2,0) A(0,1)A(0,2) A(1,1) A(2,1) A(1,2) A(2,2) A(0,0) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(1,0) A(2,0) A(0,1) A(0,2) A(2,1) A(1,2) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(1,0) A(2,0) A(0,1) A(0,2) A(1,1) A(2,1) A(1,2) A(2,2) A(0,0) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(1,1) A(2,2) A(0,0)

50. 3/02/2010CS267 Lecture 1350 Cost of Cannon’s Algorithm forall i=0 to s-1 … recall s = sqrt(p) left-circular-shift row i of A by i … cost ≤ s*(  +  *n 2 /p) forall i=0 to s-1 up-circular-shift column i of B by i … cost ≤ s*(  +  *n 2 /p) for k=0 to s-1 forall i=0 to s-1 and j=0 to s-1 C(i,j) = C(i,j) + A(i,j)*B(i,j) … cost = 2*(n/s) 3 = 2*n 3 /p 3/2 left-circular-shift each row of A by 1 … cost =  +  *n 2 /p up-circular-shift each column of B by 1 … cost =  +  *n 2 /p ° Total Time = 2*n 3 /p + 4 * s*  + 4*  *n 2 /s ° Parallel Efficiency = 2*n 3 / (p * Total Time) = 1/( 1 +  * 2*(s/n) 3 +  * 2*(s/n) ) = 1/(1 + O(sqrt(p)/n)) ° Grows to 1 as n/s = n/sqrt(p) = sqrt(data per processor) grows ° Better than 1D layout, which had Efficiency = 1/(1 + O(p/n))

51. Cannon’s Algorithm is “optimal” Optimal means Considering only O(n 3 ) matmul algs (not Strassen) Considering only O(n 2 /p) storage per processor Use communication lower bound #words =  (#flops/M 1/2 ) Sequential: a processor doing n 3 flops from matmul with a fast memory of size M must do  (n 3 / M 1/2 ) references to slow memory Parallel: a processor doing f =#flops from matmul with a local memory of size M must do  (f / M 1/2 ) references to remote memory Local memory = O(n 2 /p), f  n 3 /p for at least one proc. So f / M 1/2 =  ((n 3 /p)/ (n 2 /p) 1/2 ) =  ( n 2 /p 1/2 ) #Messages =  ( #words sent / max message size) =  ( (n 2 /p 1/2 )/(n 2 /p)) =  (p 1/2 ) 3/02/2010CS267 Lecture 1351

52. 3/02/2010CS267 Lecture 1352 Pros and Cons of Cannon So what if it’s “optimal”, is it fast? Local computation one call to (optimized) matrix-multiply Hard to generalize for p not a perfect square A and B not square Dimensions of A, B not perfectly divisible by s=sqrt(p) A and B not “aligned” in the way they are stored on processors block-cyclic layouts Memory hog (extra copies of local matrices)

53. 3/02/2010CS267 Lecture 1353 SUMMA Algorithm SUMMA = Scalable Universal Matrix Multiply Slightly less efficient, but simpler and easier to generalize Presentation from van de Geijn and Watts www.netlib.org/lapack/lawns/lawn96.ps Similar ideas appeared many times Used in practice in PBLAS = Parallel BLAS www.netlib.org/lapack/lawns/lawn100.ps

54. 3/02/2010CS267 Lecture 1354 SUMMA * = i j A(i,k) k k B(k,j) i, j represent all rows, columns owned by a processor k is a block of b  1 rows or columns C(i,j) = C(i,j) +  k A(i,k)*B(k,j) Assume a p r by p c processor grid (p r = p c = 4 above) Need not be square C(i,j)

55. 3/02/2010CS267 Lecture 1355 SUMMA For k=0 to n/b-1 … where b is the block size … b = # cols in A(i,k) and # rows in B(k,j) for all i = 1 to p r … in parallel owner of A(i,k) broadcasts it to whole processor row for all j = 1 to p c … in parallel owner of B(k,j) broadcasts it to whole processor column Receive A(i,k) into Acol Receive B(k,j) into Brow C_myproc = C_myproc + Acol * Brow * = i j A(i,k) k k B(k,j) C(i,j)

56. 3/02/2010CS267 Lecture 1356 SUMMA performance For k=0 to n/b-1 for all i = 1 to s … s = sqrt(p) owner of A(i,k) broadcasts it to whole processor row … time = log s *(  +  * b*n/s), using a tree for all j = 1 to s owner of B(k,j) broadcasts it to whole processor column … time = log s *(  +  * b*n/s), using a tree Receive A(i,k) into Acol Receive B(k,j) into Brow C_myproc = C_myproc + Acol * Brow … time = 2*(n/s) 2 *b °Total time = 2*n 3 /p +  * log p * n/b +  * log p * n 2 /s °To simplify analysis only, assume s = sqrt(p)

57. 3/02/2010CS267 Lecture 1357 SUMMA performance Total time = 2*n 3 /p +  * log p * n/b +  * log p * n 2 /s Parallel Efficiency = 1/(1 +  * log p * p / (2*b*n 2 ) +  * log p * s/(2*n) )  Same  term as Cannon, except for log p factor log p grows slowly so this is ok Latency (  ) term can be larger, depending on b When b=1, get  * log p * n As b grows to n/s, term shrinks to  * log p * s (log p times Cannon) Temporary storage grows like 2*b*n/s Can change b to tradeoff latency cost with memory

58. 3/02/2010CS267 Lecture 858 PDGEMM = PBLAS routine for matrix multiply Observations: For fixed N, as P increases Mflops increases, but less than 100% efficiency For fixed P, as N increases, Mflops (efficiency) rises DGEMM = BLAS routine for matrix multiply Maximum speed for PDGEMM = # Procs * speed of DGEMM Observations (same as above): Efficiency always at least 48% For fixed N, as P increases, efficiency drops For fixed P, as N increases, efficiency increases

59. 3/02/2010 CS267 Lecture 13 59 Summary of Parallel Matrix Multiplication 1D Layout Bus without broadcast - slower than serial Nearest neighbor communication on a ring (or bus with broadcast): Efficiency = 1/(1 + O(p/n)) 2D Layout Cannon Efficiency = 1/(1+O(  sqrt(p) /n    +  * sqrt(p) /n)) – optimal! Hard to generalize for general p, n, block cyclic, alignment SUMMA Efficiency = 1/(1 + O(  log p * p / (b*n 2 ) +  log p * sqrt(p) /n)) Very General b small => less memory, lower efficiency b large => more memory, high efficiency Used in practice (PBLAS)

60. 3/02/2010CS267 Lecture 1360 ScaLAPACK Parallel Library

61. 3/02/2010CS267 Lecture 1361 Extra Slides

62. 2/27/08CS267 Guest Lecture 162 Recursive Layouts For both cache hierarchies and parallelism, recursive layouts may be useful Z-Morton, U-Morton, and X-Morton Layout Also Hilbert layout and others What about the user’s view? Fortunately, many problems can be solved on a permutation Never need to actually change the user’s layout

63. 02/09/2006CS267 Lecture 863 Gaussian Elimination 0 x x x x...... Standard Way subtract a multiple of a row 0 x 0 0... 0 LINPACK apply sequence to a column x nb then apply nb to rest of matrix a 3 =a 3 -a 1 *a 2 a3a3 a2a2 a1a1 L a 2 = L -1 a 2 0 x 0 0... 0 nb LAPACK apply sequence to nb Slide source: Dongarra

64. 02/09/2006CS267 Lecture 864 LU Algorithm: 1: Split matrix into two rectangles (m x n/2) if only 1 column, scale by reciprocal of pivot & return 2: Apply LU Algorithm to the left part 3: Apply transformations to right part (triangular solve A 12 = L -1 A 12 and matrix multiplication A 22 = A 22 -A 21 *A 12 ) 4: Apply LU Algorithm to right part Gaussian Elimination via a Recursive Algorithm L A 12 A 21 A 22 F. Gustavson and S. Toledo Most of the work in the matrix multiply Matrices of size n/2, n/4, n/8, … Slide source: Dongarra

65. 02/09/2006CS267 Lecture 865 Recursive Factorizations Just as accurate as conventional method Same number of operations Automatic variable blocking Level 1 and 3 BLAS only ! Extreme clarity and simplicity of expression Highly efficient The recursive formulation is just a rearrangement of the point-wise LINPACK algorithm The standard error analysis applies (assuming the matrix operations are computed the “conventional” way). Slide source: Dongarra

66. 02/09/2006CS267 Lecture 866 LAPACK Recursive LU LAPACK Dual-processor Uniprocessor Slide source: Dongarra

67. 02/09/2006CS267 Lecture 867 Review: BLAS 3 (Blocked) GEPP for ib = 1 to n-1 step b … Process matrix b columns at a time end = ib + b-1 … Point to end of block of b columns apply BLAS2 version of GEPP to get A(ib:n, ib:end) = P’ * L’ * U’ … let LL denote the strict lower triangular part of A(ib:end, ib:end) + I A(ib:end, end+1:n) = LL -1 * A(ib:end, end+1:n) … update next b rows of U A(end+1:n, end+1:n ) = A(end+1:n, end+1:n ) - A(end+1:n, ib:end) * A(ib:end, end+1:n) … apply delayed updates with single matrix-multiply … with inner dimension b BLAS 3

68. 02/09/2006CS267 Lecture 868 Review: Row and Column Block Cyclic Layout processors and matrix blocks are distributed in a 2d array pcol-fold parallelism in any column, and calls to the BLAS2 and BLAS3 on matrices of size brow-by-bcol serial bottleneck is eased need not be symmetric in rows and columns

69. 02/09/2006CS267 Lecture 869 Distributed GE with a 2D Block Cyclic Layout block size b in the algorithm and the block sizes brow and bcol in the layout satisfy b=brow=bcol. shaded regions indicate busy processors or communication performed. unnecessary to have a barrier between each step of the algorithm, e.g.. step 9, 10, and 11 can be pipelined

70. 02/09/2006CS267 Lecture 870 Distributed GE with a 2D Block Cyclic Layout

71. 02/09/2006CS267 Lecture 871 Matrix multiply of green = green - blue * pink

72. 02/09/2006CS267 Lecture 872 PDGESV = ScaLAPACK parallel LU routine Since it can run no faster than its inner loop (PDGEMM), we measure: Efficiency = Speed(PDGESV)/Speed(PDGEMM) Observations: Efficiency well above 50% for large enough problems For fixed N, as P increases, efficiency decreases (just as for PDGEMM) For fixed P, as N increases efficiency increases (just as for PDGEMM) From bottom table, cost of solving Ax=b about half of matrix multiply for large enough matrices. From the flop counts we would expect it to be (2*n 3 )/(2/3*n 3 ) = 3 times faster, but communication makes it a little slower.

73. 02/09/2006CS267 Lecture 873

74. 02/09/2006CS267 Lecture 874 Scales well, nearly full machine speed

75. 02/09/2006CS267 Lecture 875 Old version, pre 1998 Gordon Bell Prize Still have ideas to accelerate Project Available! Old Algorithm, plan to abandon

76. 02/09/2006CS267 Lecture 876 Have good ideas to speedup Project available! Hardest of all to parallelize Have alternative, and would like to compare Project available!

77. 02/09/2006CS267 Lecture 877 Out-of-core means matrix lives on disk; too big for main mem Much harder to hide latency of disk QR much easier than LU because no pivoting needed for QR Moral: use QR to solve Ax=b Projects available (perhaps very hard…)

78. 02/09/2006CS267 Lecture 878 A small software project...

79. 02/09/2006CS267 Lecture 879 Work-Depth Model of Parallelism The work depth model: The simplest model is used For algorithm design, independent of a machine The work, W, is the total number of operations The depth, D, is the longest chain of dependencies The parallelism, P, is defined as W/D Specific examples include: circuit model, each input defines a graph with ops at nodes vector model, each step is an operation on a vector of elements language model, where set of operations defined by language

80. 02/09/2006CS267 Lecture 880 Latency Bandwidth Model Network of fixed number P of processors fully connected each with local memory Latency (  ) accounts for varying performance with number of messages gap (g) in logP model may be more accurate cost if messages are pipelined Inverse bandwidth (  ) accounts for performance varying with volume of data Efficiency (in any model): serial time / (p * parallel time) perfect (linear) speedup  efficiency = 1

81. 02/09/2006CS267 Lecture 881 Initial Step to Skew Matrices in Cannon Initial blocked input After skewing before initial block multiplies A(0,1)A(0,2) A(1,0) A(2,0) A(1,1)A(1,2) A(2,1)A(2,2) A(0,0) B(0,1)B(0,2) B(1,0) B(2,0) B(1,1)B(1,2) B(2,1)B(2,2) B(0,0)A(0,1)A(0,2) A(1,0) A(2,0) A(1,1)A(1,2) A(2,1)A(2,2) A(0,0) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0)

82. 02/09/2006CS267 Lecture 882 Skewing Steps in Cannon First step Second Third A(0,1)A(0,2) A(1,0) A(2,0) A(1,1)A(1,2) A(2,1)A(2,2) A(0,0) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(0,1)A(0,2) A(1,0) A(2,0) A(1,2) A(2,1) B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(0,1)A(0,2) A(1,0) A(2,0) A(1,1)A(1,2) A(2,1)A(2,2) A(0,0)B(0,1) B(0,2)B(1,0) B(2,0) B(1,1) B(1,2) B(2,1) B(2,2)B(0,0) A(1,1) A(2,2) A(0,0)

83. 2/25/2009CS267 Lecture 883 Motivation (1) 3 Basic Linear Algebra Problems 1.Linear Equations: Solve Ax=b for x 2.Least Squares: Find x that minimizes ||r|| 2   r i 2 where r=Ax-b Statistics: Fitting data with simple functions 3a. Eigenvalues: Find and x where Ax = x Vibration analysis, e.g., earthquakes, circuits 3b. Singular Value Decomposition: A T Ax=  2 x Data fitting, Information retrieval Lots of variations depending on structure of A A symmetric, positive definite, banded, …

84. 2/25/2009CS267 Lecture 884 Motivation (2) Why dense A, as opposed to sparse A? Many large matrices are sparse, but … Dense algorithms easier to understand Some applications yields large dense matrices LINPACK Benchmark (www.top500.org) “How fast is your computer?” = “How fast can you solve dense Ax=b?” Large sparse matrix algorithms often yield smaller (but still large) dense problems Do ParLab Apps most use small dense matrices?

85. 02/25/2009CS267 Lecture 8 Algorithms for 2D (3D) Poisson Equation (N = n 2 (n 3 ) vars) AlgorithmSerialPRAMMemory #Procs Dense LUN 3 NN 2 N 2 Band LUN 2 (N 7/3 )NN 3/2 (N 5/3 )N (N 4/3 ) JacobiN 2 (N 5/3 ) N (N 2/3 ) NN Explicit Inv.N 2 log NN 2 N 2 Conj.Gradients N 3/2 (N 4/3 ) N 1/2(1/3) *log NNN Red/Black SOR N 3/2 (N 4/3 ) N 1/2 (N 1/3 ) NN Sparse LUN 3/2 (N 2 ) N 1/2 N*log N(N 4/3 ) N FFTN*log Nlog NNN MultigridNlog 2 NNN Lower boundNlog NN PRAM is an idealized parallel model with zero cost communication Reference: James Demmel, Applied Numerical Linear Algebra, SIAM, 1997 (Note: corrected complexities for 3D case from last lecture!).

86. Lessons and Questions (1) Structure of the problem matters Cost of solution can vary dramatically (n 3 to n) Many other examples Some structure can be figured out automatically “A\b” can figure out symmetry, some sparsity Some structures known only to (smart) user If performance not critical, user may be happy to settle for A\b How much of this goes into the motifs? How much should we try to help user choose? Tuning, but at algorithmic choice level (SALSA) Motifs overlap Dense, sparse, (un)structured grids, spectral

87. Organizing Linear Algebra (1) By Operations Low level (eg mat-mul: BLAS) Standard level (eg solve Ax=b, Ax=λx: Sca/LAPACK) Applications level (eg systems & control: SLICOT) By Performance/accuracy tradeoffs “Direct methods” with guarantees vs “iterative methods” that may work faster and accurately enough By Structure Storage Dense –columnwise, rowwise, 2D block cyclic, recursive space-filling curves Banded, sparse (many flavors), black-box, … Mathematical Symmetries, positive definiteness, conditioning, … As diverse as the world being modeled

88. Organizing Linear Algebra (2) By Data Type Real vs Complex Floating point (fixed or varying length), other By Target Platform Serial, manycore, GPU, distributed memory, out-of- DRAM, Grid, … By programming interface Language bindings “A\b” versus access to details

89. For all linear algebra problems: Ex: LAPACK Table of Contents Linear Systems Least Squares Overdetermined, underdetermined Unconstrained, constrained, weighted Eigenvalues and vectors of Symmetric Matrices Standard (Ax = λx), Generallzed (Ax=λxB) Eigenvalues and vectors of Unsymmetric matrices Eigenvalues, Schur form, eigenvectors, invariant subspaces Standard, Generalized Singular Values and vectors (SVD) Standard, Generalized Level of detail Simple Driver Expert Drivers with error bounds, extra-precision, other options Lower level routines (“apply certain kind of orthogonal transformation”)

90. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

91. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

92. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

93. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

94. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

95. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

96. For all matrix/problem structures: Ex: LAPACK Table of Contents BD – bidiagonal GB – general banded GE – general GG – general, pair GT – tridiagonal HB – Hermitian banded HE – Hermitian HG – upper Hessenberg, pair HP – Hermitian, packed HS – upper Hessenberg OR – (real) orthogonal OP – (real) orthogonal, packed PB – positive definite, banded PO – positive definite PP – positive definite, packed PT – positive definite, tridiagonal SB – symmetric, banded SP – symmetric, packed ST – symmetric, tridiagonal SY – symmetric TB – triangular, banded TG – triangular, pair TB – triangular, banded TP – triangular, packed TR – triangular TZ – trapezoidal UN – unitary UP – unitary packed

97. For all data types: Ex: LAPACK Table of Contents Real and complex Single and double precision Arbitrary precision in progress

98. Organizing Linear Algebra (3) www.netlib.org/lapackwww.netlib.org/scalapack www.cs.utk.edu/~dongarra/etemplateswww.netlib.org/templates gams.nist.gov

99. 2/27/08CS267 Guest Lecture 199 Review of the BLAS BLAS levelEx.# mem refs# flopsq 1“Axpy”, Dot prod 3n2n 1 2/3 2Matrix- vector mult n2n2 2n 2 2 3Matrix- matrix mult 4n 2 2n 3 n/2 Building blocks for all linear algebra Parallel versions call serial versions on each processor So they must be fast! Define q = # flops / # mem refs = “computational intensity” The larger is q, the faster the algorithm can go in the presence of memory hierarchy “axpy”: y =  *x + y, where  scalar, x and y vectors