伯克利并行计算课件7.pptVIP

Sparse-Matrix-Vector-Multiply (SpMV) y = A*x Using Segmented Scan (SegScan) Segscan computes prefix sums of arbitrary segments Use CSR format of Sparse Matrix A, store x densely Create array P of all nonzero A(i,j)*x(j) = Val(k)*x(Col_Ind(k)) Create array S showing where segments (rows) start Compute SegScan( P, S ) = Extract A*x = [14 61 24 ] /afs//project/scandal/public/papers/CMU-CS-93-173.ps.Z 02/06/2014 CS267 Lecture 6+ Segscan ( [3, 1, 4, 5, 6, 1, 2, 3 ], [T, F, F, T, T, F, F, T ]) = [3, 4, 8, 5, 6, 7, 9, 3] 0 2 3 0 2 4 0 0 5 3 0 0 0 1 A = Val = [ 1 2 3 2 4 5 3 1 ] Col_Ind = [ 1 3 4 1 2 5 1 5 ] Row_Ptr=[ 1 4 7 9 ] 7 8 2 1 3 x= P = [ 7 4 3 14 32 15 21 3 ] S = [ T F F T F F T F ] [ 7 11 14 14 46 61 21 24 ] Page layout in a browser Applying layout rules to html description of a webpage is a bottleneck, scan can help Simplest example Given widths [x1, x2, … , xn] of items to display on page, where should each item go? Item j starts at x1 + x2 + … + xj-1 Real examples have complicated constraints Defined by general trees, since in html each object to display can be composed of other objects To get location of each object, need to do preorder traversal of tree, “adding up” constraints of previous objects Scan can do preorder traversal of any tree in parallel Not just binary trees Ras Bodik, Leo Meyerovich 02/06/2014 CS267 Lecture 6+ 02/06/2014 CS267 Lecture 6+ Summary of tree algorithms Lots of problems can be done quickly - in theory - using trees Some algorithms are widely used broadcasts, reductions, parallel prefix carry look ahead addition Some are of theoretical interest only Csanky’s method for matrix inversion Solving tridiagonal linear systems (without pivoting) Both numerically unstable Csanky needs too many processors Embedded in various systems MPI, Split-C, Titanium, NESL, other languages CM-5 hardware control network