1 / 12

Straight Line Code

Straight Line Code. x i = a x i = x j + x k x i = x j  x k. No loop No index as variables No if statement. Arithmetic circuit DAG. w ij : weight of edges from v i to v j. Example. x 1. x 2. 4. 3. x 1 = 4 x 2 = 3 x 3 = x 1 + x 2 x 4 = x 2 + x 3

bjone
Télécharger la présentation

Straight Line Code

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Straight Line Code xi = a xi = xj + xk xi = xj xk No loop No index as variables No if statement Arithmetic circuit DAG wij : weight of edges from vi to vj

  2. Example x1 x2 4 3 x1 = 4 x2 = 3 x3 = x1 + x2 x4 = x2 + x3 x5 = x2 * x4 x6 = x3 * x4 x7 = x4 * x6 x3 x4 + + x6 x5 * * Straight line code x7 * Arithmetic Circuit

  3. degree v1 : x1 1 v2 : x2 1 v3 : x1 + x2 1 v4 : x1 + 2x2 1 v5 : x1x2 + 2x22 2 v6 : x12 + 3x2 + 2x22 v7 : x1 + … . . . Multi Variate Polynomial Parallel Algorithm O(logN log dN) d: degree of circuit (degree of polynomial) N: Size or number of nodes in the circuit d may be large (exponential of N) useful only if d is polynomial of N

  4. Degree of poly O(N) Gaussian Elimination straight line program with O( N3) nodes Application O ( log2 N) time Construction of heap Sorting - Straight line program with O(N2) nodes O ( log2 N) time [bubble sort, insertion sort]

  5. vj vj vi vj vi vj 1 1 vk * => vk Case (a) Both vi and vj are are leaves leaf node: its value is evaluated completely Eval - Add: all inputs are leves Eval - Mult: all inputs are leaves one input is leaf Skip - Add: not all inputs are leaves Eval - Add vi vj vi vj wjk wik => vk + vk Action of Eval -Add vk = viwik + vjwjk Eval - Mult vi vi wjk= value(vi) 1 1 * vk => + vk Case (b) vi is a leaf, and vj are not leaf (exactly one input is a leaf node)

  6. Skip Add w2 w2 w3 w3 w1 w4 w1 w4 + + v w3*w5 w5 w2*w5 z z w1*w5 w4*w5 + + x Inputs of the first addition node x are promoted to become inputs to the second addition node y. Inputs to x are are maintained if x is an input to another node z.

  7. Skip-Add Multiple edges from vi to vj may be inserted. New weight from vi to vj if vi, vj are addition node i original wij must be removed(vi, vj both + nodes) vj : addition node vi : not k vj is not an addition nde j if vk, vj are addition node otherwise Skip Add: W <- W - W+ + WW+ O(n3) PEs, O(logn) time

  8. v1 6 2 v2 x1 = 6 x2 = 2 x3 = x1 + x2 x4 = x2 + x3 x5 = x3• x4 x6 = x3 + x5 x7 = x4• x5 v3 + + v4 v5 • • v7 + v6 Straight-line Code Arithmetic Circuit node polynomial degree x1 x2 x1 + x2 x1 + 2x2 x12+3x1x2+2x22 x12+3x1x2+2x22+x1+x2 x13+5x12x2+8x1x22+4x23 1 1 1 1 2 2 3 v1 v2 v3 v4 v5 v6 v7

  9. v1 v2 v1 v2 v1 v2 6 2 6 2 6 2 1 1 8 1 v3 v3 v3 8 8 8 1 1 1 v4 v4 v4 + + + 1 8 1 8 1 1 v5 1 v5 v5 1 1 1 + + • 1 1 1 1 1 v7 v7 v7 • 8 • • v6 v6 + + + v6 after Eval-Add after Eval-Mult after Skip-Add v1 v2 v1 6 2 v2 v1 6 2 v2 6 2 v3 8 v3 8 v3 8 10 v4 10 v4 10 v4 1 v5 80 v5 v5 8 1 80 80 8 1 v7 v7 v7 • 8 v6 + v6 800 + v6 800 + after Eval-Add after Eval-Mult after Skip-Add

  10. Complexity a b Define height leaf node = 2 * node : h(x) = h(a) + h(b) + node: max (a+1, m) a = max (h1,..., hi) v1,..., vi : + nodes inputs to x m = max (hi+1 ,..., hk) vi+1,....,vk : * nodes inputs to x degree and height: almost same except + node degree: max(d1,d2), height: max (h1,h2)+1 Theorem: O(logdN) phases are sufficient. Proof is based on the following two claims. Claim 1: The of an N-node circuit of degree d  (N+1)d Claim 2: After log3/2h phases, every node has been reduced to a leaf node. * x vi+1 v1 vi vk + * * + + x

  11. Proof on the complexity of straight line program Lemma: The height of an arithmetic circuit of degree d is at most (Z+2)d, where Z is the number of addition nodes in the circuit that have addition nodes as inputs. Proof: By induction on the number of nodes N in the circuit. Basis (N=1) trivially true. Inductive hypothesis: Assume that the claim is true for N. Inductive step: Let us prove for N+1. Let v be a node with degree d and max height h with inputs v1,...,vk, having degree d1,...dk, and heights h1,...,hk. Three cases. case 1: v is “*” node. Then d=d1 + ... + dk, and h=h1 + ... + hk, hi (Z+2)di by inductive hypothesis. Thus, h  (Z+2)d. case 2: v is “+” node, and max of h1 ,..., hkis not attained by “+” node. hi = max{hi} max{(Z+2)di} (Z+2)d. case 3: v is “+” node, and max of h1 ,..., hkis attained by “+” node. hi = 1 + max{hi} 1 + max{(Z+1)di} 1 + (Z+1)d  (Z+2)d. ( hi (Z+1)di follows from the fact that v itself is a type of Z, and must not be part of di subcircuit, and inductive hypothesis.) Thus, the lemma holds. Corollary: The height of N node circuit of degree d is at most (N+1)d. Proof: Since Z  N-1, it follows.

  12. v2 v1 * Proof on the complexity (cont’) Lemma: Given any circuit and node v, let h :the height of v at the beginning of a phase, h’ :the height of v at the end of the phase. If h  3, then h’  (2/3)h. Moreover, if v is *” node that gets converted into an addition node during the phase, then h’  (2/3)h - 1/3. Proof: Induction on the size of circuit. Suppose that v is *node at the beginning of phase. Case 1. v becomes a leaf during the phase. h’=2, * node h  4. Thus, h’  h/2. Case 2. v remains *. Nor v1, v2 are leaf before Eval-Mult. Thus, h1,h2 > 3 h1’  (2/3)h1, h2’  (2/3)h2, by inductive hypo. Thus, h’  h1’ + h2’  (2/3) (h1 +h2 )  (2/3)h v Case 3. v becomes “+” node during the phase. One node was leaf, the other is not. Let v1 be the input to v before the phase. After the phase, either v1 stays as the only input, or v1 participate skip-add during the phase. If v1 remains input of v: h = h1+2, h1’ (2/3)h1, thus, h’ = h1’+1  (2/3)h- 1. If v1participate skip-add during the phase, then any new input u to v after the phase should be the input to v1 before the phase. Note that the height of u is less than the height of v1, so the same argument as the above holds.

More Related