Jusitification for dahlia dse parameter space

[sgpthomas]

Thread to keep track of justifications for why we chose the parameter space we did for each benchmark

[sgpthomas]

Stencil-2d-inner

    "ORIG_D1": [1,2,3,4,5,6],
    "ORIG_D2": [1,2,3,4,5,6],
    "FILTER_D1": [1,2,3],
    "FILTER_D2": [1,2,3],
    "UR_LOOP1": [1,2,3],
    "UR_LOOP2": [1,2,3]

Filter banking and the loop unrolling can't be bigger than 3 so we are trying all banking and unroll factors up to 3. For the banking on orig, we chose 1..6 even though everything above 3 will be overpartioning because we are testing the claim that you shouldn't think about the banking of an array independently from the loop that it is used within.

18/2916 points accepted by Dahlia

[rachitnigam]

NW

    "SEQA_BANK": [1,2,4],
    "SEQB_BANK": [1,2,4],
    "PTR_BANK1": [1,2,4],
    "PTR_BANK2": [1,2,4],
    "M_BANK1":   [1,2,4],
    "M_BANK2":   [1,2,4],
    "UNROLLA": [1,2,4,8],
    "UNROLLB": [1,2,4,8]

Did not include banking factor 3 because design space becomes ~260,000 points. Unrolling factors calculated by multiplying banking factors and memory ports (2).

448/11664 points accepted by Dahlia

[sgpthomas]

MD-KNN

    "FORCE":    [1,2,3,4],
    "POS":      [1,2,3,4],
    "J_D1":     [1,2,3,4],
    "J_D2":     [1,2,3,4],
    "UR_LOOP1": [1,2,3,4,5,6,7,8],
    "UR_LOOP2": [1,2,3,4,5,6,7,8]

525/16384 points accepted by Dahlia

[sgpthomas]

MD-GRID

    "D1": [1,2,3,4],
    "D2": [1,2,3,4],
    "D3": [1,2,3,4],
    "B0X_UR": [1,2,4,5,6,7,8],
    "B0Y_UR": [1,2,4,5,6,7,8],
    "B0Z_UR": [1,2,4,5,6,7,8]

81/21952 points accepted by Dahlia