DESC/bench_sparse.jl at main · finch-tensor/DESC · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
using Finch
using BenchmarkTools
using SparseArrays
# using Random

# Random.seed!(1234)

const P = 8
const n = 10000
const dims = (n, n)


function setup_data(dims, P)
    tensors = Vector{Tensor}(undef, P)
    data = Vector{Tuple}(undef, P)
    for i in 1:P
        t = fsprand(dims[1], dims[2], 0.001)
        tensors[i] = Tensor(SparseDict(SparseDict(Element(0.0))), t)
        data[i] = ffindnz(t)
    end

    current_dim3 = 2
    I = data[1][1:length(dims)]
    I = (I..., fill(1, length(I[1])))
    V = data[1][length(data[1])]

    for i in 2:P
        cur_tens = (data[i][1:length(data[i]) - 1]..., fill(current_dim3, length(data[i][1])))
        I = map(vcat, I, cur_tens)
        V = vcat(V, data[i][length(data[i])])
        current_dim3 += 1
    end

    data_out = fsparse(I..., V)
    return tensors, data_out
end

println("Initializing Data... (Sparse, Sparse) test")
tensors, merged_tensors = setup_data(dims, P)

@info "Warmup and Init"
A = Tensor(Dense(SparseDict(SparseDict(Element(0.0)))), merged_tensors)
B = Tensor(SparseDict(SparseDict(Element(0.0))), dims[1], dims[2])

eval(@finch_kernel function serial_merge_benchmark(A, B)
    B .= 0
    for p in _, j in _, i in _
        B[i, j] += A[i, j, p]
    end
end)

serial_merge_benchmark(A, B)

tot = nnz(SparseMatrixCSC(B))
println("total nnz: $tot")


A = Tensor(Dense(SparseList(SparseList(Element(0.0)))), merged_tensors)
B = Tensor(SparseList(SparseList(Element(0.0))), B)

ptr_1 = Vector{Vector{Int64}}(undef, P)
idx_1 = Vector{Vector{Int64}}(undef, P)
ptr_2 = Vector{Vector{Int64}}(undef, P)
idx_2 = Vector{Vector{Int64}}(undef, P)
val = Vector{Vector{Float64}}(undef, P)


for i in 1:P
    current = tensors[i]

    ptr_1[i] = current.lvl.ptr
    idx_1[i] = current.lvl.idx

    ptr_2[i] = current.lvl.lvl.ptr
    idx_2[i] = current.lvl.lvl.idx

    val[i] = current.lvl.lvl.lvl.val
end

include("./mergeelement.jl")
include("./mergedense.jl")

max_dim = 1
max_idx = n

println("Benchmarking Parallel Algorithm with Balancer:")

include("./mergesplistbalance.jl")

function merge_parallel_balance(gfm, ptr, idx, ptr2, idx2, P, max_pos, max_idx, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, val, lvl_val)
    gfm1, max_pos1 = merge_splist_balance(gfm, ptr, idx, P, max_pos, max_idx, false, lvl_ptr, lvl_idx)
    gfm2, max_pos2 = merge_splist_balance(gfm1, ptr2, idx2, P, max_pos1, max_idx, false, lvl2_ptr, lvl2_idx)
    merge_element(gfm2, val, max_pos2, P, lvl_val)
end

respb = @benchmark merge_parallel_balance(gfm, $ptr_1, $idx_1, $ptr_2, $idx_2, $P, $max_dim, $n, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, $val, lvl_val) setup=(
    gfm = [[1] for _ in 1:$P];
    lvl_ptr = Vector{Int}();
    lvl_idx = Vector{Int}();
    lvl2_ptr = Vector{Int}();
    lvl2_idx = Vector{Int}();
    lvl_val = Vector{Float64}()
) evals=1

display(respb)

lvl_ptr = Vector{Int}()
lvl_idx = Vector{Int}()
lvl2_ptr = Vector{Int}()
lvl2_idx = Vector{Int}()
lvl_val = Vector{Float64}()
global_fbr_map0 = [[1] for _ in 1:P]

merge_parallel_balance(global_fbr_map0, ptr_1, idx_1, ptr_2, idx_2, P, max_dim, n, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, val, lvl_val)


@assert lvl_ptr == B.lvl.ptr
@assert lvl_idx == B.lvl.idx
@assert lvl2_ptr == B.lvl.lvl.ptr
@assert lvl2_idx == B.lvl.lvl.idx
@assert lvl_val == B.lvl.lvl.lvl.val

println("Benchmarking baseline:")
include("./baseline.jl")

max_dim = n

function merge_baseline(global_fbr_map, local_fbr_map, task_map, ptr, index, ptr2, idx2, val, P, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, lvl_val, max_dim, max_idx)
    cutoffs1 = compute_proc_cutoffs(index, P)
    merged_positions, merged_indices, local_fbr_map2, task_map2 = gen_pos_idx_map(global_fbr_map, local_fbr_map, task_map, ptr, index, cutoffs1, P)
    global_fbr_map, local_fbr_map, task_map = process_next_lvl(merged_positions, merged_indices, task_map2, local_fbr_map2, P, 1, lvl_ptr, lvl_idx)

    cutoffs2 = compute_proc_cutoffs(idx2, P)
    merged_positions, merged_indices, local_fbr_map2, task_map2 = gen_pos_idx_map(global_fbr_map, local_fbr_map, task_map, ptr2, idx2, cutoffs2, P)
    global_fbr_map, local_fbr_map, task_map = process_next_lvl(merged_positions, merged_indices, task_map2, local_fbr_map2, P, max_dim, lvl2_ptr, lvl2_idx)
    merge_element_level(global_fbr_map, local_fbr_map, task_map, val, P, lvl_val)
end

resb = @benchmark merge_baseline(global_fbr_map, local_fbr_map, task_map, $ptr_1, $idx_1, $ptr_2, $idx_2, $val, $P, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, lvl_val, $max_dim, $n) setup=(
    global_fbr_map=fill(1, $P);
    local_fbr_map=fill(1, $P);
    task_map=repeat(1:($P), 1);
    lvl_ptr=Vector{Int}();
    lvl_idx=Vector{Int}();
    lvl2_ptr=Vector{Int}();
    lvl2_idx=Vector{Int}();
    lvl_val=Vector{Float64}()
) evals=1

display(resb)

lvl_ptr = Vector{Int}()
lvl_idx = Vector{Int}()
lvl2_ptr = Vector{Int}()
lvl2_idx = Vector{Int}()
lvl_val = Vector{Float64}()

global_fbr_map0 = fill(1, P)
local_fbr_map0 = fill(1, P)
task_map0 = repeat(1:P, 1)

merge_baseline(global_fbr_map0, local_fbr_map0, task_map0, ptr_1, idx_1, ptr_2, idx_2, val, P, lvl_ptr, lvl_idx, lvl2_ptr, lvl2_idx, lvl_val, max_dim, n)

@assert lvl_ptr == B.lvl.ptr
@assert lvl_idx == B.lvl.idx
@assert lvl2_ptr == B.lvl.lvl.ptr
@assert lvl2_idx == B.lvl.lvl.idx
@assert lvl_val == B.lvl.lvl.lvl.val

@info "all equivalency checks passed!"