FastAlmostBandedMatrices.jl

A fast implementation of Almost Banded Matrices in Julia. Primarily developed for use in BoundaryValueDiffEq.jl.

SemiseparableMatrices.jl provides an alternate implementation of almost banded matrices, but that had considerable overhead for our use-case. This implementation borrows a lot of details from that repository.

High Level Examples

using FastAlmostBandedMatrices

m = 2
n = 10

A1 = AlmostBandedMatrix(brand(Float64, n, n, m + 1, m), rand(Float64, m, n))

Benchmarks

QR Factorization & Linear Solve

Click me!

using BenchmarkTools, FastAlmostBandedMatrices, SparseArrays, FillArrays, LinearAlgebra
import SemiseparableMatrices

m = 5
n = 1000

II = zeros(n, m)
II[diagind(II)] .= 1

A1 = AlmostBandedMatrix(brand(Float64, n, n, m + 1, m), rand(Float64, m, n))
A2 = Matrix(A1)
A3 = SemiseparableMatrices.AlmostBandedMatrix(copy(A1.bands),
    SemiseparableMatrices.LowRankMatrix(II, copy(A1.fill)))
A4 = sparse(A2)

@benchmark qr($A1)
# BenchmarkTools.Trial: 6959 samples with 1 evaluation.
#  Range (min … max):  680.420 μs …   2.600 ms  ┊ GC (min … max): 0.00% … 70.11%
#  Time  (median):     707.540 μs               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   716.433 μs ± 100.758 μs  ┊ GC (mean ± σ):  1.05% ±  4.81%

#            ▃▃▅▅██▃▂                                              
#   ▂▂▂▃▃▃▂▃▇█████████▆▅▄▄▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▂▁▁▂▂▂▂▂▂▁▂▂▂▂▂▁▂ ▃
#   680 μs           Histogram: frequency by time          802 μs <

#  Memory estimate: 320.70 KiB, allocs estimate: 11.

@benchmark qr($A2)
# BenchmarkTools.Trial: 173 samples with 1 evaluation.
#  Range (min … max):  23.543 ms … 39.269 ms  ┊ GC (min … max): 0.00% … 0.00%
#  Time  (median):     29.162 ms              ┊ GC (median):    0.00%
#  Time  (mean ± σ):   28.962 ms ±  2.984 ms  ┊ GC (mean ± σ):  0.75% ± 2.29%

#            ▂▁▃  █       ▂▄▄▂▂ ▁                                
#   ▅▅▁▃▃▃▅▁▇███▇▆██▅▄▃▇▄▃███████▅▄▅▄▃▃▁▅▃▁▃▃▄▅▁▃▃▄▃▁▁▃▁▁▁▁▁▁▁▃ ▃
#   23.5 ms         Histogram: frequency by time        38.4 ms <

#  Memory estimate: 8.18 MiB, allocs estimate: 6.

@benchmark qr($A3)
# BenchmarkTools.Trial: 1452 samples with 1 evaluation.
#  Range (min … max):  2.975 ms …   7.381 ms  ┊ GC (min … max): 0.00% … 21.67%
#  Time  (median):     3.206 ms               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   3.440 ms ± 625.169 μs  ┊ GC (mean ± σ):  6.43% ± 11.76%

#     ▂▆██▇▅▂                                ▁▁  ▂▁▁▁            
#   ██████████▆▅▅▄▄▄▄▁▁▄▁▄▄▁▁▁▁▄▁▄▄▄▁▄▁▁▁▁▁▆▇████████▇▅▇▅▅▅▁▄▁▅ █
#   2.97 ms      Histogram: log(frequency) by time      5.52 ms <

#  Memory estimate: 6.32 MiB, allocs estimate: 65108.

@benchmark qr($A4)
# BenchmarkTools.Trial: 362 samples with 1 evaluation.
#  Range (min … max):   8.413 ms … 104.416 ms  ┊ GC (min … max): 0.00% … 1.04%
#  Time  (median):     10.826 ms               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   13.788 ms ±   9.973 ms  ┊ GC (mean ± σ):  1.20% ± 3.47%

#    ▄█                                                           
#   ▄███▆▄▆▅▅▅▄▃▂▃▃▂▂▂▃▃▃▄▃▁▂▁▂▁▂▁▃█▇█▅▆▆▅▃▄▃▃▂▃▂▃▁▃▂▃▃▂▃▂▂▂▃▃▁▃ ▃
#   8.41 ms         Histogram: frequency by time         22.3 ms <

#  Memory estimate: 25.44 MiB, allocs estimate: 169.

b = randn(n)

@benchmark $A1 \ $b
# BenchmarkTools.Trial: 5945 samples with 1 evaluation.
#  Range (min … max):  797.407 μs …   2.972 ms  ┊ GC (min … max): 0.00% … 68.87%
#  Time  (median):     828.066 μs               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   838.834 μs ± 124.996 μs  ┊ GC (mean ± σ):  1.27% ±  5.33%

#                 ▁▂▂▇█▇▃▂▁                                        
#   ▂▂▃▄▅▆▇█▇▆█▇███████████▇█▇▆▆▅▄▄▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▁▂▂▁▁▂▁▂▂▁▁▂▂ ▄
#   797 μs           Histogram: frequency by time          901 μs <

#  Memory estimate: 367.95 KiB, allocs estimate: 431.

@benchmark $A2 \ $b
# BenchmarkTools.Trial: 586 samples with 1 evaluation.
#  Range (min … max):  7.682 ms …  15.557 ms  ┊ GC (min … max): 0.00% … 40.03%
#  Time  (median):     8.305 ms               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   8.515 ms ± 679.543 μs  ┊ GC (mean ± σ):  1.86% ±  3.97%

#        ▂█▃▂▅▆▄     ▁                                           
#   ▄▆▆▅▆█████████▅▇▆█▄▄▃▄▆▆▅▆▄▃▅▃▄▃▄▂▃▁▃▃▃▃▁▂▂▃▁▃▃▁▂▁▁▁▃▁▁▁▁▃▂ ▃
#   7.68 ms         Histogram: frequency by time          11 ms <

#  Memory estimate: 7.64 MiB, allocs estimate: 4.

@benchmark $A3 \ $b
# BenchmarkTools.Trial: 1212 samples with 1 evaluation.
#  Range (min … max):  3.458 ms …   7.806 ms  ┊ GC (min … max): 0.00% … 30.80%
#  Time  (median):     3.815 ms               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   4.118 ms ± 707.897 μs  ┊ GC (mean ± σ):  7.61% ± 12.33%

#       ▂█▇▆▆                                                    
#   ▂▃▄▅██████▆▄▃▃▃▂▂▂▂▂▂▂▂▁▂▁▂▁▂▂▁▁▁▂▂▂▂▃▃▄▄▄▃▃▃▃▃▃▃▂▂▂▃▂▁▂▂▂▂ ▃
#   3.46 ms         Histogram: frequency by time        6.17 ms <

#  Memory estimate: 7.55 MiB, allocs estimate: 78253.

@benchmark $A4 \ $b
# BenchmarkTools.Trial: 2879 samples with 1 evaluation.
#  Range (min … max):  1.627 ms …   3.120 ms  ┊ GC (min … max): 0.00% … 36.35%
#  Time  (median):     1.675 ms               ┊ GC (median):    0.00%
#  Time  (mean ± σ):   1.734 ms ± 243.465 μs  ┊ GC (mean ± σ):  2.76% ±  7.52%

#   ▂▇█▆▃                                                   ▁▁  ▁
#   ██████▇█▆▄▅▃▄▁▁▁▁▃▁▁▁▁▁▁▃▄▁▁▃▁▃▁▁▁▃▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄▇███ █
#   1.63 ms      Histogram: log(frequency) by time      2.81 ms <

#  Memory estimate: 2.21 MiB, allocs estimate: 82.

Public API

For documentation of the public API, please use the REPL help mode.

AlmostBandedMatrix
bandpart
fillpart
exclusive_bandpart
finish_part_setindex!

Some Considerations

Not all operations are meant to be fast. Currently only the following operations are fast or planned to be fast:
- Matrix Vector Multiply
- Matrix Matrix Multiply
- QR Factorization
- UpperTriangular ldiv!
- SubArrays
Broadcasting won't be fast and will materialize the array into a standard Julia Array.
ArrayInferface.jl API
- fast_scalar_indexing
- qr_instance
No support for fast_scalar_indexing == false arrays.
LinearSolve.jl has proper dispatches setup.

avik-pal / FastAlmostBandedMatrices.jl