Mercurial > repos > public > sbplib_julia
annotate stencil.jl @ 129:1aaeb46ba5f4 cell_based_test
Improve efficiency of apply by the following:
- Remove divisions in interior loop by storing and multiplying by the reciprocal of grid spacing instead.
- Add @inline to apply(::Laplace
- Remove initialization of w = 0 in apply(::Stencil) by manually unrolling first iteration of the loop.
author | Vidar Stiernström <vidar.stiernstrom@it.uu.se> |
---|---|
date | Thu, 14 Feb 2019 16:25:22 +0100 |
parents | 7c0b9bb7ab4d |
children | 8569c637d923 |
rev | line source |
---|---|
84
48079bd39969
Change to using tuples in stencils and ops
Jonatan Werpers <jonatan@werpers.com>
parents:
69
diff
changeset
|
1 struct Stencil{T<:Real,N} |
48079bd39969
Change to using tuples in stencils and ops
Jonatan Werpers <jonatan@werpers.com>
parents:
69
diff
changeset
|
2 range::Tuple{Int,Int} |
48079bd39969
Change to using tuples in stencils and ops
Jonatan Werpers <jonatan@werpers.com>
parents:
69
diff
changeset
|
3 weights::NTuple{N,T} |
8
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
4 end |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
5 |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
6 function flip(s::Stencil) |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
7 range = (-s.range[2], -s.range[1]) |
84
48079bd39969
Change to using tuples in stencils and ops
Jonatan Werpers <jonatan@werpers.com>
parents:
69
diff
changeset
|
8 return Stencil(range, reverse(s.weights)) |
8
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
9 end |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
10 |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
11 # Provides index into the Stencil based on offset for the root element |
128
7c0b9bb7ab4d
Improve stencil application code to make it more friendly to compiler optimizations
Jonatan Werpers <jonatan@werpers.com>
parents:
122
diff
changeset
|
12 @inline function Base.getindex(s::Stencil, i::Int) |
71
18d0d794d3bb
Make stencils respond to @ inbounds
Jonatan Werpers <jonatan@werpers.com>
parents:
67
diff
changeset
|
13 @boundscheck if i < s.range[1] || s.range[2] < i |
67
7fd4e7a1cd38
Make stencil more type stabeler
Jonatan Werpers <jonatan@werpers.com>
parents:
43
diff
changeset
|
14 return eltype(s.weights)(0) |
8
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
15 end |
71
18d0d794d3bb
Make stencils respond to @ inbounds
Jonatan Werpers <jonatan@werpers.com>
parents:
67
diff
changeset
|
16 return s.weights[1 + i - s.range[1]] |
8
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
17 end |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
18 |
128
7c0b9bb7ab4d
Improve stencil application code to make it more friendly to compiler optimizations
Jonatan Werpers <jonatan@werpers.com>
parents:
122
diff
changeset
|
19 Base.@propagate_inbounds @inline function apply(s::Stencil{T,N}, v::AbstractVector, i::Int) where {T,N} |
129
1aaeb46ba5f4
Improve efficiency of apply by the following:
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
128
diff
changeset
|
20 w = s.weights[1]*v[i+ s.range[1]] |
1aaeb46ba5f4
Improve efficiency of apply by the following:
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
128
diff
changeset
|
21 @simd for k ∈ 2:N |
128
7c0b9bb7ab4d
Improve stencil application code to make it more friendly to compiler optimizations
Jonatan Werpers <jonatan@werpers.com>
parents:
122
diff
changeset
|
22 w += s.weights[k]*v[i+ s.range[1] + k-1] |
8
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
23 end |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
24 return w |
433008d3b7d3
Move stencil to its own file
Jonatan Werpers <jonatan@werpers.com>
parents:
diff
changeset
|
25 end |
122
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
26 |
129
1aaeb46ba5f4
Improve efficiency of apply by the following:
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
128
diff
changeset
|
27 # TODO: Fix loop unrolling here as well. Then we can also remove Base.getindex(::Stencil) |
128
7c0b9bb7ab4d
Improve stencil application code to make it more friendly to compiler optimizations
Jonatan Werpers <jonatan@werpers.com>
parents:
122
diff
changeset
|
28 Base.@propagate_inbounds @inline function apply_backwards(s::Stencil, v::AbstractVector, i::Int) |
122
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
29 w = zero(eltype(v)) |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
30 for j ∈ s.range[2]:-1:s.range[1] |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
31 @inbounds weight = s[j] |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
32 w += weight*v[i-j] |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
33 end |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
34 return w |
6c6979ff17f4
Introduce and use apply_backwards for stencils
Jonatan Werpers <jonatan@werpers.com>
parents:
85
diff
changeset
|
35 end |