mtsokol
commented
4 months ago Overall I'm happy with the change, I would like only to investigate a bit further PlusOneVector.
To repeat one of my comments: I run SDDMM with csr sparse format. And for explicit copy I get 5s execution:
```julia
Executing:
:(function var"##compute#316"(prgm)
begin
V = ((((((((((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, SparseListLevel{Int64, Vector{Int64}, Vector{Int64}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}}
V_2 = ((((((((((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).children[1]).children[3]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
V_3 = (((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[3]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A0 = V::Tensor{DenseLevel{Int64, SparseListLevel{Int64, Vector{Int64}, Vector{Int64}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}}
A2 = V_2::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A5 = V_3::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A8 = Tensor(Dense(SparseDict(Element{0.0, Float64}())))::Tensor{DenseLevel{Int64, SparseLevel{Int64, Finch.DictTable{Int64, Int64, Vector{Int64}, Vector{Int64}, Vector{Int64}, Dict{Tuple{Int64, Int64}, Int64}}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}}
@finch mode = :fast begin
A8 .= 0.0
for i49 = _
for i47 = _
for i48 = _
A8[i48, i47] << + >>= (*)((*)(A0[i48, i47], A2[i47, i49]), A5[i48, i49])
end
end
end
return A8
end
return (swizzle(A8, 2, 1),)
end
end)
```
And for the no-copy constructor (that differs only in PlusOneVector, used diffchecker.com) it's 100s:
```julia
Executing:
:(function var"##compute#308"(prgm)
begin
V = ((((((((((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, SparseListLevel{Int64, PlusOneVector{Int32}, PlusOneVector{Int32}, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
V_2 = ((((((((((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[2]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).children[1]).children[3]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
V_3 = (((((((((((((((((((((prgm.children[1]).children[2]).children[2]).children[3]).children[1]).children[1]).children[1]).children[2]).children[1]).children[3]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[1]).children[2]).tns.val::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A0 = V::Tensor{DenseLevel{Int64, SparseListLevel{Int64, PlusOneVector{Int32}, PlusOneVector{Int32}, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A2 = V_2::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A5 = V_3::Tensor{DenseLevel{Int64, DenseLevel{Int64, ElementLevel{0.0, Float64, Int64, PyArray{Float64, 1, true, true, Float64}}}}}
A8 = Tensor(Dense(SparseDict(Element{0.0, Float64}())))::Tensor{DenseLevel{Int64, SparseLevel{Int64, Finch.DictTable{Int64, Int64, Vector{Int64}, Vector{Int64}, Vector{Int64}, Dict{Tuple{Int64, Int64}, Int64}}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}}
@finch mode = :fast begin
A8 .= 0.0
for i49 = _
for i47 = _
for i48 = _
A8[i48, i47] << + >>= (*)((*)(A0[i48, i47], A2[i47, i49]), A5[i48, i49])
end
end
end
return A8
end
return (swizzle(A8, 2, 1),)
end
end)
```
hameerabbasi
willow-ahrens
Hi @willow-ahrens,
Here's a PR with adjustments required by https://github.com/willow-ahrens/Finch.jl/pull/562.
One thing that I'm concerned about is reduced performance of code that uses
PlusOneVector(zero-copy SciPy constructors), which is way slower than copy counterparts.For example, in SDDMM, changing copy to no-copy CSR matrix introduces x10 speed reduction.