Skip to content

Inbounds decompression #355

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
blegat wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Inbounds decompression #355

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
984adab
Elide bounds checks in CSC substitution decompression
blegat Oct 19, 2026
55395bc
Check bounds
blegat Oct 28, 2026
6798b46
More targetted @inbounds
blegat Oct 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
84 changes: 64 additions & 20 deletions src/decompression.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -516,6 +516,26 @@ sparse matrix with column pointers `A_colptr`, without constructing the matrix.
The destination pattern must match the full coloring pattern (`uplo=:F`), its
lower triangle (`uplo=:L`), or its upper triangle (`uplo=:U`), with row indices
sorted in increasing order within each column. Return `nothing`.

For substitution decompression, the buffers must use one-based indexing, `B`
must have one row per vertex and one column per color, and the lengths of
`A_colptr` and `nzA` must match the selected destination pattern. These sizes
are checked before any entries are written. The following structural conditions
are assumed without checking:

- `A_colptr[1] == 1` and `A_colptr[end] == length(nzA) + 1`.
- For each column `j`, `A_colptr[j + 1] - A_colptr[j]` equals the number of
stored entries in column `j` of the selected coloring pattern, including
explicit zeros. Thus the pointers are nondecreasing.
- Entries in `nzA[A_colptr[j]:(A_colptr[j + 1] - 1)]` correspond to those
pattern entries in increasing row order.
- The index and offset arrays in `result` have not been modified since coloring.

Substitution decompression checks each destination index immediately before
writing. Only validated reads and scratch-buffer accesses use `@inbounds`.
Malformed column pointers that produce out-of-bounds destination indices throw
`BoundsError`. Pattern mismatches whose indices remain in bounds can still
produce incorrect values.
"""
function decompress_csc!(
nzA::AbstractVector,
Expand Down Expand Up @@ -640,14 +660,14 @@ end
# uplo = :L or uplo = :F
# A[i,j] is stored at index_ij = (A.colptr[j+1] - offset_L) in A.nzval
if uplo != :U
nzind = A_colptr[j + 1] - lower_triangle_offsets[counter]
nzind = @inbounds A_colptr[j + 1] - lower_triangle_offsets[counter]
nzA[nzind] = val
end

# uplo = :U or uplo = :F
# A[j,i] is stored at index_ji = (A.colptr[i] + offset_U) in A.nzval
if uplo != :L
nzind = A_colptr[i] + upper_triangle_offsets[counter]
nzind = @inbounds A_colptr[i] + upper_triangle_offsets[counter]
nzA[nzind] = val
end

Expand All @@ -656,28 +676,45 @@ end
# uplo = :U or uplo = :F
# A[i,j] is stored at index_ij = (A.colptr[j] + offset_U) in A.nzval
if uplo != :L
nzind = A_colptr[j] + upper_triangle_offsets[counter]
nzind = @inbounds A_colptr[j] + upper_triangle_offsets[counter]
nzA[nzind] = val
end

# uplo = :L or uplo = :F
# A[j,i] is stored at index_ji = (A.colptr[i+1] - offset_L) in A.nzval
if uplo != :U
nzind = A_colptr[i + 1] - lower_triangle_offsets[counter]
nzind = @inbounds A_colptr[i + 1] - lower_triangle_offsets[counter]
nzA[nzind] = val
end
end
#! format: on
return nothing
end

function check_tree_csc_buffers(nzA, A_colptr, B, result::TreeSetColoringResult, uplo)
Base.require_one_based_indexing(nzA, A_colptr, B)
uplo in (:F, :L, :U) || throw(ArgumentError("`uplo` must be :F, :L, or :U."))
S = result.ag.S
n = size(S, 2)
size(B) == (n, ncolors(result)) ||
throw(DimensionMismatch("`B` must have size ($n, $(ncolors(result)))."))
length(A_colptr) == n + 1 ||
throw(DimensionMismatch("`A_colptr` must have length $(n + 1)."))

expected_nnz = uplo == :F ? nnz(S) : nb_edges(result.ag) + result.ag.nb_self_loops
length(nzA) == expected_nnz ||
throw(DimensionMismatch("`nzA` must have length $expected_nnz."))
return nothing
end

function decompress_csc!(
nzA::AbstractVector{R},
A_colptr::AbstractVector{<:Integer},
B::AbstractMatrix{R},
result::TreeSetColoringResult,
uplo::Symbol=:F,
) where {R<:Real}
check_tree_csc_buffers(nzA, A_colptr, B, result, uplo)
(;
ag,
color,
Expand All @@ -698,24 +735,30 @@ function decompress_csc!(
buffer_right_type = similar(buffer, R)
end

# The result constructor stores valid vertex/color indices and offsets in
# concrete Vectors. The checks above cover accesses to B and A_colptr; each
# computed destination index is checked immediately before writing to nzA.
# Recover the diagonal coefficients of A
if !augmented_graph(ag)
if uplo == :L
for i in diagonal_indices
# A[i, i] is the first element in column i
nzind = A_colptr[i]
nzA[nzind] = B[i, color[i]]
val = @inbounds B[i, color[i]]
nzA[nzind] = val
end
elseif uplo == :U
for i in diagonal_indices
# A[i, i] is the last element in column i
nzind = A_colptr[i + 1] - 1
nzA[nzind] = B[i, color[i]]
val = @inbounds B[i, color[i]]
nzA[nzind] = val
end
else # uplo == :F
for (k, i) in enumerate(diagonal_indices)
nzind = diagonal_nzind[k]
nzA[nzind] = B[i, color[i]]
nzind = @inbounds diagonal_nzind[k]
val = @inbounds B[i, color[i]]
nzA[nzind] = val
end
end
end
Expand All @@ -726,15 +769,16 @@ function decompress_csc!(
# Recover the off-diagonal coefficients of A
for k in 1:nt
# Positions of the first and last edges of the tree
first = tree_edge_indices[k]
last = tree_edge_indices[k + 1] - 1
first = @inbounds tree_edge_indices[k]
last = @inbounds tree_edge_indices[k + 1] - 1

# A star needs no substitution or scratch-buffer updates.
if is_star[k]
star = @inbounds is_star[k]
if star
for pos in first:last
(i, j) = reverse_bfs_orders[pos]
(i, j) = @inbounds reverse_bfs_orders[pos]
counter += 1
val = B[i, color[j]]
val = @inbounds B[i, color[j]]
set_symmetric_csc!(
nzA,
A_colptr,
Expand All @@ -750,18 +794,18 @@ function decompress_csc!(
else
# Reset the buffer to zero for all vertices in the tree (except the root)
for pos in first:last
(vertex, _) = reverse_bfs_orders[pos]
buffer_right_type[vertex] = zero(R)
(vertex, _) = @inbounds reverse_bfs_orders[pos]
@inbounds buffer_right_type[vertex] = zero(R)
end
# Reset the buffer to zero for the root vertex
(_, root) = reverse_bfs_orders[last]
buffer_right_type[root] = zero(R)
(_, root) = @inbounds reverse_bfs_orders[last]
@inbounds buffer_right_type[root] = zero(R)

for pos in first:last
(i, j) = reverse_bfs_orders[pos]
(i, j) = @inbounds reverse_bfs_orders[pos]
counter += 1
val = B[i, color[j]] - buffer_right_type[i]
buffer_right_type[j] = buffer_right_type[j] + val
val = @inbounds B[i, color[j]] - buffer_right_type[i]
@inbounds buffer_right_type[j] = buffer_right_type[j] + val

set_symmetric_csc!(
nzA,
Expand Down
68 changes: 68 additions & 0 deletions test/decompression.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,6 +188,74 @@ end
end
end

struct ZeroBasedDecompressionArray{T,N,A<:AbstractArray{T,N}} <: AbstractArray{T,N}
parent::A
end
Base.size(A::ZeroBasedDecompressionArray) = size(A.parent)
Base.axes(A::ZeroBasedDecompressionArray) = map(n -> 0:(n - 1), size(A))

@testset "Tree CSC buffer validation" begin
A = sparse([1, 2, 2, 3, 3, 4, 4], [2, 1, 3, 2, 4, 3, 4], ones(7), 6, 6)
result = coloring(
A,
ColoringProblem(; structure=:symmetric),
GreedyColoringAlgorithm(NaturalOrder(); decompression=:substitution),
)
B = compress(A, result)
for uplo in (:F, :L, :U)
expected = if uplo == :F
A
elseif uplo == :L
tril(A)
else
triu(A)
end
values = fill(NaN, nnz(expected))
pointers = expected.colptr
fill!(result.buffer, NaN)
for bad_B in (B[1:(end - 1), :], B[:, 1:(end - 1)], hcat(B, B))
@test_throws DimensionMismatch decompress_csc!(
values, pointers, bad_B, result, uplo
)
end
for bad_values in (values[1:(end - 1)], vcat(values, NaN))
@test_throws DimensionMismatch decompress_csc!(
bad_values, pointers, B, result, uplo
)
end
@test_throws DimensionMismatch decompress_csc!(
values, pointers[1:(end - 1)], B, result, uplo
)
@test_throws ArgumentError decompress_csc!(values, pointers, B, result, :invalid)
@test_throws ArgumentError decompress_csc!(
ZeroBasedDecompressionArray(values), pointers, B, result, uplo
)
@test_throws ArgumentError decompress_csc!(
values, ZeroBasedDecompressionArray(pointers), B, result, uplo
)
@test_throws ArgumentError decompress_csc!(
values, pointers, ZeroBasedDecompressionArray(B), result, uplo
)
# Every invalid call must fail before writing output or scratch data.
@test all(isnan, values)
@test all(isnan, result.buffer)
@test decompress_csc!(values, pointers, B, result, uplo) === nothing
@test values == nonzeros(expected)

# Destination writes retain their ordinary bounds checks. A bad
# pointer can fail after earlier coefficients have already been written.
for bad_pointer in (0, 2length(values) + 2)
bad_pointers = fill(bad_pointer, length(pointers))
@test_throws BoundsError decompress_csc!(values, bad_pointers, B, result, uplo)
storage = zeros(2length(values))
values_view = @view storage[1:2:end]
@test_throws BoundsError decompress_csc!(
values_view, bad_pointers, B, result, uplo
)
end
end
end

function test_tree_bicoloring_decompression(T, Ti, order, target)
problem = ColoringProblem(; structure=:nonsymmetric, partition=:bidirectional)
algorithm = GreedyColoringAlgorithm(
Expand Down
Loading