Skip to content

Clear the output of sparse products when beta is zero #702

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

Merged
maleadt merged 1 commit into from
Oct 16, 2026
Merged

Clear the output of sparse products when beta is zero #702

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 18 additions & 0 deletions lib/mkl/wrappers_sparse.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,14 @@ const onemklSparseInt = Union{Int32, Int64}
# matrices without stored entries never get a handle: the operations short-circuit instead
_mkl_empty(A::oneAbstractSparseMatrix) = nnz(A) == 0 || any(iszero, size(A))

# `mul!` ignores the output when beta is zero, and callers pass uninitialized memory (e.g.
# from `similar`), but oneMKL computes beta * y regardless in some cases (gemv with a
# transposed CSR matrix, and gemm on older releases), letting NaNs into the result (#703).
function _clear_ignored_output!(beta::Number, y::AbstractArray)
iszero(beta) && fill!(y, zero(eltype(y)))
return y
end

# result of op(A) * x for an empty A
function _scale_output!(beta::Number, y::AbstractArray)
if iszero(beta)
Expand Down Expand Up @@ -210,6 +218,7 @@ for SparseMatrix in (:oneSparseMatrixCSR, :oneSparseMatrixCOO)
y::oneStridedVector{$elty})

_mkl_empty(A) && return _scale_output!(beta, y)
_clear_ignored_output!(beta, y)
queue = global_queue(context(x), device(x))
$fname(sycl_queue(queue), trans, alpha, sparse_matrix_handle(A), x, beta, y)
y
Expand Down Expand Up @@ -250,6 +259,7 @@ for SparseMatrix in (:oneSparseMatrixCSC,)
queue = global_queue(context(x), device(x))
m, n = size(A)
if m != 0 && n != 0
_clear_ignored_output!(beta, y)
$fname(sycl_queue(queue), flip_trans(trans), alpha, sparse_matrix_handle(A), x, beta, y)
end
y
Expand Down Expand Up @@ -284,6 +294,7 @@ for SparseMatrix in (:oneSparseMatrixCSC,)
beta = conj(beta)
end

_clear_ignored_output!(beta, y)
queue = global_queue(context(x), device(x))
$fname(sycl_queue(queue), flip_trans(trans), alpha, sparse_matrix_handle(A), x, beta, y)

Expand Down Expand Up @@ -331,6 +342,7 @@ for (fname, elty) in ((:onemklSsparse_gemm, :Float32),
ldb = max(1,stride(B,2))
ldc = max(1,stride(C,2))
queue = global_queue(context(C), device(C))
_clear_ignored_output!(beta, C)
$fname(sycl_queue(queue), 'C', transa, transb, alpha, sparse_matrix_handle(A), B, nrhs, ldb, beta, C, ldc)
C
end
Expand Down Expand Up @@ -371,6 +383,7 @@ for (fname, elty) in ((:onemklSsparse_gemm, :Float32),
ldb = max(1,stride(B,2))
ldc = max(1,stride(C,2))
queue = global_queue(context(C), device(C))
_clear_ignored_output!(beta, C)
$fname(sycl_queue(queue), 'C', flip_trans(transa), transb, alpha, sparse_matrix_handle(A), B, nrhs, ldb, beta, C, ldc)
C
end
Expand Down Expand Up @@ -437,6 +450,7 @@ for (fname, elty) in (
transb_eff = transb
end

_clear_ignored_output!(beta, C)
$fname(sycl_queue(queue), 'C', flip_trans(transa), transb_eff, alpha, sparse_matrix_handle(A), B, nrhs, ldb, beta, C, ldc)

# Undo conjugation to obtain C_new
Expand Down Expand Up @@ -480,6 +494,7 @@ for (fname, elty) in ((:onemklSsparse_symv, :Float32),

_mkl_empty(A) && return _scale_output!(beta, y)
queue = global_queue(context(y), device(y))
_clear_ignored_output!(beta, y)
$fname(sycl_queue(queue), uplo, alpha, sparse_matrix_handle(A), x, beta, y)
y
end
Expand All @@ -501,6 +516,7 @@ for (fname, elty) in ((:onemklSsparse_symv, :Float32),

_mkl_empty(A) && return _scale_output!(beta, y)
queue = global_queue(context(y), device(y))
_clear_ignored_output!(beta, y)
$fname(sycl_queue(queue), flip_uplo(uplo), alpha, sparse_matrix_handle(A), x, beta, y)
y
end
Expand All @@ -523,6 +539,7 @@ for (fname, elty) in ((:onemklSsparse_trmv, :Float32),

_mkl_empty(A) && return _empty_trmv!(diag, alpha, x, beta, y)
queue = global_queue(context(y), device(y))
_clear_ignored_output!(beta, y)
$fname(sycl_queue(queue), uplo, trans, diag, alpha, sparse_matrix_handle(A), x, beta, y)
y
end
Expand Down Expand Up @@ -567,6 +584,7 @@ for (fname, elty) in (
)
)
queue = global_queue(context(y), device(y))
_clear_ignored_output!(beta, y)
$fname(sycl_queue(queue), uplo, flip_trans(trans), diag, alpha, sparse_matrix_handle(A), x, beta, y)
return y
end
Expand Down
21 changes: 21 additions & 0 deletions test/onemkl.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1381,6 +1381,27 @@ end
@test collect(UnitLowerTriangular(E) \ oneVector(x)) ≈ x
end

@testset "beta = 0 ignores the output" begin
# empty rows and columns produce outputs without any contribution
A = SparseMatrixCSC{T, Int32}(sprand(T, 20, 10, 0.3))
A[2, :] .= 0
A[:, 3] .= 0
dropzeros!(A)
@testset "$SparseMatrix" for SparseMatrix in csr_csc_matrices
dA = SparseMatrix(A)
@testset "opa = $(nameof(opa))" for opa in (identity, transpose, adjoint)
x = rand(T, size(opa(A), 2))
y = oneVector(fill(T(NaN), size(opa(A), 1)))
mul!(y, opa(dA), oneVector(x))
@test collect(y) ≈ opa(A) * x
X = rand(T, size(opa(A), 2), 2)
Y = oneMatrix(fill(T(NaN), size(opa(A), 1), 2))
mul!(Y, opa(dA), oneMatrix(X))
@test collect(Y) ≈ opa(A) * X
end
end
end

@testset "sparse gemv" begin
@testset "$SparseMatrix" for SparseMatrix in coo_csr_csc_matrices
@testset "transa = $transa" for (transa, opa) in [('N', identity), ('T', transpose), ('C', adjoint)]
Expand Down
Loading