Interface to HiGHS as an LP solver #72
Comments
|
An interface to the LP solver HiGHS was added in #74. For QPs, it's a bit trickier because HiGHS expects a compressed sparse Hessian in one of the following shapes:
At the moment, the CSC sparse format in Uno is only column-wise upper triangular. I need to investigate this further. |
|
@cvanaret function MyTranspose(A::SparseMatrixCSC{T,Cint}) where {T}
m, n = size(A)
nnzA = nnz(A)
A_colptr = A.colptr
A_rowval = A.rowval
A_nzval = A.nzval
# Allocate storage for the column pointers and row indices of B = Aᵀ
B_colptr = zeros(Cint, m + 1)
B_rowval = Vector{Cint}(undef, nnzA)
B_nzval = Vector{T}(undef,
8000
nnzA)
# Count the number of non-zeros for each row of A.
# It corresponds to the number of non-zeros for each column of B = Aᵀ.
for k in 1:nnzA
i = A_rowval[k]
B_colptr[i] += 1
end
# Compute the cumulative sum to determine the starting positions of rows in B_rowval
counter = 1
for col in 1:m
nnz_col = B_colptr[col]
B_colptr[col] = counter
counter += nnz_col
end
B_colptr[m + 1] = counter
# Store the row indices for each column of B = Aᵀ
for j in 1:n
for index in A_colptr[j]:(A_colptr[j + 1] - 1)
i = A_rowval[index]
# Update B_rowval for the non-zero B[j,i].
# It corresponds to the non-zero A[i,j].
pos = B_colptr[i]
B_rowval[pos] = j
B_nzval[pos] = A_nzval[index]
B_colptr[i] += 1
end
end
# Fix offsets of B_colptr to restore correct starting positions
for col in m:-1:2
B_colptr[col] = B_colptr[col - 1]
end
B_colptr[1] = 1
return SparseMatrixCSC{T,Cint}(n, m, B_colptr, B_rowval, B_nzval)
endIt should not be hard to adapt in C++. |
|
Thanks a lot, much appreciated! This is a dumb situation when you think that the immense majority of the Hessian occurrences in HiGHS are matrix-vector products :) I think |
|
Yes, it's a typo. It should be |
|
How do I enable the tests with the LP solver? Is there a preset that will use the LP solver and not the QP? |
Add an interface to HiGHS as an LP solver. This should be relatively simple, since HiGHS is implemented in C++.
The text was updated successfully, but these errors were encountered: