Replace allowed_getindex with only() for GPU compatibility

Co-authored-by: albertomercurio <61953577+albertomercurio@users.noreply.github.com>

Author: Copilot

src/matrix_coo/matrix_coo.jl

@@ -408,7 +408,7 @@ function Base.:+(A::DeviceSparseMatrixCOO, B::DeviceSparseMatrixCOO)
 
     # Compute write indices using cumsum
     write_indices = _cumsum_AK(keep_mask)
-    nnz_final = allowed_getindex(write_indices, nnz_concat)
+    nnz_final = only(write_indices[nnz_concat:nnz_concat])
 
     # Allocate final arrays
     rowind_C = similar(getrowind(A), nnz_final)
@@ -514,7 +514,7 @@ for (wrapa, transa, conja, unwrapa, whereT1) in trans_adj_wrappers(:DeviceSparse
             kernel_mark!(keep_mask, rowind_sorted, colind_sorted, nnz_concat; ndrange = (nnz_concat,))
 
             write_indices = _cumsum_AK(keep_mask)
-            nnz_final = allowed_getindex(write_indices, nnz_concat)
+            nnz_final = only(write_indices[nnz_concat:nnz_concat])
 
             rowind_C = similar(getrowind(_A), nnz_final)
             colind_C = similar(getcolind(_A), nnz_final)

src/matrix_csc/matrix_csc.jl

@@ -379,7 +379,7 @@ function Base.:+(A::DeviceSparseMatrixCSC, B::DeviceSparseMatrixCSC)
     colptr_C[1:1] .= one(Ti)
 
     # Allocate result arrays
-    nnz_total = allowed_getindex(colptr_C, n + 1) - one(Ti)
+    nnz_total = only(colptr_C[n+1:n+1]) - one(Ti)
     rowval_C = similar(getrowval(A), nnz_total)
     nzval_C = similar(nonzeros(A), Tv, nnz_total)
 

src/matrix_csr/matrix_csr.jl

@@ -376,7 +376,7 @@ function Base.:+(A::DeviceSparseMatrixCSR, B::DeviceSparseMatrixCSR)
     rowptr_C[1:1] .= one(Ti)
 
     # Allocate result arrays
-    nnz_total = allowed_getindex(rowptr_C, m + 1) - one(Ti)
+    nnz_total = only(rowptr_C[m+1:m+1]) - one(Ti)
     colval_C = similar(getcolval(A), nnz_total)
     nzval_C = similar(nonzeros(A), Tv, nnz_total)