Two process-wide statics are read and written with no synchronisation, so matrix operations called from more than one thread return wrong numbers. Nothing throws, nothing is malformed, and no result looks suspicious.
I found this from the other side: AngouriMath uses GenTensor for its symbolic matrices, and a determinant computed on a background thread was coming back numerically wrong. AngouriMath's issue is #1219.
1. The scratch-matrix pool
SquareMatrixFactory<T, TWrapper>.GetMatrix hands every caller the same GenTensor for a given size, and the caller then writes into it — Inversion.GetCofactorMatrix(t, temp, ...) fills the matrix it is given. Two threads taking a determinant or an inverse of the same size get the same buffer and overwrite each other's minors between the write and the read.
Sixty 5×5 int matrices, each computed once sequentially for truth and then rebuilt and recomputed under Parallel.For:
DeterminantLaplace 53 of 60 disagree
Adjoint 60 of 60 disagree
Around 10 ms, every run. This is not a rare interleaving, it is the normal case.
DeterminantGaussianSafeDivision is unaffected — it works on its own copy and never asks the pool for anything.
The lock inside GetMatrix does not help and is itself incomplete: the list is indexed after the lock is released, so a concurrent Add can reallocate the backing array underneath the reader.
2. The compiled-operation cache
ExpressionCompiler<T, TWrapper>.storage is a plain Dictionary, written by whichever thread first asks for a given (operation, rank, parallel). Dictionary<,> is documented as safe for concurrent readers only while nobody is writing.
Reached cold from several threads it fails with the runtime's own diagnostic:
System.InvalidOperationException: Operations that change non-concurrent collections must
have exclusive access. A concurrent update was performed on this collection and corrupted
its state. The collection's state is no longer correct.
...
The given key '(Subtraction, 1, False)' was not present in the dictionary.
Worth noting that this one hides easily. My first test computed its expected values sequentially and only then went parallel, which populated every key before the threads started — the cache holds a handful of entries, so a warm-up makes the defect invisible and the test passes against the broken code.
Fix
I have opened a PR. Both are small: [ThreadStatic] for the pool, since a scratch buffer has nothing to share between threads, and ConcurrentDictionary for the cache. The Laplace determinant comes out about 5% faster, because the pool is now taken once at the top of the recursion instead of re-checked and re-indexed at every one of its n! nodes.
Two process-wide statics are read and written with no synchronisation, so matrix operations called from more than one thread return wrong numbers. Nothing throws, nothing is malformed, and no result looks suspicious.
I found this from the other side: AngouriMath uses
GenTensorfor its symbolic matrices, and a determinant computed on a background thread was coming back numerically wrong. AngouriMath's issue is #1219.1. The scratch-matrix pool
SquareMatrixFactory<T, TWrapper>.GetMatrixhands every caller the sameGenTensorfor a given size, and the caller then writes into it —Inversion.GetCofactorMatrix(t, temp, ...)fills the matrix it is given. Two threads taking a determinant or an inverse of the same size get the same buffer and overwrite each other's minors between the write and the read.Sixty 5×5
intmatrices, each computed once sequentially for truth and then rebuilt and recomputed underParallel.For:Around 10 ms, every run. This is not a rare interleaving, it is the normal case.
DeterminantGaussianSafeDivisionis unaffected — it works on its own copy and never asks the pool for anything.The lock inside
GetMatrixdoes not help and is itself incomplete: the list is indexed after the lock is released, so a concurrentAddcan reallocate the backing array underneath the reader.2. The compiled-operation cache
ExpressionCompiler<T, TWrapper>.storageis a plainDictionary, written by whichever thread first asks for a given(operation, rank, parallel).Dictionary<,>is documented as safe for concurrent readers only while nobody is writing.Reached cold from several threads it fails with the runtime's own diagnostic:
Worth noting that this one hides easily. My first test computed its expected values sequentially and only then went parallel, which populated every key before the threads started — the cache holds a handful of entries, so a warm-up makes the defect invisible and the test passes against the broken code.
Fix
I have opened a PR. Both are small:
[ThreadStatic]for the pool, since a scratch buffer has nothing to share between threads, andConcurrentDictionaryfor the cache. The Laplace determinant comes out about 5% faster, because the pool is now taken once at the top of the recursion instead of re-checked and re-indexed at every one of its n! nodes.