Greetings and thanks for your hard work!I've made it a point to limit my comments about P1385. It's adifferent design that serves a different set of users. That beingsaid, I'd like to comment on outer products and rank-1 updates.On Mon, May 1, 2023 at 5:55 AM Guy Davidson via SG19<sg19@lists.isocpp.org> wrote:
Hello everyone
I'm just putting together a first pass at the wording for P1385, A proposal to add linear algebra support to the C++ standard library. If you look at the latest revision you will infer that at Kona in November and at Issaquah in February I addressed SG6 and LEWG about withdrawing the vector class entirely and simply offering a matrix class, where a vector is a special case of a matrix, with a single row or column. There were no objections to this approach.
While there were no objections raised in the meeting, others have come in, and I want to use the reflectors to gather opinion about the matter. The heart of the problem is: what does the vector product signify? Is it an inner or outer product? Is vector orientation significant?
With my mathematician's hat on, multiplying a row vector by a column vector is an inner product, yielding a scalar value if both vectors have the same number of elements. Appearing much more rarely, multiplying a column vector by a row vector is an outer product yielding a square matrix.
Outer products have a common use case which isn't represented in theproposal: Rank-1 update of an existing matrix. If you try to spellthat using overloaded arithmetic operators, you get the following,assuming that x is a row vector and y is a column vector.
(Did you mean x is a column and y a row?)
A += x * y;
A naive implementation would always create a new temporary matrix to
hold the outer product result x * y. The only way NOT to do that, and
still retain the syntax "A += x * y," would be to use expression
templates. ("x * y" would return outer_product_expression<X, Y>, and
matrix::operator+=(outer_product_expression<X, Y>&&) would perform a
rank-1 update.)
Right. But expression templates aren’t a panacea for matrix-based operations. E.g., if y is replaced by an expression that depends on A, you run into aliasing issues.
For many users of my libraries, A is M x N where M and N are often
much bigger than 4. Creating a temporary matrix on every rank-1
update would be prohibitively expensive. BLAS 2 - style in-place LU
factorization of an N x N matrix involves N - 1 outer products, and
would therefore involve N - 1 allocations of max size (N-1)^2.
However, LU factorization only needs kN extra space for a small
positive integer k. Even for 4 x 4 matrices and length-4 vectors, it
could save (more precious SIMD) registers and instructions not to
create a temporary matrix to hold the outer product result.
Therefore, if you want to retain use of operator* for outer products,
I conclude that the library would need to promise to use expression
templates. Otherwise, I conclude that the rank-1 update use case
would be better served by a named function (e.g., outer(x, y, A)) that
updates an existing matrix in place.
Agreed: I think a functional interface is more reliable at the outset.