matrix multiplication over integer mod ring is slow

[sagetrac-dmharvey]

Sage 4.5.3, 2.6GHz Opteron, Linux

This is ok:

sage: M1 = Matrix([[randrange(3^20) for i in range(100)] for j in range(100)])
sage: M2 = Matrix([[randrange(3^20) for i in range(100)] for j in range(100)])
sage: timeit("M3 = M1 * M2")
5 loops, best of 3: 45.6 ms per loop

(That's about 4 times slower than Magma, but I can put up with that, that's a ticket for another day.)

Here is the problem:

sage: R = Integers(3^20)
sage: M1 = Matrix([[R.random_element() for i in range(100)] for j in range(100)])
sage: M2 = Matrix([[R.random_element() for i in range(100)] for j in range(100)])
sage: timeit("M3 = M1 * M2")
5 loops, best of 3: 877 ms per loop

In other words, I can multiply the matrices over R roughly 20x faster by multiplying over Z and then reducing! That's ridiculous!

Component: performance

Keywords: sd90

Issue created by migration from https://trac.sagemath.org/ticket/9888

[robertwb]

comment:1

I don't think anything has gone into non-word-sized modulus, so this is probably using totally generic per-element wrapping code :(. Should be an easy fix to get better than this, doing something real would be a bit more work.

[kedlaya]

comment:2

I just tried the timings again:

sage: sage: M1 = Matrix([[randrange(3^20) for i in range(100)] for j in range(100)])
sage: sage: M2 = Matrix([[randrange(3^20) for i in range(100)] for j in range(100)])
sage: sage: timeit("M3 = M1 * M2")
125 loops, best of 3: 5.62 ms per loop
sage: sage: R = Integers(3^20)
sage: sage: M1 = Matrix([[R.random_element() for i in range(100)] for j in range(100)])
sage: sage: M2 = Matrix([[R.random_element() for i in range(100)] for j in range(100)])
sage: sage: timeit("M3 = M1 * M2")
5 loops, best of 3: 530 ms per loop

so now the discrepancy is up to a factor of 100!

My recollection is that lifting the multiplication up to Z is in fact the correct algorithmic approach. In practice, this hands the problem off to FLINT, where (in this size range) the multiplication is done multimodular.

[kedlaya]

comment:3

See #12177 for a related discussion.

[adeines]

Changed keywords from none to sd90

[kedlaya]

comment:5

There appears to be special-purpose code using Linbox for modulus up to 2²³ in sage/matrix/matrix_modn_dense_double.pyx and sage/matrix/matrix_modn_dense_float.pyx. To handle this issue, it would be best to create a file sage/matrix/matrix_modn_dense.pyx in which we create the class Matrix_modn_dense with a special _mul_ method. But we should make sure not to create a regression by disconnecting the existing code for smaller moduli.