This is an alpha release of foreign-numeric-vector, a package of
functions and accessors for working with numeric vectors stored in the
foreign (non-Lisp) heap. The hope is that this package will be useful
for implementing foreign function interfaces to foreign libraries
(e.g. BLAS, LAPACK, FFTW). It is especially useful for wrapping
libraries portably across CL implementations which vary in their
choices of boxing or unboxing of individual elements of native arrays,
and which may or may not offer the ability to protect an array from
being moved by the CL garbage collector.
foreign-numeric-vector requires CFFI and ITERATE. The dependence on
CFFI is fundamental. The dependence on ITERATE could be removed
fairly easily. I had originally thought that a good approach would be
to use iterate to integrate the foreign-numeric-vectors with other
sequences. Currently, it is not clear to me that ITERATE makes it
sufficiently easy to generate sufficiently efficient code; this is
something I need to investigate further.
There is no real documentation, but usage is fairly straightforward.
Basic usage (showing a couple different modes of access, and using
some included utilities) looks something like:
(defun double-float-readme-example (length)
(let ((fnv (make-fnv-double length)))
(fixtimes (i length)
(setf (fnv-double-ref fnv i) (random 1.0d0)))
(let ((sum 0.0d0))
(declare (type double-float sum))
(over-fnv-double (v) fnv
(dfinc sum v))
(format t "Sum of array is ~A~%" sum))
(format t "The smallest few entries: ~A~%" fnv)
(format t "The largest few entries: ~A~%" fnv)
See the file examples.lisp for some basic examples, discussion of
efficient choices, and some benchmarks to run.
Substantial effort has been made to make foreign-numeric-vector
efficient. We make great use of type declarations and compiler macros
to promote inlining. We are of course reliant on the underlying
compilers and foreign function interfaces, and on CFFI. For instance,
on SBCL, it is much faster to use mem-ref rather than mem-aref, making
our own guarantees that the index to byte conversion involves only
fixnums. More testing is required, but preliminary testing on SBCL
indicates that for double-float operations, we are competitive with
native array accesses in all cases. For complex double-floats,
writing the complex number #C(1.0d0 1.0d0) to a foreign-numeric-vector
is about twice as slow as writing it to a native array, but adding up
a bunch of complex numbers is about the same speed. Sorting is quite
efficient, because the comparison functions are required at compile
time and get inlined. We use a template style architecture and
macro-generating-macros to make the operations on different types.
The code is released under the BSD license without the advertising
clause (see LICENSE).
As of 04 November 2006, this code is available at
All comments, suggestions, performance reports, and assistance should
be sent to
Additional maintainance by AJ Rossini <firstname.lastname@example.org>,
distributed via github (http://github.com/blindglobe/fnv.git)
(FIXME: VERIFY URL!).