Crate matrixmultiply[−][src]
Expand description
General matrix multiplication for f32, f64 matrices. Operates on matrices with general layout (they can use arbitrary row and column stride).
This crate uses the same macro/microkernel approach to matrix multiplication as the BLIS project.
We presently provide a few good microkernels, portable and for x86-64, and only one operation: the general matrix-matrix multiplication (“gemm”).
Matrix Representation
matrixmultiply supports matrices with general stride, so a matrix is passed using a pointer and four integers:
a: *const f32, pointer to the first element in the matrixm: usize, number of rowsk: usize, number of columnsrsa: isize, row stridecsa: isize, column stride
In this example, A is a m by k matrix. a is a pointer to the element at
index 0, 0.
The row stride is the pointer offset (in number of elements) to the element on the next row. It’s the distance from element i, j to i + 1, j.
The column stride is the pointer offset (in number of elements) to the element in the next column. It’s the distance from element i, j to i, j + 1.
For example for a contiguous matrix, row major strides are rsa=k, csa=1 and column major strides are rsa=1, csa=m.
Strides can be negative or even zero, but for a mutable matrix elements may not alias each other.
Portability and Performance
-
The default kernels are written in portable Rust and available on all targets. These may depend on autovectorization to perform well.
-
x86 and x86-64 features can be detected at runtime by default or compile time (if enabled), and the crate following kernel variants are implemented:
fmaavxsse2
Features
std
std is enabled by default.
This crate can be used without the standard library (#![no_std]) by
disabling the default std feature. To do so, use this in your
Cargo.toml:
matrixmultiply = { version = "0.2", default-features = false }Runtime CPU feature detection is available only when std is enabled.
Without the std feature, the crate uses special CPU features only if they
are enabled at compile time. (To enable CPU features at compile time, pass
the relevant
target-cpu
or
target-feature
option to rustc.)
threading
threading is an optional crate feature
Threading enables multithreading for the operations. The environment variable
MATMUL_NUM_THREADS decides how many threads are used at maximum. At the moment 1-4 are
supported and the default is the number of physical cpus (as detected by num_cpus).
cgemm
cgemm is an optional crate feature.
It enables the cgemm and zgemm methods for complex matrix multiplication.
This is an experimental feature and not yet as performant as the float kernels on x86.
The complex representation we use is [f64; 2].
constconf
constconf is an optional feature. When enabled, cache-sensitive parameters of
the gemm implementations can be tweaked at compile time by defining the following variables:
MATMUL_SGEMM_MC(And so on, for S, D, C, ZGEMM and with NC, KC or MC).
Other Notes
The functions in this crate are thread safe, as long as the destination matrix is distinct.
Rust Version
This version requires Rust 1.41.1 or later; the crate follows a carefully considered upgrade policy, where updating the minimum Rust version is not a breaking change.
Enums
cgemm/zgemm per-operand options