Struct Rates 

pub struct Rates {

    pub gemm_mac: Rate,
    pub gemm_single: Rate,
    pub block_apply_mac: Rate,
    pub lu_full_pivot: Rate,
    pub bunch_kaufman: Rate,
    pub qr_col_pivot: Rate,
    pub bidiag: Rate,
    pub hessenberg: Rate,
    pub tridiag: Rate,
    pub matvec: Rate,
    pub bidiag_svd_dc: Rate,
    pub tridiag_evd_dc: Rate,
    pub triangular_inverse: Rate,
    pub sparse_matmul: Rate,
    pub householder_rank_one: Rate,
    pub lu_swap: Rate,
}

The serial throughput constants of the kernel families of one host class, one per family (design record: threadteam docs/design/thresholds.md, section 2.3). The unit is what the gate of the family counts. A step of a factorization is timed inside the whole serial factorization, so the constant includes the serial parts of a step (pivot search, Householder generation) and slightly overestimates the time of the part that is split.

The row of a class is Rates::for_class; rates is the row of the host. Sources:

• x86_64-avx512: one thread on an AMD Ryzen AI 9 HX 470 (Zen 5) at 2.0 GHz, boost off, median over shapes and rounds (faer-benchmark data/probes/20261026-awake-team-linux/tables-consts-exec3.md); the divide and conquer and sparse product from the region constants of the same file (raw-exec3/rconsts-*.csv).
• aarch64-apple: one thread on an Apple M5 Max, QoS UserInitiated, L2 resident shapes, median over the shapes 96 to 768 and two processes (data/probes/20261026-threshold-recalibration).
• every other class is unmeasured and takes the generic row: per entry the smaller of the measured rows, so that an unknown host keeps work serial a little longer than it must rather than dispatching too early.

A row’s rates and the dispatch costs of the same class must come from the same machine and clock state (HostClass).

Fields

gemm_mac: Rate

GEMM and every kernel counted in multiply-adds (triangular products and solves), per multiply-add.

gemm_single: Rate

GEMM in single precision, per multiply-add: real is f32, cplx is c32.

block_apply_mac: Rate

The multiply-adds of the block Householder apply, per multiply-add.

lu_full_pivot: Rate

Full-pivoting LU, one step: rank-one update fused with the pivot search, per element of the trailing block (m-k-1)*(n-k-1).

bunch_kaufman: Rate

Bunch-Kaufman with full pivoting, one step: rank-1 or rank-2 update fused with the pivot search, per element of the trailing block rem*rem.

qr_col_pivot: Rate

QR with column pivoting, one step: fused update of the trailing block and the norms, per element (m-k-1)*(n-k-1).

bidiag: Rate

Bidiagonalization, one step: the fused two-sided update and the matrix-vector product, per element (m-k-1)*(n-k-1).

hessenberg: Rate

Hessenberg reduction (unblocked), one step: two matrix-vector products and a rank-one update, per element n*n of the trailing block.

tridiag: Rate

Tridiagonalization, one step: the symmetric fused update, per element rem*rem/2 of the triangle.

matvec: Rate

Matrix-vector product, per matrix element (the mean of the column major and row major products).

bidiag_svd_dc: Rate

Divide and conquer bidiagonal SVD (bidiag_svd, both singular vectors, non deflating data), the whole serial solve of size m per m^2; the join of size n has the two halves as work, 2 (n/2)^2. The m^2 term (secular equation, deflation, permutations) is larger than the merge products below m about 1900, so the unit is m^2; the constant is the median over m 64 to 512 (it is size dependent: 48 to 62 on the M5 fit, 145 to 178 on the Zen 5). Real types only (the complex value repeats the real one).

tridiag_evd_dc: Rate

Divide and conquer symmetric tridiagonal EVD (tridiag_evd, eigenvectors on), the whole serial solve of size m per m^2, used as bidiag_svd_dc (88 to 102 on the Zen 5 over m 64 to 512). Real types only.

triangular_inverse: Rate

Triangular inverse (invert_*_triangular), the whole serial inverse of size m per m^3; the join of size n has the two halves as work, 2 (n/2)^3. The constant is the median over m 64 to 512 and falls with the size (on the Zen 5, f64 0.047 per m^3 at 64 to 0.0145 at 512), so a join near the floor is a little overestimated.

sparse_matmul: Rate

Sparse by sparse numeric product, per multiply-add of the symbolic flop count (SparseMatMulInfo, free to read); the median over the measured densities and sizes.

householder_rank_one: Rate

Rank-one Householder apply (N == 1), per element of the target M*K.

lu_swap: Rate

One row swap of one column in the LU swap stage (random pivot rows, columns of 64 to 1024 elements, cache resident, one thread; memory bound, so the c64 value is the larger element width, not four times the f64 one).

Implementations

impl Rates

pub const X86_64_AVX512: Rates

x86_64-avx512: Zen 5 at 2.0 GHz, boost off (see the type docs).

pub const AARCH64_APPLE: Rates

aarch64-apple: Apple M5 Max (the constants of the first calibration).

pub const GENERIC: Rates

generic (every unmeasured class): per entry the smaller of the measured rows.

pub fn for_class(class: HostClass) -> &'static Rates

The row of a host class.

Trait Implementations

impl Clone for Rates

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for Rates

impl Debug for Rates

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for Rates

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Rates