auto_contractor.benchmark — Matrix Operation Benchmarks

Source: qlat/auto_contractor/benchmark.py

Note: Update this document when updating the source file.

Outline

  1. Overview

  2. Benchmark Functions

  3. Main Block

  4. Examples

Overview

benchmark provides timer-instrumented wrappers for measuring the throughput of the low-level matrix operations that dominate auto-contractor evaluation cost. Each benchmark function calls its target repeatedly in batches of 10 inside a q.TimerFork context so that per-call timing and FLOP rates are recorded by the Qlattice profiling system.

Benchmark Functions

benchmark_function_1(f, arg, ...)

Benchmark a single-argument function f(arg).

Parameter

Type

Default

Description

f

callable

—

Function to benchmark

arg

any

—

Argument passed to f

benchmark_size

int

1000

Number of outer loop iterations

benchmark_num

int

10

Number of benchmark repetitions

total_flops

int

0

FLOP count reported to the timer

benchmark_function_2(f, arg1, arg2, ...)

Benchmark a two-argument function f(arg1, arg2). Parameters are identical to benchmark_function_1 except for the extra argument.

benchmark_function_3(f, arg1, arg2, arg3, ...)

Benchmark a three-argument function f(arg1, arg2, arg3). Same parameter conventions.

Main Block

When run as a script, the module benchmarks the core spin-colour matrix operations used by the auto-contractor runtime:

Operation

Function

FLOP Count

sc * sc

mat_mul_sc_sc

13536

sc * s

mat_mul_sc_s

4320

s * sc

mat_mul_s_sc

4320

s * s

mat_mul_s_s

4320

tr(sc, sc)

mat_sc_sc_trace

480

tr(sc)

mat_sc_trace

22

Random test matrices are created via make_rand_spin_color_matrix and make_rand_spin_matrix with a fixed RngState("seed") for reproducibility.

Examples

import qlat as q
import auto_contractor as ac

q.begin_with_mpi([[1, 1, 1, 4]])

from auto_contractor.benchmark import benchmark_function_2
from auto_contractor.eval import make_rand_spin_color_matrix, get_mat, mat_mul_sc_sc

rng = q.RngState("bench-test")
m1 = make_rand_spin_color_matrix(rng)
m2 = make_rand_spin_color_matrix(rng)

benchmark_function_2(
    mat_mul_sc_sc, get_mat(m1), get_mat(m2),
    benchmark_size=100, benchmark_num=5, total_flops=13536,
)

q.end_with_mpi()