Streamline the way that test iteration count is determined

Currently, tests use a handful of constants to determine how many
iterations to perform: `NTESTS`, `AROUND`, and `MAX_CHECK_POINTS`. This
configuration is not very straightforward to adjust and needs to be
repeated everywhere it is used.

Replace this with new functions in the `run_cfg` module that determine
iteration counts in a more reusable and documented way.

This only updates `edge_cases` and `domain_logspace`, `random` is
refactored in a later commit.
This commit is contained in:
Trevor Gross
2024-12-26 07:44:54 +00:00
parent b7b22b54c0
commit 9a7a519370
5 changed files with 208 additions and 57 deletions

View File

@@ -6,41 +6,26 @@ use libm::support::{IntTy, MinInt};
use crate::domain::HasDomain;
use crate::op::OpITy;
use crate::run_cfg::{GeneratorKind, iteration_count};
use crate::{CheckCtx, MathOp, logspace};
/// Number of tests to run.
// FIXME(ntests): replace this with a more logical algorithm
const NTESTS: usize = {
if cfg!(optimizations_enabled) {
if crate::emulated()
|| !cfg!(target_pointer_width = "64")
|| cfg!(all(target_arch = "x86_64", target_vendor = "apple"))
{
// Tests are pretty slow on non-64-bit targets, x86 MacOS, and targets that run
// in QEMU.
100_000
} else {
5_000_000
}
} else {
// Without optimizations just run a quick check
800
}
};
/// Create a range of logarithmically spaced inputs within a function's domain.
///
/// This allows us to get reasonably thorough coverage without wasting time on values that are
/// NaN or out of range. Random tests will still cover values that are excluded here.
pub fn get_test_cases<Op>(_ctx: &CheckCtx) -> impl Iterator<Item = (Op::FTy,)>
pub fn get_test_cases<Op>(ctx: &CheckCtx) -> impl Iterator<Item = (Op::FTy,)>
where
Op: MathOp + HasDomain<Op::FTy>,
IntTy<Op::FTy>: TryFrom<usize>,
IntTy<Op::FTy>: TryFrom<u64>,
RangeInclusive<IntTy<Op::FTy>>: Iterator,
{
let domain = Op::DOMAIN;
let ntests = iteration_count(ctx, GeneratorKind::Domain, 0);
// We generate logspaced inputs within a specific range, excluding values that are out of
// range in order to make iterations useful (random tests still cover the full range).
let start = domain.range_start();
let end = domain.range_end();
let steps = OpITy::<Op>::try_from(NTESTS).unwrap_or(OpITy::<Op>::MAX);
let steps = OpITy::<Op>::try_from(ntests).unwrap_or(OpITy::<Op>::MAX);
logspace(start, end, steps).map(|v| (v,))
}

View File

@@ -3,18 +3,11 @@
use libm::support::Float;
use crate::domain::HasDomain;
use crate::run_cfg::{check_near_count, check_point_count};
use crate::{CheckCtx, FloatExt, MathOp};
/// Number of values near an interesting point to check.
// FIXME(ntests): replace this with a more logical algorithm
const AROUND: usize = 100;
/// Functions have infinite asymptotes, limit how many we check.
// FIXME(ntests): replace this with a more logical algorithm
const MAX_CHECK_POINTS: usize = 10;
/// Create a list of values around interesting points (infinities, zeroes, NaNs).
pub fn get_test_cases<Op, F>(_ctx: &CheckCtx) -> impl Iterator<Item = (F,)>
pub fn get_test_cases<Op, F>(ctx: &CheckCtx) -> impl Iterator<Item = (F,)>
where
Op: MathOp<FTy = F> + HasDomain<F>,
F: Float,
@@ -25,23 +18,26 @@ where
let domain_start = domain.range_start();
let domain_end = domain.range_end();
let check_points = check_point_count(ctx);
let near_points = check_near_count(ctx);
// Check near some notable constants
count_up(F::ONE, values);
count_up(F::ZERO, values);
count_up(F::NEG_ONE, values);
count_down(F::ONE, values);
count_down(F::ZERO, values);
count_down(F::NEG_ONE, values);
count_up(F::ONE, near_points, values);
count_up(F::ZERO, near_points, values);
count_up(F::NEG_ONE, near_points, values);
count_down(F::ONE, near_points, values);
count_down(F::ZERO, near_points, values);
count_down(F::NEG_ONE, near_points, values);
values.push(F::NEG_ZERO);
// Check values near the extremes
count_up(F::NEG_INFINITY, values);
count_down(F::INFINITY, values);
count_down(domain_end, values);
count_up(domain_start, values);
count_down(domain_start, values);
count_up(domain_end, values);
count_down(domain_end, values);
count_up(F::NEG_INFINITY, near_points, values);
count_down(F::INFINITY, near_points, values);
count_down(domain_end, near_points, values);
count_up(domain_start, near_points, values);
count_down(domain_start, near_points, values);
count_up(domain_end, near_points, values);
count_down(domain_end, near_points, values);
// Check some special values that aren't included in the above ranges
values.push(F::NAN);
@@ -50,9 +46,9 @@ where
// Check around asymptotes
if let Some(f) = domain.check_points {
let iter = f();
for x in iter.take(MAX_CHECK_POINTS) {
count_up(x, values);
count_down(x, values);
for x in iter.take(check_points) {
count_up(x, near_points, values);
count_down(x, near_points, values);
}
}
@@ -65,11 +61,11 @@ where
/// Add `AROUND` values starting at and including `x` and counting up. Uses the smallest possible
/// increments (1 ULP).
fn count_up<F: Float>(mut x: F, values: &mut Vec<F>) {
fn count_up<F: Float>(mut x: F, points: u64, values: &mut Vec<F>) {
assert!(!x.is_nan());
let mut count = 0;
while x < F::INFINITY && count < AROUND {
while x < F::INFINITY && count < points {
values.push(x);
x = x.next_up();
count += 1;
@@ -78,11 +74,11 @@ fn count_up<F: Float>(mut x: F, values: &mut Vec<F>) {
/// Add `AROUND` values starting at and including `x` and counting down. Uses the smallest possible
/// increments (1 ULP).
fn count_down<F: Float>(mut x: F, values: &mut Vec<F>) {
fn count_down<F: Float>(mut x: F, points: u64, values: &mut Vec<F>) {
assert!(!x.is_nan());
let mut count = 0;
while x > F::NEG_INFINITY && count < AROUND {
while x > F::NEG_INFINITY && count < points {
values.push(x);
x = x.next_down();
count += 1;

View File

@@ -12,6 +12,7 @@ use crate::{BaseName, CheckCtx, GenerateInput};
const SEED: [u8; 32] = *b"3.141592653589793238462643383279";
/// Number of tests to run.
// FIXME(ntests): clean this up when possible
const NTESTS: usize = {
if cfg!(optimizations_enabled) {
if crate::emulated()

View File

@@ -25,7 +25,7 @@ pub use libm::support::{Float, Int, IntTy, MinInt};
pub use num::{FloatExt, logspace};
pub use op::{BaseName, FloatTy, Identifier, MathOp, OpCFn, OpFTy, OpRustFn, OpRustRet, Ty};
pub use precision::{MaybeOverride, SpecialCase, default_ulp};
pub use run_cfg::{CheckBasis, CheckCtx};
pub use run_cfg::{CheckBasis, CheckCtx, EXTENSIVE_ENV, GeneratorKind};
pub use test_traits::{CheckOutput, GenerateInput, Hex, TupleCall};
/// Result type for tests is usually from `anyhow`. Most times there is no success value to

View File

@@ -1,13 +1,11 @@
//! Configuration for how tests get run.
#![allow(unused)]
use std::collections::BTreeMap;
use std::env;
use std::sync::LazyLock;
use crate::{BaseName, FloatTy, Identifier, op};
use crate::{BaseName, FloatTy, Identifier, test_log};
/// The environment variable indicating which extensive tests should be run.
pub const EXTENSIVE_ENV: &str = "LIBM_EXTENSIVE_TESTS";
/// Context passed to [`CheckOutput`].
@@ -49,3 +47,174 @@ pub enum CheckBasis {
/// Check against infinite precision (MPFR).
Mpfr,
}
/// The different kinds of generators that provide test input.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum GeneratorKind {
Domain,
Random,
}
/// A list of all functions that should get extensive tests.
///
/// This also supports the special test name `all` to run all tests, as well as `all_f16`,
/// `all_f32`, `all_f64`, and `all_f128` to run all tests for a specific float type.
static EXTENSIVE: LazyLock<Vec<Identifier>> = LazyLock::new(|| {
let var = env::var(EXTENSIVE_ENV).unwrap_or_default();
let list = var.split(",").filter(|s| !s.is_empty()).collect::<Vec<_>>();
let mut ret = Vec::new();
let append_ty_ops = |ret: &mut Vec<_>, fty: FloatTy| {
let iter = Identifier::ALL.iter().filter(move |id| id.math_op().float_ty == fty).copied();
ret.extend(iter);
};
for item in list {
match item {
"all" => ret = Identifier::ALL.to_owned(),
"all_f16" => append_ty_ops(&mut ret, FloatTy::F16),
"all_f32" => append_ty_ops(&mut ret, FloatTy::F32),
"all_f64" => append_ty_ops(&mut ret, FloatTy::F64),
"all_f128" => append_ty_ops(&mut ret, FloatTy::F128),
s => {
let id = Identifier::from_str(s)
.unwrap_or_else(|| panic!("unrecognized test name `{s}`"));
ret.push(id);
}
}
}
ret
});
/// Information about the function to be tested.
#[derive(Debug)]
struct TestEnv {
/// Tests should be reduced because the platform is slow. E.g. 32-bit or emulated.
slow_platform: bool,
/// The float cannot be tested exhaustively, `f64` or `f128`.
large_float_ty: bool,
/// Env indicates that an extensive test should be run.
should_run_extensive: bool,
/// Multiprecision tests will be run.
mp_tests_enabled: bool,
/// The number of inputs to the function.
input_count: usize,
}
impl TestEnv {
fn from_env(ctx: &CheckCtx) -> Self {
let id = ctx.fn_ident;
let op = id.math_op();
let will_run_mp = cfg!(feature = "test-multiprecision");
// Tests are pretty slow on non-64-bit targets, x86 MacOS, and targets that run in QEMU. Start
// with a reduced number on these platforms.
let slow_on_ci = crate::emulated()
|| usize::BITS < 64
|| cfg!(all(target_arch = "x86_64", target_vendor = "apple"));
let slow_platform = slow_on_ci && crate::ci();
let large_float_ty = match op.float_ty {
FloatTy::F16 | FloatTy::F32 => false,
FloatTy::F64 | FloatTy::F128 => true,
};
let will_run_extensive = EXTENSIVE.contains(&id);
let input_count = op.rust_sig.args.len();
Self {
slow_platform,
large_float_ty,
should_run_extensive: will_run_extensive,
mp_tests_enabled: will_run_mp,
input_count,
}
}
}
/// The number of iterations to run for a given test.
pub fn iteration_count(ctx: &CheckCtx, gen_kind: GeneratorKind, argnum: usize) -> u64 {
let t_env = TestEnv::from_env(ctx);
// Ideally run 5M tests
let mut domain_iter_count: u64 = 4_000_000;
// Start with a reduced number of tests on slow platforms.
if t_env.slow_platform {
domain_iter_count = 100_000;
}
// Larger float types get more iterations.
if t_env.large_float_ty {
domain_iter_count *= 4;
}
// Functions with more arguments get more iterations.
let arg_multiplier = 1 << (t_env.input_count - 1);
domain_iter_count *= arg_multiplier;
// If we will be running tests against MPFR, we don't need to test as much against musl.
// However, there are some platforms where we have to test against musl since MPFR can't be
// built.
if t_env.mp_tests_enabled && ctx.basis == CheckBasis::Musl {
domain_iter_count /= 100;
}
// Run fewer random tests than domain tests.
let random_iter_count = domain_iter_count / 100;
let mut total_iterations = match gen_kind {
GeneratorKind::Domain => domain_iter_count,
GeneratorKind::Random => random_iter_count,
};
if cfg!(optimizations_enabled) {
// Always run at least 10,000 tests.
total_iterations = total_iterations.max(10_000);
} else {
// Without optimizations, just run a quick check regardless of other parameters.
total_iterations = 800;
}
// Adjust for the number of inputs
let ntests = match t_env.input_count {
1 => total_iterations,
2 => (total_iterations as f64).sqrt().ceil() as u64,
3 => (total_iterations as f64).cbrt().ceil() as u64,
_ => panic!("test has more than three arguments"),
};
let total = ntests.pow(t_env.input_count.try_into().unwrap());
test_log(&format!(
"{gen_kind:?} {basis:?} {fn_ident} arg {arg}/{args}: {ntests} iterations \
({total} total)",
basis = ctx.basis,
fn_ident = ctx.fn_ident,
arg = argnum + 1,
args = t_env.input_count,
));
ntests
}
/// For domain tests, limit how many asymptotes or specified check points we test.
pub fn check_point_count(ctx: &CheckCtx) -> usize {
let t_env = TestEnv::from_env(ctx);
if t_env.slow_platform || !cfg!(optimizations_enabled) { 4 } else { 10 }
}
/// When validating points of interest (e.g. asymptotes, inflection points, extremes), also check
/// this many surrounding values.
pub fn check_near_count(_ctx: &CheckCtx) -> u64 {
if cfg!(optimizations_enabled) { 100 } else { 10 }
}
/// Check whether extensive actions should be run or skipped.
#[expect(dead_code, reason = "extensive tests have not yet been added")]
pub fn skip_extensive_test(ctx: &CheckCtx) -> bool {
let t_env = TestEnv::from_env(ctx);
!t_env.should_run_extensive
}