use crate::config::{DeviceSolveConfig, TimingMode};
use crate::problem::ProjectedFirstOrderProblem;
use crate::solve::ProjectedFirstOrderWorkspace;
use crate::workspace::WorkspaceFor;
use loeres::SolverError;
use loeres_backend_static::array::FixedVector;
pub(super) struct Quadratic<const N: usize> {
pub(super) target: FixedVector<f64, N>,
pub(super) lo: FixedVector<f64, N>,
pub(super) hi: FixedVector<f64, N>,
pub(super) alpha: f64,
}
impl<const N: usize> ProjectedFirstOrderProblem<f64, N> for Quadratic<N> {
type Bounds = FixedVector<f64, N>;
fn validate_boundary(&self) -> Result<(), SolverError> {
for (l, h) in self.lo.as_slice().iter().zip(self.hi.as_slice()) {
if !l.is_finite() || !h.is_finite() {
return Err(SolverError::NonFiniteInput);
}
if *l > *h {
return Err(SolverError::InvalidInput);
}
}
Ok(())
}
fn lower_bound(&self) -> &FixedVector<f64, N> {
&self.lo
}
fn upper_bound(&self) -> &FixedVector<f64, N> {
&self.hi
}
fn step_scale(&self) -> f64 {
self.alpha
}
fn gradient_at(
&self,
x: &FixedVector<f64, N>,
grad: &mut FixedVector<f64, N>,
) -> Result<(), SolverError> {
for ((g, &xi), &ti) in grad
.as_mut_slice()
.iter_mut()
.zip(x.as_slice())
.zip(self.target.as_slice())
{
*g = xi - ti;
}
Ok(())
}
fn objective_at(&self, x: &FixedVector<f64, N>) -> Result<f64, SolverError> {
let mut acc = 0.0;
for (&xi, &ti) in x.as_slice().iter().zip(self.target.as_slice()) {
let d = xi - ti;
acc += 0.5 * d * d;
}
Ok(acc)
}
}
impl<const N: usize> WorkspaceFor<Quadratic<N>> for Quadratic<N> {
type Workspace = ProjectedFirstOrderWorkspace<f64, N>;
fn required_workspace_bytes() -> usize {
core::mem::size_of::<ProjectedFirstOrderWorkspace<f64, N>>()
}
}
pub(super) fn quad2() -> Quadratic<2> {
Quadratic {
target: FixedVector::from_array([0.5, -0.5]),
lo: FixedVector::from_array([-1.0, -1.0]),
hi: FixedVector::from_array([1.0, 1.0]),
alpha: 0.5,
}
}
pub(super) fn workspace<const N: usize>() -> ProjectedFirstOrderWorkspace<f64, N> {
ProjectedFirstOrderWorkspace::new(FixedVector::from_array([0.0; N]))
}
pub(super) fn config(
max_iterations: u32,
tolerance: f64,
timing_mode: TimingMode,
) -> DeviceSolveConfig<f64> {
DeviceSolveConfig {
max_iterations,
tolerance,
timing_mode,
}
}
pub(super) struct Mismatch {
pub(super) lo: FixedVector<f64, 2>,
pub(super) hi: FixedVector<f64, 2>,
}
impl ProjectedFirstOrderProblem<f64, 3> for Mismatch {
type Bounds = FixedVector<f64, 2>;
fn validate_boundary(&self) -> Result<(), SolverError> {
Ok(())
}
fn lower_bound(&self) -> &FixedVector<f64, 2> {
&self.lo
}
fn upper_bound(&self) -> &FixedVector<f64, 2> {
&self.hi
}
fn step_scale(&self) -> f64 {
0.5
}
fn gradient_at(
&self,
_x: &FixedVector<f64, 3>,
grad: &mut FixedVector<f64, 3>,
) -> Result<(), SolverError> {
for g in grad.as_mut_slice().iter_mut() {
*g = 0.0;
}
Ok(())
}
fn objective_at(&self, _x: &FixedVector<f64, 3>) -> Result<f64, SolverError> {
Ok(0.0)
}
}
pub(super) fn quad2_with_alpha(alpha: f64) -> Quadratic<2> {
Quadratic {
target: FixedVector::from_array([0.5, -0.5]),
lo: FixedVector::from_array([-1.0, -1.0]),
hi: FixedVector::from_array([1.0, 1.0]),
alpha,
}
}
pub(super) struct NanGradient {
pub(super) lo: FixedVector<f64, 2>,
pub(super) hi: FixedVector<f64, 2>,
}
impl ProjectedFirstOrderProblem<f64, 2> for NanGradient {
type Bounds = FixedVector<f64, 2>;
fn validate_boundary(&self) -> Result<(), SolverError> {
Ok(())
}
fn lower_bound(&self) -> &FixedVector<f64, 2> {
&self.lo
}
fn upper_bound(&self) -> &FixedVector<f64, 2> {
&self.hi
}
fn step_scale(&self) -> f64 {
0.5
}
fn gradient_at(
&self,
_x: &FixedVector<f64, 2>,
grad: &mut FixedVector<f64, 2>,
) -> Result<(), SolverError> {
for g in grad.as_mut_slice().iter_mut() {
*g = f64::NAN;
}
Ok(())
}
fn objective_at(&self, _x: &FixedVector<f64, 2>) -> Result<f64, SolverError> {
Ok(0.0)
}
}
pub(super) struct LaxNanBounds {
pub(super) lo: FixedVector<f64, 2>,
pub(super) hi: FixedVector<f64, 2>,
}
impl ProjectedFirstOrderProblem<f64, 2> for LaxNanBounds {
type Bounds = FixedVector<f64, 2>;
fn validate_boundary(&self) -> Result<(), SolverError> {
Ok(())
}
fn lower_bound(&self) -> &FixedVector<f64, 2> {
&self.lo
}
fn upper_bound(&self) -> &FixedVector<f64, 2> {
&self.hi
}
fn step_scale(&self) -> f64 {
0.5
}
fn gradient_at(
&self,
x: &FixedVector<f64, 2>,
grad: &mut FixedVector<f64, 2>,
) -> Result<(), SolverError> {
for (g, &xi) in grad.as_mut_slice().iter_mut().zip(x.as_slice()) {
*g = xi;
}
Ok(())
}
fn objective_at(&self, _x: &FixedVector<f64, 2>) -> Result<f64, SolverError> {
Ok(0.0)
}
}