use super::{CpuBackend, TensorInfo, linear};
use crate::{Result, invalid};
fn integer(p: &[u8; 64], index: usize) -> i32 {
i32::from_le_bytes(std::array::from_fn(|offset| p[index * 4 + offset]))
}
fn float(p: &[u8; 64], index: usize) -> f32 {
f32::from_bits(integer(p, index) as u32)
}
impl CpuBackend {
#[expect(
clippy::too_many_lines,
reason = "The pinned rotary cache and pair traversal retain their numerical order."
)]
pub(super) fn rope_values(&self, info: &TensorInfo, input: &[f32]) -> Result<Vec<f32>> {
let p = &info.parameters;
let dims = usize::try_from(integer(p, 1))?;
let mode = integer(p, 2);
let offset = usize::try_from(integer(p, 15))?;
if dims == 0
|| !dims.is_multiple_of(2)
|| !offset.is_multiple_of(2)
|| offset
.checked_add(dims)
.is_none_or(|end| end > info.shape[0])
|| !matches!(mode, 0 | 2 | 8 | 24 | 40)
{
return Err(invalid("CPU rotary dimensions, offset or mode differ"));
}
let positions = self.values(
*info
.sources
.get(1)
.ok_or_else(|| invalid("CPU rotary positions missing"))?,
)?;
let factors = info.sources.get(2).map(|&id| self.values(id)).transpose()?;
let freq_base = float(p, 5);
let freq_scale = float(p, 6);
let extension = float(p, 7);
let mut magnitude = float(p, 8);
let theta_scale = freq_base.powf(-2.0 / dims as f32);
let sections: [usize; 4] = std::array::from_fn(|axis| integer(p, 11 + axis) as usize);
let section_count: usize = sections.iter().sum();
let multi = mode & 8 != 0;
let vision = mode == 24;
let rotated = if vision { info.shape[0] } else { dims };
if multi && (section_count == 0 || positions.len() != 4 * info.shape[2])
|| !multi && positions.len() != info.shape[2]
|| vision && (dims * 2 != info.shape[0] || offset != 0)
{
return Err(invalid("CPU rotary position or section extent differs"));
}
if factors
.as_ref()
.is_some_and(|factors| factors.len() < rotated / 2)
{
return Err(invalid("CPU rotary frequency factors are incomplete"));
}
let correction = |beta: f32| {
dims as f32 * (integer(p, 4) as f32 / (beta * 2.0 * std::f32::consts::PI)).ln()
/ (2.0 * freq_base.ln())
};
let low = correction(float(p, 9)).floor().max(0.0);
let high = correction(float(p, 10)).ceil().min(dims as f32 - 1.0);
if extension != 0.0 {
magnitude *= 1.0 + 0.1 * (1.0 / freq_scale).ln();
}
let mut output = input.to_vec();
let mut cache = vec![(0.0_f32, 0.0_f32); rotated / 2];
for sequence in 0..info.shape[2] {
let positions4: [f32; 4] = std::array::from_fn(|axis| {
if multi {
positions[sequence + axis * info.shape[2]]
} else {
positions[sequence]
}
});
let mut theta = positions4;
for (pair, entry) in cache.iter_mut().enumerate() {
let sector = if multi { pair % section_count } else { 0 };
if vision {
let mut boundary = 0;
for axis in 0..4 {
if sector == boundary {
theta[axis] = positions4[axis];
}
boundary += sections[axis];
}
}
let axis = if !multi {
0
} else if mode == 40 {
if sector % 3 == 1 && sector < 3 * sections[1] {
1
} else if sector % 3 == 2 && sector < 3 * sections[2] {
2
} else if sector.is_multiple_of(3) && sector < 3 * sections[0] {
0
} else {
3
}
} else if sector < sections[0] {
0
} else if sector < sections[0] + sections[1] {
1
} else if sector < sections[0] + sections[1] + sections[2] {
2
} else {
3
};
let extrapolated =
theta[axis] / factors.as_ref().map_or(1.0, |factors| factors[pair]);
let interpolated = freq_scale * extrapolated;
let angle = if extension == 0.0 {
interpolated
} else {
let ramp = (1.0
- ((pair as f32 - low) / (high - low).max(0.001)).clamp(0.0, 1.0))
* extension;
interpolated * (1.0 - ramp) + extrapolated * ramp
};
*entry = (
angle.cos() * magnitude,
angle.sin() * magnitude * if info.operation == 49 { -1.0 } else { 1.0 },
);
for angle in &mut theta {
*angle *= theta_scale;
}
}
for batch in 0..info.shape[3] {
for head in 0..info.shape[1] {
let base = linear([offset, head, sequence, batch], info.shape);
for (pair, &(cosine, sine)) in cache.iter().enumerate() {
let first = if mode == 0 { 2 * pair } else { pair };
let second = first
+ if mode == 0 {
1
} else if vision {
dims
} else {
dims / 2
};
let x = input[base + first];
let y = input[base + second];
output[base + first] = x * cosine - y * sine;
output[base + second] = x * sine + y * cosine;
}
}
}
}
Ok(output)
}
}