mod kernel;
pub use kernel::{matmul, matmul_accumulate};
use crate::backend_ops::BackendOps;
use coeus_core::Scalar;
use coeus_tensor::Tensor;
#[inline]
pub fn bmm<T: Scalar, B: BackendOps<T> + Default>(
a: &Tensor<T, B>,
b: &Tensor<T, B>,
backend: &B,
) -> Result<Tensor<T, B>, B::Error> {
assert_eq!(a.ndim(), 3, "bmm: a must be 3-D, got {}-D", a.ndim());
assert_eq!(b.ndim(), 3, "bmm: b must be 3-D, got {}-D", b.ndim());
let (batch, _m, _k) = (a.shape()[0], a.shape()[1], a.shape()[2]);
let (_b2, _k2, _n) = (b.shape()[0], b.shape()[1], b.shape()[2]);
assert_eq!(batch, b.shape()[0], "bmm: batch mismatch");
assert_eq!(a.shape()[2], b.shape()[1], "bmm: inner dim mismatch");
kernel::matmul(a, b, backend)
}
#[inline]
pub fn outer<T: Scalar, B: BackendOps<T> + Default>(
a: &Tensor<T, B>,
b: &Tensor<T, B>,
backend: &B,
) -> Result<Tensor<T, B>, B::Error> {
assert_eq!(a.ndim(), 1, "outer: a must be 1-D, got {}-D", a.ndim());
assert_eq!(b.ndim(), 1, "outer: b must be 1-D, got {}-D", b.ndim());
let m = a.shape()[0];
let n = b.shape()[0];
let a_col = a.clone().reshape([m, 1]);
let b_row = b.clone().reshape([1, n]);
kernel::matmul(&a_col, &b_row, backend)
}
#[cfg(test)]
mod tests {
use super::*;
use coeus_core::SequentialBackend;
#[test]
fn bmm_matches_manual_batch_matmul() {
let backend = SequentialBackend::new();
let a = Tensor::<f32, SequentialBackend>::from_slice(
[2, 2, 3],
&[0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 1.0, 0.0, 2.0, 0.0, 1.0, 3.0],
)
.expect("invariant: test backend operation succeeds");
let b = Tensor::<f32, SequentialBackend>::from_slice(
[2, 3, 2],
&[0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 1.0, 2.0, 0.0, 1.0, 3.0, 4.0],
)
.expect("invariant: test backend operation succeeds");
let out = bmm(&a, &b, &backend).expect("valid batched matmul test shapes");
assert_eq!(out.shape(), &[2, 2, 2]);
assert_eq!(
out.as_slice(),
&[10.0, 13.0, 28.0, 40.0, 7.0, 10.0, 9.0, 13.0]
);
}
#[test]
fn outer_matches_pairwise_products() {
let backend = SequentialBackend::new();
let a = Tensor::<f32, SequentialBackend>::from_slice([3], &[1.0, 2.0, 3.0])
.expect("invariant: test backend operation succeeds");
let b = Tensor::<f32, SequentialBackend>::from_slice([2], &[4.0, 5.0])
.expect("invariant: test backend operation succeeds");
let out = outer(&a, &b, &backend).expect("valid outer-product test shapes");
assert_eq!(out.shape(), &[3, 2]);
assert_eq!(out.as_slice(), &[4.0, 5.0, 8.0, 10.0, 12.0, 15.0]);
}
}