Document what compiler is doing for diffxi

This commit is contained in:
Magnus Ulimoen 2021-03-25 22:39:00 +01:00
parent 76f5291131
commit a33e1d37ba
2 changed files with 42 additions and 15 deletions

View File

@ -102,7 +102,10 @@ pub(crate) fn diff_op_1d_slice<const M: usize, const N: usize, const D: usize>(
prev: &[Float],
fut: &mut [Float],
) {
use std::convert::TryInto;
#[inline(never)]
/// This prevents code bloat, both start and end block gives
/// a matrix multiplication with the same matrix sizes
fn dedup_matmul<const M: usize, const N: usize>(
c: &mut ColVector<Float, M>,
a: &Matrix<Float, M, N>,
@ -129,7 +132,6 @@ pub(crate) fn diff_op_1d_slice<const M: usize, const N: usize, const D: usize>(
let (futb1, fut) = fut.split_at_mut(M);
let (fut, futb2) = fut.split_at_mut(nx - 2 * M);
use std::convert::TryInto;
{
let prev = ColVector::<_, N>::map_to_col(prev.array_windows::<N>().next().unwrap());
let fut = ColVector::<_, M>::map_to_col_mut(futb1.try_into().unwrap());
@ -138,10 +140,29 @@ pub(crate) fn diff_op_1d_slice<const M: usize, const N: usize, const D: usize>(
*fut *= idx;
}
{
// The window needs to be aligned to the diagonal elements,
// based on the block size
let window_elems_to_skip = M - ((D - 1) / 2);
// The compiler is pretty clever right here. It is able to
// inline the entire computation into a very tight loop.
// It seems the "optimal" way is to broadcast each element of diag
// into separate SIMD vectors.
// Then an unroll of SIMD::lanes items from fut and prev:
// f[0] = diag[0] * in[0] + diag[1] * in[1] + diag[2] * in[3] + ...
// f[1] = diag[0] * in[1] + diag[1] * in[2] + diag[2] * in[4] + ...
// \-- fma's along the vertical axis, then horizontal
//
// The resulting inner loop performs:
// one mul, D-1 fma,
// one mul (by idx),
// one store
// two integer adds (input and output offsets)
// one test + jump
//
// The compiler is clever enough to combine two such unrollings and merge
// these computations to prevent stalling
for (window, f) in prev[window_elems_to_skip..]
.array_windows::<D>()
.zip(fut.array_chunks_mut::<1>())
@ -152,6 +173,7 @@ pub(crate) fn diff_op_1d_slice<const M: usize, const N: usize, const D: usize>(
fut.matmul_float_into(&matrix.diag, prev);
*fut *= idx;
}
}
{
let prev = prev.array_windows::<N>().next_back().unwrap();

View File

@ -123,10 +123,15 @@ impl<const M: usize, const P: usize> Matrix<Float, M, P> {
) {
for i in 0..M {
for j in 0..P {
let mut t = 0.0;
for k in 0..N {
// Slightly cheaper to do first computation separately
// rather than store zero and issue all ops as fma
let mut t = if N == 0 {
0.0
} else {
lhs[(i, 0)] * rhs[(0, j)]
};
for k in 1..N {
t = Float::mul_add(lhs[(i, k)], rhs[(k, j)], t);
// t = t + lhs[(i, k)] * rhs[(k, j)];
}
self[(i, j)] = t;
}