Skip to content

perf: transpose recoder coefficient matrix for cache locality #35

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MavenRain wants to merge 1 commit into
base: main
Choose a base branch
from
Open

perf: transpose recoder coefficient matrix for cache locality #35

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 23 additions & 11 deletions src/full/recoder.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,16 +82,25 @@ impl Recoder {
let piece_byte_len = full_coded_piece_byte_len - num_pieces_coded_together;
let num_pieces_received = data.len() / full_coded_piece_byte_len;

let mut coding_vectors = Vec::with_capacity(num_pieces_received * num_pieces_coded_together);
// Store the coding coefficient matrix in column-major (transposed) order so that
// each column is contiguous in memory. The hot path in `recode_with_buf` computes a
// dot product per column; contiguous storage lets that inner loop scan sequentially,
// which is significantly more cache-friendly than the row-major stride access it
// would otherwise require.
let mut coding_vectors = vec![Gf256::default(); num_pieces_received * num_pieces_coded_together];
let mut coded_pieces = Vec::with_capacity(num_pieces_received * piece_byte_len);

data.chunks_exact(full_coded_piece_byte_len).for_each(|full_coded_piece| {
let coding_vector = &full_coded_piece[..num_pieces_coded_together];
let coded_piece = &full_coded_piece[num_pieces_coded_together..];
data.chunks_exact(full_coded_piece_byte_len)
.enumerate()
.for_each(|(piece_idx, full_coded_piece)| {
let coding_vector = &full_coded_piece[..num_pieces_coded_together];
let coded_piece = &full_coded_piece[num_pieces_coded_together..];

coding_vectors.extend(coding_vector.iter().map(|&symbol| Gf256::new(symbol)));
coded_pieces.extend_from_slice(coded_piece);
});
for (coeff_idx, &symbol) in coding_vector.iter().enumerate() {
coding_vectors[coeff_idx * num_pieces_received + piece_idx] = Gf256::new(symbol);
}
coded_pieces.extend_from_slice(coded_piece);
});

// Pre-allocate internal workspace buffers to avoid repeated allocations during recoding.
let encoder = unsafe { Encoder::without_padding(coded_pieces, num_pieces_received).unwrap_unchecked() };
Expand Down Expand Up @@ -131,13 +140,16 @@ impl Recoder {
rng.fill_bytes(&mut self.random_recoding_vector);

for (coeff_idx, coeff_val) in computed_coding_vector.iter_mut().enumerate().take(self.num_pieces_coded_together) {
// The coding coefficient matrix is stored column-major, so column
// `coeff_idx` occupies a contiguous slice. Zipping it with the
// recoding vector gives a sequential memory access pattern.
let column = &self.coding_vectors[coeff_idx * self.num_pieces_received..(coeff_idx + 1) * self.num_pieces_received];
let computed_coeff = self
.random_recoding_vector
.iter()
.enumerate()
.fold(Gf256::default(), |acc, (recoding_vec_idx, &cur)| {
let row_begins_at = recoding_vec_idx * self.num_pieces_coded_together;
acc + Gf256::new(cur) * self.coding_vectors[row_begins_at + coeff_idx]
.zip(column)
.fold(Gf256::default(), |acc, (&cur, &coeff)| {
acc + Gf256::new(cur) * coeff
});

*coeff_val = computed_coeff.get();
Expand Down