huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-18 11:37:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix the reduce_sum implementation and add some tests.

Browse Source
This commit is contained in:
laurent
2023-06-25 10:55:04 +01:00
parent 0f369dd870
commit ba0693a908
2 changed files with 29 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/cpu_backend.rs
View File

@ -178,18 +178,22 @@ impl CpuStorage {
dst_dims[sum_dim] = 1;
}
let dst_shape = Shape::from(dst_dims);
let sum_dims_and_stride: Vec<_> = src_dims
let mut sum_dims = sum_dims.to_vec();
// Sort the sum_dims as they have to be processed from left to right when converting the
// indexes.
sum_dims.sort();
let sum_dims_and_stride: Vec<_> = sum_dims
.iter()
.enumerate()
.map(|(i, d)| (d, src_dims[i + 1..].iter().product::<usize>()))
.map(|&d| (src_dims[d], src_dims[d + 1..].iter().product::<usize>()))
.collect();
let to_dst_index = |unstr_index: usize| {
// TODO: Optimize, the following does lots of slow division and modulos.
// TODO: Optimize, the following does lots of slow division.
let mut dst_index = unstr_index;
// Set the sum_dims indexes to 0.
for &(dim, stride) in sum_dims_and_stride.iter() {
let index = dst_index / stride % dim;
dst_index -= index * stride;
// The compiler is able to optimize the following in a single divmod op.
let (pre, post) = (dst_index / stride, dst_index % stride);
dst_index = (pre / dim) * stride + post;
}
dst_index
};