zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-12-16 01:49:05 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Fix slice data size (#1394)

Browse Source 
* fix slice data size and add tests

* fix contiguous flag

* simplify stride and perform copy for non-contiguous arrays

* fix cpu

* comment
This commit is contained in:
Awni Hannun
2024-09-04 19:10:43 -07:00
committed by GitHub
parent 11371fe251
commit 7cca1727af
12 changed files with 129 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
mlx/backend/common/slicing.cpp
View File

@@ -6,18 +6,16 @@ namespace mlx::core {
std::tuple<bool, int64_t, std::vector<int64_t>> prepare_slice(
const array& in,
std::vector<int>& start_indices,
std::vector<int>& strides) {
const std::vector<int>& start_indices,
const std::vector<int>& strides) {
int64_t data_offset = 0;
bool copy_needed = false;
std::vector<int64_t> inp_strides(in.ndim(), 0);
for (int i = 0; i < in.ndim(); ++i) {
data_offset += start_indices[i] * in.strides()[i];
inp_strides[i] = in.strides()[i] * strides[i];
copy_needed |= strides[i] < 0;
}
return std::make_tuple(copy_needed, data_offset, inp_strides);
}
@@ -25,26 +23,16 @@ void shared_buffer_slice(
const array& in,
const std::vector<size_t>& out_strides,
size_t data_offset,
size_t data_size,
array& out) {
// Compute row/col contiguity
auto [data_size, is_row_contiguous, is_col_contiguous] =
auto [no_bsx_size, is_row_contiguous, is_col_contiguous] =
check_contiguity(out.shape(), out_strides);
auto flags = in.flags();
flags.row_contiguous = is_row_contiguous;
flags.col_contiguous = is_col_contiguous;
if (data_size == 1) {
// Broadcasted scalar array is contiguous.
flags.contiguous = true;
} else if (data_size == in.data_size()) {
// Means we sliced a broadcasted dimension so leave the "no holes" flag
// alone.
} else {
// We sliced something. So either we are row or col contiguous or we
// punched a hole.
flags.contiguous &= flags.row_contiguous || flags.col_contiguous;
}
flags.contiguous = (no_bsx_size == data_size);
out.copy_shared_buffer(in, out_strides, flags, data_size, data_offset);
}