Skip to content

Refactor: Extract set_output Function from utils.h #4467

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
JasonHonKL wants to merge 2 commits into
base: main
Choose a base branch
from
+31 −70
Open

Refactor: Extract set_output Function from utils.h #4467

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
62e8b8d
[metal] Refactor ensure innermost contiguous function
JasonHonKL Sep 5, 2026
70c0a1b
remove unnecessary comment
JasonHonKL Sep 5, 2026
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
52 changes: 2 additions & 50 deletions mlx/backend/metal/normalization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,31 +22,7 @@ void RMSNorm::eval_gpu(
auto& out = outputs[0];

// Make sure that the last dimension is contiguous
auto set_output = [&s, &out](const array& x) {
bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
if (no_copy && x.ndim() > 1) {
auto s = x.strides()[x.ndim() - 2];
no_copy &= (s == 0 || s == x.shape().back() || x.shape(-2) == 1);
}
if (no_copy) {
if (x.is_donatable()) {
out.copy_shared_buffer(x);
} else {
out.set_data(
allocator::malloc(x.data_size() * x.itemsize()),
x.data_size(),
x.strides(),
x.flags());
}
return x;
} else {
array x_copy = contiguous_copy_gpu(x, s);
out.copy_shared_buffer(x_copy);
return x_copy;
}
};

const array x = set_output(inputs[0]);
const array x = ensure_innermost_contiguous(out, inputs[0], s, true);
const array& w = inputs[1];

auto axis_size = static_cast<uint32_t>(x.shape().back());
Expand Down Expand Up @@ -224,31 +200,7 @@ void LayerNorm::eval_gpu(
auto& out = outputs[0];

// Make sure that the last dimension is contiguous
auto set_output = [&s, &out](const array& x) {
bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
if (no_copy && x.ndim() > 1) {
auto s = x.strides()[x.ndim() - 2];
no_copy &= (s == 0 || s == x.shape().back() || x.shape(-2) == 1);
}
if (no_copy) {
if (x.is_donatable()) {
out.copy_shared_buffer(x);
} else {
out.set_data(
allocator::malloc(x.data_size() * x.itemsize()),
x.data_size(),
x.strides(),
x.flags());
}
return x;
} else {
array x_copy = contiguous_copy_gpu(x, s);
out.copy_shared_buffer(x_copy);
return x_copy;
}
};

const array x = set_output(inputs[0]);
const array x = ensure_innermost_contiguous(out, inputs[0], s, true);
const array& w = inputs[1];
const array& b = inputs[2];

Expand Down
21 changes: 1 addition & 20 deletions mlx/backend/metal/softmax.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,26 +22,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
auto& d = metal::device(s.device);

// Make sure that the last dimension is contiguous
auto set_output = [&s, &out](const array& x) {
if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
if (x.is_donatable()) {
out.copy_shared_buffer(x);
} else {
out.set_data(
allocator::malloc(x.data_size() * x.itemsize()),
x.data_size(),
x.strides(),
x.flags());
}
return x;
} else {
array x_copy = contiguous_copy_gpu(x, s);
out.copy_shared_buffer(x_copy);
return x_copy;
}
};

const array in = set_output(inputs[0]);
const array in = ensure_innermost_contiguous(out, inputs[0], s);

int axis_size = in.shape().back();
int n_rows = in.data_size() / axis_size;
Expand Down
28 changes: 28 additions & 0 deletions mlx/backend/metal/utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
#include <type_traits>

#include "mlx/array.h"
#include "mlx/backend/gpu/copy.h"
#include "mlx/backend/metal/device.h"
#include "mlx/primitives.h"

Expand Down Expand Up @@ -81,6 +82,33 @@ inline size_t ceildiv(size_t n, size_t m) {
return (n + m - 1) / m;
}

inline array ensure_innermost_contiguous(
array& out,
const array& x,
const Stream& s,
bool relax_prev_stride = false) {
bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
if (relax_prev_stride && no_copy && x.ndim() > 1) {
auto stride = x.strides()[x.ndim() - 2];
no_copy &= (stride == 0 || stride == x.shape().back() || x.shape(-2) == 1);
}
if (no_copy) {
if (x.is_donatable()) {
out.copy_shared_buffer(x);
} else {
out.set_data(
allocator::malloc(x.data_size() * x.itemsize()),
x.data_size(),
x.strides(),
x.flags());
}
return x;
}
array x_copy = contiguous_copy_gpu(x, s);
out.copy_shared_buffer(x_copy);
return x_copy;
}

inline void check_kernel_threadgroup_size(
const MTL::ComputePipelineState* kernel,
MTL::Size group_dims,
Expand Down
Loading