Skip to content

Preserve subnormal float values when casting to bool #4224

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
reckylurker wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Preserve subnormal float values when casting to bool #4224

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c071987
metal: fix subnormal float conversion to bool in copy and reduction k…
reckylurker Aug 13, 2026
dd1b57d
refactor(metal): rename mlx_cast to cast_to and remove half specializ…
reckylurker Aug 15, 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
46 changes: 23 additions & 23 deletions mlx/backend/metal/kernels/copy.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,11 +9,11 @@ template <typename T, typename U, int N = WorkPerThread<U>::n>
index *= N;
if (N > 1 && index + N > size) {
for (int i = 0; index + i < size; ++i) {
dst[index + i] = static_cast<U>(src[0]);
dst[index + i] = mlx_cast<U>(src[0]);
}
} else {
for (int i = 0; i < N; ++i) {
dst[index + i] = static_cast<U>(src[0]);
dst[index + i] = mlx_cast<U>(src[0]);
}
}
}
Expand All @@ -27,11 +27,11 @@ template <typename T, typename U, int N = WorkPerThread<U>::n>
index *= N;
if (N > 1 && index + N > size) {
for (int i = 0; index + i < size; ++i) {
dst[index + i] = static_cast<U>(src[index + i]);
dst[index + i] = mlx_cast<U>(src[index + i]);
}
} else {
for (int i = 0; i < N; ++i) {
dst[index + i] = static_cast<U>(src[index + i]);
dst[index + i] = mlx_cast<U>(src[index + i]);
}
}
}
Expand All @@ -46,11 +46,11 @@ template <typename T, typename U, int N = WorkPerThread<U>::n>
int64_t offset = N * (index.x + grid_dim.x * int64_t(index.y));
if (N > 1 && offset + N > size) {
for (int i = 0; offset + i < size; ++i) {
dst[offset + i] = static_cast<U>(src[0]);
dst[offset + i] = mlx_cast<U>(src[0]);
}
} else {
for (int i = 0; i < N; ++i) {
dst[offset + i] = static_cast<U>(src[0]);
dst[offset + i] = mlx_cast<U>(src[0]);
}
}
}
Expand All @@ -65,11 +65,11 @@ template <typename T, typename U, int N = WorkPerThread<U>::n>
int64_t offset = N * (index.x + grid_dim.x * int64_t(index.y));
if (N > 1 && offset + N > size) {
for (int i = 0; offset + i < size; ++i) {
dst[offset + i] = static_cast<U>(src[offset + i]);
dst[offset + i] = mlx_cast<U>(src[offset + i]);
}
} else {
for (int i = 0; i < N; ++i) {
dst[offset + i] = static_cast<U>(src[offset + i]);
dst[offset + i] = mlx_cast<U>(src[offset + i]);
}
}
}
Expand All @@ -81,7 +81,7 @@ template <typename T, typename U, typename IdxT = int64_t>
constant const int64_t& src_stride [[buffer(3)]],
uint index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_1<IdxT>(index, src_stride);
dst[index] = static_cast<U>(src[src_idx]);
dst[index] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -93,7 +93,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint2 grid_dim [[threads_per_grid]]) {
auto src_idx = elem_to_loc_2<IdxT>(index, src_strides);
IdxT dst_idx = index.x + IdxT(grid_dim.x) * index.y;
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -106,7 +106,7 @@ template <typename T, typename U, typename IdxT = int64_t>
auto src_idx = elem_to_loc_3<IdxT>(index, src_strides);
IdxT dst_idx =
index.x + IdxT(grid_dim.x) * (index.y + IdxT(grid_dim.y) * index.z);
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, int N = 1, typename IdxT = int64_t>
Expand All @@ -123,14 +123,14 @@ template <typename T, typename U, int N = 1, typename IdxT = int64_t>
if (N == 1) {
IdxT dst_idx =
index.x + grid_dim.x * (index.y + IdxT(grid_dim.y) * index.z);
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
return;
}
auto xshape = src_shape[ndim - 1];
IdxT dst_idx = N * index.x + xshape * (index.y + IdxT(grid_dim.y) * index.z);
auto src_xstride = src_strides[ndim - 1];
for (int i = 0; i < N && (int(N * index.x) + i) < xshape; ++i) {
dst[dst_idx + i] = static_cast<U>(src[src_idx]);
dst[dst_idx + i] = mlx_cast<U>(src[src_idx]);
src_idx += src_xstride;
}
}
Expand All @@ -144,7 +144,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_1<IdxT>(index, src_stride);
auto dst_idx = elem_to_loc_1<IdxT>(index, dst_stride);
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -156,7 +156,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint2 index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_2<IdxT>(index, src_strides);
auto dst_idx = elem_to_loc_2<IdxT>(index, dst_strides);
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -168,7 +168,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint3 index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_3<IdxT>(index, src_strides);
auto dst_idx = elem_to_loc_3<IdxT>(index, dst_strides);
dst[dst_idx] = static_cast<U>(src[src_idx]);
dst[dst_idx] = mlx_cast<U>(src[src_idx]);
}

template <typename T, typename U, int N = 1, typename IdxT = int64_t>
Expand All @@ -187,14 +187,14 @@ template <typename T, typename U, int N = 1, typename IdxT = int64_t>
dst_strides,
ndim);
if (N == 1) {
dst[idx.y] = static_cast<U>(src[idx.x]);
dst[idx.y] = mlx_cast<U>(src[idx.x]);
return;
}
IdxT src_xstride = src_strides[ndim - 1];
IdxT dst_xstride = dst_strides[ndim - 1];
auto xshape = src_shape[ndim - 1];
for (int i = 0; i < N && (int(N * index.x) + i) < xshape; ++i) {
dst[idx.y] = static_cast<U>(src[idx.x]);
dst[idx.y] = mlx_cast<U>(src[idx.x]);
idx.x += src_xstride;
idx.y += dst_xstride;
}
Expand All @@ -211,7 +211,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_1<IdxT>(index, src_stride);
auto dst_idx = elem_to_loc_1<IdxT>(index, dst_stride);
dst[dst_idx + dst_offset] = src[src_idx + src_offset];
dst[dst_idx + dst_offset] = mlx_cast<U>(src[src_idx + src_offset]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -225,7 +225,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint2 index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_2<IdxT>(index, src_strides);
auto dst_idx = elem_to_loc_2<IdxT>(index, dst_strides);
dst[dst_idx + dst_offset] = src[src_idx + src_offset];
dst[dst_idx + dst_offset] = mlx_cast<U>(src[src_idx + src_offset]);
}

template <typename T, typename U, typename IdxT = int64_t>
Expand All @@ -239,7 +239,7 @@ template <typename T, typename U, typename IdxT = int64_t>
uint3 index [[thread_position_in_grid]]) {
auto src_idx = elem_to_loc_3<IdxT>(index, src_strides);
auto dst_idx = elem_to_loc_3<IdxT>(index, dst_strides);
dst[dst_idx + dst_offset] = src[src_idx + src_offset];
dst[dst_idx + dst_offset] = mlx_cast<U>(src[src_idx + src_offset]);
}

template <typename T, typename U, int N = 1, typename IdxT = int64_t>
Expand All @@ -262,14 +262,14 @@ template <typename T, typename U, int N = 1, typename IdxT = int64_t>
dst_strides,
ndim);
if (N == 1) {
dst[idx.y] = src[idx.x];
dst[idx.y] = mlx_cast<U>(src[idx.x]);
return;
}
IdxT src_xstride = src_strides[ndim - 1];
IdxT dst_xstride = dst_strides[ndim - 1];
auto xshape = src_shape[ndim - 1];
for (int i = 0; i < N && (int(N * index.x) + i) < xshape; ++i) {
dst[idx.y] = src[idx.x];
dst[idx.y] = mlx_cast<U>(src[idx.x]);
idx.x += src_xstride;
idx.y += dst_xstride;
}
Expand Down
4 changes: 2 additions & 2 deletions mlx/backend/metal/kernels/reduction/reduce_all.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,13 +37,13 @@ template <

for (IdxT b = 0; b < blocks; b++) {
for (int i = 0; i < N_READS; i++) {
total = op(static_cast<U>(in[i]), total);
total = op(mlx_cast<U>(in[i]), total);
}
in += lsize.x * N_READS;
}
if (extra > 0) {
for (int i = 0; i < extra; i++) {
total = op(static_cast<U>(in[i]), total);
total = op(mlx_cast<U>(in[i]), total);
}
}

Expand Down
14 changes: 7 additions & 7 deletions mlx/backend/metal/kernels/reduction/reduce_col.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,13 +43,13 @@ template <typename T, typename U, typename Op, typename IdxT, int NDIMS>
row = in + loop.location();
if (safe) {
for (int i = 0; i < n_reads; i++) {
totals[i] = op(static_cast<U>(row[i]), totals[i]);
totals[i] = op(mlx_cast<U>(row[i]), totals[i]);
}
} else {
U vals[n_reads];
for (int i = 0; i < n_reads; i++) {
vals[i] =
(column + i < reduction_stride) ? static_cast<U>(row[i]) : op.init;
(column + i < reduction_stride) ? mlx_cast<U>(row[i]) : op.init;
}
for (int i = 0; i < n_reads; i++) {
totals[i] = op(vals[i], totals[i]);
Expand Down Expand Up @@ -125,7 +125,7 @@ template <typename T, typename U, typename Op, typename IdxT, int NDIMS>
for (IdxT r = gid.z * lsize.y + lid.y; r < total_rows;
r += lsize.y * gsize.z) {
row = in + loop.location();
total = op(static_cast<U>(*row), total);
total = op(mlx_cast<U>(*row), total);
loop.next(lsize.y * gsize.z, reduce_shape, reduce_strides);
}

Expand Down Expand Up @@ -207,13 +207,13 @@ template <

if (safe) {
for (int i = 0; i < n_reads; i++) {
totals[i] = op(static_cast<U>(row[i]), totals[i]);
totals[i] = op(mlx_cast<U>(row[i]), totals[i]);
}
} else {
U vals[n_reads];
for (int i = 0; i < n_reads; i++) {
vals[i] =
(column + i < reduction_stride) ? static_cast<U>(row[i]) : op.init;
(column + i < reduction_stride) ? mlx_cast<U>(row[i]) : op.init;
}
for (int i = 0; i < n_reads; i++) {
totals[i] = op(vals[i], totals[i]);
Expand Down Expand Up @@ -352,13 +352,13 @@ template <

if (safe) {
for (int i = 0; i < n_reads; i++) {
totals[i] = op(static_cast<U>(row[i]), totals[i]);
totals[i] = op(mlx_cast<U>(row[i]), totals[i]);
}
} else {
U vals[n_reads];
for (int i = 0; i < n_reads; i++) {
vals[i] =
(column + i < reduction_stride) ? static_cast<U>(row[i]) : op.init;
(column + i < reduction_stride) ? mlx_cast<U>(row[i]) : op.init;
}
for (int i = 0; i < n_reads; i++) {
totals[i] = op(vals[i], totals[i]);
Expand Down
6 changes: 3 additions & 3 deletions mlx/backend/metal/kernels/reduction/reduce_row.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ METAL_FUNC void per_thread_row_reduce(
for (int i = 0; i < blocks; i++) {
for (int j = 0; j < N_WRITES; j++) {
for (int i = 0; i < N_READS; i++) {
totals[j] = op(static_cast<U>(inputs[j][i]), totals[j]);
totals[j] = op(mlx_cast<U>(inputs[j][i]), totals[j]);
}

inputs[j] += lsize_x * N_READS;
Expand All @@ -46,13 +46,13 @@ METAL_FUNC void per_thread_row_reduce(
if (index + N_READS <= extra) {
for (int j = 0; j < N_WRITES; j++) {
for (int i = 0; i < N_READS; i++) {
totals[j] = op(static_cast<U>(inputs[j][i]), totals[j]);
totals[j] = op(mlx_cast<U>(inputs[j][i]), totals[j]);
}
}
} else {
for (int j = 0; j < N_WRITES; j++) {
for (int i = 0; index + i < extra; i++) {
totals[j] = op(static_cast<U>(inputs[j][i]), totals[j]);
totals[j] = op(mlx_cast<U>(inputs[j][i]), totals[j]);
}
}
}
Expand Down
29 changes: 29 additions & 0 deletions mlx/backend/metal/kernels/utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -446,3 +446,32 @@ template <typename T, typename U>
struct ConditionalType<true, T, U> {
using type = T;
};

///////////////////////////////////////////////////////////////////////////////
// Type casting utils
///////////////////////////////////////////////////////////////////////////////

template <typename U, typename T>
inline U mlx_cast(T val) {

@zcbenz zcbenz Aug 15, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In cuda backend there is a similar utility named cast_to, I think we can use the same name here.

return static_cast<U>(val);
}

template <>
inline bool mlx_cast<bool, float>(float val) {
return (as_type<uint32_t>(val) & 0x7FFFFFFF) != 0;
}

template <>
inline bool mlx_cast<bool, bfloat16_t>(bfloat16_t val) {
return (as_type<uint16_t>(val) & 0x7FFF) != 0;
}

template <>
inline bool mlx_cast<bool, half>(half val) {
return (as_type<uint16_t>(val) & 0x7FFF) != 0;

@zcbenz zcbenz Aug 15, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

float16 does not seem to need this, should add a test.

@reckylurker reckylurker Aug 15, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, looks like half values are not affected in MSL. Thanks for the suggestion.

}

template <>
inline bool mlx_cast<bool, complex64_t>(complex64_t val) {
return mlx_cast<bool, float>(val.real) || mlx_cast<bool, float>(val.imag);
}
11 changes: 11 additions & 0 deletions python/tests/test_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1299,6 +1299,17 @@ def test_any(self):
self.assertEqual(mx.any(a, axis=0).tolist(), [True, False])
self.assertEqual(mx.any(a, axis=1).tolist(), [True, False])

def test_subnormal_bool_cast(self):
f32_sub = mx.array(np.array([0x00000001], dtype=np.uint32)).view(mx.float32)
bf16_sub = mx.array(np.array([0x0001], dtype=np.uint16)).view(mx.bfloat16)

self.assertTrue(f32_sub.astype(mx.bool_).item())
self.assertTrue(bf16_sub.astype(mx.bool_).item())
self.assertTrue(mx.any(f32_sub).item())
self.assertTrue(mx.any(bf16_sub).item())
self.assertTrue(mx.all(f32_sub).item())
self.assertTrue(mx.all(bf16_sub).item())

def test_stop_gradient(self):
def func(x):
return mx.sum(2 * x + mx.stop_gradient(3 * x))
Expand Down
Loading