zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-06-25 01:41:17 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix multi-block sort stride management (#1169)

Browse Source 
* Fix multi-block sort stride management

* Add seed to tests
This commit is contained in:
Jagrit Digani 2024-05-31 11:10:54 -07:00 committed by GitHub
parent 9f0df51f8d
commit 76b6cece46
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 33 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
mlx/backend/metal/sort.cpp
View File

@ -222,22 +222,24 @@ void multi_block_sort(
// Copy outputs with appropriate strides // Copy outputs with appropriate strides
array strided_out_arr = argsort ? dev_idxs_out : dev_vals_out; array strided_out_arr = argsort ? dev_idxs_out : dev_vals_out;
if (axis == strided_out_arr.ndim() - 1) { if (axis == in.ndim() - 1) {
copy_gpu_inplace(strided_out_arr, out, CopyType::Vector, s); copy_gpu_inplace(strided_out_arr, out, CopyType::Vector, s);
} else { } else {
std::vector<int> strided_out_shape = strided_out_arr.shape(); std::vector<int> strided_out_shape = in.shape();
std::vector<size_t> strided_out_str = strided_out_arr.strides();
int out_axis_shape = strided_out_shape[axis]; int out_axis_shape = strided_out_shape[axis];
int out_axis_str = strided_out_str[axis];
strided_out_shape.erase(strided_out_shape.begin() + axis); strided_out_shape.erase(strided_out_shape.begin() + axis);
strided_out_str.erase(strided_out_str.begin() + axis);
strided_out_shape.push_back(out_axis_shape); strided_out_shape.push_back(out_axis_shape);
strided_out_str.push_back(out_axis_str);
array strided_out_slice(strided_out_shape, out.dtype(), nullptr, {}); std::vector<size_t> strided_out_str(in.ndim(), 1);
for (int i = in.ndim() - 2; i >= 0; --i) {
strided_out_str[i] = strided_out_str[i + 1] * strided_out_shape[i + 1];
}
strided_out_str.erase(strided_out_str.end() - 1);
strided_out_str.insert(strided_out_str.begin() + axis, 1);
array strided_out_slice(in.shape(), out.dtype(), nullptr, {});
strided_out_slice.copy_shared_buffer( strided_out_slice.copy_shared_buffer(
strided_out_arr, strided_out_arr,
strided_out_str, strided_out_str,

22
python/tests/test_ops.py
View File

@ -1754,6 +1754,9 @@ class TestOps(mlx_tests.MLXTestCase):
self.assertTrue(np.array_equal(d_np, d_mx)) self.assertTrue(np.array_equal(d_np, d_mx))
self.assertEqual(c_mx.dtype, mx.uint32) self.assertEqual(c_mx.dtype, mx.uint32)
# Set random seed
np.random.seed(0)
# Test multi-block sort # Test multi-block sort
a_np = np.random.normal(size=(32769,)).astype(np.float32) a_np = np.random.normal(size=(32769,)).astype(np.float32)
a_mx = mx.array(a_np) a_mx = mx.array(a_np)
@ -1764,6 +1767,25 @@ class TestOps(mlx_tests.MLXTestCase):
self.assertTrue(np.array_equal(b_np, b_mx)) self.assertTrue(np.array_equal(b_np, b_mx))
self.assertEqual(b_mx.dtype, a_mx.dtype) self.assertEqual(b_mx.dtype, a_mx.dtype)
# Test multi-dum multi-block sort
a_np = np.random.normal(size=(2, 4, 32769)).astype(np.float32)
a_mx = mx.array(a_np)
b_np = np.sort(a_np, axis=-1)
b_mx = mx.sort(a_mx, axis=-1)
self.assertTrue(np.array_equal(b_np, b_mx))
self.assertEqual(b_mx.dtype, a_mx.dtype)
a_np = np.random.normal(size=(2, 32769, 4)).astype(np.float32)
a_mx = mx.array(a_np)
b_np = np.sort(a_np, axis=1)
b_mx = mx.sort(a_mx, axis=1)
self.assertTrue(np.array_equal(b_np, b_mx))
self.assertEqual(b_mx.dtype, a_mx.dtype)
def test_partition(self): def test_partition(self):
shape = (3, 4, 5) shape = (3, 4, 5)
for dtype in ("int32", "float32"): for dtype in ("int32", "float32"):