Merge pull request #15377 from jakevdp:gather-slice

PiperOrigin-RevId: 524920532

jax/_src/numpy/lax_numpy.py

@@ -47,8 +47,10 @@
 
 from jax._src import api_util
 from jax._src import core
+from jax._src import dispatch
 from jax._src import dtypes
 from jax._src.api_util import _ensure_index_tuple
+from jax._src.array import ArrayImpl
 from jax._src.core import ShapedArray, ConcreteArray
 from jax._src.lax.lax import (_array_copy, _sort_lt_comparator,
                               _sort_le_comparator, PrecisionLike)
@@ -58,11 +60,11 @@
 from jax._src.numpy import ufuncs
 from jax._src.numpy import util
 from jax._src.numpy.vectorize import vectorize
+from jax._src.ops import scatter
 from jax._src.typing import Array, ArrayLike, DimSize, DType, DTypeLike, Shape
-from jax._src.util import (unzip2, subvals, safe_zip,
+from jax._src.util import (unzip2, unzip3, subvals, safe_zip,
                            ceil_of_ratio, partition_list,
                            canonicalize_axis as _canonicalize_axis)
-from jax._src.array import ArrayImpl
 
 newaxis = None
 T = TypeVar('T')
@@ -3968,39 +3970,109 @@ def replace(tup, val):
 
 ### Indexing
 
+def _is_integer_index(idx: Any) -> bool:
+  return isinstance(idx, (int, np.integer)) and not isinstance(idx, (bool, np.bool_))
+
+def _is_simple_reverse_slice(idx: Any) -> bool:
+  return (isinstance(idx, slice) and
+          idx.start is idx.stop is None and
+          isinstance(idx.step, int) and idx.step == -1)
+
+def _is_valid_integer_index_for_slice(idx, size, mode):
+  if size == 0:
+    return False
+  if _is_integer_index(idx):
+    return -size <= idx < size
+  try:
+    shape, dtype = np.shape(idx), _dtype(idx)
+  except:
+    return False
+  if shape == () and np.issubdtype(dtype, np.integer):
+    # For dynamic integer indices, dynamic_slice semantics require index clipping:
+    return mode in [None, 'promise_inbounds', 'clip']
+  return False
+
+def _is_contiguous_slice(idx):
+  return (isinstance(idx, slice) and
+          (idx.start is None or _is_integer_index(idx.start)) and
+          (idx.stop is None or _is_integer_index(idx.stop)) and
+          (idx.step is None or (_is_integer_index(idx.step) and idx.step == 1)))
+
+def _attempt_rewriting_take_via_slice(arr: Array, idx: Any, mode: Optional[str]) -> Optional[Array]:
+  # attempt to compute _rewriting_take via lax.slice(); return None if not possible.
+  idx = idx if isinstance(idx, tuple) else (idx,)
+
+  if not _all(isinstance(i, int) for i in arr.shape):
+    return None
+  if len(idx) > arr.ndim:
+    return None
+  if _any(i is None for i in idx):
+    return None  # TODO(jakevdp): handle newaxis case
+
+  simple_revs = {i for i, ind in enumerate(idx) if _is_simple_reverse_slice(ind)}
+  int_indices = {i for i, (ind, size) in enumerate(zip(idx, arr.shape))
+                 if _is_valid_integer_index_for_slice(ind, size, mode)}
+  contiguous_slices = {i for i, ind in enumerate(idx) if _is_contiguous_slice(ind)}
+
+  # For sharded inputs, indexing (like x[0]) and partial slices (like x[:2] as
+  # opposed to x[:]) lead to incorrect sharding semantics when computed via
+  # dynamic_slice, so we fall back to gather.
+  # TODO(yashkatariya): fix dynamic_slice with sharding
+  is_sharded = (isinstance(arr, ArrayImpl) and
+                not dispatch.is_single_device_sharding(arr.sharding))
+  has_partial_slices = _any(idx[i].indices(arr.shape[i]) != (0, arr.shape[i], 1)
+                            for i in contiguous_slices)
+  if is_sharded and (int_indices or has_partial_slices):
+    return None
+
+  if len(simple_revs) + len(int_indices) + len(contiguous_slices) != len(idx):
+    return None
+
+  if simple_revs:
+    arr = lax.rev(arr, tuple(simple_revs))
+    idx = tuple(slice(None) if i in simple_revs else ind
+                for i, ind in enumerate(idx))
+    contiguous_slices |= simple_revs
+
+  if not (int_indices or has_partial_slices):
+    return arr
+
+  idx += (arr.ndim - len(idx)) * (slice(None),)
+  start_indices: Sequence[ArrayLike] = []
+  slice_sizes: Sequence[int] = []
+
+  for ind, size in safe_zip(idx, arr.shape):
+    if isinstance(ind, slice):
+      start, stop, step = ind.indices(size)
+      assert step == 1  # checked above
+      start_indices.append(start)
+      slice_sizes.append(_max(0, stop - start))
+    else:
+      assert np.issubdtype(_dtype(ind), np.integer)  # checked above
+      assert np.shape(ind) == ()  # checked above
+      start_indices.append(ind)
+      slice_sizes.append(1)
+  # We must be careful with dtypes because dynamic_slice requires all
+  # start indices to have matching types.
+  if len(start_indices) > 1:
+    start_indices = util.promote_dtypes(*start_indices)
+  arr = lax.dynamic_slice(arr, start_indices=start_indices, slice_sizes=slice_sizes)
+  if int_indices:
+    arr = lax.squeeze(arr, tuple(int_indices))
+  return arr
+
+
 def _rewriting_take(arr, idx, indices_are_sorted=False, unique_indices=False,
                     mode=None, fill_value=None):
   # Computes arr[idx].
   # All supported cases of indexing can be implemented as an XLA gather,
   # followed by an optional reverse and broadcast_in_dim.
 
-  # Handle some special cases, falling back if error messages might differ.
-  if (arr.ndim > 0 and isinstance(idx, (int, np.integer)) and
-      not isinstance(idx, (bool, np.bool_)) and isinstance(arr.shape[0], int)):
-    if 0 <= idx < arr.shape[0]:
-      # Use dynamic rather than static index here to avoid slow repeated execution:
-      # See https://github.com/google/jax/issues/12198
-      return lax.dynamic_index_in_dim(arr, idx, keepdims=False)
-  if (arr.ndim > 0 and isinstance(arr.shape[0], int) and
-      isinstance(idx, slice) and
-      (type(idx.start) is int or idx.start is None) and
-      (type(idx.stop)  is int or idx.stop is  None) and
-      (type(idx.step)  is int or idx.step is  None)):
-    n = arr.shape[0]
-    start = idx.start if idx.start is not None else 0
-    stop  = idx.stop  if idx.stop  is not None else n
-    step  = idx.step  if idx.step  is not None else 1
-    if (0 <= start < n and 0 <= stop <= n and 0 < step and
-        (start, stop, step) != (0, n, 1)):
-      if _any(isinstance(d, core.Tracer) for d in arr.shape[1:]):
-        if step == 1:  # TODO(mattjj, sharadmv): handle step != 1
-          return lax.dynamic_slice_in_dim(arr, start, _max(0, stop - start), 0)
-      elif step == 1:
-        # Use dynamic rather than static slice here to avoid slow repeated execution:
-        # See https://github.com/google/jax/issues/12198
-        return lax.dynamic_slice_in_dim(arr, start, _max(0, stop - start), 0)
-      else:
-        return lax.slice_in_dim(arr, start, stop, step)
+  # For simplicity of generated primitives, we call lax.dynamic_slice in the
+  # simplest cases: i.e. non-dynamic arrays indexed with integers and slices.
+
+  if (result := _attempt_rewriting_take_via_slice(arr, idx, mode)) is not None:
+    return result
 
   # TODO(mattjj,dougalm): expand dynamic shape indexing support
   if jax.config.jax_dynamic_shapes and arr.ndim > 0:

tests/array_test.py

@@ -506,11 +506,7 @@ def _check(out, inp, shard_shape):
     _check(arr[0:6:1], np_inp[0:6:1], (2, 2, 1))
     _check(arr[:4], np_inp[:4], (2, 2, 1))
     _check(arr[::-1], np_inp[::-1], (2, 2, 1))
-
-    # TODO(yashkatariya): This returns a replicated output because the int
-    # indexing in `_rewriting_take` goes via `dynamic_index_in_dim` rather than
-    # `_gather`.
-    # _check(arr[1], np_inp[1], (2, 1))
+    _check(arr[1], np_inp[1], (2, 1))
 
   def test_array_getitem_replicated_multi_device(self):
     global_mesh = jtu.create_global_mesh((4, 2), ('x', 'y'))

tests/lax_numpy_indexing_test.py

@@ -854,6 +854,32 @@ def testJVPOfGradOfIndexing(self):
     self.assertAllClose(expected, primals)
     self.assertAllClose(np.zeros_like(x), tangents)
 
+  def testSimpleIndexingUsesSlice(self):
+    jaxpr = jax.make_jaxpr(lambda x: x[:2, :2])(jnp.ones((3, 4)))
+    self.assertEqual(len(jaxpr.jaxpr.eqns), 7)
+    self.assertEqual(jaxpr.jaxpr.eqns[-1].primitive, lax.dynamic_slice_p)
+
+    jaxpr = jax.make_jaxpr(lambda x: x[0, :2, 1])(jnp.ones((3, 4, 5)))
+    self.assertEqual(len(jaxpr.jaxpr.eqns), 11)
+    self.assertEqual(jaxpr.jaxpr.eqns[-2].primitive, lax.dynamic_slice_p)
+    self.assertEqual(jaxpr.jaxpr.eqns[-1].primitive, lax.squeeze_p)
+
+    jaxpr = jax.make_jaxpr(lambda x: x[0, 0])(jnp.ones((3, 4, 5)))
+    self.assertEqual(len(jaxpr.jaxpr.eqns), 11)
+    self.assertEqual(jaxpr.jaxpr.eqns[-2].primitive, lax.dynamic_slice_p)
+    self.assertEqual(jaxpr.jaxpr.eqns[-1].primitive, lax.squeeze_p)
+
+    jaxpr = jax.make_jaxpr(lambda x: x[:, 1])(jnp.ones((3, 4, 5)))
+    self.assertEqual(len(jaxpr.jaxpr.eqns), 11)
+    self.assertEqual(jaxpr.jaxpr.eqns[-2].primitive, lax.dynamic_slice_p)
+    self.assertEqual(jaxpr.jaxpr.eqns[-1].primitive, lax.squeeze_p)
+
+    # Simple reverses lower to lax.rev_p
+    jaxpr = jax.make_jaxpr(lambda x: x[:, ::-1])(jnp.ones((3, 4)))
+    print(jaxpr)
+    self.assertEqual(len(jaxpr.jaxpr.eqns), 1)
+    self.assertEqual(jaxpr.jaxpr.eqns[0].primitive, lax.rev_p)
+
   def testTrivialGatherIsntGenerated(self):
     # https://github.com/google/jax/issues/1621
     jaxpr = jax.make_jaxpr(lambda x: x[:, None])(np.arange(4))
@@ -862,13 +888,24 @@ def testTrivialGatherIsntGenerated(self):
 
     jaxpr = jax.make_jaxpr(lambda x: x[0:6:1])(np.arange(4))
     self.assertEqual(len(jaxpr.jaxpr.eqns), 0)
+
     jaxpr = jax.make_jaxpr(lambda x: x[:4])(np.arange(4))
     self.assertEqual(len(jaxpr.jaxpr.eqns), 0)
 
     jaxpr = jax.make_jaxpr(lambda x: x[::-1])(np.arange(4))
     self.assertEqual(len(jaxpr.jaxpr.eqns), 1)
     self.assertEqual(jaxpr.jaxpr.eqns[0].primitive, lax.rev_p)
 
+  def testOOBEmptySlice(self):
+    x = jnp.arange(4, dtype='float32')
+    self.assertArraysEqual(x[1:0], jnp.empty(0, dtype='float32'))
+    self.assertArraysEqual(x[-2:-10], jnp.empty(0, dtype='float32'))
+    self.assertArraysEqual(x[5:10], jnp.empty(0, dtype='float32'))
+
+    x = jnp.arange(6, dtype='float32').reshape(2, 3)
+    self.assertArraysEqual(x[-1:-4], jnp.empty((0, 3), dtype='float32'))
+    self.assertArraysEqual(x[:, 3:2], jnp.empty((2, 0), dtype='float32'))
+
   def testIndexingEmptyDimension(self):
     # Issue 2671: XLA error when indexing into dimension of size 0
     x = jnp.ones((2, 0))