Simplify and speedup numba DimShuffle implementation

Author: ricardoV94

pytensor/link/numba/dispatch/elemwise.py

@@ -4,6 +4,7 @@
 import numba
 import numpy as np
 from numba.core.extending import overload
+from numpy.lib.stride_tricks import as_strided
 
 from pytensor.graph.op import Op
 from pytensor.link.numba.dispatch import basic as numba_basic
@@ -411,91 +412,23 @@ def numba_funcify_CAReduce(op, node, **kwargs):
 
 @numba_funcify.register(DimShuffle)
 def numba_funcify_DimShuffle(op, node, **kwargs):
-    shuffle = tuple(op.shuffle)
-    transposition = tuple(op.transposition)
-    augment = tuple(op.augment)
+    new_order = tuple(op._new_order)
+    shape_template = (1,) * node.outputs[0].ndim
+    strides_template = (0,) * node.outputs[0].ndim
 
-    ndim_new_shape = len(shuffle) + len(augment)
-
-    no_transpose = all(i == j for i, j in enumerate(transposition))
-    if no_transpose:
-
-        @numba_basic.numba_njit
-        def transpose(x):
-            return x
-
-    else:
-
-        @numba_basic.numba_njit
-        def transpose(x):
-            return np.transpose(x, transposition)
-
-    shape_template = (1,) * ndim_new_shape
-
-    # When `len(shuffle) == 0`, the `shuffle_shape[j]` expression below
-    # is typed as `getitem(Tuple(), int)`, which has no implementation
-    # (since getting an item from an empty sequence doesn't make sense).
-    # To avoid this compile-time error, we omit the expression altogether.
-    if len(shuffle) > 0:
-        # Use the statically known shape if available
-        if all(length is not None for length in node.outputs[0].type.shape):
-            shape = node.outputs[0].type.shape
-
-            @numba_basic.numba_njit
-            def find_shape(array_shape):
-                return shape
-
-        else:
-
-            @numba_basic.numba_njit
-            def find_shape(array_shape):
-                shape = shape_template
-                j = 0
-                for i in range(ndim_new_shape):
-                    if i not in augment:
-                        length = array_shape[j]
-                        shape = numba_basic.tuple_setitem(shape, i, length)
-                        j = j + 1
-                return shape
-
-    else:
-
-        @numba_basic.numba_njit
-        def find_shape(array_shape):
-            return shape_template
-
-    if ndim_new_shape > 0:
-
-        @numba_basic.numba_njit
-        def dimshuffle_inner(x, shuffle):
-            x = transpose(x)
-            shuffle_shape = x.shape[: len(shuffle)]
-            new_shape = find_shape(shuffle_shape)
-
-            # FIXME: Numba's `array.reshape` only accepts C arrays.
-            return np.reshape(np.ascontiguousarray(x), new_shape)
+    @numba_basic.numba_njit
+    def dimshuffle(x):
+        old_shape = x.shape
+        old_strides = x.strides
 
-    else:
+        new_shape = shape_template
+        new_strides = strides_template
+        for i, o in enumerate(new_order):
+            if o != -1:
+                new_shape = numba_basic.tuple_setitem(new_shape, i, old_shape[o])
+                new_strides = numba_basic.tuple_setitem(new_strides, i, old_strides[o])
 
-        @numba_basic.numba_njit
-        def dimshuffle_inner(x, shuffle):
-            return np.reshape(np.ascontiguousarray(x), ())
-
-    # Without the following wrapper function we would see this error:
-    # E   No implementation of function Function(<built-in function getitem>) found for signature:
-    # E
-    # E    >>> getitem(UniTuple(int64 x 2), slice<a:b>)
-    # E
-    # E   There are 22 candidate implementations:
-    # E      - Of which 22 did not match due to:
-    # E      Overload of function 'getitem': File: <numerous>: Line N/A.
-    # E        With argument(s): '(UniTuple(int64 x 2), slice<a:b>)':
-    # E       No match.
-    # ...(on this line)...
-    # E           shuffle_shape = res.shape[: len(shuffle)]
-    @numba_basic.numba_njit(inline="always")
-    def dimshuffle(x):
-        return dimshuffle_inner(np.asarray(x), shuffle)
+        return as_strided(x, shape=new_shape, strides=new_strides)
 
     return dimshuffle
 

tests/link/numba/test_elemwise.py

@@ -23,6 +23,7 @@
 from tests.tensor.test_elemwise import (
     careduce_benchmark_tester,
     check_elemwise_runtime_broadcast,
+    dimshuffle_benchmark,
 )
 
 
@@ -643,3 +644,7 @@ def test_numba_careduce_benchmark(self, axis, c_contiguous, benchmark):
         return careduce_benchmark_tester(
             axis, c_contiguous, mode="NUMBA", benchmark=benchmark
         )
+
+    @pytest.mark.parametrize("c_contiguous", (True, False))
+    def test_dimshuffle(self, c_contiguous, benchmark):
+        dimshuffle_benchmark("NUMBA", c_contiguous, benchmark)

tests/tensor/test_elemwise.py

@@ -66,6 +66,30 @@ def custom_reduce(a):
     return np.apply_along_axis(custom_reduce, axis, x)
 
 
+def dimshuffle_benchmark(mode, c_contiguous, benchmark):
+    x = tensor3("x")
+    if c_contiguous:
+        x_val = np.random.random((2, 3, 4)).astype(config.floatX)
+    else:
+        x_val = np.random.random((200, 300, 400)).transpose(1, 2, 0)
+    ys = [x.transpose(t) for t in itertools.permutations((0, 1, 2))]
+    ys += [
+        x[None],
+        x[:, None],
+        x[:, :, None],
+        x[:, :, :, None],
+    ]
+    # Borrow to avoid deepcopy overhead
+    fn = pytensor.function(
+        [In(x, borrow=True)],
+        [Out(y, borrow=True) for y in ys],
+        mode=mode,
+    )
+    fn.trust_input = True
+    fn(x_val)  # JIT compile for JIT backends
+    benchmark(fn, x_val)
+
+
 class TestDimShuffle(unittest_tools.InferShapeTester):
     op = DimShuffle
     type = TensorType
@@ -218,23 +242,9 @@ def test_valid_input_ndim(self):
         with pytest.raises(TypeError, match="input_ndim must be an integer"):
             DimShuffle(input_ndim=(True, False), new_order=(1, 0))
 
-    def test_benchmark(self, benchmark):
-        x = tensor3("x")
-        x_val = np.random.random((2, 3, 4)).astype(config.floatX)
-        ys = [x.transpose(t) for t in itertools.permutations((0, 1, 2))]
-        ys += [
-            x[None],
-            x[:, None],
-            x[:, :, None],
-            x[:, :, :, None],
-        ]
-        # Borrow to avoid deepcopy overhead
-        fn = pytensor.function(
-            [In(x, borrow=True)],
-            [Out(y, borrow=True) for y in ys],
-        )
-        fn.trust_input = True
-        benchmark(fn, x_val)
+    @pytest.mark.parametrize("c_contiguous", [True, False])
+    def test_benchmark(self, c_contiguous, benchmark):
+        dimshuffle_benchmark("FAST_RUN", c_contiguous, benchmark)
 
 
 class TestBroadcast: