[TOPI] Fix nn.pool*d issue with 'vectorize' function and add unit t…

…ests (apache#8541)

* Fix issue in 'vectorize' function for 1D and 3D tensors

* Add pooling tests for channel last layouts

* Add support for more general layouts in "poolnd" implementation

* Reformat with 'black'

* Fix lint issues

python/tvm/topi/testing/poolnd_python.py

@@ -18,12 +18,60 @@
 """Ground truth max and average pooling operators in python."""
 import itertools
 import math
-from typing import List, Tuple
+from typing import List, Tuple, Optional
 
 import numpy as np
 import tvm
 
 
+def _get_supported_layout(dims: int):
+    """
+    Returns layout that is supported by poolnd_python based on number of
+    dimensions of input tensor
+    """
+    assert dims in [3, 4, 5], f"{dims}-dimensional tensor is not supported"
+    if dims == 3:
+        return "NCW"
+    if dims == 4:
+        return "NCHW"
+    # dims == 5
+    return "NCDHW"
+
+
+def _convert_to_layout(
+    input_tensor: np.ndarray,
+    layout: str,
+) -> np.ndarray:
+    """
+    Converts back to original layout after the algorithm is finished
+    """
+    supported_layout = _get_supported_layout(input_tensor.ndim)
+    if layout is not None and supported_layout != layout:
+        # Generate transpose list
+        transpose_list = []
+        for d in layout:
+            transpose_list.append(supported_layout.index(d))
+        return input_tensor.transpose(transpose_list)
+    return input_tensor
+
+
+def _convert_from_layout(
+    input_tensor: np.ndarray,
+    layout: str,
+) -> np.ndarray:
+    """
+    Converts tensor to one of suppored layouts
+    """
+    supported_layout = _get_supported_layout(input_tensor.ndim)
+    if layout is not None and supported_layout != layout:
+        # Generate transpose list
+        transpose_list = []
+        for d in supported_layout:
+            transpose_list.append(layout.index(d))
+        return input_tensor.transpose(transpose_list)
+    return input_tensor
+
+
 def get_slice(
     spatial_dimensions: int,
     pad_np: np.array,
@@ -90,8 +138,12 @@ def poolnd_python(
     count_include_pad: bool = True,
     ceil_mode: bool = False,
     dtype: str = "float32",
+    layout: Optional[str] = None,
 ) -> np.array:
     """Ground truth pooling operator impelmented in numpy."""
+
+    np_data = _convert_from_layout(np_data, layout)
+
     out_shape = [np_data.shape[0], np_data.shape[1]]
     for dim in range(2, len(np_data.shape)):
         i = dim - 2
@@ -158,4 +210,4 @@ def poolnd_python(
         else:
             raise ValueError("Pool type {} is not supported".format(pool_type))
 
-    return ret_np
+    return _convert_to_layout(ret_np, layout)

python/tvm/topi/x86/pooling.py

@@ -26,8 +26,8 @@ def vectorize(fused_axis, num_parallel_axis, vectorize_limit=64):
         reorder_axis = [fused_axis]
         for i in range(num_parallel_axis, len(sch.op.axis) - 1):
             reorder_axis.append(sch.op.axis[i])
-        kw, kh = sch.op.reduce_axis
-        fuse_k = sch.fuse(kw, kh)
+        k = sch.op.reduce_axis
+        fuse_k = sch.fuse(*k)
         c = sch.op.axis[len(sch.op.axis) - 1]
         reorder_axis += [fuse_k, c]
         sch.reorder(*reorder_axis)
@@ -83,7 +83,7 @@ def schedule_pool(outs, layout):
     def _schedule(PaddedInput, Pool):
         if isinstance(PaddedInput.op, te.tensor.ComputeOp):
             s[PaddedInput].compute_inline()
-        do_vectorize = layout[-1] not in "HWhw"
+        do_vectorize = layout[-1] not in "DHWdhw"
         _parallel_sch(s[Pool], outs[0].shape, do_vectorize)
 
     def traverse(OP):