Conv3D ONNX support and conv3D_ncdhw x86 schedules (#4949)

* Support 3d Convolution with the ONNX frontend

* add unit tests for conv3d in onnx frontend

respond to PR formatting requests

add x86 schedules to conv3d ncdhw test

fix a doc string format issue

refactor for changed upsream API

* first attempt at conv3d autotuning

add default schedule for conv3d_ncdhw

fill in autotvm integration

add a fallback for invalid schedules

fix fallback

fix reduction order to get simd working correctly

python/tvm/relay/frontend/onnx.py

@@ -91,16 +91,18 @@ def get_numpy(tensor_proto):
     return to_array(tensor_proto)
 
 
-def dimension_picker(prefix, surfix=''):
+def dimension_picker(prefix, suffix=''):
     """Check that dimensions are supported."""
     def _impl(attr):
         kernel = attr['kernel_shape']
         if len(kernel) == 1:
-            return prefix + '1d' + surfix
+            return prefix + '1d' + suffix
         if len(kernel) == 2:
-            return prefix + '2d' + surfix
-        msg = 'Only 1D and 2D kernels are supported for operator {}.'
-        op_name = prefix + '1d/2d'
+            return prefix + '2d' + suffix
+        if len(kernel) == 3:
+            return prefix + '3d' + suffix
+        msg = 'Only 1D, 2D, and 3D kernels are supported for operator {}.'
+        op_name = prefix + '1d/2d/3d'
         raise tvm.error.OpAttributeInvalid(msg.format(op_name))
 
     return _impl
@@ -155,11 +157,11 @@ def onnx_storage_order2layout(storage_order, dims=2):
 
 def dimension_constraint():
     def _dim_check(attrs):
-        if len(attrs['kernel_shape']) == 2 or len(attrs['kernel_shape']) == 1:
+        if len(attrs['kernel_shape']) in [1, 2, 3]:
             return True
         return False
 
-    return _dim_check, "Only 1d and 2d kernel supported."
+    return _dim_check, "Only 1d, 2d and 3d kernel supported."
 
 
 class OnnxOpConverter(object):

python/tvm/relay/op/strategy/x86.py

@@ -188,10 +188,9 @@ def conv3d_strategy_cpu(attrs, inputs, out_type, target):
     strategy = _op.OpStrategy()
     layout = attrs.data_layout
     if layout == "NCDHW":
-        logger.warning("conv3d with layout NCDHW is not optimized for x86.")
-        strategy.add_implementation(wrap_compute_conv3d(topi.nn.conv3d_ncdhw),
-                                    wrap_topi_schedule(topi.generic.schedule_conv3d_ncdhw),
-                                    name="conv3d_ncdhw.generic")
+        strategy.add_implementation(wrap_compute_conv3d(topi.x86.conv3d_ncdhw),
+                                    wrap_topi_schedule(topi.x86.schedule_conv3d_ncdhw),
+                                    name="conv3d_ncdhw.x86")
     elif layout == "NDHWC":
         strategy.add_implementation(wrap_compute_conv3d(topi.x86.conv3d_ndhwc),
                                     wrap_topi_schedule(topi.x86.schedule_conv3d_ndhwc),

tests/python/frontend/onnx/test_forward.py

@@ -1794,37 +1794,51 @@ def verify_conv(x_shape, w_shape, y_shape, padding, kernel_shape, strides, dilat
 
 
 def test_conv():
+    def repeat(N, D):
+        return tuple([N for _ in range(D)])
+    for D in [1, 2, 3]:
         # Convolution with padding
-    # Conv2D
-    verify_conv((1, 1, 5, 5), (1, 1, 3, 3), (1, 1, 5, 5), [1, 1, 1, 1], [3, 3], [1, 1], [1, 1])
-    # Conv1D
-    verify_conv((1, 1, 5), (1, 1, 3), (1, 1, 5), [1, 1], [3], [1], [1])
-
+        verify_conv((1, 1) + repeat(5, D),
+                    (1, 1) + repeat(3, D),
+                    (1, 1) + repeat(5, D),
+                    2 * repeat(1, D),
+                    repeat(3, D),
+                    repeat(1, D),
+                    repeat(1, D))
         # Convolution without padding
-    # Conv2D
-    verify_conv((1, 1, 5, 5), (1, 1, 3, 3), (1, 1, 3, 3), [0, 0, 0, 0], [3, 3], [1, 1], [1, 1])
-    # Conv1D
-    verify_conv((1, 1, 5), (1, 1, 3), (1, 1, 3), [0, 0], [3], [1], [1])
-
+        verify_conv((1, 1) + repeat(5, D),
+                    (1, 1) + repeat(3, D),
+                    (1, 1) + repeat(3, D),
+                    2 * repeat(0, D),
+                    repeat(3, D),
+                    repeat(1, D),
+                    repeat(1, D))
         # Convolution with autopadding
-    verify_conv((1, 1, 5, 5), (1, 1, 3, 3), (1, 1, 5, 5),
-                None, [3, 3], [1, 1], [1, 1],
+        verify_conv((1, 1) + repeat(5, D),
+                    (1, 1) + repeat(3, D),
+                    (1, 1) + repeat(5, D),
+                    None,
+                    repeat(3, D),
+                    repeat(1, D),
+                    repeat(1, D),
                     auto_pad="SAME_UPPER")
-    # Conv1D
-    verify_conv((1, 1, 5), (1, 1, 3), (1, 1, 5), None, [3], [1], [1], auto_pad="SAME_UPPER")
-
         # Convolution with non uniform stride
-    verify_conv((1, 1, 5, 5), (1, 1, 3, 3), (1, 1, 3, 3),
-                None, [3, 3], [2, 2], [1, 1],
+        verify_conv((1, 1) + repeat(5, D),
+                    (1, 1) + repeat(3, D),
+                    (1, 1) + repeat(3, D),
+                    None,
+                    repeat(3, D),
+                    repeat(2, D),
+                    repeat(1, D),
                     auto_pad="SAME_UPPER")
-    # Conv1D
-    verify_conv((1, 1, 5), (1, 1, 3), (1, 1, 3), None, [3], [2], [1], auto_pad="SAME_UPPER")
-
         # Convolution with dilation
-    verify_conv((1, 1, 5, 5), (1, 1, 3, 3), (1, 1, 5, 5), [2, 2, 2, 2], [3, 3], [1, 1], [2, 2])
-    # Conv1D
-    verify_conv((1, 1, 5), (1, 1, 3), (1, 1, 5), [2, 2], [3], [1], [2])
-
+        verify_conv((1, 1) + repeat(5, D),
+                    (1, 1) + repeat(3, D),
+                    (1, 1) + repeat(5, D),
+                    2 * repeat(2, D),
+                    repeat(3, D),
+                    repeat(1, D),
+                    repeat(2, D))
 
 def verify_convtranspose(x_shape, w_shape, y_shape, p):
     node = onnx.helper.make_node("ConvTranspose",

topi/tests/python/test_topi_conv3d_ncdhw.py

@@ -30,6 +30,7 @@ from common import get_all_backend
 
 _conv3d_ncdhw_implement = {
     "generic": (topi.nn.conv3d_ncdhw, topi.generic.schedule_conv3d_ncdhw),
+    "cpu": (topi.x86.conv3d_ncdhw, topi.x86.schedule_conv3d_ncdhw),
     "gpu": (topi.cuda.conv3d_ncdhw, topi.cuda.schedule_conv3d_ncdhw),
 }