[Relay][Pass] Fix Depthwise AlterLayout

python/tvm/relay/op/nn/_nn.py

@@ -345,3 +345,28 @@ def schedule_contrib_conv2d_NCHWc(attrs, outs, target):
 
 reg.register_pattern("nn.contrib_conv2d_NCHWc",
                      OpPattern.OUT_ELEMWISE_FUSABLE)
+
+@reg.register_compute("nn.contrib_depthwise_conv2d_NCHWc")
+def compute_contrib_depthwise_conv2d_NCHWc(attrs, inputs, out_dtype, target):
+    """Compute definition of depthwise conv2d NCHWc"""
+    # pylint: disable=assignment-from-no-return
+    padding = attrs.get_int_tuple("padding")
+    strides = attrs.get_int_tuple("strides")
+    dilation = attrs.get_int_tuple("dilation")
+    data_layout = attrs.get_str("data_layout")
+    out_layout = attrs.get_str("out_layout")
+    out_dtype = attrs.get_str("out_dtype")
+    out_dtype = inputs[0].dtype if out_dtype == "" else out_dtype
+
+    out = topi.nn.depthwise_conv2d_NCHWc(inputs[0], inputs[1], strides, padding, dilation,
+                                         data_layout, out_layout, out_dtype)
+    return [out]
+
+@reg.register_schedule("nn.contrib_depthwise_conv2d_NCHWc")
+def schedule_contrib_depthwise_conv2d_NCHWc(attrs, outs, target):
+    """Schedule definition of contrib_conv2d_NCHWc"""
+    with target:
+        return topi.generic.schedule_depthwise_conv2d_NCHWc(outs)
+
+reg.register_pattern("nn.contrib_depthwise_conv2d_NCHWc",
+                     OpPattern.OUT_ELEMWISE_FUSABLE)

python/tvm/relay/op/nn/nn.py

@@ -927,6 +927,70 @@ def contrib_conv2d_nchwc(data,
                                       groups, channels, kernel_size, data_layout,
                                       kernel_layout, out_layout, out_dtype)
 
+def contrib_depthwise_conv2d_nchwc(data,
+                                   kernel,
+                                   strides=(1, 1),
+                                   padding=(0, 0),
+                                   dilation=(1, 1),
+                                   groups=1,
+                                   channels=None,
+                                   kernel_size=None,
+                                   data_layout="NCHW8c",
+                                   kernel_layout="OIHW",
+                                   out_layout="",
+                                   out_dtype=""):
+    r"""Variant of 2D depthwise convolution.
+
+    This operator takes the weight as the depthwise convolution kernel
+    and depthwise convolves it with data to produce an output, following a specialized
+    NCHWc data layout.
+
+    Parameters
+    ----------
+    data : tvm.relay.Expr
+        The input data to the operator.
+
+    kernel : tvm.relay.Expr
+        The kernel expressions.
+
+    strides : tuple of int, optional
+        The strides of convoltution.
+
+    padding : tuple of int, optional
+        The padding of convolution on both sides of inputs before convolution.
+
+    dilation : tuple of int, optional
+        Specifies the dilation rate to be used for dilated convolution.
+
+    groups : int, optional
+        Number of groups for grouped convolution.
+
+    channels : int, optional
+        Number of output channels of this convolution.
+
+    kernel_size : tuple of int, optional
+        The spatial of the convolution kernel.
+
+    data_layout : str, optional
+        Layout of the input.
+
+    kernel_layout : str, optional
+        Layout of the weight.
+
+    out_layout : str, optional
+        Layout of the output, by default, out_layout is the same as data_layout
+
+    out_dtype : str, optional
+        Specifies the output data type for mixed precision conv2d.
+
+    Returns
+    -------
+    result : tvm.relay.Expr
+        The computed result.
+    """
+    return _make.contrib_depthwise_conv2d_NCHWc(data, kernel, strides, padding, dilation,
+                                                groups, channels, kernel_size, data_layout,
+                                                kernel_layout, out_layout, out_dtype)
 
 def contrib_conv2d_winograd_weight_transform(weight,
                                              tile_size):

src/relay/op/nn/convolution.cc

@@ -582,5 +582,57 @@ RELAY_REGISTER_OP("nn.contrib_conv2d_NCHWc")
         Conv2DInferCorrectLayout<Conv2DAttrs>);
 
 
+// Positional relay function to create depthwise conv2d NCHWc operator
+// used by frontend FFI.
+Expr MakeDepthwiseConv2DNCHWc(Expr data,
+                     Expr kernel,
+                     Array<IndexExpr> strides,
+                     Array<IndexExpr> padding,
+                     Array<IndexExpr> dilation,
+                     int groups,
+                     IndexExpr channels,
+                     Array<IndexExpr> kernel_size,
+                     std::string data_layout,
+                     std::string kernel_layout,
+                     std::string out_layout,
+                     DataType out_dtype) {
+  auto attrs = make_node<Conv2DAttrs>();
+  attrs->strides = std::move(strides);
+  attrs->padding = std::move(padding);
+  attrs->dilation = std::move(dilation);
+  attrs->groups = groups;
+  attrs->channels = channels;
+  attrs->kernel_size = std::move(kernel_size);
+  attrs->data_layout = std::move(data_layout);
+  attrs->kernel_layout = std::move(kernel_layout);
+  attrs->out_layout = std::move(out_layout);
+  attrs->out_dtype = std::move(out_dtype);
+  static const Op& op = Op::Get("nn.contrib_depthwise_conv2d_NCHWc");
+  return CallNode::make(op, {data, kernel}, Attrs(attrs), {});
+}
+
+TVM_REGISTER_API("relay.op.nn._make.contrib_depthwise_conv2d_NCHWc")
+.set_body([](const TVMArgs& args, TVMRetValue* rv) {
+    runtime::detail::unpack_call<Expr, 12>(MakeDepthwiseConv2DNCHWc, args, rv);
+  });
+
+
+RELAY_REGISTER_OP("nn.contrib_depthwise_conv2d_NCHWc")
+.describe(R"code(Compute conv2d with NCHWc data layout. Only supports NCHW layout.
+- **data**: Input is 5D packed tensor.
+- **weight**: 6D packed tensor.
+
+- **out**:  Output is 5D packed tensor
+)code" TVM_ADD_FILELINE)
+.set_attrs_type_key("relay.attrs.DepthwiseConv2D")
+.set_num_inputs(2)
+.add_argument("data", "Tensor", "The input tensor.")
+.add_argument("weight", "Tensor", "The weight tensor.")
+.set_support_level(10)
+.add_type_rel("Conv2D", Conv2DRel)
+.set_attr<FInferCorrectLayout>("FInferCorrectLayout",
+        Conv2DInferCorrectLayout<Conv2DAttrs>);
+
+
 }  // namespace relay
 }  // namespace tvm

tests/lint/pylintrc

@@ -114,7 +114,7 @@ single-line-if-stmt=no
 no-space-check=trailing-comma,dict-separator
 
 # Maximum number of lines in a module
-max-module-lines=1000
+max-module-lines=1500
 
 # String used as indentation unit. This is usually "    " (4 spaces) or "\t" (1
 # tab).

topi/python/topi/x86/conv2d.py

@@ -290,7 +290,7 @@ def _alter_conv2d_layout(attrs, inputs, tinfo, F):
     batch_size, in_channel, height, width = get_const_tuple(data.shape)
 
     groups = attrs.get_int("groups")
-    out_channel = attrs.get_int("channels")
+    out_channel = attrs.get_int("channels") if F == sym else attrs.get_int("channels").value
     padding = attrs.get_int_tuple("padding")
     strides = attrs.get_int_tuple("strides")
     dilation = attrs.get_int_tuple("dilation")
@@ -330,16 +330,11 @@ def _alter_conv2d_layout(attrs, inputs, tinfo, F):
 
     new_data = tvm.placeholder((batch_size, in_channel//ic_bn, height, width, ic_bn),
                                dtype=data.dtype)
-    if is_depthwise:
-        # channel, channel_multiplier, kh, kw -> out_channel_chunk, kh, kw, out_channel_block
-        # in which out_channel = merge(channel, channel_multiplier)
-        kernel_sym = copy_inputs[1]
-        kernel_sym = sym.reshape(kernel_sym, shape=(out_channel//oc_bn, oc_bn, kh, kw))
-        kernel_sym = sym.transpose(kernel_sym, axes=(0, 2, 3, 1))
-        copy_inputs[1] = kernel_sym
 
+    if is_depthwise:
+        new_attrs['kernel_layout'] = 'OIHW1i%do' % oc_bn
         # Store altered operator's config
-        new_kernel = tvm.placeholder((out_channel//oc_bn, kh, kw, oc_bn), dtype=kernel.dtype)
+        new_kernel = tvm.placeholder((out_channel//oc_bn, 1, kh, kw, 1, oc_bn), dtype=kernel.dtype)
         new_workload = autotvm.task.args_to_workload(
             [new_data, new_kernel, strides, padding, dilation, new_attrs[layout_name],
              new_attrs['out_layout'], out_dtype], depthwise_conv2d_NCHWc)
@@ -356,9 +351,16 @@ def _alter_conv2d_layout(attrs, inputs, tinfo, F):
              new_attrs['out_layout'], out_dtype], conv2d_NCHWc)
 
     dispatch_ctx.update(target, new_workload, cfg)
-    if F == sym:
-        return F.contrib.conv2d_NCHWc(*copy_inputs, **new_attrs)
-    return F.nn.contrib_conv2d_nchwc(*copy_inputs, **new_attrs)
+
+    if is_depthwise:
+        if F == sym:
+            raise Exception("NNVM is no longer supported.")
+        else:
+            return F.nn.contrib_depthwise_conv2d_nchwc(*copy_inputs, **new_attrs)
+    else:
+        if F == sym:
+            return F.contrib.conv2d_NCHWc(*copy_inputs, **new_attrs)
+        return F.nn.contrib_conv2d_nchwc(*copy_inputs, **new_attrs)
 
 
 @autotvm.register_topi_compute(conv2d_NCHWc, 'cpu', 'direct')

topi/python/topi/x86/depthwise_conv2d.py

@@ -58,7 +58,7 @@ def _depthwise_conv2d_NCHWc_cpu(cfg, data, kernel, strides, padding, dilation,
                                 layout, out_layout, out_dtype=None):
     out_dtype = data.dtype if out_dtype is None else out_dtype
     batch, in_channel_chunk, in_height, in_width, in_channel_block = get_const_tuple(data.shape)
-    out_channel_chunk, filter_height, filter_width, out_channel_block \
+    out_channel_chunk, _, filter_height, filter_width, __, out_channel_block \
         = get_const_tuple(kernel.shape)
 
     strides = strides if isinstance(strides, (tuple, list)) else (strides, strides)
@@ -102,7 +102,7 @@ def _depthwise_conv2d_NCHWc_cpu(cfg, data, kernel, strides, padding, dilation,
                       oh*HSTR+kh, ow*WSTR+kw,
                       ((oco * out_channel_block + oci) // channel_multiplier) % in_channel_block]
              .astype(out_dtype) *
-             kernel[oco, kh, kw, oci].astype(out_dtype)),
+             kernel[oco, 0, kh, kw, 0, oci].astype(out_dtype)),
             axis=[kh, kw]),
         name='DepthwiseConv2d', tag="depthwise_conv2d_NCHWc")
     return Output