MLBEDSW-3499: Support for PAD operator
Replaces the PAD operator by hardware padding when possible.
Change-Id: I9dce0885e51a4a73715824d7368637222e39b2b3
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
diff --git a/ethosu/vela/graph_optimiser.py b/ethosu/vela/graph_optimiser.py
index 3759d3b..00edf83 100644
--- a/ethosu/vela/graph_optimiser.py
+++ b/ethosu/vela/graph_optimiser.py
@@ -156,7 +156,7 @@
return total_padding
-def calc_padding_and_skirt(padding_type, kernel_size, stride, input_dims):
+def calc_padding_and_skirt(padding_type, kernel_size, stride, input_dims, explicit_padding):
ypad = needed_total_padding(int(input_dims[1]), int(stride[1]), int(kernel_size[0]))
xpad = needed_total_padding(int(input_dims[2]), int(stride[2]), int(kernel_size[1]))
if padding_type == Padding.SAME:
@@ -169,6 +169,12 @@
right_pad = 0
top_pad = 0
bottom_pad = 0
+ elif padding_type == Padding.EXPLICIT:
+ # Padding is specified in a PAD operator which has been bypassed.
+ # The top and left padding are taken from the PAD; bottom and right are calculated.
+ top_pad, left_pad, _, _ = explicit_padding
+ bottom_pad = ypad - top_pad
+ right_pad = xpad - left_pad
else:
raise UnsupportedFeatureError(f"Unknown padding")
padding = (top_pad, left_pad, bottom_pad, right_pad)
@@ -537,7 +543,11 @@
dilation_h, dilation_w = op.get_dilation_h_w()
dilated_kernel_size = [dilation_h * (kernel_size[0] - 1) + 1, dilation_w * (kernel_size[1] - 1) + 1]
padding, skirt = calc_padding_and_skirt(
- op.attrs["padding"], dilated_kernel_size, op.attrs["strides"], input_shape
+ op.attrs["padding"],
+ dilated_kernel_size,
+ op.attrs["strides"],
+ input_shape,
+ op.attrs.get("explicit_padding"),
)
op.attrs["explicit_padding"] = padding
@@ -1122,6 +1132,30 @@
return op
+def optimise_pad(op, arch, nng):
+ """
+ Converts tens1 -> PAD -> tens2 -> CONV to tens1 -> CONV
+ if both operations can be run on the NPU.
+ """
+ if (
+ (op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op())
+ and op.run_on_npu
+ and op.attrs["padding"] == Padding.VALID
+ ):
+ pad_op = op.ifm.ops[0]
+ if pad_op.type != Op.Pad or not pad_op.run_on_npu:
+ return op
+ # Bypass the PAD operator
+ op.set_input_tensor(pad_op.ifm, 0)
+ # Adjust the padding attributes of the convolution operator
+ op.attrs["padding"] = Padding.EXPLICIT
+ padding = pad_op.inputs[1].values # 4x2 tensor, first dimension is N, H, W, C
+ top, left, bottom, right = (padding[1][0], padding[2][0], padding[1][1], padding[2][1])
+ op.attrs["explicit_padding"] = (top, left, bottom, right)
+ op.set_ifm_ofm_shapes()
+ return op
+
+
def add_attrs_to_resizebilinear(op, arch, nng):
if op.type == Op.ResizeBilinear and op.run_on_npu:
input_tensor = op.inputs[0]
@@ -1213,7 +1247,11 @@
for idx, sg in enumerate(nng.subgraphs):
# remove passthrough tensors and attempt further optimizations
nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng, sg, arch, [remove_passthrough_tensor], [fuse_activation_function_with_prev, add_padding_fields],
+ nng,
+ sg,
+ arch,
+ [remove_passthrough_tensor],
+ [fuse_activation_function_with_prev, optimise_pad, add_padding_fields],
)
# Post-optimisation operator debug tracing