MLBEDSW-6686: Resize bilinear HPC with tile padding
- Added support for Resize Bilinear with half pixel centers for int8 and
uint8.
- Utilizes the new "TILE" padding mode.
- Utilizes ofm stride multipliers and modified tile base offsets to
write OFMs interleaved.
Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>
Change-Id: I37fa77c022a368f05fda0ead75d8696c9205f833
diff --git a/ethosu/vela/operation.py b/ethosu/vela/operation.py
index e162204..af2205c 100644
--- a/ethosu/vela/operation.py
+++ b/ethosu/vela/operation.py
@@ -1,4 +1,4 @@
-# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
+# Copyright (C) 2020-2022 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -474,7 +474,7 @@
__slots__ = (
"type",
- "original_type",
+ "_original_type",
"name",
"op_index",
"attrs",
@@ -501,12 +501,14 @@
"write_offset",
"write_shape",
"ifm_resampling_mode",
+ "tile_base_offsets_ifm",
+ "tile_base_offsets_ofm",
"ofm_stride_multiplier",
)
def __init__(self, op_type: Op, name: str):
self.type = op_type
- self.original_type = op_type
+ self._original_type = op_type # the original type of the operation. once set this shouldn't be changed
self.name = name
self.attrs: Dict[str, Any] = {}
self.inputs: List[Optional[Tensor]] = []
@@ -546,6 +548,10 @@
# write_offset 0,9,0,0, write_shape 1,1,8,1
self.write_shape: Optional[Shape4D] = None
self.ifm_resampling_mode: resampling_mode = resampling_mode.NONE
+ # ifm (nhwc), ifm2 (nhwc)
+ self.tile_base_offsets_ifm: List[List[int]] = [[0, 0, 0, 0], [0, 0, 0, 0]]
+ # ofm (nhwc)
+ self.tile_base_offsets_ofm: List[int] = [0, 0, 0, 0]
# For interleaved/sparse outputs - stride is multiplied with the stride factor of the corresponding axis
# Order is [C, H, W] - default is no multiplication
self.ofm_stride_multiplier: List[int] = [1, 1, 1]
@@ -553,6 +559,9 @@
def clone(self, suffix="_clone"):
res = Operation(self.type, self.name + suffix)
+ # maintain the original type, in cases where the type was changed to something different
+ res._original_type = self._original_type
+
res.attrs = dict(self.attrs)
res.inputs = list(self.inputs)
res.outputs = list(self.outputs)
@@ -567,11 +576,15 @@
res.op_index = None # not relevant as not part of input network
res.read_offsets = list(self.read_offsets)
res.read_shapes = list(self.read_shapes)
+ res.write_offset = Shape4D(*self.write_offset) if self.write_offset else None
+ res.write_shape = Shape4D(*self.write_shape) if self.write_shape else None
res.rounding_mode = self.rounding_mode
res.explicit_scaling = self.explicit_scaling
res.low_precision_scaling = self.low_precision_scaling
res.rescale = self.rescale
res.ifm_resampling_mode = self.ifm_resampling_mode
+ res.tile_base_offsets_ifm = [_ifm.copy() for _ifm in self.tile_base_offsets_ifm]
+ res.tile_base_offsets_ofm = self.tile_base_offsets_ofm.copy()
res.ofm_stride_multiplier = self.ofm_stride_multiplier.copy()
return res
@@ -581,6 +594,10 @@
__repr__ = __str__
+ @property
+ def original_type(self):
+ return self._original_type
+
def get_kernel_size(self):
weights = self.weights
if weights and self.type.npu_block_type in (NpuBlockType.ConvolutionDepthWise, NpuBlockType.ConvolutionMxN):