Fix BatchToSpaceFixture
* Use a vector to represent the (static) block shape instead of an N-D
Tensor. The previous use of ND Tensor as block shape was wrong, not
adhering to the specification, and non-functional (only first dim was
used anyway).
* The fixture now accepts a static block shape, because the dynamic
case is not properly implemented and will be deprecated for now.
* Fix an assertion error in reference implementation.
Partially resolves COMPMID-5918
Change-Id: I5221e52ccc05e7c1249dec3a42426f954a73729a
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9357
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-by: Omar Al Khatib <omar.alkhatib@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/datasets/BatchToSpaceDataset.h b/tests/datasets/BatchToSpaceDataset.h
index 1edd457..2670af5 100644
--- a/tests/datasets/BatchToSpaceDataset.h
+++ b/tests/datasets/BatchToSpaceDataset.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -38,15 +38,17 @@
class BatchToSpaceLayerDataset
{
public:
- using type = std::tuple<TensorShape, TensorShape, TensorShape>;
+ using type = std::tuple<TensorShape, std::vector<int32_t>, CropInfo, TensorShape>;
struct iterator
{
- iterator(std::vector<TensorShape>::const_iterator src_it,
- std::vector<TensorShape>::const_iterator block_shape_it,
- std::vector<TensorShape>::const_iterator dst_it)
+ iterator(std::vector<TensorShape>::const_iterator src_it,
+ std::vector<std::vector<int32_t>>::const_iterator block_shape_it,
+ std::vector<CropInfo>::const_iterator crop_info_it,
+ std::vector<TensorShape>::const_iterator dst_it)
: _src_it{ std::move(src_it) },
_block_shape_it{ std::move(block_shape_it) },
+ _crop_info_it{ std::move(crop_info_it) },
_dst_it{ std::move(dst_it) }
{
}
@@ -56,44 +58,48 @@
std::stringstream description;
description << "In=" << *_src_it << ":";
description << "BlockShape=" << *_block_shape_it << ":";
+ description << "CropInfo=" << *_crop_info_it << ":";
description << "Out=" << *_dst_it;
return description.str();
}
BatchToSpaceLayerDataset::type operator*() const
{
- return std::make_tuple(*_src_it, *_block_shape_it, *_dst_it);
+ return std::make_tuple(*_src_it, *_block_shape_it, *_crop_info_it, *_dst_it);
}
iterator &operator++()
{
++_src_it;
++_block_shape_it;
+ ++_crop_info_it;
++_dst_it;
return *this;
}
private:
- std::vector<TensorShape>::const_iterator _src_it;
- std::vector<TensorShape>::const_iterator _block_shape_it;
- std::vector<TensorShape>::const_iterator _dst_it;
+ std::vector<TensorShape>::const_iterator _src_it;
+ std::vector<std::vector<int32_t>>::const_iterator _block_shape_it;
+ std::vector<CropInfo>::const_iterator _crop_info_it;
+ std::vector<TensorShape>::const_iterator _dst_it;
};
iterator begin() const
{
- return iterator(_src_shapes.begin(), _block_shape_shapes.begin(), _dst_shapes.begin());
+ return iterator(_src_shapes.begin(), _block_shapes.begin(), _crop_infos.begin(), _dst_shapes.begin());
}
int size() const
{
- return std::min(_src_shapes.size(), std::min(_block_shape_shapes.size(), _dst_shapes.size()));
+ return std::min(std::min(std::min(_src_shapes.size(), _block_shapes.size()), _crop_infos.size()), _dst_shapes.size());
}
- void add_config(TensorShape src, TensorShape block_shape, TensorShape dst)
+ void add_config(const TensorShape &src, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &dst)
{
_src_shapes.emplace_back(std::move(src));
- _block_shape_shapes.emplace_back(std::move(block_shape));
+ _block_shapes.emplace_back(std::move(block_shape));
+ _crop_infos.emplace_back(std::move(crop_info));
_dst_shapes.emplace_back(std::move(dst));
}
@@ -102,35 +108,60 @@
BatchToSpaceLayerDataset(BatchToSpaceLayerDataset &&) = default;
private:
- std::vector<TensorShape> _src_shapes{};
- std::vector<TensorShape> _block_shape_shapes{};
- std::vector<TensorShape> _dst_shapes{};
+ std::vector<TensorShape> _src_shapes{};
+ std::vector<std::vector<int32_t>> _block_shapes{};
+ std::vector<CropInfo> _crop_infos{};
+ std::vector<TensorShape> _dst_shapes{};
};
+/** Follow NCHW data layout across all datasets. I.e.
+ * TensorShape(Width(X), Height(Y), Channel(Z), Batch(W))
+ */
+
class SmallBatchToSpaceLayerDataset final : public BatchToSpaceLayerDataset
{
public:
SmallBatchToSpaceLayerDataset()
{
- add_config(TensorShape(1U, 1U, 1U, 4U), TensorShape(2U), TensorShape(2U, 2U, 1U, 1U));
- add_config(TensorShape(3U, 1U, 1U, 4U), TensorShape(2U), TensorShape(6U, 2U, 1U, 1U));
- add_config(TensorShape(1U, 2U, 2U, 4U), TensorShape(2U), TensorShape(2U, 4U, 2U, 1U));
- add_config(TensorShape(1U, 3U, 1U, 8U), TensorShape(2U), TensorShape(2U, 6U, 1U, 2U));
- add_config(TensorShape(3U, 4U, 1U, 4U), TensorShape(2U), TensorShape(6U, 8U, 1U, 1U));
- add_config(TensorShape(1U, 1U, 1U, 8U), TensorShape(4U, 2U), TensorShape(4U, 2U, 1U, 1U));
- add_config(TensorShape(3U, 1U, 1U, 8U), TensorShape(2U, 4U), TensorShape(6U, 4U, 1U, 1U));
+ // Block size = 1 (effectively no batch to space)
+ add_config(TensorShape(1U, 1U, 1U, 4U), { 1U, 1U }, CropInfo(), TensorShape(1U, 1U, 1U, 4U));
+ add_config(TensorShape(8U, 2U, 4U, 3U), { 1U, 1U }, CropInfo(), TensorShape(8U, 2U, 4U, 3U));
+ // Same block size in both x and y
+ add_config(TensorShape(3U, 2U, 1U, 4U), { 2U, 2U }, CropInfo(), TensorShape(6U, 4U, 1U, 1U));
+ add_config(TensorShape(1U, 3U, 2U, 9U), { 3U, 3U }, CropInfo(), TensorShape(3U, 9U, 2U, 1U));
+ // Different block size in x and y
+ add_config(TensorShape(5U, 7U, 7U, 4U), { 2U, 1U }, CropInfo(), TensorShape(10U, 7U, 7U, 2U));
+ add_config(TensorShape(3U, 3U, 1U, 8U), { 1U, 2U }, CropInfo(), TensorShape(3U, 6U, 1U, 4U));
+ add_config(TensorShape(5U, 2U, 2U, 6U), { 3U, 2U }, CropInfo(), TensorShape(15U, 4U, 2U, 1U));
}
};
+/** Relative small shapes that are still large enough to leave room for testing cropping of the output shape
+ */
+class SmallBatchToSpaceLayerWithCroppingDataset final : public BatchToSpaceLayerDataset
+{
+public:
+ SmallBatchToSpaceLayerWithCroppingDataset()
+ {
+ // Crop in both dims
+ add_config(TensorShape(5U, 3U, 2U, 8U), { 2U, 2U }, CropInfo(1U, 1U, 2U, 1U), TensorShape(8U, 3U, 2U, 2U));
+ // Left crop in x dim
+ add_config(TensorShape(1U, 1U, 1U, 20U), { 4U, 5U }, CropInfo(2U, 1U, 0U, 2U), TensorShape(1U, 3U, 1U, 1U));
+ // Left crop in y dim
+ add_config(TensorShape(3U, 1U, 1U, 8U), { 2U, 4U }, CropInfo(0U, 0U, 2U, 1U), TensorShape(6U, 1U, 1U, 1U));
+ }
+};
class LargeBatchToSpaceLayerDataset final : public BatchToSpaceLayerDataset
{
public:
LargeBatchToSpaceLayerDataset()
{
- add_config(TensorShape(64U, 32U, 2U, 4U), TensorShape(2U), TensorShape(128U, 64U, 2U, 1U));
- add_config(TensorShape(128U, 16U, 2U, 16U), TensorShape(2U), TensorShape(512U, 64U, 2U, 1U));
- add_config(TensorShape(16U, 8U, 2U, 8U), TensorShape(4U, 2U), TensorShape(64U, 16U, 2U, 1U));
- add_config(TensorShape(8U, 16U, 2U, 8U), TensorShape(2U, 4U), TensorShape(16U, 64U, 2U, 1U));
+ // Same block size in both x and y
+ add_config(TensorShape(64U, 32U, 2U, 4U), { 2U, 2U }, CropInfo(), TensorShape(128U, 64U, 2U, 1U));
+ add_config(TensorShape(128U, 16U, 2U, 18U), { 3U, 3U }, CropInfo(), TensorShape(384U, 48U, 2U, 2U));
+ // Different block size in x and y
+ add_config(TensorShape(16U, 8U, 2U, 8U), { 4U, 1U }, CropInfo(), TensorShape(64U, 8U, 2U, 2U));
+ add_config(TensorShape(8U, 16U, 2U, 8U), { 2U, 4U }, CropInfo(), TensorShape(16U, 64U, 2U, 1U));
}
};
} // namespace datasets