Skip to content

ARROW-3515: [C++] Introduce NumericTensor class #2759

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
mrkn wants to merge 3 commits into from
Closed

ARROW-3515: [C++] Introduce NumericTensor class #2759

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1646461
Introduce NumericTensor class
mrkn Oct 12, 2018
14fa527
Remove needless cases
mrkn Oct 24, 2018
37f0bb4
Add tests for column-major strides
mrkn Oct 24, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
87 changes: 87 additions & 0 deletions cpp/src/arrow/tensor-test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,4 +104,91 @@ TEST(TestTensor, ZeroDimensionalTensor) {
ASSERT_EQ(t.strides().size(), 1);
}

TEST(TestNumericTensor, ElementAccess) {
std::vector<int64_t> shape = {3, 4};

std::vector<int64_t> values_i64 = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
std::shared_ptr<Buffer> buffer_i64(Buffer::Wrap(values_i64));
NumericTensor<Int64Type> t_i64(buffer_i64, shape);

ASSERT_EQ(1, t_i64.Value({0, 0}));
ASSERT_EQ(5, t_i64.Value({1, 0}));
ASSERT_EQ(6, t_i64.Value({1, 1}));
ASSERT_EQ(11, t_i64.Value({2, 2}));

std::vector<float> values_f32 = {1.1f, 2.1f, 3.1f, 4.1f, 5.1f, 6.1f,
7.1f, 8.1f, 9.1f, 10.1f, 11.1f, 12.1f};
std::shared_ptr<Buffer> buffer_f32(Buffer::Wrap(values_f32));
NumericTensor<FloatType> t_f32(buffer_f32, shape);

ASSERT_EQ(1.1f, t_f32.Value({0, 0}));
ASSERT_EQ(5.1f, t_f32.Value({1, 0}));
ASSERT_EQ(6.1f, t_f32.Value({1, 1}));
ASSERT_EQ(11.1f, t_f32.Value({2, 2}));
}

TEST(TestNumericTensor, ElementAccessWithRowMajorStrides) {
std::vector<int64_t> shape = {3, 4};

const int64_t i64_size = sizeof(int64_t);
std::vector<int64_t> values_i64 = {1, 2, 3, 4, 0, 0, 5, 6, 7,
8, 0, 0, 9, 10, 11, 12, 0, 0};
std::vector<int64_t> strides_i64 = {i64_size * 6, i64_size};
std::shared_ptr<Buffer> buffer_i64(Buffer::Wrap(values_i64));
NumericTensor<Int64Type> t_i64(buffer_i64, shape, strides_i64);

ASSERT_EQ(1, t_i64.Value({0, 0}));
ASSERT_EQ(2, t_i64.Value({0, 1}));
ASSERT_EQ(4, t_i64.Value({0, 3}));
ASSERT_EQ(5, t_i64.Value({1, 0}));
ASSERT_EQ(6, t_i64.Value({1, 1}));
ASSERT_EQ(11, t_i64.Value({2, 2}));

const int64_t f32_size = sizeof(float);
std::vector<float> values_f32 = {1.1f, 2.1f, 3.1f, 4.1f, 0.0f, 0.0f,
5.1f, 6.1f, 7.1f, 8.1f, 0.0f, 0.0f,
9.1f, 10.1f, 11.1f, 12.1f, 0.0f, 0.0f};
std::vector<int64_t> strides_f32 = {f32_size * 6, f32_size};
std::shared_ptr<Buffer> buffer_f32(Buffer::Wrap(values_f32));
NumericTensor<FloatType> t_f32(buffer_f32, shape, strides_f32);

ASSERT_EQ(1.1f, t_f32.Value({0, 0}));
ASSERT_EQ(2.1f, t_f32.Value({0, 1}));
ASSERT_EQ(4.1f, t_f32.Value({0, 3}));
ASSERT_EQ(5.1f, t_f32.Value({1, 0}));
ASSERT_EQ(6.1f, t_f32.Value({1, 1}));
ASSERT_EQ(11.1f, t_f32.Value({2, 2}));
}

TEST(TestNumericTensor, ElementAccessWithColumnMajorStrides) {
std::vector<int64_t> shape = {3, 4};

const int64_t i64_size = sizeof(int64_t);
std::vector<int64_t> values_i64 = {1, 5, 9, 0, 2, 6, 10, 0, 3, 7, 11, 0, 4, 8, 12, 0};
std::vector<int64_t> strides_i64 = {i64_size, i64_size * 4};
std::shared_ptr<Buffer> buffer_i64(Buffer::Wrap(values_i64));
NumericTensor<Int64Type> t_i64(buffer_i64, shape, strides_i64);

ASSERT_EQ(1, t_i64.Value({0, 0}));
ASSERT_EQ(2, t_i64.Value({0, 1}));
ASSERT_EQ(4, t_i64.Value({0, 3}));
ASSERT_EQ(5, t_i64.Value({1, 0}));
ASSERT_EQ(6, t_i64.Value({1, 1}));
ASSERT_EQ(11, t_i64.Value({2, 2}));

const int64_t f32_size = sizeof(float);
std::vector<float> values_f32 = {1.1f, 5.1f, 9.1f, 0.0f, 2.1f, 6.1f, 10.1f, 0.0f,
3.1f, 7.1f, 11.1f, 0.0f, 4.1f, 8.1f, 12.1f, 0.0f};
std::vector<int64_t> strides_f32 = {f32_size, f32_size * 4};
std::shared_ptr<Buffer> buffer_f32(Buffer::Wrap(values_f32));
NumericTensor<FloatType> t_f32(buffer_f32, shape, strides_f32);

ASSERT_EQ(1.1f, t_f32.Value({0, 0}));
ASSERT_EQ(2.1f, t_f32.Value({0, 1}));
ASSERT_EQ(4.1f, t_f32.Value({0, 3}));
ASSERT_EQ(5.1f, t_f32.Value({1, 0}));
ASSERT_EQ(6.1f, t_f32.Value({1, 1}));
ASSERT_EQ(11.1f, t_f32.Value({2, 2}));
}

} // namespace arrow
47 changes: 47 additions & 0 deletions cpp/src/arrow/tensor.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@

#include "arrow/compare.h"
#include "arrow/type.h"
#include "arrow/type_traits.h"
#include "arrow/util/checked_cast.h"
#include "arrow/util/logging.h"

Expand Down Expand Up @@ -121,4 +122,50 @@ Type::type Tensor::type_id() const { return type_->id(); }

bool Tensor::Equals(const Tensor& other) const { return TensorEquals(*this, other); }

// ----------------------------------------------------------------------
// NumericTensor

template <typename TYPE>
NumericTensor<TYPE>::NumericTensor(const std::shared_ptr<Buffer>& data,
const std::vector<int64_t>& shape)
: NumericTensor(data, shape, {}, {}) {}

template <typename TYPE>
NumericTensor<TYPE>::NumericTensor(const std::shared_ptr<Buffer>& data,
const std::vector<int64_t>& shape,
const std::vector<int64_t>& strides)
: NumericTensor(data, shape, strides, {}) {}

template <typename TYPE>
NumericTensor<TYPE>::NumericTensor(const std::shared_ptr<Buffer>& data,
const std::vector<int64_t>& shape,
const std::vector<int64_t>& strides,
const std::vector<std::string>& dim_names)
: Tensor(TypeTraits<TYPE>::type_singleton(), data, shape, strides, dim_names) {}

template <typename TYPE>
int64_t NumericTensor<TYPE>::CalculateValueOffset(
const std::vector<int64_t>& index) const {
int64_t offset = 0;
for (size_t i = 0; i < index.size(); ++i) {
offset += index[i] * strides_[i];
}
return offset;
}

// ----------------------------------------------------------------------
// Instantiate templates

template class ARROW_TEMPLATE_EXPORT NumericTensor<UInt8Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<UInt16Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<UInt32Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<UInt64Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<Int8Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<Int16Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<Int32Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<Int64Type>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<HalfFloatType>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<FloatType>;
template class ARROW_TEMPLATE_EXPORT NumericTensor<DoubleType>;

} // namespace arrow
30 changes: 29 additions & 1 deletion cpp/src/arrow/tensor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,7 @@ class ARROW_EXPORT Tensor {
Tensor(const std::shared_ptr<DataType>& type, const std::shared_ptr<Buffer>& data,
const std::vector<int64_t>& shape, const std::vector<int64_t>& strides);

/// Constructor with strides and dimension names
/// Constructor with non-negative strides and dimension names
Tensor(const std::shared_ptr<DataType>& type, const std::shared_ptr<Buffer>& data,
const std::vector<int64_t>& shape, const std::vector<int64_t>& strides,
const std::vector<std::string>& dim_names);
Expand Down Expand Up @@ -114,6 +114,34 @@ class ARROW_EXPORT Tensor {
ARROW_DISALLOW_COPY_AND_ASSIGN(Tensor);
};

template <typename TYPE>
class ARROW_EXPORT NumericTensor : public Tensor {
public:
using TypeClass = TYPE;
using value_type = typename TypeClass::c_type;

/// Constructor with no dimension names or strides, data assumed to be row-major
NumericTensor(const std::shared_ptr<Buffer>& data, const std::vector<int64_t>& shape);

/// Constructor with non-negative strides
NumericTensor(const std::shared_ptr<Buffer>& data, const std::vector<int64_t>& shape,
const std::vector<int64_t>& strides);

/// Constructor with non-negative strides and dimension names
NumericTensor(const std::shared_ptr<Buffer>& data, const std::vector<int64_t>& shape,
const std::vector<int64_t>& strides,
const std::vector<std::string>& dim_names);

const value_type& Value(const std::vector<int64_t>& index) const {
int64_t offset = CalculateValueOffset(index);
const value_type* ptr = reinterpret_cast<const value_type*>(raw_data() + offset);
return *ptr;
Copy link
Member

@pitrou pitrou Oct 15, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

On some CPU architectures, this assumes the data is naturally aligned. It probably doesn't matter for now.

Copy link
Contributor

@fsaintjacques fsaintjacques Oct 16, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One way to tackle this is to make CalculateValueOffset return an offset in units of TYPE instead of bytes. It'll also make it easier to work with non-fixed-bytes in the future (also possibly correcting the problem with BOOL).

Copy link
Member Author

@mrkn mrkn Oct 16, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I changed this function to return an offset in units of TYPE.

Copy link
Member

@pitrou pitrou Oct 16, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please undo the change (sorry). The strides are in bytes units, so the offset has to be calculated in bytes units. When you're dividing the bytes offset by itemsize, you have no guarantee that there is no remainder (though that will be the common case).

In any case, the change didn't fix the issue, as raw_data() could be misaligned.

Copy link
Member Author

@mrkn mrkn Oct 16, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see.

}

protected:
int64_t CalculateValueOffset(const std::vector<int64_t>& index) const;
};

} // namespace arrow

#endif // ARROW_TENSOR_H