diff --git a/include/aidge/backend/TensorImpl.hpp b/include/aidge/backend/TensorImpl.hpp
index a27f0317c59916facef970a3c1b91704fb485cd4..62f13acb3db81954a4fbb753a3e68e1c5a516402 100644
--- a/include/aidge/backend/TensorImpl.hpp
+++ b/include/aidge/backend/TensorImpl.hpp
@@ -67,19 +67,13 @@ private:
class TensorImpl {
public:
TensorImpl() = delete;
- TensorImpl(const char *backend, DeviceIdx_t device = 0) : mBackend(backend), mDevice(device){};
+ TensorImpl(const char *backend, DeviceIdx_t device, NbElts_t length) : mBackend(backend), mDevice(device), mNbElts(length) {};
/**
* Return the (backend, device) pair for this implementation.
*/
std::pair<std::string, DeviceIdx_t> device() const { return std::make_pair(mBackend, mDevice); }
- /**
- * Set the device ID for current backend.
- * @param device New device ID on current backend.
- */
- virtual void setDevice(DeviceIdx_t device) = 0;
-
/**
* Copy data from the same device.
* @param src Pointer on current implementation device.
@@ -93,30 +87,34 @@ public:
* @param srcDt Source data type.
* @param src Pointer on current implementation device.
* @param length Number of elements to copy.
+ * @param offset Destination offset (in number of elements).
*/
- virtual void copyCast(const void *src, NbElts_t length, const DataType srcDt) = 0;
+ virtual void copyCast(const void *src, const DataType srcDt, NbElts_t length, NbElts_t offset = 0) = 0;
/**
* Copy data from an other device on the same backend.
* @param device (backend, device) pair to copy from. The backend must match current implementation backend.
* @param src Pointer on current implementation backend.
* @param length Number of elements to copy.
+ * @param offset Destination offset (in number of elements).
*/
- virtual void copyFromDevice(const void *src, NbElts_t length, const std::pair<std::string, DeviceIdx_t>& device) = 0;
+ virtual void copyFromDevice(const void *src, const std::pair<std::string, DeviceIdx_t>& device, NbElts_t length, NbElts_t offset = 0) = 0;
/**
* Copy data from host.
* @param src Host pointer to copy from.
* @param length Number of elements to copy.
+ * @param offset Destination offset (in number of elements).
*/
- virtual void copyFromHost(const void *src, NbElts_t length) = 0;
+ virtual void copyFromHost(const void *src, NbElts_t length, NbElts_t offset = 0) = 0;
/**
* Copy data to host.
* @param src Host pointer to copy to.
* @param length Number of elements to copy.
+ * @param offset Source offset (in number of elements).
*/
- virtual void copyToHost(void *dst, NbElts_t length) const = 0;
+ virtual void copyToHost(void *dst, NbElts_t length, NbElts_t offset = 0) const = 0;
/**
* Return the raw device pointer.
@@ -146,8 +144,22 @@ public:
AIDGE_THROW_OR_ABORT(std::runtime_error, "Cannot set raw pointer for backend %s", mBackend);
};
- virtual std::size_t size() const = 0; // Storage size
- virtual std::size_t scalarSize() const = 0; // Size of one scalar (in bytes)
+ /**
+ * Set the size, in number of elements, that must be stored.
+ */
+ void resize(NbElts_t length) {
+ mNbElts = length;
+ }
+
+ /**
+ * Return the number of elements stored.
+ */
+ inline std::size_t size() const noexcept { return mNbElts; }
+
+ /**
+ * Return the size (in bytes) of one element (scalar).
+ */
+ virtual std::size_t scalarSize() const noexcept = 0;
constexpr const char *backend() const { return mBackend; }
virtual ~TensorImpl() = default;
virtual bool operator==(const TensorImpl &othImpl) const = 0;
@@ -156,12 +168,16 @@ public:
* Copy from another backend.
* @param srcImpl Source TensorImpl to copy from.
* @param length Number of elements of size scalarSize() to copy
+ * @param srcOffset Source offset (in number of elements).
+ * @param dstOffset Destination offset (in number of elements).
*/
- void copyFrom(const TensorImpl& srcImpl, NbElts_t length);
+ void copyFrom(const TensorImpl& srcImpl, NbElts_t length, NbElts_t srcOffset = 0, NbElts_t dstOffset = 0);
protected:
const char *mBackend;
- DeviceIdx_t mDevice;
+ const DeviceIdx_t mDevice;
+ /// Number of elements (to be) stored
+ NbElts_t mNbElts;
};
} // namespace Aidge
diff --git a/include/aidge/data/Tensor.hpp b/include/aidge/data/Tensor.hpp
index 8129a900718169861dc2df4213cd3533d1dfe570..658c0b497d9753f1bdfd42a274dbb48970cb6d6b 100644
--- a/include/aidge/data/Tensor.hpp
+++ b/include/aidge/data/Tensor.hpp
@@ -32,15 +32,18 @@ namespace Aidge {
* Contains a pointer to an actual contiguous implementation of data.
*/
class Tensor : public Data,
- public Registrable<Tensor, std::tuple<std::string, DataType>, std::unique_ptr<TensorImpl>(const Tensor &)> {
+ public Registrable<Tensor, std::tuple<std::string, DataType>, std::shared_ptr<TensorImpl>(DeviceIdx_t device, NbElts_t length)> {
private:
DataType mDataType; /** enum to specify data type. */
std::vector<DimSize_t> mDims; /** Dimensions of the tensor. */
- std::unique_ptr<TensorImpl> mImpl; /** Pointer to the actual data implementation. */
+ std::vector<DimSize_t> mStrides; /** Stride dimensions of the tensor. */
+ std::shared_ptr<TensorImpl> mImpl; /** Pointer to the actual data implementation. */
+ std::size_t mImplOffset = 0;
std::shared_ptr<Tensor> mGrad; /** Pointer to the associated gradient Tensor instance. */
// Cached data
std::size_t mSize = 0; /** Number of elements in the Tensor. */
+ bool mContiguous = true;
public:
static constexpr const char *Type = "Tensor";
@@ -57,21 +60,29 @@ class Tensor : public Data,
}
/**
- * @brief Construct a new Tensor object copied from another one.
+ * @brief Construct a new Tensor object from another one (shallow copy).
+ * Data memory is not copied, but shared between the new Tensor and the
+ * initial one.
+ *
* @param otherTensor
*/
- Tensor(const Tensor& otherTensor)
- : Data(Type),
- mDataType(otherTensor.mDataType),
- mDims(otherTensor.mDims),
- mSize(otherTensor.mSize)
- {
- if (otherTensor.hasImpl()) {
- mImpl = Registrar<Tensor>::create({otherTensor.mImpl->backend(), dataType()})(*this);
- mImpl->setDevice(otherTensor.mImpl->device().second);
- // Same backend, same device => directly use copy()
- mImpl->copy(otherTensor.mImpl->rawPtr(), mSize);
+ Tensor(const Tensor&) = default;
+ Tensor(Tensor&&) = default;
+
+ /**
+ * Perform a deep copy of the tensor.
+ */
+ Tensor clone() const {
+ Tensor newTensor(*this);
+ if (!newTensor.isContiguous()) {
+ newTensor.makeContiguous();
+ }
+ else {
+ std::shared_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({mImpl->backend(), mDataType})(mImpl->device().second, mSize);
+ newImpl->copy(mImpl->rawPtr(mImplOffset), mSize);
+ newTensor.setImpl(newImpl);
}
+ return newTensor;
}
/**
@@ -84,7 +95,8 @@ class Tensor : public Data,
: Data(Type),
mDataType(NativeType<T>::type),
mDims({SIZE_0}),
- mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this)),
+ mStrides({1}),
+ mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0)),
mSize(SIZE_0) {
mImpl->copyFromHost(&arr.data[0], SIZE_0);
}
@@ -93,9 +105,9 @@ class Tensor : public Data,
constexpr Tensor &operator=(Array1D<T, SIZE_0> &&arr) {
resize({SIZE_0});
if (!mImpl) {
- mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this);
+ mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0);
}
- mImpl->copyFromHost(&arr.data[0], SIZE_0);
+ mImpl->copyFromHost(&arr.data[0], SIZE_0, mImplOffset);
return *this;
}
@@ -110,7 +122,8 @@ class Tensor : public Data,
: Data(Type),
mDataType(NativeType<T>::type),
mDims({SIZE_0, SIZE_1}),
- mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this)),
+ mStrides({SIZE_1, 1}),
+ mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1)),
mSize(SIZE_0 * SIZE_1) {
mImpl->copyFromHost(&arr.data[0][0], SIZE_0 * SIZE_1);
}
@@ -119,9 +132,9 @@ class Tensor : public Data,
constexpr Tensor &operator=(Array2D<T, SIZE_0, SIZE_1> &&arr) {
resize({SIZE_0, SIZE_1});
if (!mImpl) {
- mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this);
+ mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1);
}
- mImpl->copyFromHost(&arr.data[0][0], SIZE_0 * SIZE_1);
+ mImpl->copyFromHost(&arr.data[0][0], SIZE_0 * SIZE_1, mImplOffset);
return *this;
}
@@ -137,7 +150,8 @@ class Tensor : public Data,
: Data(Type),
mDataType(NativeType<T>::type),
mDims({SIZE_0, SIZE_1, SIZE_2}),
- mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this)),
+ mStrides({SIZE_1 * SIZE_2, SIZE_2, 1}),
+ mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1 * SIZE_2)),
mSize(SIZE_0 * SIZE_1 * SIZE_2) {
mImpl->copyFromHost(&arr.data[0][0][0], SIZE_0 * SIZE_1 * SIZE_2);
}
@@ -146,9 +160,9 @@ class Tensor : public Data,
constexpr Tensor &operator=(Array3D<T, SIZE_0, SIZE_1, SIZE_2> &&arr) {
resize({SIZE_0, SIZE_1, SIZE_2});
if (!mImpl) {
- mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this);
+ mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1 * SIZE_2);
}
- mImpl->copyFromHost(&arr.data[0][0][0], SIZE_0 * SIZE_1 * SIZE_2);
+ mImpl->copyFromHost(&arr.data[0][0][0], SIZE_0 * SIZE_1 * SIZE_2, mImplOffset);
return *this;
}
@@ -165,7 +179,8 @@ class Tensor : public Data,
: Data(Type),
mDataType(NativeType<T>::type),
mDims({SIZE_0, SIZE_1, SIZE_2, SIZE_3}),
- mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this)),
+ mStrides({SIZE_1 * SIZE_2 * SIZE_3, SIZE_2 * SIZE_3, SIZE_3, 1}),
+ mImpl(Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3)),
mSize(SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3) {
mImpl->copyFromHost(&arr.data[0][0][0][0], SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3);
}
@@ -174,33 +189,35 @@ class Tensor : public Data,
constexpr Tensor &operator=(Array4D<T, SIZE_0, SIZE_1, SIZE_2, SIZE_3> &&arr) {
resize({SIZE_0, SIZE_1, SIZE_2, SIZE_3});
if (!mImpl) {
- mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(*this);
+ mImpl = Registrar<Tensor>::create({"cpu", NativeType<T>::type})(0, SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3);
}
- mImpl->copyFromHost(&arr.data[0][0][0][0], SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3);
+ mImpl->copyFromHost(&arr.data[0][0][0][0], SIZE_0 * SIZE_1 * SIZE_2 * SIZE_3, mImplOffset);
return *this;
}
/**
- * @brief Copy dimensions, datatype and data of another Tensor.
+ * @brief Copy dimensions, datatype and data from another Tensor.
+ * If current Tensor already has an implementation, data is copied to the
+ * existing implementation. Tensor backend/device remain untouched.
+ * If current Tensor does not have an implementation, only a shallow copy
+ * is performed and the Tensor will share data with t.
* @param t other Tensor object.
* @return Tensor&
*/
Tensor &operator=(const Tensor &t) {
- resize(t.dims());
- setDataType(t.dataType());
+ resize(t.dims(), t.strides());
+ setDataType(t.dataType(), false); // do not convert existing data
if (t.hasImpl()) {
if (hasImpl()) {
- copyCastFrom(t);
+ copyFrom(t);
}
else {
- mImpl = Registrar<Tensor>::create({t.mImpl->backend(), dataType()})(*this);
- mImpl->setDevice(t.mImpl->device().second);
- // Same backend, same device => directly use copy()
- mImpl->copy(t.mImpl->rawPtr(), mSize);
+ // Perform a shallow copy only
+ setImpl(t.mImpl, t.mImplOffset);
}
}
else {
- mImpl = nullptr;
+ setImpl(nullptr);
}
return *this;
}
@@ -233,17 +250,15 @@ class Tensor : public Data,
if (mImpl->device() != std::make_pair(name, device)) {
// Backend change: create new impl, copy from old to new and replace
// impl
- std::unique_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({name, mDataType})(*this);
- newImpl->setDevice(device);
+ std::shared_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({name, mDataType})(device, mImpl->size());
if (copyFrom) {
- newImpl->copyFrom(*mImpl, size());
+ newImpl->copyFrom(*mImpl, mImpl->size(), mImplOffset, 0);
}
- mImpl = std::move(newImpl);
+ setImpl(newImpl);
}
}
else {
- mImpl = Registrar<Tensor>::create({name, mDataType})(*this);
- mImpl->setDevice(device);
+ mImpl = Registrar<Tensor>::create({name, mDataType})(device, mSize);
}
}
@@ -273,21 +288,32 @@ class Tensor : public Data,
*/
void setDataType(const DataType dt, bool copyCast = true) {
if (mImpl && (dataType() != dt)) {
- std::unique_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({mImpl->backend(), dt})(*this);
+ std::shared_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({mImpl->backend(), dt})(mImpl->device().second, mImpl->size());
if (copyCast) {
- newImpl->copyCast(mImpl->rawPtr(), size(), mDataType);
+ newImpl->copyCast(mImpl->rawPtr(mImplOffset), mDataType, mImpl->size());
}
- mImpl = std::move(newImpl);
+ setImpl(newImpl);
}
mDataType = dt;
}
/**
* @brief Get the Impl object
- * @return constexpr const std::unique_ptr<TensorImpl>&
+ * @return constexpr const std::shared_ptr<TensorImpl>&
*/
- constexpr const std::unique_ptr<TensorImpl> &getImpl() { return mImpl; }
- constexpr const std::unique_ptr<TensorImpl> &getImpl() const { return mImpl; }
+ constexpr const std::shared_ptr<TensorImpl> &getImpl() const { return mImpl; }
+ constexpr std::size_t getImplOffset() const { return mImplOffset; }
+
+ /**
+ * @brief Set the Impl object
+ *
+ * @param impl New impl shared pointer
+ * @param implOffset Storage offset in this new impl for this Tensor
+ */
+ void setImpl(std::shared_ptr<TensorImpl> impl, std::size_t implOffset = 0) {
+ mImpl = impl;
+ mImplOffset = implOffset;
+ }
/**
* @brief Return if an implementaiton has been associated.
@@ -319,6 +345,18 @@ class Tensor : public Data,
*/
constexpr const std::vector<DimSize_t> &dims() const { return mDims; }
+ /**
+ * @brief Get strides of the Tensor object.
+ * @return constexpr const std::vector<DimSize_t>&
+ */
+ constexpr const std::vector<DimSize_t> &strides() const { return mStrides; }
+
+ /**
+ * @brief Return true if Tensor is contiguous in memory.
+ * @return bool
+ */
+ constexpr bool isContiguous() const { return mContiguous; }
+
/**
* @brief Get the number of elements in the Tensor object.
* @return constexpr std::size_t
@@ -350,10 +388,49 @@ class Tensor : public Data,
* one, all previous data is invalided. Otherwise, previous data may or may
* not remain valid, depending on the backend implementation.
* @param dims New dimensions
+ * @param strides Stride of the tensor (if not specified, "nested" stride is used)
*/
- void resize(const std::vector<DimSize_t> &dims) {
- mDims = dims;
- computeSize();
+ void resize(const std::vector<DimSize_t> &dims, std::vector<DimSize_t> strides = std::vector<DimSize_t>()) {
+ bool checkContiguous = true;
+ if (strides.empty()) {
+ strides.resize(dims.size());
+ size_t expectedStride = 1;
+ for (int dim = dims.size() - 1; dim >= 0; --dim) {
+ strides[dim] = expectedStride;
+ expectedStride*= dims[dim];
+ }
+ checkContiguous = false;
+ }
+ else {
+ AIDGE_ASSERT(strides.size() == dims.size(), "Number of strides must match number of dims");
+ }
+
+ if (mImpl.use_count() > 1) {
+ // Here we could also create a new storage for this tensor in this case
+ // But, is it more likely that the user really wants this, or that he did a mistake?
+ AIDGE_ASSERT(dims == mDims && strides == mStrides, "Cannot resize Tensor with shared storage");
+ }
+ else {
+ mDims = dims;
+ mStrides = strides;
+
+ mContiguous = true;
+ if (checkContiguous) {
+ size_t expectedStride = 1;
+ for (int dim = dims.size() - 1; dim >= 0; --dim) {
+ if (strides[dim] != expectedStride) {
+ mContiguous = false;
+ break;
+ }
+ expectedStride*= dims[dim];
+ }
+ }
+
+ computeSize();
+ if (mImpl) {
+ mImpl->resize(mSize);
+ }
+ }
}
/**
@@ -367,25 +444,25 @@ class Tensor : public Data,
const expectedType& get(std::size_t idx) const {
AIDGE_ASSERT(NativeType<expectedType>::type == mDataType, "wrong data type");
AIDGE_ASSERT(idx < mSize, "idx out of range");
- return *reinterpret_cast<expectedType *>(mImpl->hostPtr(idx));
+ return *reinterpret_cast<expectedType *>(mImpl->hostPtr(mImplOffset + idx));
}
template <typename expectedType>
const expectedType& get(std::vector<std::size_t> coordIdx) const {
- return get<expectedType>(getIdx(coordIdx));
+ return get<expectedType>(getStorageIdx(coordIdx));
}
template <typename expectedType>
void set(std::size_t idx, expectedType value){
AIDGE_ASSERT(NativeType<expectedType>::type == mDataType, "wrong data type");
AIDGE_ASSERT(idx < mSize, "idx out of range");
- expectedType* dataPtr = static_cast<expectedType*>(mImpl->hostPtr(idx));
+ expectedType* dataPtr = static_cast<expectedType*>(mImpl->hostPtr(mImplOffset + idx));
*dataPtr = value;
}
template <typename expectedType>
void set(std::vector<std::size_t> coordIdx, expectedType value){
- set<expectedType>(getIdx(coordIdx), value);
+ set<expectedType>(getStorageIdx(coordIdx), value);
}
@@ -449,9 +526,9 @@ class Tensor : public Data,
for (; dimVals[dim] < static_cast<std::size_t>(dims()[dim]); ++dimVals[dim]) {
res += spaceString + "{";
for (DimSize_t j = 0; j < dims()[dim + 1] - 1; ++j) {
- res += " " + ptrToString(mDataType, mImpl->hostPtr(), counter++) + ",";
+ res += " " + ptrToString(mDataType, mImpl->hostPtr(mImplOffset), counter++) + ",";
}
- res += " " + ptrToString(mDataType, mImpl->hostPtr(), counter++) + "}";
+ res += " " + ptrToString(mDataType, mImpl->hostPtr(mImplOffset), counter++) + "}";
if (dimVals[dim] < static_cast<std::size_t>(dims()[dim] - 1)) {
res += ",";
}
@@ -471,7 +548,7 @@ class Tensor : public Data,
} else {
res += "{";
for (DimSize_t j = 0; j < dims()[0]; ++j) {
- res += " " + ptrToString(mDataType, mImpl->hostPtr(), j) + ((j < dims()[0]-1) ? "," : " ");
+ res += " " + ptrToString(mDataType, mImpl->hostPtr(mImplOffset), j) + ((j < dims()[0]-1) ? "," : " ");
}
}
res += "}";
@@ -493,6 +570,7 @@ class Tensor : public Data,
/**
* @brief From the the 1D contiguous index, return the coordinate of an element in the tensor.
+ * Beware: do not use this function with the storage index!
*
* @param flatIdx 1D contiguous index of the value considering a flatten, contiguous, tensor.
* @return std::vector<DimSize_t>
@@ -512,6 +590,8 @@ class Tensor : public Data,
* @brief From the coordinate returns the 1D contiguous index of an element in the tensor.
* If the number of coordinates is inferior to the number of dimensions,
* the remaining coordinates are assumed to be 0.
+ * Beware: the contiguous index will only correspond to the storage index
+ * if the tensor is contiguous!
*
* @param coordIdx Coordinate to an element in the tensor
* @return DimSize_t Contiguous index
@@ -527,6 +607,51 @@ class Tensor : public Data,
return flatIdx + coordIdx[i];
}
+ /**
+ * @brief From the coordinate returns the 1D storage index of an element in the tensor.
+ * If the number of coordinates is inferior to the number of dimensions,
+ * the remaining coordinates are assumed to be 0.
+ *
+ * @param coordIdx Coordinate to an element in the tensor
+ * @return DimSize_t Storage index
+ */
+ std::size_t getStorageIdx(const std::vector<std::size_t>& coordIdx) const {
+ AIDGE_ASSERT(coordIdx.size() <= mDims.size(), "Coordinates does not match number of dimensions");
+ return std::inner_product(coordIdx.begin(), coordIdx.end(), mStrides.begin(), DimSize_t(0));
+ }
+
+ /**
+ * Returns a sub-tensor with one or more dimension less.
+ * For instance, t.extract({1}) on a CHW tensor will return the HW tensor
+ * of channel #1.
+ * Likewise, t.extract({0, 1}) on a NCHW tensor will return the HW tensor
+ * of batch #0 and channel #1.
+ * No memory copy is performed, the returned tensor does not own the memory.
+ * If the number of coordinates matches the number of dimensions, an empty
+ * tensor is returned.
+ * It current tensor was contiguous, the returned tensor is garanteed to be
+ * contiguous as well.
+ *
+ * @param coordIdx Coordinates of the sub-tensor to extract
+ * @return Tensor Sub-tensor.
+ */
+ Tensor extract(const std::vector<std::size_t>& coordIdx) const;
+
+ /**
+ * Returns a sub-tensor at some coordinate and with some dimension.
+ *
+ * @param coordIdx First coordinates of the sub-tensor to extract
+ * @param dims Dimensions of the sub-tensor to extract
+ * @return Tensor Sub-tensor.
+ */
+ Tensor extract(const std::vector<std::size_t>& coordIdx, const std::vector<std::size_t>& dims) const;
+
+ /**
+ * Make the tensor's storage contiguous, if it is not already the case.
+ * If not contiguous, a new memory space is allocated.
+ */
+ void makeContiguous();
+
/**
* Copy-cast data from a Tensor on the same device.
* If current tensor backend/device is set and is different from src, an
@@ -572,6 +697,20 @@ class Tensor : public Data,
copyCastFrom(src, movedSrc);
}
+ /**
+ * Return a reference to a Tensor that is garanteed to be contiguous:
+ * - itself, if already contiguous;
+ * - the provided Tensor, overwritten with the copied data.
+ * The data type, backend and device stay the same.
+ * @param fallback A shared_ptr to Tensor ready to be overwritten if necessary.
+ * The shared_ptr does not need to be initialized. No new memory allocation
+ * will occur if fallback has already been allocated with the right
+ * type/size/device.
+ * @return Reference to either itself or to fallback.
+ */
+ Tensor& refContiguous(std::shared_ptr<Tensor>& fallback);
+ const Tensor& refContiguous(std::shared_ptr<Tensor>& fallback) const;
+
/**
* Return a reference to a Tensor casted to the desired data type:
* - itself, if already at the right data type;
@@ -642,6 +781,43 @@ class Tensor : public Data,
return refCastFrom(fallback, targetReqs.dataType(), device.first, device.second);
}
+ /**
+ * Return a reference to a Tensor on desired data type and backend/device:
+ * - itself, if already with the right characteristics;
+ * - the provided Tensor, overwritten with the right characteristics.
+ * NOTE: no data is copy-casted. If it was so in a previous refCastFrom() on
+ * the same fallback, it remains valid, otherwise, data is invalid.
+ * @param fallback A shared_ptr to Tensor ready to be overwritten if necessary.
+ * The shared_ptr does not need to be initialized. No new memory allocation
+ * will occur if fallback has already been allocated with the right
+ * type/size/device.
+ * @param dt The desired data type.
+ * @param backend The desired backend.
+ * @param device The desired device.
+ * @return Reference to either itself or to fallback.
+ */
+ Tensor& ref(std::shared_ptr<Tensor>& fallback, const Aidge::DataType& dt, const std::string &backend, DeviceIdx_t device = 0);
+ const Tensor& ref(std::shared_ptr<Tensor>& fallback, const Aidge::DataType& dt, const std::string &backend, DeviceIdx_t device = 0) const;
+
+ /**
+ * Return a reference to a Tensor with same characteristics
+ * (data type, backend/device) as targetReqs Tensor:
+ * - itself, if already with the right characteristics;
+ * - the provided Tensor, overwritten with the right characteristics.
+ * NOTE: no data is copy-casted. If it was so in a previous refCastFrom() on
+ * the same fallback, it remains valid, otherwise, data is invalid.
+ * @param fallback A shared_ptr to Tensor ready to be overwritten if necessary.
+ * The shared_ptr does not need to be initialized. No new memory allocation
+ * will occur if fallback has already been allocated with the right
+ * type/size/device.
+ * @param targetReqs Tensor with the desired target characteristics.
+ * @return Reference to either itself or to fallback.
+ */
+ Tensor& ref(std::shared_ptr<Tensor>& fallback, const Tensor& targetReqs) {
+ const auto& device = targetReqs.getImpl()->device();
+ return ref(fallback, targetReqs.dataType(), device.first, device.second);
+ }
+
private:
///\bug not protected against overflow
void computeSize() {
diff --git a/include/aidge/operator/Gather.hpp b/include/aidge/operator/Gather.hpp
index 20082eed28825ade9d62fb5d4e081840d3bd4442..f6647f99151304d0cf083aed109cc642c9f1ecc2 100644
--- a/include/aidge/operator/Gather.hpp
+++ b/include/aidge/operator/Gather.hpp
@@ -27,25 +27,26 @@
#include "aidge/utils/Types.h"
namespace Aidge {
-enum class GatherAttr { Axis };
+enum class GatherAttr { Indices, GatheredShape, Axis };
class Gather_Op : public OperatorTensor,
public Registrable<Gather_Op,
std::string,
std::unique_ptr<OperatorImpl>(const Gather_Op&)>,
- public StaticAttributes<GatherAttr, int> {
+ public StaticAttributes<GatherAttr, std::vector<std::int64_t>, std::vector<DimSize_t>, std::int64_t> {
public:
static const std::string Type;
Gather_Op() = delete;
-
- using Attributes_ = StaticAttributes<GatherAttr, int>;
+ using Attributes_ = StaticAttributes<GatherAttr, std::vector<std::int64_t>, std::vector<DimSize_t>, std::int64_t>;
template <GatherAttr e> using attr = typename Attributes_::template attr<e>;
- Gather_Op(int axis)
- : OperatorTensor(Type, 2, 0, 1),
+ Gather_Op(const std::vector<std::int64_t>& indices, const std::vector<DimSize_t>& gatheredShape, std::int64_t axis)
+ : OperatorTensor(Type, 1, 0, 1),
Attributes_(
+ attr<GatherAttr::Indices>(indices),
+ attr<GatherAttr::GatheredShape>(gatheredShape),
attr<GatherAttr::Axis>(axis))
{}
@@ -76,21 +77,21 @@ public:
}
static const std::vector<std::string> getInputsName(){
- return {"data_input", "indexes"};
+ return {"data_input"};
}
static const std::vector<std::string> getOutputsName(){
return {"data_output"};
}
};
-inline std::shared_ptr<Node> Gather(int axis = 0, const std::string& name = "") {
- return std::make_shared<Node>(std::make_shared<Gather_Op>(axis), name);
+inline std::shared_ptr<Node> Gather( const std::vector<std::int64_t>& indices, const std::vector<DimSize_t>& gatheredShape, std::int64_t axis = 0, const std::string& name = "") {
+ return std::make_shared<Node>(std::make_shared<Gather_Op>(indices, gatheredShape, axis), name);
}
} // namespace Aidge
namespace {
template <>
-const char *const EnumStrings<Aidge::GatherAttr>::data[] = {"Axis"};
+const char *const EnumStrings<Aidge::GatherAttr>::data[] = {"Indices", "GatheredShape", "Axis"};
}
#endif /* AIDGE_CORE_OPERATOR_GATHER_H_ */
diff --git a/include/aidge/operator/Slice.hpp b/include/aidge/operator/Slice.hpp
index 12a7425f3339b7fbc0ae010639aacf23d97b0f5f..4a073bc525640846c28d718d09741a67d499830e 100644
--- a/include/aidge/operator/Slice.hpp
+++ b/include/aidge/operator/Slice.hpp
@@ -29,17 +29,17 @@ enum class SliceAttr { Starts, Ends, Axes };
class Slice_Op
: public OperatorTensor,
public Registrable<Slice_Op, std::string, std::unique_ptr<OperatorImpl>(const Slice_Op &)>,
- public StaticAttributes<SliceAttr, std::vector<std::int32_t>, std::vector<std::int32_t>, std::vector<std::int32_t>> {
+ public StaticAttributes<SliceAttr, std::vector<std::int64_t>, std::vector<std::int64_t>, std::vector<std::int64_t>> {
public:
static const std::string Type;
Slice_Op() = delete;
- using Attributes_ = StaticAttributes<SliceAttr, std::vector<std::int32_t>, std::vector<std::int32_t>, std::vector<std::int32_t>>;
+ using Attributes_ = StaticAttributes<SliceAttr, std::vector<std::int64_t>, std::vector<std::int64_t>, std::vector<std::int64_t>>;
template <SliceAttr e>
using attr = typename Attributes_::template attr<e>;
- Slice_Op(const std::vector<std::int32_t>& starts, const std::vector<std::int32_t>& ends, const std::vector<std::int32_t>& axes)
+ Slice_Op(const std::vector<std::int64_t>& starts, const std::vector<std::int64_t>& ends, const std::vector<std::int64_t>& axes)
: OperatorTensor(Type, 1, 0, 1),
Attributes_(attr<SliceAttr::Starts>(starts),
attr<SliceAttr::Ends>(ends),
@@ -94,9 +94,9 @@ public:
* @param name Name of the Operator.
* @return std::shared_ptr<Node> A Node containing the Operator.
*/
-inline std::shared_ptr<Node> Slice(const std::vector<std::int32_t> starts,
- const std::vector<std::int32_t> ends,
- const std::vector<std::int32_t> axes,
+inline std::shared_ptr<Node> Slice(const std::vector<std::int64_t> starts,
+ const std::vector<std::int64_t> ends,
+ const std::vector<std::int64_t> axes,
const std::string &name = "") {
// FIXME: properly handle default w&b initialization in every cases
return std::make_shared<Node>(std::make_shared<Slice_Op>(starts, ends, axes), name);
diff --git a/python_binding/data/pybind_Tensor.cpp b/python_binding/data/pybind_Tensor.cpp
index 6d6f20ebe9377ce177d936c00f097fee76954bd9..9fbf08d0b782b6f39b2bef3d0b3ab918f6789ac0 100644
--- a/python_binding/data/pybind_Tensor.cpp
+++ b/python_binding/data/pybind_Tensor.cpp
@@ -30,7 +30,7 @@ void addCtor(py::class_<Tensor,
Data,
Registrable<Tensor,
std::tuple<std::string, DataType>,
- std::unique_ptr<TensorImpl>(const Tensor&)>>& mTensor){
+ std::shared_ptr<TensorImpl>(DeviceIdx_t device, NbElts_t length)>>& mTensor){
mTensor.def(py::init([](
py::array_t<T, py::array::c_style | py::array::forcecast> b,
std::string backend = "cpu") {
@@ -60,16 +60,16 @@ void addCtor(py::class_<Tensor,
void init_Tensor(py::module& m){
py::class_<Registrable<Tensor,
std::tuple<std::string, DataType>,
- std::unique_ptr<TensorImpl>(const Tensor&)>,
+ std::shared_ptr<TensorImpl>(DeviceIdx_t device, NbElts_t length)>,
std::shared_ptr<Registrable<Tensor,
std::tuple<std::string, DataType>,
- std::unique_ptr<TensorImpl>(const Tensor&)>>>(m,"TensorRegistrable");
+ std::shared_ptr<TensorImpl>(DeviceIdx_t device, NbElts_t length)>>>(m,"TensorRegistrable");
py::class_<Tensor, std::shared_ptr<Tensor>,
Data,
Registrable<Tensor,
std::tuple<std::string, DataType>,
- std::unique_ptr<TensorImpl>(const Tensor&)>> pyClassTensor
+ std::shared_ptr<TensorImpl>(DeviceIdx_t device, NbElts_t length)>> pyClassTensor
(m,"Tensor", py::multiple_inheritance(), py::buffer_protocol());
pyClassTensor.def(py::init<>())
@@ -78,7 +78,7 @@ void init_Tensor(py::module& m){
.def("dims", (const std::vector<DimSize_t>& (Tensor::*)()const) &Tensor::dims)
.def("dtype", &Tensor::dataType)
.def("size", &Tensor::size)
- .def("resize", (void (Tensor::*)(const std::vector<DimSize_t>&)) &Tensor::resize)
+ .def("resize", (void (Tensor::*)(const std::vector<DimSize_t>&, std::vector<DimSize_t>)) &Tensor::resize)
.def("has_impl", &Tensor::hasImpl)
.def("get_coord", &Tensor::getCoord)
.def("get_idx", &Tensor::getIdx)
@@ -120,7 +120,7 @@ void init_Tensor(py::module& m){
}
})
.def_buffer([](Tensor& b) -> py::buffer_info {
- const std::unique_ptr<TensorImpl>& tensorImpl = b.getImpl();
+ const std::shared_ptr<TensorImpl>& tensorImpl = b.getImpl();
std::vector<size_t> dims;
std::vector<size_t> strides;
diff --git a/python_binding/operator/pybind_Gather.cpp b/python_binding/operator/pybind_Gather.cpp
index 92a93c3005fca5ff08ab255d2ded530424883a0d..f0d55e2f40bd89269c96564cea6b5a002b477b8b 100644
--- a/python_binding/operator/pybind_Gather.cpp
+++ b/python_binding/operator/pybind_Gather.cpp
@@ -24,6 +24,6 @@ void init_Gather(py::module& m) {
.def("get_outputs_name", &Gather_Op::getOutputsName)
.def("attributes_name", &Gather_Op::staticGetAttrsName);
- m.def("Gather", &Gather, py::arg("axis"), py::arg("name") = "");
+ m.def("Gather", &Gather, py::arg("indices"), py::arg("gathered_shape"), py::arg("axis"), py::arg("name") = "");
}
} // namespace Aidge
diff --git a/src/backend/TensorImpl.cpp b/src/backend/TensorImpl.cpp
index 3982ee1fed9c9198b539bf9a28edd461992b791f..ee2f82a9cf847bfc6fe51e8d8b621e53a4c93cf4 100644
--- a/src/backend/TensorImpl.cpp
+++ b/src/backend/TensorImpl.cpp
@@ -14,23 +14,23 @@
#include "aidge/utils/Types.h"
#include "aidge/utils/ErrorHandling.hpp"
-void Aidge::TensorImpl::copyFrom(const TensorImpl& srcImpl, NbElts_t length) {
- if (&srcImpl == this) {
+void Aidge::TensorImpl::copyFrom(const TensorImpl& srcImpl, NbElts_t length, NbElts_t srcOffset, NbElts_t dstOffset) {
+ if (&srcImpl == this && srcOffset == dstOffset) {
return;
}
if (srcImpl.device() != device()) {
if (srcImpl.backend() == backend()) {
// Same backend, but different device
- copyFromDevice(srcImpl.rawPtr(), length, srcImpl.device());
+ copyFromDevice(srcImpl.rawPtr(srcOffset), srcImpl.device(), length, dstOffset);
}
else if (srcImpl.hostPtr() != nullptr) {
// Different backend, but input is valid on host
- copyFromHost(srcImpl.hostPtr(), length);
+ copyFromHost(srcImpl.hostPtr(srcOffset), length, dstOffset);
}
else if (hostPtr() != nullptr) {
// Different backend, but dst is valid on host
- srcImpl.copyToHost(hostPtr(), length);
+ srcImpl.copyToHost(hostPtr(srcOffset), length, dstOffset);
}
else {
// No direct link possible from src to dst device
@@ -40,12 +40,12 @@ void Aidge::TensorImpl::copyFrom(const TensorImpl& srcImpl, NbElts_t length) {
// - There is currently no concrete use case
// - Just providing a pointer would be unsafe (risk of buffer overflow...)
auto tmpHostBuffer = std::unique_ptr<char[]>(new char[scalarSize() * length]);
- srcImpl.copyToHost(tmpHostBuffer.get(), length);
- copyFromHost(tmpHostBuffer.get(), length);
+ srcImpl.copyToHost(tmpHostBuffer.get(), length, srcOffset);
+ copyFromHost(tmpHostBuffer.get(), length, dstOffset);
}
}
else {
// Same device: simple copy on device
- copy(srcImpl.rawPtr(), length);
+ copy(srcImpl.rawPtr(srcOffset), length, dstOffset);
}
}
diff --git a/src/data/Tensor.cpp b/src/data/Tensor.cpp
index da0c626d78dd1cc4452bfc07bf6c6a7f58b8d1e4..d45dee5639a6bc082871e1110657392fb97c15ec 100644
--- a/src/data/Tensor.cpp
+++ b/src/data/Tensor.cpp
@@ -13,11 +13,72 @@
#include "aidge/utils/Types.h"
#include "aidge/utils/ErrorHandling.hpp"
+Aidge::Tensor Aidge::Tensor::extract(const std::vector<std::size_t>& coordIdx) const {
+ AIDGE_ASSERT(isContiguous(), "Tensor must be contiguous");
+ AIDGE_ASSERT(coordIdx.size() <= mDims.size(), "Number of coordinates is higher than number of dimensions");
+
+ Tensor subTensor(mDataType);
+ subTensor.resize(std::vector<size_t>(mDims.begin() + coordIdx.size(), mDims.end()),
+ std::vector<size_t>(mStrides.begin() + coordIdx.size(), mStrides.end()));
+ subTensor.setBackend(mImpl->backend(), mImpl->device().second);
+ subTensor.setImpl(mImpl, mImplOffset + getStorageIdx(coordIdx));
+ return subTensor;
+}
+
+Aidge::Tensor Aidge::Tensor::extract(const std::vector<std::size_t>& coordIdx, const std::vector<std::size_t>& dims) const {
+ AIDGE_ASSERT(isContiguous(), "Tensor must be contiguous");
+ AIDGE_ASSERT(coordIdx.size() == mDims.size(), "Coordinates does not match number of dimensions");
+
+ Tensor subTensor(mDataType);
+ subTensor.resize(dims, mStrides);
+ subTensor.setBackend(mImpl->backend(), mImpl->device().second);
+ subTensor.setImpl(mImpl, mImplOffset + getStorageIdx(coordIdx));
+ return subTensor;
+}
+
+void Aidge::Tensor::makeContiguous() {
+ if (!mImpl || isContiguous()) {
+ return;
+ }
+
+ // Block so that mImpl ref count is 1 for resize()
+ {
+ // Create a new storage that will be contiguous
+ std::shared_ptr<TensorImpl> newImpl = Registrar<Tensor>::create({mImpl->backend(), mDataType})(mImpl->device().second, mSize);
+ // Copy elements from old to new storage
+ size_t idx = 0;
+ while (idx < mSize) {
+ const size_t storageIdx = getStorageIdx(getCoord(idx));
+
+ // Determine the size of the contiguous chunk
+ size_t copySize = 1;
+ while (idx + copySize < mSize &&
+ getStorageIdx(getCoord(idx + copySize)) == storageIdx + copySize)
+ {
+ ++copySize;
+ }
+
+ // Perform a single copy for the contiguous chunk
+ newImpl->copy(mImpl->rawPtr(mImplOffset + storageIdx), copySize, idx);
+
+ // Move to the next index after the contiguous chunk
+ idx += copySize;
+ }
+ // Replace old storage by new, contiguous, storage
+ setImpl(newImpl);
+ }
+
+ // Resize tensor without strides => tensor is now contiguous
+ resize(mDims);
+}
+
void Aidge::Tensor::copyCast(const Tensor& src) {
if (&src == this) {
return;
}
+ AIDGE_ASSERT(src.isContiguous(), "cannot copy-cast non-contiguous tensor");
+
// Current Tensor has necessarily a data type, but may not have backend
if (!getImpl()) {
// If no backend was set for the current tensor, use the same as src
@@ -27,7 +88,7 @@ void Aidge::Tensor::copyCast(const Tensor& src) {
resize(src.dims());
AIDGE_ASSERT(src.getImpl()->device() == getImpl()->device(), "cannot copy-cast from a different backend/device");
- getImpl()->copyCast(src.getImpl()->rawPtr(), src.size(), src.dataType());
+ getImpl()->copyCast(src.getImpl()->rawPtr(src.mImplOffset), src.dataType(), src.size(), mImplOffset);
}
void Aidge::Tensor::copyFrom(const Tensor& src) {
@@ -35,6 +96,8 @@ void Aidge::Tensor::copyFrom(const Tensor& src) {
return;
}
+ AIDGE_ASSERT(src.isContiguous(), "cannot copy from non-contiguous tensor");
+
// Current Tensor has necessarily a data type, but may not have backend
if (!getImpl()) {
// If no backend was set for the current tensor, use the same as src
@@ -44,7 +107,7 @@ void Aidge::Tensor::copyFrom(const Tensor& src) {
resize(src.dims());
AIDGE_ASSERT(src.dataType() == dataType(), "cannot copy from a different data type");
- getImpl()->copyFrom(*(src.getImpl()), src.size());
+ getImpl()->copyFrom(*(src.getImpl()), src.size(), src.mImplOffset, mImplOffset);
}
void Aidge::Tensor::copyCastFrom(const Tensor& src, std::shared_ptr<Tensor>& movedSrcPtr) {
@@ -52,6 +115,8 @@ void Aidge::Tensor::copyCastFrom(const Tensor& src, std::shared_ptr<Tensor>& mov
return;
}
+ AIDGE_ASSERT(src.isContiguous(), "cannot copy-cast from non-contiguous tensor");
+
// Current Tensor has necessarily a data type, but may not have backend
if (!getImpl()) {
// If no backend was set for the current tensor, use the same as src
@@ -65,12 +130,35 @@ void Aidge::Tensor::copyCastFrom(const Tensor& src, std::shared_ptr<Tensor>& mov
const auto device = getImpl()->device();
const Tensor& movedSrc = src.refFrom(movedSrcPtr, device.first, device.second);
// Second, copy-cast data (necessary)
- getImpl()->copyCast(movedSrc.getImpl()->rawPtr(), movedSrc.size(), movedSrc.dataType());
+ getImpl()->copyCast(movedSrc.getImpl()->rawPtr(movedSrc.mImplOffset), movedSrc.dataType(), movedSrc.size(), mImplOffset);
}
else {
// Directly copy, no conversion necessary
// Avoid making a double copy if both data type and device are the same
- getImpl()->copyFrom(*(src.getImpl()), src.size());
+ getImpl()->copyFrom(*(src.getImpl()), src.size(), src.mImplOffset, mImplOffset);
+ }
+}
+
+Aidge::Tensor& Aidge::Tensor::refContiguous(std::shared_ptr<Tensor>& fallback) {
+ // Scott Meyers' solution to avoid code duplication
+ return const_cast<Tensor&>(static_cast<const Tensor&>(*this).refContiguous(fallback));
+}
+
+const Aidge::Tensor& Aidge::Tensor::refContiguous(std::shared_ptr<Tensor>& fallback) const {
+ AIDGE_ASSERT(getImpl(), "no backend was set for tensor, cannot refCast() it");
+
+ if (isContiguous()) {
+ return *this;
+ }
+ else {
+ if (this != fallback.get()) {
+ // Shallow copy to fallback
+ *fallback = *this;
+ }
+
+ // Make fallback contiguous
+ fallback->makeContiguous();
+ return *fallback;
}
}
@@ -91,6 +179,8 @@ const Aidge::Tensor& Aidge::Tensor::refCast(std::shared_ptr<Tensor>& fallback, c
fallback->setDataType(dt);
}
else {
+ AIDGE_ASSERT(isContiguous(), "cannot refCast non-contiguous tensor");
+
if (!fallback) {
fallback = std::make_shared<Tensor>(dt);
}
@@ -101,7 +191,7 @@ const Aidge::Tensor& Aidge::Tensor::refCast(std::shared_ptr<Tensor>& fallback, c
const auto device = getImpl()->device();
fallback->setBackend(device.first, device.second, false); // don't keep previous data (no copy)
fallback->resize(dims());
- fallback->getImpl()->copyCast(getImpl()->rawPtr(), size(), dataType());
+ fallback->getImpl()->copyCast(getImpl()->rawPtr(mImplOffset), dataType(), size(), fallback->mImplOffset);
}
return *fallback;
}
@@ -124,6 +214,8 @@ const Aidge::Tensor& Aidge::Tensor::refFrom(std::shared_ptr<Tensor>& fallback, c
fallback->setBackend(backend, device);
}
else {
+ AIDGE_ASSERT(isContiguous(), "cannot refFrom non-contiguous tensor");
+
if (!fallback) {
fallback = std::make_shared<Tensor>(dataType());
}
@@ -133,8 +225,34 @@ const Aidge::Tensor& Aidge::Tensor::refFrom(std::shared_ptr<Tensor>& fallback, c
fallback->setBackend(backend, device, false); // don't keep previous data (no copy)
fallback->resize(dims());
- fallback->getImpl()->copyFrom(*getImpl(), size());
+ fallback->getImpl()->copyFrom(*getImpl(), size(), mImplOffset, fallback->mImplOffset);
+ }
+ return *fallback;
+ }
+}
+
+Aidge::Tensor& Aidge::Tensor::ref(std::shared_ptr<Tensor>& fallback, const Aidge::DataType& dt, const std::string &backend, DeviceIdx_t device) {
+ // Scott Meyers' solution to avoid code duplication
+ return const_cast<Tensor&>(static_cast<const Tensor&>(*this).ref(fallback, dt, backend, device));
+}
+
+const Aidge::Tensor& Aidge::Tensor::ref(std::shared_ptr<Tensor>& fallback, const Aidge::DataType& dt, const std::string &backend, DeviceIdx_t device) const {
+ AIDGE_ASSERT(getImpl(), "no backend was set for tensor, cannot ref() it");
+
+ if (dt == dataType() && std::make_pair(backend, device) == getImpl()->device()) {
+ return *this;
+ }
+ else {
+ // Change fallback type, backend & device, without any data copy
+ if (!fallback) {
+ fallback = std::make_shared<Tensor>(dt);
}
+ else {
+ fallback->setDataType(dt, false); // don't keep previous data (no copy)
+ }
+
+ fallback->setBackend(backend, device, false); // don't keep previous data (no copy)
+ fallback->resize(dims());
return *fallback;
}
}
diff --git a/src/operator/Gather.cpp b/src/operator/Gather.cpp
index 30804994b6084a5a5558f106a38a6087e54471bc..b5f9d738a0280b3bacdb2ce201c8303b2b4d0a1f 100644
--- a/src/operator/Gather.cpp
+++ b/src/operator/Gather.cpp
@@ -9,8 +9,8 @@
*
********************************************************************************/
-#include <cassert>
#include <cstddef>
+#include <cstdint>
#include <string>
#include <vector>
@@ -22,18 +22,26 @@ const std::string Aidge::Gather_Op::Type = "Gather";
void Aidge::Gather_Op::computeOutputDims() {
// check inputs have been associated
- if (!getInput(0) || !getInput(1)) {
- AIDGE_THROW_OR_ABORT(std::runtime_error, "At least one input was not connected");
+ if (!getInput(0)) {
+ AIDGE_THROW_OR_ABORT(std::runtime_error, "Input was not connected");
}
- if (getInput(1)->nbDims()!=2){
- AIDGE_THROW_OR_ABORT(std::runtime_error, "Indices input must be a 2D Tensor");
- }
+ if (!getInput(0)->empty()) {
+ std::vector<DimSize_t> outDims = getInput(0)->dims();
+ const std::vector<DimSize_t> gatheredShape = this->template getAttr<GatherAttr::GatheredShape>();
+ // TODO: check indices and gatheredShape
+
+ const std::int64_t axisIdx = this->template getAttr<GatherAttr::Axis>() >= 0 ?
+ this->template getAttr<GatherAttr::Axis>() :
+ this->template getAttr<GatherAttr::Axis>() + outDims.size();
+ outDims.erase(outDims.begin() + static_cast<std::size_t>(axisIdx));
+ if (!gatheredShape.empty())
+ {
+ outDims.insert(outDims.cbegin() + static_cast<std::size_t>(axisIdx),
+ gatheredShape.cbegin(),
+ gatheredShape.cend());
+ }
- std::vector<DimSize_t> outDims = getInput(0)->dims();
- std::vector<DimSize_t> indexesDims = getInput(1)->dims();
- int axisIdx = this->template getAttr<GatherAttr::Axis>()>=0?this->template getAttr<GatherAttr::Axis>():this->template getAttr<GatherAttr::Axis>()+outDims.size();
- outDims.erase(outDims.begin() + static_cast<std::size_t>(axisIdx));
- outDims.insert(outDims.begin() + static_cast<std::size_t>(axisIdx), indexesDims.begin(),indexesDims.end());
- mOutputs[0]->resize(outDims);
+ mOutputs[0]->resize(outDims);
+ }
}
\ No newline at end of file
diff --git a/src/operator/Reshape.cpp b/src/operator/Reshape.cpp
index b0eea3c1f9f7054021b631c85e0f80e7f8845da6..30b060cd2a58d7995a7447bd9b85b9bc0026a7f7 100644
--- a/src/operator/Reshape.cpp
+++ b/src/operator/Reshape.cpp
@@ -9,39 +9,50 @@
*
********************************************************************************/
-#include <cstddef>
+#include <cstddef> // std::size_t
+#include <cstdint> // std::int64_t
+#include <stdexcept> // std::runtime_error
#include <string>
#include <vector>
#include "aidge/operator/Reshape.hpp"
-#include "aidge/utils/Types.h"
#include "aidge/utils/ErrorHandling.hpp"
+#include "aidge/utils/Types.h"
const std::string Aidge::Reshape_Op::Type = "Reshape";
void Aidge::Reshape_Op::computeOutputDims() {
- // check inputs have been associated
+ // check input has been associated
if (!getInput(0)) {
AIDGE_THROW_OR_ABORT(std::runtime_error, "Input was not connected");
}
- DimSize_t nbOutDims = this->template getAttr<ReshapeAttr::Shape>().size();
- std::vector<DimSize_t> outDims;
- std::size_t outSize = 1;
- for(std::size_t i=0; i<nbOutDims; ++i)
- {
- int dimSize = this->template getAttr<ReshapeAttr::Shape>()[i];
- if (dimSize < 1)
+ if (!getInput(0)->empty()) {
+ std::vector<DimSize_t> outDims;
+ // variables to handle a negative dimension
+ bool foundNegativeDimension = false;
+ std::size_t outSize = 1;
+ DimIdx_t negativeIndex = 0;
+
+ for(std::size_t i = 0; i < this->template getAttr<ReshapeAttr::Shape>().size(); ++i)
{
- AIDGE_THROW_OR_ABORT(std::runtime_error, "bad dimension value");
+ std::int64_t dimSize = this->template getAttr<ReshapeAttr::Shape>()[i];
+ if (dimSize < 0) {
+ if (foundNegativeDimension) {
+ AIDGE_THROW_OR_ABORT(std::runtime_error, "Found more than one negative dimension in Reshape Operator.");
+ }
+ foundNegativeDimension = true;
+ dimSize = 1;
+ negativeIndex = static_cast<DimIdx_t>(i);
+ }
+ outDims.push_back(static_cast<DimSize_t>(dimSize));
+ outSize *= static_cast<DimSize_t>(dimSize);
}
- outDims.push_back(dimSize);
- outSize *= dimSize;
- }
- if (getInput(0)->size() != outSize){
- AIDGE_THROW_OR_ABORT(std::runtime_error, "Output shape must give the same size as input");
- }
+ if (foundNegativeDimension) {
+ outDims[negativeIndex] = (getInput(0) -> size()) / outSize;
+ }
- mOutputs[0]->resize(outDims);
+ mOutputs[0]->resize(outDims);
+ }
}
\ No newline at end of file
diff --git a/src/operator/Slice.cpp b/src/operator/Slice.cpp
index 139e84b561a48c2f6a5ecd14ed9d6905d66dec20..11d91a1fcd4c1d4ee6bcc5f9d830870fa6e732e5 100644
--- a/src/operator/Slice.cpp
+++ b/src/operator/Slice.cpp
@@ -30,21 +30,23 @@ void Aidge::Slice_Op::computeOutputDims() {
AIDGE_THROW_OR_ABORT(std::runtime_error, "Every input should be associated with a Tensor");
}
- DimSize_t nbAxes = this->template getAttr<SliceAttr::Axes>().size();
+ const DimSize_t nbAxes = this->template getAttr<SliceAttr::Axes>().size();
std::vector<DimSize_t> outDims = getInput(0)->dims();
for (std::size_t i = 0; i < nbAxes; ++i) {
// For each slice operation get the params and cast them to size_t
const std::int64_t axis_ = this->template getAttr<SliceAttr::Axes>()[i];
const std::int64_t start_ = this->template getAttr<SliceAttr::Starts>()[i];
const std::int64_t end_ = this->template getAttr<SliceAttr::Ends>()[i];
- const std::size_t axis = axis_ >= 0 ? static_cast<std::size_t>(axis_) : axis_ + getInput(0)->nbDims();
- const std::size_t start = start_ >= 0 ? static_cast<std::size_t>(start_) : start_ + getInput(0)->dims()[axis];
- const std::size_t end = end_ >= 0 ? static_cast<std::size_t>(end_) : end_ + getInput(0)->dims()[axis];
+ const std::size_t axis = axis_ >= 0 ? static_cast<std::size_t>(axis_) : static_cast<std::size_t>(axis_) + getInput(0)->nbDims();
+ const std::size_t start = start_ >= 0 ? static_cast<std::size_t>(start_) : static_cast<std::size_t>(start_) + getInput(0)->dims()[axis];
+ const std::size_t end = end_ >= 0 ? static_cast<std::size_t>(end_) : static_cast<std::size_t>(end_) + getInput(0)->dims()[axis];
const std::size_t sliceLength = end - start + 1;
// Check if slice length is valid
if (sliceLength > getInput(0)->dims()[axis])
+ {
AIDGE_THROW_OR_ABORT(std::runtime_error, "ROI of Slice operator out of bounds");
+ }
outDims[axis] = sliceLength;
}
mOutputs[0]->resize(outDims);
diff --git a/src/recipies/HorizontalTiling.cpp b/src/recipies/HorizontalTiling.cpp
index 6cc34eba076934b884b336ce40081a855d917182..7d3fafc0a15d1b797fdfb1a2884b62d2d8d766c5 100644
--- a/src/recipies/HorizontalTiling.cpp
+++ b/src/recipies/HorizontalTiling.cpp
@@ -82,16 +82,16 @@ std::set<std::shared_ptr<Aidge::Node>> Aidge::getConvHorizontalTiling(const std:
clonedInputs[1] -> addChild(newNode, 0, 1);
clonedInputs[2] -> addChild(newNode, 0, 2);
// Slice for input and each parameter
- std::vector<std::int32_t> inputDimsEnd(inputDims[0].first.size());
+ std::vector<std::int64_t> inputDimsEnd(inputDims[0].first.size());
for (std::size_t dim = 0; dim < inputDimsEnd.size(); ++dim) {
- inputDimsEnd[dim] = static_cast<std::int32_t>(inputDims[0].first[dim] + inputDims[0].second[dim]) - 1;
+ inputDimsEnd[dim] = static_cast<std::int64_t>(inputDims[0].first[dim] + inputDims[0].second[dim]) - 1;
}
- std::vector<std::int32_t> inputDimsStart(inputDims[0].first.size());
+ std::vector<std::int64_t> inputDimsStart(inputDims[0].first.size());
for (std::size_t dim = 0; dim < inputDimsStart.size(); ++dim) {
- inputDimsStart[dim] = static_cast<std::int32_t>(inputDims[0].first[dim]);
+ inputDimsStart[dim] = static_cast<std::int64_t>(inputDims[0].first[dim]);
}
- std::vector<std::int32_t> usedDims(inputDimsEnd.size());
- std::iota(usedDims.begin(), usedDims.end(), static_cast<std::int32_t>(0));
+ std::vector<std::int64_t> usedDims(inputDimsEnd.size());
+ std::iota(usedDims.begin(), usedDims.end(), static_cast<std::int64_t>(0));
auto slice = Slice(inputDimsStart, inputDimsEnd, usedDims, "Slice_" + std::to_string(currentFirstDims[axis]));
slice -> addChild(newNode, 0, 0);
newNode -> addChild(concat, 0, i);