Merge remote-tracking branch 'origin/dev' into dataloader

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

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) {
         /* Request a buffer descriptor from Python */
         py::buffer_info info = b.request();
@@ -58,16 +58,16 @@ 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<>())
@@ -76,7 +76,7 @@ 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)
@@ -118,7 +118,7 @@ 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;

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);
     }
 }

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;
     }
 }