devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

[SYCL] This PR adds the minimum modifications to the Eigen unsupported module required to run it on devices supporting SYCL.

Browse Source 
* Abstracting the pointer type so that both SYCL memory and pointer can be captured.
* Converting SYCL virtual pointer to SYCL device memory in Eigen evaluator class.
* Binding SYCL placeholder accessor to command group handler by using bind method in Eigen evaluator node.
* Adding SYCL macro for controlling loop unrolling.
* Modifying the TensorDeviceSycl.h and SYCL executor method to adopt the above changes.
This commit is contained in:
Mehdi Goli
2019-06-28 10:08:23 +01:00
parent 16a56b2ddd
commit 7d08fa805a
47 changed files with 1983 additions and 951 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h
View File

@@ -109,6 +109,8 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
typedef StorageMemory<CoeffReturnType, Device> Storage;
typedef typename Storage::Type EvaluatorPointerType;
enum {
IsAligned = false,
@@ -130,8 +132,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op,
const Device& device)
: m_device(device),
m_impl(op.expression(), device),
m_shuffle(op.shufflePermutation())
m_impl(op.expression(), device)
{
const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
const Shuffle& shuffle = op.shufflePermutation();
@@ -172,7 +173,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
m_impl.evalSubExprsIfNeeded(NULL);
return true;
}
@@ -194,6 +195,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
static PacketReturnType Run(const Self& self, Index index) {
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
EIGEN_UNROLL_LOOP
for (int i = 0; i < PacketSize; ++i) {
values[i] = self.coeff(index + i);
}
@@ -210,6 +212,7 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
return self.m_impl.template packet<LoadMode>(index);
} else {
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
EIGEN_UNROLL_LOOP
for (int i = 0; i < PacketSize; ++i) {
values[i] = self.coeff(index + i);
}
@@ -330,13 +333,14 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
TensorOpCost(0, 0, compute_cost, m_is_identity /* vectorized */, PacketSize);
}
EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
// required by sycl
EIGEN_STRONG_INLINE const Shuffle& shufflePermutation() const {return m_shuffle;}
// required by sycl
EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const {return m_impl;}
EIGEN_DEVICE_FUNC typename Storage::Type data() const { return NULL; }
#ifdef EIGEN_USE_SYCL
// binding placeholder accessors to a command group handler for SYCL
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
m_impl.bind(cgh);
}
#endif
protected:
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index GetBlockOutputIndex(
Index input_index,
@@ -389,10 +393,8 @@ struct TensorEvaluator<const TensorShufflingOp<Shuffle, ArgType>, Device>
array<Index, NumDims> m_inputStrides;
array<Index, NumDims> m_unshuffledInputStrides;
const Device& m_device;
const Device EIGEN_DEVICE_REF m_device;
TensorEvaluator<ArgType, Device> m_impl;
/// required by sycl
Shuffle m_shuffle;
};
@@ -444,6 +446,7 @@ struct TensorEvaluator<TensorShufflingOp<Shuffle, ArgType>, Device>
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
EIGEN_UNROLL_LOOP
for (int i = 0; i < PacketSize; ++i) {
this->coeffRef(index+i) = values[i];
}