Partial OpenCL support via SYCL compatible with ComputeCpp CE.

This commit is contained in:
Luke Iwanski
2016-09-19 12:44:13 +01:00
parent 59bacfe520
commit cb81975714
34 changed files with 3652 additions and 64 deletions

View File

@@ -20,8 +20,8 @@ namespace Eigen {
*
*/
namespace internal {
template<typename XprType>
struct traits<TensorEvalToOp<XprType> >
template<typename XprType, template <class> class MakePointer_>
struct traits<TensorEvalToOp<XprType, MakePointer_> >
{
// Type promotion to handle the case where the types of the lhs and the rhs are different.
typedef typename XprType::Scalar Scalar;
@@ -36,16 +36,20 @@ struct traits<TensorEvalToOp<XprType> >
enum {
Flags = 0
};
template <class T>
struct MakePointer {
typedef typename MakePointer_<T>::Type Type;
};
};
template<typename XprType>
struct eval<TensorEvalToOp<XprType>, Eigen::Dense>
template<typename XprType, template <class> class MakePointer_>
struct eval<TensorEvalToOp<XprType, MakePointer_>, Eigen::Dense>
{
typedef const TensorEvalToOp<XprType>& type;
};
template<typename XprType>
struct nested<TensorEvalToOp<XprType>, 1, typename eval<TensorEvalToOp<XprType> >::type>
template<typename XprType, template <class> class MakePointer_>
struct nested<TensorEvalToOp<XprType, MakePointer_>, 1, typename eval<TensorEvalToOp<XprType, MakePointer_> >::type>
{
typedef TensorEvalToOp<XprType> type;
};
@@ -55,37 +59,38 @@ struct nested<TensorEvalToOp<XprType>, 1, typename eval<TensorEvalToOp<XprType>
template<typename XprType>
class TensorEvalToOp : public TensorBase<TensorEvalToOp<XprType>, ReadOnlyAccessors>
template<typename XprType, template <class> class MakePointer_>
class TensorEvalToOp : public TensorBase<TensorEvalToOp<XprType, MakePointer_>, ReadOnlyAccessors>
{
public:
typedef typename Eigen::internal::traits<TensorEvalToOp>::Scalar Scalar;
typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename MakePointer_<CoeffReturnType>::Type PointerType;
typedef typename Eigen::internal::nested<TensorEvalToOp>::type Nested;
typedef typename Eigen::internal::traits<TensorEvalToOp>::StorageKind StorageKind;
typedef typename Eigen::internal::traits<TensorEvalToOp>::Index Index;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(CoeffReturnType* buffer, const XprType& expr)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(PointerType buffer, const XprType& expr)
: m_xpr(expr), m_buffer(buffer) {}
EIGEN_DEVICE_FUNC
const typename internal::remove_all<typename XprType::Nested>::type&
expression() const { return m_xpr; }
EIGEN_DEVICE_FUNC CoeffReturnType* buffer() const { return m_buffer; }
EIGEN_DEVICE_FUNC PointerType buffer() const { return m_buffer; }
protected:
typename XprType::Nested m_xpr;
CoeffReturnType* m_buffer;
PointerType m_buffer;
};
template<typename ArgType, typename Device>
struct TensorEvaluator<const TensorEvalToOp<ArgType>, Device>
template<typename ArgType, typename Device, template <class> class MakePointer_>
struct TensorEvaluator<const TensorEvalToOp<ArgType, MakePointer_>, Device>
{
typedef TensorEvalToOp<ArgType> XprType;
typedef TensorEvalToOp<ArgType, MakePointer_> XprType;
typedef typename ArgType::Scalar Scalar;
typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
typedef typename XprType::Index Index;
@@ -102,15 +107,22 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType>, Device>
};
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
: m_impl(op.expression(), device), m_device(device), m_buffer(op.buffer())
: m_impl(op.expression(), device), m_device(device),
m_buffer(op.buffer()), m_op(op), m_expression(op.expression())
{ }
// Used for accessor extraction in SYCL Managed TensorMap:
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& op() const {
return m_op;
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() {
}
typedef typename internal::traits<const TensorEvalToOp<ArgType, MakePointer_>>::template MakePointer<CoeffReturnType>::Type DevicePointer;
EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* scalar) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(DevicePointer scalar) {
EIGEN_UNUSED_VARIABLE(scalar);
eigen_assert(scalar == NULL);
return m_impl.evalSubExprsIfNeeded(m_buffer);
@@ -145,12 +157,20 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType>, Device>
TensorOpCost(0, sizeof(CoeffReturnType), 0, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_buffer; }
EIGEN_DEVICE_FUNC DevicePointer data() const { return m_buffer; }
ArgType expression() const { return m_expression; }
/// required by sycl in order to extract the accessor
const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
/// added for sycl in order to construct the buffer from the sycl device
const Device& device() const{return m_device;}
private:
TensorEvaluator<ArgType, Device> m_impl;
const Device& m_device;
CoeffReturnType* m_buffer;
DevicePointer m_buffer;
const XprType& m_op;
const ArgType m_expression;
};