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

51
unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
View File

@@ -91,24 +91,6 @@ class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows,
m_padding_left(padding_left), m_padding_right(padding_right),
m_padding_type(PADDING_VALID), m_padding_value(padding_value) {}
#ifdef EIGEN_USE_SYCL // this is work around for sycl as Eigen could not use c++11 deligate constructor feature
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides,
DenseIndex in_plane_strides, DenseIndex in_row_strides, DenseIndex in_col_strides,
DenseIndex plane_inflate_strides, DenseIndex row_inflate_strides, DenseIndex col_inflate_strides,
bool padding_explicit, DenseIndex padding_top_z, DenseIndex padding_bottom_z,
DenseIndex padding_top, DenseIndex padding_bottom, DenseIndex padding_left,
DenseIndex padding_right, PaddingType padding_type, Scalar padding_value)
: m_xpr(expr), m_patch_planes(patch_planes), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
m_plane_strides(plane_strides), m_row_strides(row_strides), m_col_strides(col_strides),
m_in_plane_strides(in_plane_strides), m_in_row_strides(in_row_strides), m_in_col_strides(in_col_strides),
m_plane_inflate_strides(plane_inflate_strides), m_row_inflate_strides(row_inflate_strides),
m_col_inflate_strides(col_inflate_strides), m_padding_explicit(padding_explicit), m_padding_top_z(padding_top_z),
m_padding_bottom_z(padding_bottom_z), m_padding_top(padding_top), m_padding_bottom(padding_bottom), m_padding_left(padding_left),
m_padding_right(padding_right), m_padding_type(padding_type), m_padding_value(padding_value) {}
#endif
EIGEN_DEVICE_FUNC
DenseIndex patch_planes() const { return m_patch_planes; }
EIGEN_DEVICE_FUNC
@@ -195,6 +177,8 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
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,
@@ -205,16 +189,9 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
CoordAccess = false,
RawAccess = false
};
#ifdef __SYCL_DEVICE_ONLY__
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType op, const Device& device)
#else
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator( const XprType& op, const Device& device)
#endif
: m_impl(op.expression(), device)
#ifdef EIGEN_USE_SYCL
, m_op(op)
#endif
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) :
m_impl(op.expression(), device)
{
EIGEN_STATIC_ASSERT((NumDims >= 5), YOU_MADE_A_PROGRAMMING_MISTAKE);
@@ -368,9 +345,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims-1]);
}
}
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;
}
@@ -531,14 +509,10 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
return TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
#ifdef EIGEN_USE_SYCL
// Required by SYCL in order to construct the expression on the device
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const XprType& xpr() const { return m_op; }
#endif
Index planePaddingTop() const { return m_planePaddingTop; }
Index rowPaddingTop() const { return m_rowPaddingTop; }
@@ -556,10 +530,17 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
Index rowInflateStride() const { return m_row_inflate_strides; }
Index colInflateStride() const { return m_col_inflate_strides; }
#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 PacketReturnType packetWithPossibleZero(Index index) const
{
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
EIGEN_UNROLL_LOOP
for (int i = 0; i < PacketSize; ++i) {
values[i] = coeff(index+i);
}
@@ -635,10 +616,6 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
TensorEvaluator<ArgType, Device> m_impl;
#ifdef EIGEN_USE_SYCL
// Required by SYCL in order to construct the expression on the device
XprType m_op;
#endif
};