mirror of
https://gitlab.com/libeigen/eigen.git
synced 2024-12-27 07:29:52 +08:00
Adding TensorChippingOP for sycl backend; fixing the index value in the verification operation for cxx11_tensorChipping.cpp test
This commit is contained in:
parent
fad776492f
commit
0d153ded29
@ -150,7 +150,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
|
||||
};
|
||||
|
||||
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
|
||||
: m_impl(op.expression(), device), m_dim(op.dim()), m_device(device)
|
||||
: m_impl(op.expression(), device), m_dim(op.dim()), m_device(device), m_offset(op.offset())
|
||||
{
|
||||
EIGEN_STATIC_ASSERT((NumInputDims >= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
|
||||
eigen_assert(NumInputDims > m_dim.actualDim());
|
||||
@ -206,7 +206,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
|
||||
eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
|
||||
|
||||
if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == 0) ||
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
|
||||
// m_stride is equal to 1, so let's avoid the integer division.
|
||||
eigen_assert(m_stride == 1);
|
||||
Index inputIndex = index * m_inputStride + m_inputOffset;
|
||||
@ -218,7 +218,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
|
||||
PacketReturnType rslt = internal::pload<PacketReturnType>(values);
|
||||
return rslt;
|
||||
} else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims - 1) ||
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
|
||||
// m_stride is aways greater than index, so let's avoid the integer division.
|
||||
eigen_assert(m_stride > index);
|
||||
return m_impl.template packet<LoadMode>(index + m_inputOffset);
|
||||
@ -274,17 +274,29 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
|
||||
}
|
||||
}
|
||||
|
||||
/// used by sycl
|
||||
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex dimId() const {
|
||||
return m_dim.actualDim();
|
||||
}
|
||||
|
||||
/// used by sycl
|
||||
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const DenseIndex& offset() const {
|
||||
return m_offset;
|
||||
}
|
||||
/// required by sycl in order to extract the accessor
|
||||
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
|
||||
|
||||
protected:
|
||||
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
|
||||
{
|
||||
Index inputIndex;
|
||||
if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == 0) ||
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
|
||||
// m_stride is equal to 1, so let's avoid the integer division.
|
||||
eigen_assert(m_stride == 1);
|
||||
inputIndex = index * m_inputStride + m_inputOffset;
|
||||
} else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims-1) ||
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
|
||||
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
|
||||
// m_stride is aways greater than index, so let's avoid the integer division.
|
||||
eigen_assert(m_stride > index);
|
||||
inputIndex = index + m_inputOffset;
|
||||
@ -304,6 +316,9 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
|
||||
TensorEvaluator<ArgType, Device> m_impl;
|
||||
const internal::DimensionId<DimId> m_dim;
|
||||
const Device& m_device;
|
||||
// required by sycl
|
||||
const DenseIndex m_offset;
|
||||
|
||||
};
|
||||
|
||||
|
||||
@ -344,7 +359,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
|
||||
EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
|
||||
|
||||
if ((static_cast<int>(this->Layout) == static_cast<int>(ColMajor) && this->m_dim.actualDim() == 0) ||
|
||||
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == NumInputDims-1)) {
|
||||
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == NumInputDims-1)) {
|
||||
// m_stride is equal to 1, so let's avoid the integer division.
|
||||
eigen_assert(this->m_stride == 1);
|
||||
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
|
||||
@ -355,7 +370,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
|
||||
inputIndex += this->m_inputStride;
|
||||
}
|
||||
} else if ((static_cast<int>(this->Layout) == static_cast<int>(ColMajor) && this->m_dim.actualDim() == NumInputDims-1) ||
|
||||
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == 0)) {
|
||||
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == 0)) {
|
||||
// m_stride is aways greater than index, so let's avoid the integer division.
|
||||
eigen_assert(this->m_stride > index);
|
||||
this->m_impl.template writePacket<StoreMode>(index + this->m_inputOffset, x);
|
||||
|
@ -146,6 +146,18 @@ KERNELBROKERCONVERTERSLICESTRIDEOP()
|
||||
#undef KERNELBROKERCONVERTERSLICESTRIDEOP
|
||||
|
||||
|
||||
/// specialisation of the \ref ConvertToDeviceExpression struct when the node type is TensorChippingOp
|
||||
#define KERNELBROKERCONVERTCHIPPINGOP(CVQual)\
|
||||
template <DenseIndex DimId, typename Expr>\
|
||||
struct ConvertToDeviceExpression<CVQual TensorChippingOp<DimId, Expr> > {\
|
||||
typedef CVQual TensorChippingOp<DimId, typename ConvertToDeviceExpression<Expr>::Type> Type;\
|
||||
};
|
||||
KERNELBROKERCONVERTCHIPPINGOP(const)
|
||||
KERNELBROKERCONVERTCHIPPINGOP()
|
||||
#undef KERNELBROKERCONVERTCHIPPINGOP
|
||||
|
||||
|
||||
|
||||
} // namespace internal
|
||||
} // namespace TensorSycl
|
||||
} // namespace Eigen
|
||||
|
@ -368,6 +368,23 @@ SYCLPADDINGOPEXPRCONST(TensorPaddingOp, )
|
||||
#undef SYCLPADDINGOPEXPRCONST
|
||||
|
||||
|
||||
// TensorChippingOp
|
||||
#define SYCLTENSORCHIPPINGOPEXPR(CVQual)\
|
||||
template<DenseIndex DimId, typename OrigXprType, typename XprType, typename... Params>\
|
||||
struct ExprConstructor<CVQual TensorChippingOp <DimId, OrigXprType> , CVQual TensorChippingOp<DimId, XprType>, Params... >{\
|
||||
typedef ExprConstructor<OrigXprType, XprType, Params...> my_xpr_type;\
|
||||
typedef CVQual TensorChippingOp<DimId, typename my_xpr_type::Type> Type;\
|
||||
my_xpr_type xprExpr;\
|
||||
Type expr;\
|
||||
template <typename FuncDetector>\
|
||||
ExprConstructor(FuncDetector &funcD, const utility::tuple::Tuple<Params...> &t)\
|
||||
: xprExpr(funcD.xprExpr, t), expr(xprExpr.expr, funcD.offset(), funcD.dimId()) {}\
|
||||
};
|
||||
|
||||
SYCLTENSORCHIPPINGOPEXPR(const)
|
||||
SYCLTENSORCHIPPINGOPEXPR()
|
||||
#undef SYCLTENSORCHIPPINGOPEXPR
|
||||
|
||||
|
||||
/// template deduction for \ref ExprConstructor struct
|
||||
template <typename OrigExpr, typename IndexExpr, typename FuncD, typename... Params>
|
||||
|
@ -188,7 +188,7 @@ SYCLCONTRACTIONCONVOLUTIONEXTACC(,TensorConvolutionOp)
|
||||
|
||||
|
||||
/// specialisation of the \ref ExtractAccessor struct when the node type is
|
||||
/// const TensorSlicingOp.
|
||||
/// const TensorSlicingOp.
|
||||
#define SYCLSLICEOPEXTACC(CVQual)\
|
||||
template <typename StartIndices, typename Sizes, typename XprType, typename Dev>\
|
||||
struct ExtractAccessor<TensorEvaluator<CVQual TensorSlicingOp<StartIndices, Sizes, XprType>, Dev> > {\
|
||||
@ -200,7 +200,7 @@ SYCLSLICEOPEXTACC(const)
|
||||
SYCLSLICEOPEXTACC()
|
||||
#undef SYCLSLICEOPEXTACC
|
||||
// specialisation of the \ref ExtractAccessor struct when the node type is
|
||||
/// const TensorStridingSlicingOp.
|
||||
/// TensorStridingSlicingOp.
|
||||
#define SYCLSLICESTRIDEOPEXTACC(CVQual)\
|
||||
template<typename StartIndices, typename StopIndices, typename Strides, typename XprType, typename Dev>\
|
||||
struct ExtractAccessor<TensorEvaluator<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, Dev> >{\
|
||||
@ -212,6 +212,19 @@ SYCLSLICESTRIDEOPEXTACC(const)
|
||||
SYCLSLICESTRIDEOPEXTACC()
|
||||
#undef SYCLSLICESTRIDEOPEXTACC
|
||||
|
||||
// specialisation of the \ref ExtractAccessor struct when the node type is
|
||||
/// TensorChippingOp.
|
||||
#define SYCLTENSORCHIPPINGOPEXTACC(CVQual)\
|
||||
template<DenseIndex DimId, typename XprType, typename Dev>\
|
||||
struct ExtractAccessor<TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Dev> >{\
|
||||
static inline auto getTuple(cl::sycl::handler& cgh, const TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Dev>& eval)\
|
||||
RETURN_CPP11(AccessorConstructor::getTuple(cgh, eval.impl()))\
|
||||
};
|
||||
|
||||
SYCLTENSORCHIPPINGOPEXTACC(const)
|
||||
SYCLTENSORCHIPPINGOPEXTACC()
|
||||
#undef SYCLTENSORCHIPPINGOPEXTACC
|
||||
|
||||
|
||||
/// template deduction for \ref ExtractAccessor
|
||||
template <typename Evaluator>
|
||||
|
@ -290,6 +290,22 @@ SYCLEXTRFUNCCONTRACTCONCAT(TensorConcatenationOp, axis(), const)
|
||||
SYCLEXTRFUNCCONTRACTCONCAT(TensorConcatenationOp, axis(),)
|
||||
#undef SYCLEXTRFUNCCONTRACTCONCAT
|
||||
|
||||
//TensorChippingOp
|
||||
#define SYCLEXTRFUNCCHIPPINGOP(CVQual)\
|
||||
template<DenseIndex DimId, typename XprType, typename Device>\
|
||||
struct FunctorExtractor<TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Device>>{\
|
||||
FunctorExtractor<Eigen::TensorEvaluator<XprType, Device> > xprExpr;\
|
||||
const DenseIndex m_dim;\
|
||||
const DenseIndex m_offset;\
|
||||
EIGEN_STRONG_INLINE const DenseIndex& dimId() const { return m_dim; }\
|
||||
EIGEN_STRONG_INLINE const DenseIndex& offset() const { return m_offset; }\
|
||||
FunctorExtractor(const TensorEvaluator<CVQual TensorChippingOp<DimId, XprType>, Device>& expr)\
|
||||
: xprExpr(expr.impl()), m_dim(expr.dimId()), m_offset(expr.offset()) {}\
|
||||
};
|
||||
|
||||
SYCLEXTRFUNCCHIPPINGOP(const)
|
||||
SYCLEXTRFUNCCHIPPINGOP()
|
||||
#undef SYCLEXTRFUNCCHIPPINGOP
|
||||
|
||||
/// template deduction function for FunctorExtractor
|
||||
template <typename Evaluator>
|
||||
|
@ -139,6 +139,17 @@ SLICEOPLEAFCOUNT(const)
|
||||
SLICEOPLEAFCOUNT()
|
||||
#undef SLICEOPLEAFCOUNT
|
||||
|
||||
|
||||
/// specialisation of the \ref LeafCount struct when the node type is TensorChippingOp
|
||||
#define CHIPPINGOPLEAFCOUNT(CVQual)\
|
||||
template <DenseIndex DimId, typename XprType>\
|
||||
struct LeafCount<CVQual TensorChippingOp<DimId, XprType> >:CategoryCount<XprType>{};
|
||||
|
||||
CHIPPINGOPLEAFCOUNT(const)
|
||||
CHIPPINGOPLEAFCOUNT()
|
||||
#undef CHIPPINGOPLEAFCOUNT
|
||||
|
||||
|
||||
#define SLICESTRIDEOPLEAFCOUNT(CVQual)\
|
||||
template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>\
|
||||
struct LeafCount<CVQual TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >:CategoryCount<XprType>{};
|
||||
|
@ -156,6 +156,18 @@ EVALTO()
|
||||
#undef EVALTO
|
||||
|
||||
|
||||
/// specialisation of the \ref PlaceHolderExpression when the node is
|
||||
/// TensorChippingOp
|
||||
#define CHIPPINGOP(CVQual)\
|
||||
template <DenseIndex DimId, typename Expr, size_t N>\
|
||||
struct PlaceHolderExpression<CVQual TensorChippingOp<DimId, Expr>, N> {\
|
||||
typedef CVQual TensorChippingOp< DimId, typename CalculateIndex <N, Expr>::ArgType> Type;\
|
||||
};
|
||||
|
||||
CHIPPINGOP(const)
|
||||
CHIPPINGOP()
|
||||
#undef CHIPPINGOP
|
||||
|
||||
/// specialisation of the \ref PlaceHolderExpression when the node is
|
||||
/// TensorReductionOp
|
||||
#define SYCLREDUCTION(CVQual)\
|
||||
|
@ -166,6 +166,7 @@ if(EIGEN_TEST_CXX11)
|
||||
ei_add_test_sycl(cxx11_tensor_reverse_sycl "-std=c++11")
|
||||
ei_add_test_sycl(cxx11_tensor_convolution_sycl "-std=c++11")
|
||||
ei_add_test_sycl(cxx11_tensor_striding_sycl "-std=c++11")
|
||||
ei_add_test_sycl(cxx11_tensor_chipping_sycl "-std=c++11")
|
||||
endif(EIGEN_TEST_SYCL)
|
||||
# It should be safe to always run these tests as there is some fallback code for
|
||||
# older compiler that don't support cxx11.
|
||||
|
@ -43,7 +43,7 @@ static void test_simple_chip()
|
||||
VERIFY_IS_EQUAL(chip2.dimension(2), 7);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(3), 11);
|
||||
for (int i = 0; i < 2; ++i) {
|
||||
for (int j = 0; j < 3; ++j) {
|
||||
for (int j = 0; j < 5; ++j) {
|
||||
for (int k = 0; k < 7; ++k) {
|
||||
for (int l = 0; l < 11; ++l) {
|
||||
VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1,j,k,l));
|
||||
@ -75,7 +75,7 @@ static void test_simple_chip()
|
||||
for (int i = 0; i < 2; ++i) {
|
||||
for (int j = 0; j < 3; ++j) {
|
||||
for (int k = 0; k < 5; ++k) {
|
||||
for (int l = 0; l < 7; ++l) {
|
||||
for (int l = 0; l < 11; ++l) {
|
||||
VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5,l));
|
||||
}
|
||||
}
|
||||
@ -126,7 +126,7 @@ static void test_dynamic_chip()
|
||||
VERIFY_IS_EQUAL(chip2.dimension(2), 7);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(3), 11);
|
||||
for (int i = 0; i < 2; ++i) {
|
||||
for (int j = 0; j < 3; ++j) {
|
||||
for (int j = 0; j < 5; ++j) {
|
||||
for (int k = 0; k < 7; ++k) {
|
||||
for (int l = 0; l < 11; ++l) {
|
||||
VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1,j,k,l));
|
||||
@ -158,7 +158,7 @@ static void test_dynamic_chip()
|
||||
for (int i = 0; i < 2; ++i) {
|
||||
for (int j = 0; j < 3; ++j) {
|
||||
for (int k = 0; k < 5; ++k) {
|
||||
for (int l = 0; l < 7; ++l) {
|
||||
for (int l = 0; l < 11; ++l) {
|
||||
VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5,l));
|
||||
}
|
||||
}
|
||||
|
622
unsupported/test/cxx11_tensor_chipping_sycl.cpp
Normal file
622
unsupported/test/cxx11_tensor_chipping_sycl.cpp
Normal file
@ -0,0 +1,622 @@
|
||||
// This file is part of Eigen, a lightweight C++ template library
|
||||
// for linear algebra.
|
||||
//
|
||||
// Copyright (C) 2016
|
||||
// Mehdi Goli Codeplay Software Ltd.
|
||||
// Ralph Potter Codeplay Software Ltd.
|
||||
// Luke Iwanski Codeplay Software Ltd.
|
||||
// Contact: <eigen@codeplay.com>
|
||||
// Benoit Steiner <benoit.steiner.goog@gmail.com>
|
||||
//
|
||||
// This Source Code Form is subject to the terms of the Mozilla
|
||||
// Public License v. 2.0. If a copy of the MPL was not distributed
|
||||
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
|
||||
|
||||
|
||||
#define EIGEN_TEST_NO_LONGDOUBLE
|
||||
#define EIGEN_TEST_NO_COMPLEX
|
||||
#define EIGEN_TEST_FUNC cxx11_tensor_chipping_sycl
|
||||
#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
|
||||
#define EIGEN_USE_SYCL
|
||||
|
||||
#include "main.h"
|
||||
|
||||
#include <Eigen/CXX11/Tensor>
|
||||
|
||||
using Eigen::Tensor;
|
||||
|
||||
template <typename DataType, int DataLayout, typename IndexType>
|
||||
static void test_static_chip_sycl(const Eigen::SyclDevice& sycl_device)
|
||||
{
|
||||
IndexType sizeDim1 = 2;
|
||||
IndexType sizeDim2 = 3;
|
||||
IndexType sizeDim3 = 5;
|
||||
IndexType sizeDim4 = 7;
|
||||
IndexType sizeDim5 = 11;
|
||||
|
||||
array<IndexType, 5> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
array<IndexType, 4> chip1TensorRange = {{sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
|
||||
Tensor<DataType, 5, DataLayout,IndexType> tensor(tensorRange);
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip1(chip1TensorRange);
|
||||
|
||||
tensor.setRandom();
|
||||
|
||||
const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
|
||||
const size_t chip1TensorBuffSize =chip1.size()*sizeof(DataType);
|
||||
DataType* gpu_data_tensor = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
DataType* gpu_data_chip1 = static_cast<DataType*>(sycl_device.allocate(chip1TensorBuffSize));
|
||||
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip1(gpu_data_chip1, chip1TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_tensor, tensor.data(), tensorBuffSize);
|
||||
gpu_chip1.device(sycl_device)=gpu_tensor.template chip<0l>(1l);
|
||||
sycl_device.memcpyDeviceToHost(chip1.data(), gpu_data_chip1, chip1TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip1.dimension(0), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(1), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim2; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim3; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip1(i,j,k,l), tensor(1l,i,j,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip2TensorRange = {{sizeDim1, sizeDim3, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip2(chip2TensorRange);
|
||||
const size_t chip2TensorBuffSize =chip2.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip2 = static_cast<DataType*>(sycl_device.allocate(chip2TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip2(gpu_data_chip2, chip2TensorRange);
|
||||
|
||||
gpu_chip2.device(sycl_device)=gpu_tensor.template chip<1l>(1l);
|
||||
sycl_device.memcpyDeviceToHost(chip2.data(), gpu_data_chip2, chip2TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip2.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(1), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim3; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1l,j,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip3TensorRange = {{sizeDim1, sizeDim2, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip3(chip3TensorRange);
|
||||
const size_t chip3TensorBuffSize =chip3.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip3 = static_cast<DataType*>(sycl_device.allocate(chip3TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip3(gpu_data_chip3, chip3TensorRange);
|
||||
|
||||
gpu_chip3.device(sycl_device)=gpu_tensor.template chip<2l>(2l);
|
||||
sycl_device.memcpyDeviceToHost(chip3.data(), gpu_data_chip3, chip3TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip3.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip3(i,j,k,l), tensor(i,j,2l,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip4TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip4(chip4TensorRange);
|
||||
const size_t chip4TensorBuffSize =chip4.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip4 = static_cast<DataType*>(sycl_device.allocate(chip4TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip4(gpu_data_chip4, chip4TensorRange);
|
||||
|
||||
gpu_chip4.device(sycl_device)=gpu_tensor.template chip<3l>(5l);
|
||||
sycl_device.memcpyDeviceToHost(chip4.data(), gpu_data_chip4, chip4TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip4.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(2), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim3; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5l,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
array<IndexType, 4> chip5TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip5(chip5TensorRange);
|
||||
const size_t chip5TensorBuffSize =chip5.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip5 = static_cast<DataType*>(sycl_device.allocate(chip5TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip5(gpu_data_chip5, chip5TensorRange);
|
||||
|
||||
gpu_chip5.device(sycl_device)=gpu_tensor.template chip<4l>(7l);
|
||||
sycl_device.memcpyDeviceToHost(chip5.data(), gpu_data_chip5, chip5TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip5.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(2), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(3), sizeDim4);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim3; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim4; ++l) {
|
||||
VERIFY_IS_EQUAL(chip5(i,j,k,l), tensor(i,j,k,l,7l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
sycl_device.deallocate(gpu_data_tensor);
|
||||
sycl_device.deallocate(gpu_data_chip1);
|
||||
sycl_device.deallocate(gpu_data_chip2);
|
||||
sycl_device.deallocate(gpu_data_chip3);
|
||||
sycl_device.deallocate(gpu_data_chip4);
|
||||
sycl_device.deallocate(gpu_data_chip5);
|
||||
}
|
||||
|
||||
template <typename DataType, int DataLayout, typename IndexType>
|
||||
static void test_dynamic_chip_sycl(const Eigen::SyclDevice& sycl_device)
|
||||
{
|
||||
IndexType sizeDim1 = 2;
|
||||
IndexType sizeDim2 = 3;
|
||||
IndexType sizeDim3 = 5;
|
||||
IndexType sizeDim4 = 7;
|
||||
IndexType sizeDim5 = 11;
|
||||
|
||||
array<IndexType, 5> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
array<IndexType, 4> chip1TensorRange = {{sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
|
||||
Tensor<DataType, 5, DataLayout,IndexType> tensor(tensorRange);
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip1(chip1TensorRange);
|
||||
|
||||
tensor.setRandom();
|
||||
|
||||
const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
|
||||
const size_t chip1TensorBuffSize =chip1.size()*sizeof(DataType);
|
||||
DataType* gpu_data_tensor = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
DataType* gpu_data_chip1 = static_cast<DataType*>(sycl_device.allocate(chip1TensorBuffSize));
|
||||
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip1(gpu_data_chip1, chip1TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_tensor, tensor.data(), tensorBuffSize);
|
||||
gpu_chip1.device(sycl_device)=gpu_tensor.chip(1l,0l);
|
||||
sycl_device.memcpyDeviceToHost(chip1.data(), gpu_data_chip1, chip1TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip1.dimension(0), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(1), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip1.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim2; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim3; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip1(i,j,k,l), tensor(1l,i,j,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip2TensorRange = {{sizeDim1, sizeDim3, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip2(chip2TensorRange);
|
||||
const size_t chip2TensorBuffSize =chip2.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip2 = static_cast<DataType*>(sycl_device.allocate(chip2TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip2(gpu_data_chip2, chip2TensorRange);
|
||||
|
||||
gpu_chip2.device(sycl_device)=gpu_tensor.chip(1l,1l);
|
||||
sycl_device.memcpyDeviceToHost(chip2.data(), gpu_data_chip2, chip2TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip2.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(1), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip2.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim3; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1l,j,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip3TensorRange = {{sizeDim1, sizeDim2, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip3(chip3TensorRange);
|
||||
const size_t chip3TensorBuffSize =chip3.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip3 = static_cast<DataType*>(sycl_device.allocate(chip3TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip3(gpu_data_chip3, chip3TensorRange);
|
||||
|
||||
gpu_chip3.device(sycl_device)=gpu_tensor.chip(2l,2l);
|
||||
sycl_device.memcpyDeviceToHost(chip3.data(), gpu_data_chip3, chip3TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip3.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(2), sizeDim4);
|
||||
VERIFY_IS_EQUAL(chip3.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim4; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip3(i,j,k,l), tensor(i,j,2l,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 4> chip4TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip4(chip4TensorRange);
|
||||
const size_t chip4TensorBuffSize =chip4.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip4 = static_cast<DataType*>(sycl_device.allocate(chip4TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip4(gpu_data_chip4, chip4TensorRange);
|
||||
|
||||
gpu_chip4.device(sycl_device)=gpu_tensor.chip(5l,3l);
|
||||
sycl_device.memcpyDeviceToHost(chip4.data(), gpu_data_chip4, chip4TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip4.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(2), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip4.dimension(3), sizeDim5);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim3; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim5; ++l) {
|
||||
VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5l,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
array<IndexType, 4> chip5TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip5(chip5TensorRange);
|
||||
const size_t chip5TensorBuffSize =chip5.size()*sizeof(DataType);
|
||||
DataType* gpu_data_chip5 = static_cast<DataType*>(sycl_device.allocate(chip5TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip5(gpu_data_chip5, chip5TensorRange);
|
||||
|
||||
gpu_chip5.device(sycl_device)=gpu_tensor.chip(7l,4l);
|
||||
sycl_device.memcpyDeviceToHost(chip5.data(), gpu_data_chip5, chip5TensorBuffSize);
|
||||
|
||||
VERIFY_IS_EQUAL(chip5.dimension(0), sizeDim1);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(1), sizeDim2);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(2), sizeDim3);
|
||||
VERIFY_IS_EQUAL(chip5.dimension(3), sizeDim4);
|
||||
|
||||
for (IndexType i = 0; i < sizeDim1; ++i) {
|
||||
for (IndexType j = 0; j < sizeDim2; ++j) {
|
||||
for (IndexType k = 0; k < sizeDim3; ++k) {
|
||||
for (IndexType l = 0; l < sizeDim4; ++l) {
|
||||
VERIFY_IS_EQUAL(chip5(i,j,k,l), tensor(i,j,k,l,7l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
sycl_device.deallocate(gpu_data_tensor);
|
||||
sycl_device.deallocate(gpu_data_chip1);
|
||||
sycl_device.deallocate(gpu_data_chip2);
|
||||
sycl_device.deallocate(gpu_data_chip3);
|
||||
sycl_device.deallocate(gpu_data_chip4);
|
||||
sycl_device.deallocate(gpu_data_chip5);
|
||||
}
|
||||
|
||||
template <typename DataType, int DataLayout, typename IndexType>
|
||||
static void test_chip_in_expr(const Eigen::SyclDevice& sycl_device) {
|
||||
|
||||
IndexType sizeDim1 = 2;
|
||||
IndexType sizeDim2 = 3;
|
||||
IndexType sizeDim3 = 5;
|
||||
IndexType sizeDim4 = 7;
|
||||
IndexType sizeDim5 = 11;
|
||||
|
||||
array<IndexType, 5> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
array<IndexType, 4> chip1TensorRange = {{sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
|
||||
Tensor<DataType, 5, DataLayout,IndexType> tensor(tensorRange);
|
||||
|
||||
Tensor<DataType, 4, DataLayout,IndexType> chip1(chip1TensorRange);
|
||||
Tensor<DataType, 4, DataLayout,IndexType> tensor1(chip1TensorRange);
|
||||
tensor.setRandom();
|
||||
tensor1.setRandom();
|
||||
|
||||
const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
|
||||
const size_t chip1TensorBuffSize =chip1.size()*sizeof(DataType);
|
||||
DataType* gpu_data_tensor = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
DataType* gpu_data_chip1 = static_cast<DataType*>(sycl_device.allocate(chip1TensorBuffSize));
|
||||
DataType* gpu_data_tensor1 = static_cast<DataType*>(sycl_device.allocate(chip1TensorBuffSize));
|
||||
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_chip1(gpu_data_chip1, chip1TensorRange);
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_tensor1(gpu_data_tensor1, chip1TensorRange);
|
||||
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_tensor, tensor.data(), tensorBuffSize);
|
||||
sycl_device.memcpyHostToDevice(gpu_data_tensor1, tensor1.data(), chip1TensorBuffSize);
|
||||
gpu_chip1.device(sycl_device)=gpu_tensor.template chip<0l>(0l) + gpu_tensor1;
|
||||
sycl_device.memcpyDeviceToHost(chip1.data(), gpu_data_chip1, chip1TensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim2; ++i) {
|
||||
for (int j = 0; j < sizeDim3; ++j) {
|
||||
for (int k = 0; k < sizeDim4; ++k) {
|
||||
for (int l = 0; l < sizeDim5; ++l) {
|
||||
float expected = tensor(0l,i,j,k,l) + tensor1(i,j,k,l);
|
||||
VERIFY_IS_EQUAL(chip1(i,j,k,l), expected);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
array<IndexType, 3> chip2TensorRange = {{sizeDim2, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 3, DataLayout,IndexType> tensor2(chip2TensorRange);
|
||||
Tensor<DataType, 3, DataLayout,IndexType> chip2(chip2TensorRange);
|
||||
tensor2.setRandom();
|
||||
const size_t chip2TensorBuffSize =tensor2.size()*sizeof(DataType);
|
||||
DataType* gpu_data_tensor2 = static_cast<DataType*>(sycl_device.allocate(chip2TensorBuffSize));
|
||||
DataType* gpu_data_chip2 = static_cast<DataType*>(sycl_device.allocate(chip2TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 3, DataLayout,IndexType>> gpu_tensor2(gpu_data_tensor2, chip2TensorRange);
|
||||
TensorMap<Tensor<DataType, 3, DataLayout,IndexType>> gpu_chip2(gpu_data_chip2, chip2TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_tensor2, tensor2.data(), chip2TensorBuffSize);
|
||||
gpu_chip2.device(sycl_device)=gpu_tensor.template chip<0l>(0l).template chip<1l>(2l) + gpu_tensor2;
|
||||
sycl_device.memcpyDeviceToHost(chip2.data(), gpu_data_chip2, chip2TensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim2; ++i) {
|
||||
for (int j = 0; j < sizeDim4; ++j) {
|
||||
for (int k = 0; k < sizeDim5; ++k) {
|
||||
float expected = tensor(0l,i,2l,j,k) + tensor2(i,j,k);
|
||||
VERIFY_IS_EQUAL(chip2(i,j,k), expected);
|
||||
}
|
||||
}
|
||||
}
|
||||
sycl_device.deallocate(gpu_data_tensor);
|
||||
sycl_device.deallocate(gpu_data_tensor1);
|
||||
sycl_device.deallocate(gpu_data_chip1);
|
||||
sycl_device.deallocate(gpu_data_tensor2);
|
||||
sycl_device.deallocate(gpu_data_chip2);
|
||||
}
|
||||
|
||||
template <typename DataType, int DataLayout, typename IndexType>
|
||||
static void test_chip_as_lvalue_sycl(const Eigen::SyclDevice& sycl_device)
|
||||
{
|
||||
|
||||
IndexType sizeDim1 = 2;
|
||||
IndexType sizeDim2 = 3;
|
||||
IndexType sizeDim3 = 5;
|
||||
IndexType sizeDim4 = 7;
|
||||
IndexType sizeDim5 = 11;
|
||||
|
||||
array<IndexType, 5> tensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
array<IndexType, 4> input2TensorRange = {{sizeDim2, sizeDim3, sizeDim4, sizeDim5}};
|
||||
|
||||
Tensor<DataType, 5, DataLayout,IndexType> tensor(tensorRange);
|
||||
Tensor<DataType, 5, DataLayout,IndexType> input1(tensorRange);
|
||||
Tensor<DataType, 4, DataLayout,IndexType> input2(input2TensorRange);
|
||||
input1.setRandom();
|
||||
input2.setRandom();
|
||||
|
||||
|
||||
const size_t tensorBuffSize =tensor.size()*sizeof(DataType);
|
||||
const size_t input2TensorBuffSize =input2.size()*sizeof(DataType);
|
||||
DataType* gpu_data_tensor = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
DataType* gpu_data_input1 = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
DataType* gpu_data_input2 = static_cast<DataType*>(sycl_device.allocate(input2TensorBuffSize));
|
||||
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_tensor(gpu_data_tensor, tensorRange);
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_input1(gpu_data_input1, tensorRange);
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_input2(gpu_data_input2, input2TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input1, input1.data(), tensorBuffSize);
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input2, input2.data(), input2TensorBuffSize);
|
||||
gpu_tensor.template chip<0l>(1l).device(sycl_device)=gpu_input2;
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k < sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (i != 1) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input2(j,k,l,m));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
array<IndexType, 4> input3TensorRange = {{sizeDim1, sizeDim3, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> input3(input3TensorRange);
|
||||
input3.setRandom();
|
||||
|
||||
const size_t input3TensorBuffSize =input3.size()*sizeof(DataType);
|
||||
DataType* gpu_data_input3 = static_cast<DataType*>(sycl_device.allocate(input3TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_input3(gpu_data_input3, input3TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input3, input3.data(), input3TensorBuffSize);
|
||||
gpu_tensor.template chip<1l>(1l).device(sycl_device)=gpu_input3;
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k <sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (j != 1) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input3(i,k,l,m));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
array<IndexType, 4> input4TensorRange = {{sizeDim1, sizeDim2, sizeDim4, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> input4(input4TensorRange);
|
||||
input4.setRandom();
|
||||
|
||||
const size_t input4TensorBuffSize =input4.size()*sizeof(DataType);
|
||||
DataType* gpu_data_input4 = static_cast<DataType*>(sycl_device.allocate(input4TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_input4(gpu_data_input4, input4TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input4, input4.data(), input4TensorBuffSize);
|
||||
gpu_tensor.template chip<2l>(3l).device(sycl_device)=gpu_input4;
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k <sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (k != 3) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input4(i,j,l,m));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
array<IndexType, 4> input5TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim5}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> input5(input5TensorRange);
|
||||
input5.setRandom();
|
||||
|
||||
const size_t input5TensorBuffSize =input5.size()*sizeof(DataType);
|
||||
DataType* gpu_data_input5 = static_cast<DataType*>(sycl_device.allocate(input5TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_input5(gpu_data_input5, input5TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input5, input5.data(), input5TensorBuffSize);
|
||||
gpu_tensor.template chip<3l>(4l).device(sycl_device)=gpu_input5;
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k <sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (l != 4) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input5(i,j,k,m));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
array<IndexType, 4> input6TensorRange = {{sizeDim1, sizeDim2, sizeDim3, sizeDim4}};
|
||||
Tensor<DataType, 4, DataLayout,IndexType> input6(input6TensorRange);
|
||||
input6.setRandom();
|
||||
|
||||
const size_t input6TensorBuffSize =input6.size()*sizeof(DataType);
|
||||
DataType* gpu_data_input6 = static_cast<DataType*>(sycl_device.allocate(input6TensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 4, DataLayout,IndexType>> gpu_input6(gpu_data_input6, input6TensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input6, input6.data(), input6TensorBuffSize);
|
||||
gpu_tensor.template chip<4l>(5l).device(sycl_device)=gpu_input6;
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k <sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (m != 5) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input6(i,j,k,l));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
gpu_tensor.device(sycl_device)=gpu_input1;
|
||||
Tensor<DataType, 5, DataLayout,IndexType> input7(tensorRange);
|
||||
input7.setRandom();
|
||||
|
||||
DataType* gpu_data_input7 = static_cast<DataType*>(sycl_device.allocate(tensorBuffSize));
|
||||
TensorMap<Tensor<DataType, 5, DataLayout,IndexType>> gpu_input7(gpu_data_input7, tensorRange);
|
||||
|
||||
sycl_device.memcpyHostToDevice(gpu_data_input7, input7.data(), tensorBuffSize);
|
||||
gpu_tensor.chip(0l,0l).device(sycl_device)=gpu_input7.chip(0l,0l);
|
||||
sycl_device.memcpyDeviceToHost(tensor.data(), gpu_data_tensor, tensorBuffSize);
|
||||
|
||||
for (int i = 0; i < sizeDim1; ++i) {
|
||||
for (int j = 0; j < sizeDim2; ++j) {
|
||||
for (int k = 0; k <sizeDim3; ++k) {
|
||||
for (int l = 0; l < sizeDim4; ++l) {
|
||||
for (int m = 0; m < sizeDim5; ++m) {
|
||||
if (i != 0) {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m));
|
||||
} else {
|
||||
VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input7(i,j,k,l,m));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
sycl_device.deallocate(gpu_data_tensor);
|
||||
sycl_device.deallocate(gpu_data_input1);
|
||||
sycl_device.deallocate(gpu_data_input2);
|
||||
sycl_device.deallocate(gpu_data_input3);
|
||||
sycl_device.deallocate(gpu_data_input4);
|
||||
sycl_device.deallocate(gpu_data_input5);
|
||||
sycl_device.deallocate(gpu_data_input6);
|
||||
sycl_device.deallocate(gpu_data_input7);
|
||||
|
||||
}
|
||||
|
||||
template<typename DataType, typename dev_Selector> void sycl_chipping_test_per_device(dev_Selector s){
|
||||
QueueInterface queueInterface(s);
|
||||
auto sycl_device = Eigen::SyclDevice(&queueInterface);
|
||||
test_static_chip_sycl<DataType, RowMajor, int64_t>(sycl_device);
|
||||
test_static_chip_sycl<DataType, ColMajor, int64_t>(sycl_device);
|
||||
test_dynamic_chip_sycl<DataType, RowMajor, int64_t>(sycl_device);
|
||||
test_dynamic_chip_sycl<DataType, ColMajor, int64_t>(sycl_device);
|
||||
test_chip_in_expr<DataType, RowMajor, int64_t>(sycl_device);
|
||||
test_chip_in_expr<DataType, ColMajor, int64_t>(sycl_device);
|
||||
test_chip_as_lvalue_sycl<DataType, RowMajor, int64_t>(sycl_device);
|
||||
test_chip_as_lvalue_sycl<DataType, ColMajor, int64_t>(sycl_device);
|
||||
}
|
||||
void test_cxx11_tensor_chipping_sycl()
|
||||
{
|
||||
for (const auto& device :Eigen::get_sycl_supported_devices()) {
|
||||
CALL_SUBTEST(sycl_chipping_test_per_device<float>(device));
|
||||
}
|
||||
}
|
Loading…
Reference in New Issue
Block a user