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

Adding TensorIndexTuple and TensorTupleReduceOP backend (ArgMax/Min) for sycl; fixing the address space issue for const TensorMap; converting all discard_write to write due to data missmatch.

Browse Source
This commit is contained in:
Mehdi Goli
2017-03-07 14:27:10 +00:00
parent 8296b87d7b
commit f84963ed95
19 changed files with 813 additions and 226 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
unsupported/Eigen/CXX11/src/Tensor/TensorConvolutionSycl.h
View File

@@ -45,7 +45,7 @@ EigenConvolutionKernel1D(internal::IndexMapper<Index, InputDims, 1, Eigen::inter
void operator()(cl::sycl::nd_item<2> itemID) {
typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -103,7 +103,7 @@ EigenConvolutionKernel2D(internal::IndexMapper<Index, InputDims, 2, Eigen::inter
void operator()(cl::sycl::nd_item<3> itemID) {
typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -173,7 +173,7 @@ EigenConvolutionKernel3D(internal::IndexMapper<Index, InputDims, 3, Eigen::inter
void operator()(cl::sycl::nd_item<3> itemID) {
typedef typename TensorSycl::internal::ConvertToDeviceExpression<HostExpr>::Type DevExpr;
auto device_expr =TensorSycl::internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
auto device_evaluator = Eigen::TensorEvaluator<DevExpr, Eigen::SyclKernelDevice>(device_expr.expr, Eigen::SyclKernelDevice());
auto buffer_ptr = ConvertToActualTypeSycl(CoeffReturnType, buffer_acc);
auto kernel_ptr = ConvertToActualTypeSycl(KernelType, kernel_filter);
@@ -339,8 +339,8 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
// create input tuple of accessors
InputTupleType tuple_of_accessors = Eigen::TensorSycl::internal::createTupleOfAccessors<InputEvaluator>(cgh, m_inputImpl);
typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> OutputAccessorType;
OutputAccessorType out_res= m_device. template get_sycl_accessor<cl::sycl::access::mode::discard_write>(cgh, data);
typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::write, cl::sycl::access::target::global_buffer> OutputAccessorType;
OutputAccessorType out_res= m_device. template get_sycl_accessor<cl::sycl::access::mode::write>(cgh, data);
typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer> KernelAccessorType;
KernelAccessorType kernel_acc= m_device. template get_sycl_accessor<cl::sycl::access::mode::read>(cgh, m_kernel);