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

Converting all parallel for lambda to functor in order to prevent kernel duplication name error; adding tensorConcatinationOp backend for sycl.

Browse Source
This commit is contained in:
Mehdi Goli
2016-12-16 19:46:45 +00:00
parent 7949849ebc
commit 35bae513a0
16 changed files with 588 additions and 264 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
View File

@@ -25,6 +25,31 @@
namespace Eigen {
namespace TensorSycl {
template<typename Expr, typename FunctorExpr, typename TupleType > struct ExecExprFunctorKernel{
typedef typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr;
typedef typename Expr::Index Index;
Index range;
FunctorExpr functors;
TupleType tuple_of_accessors;
ExecExprFunctorKernel(Index range_
,
FunctorExpr functors_, TupleType tuple_of_accessors_
)
:range(range_)
, functors(functors_), tuple_of_accessors(tuple_of_accessors_)
{}
void operator()(cl::sycl::nd_item<1> itemID) {
typedef typename internal::ConvertToDeviceExpression<Expr>::Type DevExpr;
auto device_expr =internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
typename DevExpr::Index gId = static_cast<typename DevExpr::Index>(itemID.get_global_linear_id());
if (gId < range)
device_evaluator.evalScalar(gId);
}
};
/// The run function in tensor sycl convert the expression tree to a buffer
/// based expression tree;
/// creates the expression tree for the device with accessor to buffers;
@@ -34,25 +59,19 @@ void run(Expr &expr, Dev &dev) {
Eigen::TensorEvaluator<Expr, Dev> evaluator(expr, dev);
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) {
typedef typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr;
auto functors = internal::extractFunctors(evaluator);
typedef decltype(functors) FunctorExpr;
dev.sycl_queue().submit([&](cl::sycl::handler &cgh) {
// create a tuple of accessors from Evaluator
auto tuple_of_accessors = internal::createTupleOfAccessors<decltype(evaluator)>(cgh, evaluator);
typedef decltype(tuple_of_accessors) TupleType;
typename Expr::Index range, GRange, tileSize;
dev.parallel_for_setup(static_cast<typename Expr::Index>(evaluator.dimensions().TotalSize()), tileSize, range, GRange);
// run the kernel
cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
typedef typename internal::ConvertToDeviceExpression<Expr>::Type DevExpr;
auto device_expr =internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
typename DevExpr::Index gId = static_cast<typename DevExpr::Index>(itemID.get_global_linear_id());
if (gId < range) {
device_evaluator.evalScalar(gId);
}
});
cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)),
ExecExprFunctorKernel<Expr,FunctorExpr,TupleType>(range
, functors, tuple_of_accessors
));
});
dev.asynchronousExec();
}