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 Backend]

Browse Source 
* Adding Missing operations for vector comparison in SYCL. This caused compiler error for vector comparison when compiling SYCL
 * Fixing the compiler error for placement new in TensorForcedEval.h This caused compiler error when compiling SYCL backend
 * Reducing the SYCL warning by  removing the abort function inside the kernel
 * Adding Strong inline to functions inside SYCL interop.
This commit is contained in:
mehdi-goli
2020-01-07 15:13:37 +00:00
parent eedb7eeacf
commit d0ae052da4
3 changed files with 149 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
View File

@@ -77,6 +77,28 @@ class TensorForcedEvalOp : public TensorBase<TensorForcedEvalOp<XprType>, ReadOn
typename XprType::Nested m_xpr;
};
namespace internal {
template <typename Device, typename CoeffReturnType>
struct non_integral_type_placement_new{
template <typename StorageType>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index numValues, StorageType m_buffer) {
// Initialize non-trivially constructible types.
if (!internal::is_arithmetic<CoeffReturnType>::value) {
for (Index i = 0; i < numValues; ++i) new (m_buffer + i) CoeffReturnType();
}
}
};
// SYCL does not support non-integral types
// having new (m_buffer + i) CoeffReturnType() causes the following compiler error for SYCL Devices
// no matching function for call to 'operator new'
template <typename CoeffReturnType>
struct non_integral_type_placement_new<Eigen::SyclDevice, CoeffReturnType> {
template <typename StorageType>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index, StorageType) {
}
};
} // end namespace internal
template<typename ArgType_, typename Device>
struct TensorEvaluator<const TensorForcedEvalOp<ArgType_>, Device>
@@ -127,10 +149,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType_>, Device>
const Index numValues = internal::array_prod(m_impl.dimensions());
m_buffer = m_device.get((CoeffReturnType*)m_device.allocate_temp(numValues * sizeof(CoeffReturnType)));
// Initialize non-trivially constructible types.
if (!internal::is_arithmetic<CoeffReturnType>::value) {
for (Index i = 0; i < numValues; ++i) new (m_buffer + i) CoeffReturnType();
}
internal::non_integral_type_placement_new<Device, CoeffReturnType>()(numValues, m_buffer);
typedef TensorEvalToOp< const typename internal::remove_const<ArgType>::type > EvalTo;
EvalTo evalToTmp(m_device.get(m_buffer), m_op);