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

23
unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h
View File

@@ -131,6 +131,8 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
typedef OutputScalar CoeffReturnType;
typedef typename PacketType<OutputScalar, Device>::type PacketReturnType;
static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
typedef StorageMemory<CoeffReturnType, Device> Storage;
typedef typename Storage::Type EvaluatorPointerType;
enum {
IsAligned = false,
@@ -167,13 +169,13 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
return m_dimensions;
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(OutputScalar* data) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
m_impl.evalSubExprsIfNeeded(NULL);
if (data) {
evalToBuf(data);
return false;
} else {
m_data = (CoeffReturnType*)m_device.allocate(sizeof(CoeffReturnType) * m_size);
m_data = (EvaluatorPointerType)m_device.get((CoeffReturnType*)(m_device.allocate_temp(sizeof(CoeffReturnType) * m_size)));
evalToBuf(m_data);
return true;
}
@@ -202,11 +204,16 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC Scalar* data() const { return m_data; }
EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_data; }
#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_data.bind(cgh);
}
#endif
private:
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalToBuf(OutputScalar* data) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalToBuf(EvaluatorPointerType data) {
const bool write_to_out = internal::is_same<OutputScalar, ComplexScalar>::value;
ComplexScalar* buf = write_to_out ? (ComplexScalar*)data : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * m_size);
@@ -576,12 +583,12 @@ struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, D
protected:
Index m_size;
const FFT& m_fft;
const FFT EIGEN_DEVICE_REF m_fft;
Dimensions m_dimensions;
array<Index, NumDims> m_strides;
TensorEvaluator<ArgType, Device> m_impl;
CoeffReturnType* m_data;
const Device& m_device;
EvaluatorPointerType m_data;
const Device EIGEN_DEVICE_REF m_device;
// This will support a maximum FFT size of 2^32 for each dimension
// m_sin_PI_div_n_LUT[i] = (-2) * std::sin(M_PI / std::pow(2,i)) ^ 2;