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

merge

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2018-08-23 11:36:49 -07:00
parent fa0bcbf230 a709c8efb4
commit d35880ed91
19 changed files with 236 additions and 228 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
unsupported/Eigen/CXX11/src/Tensor/TensorBlock.h
View File

@@ -91,7 +91,7 @@ EIGEN_STRONG_INLINE void MergeResourceRequirements(
*block_total_size = resources[0].block_total_size;
for (std::vector<TensorOpResourceRequirements>::size_type i = 1; i < resources.size(); ++i) {
if (resources[i].block_shape == kSkewedInnerDims &&
*block_shape ! kSkewedInnerDims) {
*block_shape != kSkewedInnerDims) {
*block_shape = kSkewedInnerDims;
}
*block_total_size =
@@ -152,11 +152,11 @@ struct TensorBlockCopyOp {
const Scalar* src_base = &src_data[src_index];
Scalar* dst_base = &dst_data[dst_index];
typedef const Eigen::Array<Scalar, Dynamic, 1> Src;
typedef Eigen::Array<Scalar, Dynamic, 1> Dst;
typedef const Array<Scalar, Dynamic, 1> Src;
typedef Array<Scalar, Dynamic, 1> Dst;
typedef Eigen::Map<Src, 0, InnerStride<> > SrcMap;
typedef Eigen::Map<Dst, 0, InnerStride<> > DstMap;
typedef Map<Src, 0, InnerStride<> > SrcMap;
typedef Map<Dst, 0, InnerStride<> > DstMap;
const SrcMap src(src_base, num_coeff_to_copy, InnerStride<>(src_stride));
DstMap dst(dst_base, num_coeff_to_copy, InnerStride<>(dst_stride));
@@ -178,10 +178,8 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout,
bool BlockRead>
class TensorBlockIO {
public:
typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
TensorBlock;
typedef typename TensorBlockCopyOp<Scalar, StorageIndex>
TensorBlockCopyOp;
typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
typedef TensorBlockCopyOp<Scalar, StorageIndex> BlockCopyOp;
protected:
struct BlockIteratorState {
@@ -194,7 +192,7 @@ class TensorBlockIO {
};
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Copy(
const TensorBlock& block, StorageIndex first_coeff_index,
const Block& block, StorageIndex first_coeff_index,
const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data,
Scalar* dst_data) {
@@ -290,8 +288,8 @@ class TensorBlockIO {
const StorageIndex block_total_size =
NumDims == 0 ? 1 : block.block_sizes().TotalSize();
for (StorageIndex i = 0; i < block_total_size; i += block_inner_dim_size) {
TensorBlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
dst_data, inputIndex, input_stride, src_data);
BlockCopyOp::Run(block_inner_dim_size, outputIndex, output_stride,
dst_data, inputIndex, input_stride, src_data);
// Update index.
for (int j = 0; j < num_squeezed_dims; ++j) {
if (++block_iter_state[j].count < block_iter_state[j].size) {
@@ -320,13 +318,11 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
Layout, /*BlockRead=*/true> {
public:
typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
TensorBlock;
typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true>
Base;
typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/true> Base;
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
TensorBlock* block, const Scalar* src_data) {
Block* block, const Scalar* src_data) {
array<StorageIndex, NumDims> tensor_to_block_dim_map;
for (int i = 0; i < NumDims; ++i) {
tensor_to_block_dim_map[i] = i;
@@ -336,7 +332,7 @@ class TensorBlockReader : public TensorBlockIO<Scalar, StorageIndex, NumDims,
}
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
TensorBlock* block, StorageIndex first_coeff_index,
Block* block, StorageIndex first_coeff_index,
const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
const array<StorageIndex, NumDims>& tensor_strides, const Scalar* src_data) {
Base::Copy(*block, first_coeff_index, tensor_to_block_dim_map,
@@ -357,13 +353,11 @@ template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
Layout, /*BlockRead=*/false> {
public:
typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
TensorBlock;
typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false>
Base;
typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
typedef TensorBlockIO<Scalar, StorageIndex, NumDims, Layout, /*BlockRead=*/false> Base;
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
const TensorBlock& block, Scalar* dst_data) {
const Block& block, Scalar* dst_data) {
array<StorageIndex, NumDims> tensor_to_block_dim_map;
for (int i = 0; i < NumDims; ++i) {
tensor_to_block_dim_map[i] = i;
@@ -373,7 +367,7 @@ class TensorBlockWriter : public TensorBlockIO<Scalar, StorageIndex, NumDims,
}
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void Run(
const TensorBlock& block, StorageIndex first_coeff_index,
const Block& block, StorageIndex first_coeff_index,
const array<StorageIndex, NumDims>& tensor_to_block_dim_map,
const array<StorageIndex, NumDims>& tensor_strides, Scalar* dst_data) {
Base::Copy(block, first_coeff_index, tensor_to_block_dim_map,
@@ -401,13 +395,13 @@ struct TensorBlockCwiseBinaryOp {
const StorageIndex left_stride, const LeftScalar* left_data,
const StorageIndex right_index, const StorageIndex right_stride,
const RightScalar* right_data) {
typedef const Eigen::Array<LeftScalar, Dynamic, 1> Lhs;
typedef const Eigen::Array<RightScalar, Dynamic, 1> Rhs;
typedef Eigen::Array<OutputScalar, Dynamic, 1> Out;
typedef const Array<LeftScalar, Dynamic, 1> Lhs;
typedef const Array<RightScalar, Dynamic, 1> Rhs;
typedef Array<OutputScalar, Dynamic, 1> Out;
typedef Eigen::Map<Lhs, 0, InnerStride<> > LhsMap;
typedef Eigen::Map<Rhs, 0, InnerStride<> > RhsMap;
typedef Eigen::Map<Out, 0, InnerStride<> > OutMap;
typedef Map<Lhs, 0, InnerStride<> > LhsMap;
typedef Map<Rhs, 0, InnerStride<> > RhsMap;
typedef Map<Out, 0, InnerStride<> > OutMap;
const LeftScalar* lhs_base = &left_data[left_index];
const RightScalar* rhs_base = &right_data[right_index];
@@ -417,8 +411,7 @@ struct TensorBlockCwiseBinaryOp {
const RhsMap rhs(rhs_base, num_coeff, InnerStride<>(right_stride));
OutMap out(out_base, num_coeff, InnerStride<>(output_stride));
out =
Eigen::CwiseBinaryOp<BinaryFunctor, LhsMap, RhsMap>(lhs, rhs, functor);
out = CwiseBinaryOp<BinaryFunctor, LhsMap, RhsMap>(lhs, rhs, functor);
}
};
@@ -434,8 +427,7 @@ struct TensorBlockCwiseBinaryOp {
template <typename BinaryFunctor, typename StorageIndex, typename OutputScalar,
int NumDims, int Layout>
struct TensorBlockCwiseBinaryIO {
typedef typename TensorBlock<OutputScalar, StorageIndex, NumDims,
Layout>::Dimensions Dimensions;
typedef typename TensorBlock<OutputScalar, StorageIndex, NumDims, Layout>::Dimensions Dimensions;
struct BlockIteratorState {
StorageIndex output_stride, output_span;
@@ -627,8 +619,7 @@ struct TensorBlockView {
template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
class TensorBlockMapper {
public:
typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
TensorBlock;
typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
typedef DSizes<StorageIndex, NumDims> Dimensions;
TensorBlockMapper(const Dimensions& dims,
@@ -663,7 +654,7 @@ class TensorBlockMapper {
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Block
GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
StorageIndex first_coeff_index = 0;
DSizes<StorageIndex, NumDims> coords;
@@ -711,8 +702,7 @@ class TensorBlockMapper {
}
}
return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
data);
return Block(first_coeff_index, sizes, strides, m_tensor_strides, data);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {
@@ -818,8 +808,7 @@ class TensorBlockMapper {
template <typename Scalar, typename StorageIndex, int NumDims, int Layout>
class TensorSliceBlockMapper {
public:
typedef typename TensorBlock<Scalar, StorageIndex, NumDims, Layout>
TensorBlock;
typedef TensorBlock<Scalar, StorageIndex, NumDims, Layout> Block;
typedef DSizes<StorageIndex, NumDims> Dimensions;
TensorSliceBlockMapper(const Dimensions& tensor_dims,
@@ -860,7 +849,7 @@ class TensorSliceBlockMapper {
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Block
GetBlockForIndex(StorageIndex block_index, Scalar* data) const {
StorageIndex first_coeff_index = 0;
DSizes<StorageIndex, NumDims> coords;
@@ -917,8 +906,7 @@ class TensorSliceBlockMapper {
}
}
return TensorBlock(first_coeff_index, sizes, strides, m_tensor_strides,
data);
return Block(first_coeff_index, sizes, strides, m_tensor_strides, data);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE StorageIndex total_block_count() const {

12
unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
View File

@@ -152,13 +152,7 @@ struct TensorContractionParams {
// 1. Elementwise Relu transformation following Conv2D.
// 2. AddBias to the Conv2D output channels dimension.
//
// See expected implementation in NoOpOutputKernel.
struct OutputKernel {
template <typename Index, typename Scalar>
typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
};
// Output kernel that does absolutely nothing.
// The NoOpOutputKernel implements an output kernel that does absolutely nothing.
struct NoOpOutputKernel {
/**
* Tensor contraction evaluator calls this kernel after finishing each block
@@ -177,7 +171,7 @@ struct NoOpOutputKernel {
*/
template <typename Index, typename Scalar>
EIGEN_ALWAYS_INLINE void operator()(
const OutputKernel::OutputMapper<Index, Scalar>& /*output_mapper*/,
const internal::blas_data_mapper<Scalar, Index, ColMajor>& /*output_mapper*/,
const TensorContractionParams& /*params*/, Index /*i*/,
Index /*j*/, Index /*num_rows*/, Index /*num_cols*/) const {}
};
@@ -666,7 +660,7 @@ struct TensorContractionEvaluatorBase
// call gebp (matrix kernel)
// The parameters here are copied from Eigen's GEMM implementation
const auto output_mapper = output.getSubMapper(i2, j2);
const OutputMapper output_mapper = output.getSubMapper(i2, j2);
gebp(output_mapper, blockA, blockB, actual_mc, actual_kc, actual_nc,
Scalar(1), -1, -1, 0, 0);

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

@@ -88,6 +88,7 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
typedef typename PointerType<CoeffReturnType, Device>::Type PointerT;
enum {
IsAligned = false,
@@ -106,12 +107,12 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerT data) {
if (data) {
evalTo(data);
return false;
} else {
m_result = static_cast<CoeffReturnType*>(
m_result = static_cast<PointerT>(
m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
evalTo(m_result);
return true;
@@ -139,23 +140,22 @@ struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Devi
return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return m_result; }
EIGEN_DEVICE_FUNC PointerT data() const { return m_result; }
#ifdef EIGEN_USE_SYCL
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
#endif
protected:
EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(
data, m_dimensions);
EIGEN_DEVICE_FUNC void evalTo(PointerT data) {
TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(data, m_dimensions);
m_op.func().eval(m_op.expression(), result, m_device);
}
Dimensions m_dimensions;
const ArgType m_op;
const Device& m_device;
CoeffReturnType* m_result;
PointerT m_result;
};
@@ -250,6 +250,7 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
typedef typename PointerType<CoeffReturnType, Device>::Type PointerT;
enum {
IsAligned = false,
@@ -268,12 +269,12 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(PointerT data) {
if (data) {
evalTo(data);
return false;
} else {
m_result = static_cast<Scalar *>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar)));
m_result = static_cast<PointerT>(m_device.allocate_temp(dimensions().TotalSize() * sizeof(CoeffReturnType)));
evalTo(m_result);
return true;
}
@@ -300,22 +301,22 @@ struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType,
return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
}
EIGEN_DEVICE_FUNC typename internal::traits<XprType>::PointerType data() const { return m_result; }
EIGEN_DEVICE_FUNC PointerT data() const { return m_result; }
#ifdef EIGEN_USE_SYCL
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Device& device() const { return m_device; }
#endif
protected:
EIGEN_DEVICE_FUNC void evalTo(Scalar* data) {
TensorMap<Tensor<Scalar, NumDims, Layout> > result(data, m_dimensions);
EIGEN_DEVICE_FUNC void evalTo(PointerT data) {
TensorMap<Tensor<CoeffReturnType, NumDims, Layout> > result(data, m_dimensions);
m_op.func().eval(m_op.lhsExpression(), m_op.rhsExpression(), result, m_device);
}
Dimensions m_dimensions;
const XprType m_op;
const Device& m_device;
CoeffReturnType* m_result;
PointerT m_result;
};

8
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@@ -132,7 +132,7 @@ class TensorExecutor<Expression, DefaultDevice, Vectorizable,
if (needs_assign) {
// Size tensor blocks to fit in cache (or requested target block size).
Index block_total_size = numext::mini(cache_size, total_size);
TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
TensorBlockShapeType block_shape = kSkewedInnerDims;
// Query expression tree for desired block size/shape.
std::vector<TensorOpResourceRequirements> resources;
evaluator.getResourceRequirements(&resources);
@@ -229,10 +229,6 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, Tileable> {
Evaluator evaluator(expr, device);
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) {
const StorageIndex PacketSize =
Vectorizable
? unpacket_traits<typename Evaluator::PacketReturnType>::size
: 1;
const StorageIndex size = array_prod(evaluator.dimensions());
device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
EvalRange::alignBlockSize,
@@ -272,7 +268,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable, /*Tileable*/ tr
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign) {
TensorBlockShapeType block_shape = TensorBlockShapeType::kSkewedInnerDims;
TensorBlockShapeType block_shape = kSkewedInnerDims;
Index block_total_size = 0;
// Query expression tree for desired block size/shape.
std::vector<internal::TensorOpResourceRequirements> resources;

8
unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
View File

@@ -24,6 +24,14 @@ template<typename T> struct MakePointer {
typedef T ScalarType;
};
// The PointerType class is a container of the device specefic pointer
// used for refering to a Pointer on TensorEvaluator class. While the TensorExpression
// is a device-agnostic type and need MakePointer class for type conversion,
// the TensorEvaluator calss can be specialized for a device, hence it is possible
// to construct different types of temproray storage memory in TensorEvaluator
// for different devices by specializing the following PointerType class.
template<typename T, typename Device> struct PointerType : MakePointer<T>{};
namespace internal{
template<typename A, typename B> struct Pointer_type_promotion {
static const bool val=false;