mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-15 07:10:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

Created a mechanism to enable contraction mappers to determine the best blocking strategy.

Browse Source
This commit is contained in:
Benoit Steiner 2016-01-22 14:37:26 -08:00
parent 5358c38589
commit 4beb447e27
2 changed files with 59 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
unsupported/Eigen/CXX11/Tensor
View File

@ -89,6 +89,7 @@ typedef unsigned __int64 uint64_t;
#include "src/Tensor/TensorArgMax.h"
#include "src/Tensor/TensorConcatenation.h"
#include "src/Tensor/TensorContractionMapper.h"
#include "src/Tensor/TensorContractionBlocking.h"
#include "src/Tensor/TensorContraction.h"
#include "src/Tensor/TensorContractionThreadPool.h"
#include "src/Tensor/TensorContractionCuda.h"

58
unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h Normal file
View File

@ -0,0 +1,58 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H
#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H
namespace Eigen {
namespace internal {
enum {
ShardByRow = 0,
ShardByCol = 1
};
// Default Blocking Strategy
template <typename LhsMapper, typename RhsMapper, typename Index, int ShardingType=ShardByCol>
class TensorContractionBlocking {
public:
typedef typename LhsMapper::Scalar LhsScalar;
typedef typename RhsMapper::Scalar RhsScalar;
TensorContractionBlocking(Index k, Index m, Index n, Index num_threads = 1) :
kc_(k), mc_(m), nc_(n)
{
if (ShardingType == ShardByCol) {
computeProductBlockingSizes<LhsScalar, RhsScalar, 1>(kc_, mc_, nc_, num_threads);
}
else {
if (kc_ && mc_ && nc_) {
mc_ = (((m / num_threads) + 15) / 16) * 16;
}
}
}
EIGEN_ALWAYS_INLINE Index kc() const { return kc_; }
EIGEN_ALWAYS_INLINE Index mc() const { return mc_; }
EIGEN_ALWAYS_INLINE Index nc() const { return nc_; }
private:
Index kc_;
Index mc_;
Index nc_;
};
} // end namespace internal
} // end namespace Eigen
#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_BLOCKING_H