提交 53b0e427 编写于 作者: H hedaoyuan

Add EigenGemm.

上级 9eaef753
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <glog/logging.h>
#include "unsupported/Eigen/CXX11/Tensor"
namespace paddle {
template <class T>
struct EigenBlasGemm {
typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
Eigen::Aligned>
Matrix;
static void compute(const bool transA,
const bool transB,
const int M,
const int N,
const int K,
const T alpha,
const T* A,
const int lda,
const T* B,
const int ldb,
const T beta,
T* C,
const int ldc) {
Eigen::array<int, 2> sizeA;
if (transA) {
sizeA[0] = K;
sizeA[1] = M;
CHECK_EQ(M, lda);
} else {
sizeA[0] = M;
sizeA[1] = K;
CHECK_EQ(K, lda);
}
Eigen::array<int, 2> sizeB;
if (transB) {
sizeB[0] = N;
sizeB[1] = K;
CHECK_EQ(K, ldb);
} else {
sizeB[0] = K;
sizeB[1] = N;
CHECK_EQ(N, ldb);
}
Eigen::array<int, 2> sizeC;
sizeC[0] = M;
sizeC[1] = N;
CHECK_EQ(N, ldc);
const Matrix a(const_cast<T*>(A), sizeA);
const Matrix b(const_cast<T*>(B), sizeB);
Matrix c(C, sizeC);
typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
Eigen::array<DimPair, 1> dims;
dims[0] = DimPair(1, 0);
dims[0].first = transA ? 0 : 1;
dims[0].second = transB ? 1 : 0;
Eigen::DefaultDevice device;
if (alpha == T(1) && beta == T(0)) {
c.device(device) = a.contract(b, dims);
} else if (alpha == T(1) && beta == T(1)) {
c.device(device) += a.contract(b, dims);
} else {
c.device(device) =
c.constant(alpha) * a.contract(b, dims) + c.constant(beta) * c;
}
}
};
#ifdef PADDLE_TYPE_DOUBLE
template class EigenBlasGemm<double>;
#else
template class EigenBlasGemm<float>;
#endif
} // namespace paddle
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "GemmFunctor.h"
#include "paddle/math/MathFunctions.h"
namespace paddle {
template <class T>
struct BlasGemm<DEVICE_TYPE_CPU, T> {
static void compute(const bool transA,
const bool transB,
const int M,
const int N,
const int K,
const T alpha,
const T* A,
const int lda,
const T* B,
const int ldb,
const T beta,
T* C,
const int ldc) {
gemm<T>(transA == false ? CblasNoTrans : CblasTrans,
transB == false ? CblasNoTrans : CblasTrans,
M,
N,
K,
alpha,
A,
lda,
B,
ldb,
beta,
C,
ldc);
}
};
template <class T>
struct BlasGemm<DEVICE_TYPE_GPU, T> {
static void compute(const bool transA,
const bool transB,
const int M,
const int N,
const int K,
const T alpha,
const T* A,
const int lda,
const T* B,
const int ldb,
const T beta,
T* C,
const int ldc) {
hl_matrix_mul((T*)A,
transA == false ? HPPL_OP_N : HPPL_OP_T,
(T*)B,
transB == false ? HPPL_OP_N : HPPL_OP_T,
C,
M,
N,
K,
alpha,
beta,
lda,
ldb,
ldc);
}
};
template class BlasGemm<DEVICE_TYPE_CPU, real>;
template class BlasGemm<DEVICE_TYPE_GPU, real>;
} // namespace paddle
...@@ -14,7 +14,7 @@ limitations under the License. */ ...@@ -14,7 +14,7 @@ limitations under the License. */
#pragma once #pragma once
#include "paddle/math/MathFunctions.h" #include "TensorType.h"
namespace paddle { namespace paddle {
...@@ -24,10 +24,9 @@ namespace paddle { ...@@ -24,10 +24,9 @@ namespace paddle {
// of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul
// interface. // interface.
template <DeviceType Device, class T> template <DeviceType Device, class T>
class GemmFunctor { struct BlasGemm {
public: static void compute(const bool transA,
void operator()(const CBLAS_TRANSPOSE transA, const bool transB,
const CBLAS_TRANSPOSE TransB,
const int M, const int M,
const int N, const int N,
const int K, const int K,
...@@ -41,11 +40,15 @@ public: ...@@ -41,11 +40,15 @@ public:
const int ldc); const int ldc);
}; };
// TODO(hedaoyuan): Since the definition of the real type in the Paddle
// conflicts with the Eigen library, so compile the Eigen code can not
// include the Paddle header file. And need an EigenBlasGemm template class
// that does not contain the DeviceType parameter.
// I will fix this problem and merge BlasGemm and EigenBlasGemm into one.
template <class T> template <class T>
class GemmFunctor<DEVICE_TYPE_CPU, T> { struct EigenBlasGemm {
public: static void compute(const bool transA,
void operator()(const CBLAS_TRANSPOSE transA, const bool transB,
const CBLAS_TRANSPOSE TransB,
const int M, const int M,
const int N, const int N,
const int K, const int K,
...@@ -56,41 +59,7 @@ public: ...@@ -56,41 +59,7 @@ public:
const int ldb, const int ldb,
const T beta, const T beta,
T* C, T* C,
const int ldc) { const int ldc);
gemm<T>(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
}
};
template <class T>
class GemmFunctor<DEVICE_TYPE_GPU, T> {
public:
void operator()(const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE TransB,
const int M,
const int N,
const int K,
const T alpha,
const T* A,
const int lda,
const T* B,
const int ldb,
const T beta,
T* C,
const int ldc) {
hl_matrix_mul((T*)A,
transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
(T*)B,
TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
C,
M,
N,
K,
alpha,
beta,
lda,
ldb,
ldc);
}
}; };
} // namespace paddle } // namespace paddle
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册