more acml removal
This commit is contained in:
Wolfgang Manousek
Родитель
40d247841a
Коммит
79cfcf7d4f
|
|
@ -37,14 +37,8 @@
|
|||
#pragma warning(disable : 4244) // unreachable code; triggered for unknown reasons
|
||||
#pragma warning(disable : 4702) // conversion from 'double' to 'float'
|
||||
|
||||
#ifdef USE_ACML
|
||||
// Download ACML 5.3.1 (e.g., acml5.3.1-ifort64.exe) or above
|
||||
// from http://developer.amd.com/tools/cpu-development/amd-core-math-library-acml/acml-downloads-resources/
|
||||
// Install the ifort64_mp variant (compiled with intel compiler) of the library
|
||||
// Set Environment variable ACML_PATH to C:\AMD\acml5.3.1\ifort64_mp or the folder you installed acml
|
||||
// to point to your folder for the include file and link library
|
||||
#include <acml.h> // requires ACML 5.3.1 and above
|
||||
#elif defined(USE_MKL)
|
||||
|
||||
#ifdef defined(USE_MKL)
|
||||
// requires MKL 10.0 and above
|
||||
#include <mkl.h>
|
||||
#else
|
||||
|
|
@ -57,12 +51,6 @@
|
|||
#include <lapacke.h>
|
||||
#endif
|
||||
|
||||
#ifdef USE_ACML // MKL has one additional parameter for different matrix order
|
||||
#define BLAS_COLMAJOR
|
||||
#else
|
||||
#define BLAS_COLMAJOR (int) MatrixOrder::ColMajor,
|
||||
#endif
|
||||
|
||||
#define SWAP(a, b) \
|
||||
{ \
|
||||
(a) ^= (b); \
|
||||
|
|
@ -912,11 +900,7 @@ void CPUMatrix<ElemType>::SetValue(const size_t numRows, const size_t numCols, E
|
|||
#pragma omp parallel for
|
||||
foreach_column (j, us)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
dcopy((int) numRows, reinterpret_cast<double*>(pArray + j), (int) numCols, reinterpret_cast<double*>(bufPtr + LocateColumn(j)), 1);
|
||||
#else
|
||||
cblas_dcopy((int) numRows, reinterpret_cast<double*>(pArray + j), (int) numCols, reinterpret_cast<double*>(bufPtr + LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -926,11 +910,7 @@ void CPUMatrix<ElemType>::SetValue(const size_t numRows, const size_t numCols, E
|
|||
{
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
scopy((int) numRows, reinterpret_cast<float*>(pArray + j), (int) numCols, reinterpret_cast<float*>(bufPtr + LocateColumn(j)), 1);
|
||||
#else
|
||||
cblas_scopy((int) numRows, reinterpret_cast<float*>(pArray + j), (int) numCols, reinterpret_cast<float*>(bufPtr + LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -2844,20 +2824,12 @@ ElemType CPUMatrix<ElemType>::SumOfAbsElements() const
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
return (ElemType) dasum((int) GetNumElements(), reinterpret_cast<double*>(Data()), 1);
|
||||
#else
|
||||
return (ElemType) cblas_dasum((int) GetNumElements(), reinterpret_cast<double*>(Data()), 1);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
return sasum((int) GetNumElements(), reinterpret_cast<float*>(Data()), 1);
|
||||
#else
|
||||
return cblas_sasum((int) GetNumElements(), reinterpret_cast<float*>(Data()), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -3028,11 +3000,7 @@ void CPUMatrix<ElemType>::VectorNorm2(CPUMatrix<ElemType>& c, const bool isColWi
|
|||
#pragma omp parallel for
|
||||
foreach_column (j, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = (ElemType) dnrm2(m, reinterpret_cast<double*>(bufPtr + us.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = (ElemType) cblas_dnrm2(m, reinterpret_cast<double*>(bufPtr + us.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -3041,11 +3009,7 @@ void CPUMatrix<ElemType>::VectorNorm2(CPUMatrix<ElemType>& c, const bool isColWi
|
|||
foreach_column (j, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = snrm2(m, reinterpret_cast<float*>(bufPtr + us.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = cblas_snrm2(m, reinterpret_cast<float*>(bufPtr + us.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -3058,11 +3022,7 @@ void CPUMatrix<ElemType>::VectorNorm2(CPUMatrix<ElemType>& c, const bool isColWi
|
|||
#pragma omp parallel for
|
||||
foreach_row (i, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = dnrm2(n, reinterpret_cast<double*>(bufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = cblas_dnrm2(n, reinterpret_cast<double*>(bufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -3071,11 +3031,7 @@ void CPUMatrix<ElemType>::VectorNorm2(CPUMatrix<ElemType>& c, const bool isColWi
|
|||
foreach_row (i, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = snrm2(n, reinterpret_cast<float*>(bufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = cblas_snrm2(n, reinterpret_cast<float*>(bufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -4486,34 +4442,22 @@ void CPUMatrix<ElemType>::MultiplyAndWeightedAdd(ElemType alpha, const CPUMatrix
|
|||
|
||||
int m, n, k, l;
|
||||
int lda, ldb, ldc;
|
||||
#ifdef USE_ACML
|
||||
char transA, transB;
|
||||
#else
|
||||
CBLAS_TRANSPOSE mklTransA;
|
||||
CBLAS_TRANSPOSE mklTransB;
|
||||
#endif
|
||||
|
||||
if (transposeA)
|
||||
{
|
||||
m = (int) a.GetNumCols();
|
||||
k = (int) a.GetNumRows();
|
||||
lda = k;
|
||||
#ifdef USE_ACML
|
||||
transA = (char) MatrixTranspose::Trans;
|
||||
#else
|
||||
mklTransA = CBLAS_TRANSPOSE::CblasTrans;
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
m = (int) a.GetNumRows();
|
||||
k = (int) a.GetNumCols();
|
||||
lda = m;
|
||||
#ifdef USE_ACML
|
||||
transA = (char) MatrixTranspose::NoTrans;
|
||||
#else
|
||||
mklTransA = CBLAS_TRANSPOSE::CblasNoTrans;
|
||||
#endif
|
||||
}
|
||||
|
||||
if (transposeB)
|
||||
|
|
@ -4521,22 +4465,14 @@ void CPUMatrix<ElemType>::MultiplyAndWeightedAdd(ElemType alpha, const CPUMatrix
|
|||
l = (int) b.GetNumCols();
|
||||
n = (int) b.GetNumRows();
|
||||
ldb = n;
|
||||
#ifdef USE_ACML
|
||||
transB = (char) MatrixTranspose::Trans;
|
||||
#else
|
||||
mklTransB = CBLAS_TRANSPOSE::CblasTrans;
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
l = (int) b.GetNumRows();
|
||||
n = (int) b.GetNumCols();
|
||||
ldb = l;
|
||||
#ifdef USE_ACML
|
||||
transB = (char) MatrixTranspose::NoTrans;
|
||||
#else
|
||||
mklTransB = CBLAS_TRANSPOSE::CblasNoTrans;
|
||||
#endif
|
||||
}
|
||||
|
||||
assert(m > 0 && k > 0 && l > 0 && n > 0); // converting from size_t to int may cause overflow
|
||||
|
|
@ -4553,20 +4489,12 @@ void CPUMatrix<ElemType>::MultiplyAndWeightedAdd(ElemType alpha, const CPUMatrix
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
dgemm(transA, transB, m, n, k, alpha, reinterpret_cast<double*>(a.Data()), lda, reinterpret_cast<double*>(b.Data()), ldb, beta, reinterpret_cast<double*>(c.Data()), ldc);
|
||||
#else
|
||||
cblas_dgemm((CBLAS_ORDER) BLAS_COLMAJOR mklTransA, mklTransB, m, n, k, alpha, reinterpret_cast<double*>(a.Data()), lda, reinterpret_cast<double*>(b.Data()), ldb, beta, reinterpret_cast<double*>(c.Data()), ldc);
|
||||
#endif
|
||||
cblas_dgemm((CBLAS_ORDER) (int)MatrixOrder::ColMajor, mklTransA, mklTransB, m, n, k, alpha, reinterpret_cast<double*>(a.Data()), lda, reinterpret_cast<double*>(b.Data()), ldb, beta, reinterpret_cast<double*>(c.Data()), ldc);
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
sgemm(BLAS_COLMAJOR transA, transB, m, n, k, alpha, reinterpret_cast<float*>(a.Data()), lda, reinterpret_cast<float*>(b.Data()), ldb, beta, reinterpret_cast<float*>(c.Data()), ldc);
|
||||
#else
|
||||
cblas_sgemm((CBLAS_ORDER) BLAS_COLMAJOR mklTransA, mklTransB, m, n, k, alpha, reinterpret_cast<float*>(a.Data()), lda, reinterpret_cast<float*>(b.Data()), ldb, beta, reinterpret_cast<float*>(c.Data()), ldc);
|
||||
#endif
|
||||
cblas_sgemm((CBLAS_ORDER) (int)MatrixOrder::ColMajor, mklTransA, mklTransB, m, n, k, alpha, reinterpret_cast<float*>(a.Data()), lda, reinterpret_cast<float*>(b.Data()), ldb, beta, reinterpret_cast<float*>(c.Data()), ldc);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -4611,9 +4539,7 @@ void CPUMatrix<ElemType>::SVD(const CPUMatrix<ElemType>& A, CPUMatrix<ElemType>&
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
dgesvd('A', 'A', (int) m, (int) n, reinterpret_cast<double*>(A.Data()), (int) lda, reinterpret_cast<double*>(SIGMA.Data()), reinterpret_cast<double*>(U.Data()), (int) ldu, reinterpret_cast<double*>(VT.Data()), (int) ldvt, &info);
|
||||
#elif defined(USE_MKL)
|
||||
#ifdef USE_MKL
|
||||
double wkopt;
|
||||
int lwork = -1;
|
||||
dgesvd("All", "All", &m, &n, reinterpret_cast<double*>(A.Data()), &lda, reinterpret_cast<double*>(SIGMA.Data()), reinterpret_cast<double*>(U.Data()), &ldu, reinterpret_cast<double*>(VT.Data()), &ldvt, &wkopt, &lwork, &info);
|
||||
|
|
@ -4622,16 +4548,13 @@ void CPUMatrix<ElemType>::SVD(const CPUMatrix<ElemType>& A, CPUMatrix<ElemType>&
|
|||
dgesvd("All", "All", &m, &n, reinterpret_cast<double*>(A.Data()), &lda, reinterpret_cast<double*>(SIGMA.Data()), reinterpret_cast<double*>(U.Data()), &ldu, reinterpret_cast<double*>(VT.Data()), &ldvt, reinterpret_cast<double*>(W.Data()), &lwork, &info);
|
||||
#else
|
||||
std::vector<double> superb(std::max(std::min(m, n) - 1, 1));
|
||||
info = LAPACKE_dgesvd(BLAS_COLMAJOR 'A', 'A', (int) m, (int) n, reinterpret_cast<double*>(A.Data()), (int) lda, reinterpret_cast<double*>(SIGMA.Data()),
|
||||
info = LAPACKE_dgesvd((int) MatrixOrder::ColMajor, 'A', 'A', (int) m, (int) n, reinterpret_cast<double*>(A.Data()), (int) lda, reinterpret_cast<double*>(SIGMA.Data()),
|
||||
reinterpret_cast<double*>(U.Data()), (int) ldu, reinterpret_cast<double*>(VT.Data()), (int) ldvt, &superb[0]);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
#pragma warning(suppress : 4244)
|
||||
sgesvd('A', 'A', (int) m, (int) n, reinterpret_cast<float*>(A.Data()), (int) lda, reinterpret_cast<float*>(SIGMA.Data()), reinterpret_cast<float*>(U.Data()), (int) ldu, reinterpret_cast<float*>(VT.Data()), (int) ldvt, &info);
|
||||
#elif defined(USE_MKL)
|
||||
#ifdef USE_MKL
|
||||
float wkopt;
|
||||
int lwork = -1;
|
||||
sgesvd("All", "All", &m, &n, reinterpret_cast<float*>(A.Data()), &lda, reinterpret_cast<float*>(SIGMA.Data()), reinterpret_cast<float*>(U.Data()), &ldu, reinterpret_cast<float*>(VT.Data()), &ldvt, &wkopt, &lwork, &info);
|
||||
|
|
@ -4640,7 +4563,7 @@ void CPUMatrix<ElemType>::SVD(const CPUMatrix<ElemType>& A, CPUMatrix<ElemType>&
|
|||
sgesvd("All", "All", &m, &n, reinterpret_cast<float*>(A.Data()), &lda, reinterpret_cast<float*>(SIGMA.Data()), reinterpret_cast<float*>(U.Data()), &ldu, reinterpret_cast<float*>(VT.Data()), &ldvt, reinterpret_cast<float*>(W.Data()), &lwork, &info);
|
||||
#else
|
||||
std::vector<float> superb(std::max(std::min(m, n) - 1, 1));
|
||||
info = LAPACKE_sgesvd(BLAS_COLMAJOR 'A', 'A', (int) m, (int) n, reinterpret_cast<float*>(A.Data()), (int) lda, reinterpret_cast<float*>(SIGMA.Data()),
|
||||
info = LAPACKE_sgesvd((int) MatrixOrder::ColMajor, 'A', 'A', (int) m, (int) n, reinterpret_cast<float*>(A.Data()), (int) lda, reinterpret_cast<float*>(SIGMA.Data()),
|
||||
reinterpret_cast<float*>(U.Data()), (int) ldu, reinterpret_cast<float*>(VT.Data()), (int) ldvt, &superb[0]);
|
||||
#endif
|
||||
}
|
||||
|
|
@ -4837,20 +4760,12 @@ void CPUMatrix<ElemType>::ScaleAndAdd(ElemType alpha, const CPUMatrix<ElemType>&
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
daxpy(len, alpha, reinterpret_cast<double*>(a.Data()), incx, reinterpret_cast<double*>(c.Data()), incy);
|
||||
#else
|
||||
cblas_daxpy(len, alpha, reinterpret_cast<double*>(a.Data()), incx, reinterpret_cast<double*>(c.Data()), incy);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
saxpy(len, alpha, reinterpret_cast<float*>(a.Data()), incx, reinterpret_cast<float*>(c.Data()), incy);
|
||||
#else
|
||||
cblas_saxpy(len, alpha, reinterpret_cast<float*>(a.Data()), incx, reinterpret_cast<float*>(c.Data()), incy);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else if (a.GetNumElements() == 1) // scalar, add to all elements
|
||||
|
|
@ -4889,11 +4804,7 @@ void CPUMatrix<ElemType>::ScaleAndAdd(ElemType alpha, const CPUMatrix<ElemType>&
|
|||
#pragma omp parallel for
|
||||
foreach_column (j, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
daxpy(m, alpha, reinterpret_cast<double*>(aBufPtr), 1, reinterpret_cast<double*>(cBufPtr + c.LocateColumn(j)), 1);
|
||||
#else
|
||||
cblas_daxpy(m, alpha, reinterpret_cast<double*>(aBufPtr), 1, reinterpret_cast<double*>(cBufPtr + c.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -4902,11 +4813,7 @@ void CPUMatrix<ElemType>::ScaleAndAdd(ElemType alpha, const CPUMatrix<ElemType>&
|
|||
foreach_column (j, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
saxpy(m, alpha, reinterpret_cast<float*>(aBufPtr), 1, reinterpret_cast<float*>(cBufPtr + c.LocateColumn(j)), 1);
|
||||
#else
|
||||
cblas_saxpy(m, alpha, reinterpret_cast<float*>(aBufPtr), 1, reinterpret_cast<float*>(cBufPtr + c.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -4925,11 +4832,7 @@ void CPUMatrix<ElemType>::ScaleAndAdd(ElemType alpha, const CPUMatrix<ElemType>&
|
|||
#pragma omp parallel for
|
||||
foreach_row (i, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
daxpy(n, alpha, reinterpret_cast<double*>(aBufPtr), 1, reinterpret_cast<double*>(cBufPtr + i), m);
|
||||
#else
|
||||
cblas_daxpy(n, alpha, reinterpret_cast<double*>(aBufPtr), 1, reinterpret_cast<double*>(cBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -4938,11 +4841,7 @@ void CPUMatrix<ElemType>::ScaleAndAdd(ElemType alpha, const CPUMatrix<ElemType>&
|
|||
foreach_row (i, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
saxpy(n, alpha, reinterpret_cast<float*>(aBufPtr), 1, reinterpret_cast<float*>(cBufPtr + i), m);
|
||||
#else
|
||||
cblas_saxpy(n, alpha, reinterpret_cast<float*>(aBufPtr), 1, reinterpret_cast<float*>(cBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -5163,20 +5062,12 @@ template <class ElemType>
|
|||
}
|
||||
else if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
dscal(len, alpha, reinterpret_cast<double*>(a.Data()), incx); // TODO: Use overloads.
|
||||
#else
|
||||
cblas_dscal(len, alpha, reinterpret_cast<double*>(a.Data()), incx);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
sscal(len, alpha, reinterpret_cast<float*>(a.Data()), incx);
|
||||
#else
|
||||
cblas_sscal(len, alpha, reinterpret_cast<float*>(a.Data()), incx);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -5224,11 +5115,7 @@ void CPUMatrix<ElemType>::InnerProduct(const CPUMatrix<ElemType>& a, const CPUMa
|
|||
#pragma omp parallel for
|
||||
foreach_column (j, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = (ElemType) ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = (ElemType) cblas_ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -5237,11 +5124,7 @@ void CPUMatrix<ElemType>::InnerProduct(const CPUMatrix<ElemType>& a, const CPUMa
|
|||
foreach_column (j, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = (ElemType) sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = (ElemType) cblas_sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -5256,11 +5139,7 @@ void CPUMatrix<ElemType>::InnerProduct(const CPUMatrix<ElemType>& a, const CPUMa
|
|||
#pragma omp parallel for
|
||||
foreach_row (i, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = ddot(n, reinterpret_cast<double*>(aBufPtr + i), m, reinterpret_cast<double*>(bBufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = cblas_ddot(n, reinterpret_cast<double*>(aBufPtr + i), m, reinterpret_cast<double*>(bBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -5269,11 +5148,7 @@ void CPUMatrix<ElemType>::InnerProduct(const CPUMatrix<ElemType>& a, const CPUMa
|
|||
foreach_row (i, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = sdot(n, reinterpret_cast<float*>(aBufPtr + i), m, reinterpret_cast<float*>(bBufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = cblas_sdot(n, reinterpret_cast<float*>(aBufPtr + i), m, reinterpret_cast<float*>(bBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -5298,20 +5173,12 @@ ElemType CPUMatrix<ElemType>::InnerProductOfMatrices(const CPUMatrix<ElemType>&
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
return (ElemType) ddot((int) a.GetNumElements(), reinterpret_cast<double*>(a.Data()), 1, reinterpret_cast<double*>(b.Data()), 1);
|
||||
#else
|
||||
return (ElemType) cblas_ddot((int) a.GetNumElements(), reinterpret_cast<double*>(a.Data()), 1, reinterpret_cast<double*>(b.Data()), 1);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
return (ElemType) sdot((int) a.GetNumElements(), reinterpret_cast<float*>(a.Data()), 1, reinterpret_cast<float*>(b.Data()), 1);
|
||||
#else
|
||||
return (ElemType) cblas_sdot((int) a.GetNumElements(), reinterpret_cast<float*>(a.Data()), 1, reinterpret_cast<float*>(b.Data()), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -5539,21 +5406,13 @@ void CPUMatrix<ElemType>::InnerProductWithShiftNeg(const CPUMatrix<ElemType>& a,
|
|||
{
|
||||
for (long j = 0; j < n; j++)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = (ElemType) ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = (ElemType) cblas_ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
for (long j = 0; j < n; j++)
|
||||
{
|
||||
for (long i = 1; i < negnumber + 1; i++)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(i, j) = (ElemType) ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn((j + shift + i - 1) % n)), 1);
|
||||
#else
|
||||
c(i, j) = (ElemType) cblas_ddot(m, reinterpret_cast<double*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<double*>(bBufPtr + b.LocateColumn((j + shift + i - 1) % n)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -5561,21 +5420,13 @@ void CPUMatrix<ElemType>::InnerProductWithShiftNeg(const CPUMatrix<ElemType>& a,
|
|||
{
|
||||
for (long j = 0; j < n; j++)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(0, j) = (ElemType) sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#else
|
||||
c(0, j) = (ElemType) cblas_sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn(j)), 1);
|
||||
#endif
|
||||
}
|
||||
for (long j = 0; j < n; j++)
|
||||
{
|
||||
for (long i = 1; i < negnumber + 1; i++)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(i, j) = (ElemType) sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn((j + shift + i - 1) % n)), 1);
|
||||
#else
|
||||
c(i, j) = (ElemType) cblas_sdot(m, reinterpret_cast<float*>(aBufPtr + a.LocateColumn(j)), 1, reinterpret_cast<float*>(bBufPtr + b.LocateColumn((j + shift + i - 1) % n)), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -5593,11 +5444,7 @@ void CPUMatrix<ElemType>::InnerProductWithShiftNeg(const CPUMatrix<ElemType>& a,
|
|||
#pragma omp parallel for
|
||||
foreach_row (i, c)
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = (ElemType) ddot(n, reinterpret_cast<double*>(aBufPtr + i), m, reinterpret_cast<double*>(bBufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = (ElemType) cblas_ddot(n, reinterpret_cast<double*>(aBufPtr + i), m, reinterpret_cast<double*>(bBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
else
|
||||
|
|
@ -5606,9 +5453,6 @@ void CPUMatrix<ElemType>::InnerProductWithShiftNeg(const CPUMatrix<ElemType>& a,
|
|||
foreach_row (i, c)
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
c(i, 0) = sdot(n, reinterpret_cast<float*>(aBufPtr + i), m, reinterpret_cast<float*>(bBufPtr + i), m);
|
||||
#else
|
||||
c(i, 0) = cblas_sdot(n, reinterpret_cast<float*>(aBufPtr + i), m, reinterpret_cast<float*>(bBufPtr + i), m);
|
||||
#endif
|
||||
}
|
||||
|
|
@ -6025,13 +5869,11 @@ int CPUMatrix<ElemType>::SetNumThreads(int numThreads)
|
|||
omp_set_num_threads(numThreads);
|
||||
numThreads = omp_get_max_threads();
|
||||
|
||||
#ifdef USE_ACML
|
||||
acmlsetnumthreads(numThreads);
|
||||
#elif defined(USE_MKL)
|
||||
mkl_set_num_threads(numThreads);
|
||||
#elif defined(USE_OPENBLAS)
|
||||
openblas_set_num_threads(numThreads);
|
||||
#endif
|
||||
#ifdef USE_MKL
|
||||
mkl_set_num_threads(numThreads);
|
||||
#elif defined(USE_OPENBLAS)
|
||||
openblas_set_num_threads(numThreads);
|
||||
#endif
|
||||
#endif
|
||||
return numThreads;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -23,15 +23,7 @@
|
|||
|
||||
#pragma warning(disable : 4127) // conditional expression is constant; "if (sizeof(ElemType)==sizeof(float))" triggers this
|
||||
|
||||
#ifdef USE_ACML
|
||||
// use ACML as default.
|
||||
// Download ACML 5.3.0 (e.g., acml5.3.0-ifort64.exe) or above
|
||||
// from http://developer.amd.com/tools/cpu-development/amd-core-math-library-acml/acml-downloads-resources/
|
||||
// Install the ifort64 variant (compiled with intel compiler) of the library
|
||||
// Set Environment variable ACML_PATH to C:\AMD\acml5.3.0\ifort64_mp or the folder you installed acml
|
||||
// to point to your folder for the include file and link library
|
||||
#include <acml.h> // requires ACML 5.3.0 and above
|
||||
#elif defined(USE_MKL)
|
||||
#ifdef USE_MKL
|
||||
// requires MKL 10.0 and above
|
||||
#include <mkl.h>
|
||||
#else
|
||||
|
|
@ -53,12 +45,6 @@
|
|||
// return 42;
|
||||
//}
|
||||
|
||||
#ifdef USE_ACML // MKL has one additional parameter for different matrix order
|
||||
#define BLAS_COLMAJOR
|
||||
#else
|
||||
#define BLAS_COLMAJOR (int) MatrixOrder::ColMajor,
|
||||
#endif
|
||||
|
||||
// TODO: Move to CommonMatrix.h
|
||||
#define IDX2C(i, j, ld) (((j) * (ld)) + (i)) // 0 based indexing
|
||||
|
||||
|
|
@ -1340,20 +1326,12 @@ ElemType CPUSparseMatrix<ElemType>::SumOfAbsElements() const
|
|||
|
||||
if (sizeof(ElemType) == sizeof(double))
|
||||
{
|
||||
#ifdef USE_ACML
|
||||
return (ElemType) dasum((int) this->NzCount(), reinterpret_cast<double*>(Data()), 1);
|
||||
#else
|
||||
return (ElemType) cblas_dasum((int) this->NzCount(), reinterpret_cast<double*>(Data()), 1);
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
#pragma warning(suppress : 4244)
|
||||
#ifdef USE_ACML
|
||||
return sasum((int) this->NzCount(), reinterpret_cast<float*>(Data()), 1);
|
||||
#else
|
||||
return cblas_sasum((int) this->NzCount(), reinterpret_cast<float*>(Data()), 1);
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -227,6 +227,5 @@
|
|||
<Target Name="CheckDependencies">
|
||||
<Error Condition="'$(MathLibrary)' == 'MKL' And '$(CNTK_MKL_PATH)' == ''" Text="CNTK custom MKL location not specified, see https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#optional-mkl for instructions." />
|
||||
<Error Condition="'$(MathLibrary)' == 'MKL' And !Exists('$(CNTKCustomMKLPath)')" Text="CNTK custom MKL not found. See https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#optional-mkl for instructions." />
|
||||
<Error Condition="'$(MathLibrary)' == 'ACML' And !Exists('$(ACML_PATH)')" Text="ACML not found. See https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#acml for instructions." />
|
||||
</Target>
|
||||
</Project>
|
||||
|
|
|
|||
|
|
@ -17,7 +17,7 @@ RANDOM_OUTPUT=0
|
|||
CODE_COVERAGE=no
|
||||
FLAVORS="debug:release"
|
||||
TARGETS="cpu:gpu"
|
||||
MATH_LIBRARY="acml"
|
||||
MATH_LIBRARY="mkl"
|
||||
TESTTARGETS="cpu:gpu"
|
||||
|
||||
# parsing command line arguments:
|
||||
|
|
@ -76,9 +76,6 @@ case $key in
|
|||
;;
|
||||
-m|--math-library)
|
||||
case ${2,,} in
|
||||
acml)
|
||||
MATH_LIBRARY_OPTION="--with-acml=$ACML_PATH"
|
||||
;;
|
||||
mkl)
|
||||
MATH_LIBRARY_OPTION="--with-mkl=$MKL_PATH"
|
||||
;;
|
||||
|
|
@ -134,12 +131,6 @@ if [[ $OS == "Windows_NT" && $OSTYPE == "cygwin" ]]; then
|
|||
exit 1
|
||||
fi
|
||||
|
||||
if [[ $ACML_PATH == "" ]]; then
|
||||
echo "============ ACML path not set ============"
|
||||
echo "============ ACML libraries are needed to successfully build CNTK ============"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
if [[ "${TARGETS,,}" =~ "1bitsgd" && "${TARGETS,,}" =~ "gpu" ]]; then
|
||||
echo "============ Cannot specify both GPU and 1bit-SGD as targets, please choose one ============"
|
||||
exit 1
|
||||
|
|
|
|||
|
|
@ -11,7 +11,7 @@
|
|||
# BUILDTYPE (release/debug)
|
||||
# BUILDTYPE (GPU/CPU-only)
|
||||
# WITH_1BITSGD (whether 1bit-SGD support was enabled)
|
||||
# MATHLIB (MKL/ACML)
|
||||
# MATHLIB (MKL)
|
||||
# CUDA_PATH (if exists, i.e., for GPU builds)
|
||||
# CUB_PATH (if exists, i.e., for GPU builds)
|
||||
# CUDNN_PATH (if exists, i.e., only for GPU builds)
|
||||
|
|
|
|||
|
|
@ -15,10 +15,6 @@ cuda_path=
|
|||
cuda_check=include/cuda.h
|
||||
enable_cuda=
|
||||
|
||||
have_acml=no
|
||||
acml_path=
|
||||
acml_check=include/acml.h
|
||||
|
||||
# CNTK Custom MKL Version
|
||||
cntk_custom_mkl_version=1
|
||||
|
||||
|
|
@ -79,7 +75,6 @@ enable_code_coverage=$default_use_code_coverage
|
|||
default_path_list="/usr /usr/local /opt /opt/local"
|
||||
|
||||
# List from best to worst choice
|
||||
default_acmls="acml5.3.1/ifort64_mp"
|
||||
default_mkls="CNTKCustomMKL"
|
||||
default_openblas=""
|
||||
|
||||
|
|
@ -131,11 +126,6 @@ function find_dir ()
|
|||
done
|
||||
}
|
||||
|
||||
function find_acml ()
|
||||
{
|
||||
find_dir "$default_acmls" "$acml_check"
|
||||
}
|
||||
|
||||
function find_mkl ()
|
||||
{
|
||||
find_dir "$default_mkls" "$mkl_check"
|
||||
|
|
@ -237,7 +227,6 @@ function show_help ()
|
|||
echo " --with-gdk-include[=directory] $(show_default $(find_gdk_include))"
|
||||
echo " --with-gdk-nvml-lib[=directory] $(show_default $(find_gdk_nvml_lib))"
|
||||
echo " --with-cudnn[=directory] $(show_default $(find_cudnn))"
|
||||
echo " --with-acml[=directory] $(show_default $(find_acml))"
|
||||
echo " --with-mkl[=directory] $(show_default $(find_mkl))"
|
||||
echo " --with-mkl-sequential[=directory] $(show_default $(find_mkl))"
|
||||
echo " --with-openblas[=directory] (experimental) $(show_default $(find_openblas))"
|
||||
|
|
@ -422,28 +411,6 @@ do
|
|||
fi
|
||||
fi
|
||||
;;
|
||||
--with-acml*)
|
||||
have_acml=yes
|
||||
mathlib=acml
|
||||
if test x$optarg = x
|
||||
then
|
||||
acml_path=$(find_acml)
|
||||
if test x$acml_path = x
|
||||
then
|
||||
echo "Cannot find acml directory"
|
||||
echo "Please specify a value for --with-acml"
|
||||
exit 1
|
||||
fi
|
||||
else
|
||||
if test $(check_dir $optarg $acml_check) = yes
|
||||
then
|
||||
acml_path=$optarg
|
||||
else
|
||||
echo "Invalid acml directory $optarg"
|
||||
exit 1
|
||||
fi
|
||||
fi
|
||||
;;
|
||||
--with-mkl*)
|
||||
have_mkl=yes
|
||||
mathlib=mkl
|
||||
|
|
@ -603,24 +570,18 @@ then
|
|||
echo Defaulting to --with-buildtype=release
|
||||
fi
|
||||
|
||||
# If no math library was specified, search for acml and then mkl
|
||||
if test x$have_acml = xno && test x$have_mkl = xno && test x$have_openblas = xno
|
||||
# If no math library was specified, search for mkl
|
||||
if test x$have_mkl = xno && test x$have_openblas = xno
|
||||
then
|
||||
acml_path=$(find_acml)
|
||||
if test x$acml_path = x
|
||||
then
|
||||
mkl_path=$(find_mkl)
|
||||
if test x$mkl_path = x
|
||||
then
|
||||
echo "Cannot find a CPU math library."
|
||||
echo "Please specify --with-acml, --with-mkl, --with-mkl-sequential, --with-openblas with a path."
|
||||
exit 1
|
||||
else
|
||||
mathlib=mkl
|
||||
fi
|
||||
else
|
||||
mathlib=acml
|
||||
fi
|
||||
mkl_path=$(find_mkl)
|
||||
if test x$mkl_path = x
|
||||
then
|
||||
echo "Cannot find a CPU math library."
|
||||
echo "Please specify --with-mkl, --with-mkl-sequential, --with-openblas with a path."
|
||||
exit 1
|
||||
else
|
||||
mathlib=mkl
|
||||
fi
|
||||
fi
|
||||
|
||||
# If no cuda library specified, search for one
|
||||
|
|
@ -735,9 +696,6 @@ echo "#Configuration file for cntk" > $config
|
|||
echo BUILDTYPE=$buildtype >> $config
|
||||
echo MATHLIB=$mathlib >> $config
|
||||
case $mathlib in
|
||||
acml)
|
||||
echo ACML_PATH=$acml_path >> $config
|
||||
;;
|
||||
mkl)
|
||||
echo MKL_PATH=$mkl_path >> $config
|
||||
echo MKL_THREADING=$mkl_threading >> $config
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче