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

Fix 4x4 inverse via SSE for submatrices

Browse Source
This commit is contained in:
Gael Guennebaud
2014-07-31 16:24:29 +02:00
parent db183ca7b3
commit 26d2cdefd4
3 changed files with 19 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
Eigen/src/LU/arch/Inverse_SSE.h
View File

@@ -39,9 +39,11 @@ struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType>
ResultAlignment = bool(ResultType::Flags&AlignedBit),
StorageOrdersMatch = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
};
typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
static void run(const MatrixType& matrix, ResultType& result)
static void run(const MatrixType& mat, ResultType& result)
{
ActualMatrixType matrix(mat);
EIGEN_ALIGN16 const unsigned int _Sign_PNNP[4] = { 0x00000000, 0x80000000, 0x80000000, 0x00000000 };
// Load the full matrix into registers
@@ -167,14 +169,17 @@ struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType>
ResultAlignment = bool(ResultType::Flags&AlignedBit),
StorageOrdersMatch = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
};
static void run(const MatrixType& matrix, ResultType& result)
typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
static void run(const MatrixType& mat, ResultType& result)
{
ActualMatrixType matrix(mat);
const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
// The inverse is calculated using "Divide and Conquer" technique. The
// original matrix is divide into four 2x2 sub-matrices. Since each
// register of the matrix holds two element, the smaller matrices are
// register of the matrix holds two elements, the smaller matrices are
// consisted of two registers. Hence we get a better locality of the
// calculations.