potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 07:13:02 +04:00
Code Packages Releases Wiki Activity

Add QR decomposition to HAL

Browse Source
This commit is contained in:
Vladislav Sovrasov
2016-08-23 16:35:03 +03:00
parent d102ea96c0
commit dfe4519c07
10 changed files with 348 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/hal_internal.cpp
+116
View File
@@ -61,10 +61,12 @@
#include <typeinfo>
#include <limits>
#include <complex>
#include <vector>
#define HAL_GEMM_SMALL_COMPLEX_MATRIX_THRESH 100
#define HAL_GEMM_SMALL_MATRIX_THRESH 100
#define HAL_SVD_SMALL_MATRIX_THRESH 25
#define HAL_QR_SMALL_MATRIX_THRESH 30
#define HAL_LU_SMALL_MATRIX_THRESH 100
#define HAL_CHOLESKY_SMALL_MATRIX_THRESH 100
@@ -259,6 +261,106 @@ lapack_SVD(fptype* a, size_t a_step, fptype *w, fptype* u, size_t u_step, fptype
return CV_HAL_ERROR_OK;
}
template <typename fptype> static inline int
lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_step, fptype* dst, int* info)
{
int lda = a_step / sizeof(fptype);
char mode[] = { 'N', '\0' };
if(m < n)
return CV_HAL_ERROR_NOT_IMPLEMENTED;
std::vector<fptype> tmpAMemHolder;
fptype* tmpA;
int ldtmpA;
if (m == n)
{
transpose_square_inplace(a, lda, m);
tmpA = a;
ldtmpA = lda;
}
else
{
tmpAMemHolder.resize(m*n);
tmpA = &tmpAMemHolder.front();
ldtmpA = m;
transpose(a, lda, tmpA, m, m, n);
}
int lwork = -1;
fptype work1 = 0.;
if (b)
{
if (k == 1 && b_step == sizeof(fptype))
{
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)&work1, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)&work1, &lwork, info);
lwork = (int)round(work1); //optimal buffer size
std::vector<fptype> workBufMemHolder(lwork + 1);
fptype* buffer = &workBufMemHolder.front();
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)buffer, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)buffer, &lwork, info);
}
else
{
std::vector<fptype> tmpBMemHolder(m*k);
fptype* tmpB = &tmpBMemHolder.front();
int ldb = b_step / sizeof(fptype);
transpose(b, ldb, tmpB, m, m, k);
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)&work1, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)&work1, &lwork, info);
lwork = (int)round(work1); //optimal buffer size
std::vector<fptype> workBufMemHolder(lwork + 1);
fptype* buffer = &workBufMemHolder.front();
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)buffer, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)buffer, &lwork, info);
transpose(tmpB, m, b, ldb, k, m);
}
}
else
{
if (typeid(fptype) == typeid(float))
sgeqrf_(&m, &n, (float*)tmpA, &ldtmpA, (float*)dst, (float*)&work1, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgeqrf_(&m, &n, (double*)tmpA, &ldtmpA, (double*)dst, (double*)&work1, &lwork, info);
lwork = (int)round(work1); //optimal buffer size
std::vector<fptype> workBufMemHolder(lwork + 1);
fptype* buffer = &workBufMemHolder.front();
if (typeid(fptype) == typeid(float))
sgeqrf_(&m, &n, (float*)tmpA, &ldtmpA, (float*)dst, (float*)buffer, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgeqrf_(&m, &n, (double*)tmpA, &ldtmpA, (double*)dst, (double*)buffer, &lwork, info);
}
if (m == n)
transpose_square_inplace(a, lda, m);
else
transpose(tmpA, m, a, lda, n, m);
if (*info != 0)
*info = 0;
else
*info = 1;
return CV_HAL_ERROR_OK;
}
template <typename fptype> static inline int
lapack_gemm(const fptype *src1, size_t src1_step, const fptype *src2, size_t src2_step, fptype alpha,
const fptype *src3, size_t src3_step, fptype beta, fptype *dst, size_t dst_step, int a_m, int a_n, int d_n, int flags)
@@ -459,6 +561,20 @@ int lapack_SVD64f(double* a, size_t a_step, double *w, double* u, size_t u_step,
return lapack_SVD(a, a_step, w, u, u_step, vt, v_step, m, n, flags, &info);
}
int lapack_QR32f(float* src1, size_t src1_step, int m, int n, int k, float* src2, size_t src2_step, float* dst, int* info)
{
if (m < HAL_QR_SMALL_MATRIX_THRESH)
return CV_HAL_ERROR_NOT_IMPLEMENTED;
return lapack_QR(src1, src1_step, m, n, k, src2, src2_step, dst, info);
}
int lapack_QR64f(double* src1, size_t src1_step, int m, int n, int k, double* src2, size_t src2_step, double* dst, int* info)
{
if (m < HAL_QR_SMALL_MATRIX_THRESH)
return CV_HAL_ERROR_NOT_IMPLEMENTED;
return lapack_QR(src1, src1_step, m, n, k, src2, src2_step, dst, info);
}
int lapack_gemm32f(const float *src1, size_t src1_step, const float *src2, size_t src2_step, float alpha,
const float *src3, size_t src3_step, float beta, float *dst, size_t dst_step, int m, int n, int k, int flags)
{