// copy construct into state 1, always.
// This is a choice, even if X is state 2 (a numpy).
// We copy a numpy into a regular C++ array, which can then be used at max speed.
mem_block (mem_block const & X): size_(X.size()), py_numpy(nullptr), py_guard(nullptr) {
try { p = new ValueType[X.size()];}
catch (std::bad_alloc& ba) { TRIQS_RUNTIME_ERROR<< "Memory allocation error in memblock copy construction. Size :"<<X.size() << " bad_alloc error : "<< ba.what();}
TRACE_MEM_DEBUG("Allocating from C++ a block of size "<< X.size() << " at address " <<p);
TRIQS_MEMORY_USED_INC(X.size());
ref_count=1;
weak_ref_count =0;
// now we copy the data
#ifndef TRIQS_WITH_PYTHON_SUPPORT
copy_from(X);
#else
// if X is in state 1 or 3
if (X.py_numpy==nullptr) { copy_from(X); }
else { // X was in state 2
// else make a new copy of the numpy ...
import_numpy_array();
if (!is_scalar_or_pod<ValueType>::value) TRIQS_RUNTIME_ERROR << "Internal Error : memcpy on non-scalar";
#ifdef TRIQS_NUMPY_VERSION_LT_17
PyObject * arr3 = X.py_numpy;
#else
// STRANGE : uncommenting this leads to a segfault on mac ???
// TO BE INVESTIGATED, IT IS NOT NORMAL
//if (!PyArray_Check(X.py_numpy)) TRIQS_RUNTIME_ERROR<<"Internal error : is not an array";
PyArrayObject * arr3 = (PyArrayObject *)(X.py_numpy);
#endif
// if we can make a memcpy, do it.
if ( ( PyArray_ISFORTRAN(arr3)) || (PyArray_ISCONTIGUOUS(arr3))) {
memcpy (p,PyArray_DATA(arr3),size_ * sizeof(ValueType));
}
else { // if the X.py_numpy is not contiguous, first let numpy copy it properly, then memcpy
PyObject * na = PyObject_CallMethod(X.py_numpy,(char *)"copy",nullptr);
assert(na);
#ifdef TRIQS_NUMPY_VERSION_LT_17
PyObject * arr = na;
#else
if (!PyArray_Check(na)) TRIQS_RUNTIME_ERROR<<"Internal error : is not an array";
PyArrayObject * arr = (PyArrayObject *)(na);
#endif
assert( ( PyArray_ISFORTRAN(arr)) || (PyArray_ISCONTIGUOUS(arr)));
memcpy (p,PyArray_DATA(arr),size_ * sizeof(ValueType));
Py_DECREF(na);
}
}
#endif
}
static mxArray *makeMxFromNumeric(const PyArrayObject *pSrc)
{
npy_intp lRows=0, lCols=0;
bool lIsComplex;
bool lIsNotAMatrix = false;
double *lR = NULL;
double *lI = NULL;
mxArray *lRetval = NULL;
mwSize dims[NPY_MAXDIMS];
mwSize nDims = pSrc->nd;
const PyArrayObject *ap=NULL;
switch (pSrc->nd) {
case 0: // XXX the evil 0D
lRows = 1;
lCols = 1;
lIsNotAMatrix = true;
break;
case 1:
lRows = pSrc->dimensions[0];
lCols = min(1, lRows); // for array([]): to avoid zeros((0,1)) !
lIsNotAMatrix = true;
break;
default:
for (mwSize i = 0;i != nDims; i++) {
dims[i]=(mwSize)pSrc->dimensions[i];
}
break;
}
switch (pSrc->descr->type_num) {
case PyArray_OBJECT:
PyErr_SetString(PyExc_TypeError, "Non-numeric array types not supported");
return NULL;
case PyArray_CFLOAT:
case PyArray_CDOUBLE:
lIsComplex = true;
break;
default:
lIsComplex = false;
}
// converts to fortran order if not already
if(!PyArray_ISFORTRAN(pSrc)){
ap = (PyArrayObject * const)PyArray_FromArray((PyArrayObject*)pSrc,NULL,NPY_ALIGNED|NPY_F_CONTIGUOUS);
}
else{
ap = pSrc;
}
if(lIsNotAMatrix)
lRetval = mxCreateDoubleMatrix(lRows, lCols, lIsComplex ? mxCOMPLEX : mxREAL);
else
lRetval = mxCreateNumericArray(nDims,dims,mxDOUBLE_CLASS,lIsComplex ? mxCOMPLEX : mxREAL);
if (lRetval == NULL) return NULL;
lR = mxGetPr(lRetval);
lI = mxGetPi(lRetval);
if (lIsNotAMatrix) {
void *p = PyArray_DATA(ap);
switch (ap->descr->type_num) {
case PyArray_CHAR:
copyNumericVector2Mx((char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_UBYTE:
copyNumericVector2Mx((unsigned char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_SBYTE:
copyNumericVector2Mx((signed char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_SHORT:
copyNumericVector2Mx((short *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_INT:
copyNumericVector2Mx((int *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_LONG:
copyNumericVector2Mx((long *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_FLOAT:
copyNumericVector2Mx((float *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_DOUBLE:
copyNumericVector2Mx((double *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_CFLOAT:
copyCplxNumericVector2Mx((float *)(p), lRows, lR, lI, pSrc->strides);
break;
case PyArray_CDOUBLE:
copyCplxNumericVector2Mx((double *)(p), lRows, lR, lI, pSrc->strides);
break;
}
} else {
void *p = PyArray_DATA(ap);
npy_intp size = PyArray_SIZE(pSrc);
switch (pSrc->descr->type_num) {
case PyArray_CHAR:
copyNumeric2Mx((char *)p,size,lR);
break;
case PyArray_UBYTE:
copyNumeric2Mx((unsigned char *)p,size,lR);
break;
case PyArray_SBYTE:
copyNumeric2Mx((signed char *)p,size,lR);
break;
case PyArray_SHORT:
copyNumeric2Mx((short *)p,size,lR);
break;
case PyArray_INT:
copyNumeric2Mx((int *)p,size,lR);
break;
case PyArray_LONG:
copyNumeric2Mx((long *)p,size,lR);
break;
case PyArray_FLOAT:
copyNumeric2Mx((float *)p,size,lR);
break;
case PyArray_DOUBLE:
copyNumeric2Mx((double *)p,size,lR);
break;
case PyArray_CFLOAT:
copyCplxNumeric2Mx((float *)p,size,lR,lI);
break;
case PyArray_CDOUBLE:
copyCplxNumeric2Mx((double *)p,size,lR,lI);
break;
}
}
if(ap != pSrc){
Py_DECREF(const_cast<PyArrayObject *>(ap));
}
return lRetval;
}
numpy_extractor ( PyObject * X, bool allow_copy) {
if (X==NULL) TRIQS_RUNTIME_ERROR<<"numpy interface : the python object is NULL !";
if (_import_array()!=0) TRIQS_RUNTIME_ERROR <<"Internal Error in importing numpy";
// make sure IndexMap is cuboid ?s
static const char Order = IndexMapType::index_order_type::C_or_F;
static_assert( ((Order=='F')||(Order=='C')), "Ordering must be C or Fortran");
if ((!PyArray_Check(X)) && (Order=='D'))
TRIQS_RUNTIME_ERROR<<"numpy interface : the python object is not a numpy and you ask me to deduce the ordering in memory !";
const int elementsType (numpy_to_C_type<typename boost::remove_const<ValueType>::type>::arraytype);
int rank = IndexMapType::rank;
static const char * error_msg = " A deep copy of the object would be necessary while views are supposed to guarantee to present a *view* of the python data.\n";
if (!allow_copy) {
// in case of a view, we decide ourselves if we can do it...
// a previous uses PyArray_FromAny, but behaviour changes between different version of numpy, so it is not portable...
if (!PyArray_Check(X))
throw copy_exception () << error_msg<<" Indeed the object was not even an array !\n";
if ( elementsType != PyArray_TYPE((PyArrayObject*)X))
throw copy_exception () << error_msg<<" The deep copy is caused by a type mismatch of the elements. \n";
PyArrayObject *arr = (PyArrayObject *)X;
if ( arr->nd != rank)
throw copy_exception () << error_msg<<" Rank mismatch . numpy array is of rank "<< arr->nd << "while you ask for rank "<< rank<<". \n";
if ((Order == 'C') && (PyArray_ISFORTRAN(arr)))
throw copy_exception () << error_msg<<" The numpy is in Fortran order while it is expected in C order. \n";
if ((Order == 'F') && (!PyArray_ISFORTRAN(arr)))
throw copy_exception () << error_msg<<" The numpy is not in Fortran order as it is expected. \n";
numpy_obj = X; Py_INCREF(X);
}
else {
// From X, we ask the numpy library to make a numpy, and of the correct type.
// This handles automatically the cases where :
// - we have list, or list of list/tuple
// - the numpy type is not the one we want.
// - adjust the dimension if needed
// If X is an array :
// - if Order is same, don't change it
// - else impose it (may provoque a copy).
// if X is not array :
// - Order = FortranOrder or SameOrder - > Fortran order otherwise C
bool ForceCast = false;// Unless FORCECAST is present in flags, this call will generate an error if the data type cannot be safely obtained from the object.
int flags = (ForceCast ? NPY_FORCECAST : 0) ;// do NOT force a copy | (make_copy ? NPY_ENSURECOPY : 0);
if (!(PyArray_Check(X) && (Order=='D'))) flags |= (Order =='F' ? NPY_F_CONTIGUOUS : NPY_C_CONTIGUOUS); //impose mem order
//if (!(PyArray_Check(X) && (Order=='D'))) flags |= (Order =='F' ? NPY_FARRAY : NPY_CARRAY); //impose mem order
numpy_obj= PyArray_FromAny(X,PyArray_DescrFromType(elementsType), rank,rank, flags , NULL );
// do several checks
if (!numpy_obj) {// The convertion of X to a numpy has failed !
if (PyErr_Occurred()) {PyErr_Print();PyErr_Clear();}
TRIQS_RUNTIME_ERROR<<"numpy interface : the python object is not convertible to a numpy. ";
}
assert (PyArray_Check(numpy_obj)); assert((numpy_obj->ob_refcnt==1) || ((numpy_obj ==X)));
arr_obj = (PyArrayObject *)numpy_obj;
try {
if (arr_obj->nd!=rank) TRIQS_RUNTIME_ERROR<<"numpy interface : internal error : dimensions do not match";
if (arr_obj->descr->type_num != elementsType)
TRIQS_RUNTIME_ERROR<<"numpy interface : internal error : incorrect type of element :" <<arr_obj->descr->type_num <<" vs "<<elementsType;
if (Order == 'F') { if (!PyArray_ISFORTRAN(numpy_obj)) TRIQS_RUNTIME_ERROR<<"numpy interface : internal error : should be Fortran array";}
else {if (!PyArray_ISCONTIGUOUS(numpy_obj)) TRIQS_RUNTIME_ERROR<<"numpy interface : internal error : should be contiguous";}
}
catch(...) { Py_DECREF(numpy_obj); throw;} // make sure that in case of problem, the reference counting of python is still ok...
}
arr_obj = (PyArrayObject *)numpy_obj;
}
static PyObject *product_matrix_vector_(PyObject *self, PyObject *args)
{
PyArrayObject *matrix, *vector, *result;
double alpha = 1.;
double beta = 0.;
int int_one = 1;
int m, n;
char trans[2] = "T";
extern void dgemv_(char *trans, int *m, int *n,
double *alpha, double *a, int *lda,
double *x, int *incx,
double *beta, double *Y, int *incy);
if (!PyArg_ParseTuple(args, "O!O!O!",
&PyArray_Type, &matrix,
&PyArray_Type, &vector,
&PyArray_Type, &result)) return NULL;
if ( (NULL == matrix) || (NULL == vector) || (NULL == result) ) return NULL;
if ( (matrix->descr->type_num != NPY_DOUBLE) ||
(vector->descr->type_num != NPY_DOUBLE) ||
(result->descr->type_num != NPY_DOUBLE) ||
!PyArray_CHKFLAGS(matrix,NPY_ALIGNED) ||
!(PyArray_CHKFLAGS(matrix,NPY_C_CONTIGUOUS) || PyArray_CHKFLAGS(matrix,NPY_F_CONTIGUOUS)) ||
!PyArray_CHKFLAGS(vector,NPY_C_CONTIGUOUS|NPY_ALIGNED) ||
!PyArray_CHKFLAGS(result,NPY_C_CONTIGUOUS|NPY_ALIGNED|NPY_WRITEABLE) ) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_vector: all arguments must be of type double, contiguous and aligned, and targets should be writeable");
return NULL;
}
if ( (matrix->nd != 2) || (vector->nd != 1) || (result->nd != 1) )
{
PyErr_SetString(PyExc_ValueError,
"In product_matrix_vector: not all arguments have the right dimensionality");
return NULL;
}
/* Figure out correct values for args to blas*/
if (PyArray_ISFORTRAN(matrix)) {
m = matrix->dimensions[0];
n = matrix->dimensions[1];
trans[0] = 'N';
}
else /*if (PyArray_ISCONTIGUOUS(matrix))*/ {
/* Equivalent to matrix being the transposed of a Fortran matrix*/
m = matrix->dimensions[1];
n = matrix->dimensions[0];
}
/* Check if dimensions are compatible */
if (matrix->dimensions[1] != vector->dimensions[0]) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_vector: input dimensions are not compatible");
return NULL;
}
if (result->dimensions[0] != matrix->dimensions[0]) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_vector: target dimension is not compatible");
return NULL;
}
dgemv_(trans, &m, &n, &alpha, (double *)matrix->data, &m,
(double *)vector->data, &int_one,
&beta, (double *)result->data, &int_one);
Py_RETURN_NONE;
}
static mxArray *makeMxFromNumeric(const PyArrayObject *pSrc, bool &is_reference)
{
npy_intp lRows=0, lCols=0;
bool lIsComplex;
bool lIsNotAMatrix = false;
double *lR = NULL;
double *lI = NULL;
mxArray *lRetval = NULL;
mwSize dims[NPY_MAXDIMS];
mwSize dimsEmpty [2];
memset(&dimsEmpty, 0, 2 * sizeof(mwSize));
mwSize nDims = pSrc->nd;
const PyArrayObject *ap=NULL;
mxClassID classID = mxUNKNOWN_CLASS;
switch (pSrc->nd) {
case 0: // XXX the evil 0D
lRows = 1;
lCols = 1;
lIsNotAMatrix = true;
break;
case 1:
lRows = pSrc->dimensions[0];
lCols = min(1, lRows); // for array([]): to avoid zeros((0,1)) !
lIsNotAMatrix = true;
break;
default:
for (mwSize i = 0;i != nDims; i++) {
dims[i]=(mwSize)pSrc->dimensions[i];
}
break;
}
switch (pSrc->descr->type_num) {
case PyArray_OBJECT:
PyErr_SetString(PyExc_TypeError, "Non-numeric array types not supported");
return NULL;
case PyArray_CFLOAT:
case PyArray_CDOUBLE:
lIsComplex = true;
break;
default:
lIsComplex = false;
}
// converts to fortran order if not already
if(!PyArray_ISFORTRAN(pSrc)){
ap = (PyArrayObject * const)PyArray_FromArray((PyArrayObject*)pSrc,NULL,NPY_ALIGNED|NPY_F_CONTIGUOUS);
}
else{
ap = pSrc;
}
if(lIsNotAMatrix)
lRetval = mxCreateDoubleMatrix(lRows, lCols, lIsComplex ? mxCOMPLEX : mxREAL);
else {
switch (ap->descr->type_num) {
case PyArray_CFLOAT:
case PyArray_CDOUBLE:
classID = mxDOUBLE_CLASS;
is_reference = false;
break;
case PyArray_BOOL:
classID = mxLOGICAL_CLASS;
is_reference = true;
break;
case PyArray_CHAR:
classID = mxCHAR_CLASS;
is_reference = true;
break;
case PyArray_DOUBLE:
classID = mxDOUBLE_CLASS;
is_reference = true;
break;
case PyArray_FLOAT:
classID = mxSINGLE_CLASS;
is_reference = true;
break;
case PyArray_INT8:
classID = mxINT8_CLASS;
is_reference = true;
break;
case PyArray_UINT8:
classID = mxUINT8_CLASS;
is_reference = true;
break;
case PyArray_INT16:
classID = mxINT16_CLASS;
is_reference = true;
break;
case PyArray_UINT16:
classID = mxUINT16_CLASS;
is_reference = true;
break;
case PyArray_INT32:
classID = mxINT32_CLASS;
is_reference = true;
break;
case PyArray_UINT32:
classID = mxUINT32_CLASS;
is_reference = true;
break;
#ifdef NPY_INT64
case PyArray_INT64:
classID = mxINT64_CLASS;
is_reference = true;
break;
case PyArray_UINT64:
classID = mxUINT64_CLASS;
is_reference = true;
break;
#endif
}
if (!is_reference) {
lRetval = mxCreateNumericArray(nDims,dims,classID,lIsComplex ? mxCOMPLEX : mxREAL);
} else {
/*
// code for create and mxArray ref on the numpy object
lRetval = mxCreateNumericArray(2, dimsEmpty, classID, lIsComplex ? mxCOMPLEX : mxREAL);
if (mxSetDimensions(lRetval, dims, nDims) == 1) {
// failed to reset the dimensions
mxDestroyArray(lRetval);
lRetval = NULL;
} else {
mxSetData(lRetval, PyArray_DATA(ap));
}
*/
// uses memcopy for faster copying
is_reference = false;
lRetval = mxCreateNumericArray(nDims, dims, classID, lIsComplex ? mxCOMPLEX : mxREAL);
// this is a sanity check, it should not fail
int matlab_nbytes = mxGetNumberOfElements(lRetval) * mxGetElementSize(lRetval);
if (matlab_nbytes == PyArray_NBYTES(pSrc)) {
memcpy(mxGetData(lRetval), PyArray_DATA(ap), PyArray_NBYTES(pSrc));
} else {
PyErr_SetString(PyExc_TypeError, "Number of bytes do not match");
}
}
}
if (lRetval == NULL) return NULL;
lR = mxGetPr(lRetval);
lI = mxGetPi(lRetval);
if (lIsNotAMatrix) {
void *p = PyArray_DATA(ap);
switch (ap->descr->type_num) {
case PyArray_CHAR:
copyNumericVector2Mx((char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_UBYTE:
copyNumericVector2Mx((unsigned char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_SBYTE:
copyNumericVector2Mx((signed char *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_SHORT:
copyNumericVector2Mx((short *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_INT:
copyNumericVector2Mx((int *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_LONG:
copyNumericVector2Mx((long *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_FLOAT:
copyNumericVector2Mx((float *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_DOUBLE:
copyNumericVector2Mx((double *)(p), lRows, lR, pSrc->strides);
break;
case PyArray_CFLOAT:
copyCplxNumericVector2Mx((float *)(p), lRows, lR, lI, pSrc->strides);
break;
case PyArray_CDOUBLE:
copyCplxNumericVector2Mx((double *)(p), lRows, lR, lI, pSrc->strides);
break;
}
} else {
void *p = PyArray_DATA(ap);
npy_intp size = PyArray_SIZE(pSrc);
switch (pSrc->descr->type_num) {
case PyArray_CFLOAT:
copyCplxNumeric2Mx((float *)p,size,lR,lI);
break;
case PyArray_CDOUBLE:
copyCplxNumeric2Mx((double *)p,size,lR,lI);
break;
}
}
if(ap != pSrc){
Py_DECREF(const_cast<PyArrayObject *>(ap));
}
return lRetval;
}
static PyObject *product_matrix_matrix_(PyObject *self, PyObject *args)
{
PyArrayObject *matrix1, *matrix2, *result, *A, *B;
double alpha = 1.;
double beta = 0.;
int m, n, k; /* Watch out: m and n have a different meaning from that for dgemv, where
it is the dimensionality BEFORE the transpose, whereas
dgemm requires the dimensionality AFTER. */
int lda, ldb, ldc;
char transa[2] = "T";
char transb[2] = "T";
extern void dgemm_(char* transa, char* transb, int* m,
int* n, int* k, double *alpha,
double *a, int* lda, double *b,
int *ldb, double *beta, double *c, int* ldc);
if (!PyArg_ParseTuple(args, "O!O!O!",
&PyArray_Type, &matrix1,
&PyArray_Type, &matrix2,
&PyArray_Type, &result)) return NULL;
if ( (NULL == matrix1) || (NULL == matrix2) || (NULL == result) ) return NULL;
if ( (matrix1->descr->type_num != NPY_DOUBLE) ||
(matrix2->descr->type_num != NPY_DOUBLE) ||
(result->descr->type_num != NPY_DOUBLE) ||
!PyArray_CHKFLAGS(matrix1,NPY_ALIGNED) ||
!(PyArray_CHKFLAGS(matrix1,NPY_C_CONTIGUOUS) || PyArray_CHKFLAGS(matrix1,NPY_F_CONTIGUOUS)) ||
!PyArray_CHKFLAGS(matrix2,NPY_ALIGNED) ||
!(PyArray_CHKFLAGS(matrix2,NPY_C_CONTIGUOUS) || PyArray_CHKFLAGS(matrix2,NPY_F_CONTIGUOUS)) ||
!PyArray_CHKFLAGS(result,NPY_ALIGNED|NPY_WRITEABLE) ||
!(PyArray_CHKFLAGS(result,NPY_C_CONTIGUOUS) || PyArray_CHKFLAGS(result,NPY_F_CONTIGUOUS)) ) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_matrix: all arguments must be of type double, contiguous and aligned, and targets should be writeable");
return NULL;
}
if ( (matrix1->nd != 2) || (matrix2->nd != 2) || (result->nd != 2) )
{
PyErr_SetString(PyExc_ValueError,
"In product_matrix_matrix: not all arguments have the right dimensionality");
return NULL;
}
/* Figure out correct values for args to blas*/
if (PyArray_ISFORTRAN(result)) {
/* Matrix result has Fortran order */
A = matrix1;
B = matrix2;
m = matrix1->dimensions[0];
k = matrix1->dimensions[1];
if (PyArray_ISFORTRAN(matrix1)) {
transa[0] = 'N';
lda = m;
}
else /*if (PyArray_ISCONTIGUOUS(matrix1))*/ {
/* Equivalent to matrix being the transposed of a Fortran matrix*/
lda = k;
}
n = matrix2->dimensions[1];
k = matrix2->dimensions[0];
if (PyArray_ISFORTRAN(matrix2)) {
transb[0] = 'N';
ldb = k;
}
else /*if (PyArray_ISCONTIGUOUS(matrix2))*/ {
/* Equivalent to matrix being the transposed of a Fortran matrix*/
ldb = n;
}
}
else /*if (PyArray_ISCONTIGUOUS(result))*/ {
/* Matrix result has C order! Must be careful how blas is called!
Must interchange matrix1 with matrix2, and interchange
how the C and Fortran matrices are processed. */
A = matrix2;
B = matrix1;
m = matrix2->dimensions[1];
k = matrix2->dimensions[0];
if (PyArray_ISFORTRAN(matrix2)) {
lda = k;
}
else /*if (PyArray_ISCONTIGUOUS(matrix2))*/ {
/* Equivalent to matrix being the transposed of a Fortran matrix*/
transa[0] = 'N';
lda = m;
}
n = matrix1->dimensions[0];
k = matrix1->dimensions[1];
if (PyArray_ISFORTRAN(matrix1)) {
ldb = n;
}
else /*if (PyArray_ISCONTIGUOUS(matrix1))*/ {
/* Equivalent to matrix being the transposed of a Fortran matrix*/
transb[0] = 'N';
ldb = k;
}
}
ldc = m;
if ( matrix1->dimensions[1] != matrix2->dimensions[0] ) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_matrix: input matrices dimensions are not compatible");
return NULL;
}
if ( (matrix1->dimensions[0] != result->dimensions[0]) ||
(matrix2->dimensions[1] != result->dimensions[1]) ) {
PyErr_SetString(PyExc_ValueError,
"In product_matrix_matrix: target dimensions are not compatible");
return NULL;
}
dgemm_(transa, transb, &m, &n, &k, &alpha, (double *)A->data, &lda,
(double *)B->data, &ldb, &beta,
(double *)result->data, &ldc );
Py_RETURN_NONE;
}
/*
* optimize float array or complex array to a scalar power
* returns 0 on success, -1 if no optimization is possible
* the result is in value (can be NULL if an error occurred)
*/
static int
fast_scalar_power(PyArrayObject *a1, PyObject *o2, int inplace,
PyObject **value)
{
double exponent;
NPY_SCALARKIND kind; /* NPY_NOSCALAR is not scalar */
if (PyArray_Check(a1) &&
!PyArray_ISOBJECT(a1) &&
((kind=is_scalar_with_conversion(o2, &exponent))>0)) {
PyObject *fastop = NULL;
if (PyArray_ISFLOAT(a1) || PyArray_ISCOMPLEX(a1)) {
if (exponent == 1.0) {
fastop = n_ops.positive;
}
else if (exponent == -1.0) {
fastop = n_ops.reciprocal;
}
else if (exponent == 0.0) {
fastop = n_ops._ones_like;
}
else if (exponent == 0.5) {
fastop = n_ops.sqrt;
}
else if (exponent == 2.0) {
fastop = n_ops.square;
}
else {
return -1;
}
if (inplace || can_elide_temp_unary(a1)) {
*value = PyArray_GenericInplaceUnaryFunction(a1, fastop);
}
else {
*value = PyArray_GenericUnaryFunction(a1, fastop);
}
return 0;
}
/* Because this is called with all arrays, we need to
* change the output if the kind of the scalar is different
* than that of the input and inplace is not on ---
* (thus, the input should be up-cast)
*/
else if (exponent == 2.0) {
fastop = n_ops.square;
if (inplace) {
*value = PyArray_GenericInplaceUnaryFunction(a1, fastop);
}
else {
/* We only special-case the FLOAT_SCALAR and integer types */
if (kind == NPY_FLOAT_SCALAR && PyArray_ISINTEGER(a1)) {
PyArray_Descr *dtype = PyArray_DescrFromType(NPY_DOUBLE);
a1 = (PyArrayObject *)PyArray_CastToType(a1, dtype,
PyArray_ISFORTRAN(a1));
if (a1 != NULL) {
/* cast always creates a new array */
*value = PyArray_GenericInplaceUnaryFunction(a1, fastop);
Py_DECREF(a1);
}
}
else {
*value = PyArray_GenericUnaryFunction(a1, fastop);
}
}
return 0;
}
}
/* no fast operation found */
return -1;
}
/* optimize float array or complex array to a scalar power */
static PyObject *
fast_scalar_power(PyArrayObject *a1, PyObject *o2, int inplace)
{
double exponent;
NPY_SCALARKIND kind; /* NPY_NOSCALAR is not scalar */
if (PyArray_Check(a1) && ((kind=is_scalar_with_conversion(o2, &exponent))>0)) {
PyObject *fastop = NULL;
if (PyArray_ISFLOAT(a1) || PyArray_ISCOMPLEX(a1)) {
if (exponent == 1.0) {
/* we have to do this one special, as the
"copy" method of array objects isn't set
up early enough to be added
by PyArray_SetNumericOps.
*/
if (inplace) {
Py_INCREF(a1);
return (PyObject *)a1;
} else {
return PyArray_Copy(a1);
}
}
else if (exponent == -1.0) {
fastop = n_ops.reciprocal;
}
else if (exponent == 0.0) {
fastop = n_ops._ones_like;
}
else if (exponent == 0.5) {
fastop = n_ops.sqrt;
}
else if (exponent == 2.0) {
fastop = n_ops.square;
}
else {
return NULL;
}
if (inplace) {
return PyArray_GenericInplaceUnaryFunction(a1, fastop);
} else {
return PyArray_GenericUnaryFunction(a1, fastop);
}
}
/* Because this is called with all arrays, we need to
* change the output if the kind of the scalar is different
* than that of the input and inplace is not on ---
* (thus, the input should be up-cast)
*/
else if (exponent == 2.0) {
fastop = n_ops.multiply;
if (inplace) {
return PyArray_GenericInplaceBinaryFunction
(a1, (PyObject *)a1, fastop);
}
else {
PyArray_Descr *dtype = NULL;
PyObject *res;
/* We only special-case the FLOAT_SCALAR and integer types */
if (kind == NPY_FLOAT_SCALAR && PyArray_ISINTEGER(a1)) {
dtype = PyArray_DescrFromType(NPY_DOUBLE);
a1 = (PyArrayObject *)PyArray_CastToType(a1, dtype,
PyArray_ISFORTRAN(a1));
if (a1 == NULL) {
return NULL;
}
}
else {
Py_INCREF(a1);
}
res = PyArray_GenericBinaryFunction(a1, (PyObject *)a1, fastop);
Py_DECREF(a1);
return res;
}
}
}
return NULL;
}
/*NUMPY_API
* Clip
*/
NPY_NO_EXPORT PyObject *
PyArray_Clip(PyArrayObject *self, PyObject *min, PyObject *max, PyArrayObject *out)
{
PyArray_FastClipFunc *func;
int outgood = 0, ingood = 0;
PyArrayObject *maxa = NULL;
PyArrayObject *mina = NULL;
PyArrayObject *newout = NULL, *newin = NULL;
PyArray_Descr *indescr = NULL, *newdescr = NULL;
char *max_data, *min_data;
PyObject *zero;
/* Treat None the same as NULL */
if (min == Py_None) {
min = NULL;
}
if (max == Py_None) {
max = NULL;
}
if ((max == NULL) && (min == NULL)) {
PyErr_SetString(PyExc_ValueError,
"array_clip: must set either max or min");
return NULL;
}
func = PyArray_DESCR(self)->f->fastclip;
if (func == NULL
|| (min != NULL && !PyArray_CheckAnyScalar(min))
|| (max != NULL && !PyArray_CheckAnyScalar(max))
|| PyArray_ISBYTESWAPPED(self)
|| (out && PyArray_ISBYTESWAPPED(out))) {
return _slow_array_clip(self, min, max, out);
}
/* Use the fast scalar clip function */
/* First we need to figure out the correct type */
if (min != NULL) {
indescr = PyArray_DescrFromObject(min, NULL);
if (indescr == NULL) {
goto fail;
}
}
if (max != NULL) {
newdescr = PyArray_DescrFromObject(max, indescr);
Py_XDECREF(indescr);
indescr = NULL;
if (newdescr == NULL) {
goto fail;
}
}
else {
/* Steal the reference */
newdescr = indescr;
indescr = NULL;
}
/*
* Use the scalar descriptor only if it is of a bigger
* KIND than the input array (and then find the
* type that matches both).
*/
if (PyArray_ScalarKind(newdescr->type_num, NULL) >
PyArray_ScalarKind(PyArray_DESCR(self)->type_num, NULL)) {
indescr = PyArray_PromoteTypes(newdescr, PyArray_DESCR(self));
if (indescr == NULL) {
goto fail;
}
func = indescr->f->fastclip;
if (func == NULL) {
Py_DECREF(indescr);
return _slow_array_clip(self, min, max, out);
}
}
else {
indescr = PyArray_DESCR(self);
Py_INCREF(indescr);
}
Py_DECREF(newdescr);
newdescr = NULL;
if (!PyDataType_ISNOTSWAPPED(indescr)) {
PyArray_Descr *descr2;
descr2 = PyArray_DescrNewByteorder(indescr, '=');
Py_DECREF(indescr);
indescr = NULL;
if (descr2 == NULL) {
goto fail;
}
indescr = descr2;
}
/* Convert max to an array */
if (max != NULL) {
Py_INCREF(indescr);
maxa = (PyArrayObject *)PyArray_FromAny(max, indescr, 0, 0,
NPY_ARRAY_DEFAULT, NULL);
if (maxa == NULL) {
goto fail;
}
}
/*
* If we are unsigned, then make sure min is not < 0
* This is to match the behavior of _slow_array_clip
*
* We allow min and max to go beyond the limits
* for other data-types in which case they
* are interpreted as their modular counterparts.
*/
if (min != NULL) {
if (PyArray_ISUNSIGNED(self)) {
int cmp;
zero = PyInt_FromLong(0);
cmp = PyObject_RichCompareBool(min, zero, Py_LT);
if (cmp == -1) {
Py_DECREF(zero);
goto fail;
}
if (cmp == 1) {
min = zero;
}
else {
Py_DECREF(zero);
Py_INCREF(min);
}
}
else {
Py_INCREF(min);
}
/* Convert min to an array */
Py_INCREF(indescr);
mina = (PyArrayObject *)PyArray_FromAny(min, indescr, 0, 0,
NPY_ARRAY_DEFAULT, NULL);
Py_DECREF(min);
if (mina == NULL) {
goto fail;
}
}
/*
* Check to see if input is single-segment, aligned,
* and in native byteorder
*/
if (PyArray_ISONESEGMENT(self) &&
PyArray_CHKFLAGS(self, NPY_ARRAY_ALIGNED) &&
PyArray_ISNOTSWAPPED(self) &&
(PyArray_DESCR(self) == indescr)) {
ingood = 1;
}
if (!ingood) {
int flags;
if (PyArray_ISFORTRAN(self)) {
flags = NPY_ARRAY_FARRAY;
}
else {
flags = NPY_ARRAY_CARRAY;
}
Py_INCREF(indescr);
newin = (PyArrayObject *)PyArray_FromArray(self, indescr, flags);
if (newin == NULL) {
goto fail;
}
}
else {
newin = self;
Py_INCREF(newin);
}
/*
* At this point, newin is a single-segment, aligned, and correct
* byte-order array of the correct type
*
* if ingood == 0, then it is a copy, otherwise,
* it is the original input.
*/
/*
* If we have already made a copy of the data, then use
* that as the output array
*/
if (out == NULL && !ingood) {
out = newin;
}
/*
* Now, we know newin is a usable array for fastclip,
* we need to make sure the output array is available
* and usable
*/
if (out == NULL) {
Py_INCREF(indescr);
out = (PyArrayObject*)PyArray_NewFromDescr(Py_TYPE(self),
indescr, PyArray_NDIM(self),
PyArray_DIMS(self),
NULL, NULL,
PyArray_ISFORTRAN(self),
(PyObject *)self);
if (out == NULL) {
goto fail;
}
outgood = 1;
}
else Py_INCREF(out);
/* Input is good at this point */
if (out == newin) {
outgood = 1;
}
/* make sure the shape of the output array is the same */
if (!PyArray_SAMESHAPE(newin, out)) {
PyErr_SetString(PyExc_ValueError, "clip: Output array must have the"
"same shape as the input.");
goto fail;
}
if (!outgood && PyArray_EQUIVALENTLY_ITERABLE(
self, out, PyArray_TRIVIALLY_ITERABLE_OP_READ,
PyArray_TRIVIALLY_ITERABLE_OP_NOREAD) &&
PyArray_CHKFLAGS(out, NPY_ARRAY_ALIGNED) &&
PyArray_ISNOTSWAPPED(out) &&
PyArray_EquivTypes(PyArray_DESCR(out), indescr)) {
outgood = 1;
}
/*
* Do we still not have a suitable output array?
* Create one, now. No matter why the array is not suitable a copy has
* to be made. This may be just to avoid memory overlap though.
*/
if (!outgood) {
int oflags;
if (PyArray_ISFORTRAN(self)) {
oflags = NPY_ARRAY_FARRAY;
}
else {
oflags = NPY_ARRAY_CARRAY;
}
oflags |= (NPY_ARRAY_WRITEBACKIFCOPY | NPY_ARRAY_FORCECAST |
NPY_ARRAY_ENSURECOPY);
Py_INCREF(indescr);
newout = (PyArrayObject*)PyArray_FromArray(out, indescr, oflags);
if (newout == NULL) {
goto fail;
}
}
else {
newout = out;
Py_INCREF(newout);
}
/* Now we can call the fast-clip function */
min_data = max_data = NULL;
if (mina != NULL) {
min_data = PyArray_DATA(mina);
}
if (maxa != NULL) {
max_data = PyArray_DATA(maxa);
}
func(PyArray_DATA(newin), PyArray_SIZE(newin), min_data, max_data, PyArray_DATA(newout));
/* Clean up temporary variables */
Py_XDECREF(indescr);
Py_XDECREF(newdescr);
Py_XDECREF(mina);
Py_XDECREF(maxa);
Py_DECREF(newin);
/* Copy back into out if out was not already a nice array. */
PyArray_ResolveWritebackIfCopy(newout);
Py_DECREF(newout);
return (PyObject *)out;
fail:
Py_XDECREF(indescr);
Py_XDECREF(newdescr);
Py_XDECREF(maxa);
Py_XDECREF(mina);
Py_XDECREF(newin);
PyArray_DiscardWritebackIfCopy(newout);
Py_XDECREF(newout);
return NULL;
}
/*NUMPY_API
* Round
*/
NPY_NO_EXPORT PyObject *
PyArray_Round(PyArrayObject *a, int decimals, PyArrayObject *out)
{
PyObject *f, *ret = NULL, *tmp, *op1, *op2;
int ret_int=0;
PyArray_Descr *my_descr;
if (out && (PyArray_SIZE(out) != PyArray_SIZE(a))) {
PyErr_SetString(PyExc_ValueError,
"invalid output shape");
return NULL;
}
if (PyArray_ISCOMPLEX(a)) {
PyObject *part;
PyObject *round_part;
PyObject *arr;
int res;
if (out) {
arr = (PyObject *)out;
Py_INCREF(arr);
}
else {
arr = PyArray_Copy(a);
if (arr == NULL) {
return NULL;
}
}
/* arr.real = a.real.round(decimals) */
part = PyObject_GetAttrString((PyObject *)a, "real");
if (part == NULL) {
Py_DECREF(arr);
return NULL;
}
part = PyArray_EnsureAnyArray(part);
round_part = PyArray_Round((PyArrayObject *)part,
decimals, NULL);
Py_DECREF(part);
if (round_part == NULL) {
Py_DECREF(arr);
return NULL;
}
res = PyObject_SetAttrString(arr, "real", round_part);
Py_DECREF(round_part);
if (res < 0) {
Py_DECREF(arr);
return NULL;
}
/* arr.imag = a.imag.round(decimals) */
part = PyObject_GetAttrString((PyObject *)a, "imag");
if (part == NULL) {
Py_DECREF(arr);
return NULL;
}
part = PyArray_EnsureAnyArray(part);
round_part = PyArray_Round((PyArrayObject *)part,
decimals, NULL);
Py_DECREF(part);
if (round_part == NULL) {
Py_DECREF(arr);
return NULL;
}
res = PyObject_SetAttrString(arr, "imag", round_part);
Py_DECREF(round_part);
if (res < 0) {
Py_DECREF(arr);
return NULL;
}
return arr;
}
/* do the most common case first */
if (decimals >= 0) {
if (PyArray_ISINTEGER(a)) {
if (out) {
if (PyArray_AssignArray(out, a,
NULL, NPY_DEFAULT_ASSIGN_CASTING) < 0) {
return NULL;
}
Py_INCREF(out);
return (PyObject *)out;
}
else {
Py_INCREF(a);
return (PyObject *)a;
}
}
if (decimals == 0) {
if (out) {
return PyObject_CallFunction(n_ops.rint, "OO", a, out);
}
return PyObject_CallFunction(n_ops.rint, "O", a);
}
op1 = n_ops.multiply;
op2 = n_ops.true_divide;
}
else {
op1 = n_ops.true_divide;
op2 = n_ops.multiply;
decimals = -decimals;
}
if (!out) {
if (PyArray_ISINTEGER(a)) {
ret_int = 1;
my_descr = PyArray_DescrFromType(NPY_DOUBLE);
}
else {
Py_INCREF(PyArray_DESCR(a));
my_descr = PyArray_DESCR(a);
}
out = (PyArrayObject *)PyArray_Empty(PyArray_NDIM(a), PyArray_DIMS(a),
my_descr,
PyArray_ISFORTRAN(a));
if (out == NULL) {
return NULL;
}
}
else {
Py_INCREF(out);
}
f = PyFloat_FromDouble(power_of_ten(decimals));
if (f == NULL) {
return NULL;
}
ret = PyObject_CallFunction(op1, "OOO", a, f, out);
if (ret == NULL) {
goto finish;
}
tmp = PyObject_CallFunction(n_ops.rint, "OO", ret, ret);
if (tmp == NULL) {
Py_DECREF(ret);
ret = NULL;
goto finish;
}
Py_DECREF(tmp);
tmp = PyObject_CallFunction(op2, "OOO", ret, f, ret);
if (tmp == NULL) {
Py_DECREF(ret);
ret = NULL;
goto finish;
}
Py_DECREF(tmp);
finish:
Py_DECREF(f);
Py_DECREF(out);
if (ret_int) {
Py_INCREF(PyArray_DESCR(a));
tmp = PyArray_CastToType((PyArrayObject *)ret,
PyArray_DESCR(a), PyArray_ISFORTRAN(a));
Py_DECREF(ret);
return tmp;
}
return ret;
}
/*NUMPY_API*/
NPY_NO_EXPORT int
PyArray_CopyObject(PyArrayObject *dest, PyObject *src_object)
{
int ret;
PyArrayObject *src;
PyArray_Descr *dtype = NULL;
int ndim = 0;
npy_intp dims[NPY_MAXDIMS];
Py_INCREF(src_object);
/*
* Special code to mimic Numeric behavior for
* character arrays.
*/
if (dest->descr->type == PyArray_CHARLTR && dest->nd > 0 \
&& PyString_Check(src_object)) {
npy_intp n_new, n_old;
char *new_string;
PyObject *tmp;
n_new = dest->dimensions[dest->nd-1];
n_old = PyString_Size(src_object);
if (n_new > n_old) {
new_string = (char *)malloc(n_new);
memmove(new_string, PyString_AS_STRING(src_object), n_old);
memset(new_string + n_old, ' ', n_new - n_old);
tmp = PyString_FromStringAndSize(new_string, n_new);
free(new_string);
Py_DECREF(src_object);
src_object = tmp;
}
}
/*
* Get either an array object we can copy from, or its parameters
* if there isn't a convenient array available.
*/
if (PyArray_GetArrayParamsFromObject(src_object, PyArray_DESCR(dest),
0, &dtype, &ndim, dims, &src, NULL) < 0) {
Py_DECREF(src_object);
return -1;
}
/* If it's not an array, either assign from a sequence or as a scalar */
if (src == NULL) {
/* If the input is scalar */
if (ndim == 0) {
/* If there's one dest element and src is a Python scalar */
if (PyArray_IsScalar(src_object, Generic)) {
src = (PyArrayObject *)PyArray_FromScalar(src_object, dtype);
if (src == NULL) {
Py_DECREF(src_object);
return -1;
}
}
else {
if (PyArray_SIZE(dest) == 1) {
Py_DECREF(dtype);
return PyArray_DESCR(dest)->f->setitem(src_object,
PyArray_DATA(dest), dest);
}
else {
src = (PyArrayObject *)PyArray_NewFromDescr(&PyArray_Type,
dtype, 0, NULL, NULL,
NULL, 0, NULL);
if (src == NULL) {
Py_DECREF(src_object);
return -1;
}
if (PyArray_DESCR(src)->f->setitem(src_object,
PyArray_DATA(src), src) < 0) {
Py_DECREF(src_object);
Py_DECREF(src);
return -1;
}
}
}
}
else {
/*
* If there are more than enough dims, use AssignFromSequence
* because it can handle this style of broadcasting.
*/
if (ndim >= PyArray_NDIM(dest)) {
int res;
Py_DECREF(dtype);
res = PyArray_AssignFromSequence(dest, src_object);
Py_DECREF(src_object);
return res;
}
/* Otherwise convert to an array and do an array-based copy */
src = (PyArrayObject *)PyArray_NewFromDescr(&PyArray_Type,
dtype, ndim, dims, NULL, NULL,
PyArray_ISFORTRAN(dest), NULL);
if (src == NULL) {
Py_DECREF(src_object);
return -1;
}
if (PyArray_AssignFromSequence(src, src_object) < 0) {
Py_DECREF(src);
Py_DECREF(src_object);
return -1;
}
}
}
/* If it's an array, do a move (handling possible overlapping data) */
ret = PyArray_MoveInto(dest, src);
Py_DECREF(src);
Py_DECREF(src_object);
return ret;
}
