Mentions légales du service

Skip to content
Snippets Groups Projects
Commit e80c8a80 authored by hhakim's avatar hhakim
Browse files

Move non needed template Transform function specializations into template... 

Move non needed template Transform function specializations into template functions (from .cpp.in to hpp).
parent b10ca730
Branches
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
src/faust_linear_operator/GPU2/faust_Transform_gpu.cpp.in
+ 4
− 395
View file @ e80c8a80
#include "faust_Transform_gpu.h" #include "faust_Transform_gpu.h"
namespace Faust namespace Faust
{ {
template<>
Faust::RefManager Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::ref_man([](void *fact)
{
#ifdef DEBUG
std::cout << "Faust::Transform delete_fact" << std::endl;
#endif
delete static_cast<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*>(fact);
});
template<> template<>
int32_t Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::getNbRow() const int32_t Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::getNbRow() const
...@@ -31,7 +23,7 @@ namespace Faust ...@@ -31,7 +23,7 @@ namespace Faust
} }
template<> template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Transform() : gpu_mat_arr(nullptr), dtor_delete_data(false), dtor_disabled(false), data(std::vector<Faust::MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*>()) Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Transform() : gpu_mat_arr(nullptr), dtor_delete_data(false), dtor_disabled(false), data(std::vector<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*>())
{ {
} }
...@@ -84,59 +76,6 @@ namespace Faust ...@@ -84,59 +76,6 @@ namespace Faust
return marr_funcs->size(gpu_mat_arr); return marr_funcs->size(gpu_mat_arr);
} }
template<>
MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>* Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_fact(int32_t id, bool cloning_fact) const
{
// auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0));
// auto is_sparse = marr_funcs->is_sparse_at(this->gpu_mat_arr, id);
// MatGeneric<@FAUST_SCALAR_FOR_GM@, GPU2>* M = nullptr;
// if(is_sparse)
// {
// M = new MatSparse<@FAUST_SCALAR_FOR_GM@,GPU2>();
// M->set_gpu_mat_ptr(marr_funcs->get_spm(this->gpu_mat_arr, id));
// }
// else
// {
// M = new MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>();
// M->set_gpu_mat_ptr(marr_funcs->get_dsm(this->gpu_mat_arr, id));
// }
// if(cloning_fact)
// {
// auto tmp = M;
// M = M->clone();
// delete tmp;
// }
// return M;
if(cloning_fact)
return data[id]->clone();
else
return data[id];
}
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::update(const MatGeneric<@FAUST_SCALAR_FOR_GM@, GPU2>& M, const faust_unsigned_int id)
{
auto fact = get_fact(id, false);
auto fact_type = fact->getType();
if(M.getType() != fact_type)
throw std::runtime_error("The factor matrix to update is not of the same type (dense or sparse) as the input matrix.");
if(fact_type == Dense)
{
// fact to update is dense
auto dfact = dynamic_cast<MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>*>(fact);
auto dM = dynamic_cast<const MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>*>(&M);
*dfact = *dM;
}
else
{
// fact to update is sparse
auto sfact = dynamic_cast<MatSparse<@FAUST_SCALAR_FOR_GM@,GPU2>*>(fact);
auto sM = dynamic_cast<const MatSparse<@FAUST_SCALAR_FOR_GM@,GPU2>*>(&M);
*sfact = *sM;
}
// fact->set_id(M.is_id()); //TODO: in MatGeneric<FPP,GPU2> add is_id/set_id
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Display(bool transpose/*=false*/) const void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Display(bool transpose/*=false*/) const
{ {
...@@ -144,12 +83,6 @@ namespace Faust ...@@ -144,12 +83,6 @@ namespace Faust
marr_funcs->display_op(gpu_mat_arr, transpose?OP_TRANSP:OP_NOTRANSP); marr_funcs->display_op(gpu_mat_arr, transpose?OP_TRANSP:OP_NOTRANSP);
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::update_total_nnz() const
{
//nothing to do get_total_nnz doesn't use any cache by now
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::replace(const MatGeneric<@FAUST_SCALAR_FOR_GM@, GPU2>* new_mat, const faust_unsigned_int id) void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::replace(const MatGeneric<@FAUST_SCALAR_FOR_GM@, GPU2>* new_mat, const faust_unsigned_int id)
{ {
...@@ -162,21 +95,12 @@ namespace Faust ...@@ -162,21 +95,12 @@ namespace Faust
delete data[id]; delete data[id];
else else
ref_man.release(data[id]); ref_man.release(data[id]);
data[id] = const_cast<Faust::MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*>(new_mat); data[id] = const_cast<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*>(new_mat);
if(! dtor_delete_data) if(! dtor_delete_data)
ref_man.acquire(data[id]); ref_man.acquire(data[id]);
this->update_total_nnz(); this->update_total_nnz();
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::get_facts(std::vector<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*> &factors, bool cloning_facts/*=true*/) const
{
for(int i=0;i < size(); i++)
{
factors.push_back(get_fact(i, cloning_facts));
}
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::clear() void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::clear()
{ {
...@@ -196,22 +120,6 @@ namespace Faust ...@@ -196,22 +120,6 @@ namespace Faust
gpu_mat_arr = marr_funcs->create(); gpu_mat_arr = marr_funcs->create();
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::operator=(const Transform<@FAUST_SCALAR_FOR_GM@,GPU2>& t)
{
this->clear();
for(int i=0;i<t.size();i++)
{
this->push_back(t.get_fact(i, /*cloning*/false), /*copying*/ true);
}
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Transform(const Transform<@FAUST_SCALAR_FOR_GM@,GPU2> & t) : Transform<@FAUST_SCALAR_FOR_GM@,GPU2>()
{
*this = t;
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::insert(int32_t id, const MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>* M, bool copying/*=true*/) void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::insert(int32_t id, const MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>* M, bool copying/*=true*/)
{ {
...@@ -226,12 +134,6 @@ namespace Faust ...@@ -226,12 +134,6 @@ namespace Faust
if(!dtor_delete_data) ref_man.acquire(ins_M); if(!dtor_delete_data) ref_man.acquire(ins_M);
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::push_first(const MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>* M, bool copying/*=true*/)
{
this->insert(0, M, copying);
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::erase(int32_t id) void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::erase(int32_t id)
{ {
...@@ -244,12 +146,6 @@ namespace Faust ...@@ -244,12 +146,6 @@ namespace Faust
data.erase(data.begin()+id); data.erase(data.begin()+id);
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::pop_front()
{
this->erase(0);
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::transpose() void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::transpose()
{ {
...@@ -261,12 +157,6 @@ namespace Faust ...@@ -261,12 +157,6 @@ namespace Faust
std::reverse(data.begin(),data.end()); std::reverse(data.begin(),data.end());
} }
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::pop_back()
{
this->erase(size()-1);
}
template<> template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Transform(const std::vector<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*> &factors, const @FAUST_SCALAR_FOR_GM@ lambda_ /*= (FPP)1.0*/, const bool optimizedCopy/*=false*/, const bool cloning_fact/*=true*/) : Transform() Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::Transform(const std::vector<MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>*> &factors, const @FAUST_SCALAR_FOR_GM@ lambda_ /*= (FPP)1.0*/, const bool optimizedCopy/*=false*/, const bool cloning_fact/*=true*/) : Transform()
{ {
...@@ -315,14 +205,6 @@ namespace Faust ...@@ -315,14 +205,6 @@ namespace Faust
if(isConj && opThis == 'N') M.conjugate(); if(isConj && opThis == 'N') M.conjugate();
} }
template<>
MatDense<@FAUST_SCALAR_FOR_GM@,GPU2> Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_product(const char opThis/*='N'*/, const bool isConj/*=false*/) const
{
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> M;
this->get_product(M, opThis, isConj);
return M;
}
template<> template<>
faust_unsigned_int Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_total_nnz() const faust_unsigned_int Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_total_nnz() const
{ {
...@@ -361,30 +243,7 @@ namespace Faust ...@@ -361,30 +243,7 @@ namespace Faust
} }
template<> template<>
void Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::multiply(const Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2> & A) MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::sliceMultiply(const Slice s[2], MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>& gpu_X, const char opThis) const
{
if (A.size() != 0)
{
if (size() == 0)
{
(*this)=A;
}
else
{
if (getNbCol() != A.getNbRow())
{
throw std::runtime_error("Dimensions must agree");
}
for (int i=0;i<A.size();i++)
{
this->push_back(A.get_fact(i, /*cloning*/ false)); // push_back clones it afterward
}
}
}
}
template<>
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::sliceMultiply(const Slice s[2], MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>& gpu_X, const char opThis) const
{ {
auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0)); auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0));
if(gpu_mat_arr == nullptr) throw std::runtime_error("gpu_mat_arr is nullptr"); if(gpu_mat_arr == nullptr) throw std::runtime_error("gpu_mat_arr is nullptr");
...@@ -419,7 +278,7 @@ namespace Faust ...@@ -419,7 +278,7 @@ namespace Faust
} }
template<> template<>
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::indexMultiply(faust_unsigned_int* ids[2], size_t id_lens[2], MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>& gpu_X, const char opThis) const MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::indexMultiply(faust_unsigned_int* ids[2], size_t id_lens[2], MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>& gpu_X, const char opThis) const
{ {
auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0)); auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0));
if(gpu_mat_arr == nullptr) throw std::runtime_error("gpu_mat_arr is nullptr"); if(gpu_mat_arr == nullptr) throw std::runtime_error("gpu_mat_arr is nullptr");
...@@ -446,260 +305,10 @@ namespace Faust ...@@ -446,260 +305,10 @@ namespace Faust
return out; return out;
} }
template<>
void Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::multiplyLeft(const Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2> & A)
{
if (A.size() != 0)
{
if (size() == 0)
{
(*this)=A;
}
else
{
if (A.getNbCol() != getNbRow())
{
throw std::runtime_error("Dimensions must agree");
}
for (int i=A.size()-1;i>-1;i--)
{
this->push_first(A.get_fact(i, /*cloning*/ false)); // push_back clones it afterward
}
}
}
}
template<> template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::multiply(const @FAUST_SCALAR_FOR_GM@& a, const int32_t id/*=-1*/) void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::multiply(const @FAUST_SCALAR_FOR_GM@& a, const int32_t id/*=-1*/)
{ {
auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0)); auto marr_funcs = GPUModHandler::get_singleton()->marr_funcs((@FAUST_SCALAR_FOR_GM@)(0));
marr_funcs->scalar_mul_id(this->gpu_mat_arr, reinterpret_cast<const @GM_SCALAR@*>(&a), id); marr_funcs->scalar_mul_id(this->gpu_mat_arr, reinterpret_cast<const @GM_SCALAR@*>(&a), id);
} }
template<>
bool Faust::Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::is_fact_sparse(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == Sparse;
}
template<>
bool Faust::Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::is_fact_dense(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == Dense;
}
template<>
bool Faust::Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::is_fact_bsr(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == BSR;
}
template<>
void Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_fact(const faust_unsigned_int &id,
@FAUST_SCALAR_FOR_GM@* elts,
faust_unsigned_int* num_rows,
faust_unsigned_int* num_cols,
const bool transpose /*=false*/) const
{
if(! is_fact_dense(id))
throw std::runtime_error("faust_Transform_gpu: this get_fact function signature is for MatDense only.");
auto gen_mat = get_fact(id, /*cloning*/ false);
auto dmat = dynamic_cast<MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>*>(gen_mat);
@FAUST_SCALAR_FOR_GM@ tmp;
*num_cols = gen_mat->getNbCol();
*num_rows = gen_mat->getNbRow();
/******* failsafe copy **/
auto cpu_mdense = new MatDense<@FAUST_SCALAR_FOR_GM@,Cpu>(dmat->getNbRow(), dmat->getNbCol());
dmat->tocpu(*cpu_mdense);
memcpy(elts, cpu_mdense->getData(),cpu_mdense->getNbCol()*cpu_mdense->getNbRow()*sizeof(@FAUST_SCALAR_FOR_GM@));
delete cpu_mdense;
/********* efficient copy that should work but doesn't */
// TODO: normally it must be possible to copy directly to elts but I noticed a segfault (and I couldn't figure out how to fix it)
// dmat->tocpu(elts);
/***************************/
if(transpose)
{
// transpose in-place
// the matrix is in column-major order
for(int j=0;j<*num_cols;j++)
for(int i=0; i<*num_rows;i++)
{
tmp = elts[i**num_cols+j];
elts[i**num_cols+j] = elts[j**num_rows+i];
elts[j**num_rows+i] = tmp;
}
// swap num_cols and num_rows
// (with only these 2 variables -- F2 arithmetic trick)
*num_cols = *num_cols^*num_rows;
*num_rows = *num_cols^*num_rows;
*num_cols = *num_cols^*num_rows;
}
}
template<>
void Faust::Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::get_fact(const faust_unsigned_int id,
int* d_outer_count_ptr, int* d_inner_ptr, @FAUST_SCALAR_FOR_GM@* d_elts,
faust_unsigned_int* nnz,
faust_unsigned_int* num_rows, faust_unsigned_int* num_cols,
bool transpose /* default to false */) const
{
if(! is_fact_sparse(id))
throw std::runtime_error("faust_Transform_gpu: this get_fact function signature is for MatSparse only.");
auto gen_mat = get_fact(id, /*cloning*/ false);
auto smat = dynamic_cast<MatSparse<@FAUST_SCALAR_FOR_GM@,GPU2>*>(gen_mat);
MatSparse<@FAUST_SCALAR_FOR_GM@,Cpu> cpu_smat;
if(transpose)
{
auto t_smat = smat->clone();
t_smat->transpose();
t_smat->tocpu(d_outer_count_ptr, d_inner_ptr, d_elts, (int32_t*) num_rows, (int32_t*) num_cols, (int32_t*) nnz);
delete t_smat;
}
else
{
smat->tocpu(d_outer_count_ptr, d_inner_ptr, d_elts, (int32_t*) num_rows, (int32_t*) num_cols, (int32_t*) nnz);
}
}
template<>
Vect<@FAUST_SCALAR_FOR_GM@,GPU2> Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::multiply(const Vect<@FAUST_SCALAR_FOR_GM@,GPU2>& x, const char opThis/*='N'*/)
{
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> out = this->multiply((const MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>)x, opThis);
Vect<@FAUST_SCALAR_FOR_GM@,GPU2> v_out;
v_out.gpu_mat = out.gpu_mat;
out.gpu_mat = nullptr; // avoid freeing v_out.gpu_mat when out of scope
return v_out;
}
template<>
MatGeneric<@FAUST_SCALAR_FOR_GM@,GPU2>* Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator*() const
{
return container.get_fact(index, /*cloning_fact*/ false);
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator& Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator++()
{
index++;
return *this;
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator+(int arg)
{
iterator copy(*this);
copy.index = this->index+arg;
return copy;
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator-(int arg)
{
iterator copy(*this);
copy.index = this->index-arg;
return copy;
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator++(int)
{
iterator copy(*this);
this->index++;
return copy;
}
template<>
bool Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator<(const Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator& it)
{
return this->index < it.index;
}
template<>
bool Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::operator!=(const Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator& it)
{
return this->index != it.index;
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator::iterator(const Transform<@FAUST_SCALAR_FOR_GM@, GPU2>& container, size_t index) : index(index), container(container)
{
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::begin() const
{
return Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator(*this, 0);
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::end() const
{
return Transform<@FAUST_SCALAR_FOR_GM@,GPU2>::iterator(*this, size());
}
template<>
void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::tocpu(Transform<@FAUST_SCALAR_FOR_GM@, Cpu>& cpu_transf) const
{
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>* gpu_mdense = nullptr;
MatSparse<@FAUST_SCALAR_FOR_GM@, GPU2>* gpu_msparse = nullptr;
MatBSR<@FAUST_SCALAR_FOR_GM@, GPU2>* gpu_mbsr = nullptr;
for(auto gpu_mat: data)
{
if(gpu_mdense = dynamic_cast<MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>*>(gpu_mat))
{
auto cpu_mdense = new MatDense<@FAUST_SCALAR_FOR_GM@,Cpu>(gpu_mdense->getNbRow(), gpu_mdense->getNbCol());
gpu_mdense->tocpu(*cpu_mdense);
cpu_transf.push_back(cpu_mdense, false, false);
}
else if(gpu_msparse = dynamic_cast<MatSparse<@FAUST_SCALAR_FOR_GM@, GPU2>*>(gpu_mat))
{
auto cpu_msparse = new MatSparse<@FAUST_SCALAR_FOR_GM@,Cpu>();
cpu_msparse->resize(gpu_msparse->getNonZeros(), gpu_msparse->getNbRow(), gpu_msparse->getNbCol());
gpu_msparse->tocpu(*cpu_msparse);
cpu_transf.push_back(cpu_msparse, false, false, false);
}
else if(gpu_mbsr = dynamic_cast<MatBSR<@FAUST_SCALAR_FOR_GM@, GPU2>*>(gpu_mat))
{
auto cpu_mbsr = new MatBSR<@FAUST_SCALAR_FOR_GM@,Cpu>();
gpu_mbsr->tocpu(*cpu_mbsr);
cpu_transf.push_back(cpu_mbsr, false, false, false);
}
else
throw std::runtime_error("Invalid matrix pointer");
}
}
template<>
Transform<@FAUST_SCALAR_FOR_GM@, Cpu> Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::tocpu() const
{
Transform<@FAUST_SCALAR_FOR_GM@, Cpu> cpu_transf;
tocpu(cpu_transf);
return cpu_transf;
}
template<>
void Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::save_mat_file(const char* filename, const bool transpose, const bool conjugate) const
{
Transform<@FAUST_SCALAR_FOR_GM@,Cpu> cpu_transf;
this->tocpu(cpu_transf);
cpu_transf.save_mat_file(filename, transpose, conjugate);
}
template<>
Real<@FAUST_SCALAR_FOR_GM@> Faust::Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::normL1(const bool transpose /* = false */, const bool full_array/*=true*/, const int batch_size/*=1*/) const
{
double norm;
MatDense<@FAUST_SCALAR_FOR_GM@, GPU2> full = get_product(transpose?'T':'N');
norm = std::abs(full.normL1(/*transpose*/)); //transpose not necessary because full is already transposed if needed
return norm;
}
template<>
faust_unsigned_int Transform<@FAUST_SCALAR_FOR_GM@, GPU2>::get_fact_nnz(const faust_unsigned_int id) const
{
return this->get_fact(id, false)->getNonZeros();
}
} }
src/faust_linear_operator/GPU2/faust_Transform_gpu.hpp
+ 374
− 0
View file @ e80c8a80
namespace Faust namespace Faust
{ {
template<typename FPP>
RefManager Transform<FPP,GPU2>::ref_man([](void *fact)
{
#ifdef DEBUG
std::cout << "Transform delete_fact" << std::endl;
#endif
delete static_cast<MatGeneric<FPP,GPU2>*>(fact);
});
template<typename FPP>
MatGeneric<FPP,GPU2>* Transform<FPP,GPU2>::get_fact(int32_t id, bool cloning_fact) const
{
if(cloning_fact)
return data[id]->clone();
else
return data[id];
}
template<typename FPP>
void Transform<FPP,GPU2>::update(const MatGeneric<FPP, GPU2>& M, const faust_unsigned_int id)
{
auto fact = get_fact(id, false);
auto fact_type = fact->getType();
if(M.getType() != fact_type)
throw std::runtime_error("The factor matrix to update is not of the same type (dense or sparse) as the input matrix.");
if(fact_type == Dense)
{
// fact to update is dense
auto dfact = dynamic_cast<MatDense<FPP,GPU2>*>(fact);
auto dM = dynamic_cast<const MatDense<FPP,GPU2>*>(&M);
*dfact = *dM;
}
else
{
// fact to update is sparse
auto sfact = dynamic_cast<MatSparse<FPP,GPU2>*>(fact);
auto sM = dynamic_cast<const MatSparse<FPP,GPU2>*>(&M);
*sfact = *sM;
}
// fact->set_id(M.is_id()); //TODO: in MatGeneric<FPP,GPU2> add is_id/set_id
}
template<typename FPP>
void Transform<FPP,GPU2>::update_total_nnz() const
{
//nothing to do get_total_nnz doesn't use any cache by now
}
template<typename FPP>
void Transform<FPP, GPU2>::get_facts(std::vector<MatGeneric<FPP,GPU2>*> &factors, bool cloning_facts/*=true*/) const
{
for(int i=0;i < size(); i++)
{
factors.push_back(get_fact(i, cloning_facts));
}
}
template<typename FPP>
void Transform<FPP, GPU2>::operator=(const Transform<FPP,GPU2>& t)
{
this->clear();
for(int i=0;i<t.size();i++)
{
this->push_back(t.get_fact(i, /*cloning*/false), /*copying*/ true);
}
}
template<typename FPP>
Transform<FPP,GPU2>::Transform(const Transform<FPP,GPU2> & t) : Transform<FPP,GPU2>()
{
*this = t;
}
template<typename FPP>
void Transform<FPP,GPU2>::push_first(const MatGeneric<FPP,GPU2>* M, bool copying/*=true*/)
{
this->insert(0, M, copying);
}
template<typename FPP>
void Transform<FPP,GPU2>::pop_front()
{
this->erase(0);
}
template<typename FPP>
void Transform<FPP,GPU2>::pop_back()
{
this->erase(size()-1);
}
template<typename FPP>
MatDense<FPP,GPU2> Transform<FPP,GPU2>::get_product(const char opThis/*='N'*/, const bool isConj/*=false*/) const
{
MatDense<FPP, GPU2> M;
this->get_product(M, opThis, isConj);
return M;
}
template<typename FPP>
void Transform<FPP,GPU2>::multiply(const Transform<FPP,GPU2> & A)
{
if (A.size() != 0)
{
if (size() == 0)
{
(*this)=A;
}
else
{
if (getNbCol() != A.getNbRow())
{
throw std::runtime_error("Dimensions must agree");
}
for (int i=0;i<A.size();i++)
{
this->push_back(A.get_fact(i, /*cloning*/ false)); // push_back clones it afterward
}
}
}
}
template<typename FPP>
void Transform<FPP,GPU2>::multiplyLeft(const Transform<FPP,GPU2> & A)
{
if (A.size() != 0)
{
if (size() == 0)
{
(*this)=A;
}
else
{
if (A.getNbCol() != getNbRow())
{
throw std::runtime_error("Dimensions must agree");
}
for (int i=A.size()-1;i>-1;i--)
{
this->push_first(A.get_fact(i, /*cloning*/ false)); // push_back clones it afterward
}
}
}
}
template<typename FPP>
bool Transform<FPP, GPU2>::is_fact_sparse(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == Sparse;
}
template<typename FPP>
bool Transform<FPP, GPU2>::is_fact_dense(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == Dense;
}
template<typename FPP>
bool Transform<FPP, GPU2>::is_fact_bsr(int id) const
{
return get_fact(id, /*cloning*/ false)->getType() == BSR;
}
template<typename FPP>
void Transform<FPP,GPU2>::get_fact(const faust_unsigned_int &id,
FPP* elts,
faust_unsigned_int* num_rows,
faust_unsigned_int* num_cols,
const bool transpose /*=false*/) const
{
if(! is_fact_dense(id))
throw std::runtime_error("faust_Transform_gpu: this get_fact function signature is for MatDense only.");
auto gen_mat = get_fact(id, /*cloning*/ false);
auto dmat = dynamic_cast<MatDense<FPP,GPU2>*>(gen_mat);
FPP tmp;
*num_cols = gen_mat->getNbCol();
*num_rows = gen_mat->getNbRow();
/******* failsafe copy **/
auto cpu_mdense = new MatDense<FPP,Cpu>(dmat->getNbRow(), dmat->getNbCol());
dmat->tocpu(*cpu_mdense);
memcpy(elts, cpu_mdense->getData(),cpu_mdense->getNbCol()*cpu_mdense->getNbRow()*sizeof(FPP));
delete cpu_mdense;
/********* efficient copy that should work but doesn't */
// TODO: normally it must be possible to copy directly to elts but I noticed a segfault (and I couldn't figure out how to fix it)
// dmat->tocpu(elts);
/***************************/
if(transpose)
{
// transpose in-place
// the matrix is in column-major order
for(int j=0;j<*num_cols;j++)
for(int i=0; i<*num_rows;i++)
{
tmp = elts[i**num_cols+j];
elts[i**num_cols+j] = elts[j**num_rows+i];
elts[j**num_rows+i] = tmp;
}
// swap num_cols and num_rows
// (with only these 2 variables -- F2 arithmetic trick)
*num_cols = *num_cols^*num_rows;
*num_rows = *num_cols^*num_rows;
*num_cols = *num_cols^*num_rows;
}
}
template<typename FPP>
void Transform<FPP,GPU2>::get_fact(const faust_unsigned_int id,
int* d_outer_count_ptr, int* d_inner_ptr, FPP* d_elts,
faust_unsigned_int* nnz,
faust_unsigned_int* num_rows, faust_unsigned_int* num_cols,
bool transpose /* default to false */) const
{
if(! is_fact_sparse(id))
throw std::runtime_error("faust_Transform_gpu: this get_fact function signature is for MatSparse only.");
auto gen_mat = get_fact(id, /*cloning*/ false);
auto smat = dynamic_cast<MatSparse<FPP,GPU2>*>(gen_mat);
MatSparse<FPP,Cpu> cpu_smat;
if(transpose)
{
auto t_smat = smat->clone();
t_smat->transpose();
t_smat->tocpu(d_outer_count_ptr, d_inner_ptr, d_elts, (int32_t*) num_rows, (int32_t*) num_cols, (int32_t*) nnz);
delete t_smat;
}
else
{
smat->tocpu(d_outer_count_ptr, d_inner_ptr, d_elts, (int32_t*) num_rows, (int32_t*) num_cols, (int32_t*) nnz);
}
}
template<typename FPP>
Vect<FPP,GPU2> Transform<FPP,GPU2>::multiply(const Vect<FPP,GPU2>& x, const char opThis/*='N'*/)
{
MatDense<FPP, GPU2> out = this->multiply((const MatDense<FPP,GPU2>)x, opThis);
Vect<FPP,GPU2> v_out;
v_out.gpu_mat = out.gpu_mat;
out.gpu_mat = nullptr; // avoid freeing v_out.gpu_mat when out of scope
return v_out;
}
template<typename FPP>
MatGeneric<FPP,GPU2>* Transform<FPP,GPU2>::iterator::operator*() const
{
return container.get_fact(index, /*cloning_fact*/ false);
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator& Transform<FPP,GPU2>::iterator::operator++()
{
index++;
return *this;
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator Transform<FPP,GPU2>::iterator::operator+(int arg)
{
iterator copy(*this);
copy.index = this->index+arg;
return copy;
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator Transform<FPP,GPU2>::iterator::operator-(int arg)
{
iterator copy(*this);
copy.index = this->index-arg;
return copy;
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator Transform<FPP,GPU2>::iterator::operator++(int)
{
iterator copy(*this);
this->index++;
return copy;
}
template<typename FPP>
bool Transform<FPP,GPU2>::iterator::operator<(const Transform<FPP,GPU2>::iterator& it)
{
return this->index < it.index;
}
template<typename FPP>
bool Transform<FPP,GPU2>::iterator::operator!=(const Transform<FPP,GPU2>::iterator& it)
{
return this->index != it.index;
}
template<typename FPP>
Transform<FPP,GPU2>::iterator::iterator(const Transform<FPP, GPU2>& container, size_t index) : index(index), container(container)
{
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator Transform<FPP,GPU2>::begin() const
{
return Transform<FPP,GPU2>::iterator(*this, 0);
}
template<typename FPP>
typename Transform<FPP,GPU2>::iterator Transform<FPP,GPU2>::end() const
{
return Transform<FPP,GPU2>::iterator(*this, size());
}
template<typename FPP>
void Transform<FPP, GPU2>::tocpu(Transform<FPP, Cpu>& cpu_transf) const
{
MatDense<FPP, GPU2>* gpu_mdense = nullptr;
MatSparse<FPP, GPU2>* gpu_msparse = nullptr;
MatBSR<FPP, GPU2>* gpu_mbsr = nullptr;
for(auto gpu_mat: data)
{
if(gpu_mdense = dynamic_cast<MatDense<FPP, GPU2>*>(gpu_mat))
{
auto cpu_mdense = new MatDense<FPP,Cpu>(gpu_mdense->getNbRow(), gpu_mdense->getNbCol());
gpu_mdense->tocpu(*cpu_mdense);
cpu_transf.push_back(cpu_mdense, false, false);
}
else if(gpu_msparse = dynamic_cast<MatSparse<FPP, GPU2>*>(gpu_mat))
{
auto cpu_msparse = new MatSparse<FPP,Cpu>();
cpu_msparse->resize(gpu_msparse->getNonZeros(), gpu_msparse->getNbRow(), gpu_msparse->getNbCol());
gpu_msparse->tocpu(*cpu_msparse);
cpu_transf.push_back(cpu_msparse, false, false, false);
}
else if(gpu_mbsr = dynamic_cast<MatBSR<FPP, GPU2>*>(gpu_mat))
{
auto cpu_mbsr = new MatBSR<FPP,Cpu>();
gpu_mbsr->tocpu(*cpu_mbsr);
cpu_transf.push_back(cpu_mbsr, false, false, false);
}
else
throw std::runtime_error("Invalid matrix pointer");
}
}
//TODO: refactor to generic CPU/GPU code (using if needed a non-member function on Transform<FPP, DEV>) //TODO: refactor to generic CPU/GPU code (using if needed a non-member function on Transform<FPP, DEV>)
template<typename FPP> template<typename FPP>
MatDense<FPP,GPU2> Transform<FPP,GPU2>::multiply(const MatDense<FPP,GPU2> &A, const char opThis) /*const*/ //TODO: should be const MatDense<FPP,GPU2> Transform<FPP,GPU2>::multiply(const MatDense<FPP,GPU2> &A, const char opThis) /*const*/ //TODO: should be const
...@@ -19,4 +360,37 @@ namespace Faust ...@@ -19,4 +360,37 @@ namespace Faust
return mat; return mat;
} }
template<typename FPP>
Transform<FPP, Cpu> Transform<FPP, GPU2>::tocpu() const
{
Transform<FPP, Cpu> cpu_transf;
tocpu(cpu_transf);
return cpu_transf;
}
template<typename FPP>
void Transform<FPP, GPU2>::save_mat_file(const char* filename, const bool transpose, const bool conjugate) const
{
Transform<FPP,Cpu> cpu_transf;
this->tocpu(cpu_transf);
cpu_transf.save_mat_file(filename, transpose, conjugate);
}
template<typename FPP>
Real<FPP> Transform<FPP, GPU2>::normL1(const bool transpose /* = false */, const bool full_array/*=true*/, const int batch_size/*=1*/) const
{
double norm;
MatDense<FPP, GPU2> full = get_product(transpose?'T':'N');
norm = std::abs(full.normL1(/*transpose*/)); //transpose not necessary because full is already transposed if needed
return norm;
}
template<typename FPP>
faust_unsigned_int Transform<FPP, GPU2>::get_fact_nnz(const faust_unsigned_int id) const
{
return this->get_fact(id, false)->getNonZeros();
}
} }
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment