CNTK/Source/Math/cudalattice.h

132 строки
7.4 KiB
C++

// cudalattice.h -- lattice forward/backward functions CUDA execution
//
// F. Seide, V-hansu
#pragma once
#include <stdexcept> // (for NOCUDA version only)
#include "latticestorage.h" // basic data types for storing lattices
#define __kernel_emulation__ // allow the compilation of CUDA kernels on the CPU
#include "latticefunctionskernels.h" // for data types passed to kernel functions
#undef __kernel_emulation__
#include "Matrix.h"
namespace msra { namespace cuda {
// ---------------------------------------------------------------------------
// how to add vector types here:
// - for types with methods that operate on the vector, follow the example of 'somedatavector'
// - for types without extra methods, follow the example of 'ushortvector'
// - the total set of classes that need to be defined is:
// - XXXvector --the interface that the calling code sees (in this header file)
// - XXXvectorimpl --glue for each method: switch CUDA device, then call kernel (in XXXvectorops, cudalattice.cpp)
// - XXXvectorops --contains the actual kernel functions (cudalatticeops.h, cudalatticeops.cu.h)
// - comments:
// - follow precisely this pattern
// - for non-vector elements you need to pass by reference, you can still use this with a size-1 vector
// - and for functions that have no 'this', use a size-0 vector
// - note that you cannot add data members to XXXvector, as there is no mechanism to sync between GPU and CPU; only vector elements are synced
// - remember to Get Latest in both projects to see the change, for cudalattice.h and latticestorage.h
// ---------------------------------------------------------------------------
// a vector type; use this as a basetype
// Note that this will only be instantiated for types known inside this lib, and a newvector<> function must be exported for each.
template <typename ELEMTYPE>
struct /*interface*/ vectorbase
{
virtual ~vectorbase()
{
}
virtual void allocate(size_t n) = 0;
virtual size_t size() const throw() = 0;
virtual void assign(const ELEMTYPE* p, size_t n, bool synchronize) = 0;
template <class VECTOR>
void assign(const VECTOR& v, bool synchronize)
{
allocate(v.size());
if (!v.empty())
assign(&v[0], v.size(), synchronize);
}
virtual void fetch(ELEMTYPE* p, size_t n, bool synchronize) const = 0;
template <class VECTOR>
void fetch(VECTOR& v, bool synchronize) const
{
v.resize(size());
if (!v.empty())
fetch(&v[0], v.size(), synchronize);
}
typedef ELEMTYPE elemtype;
};
// ---------------------------------------------------------------------------
// vectors of these with custom functions
// The XXXvector classes must derive from vectorbase<XXX>.
// ---------------------------------------------------------------------------
struct somedatavector : public vectorbase<msra::lattices::somedata>
{
// must implement all members of vectorbase<>, and can add operations here
virtual int somedataoperation(size_t arg) = 0;
};
typedef vectorbase<unsigned short> ushortvector;
typedef vectorbase<float> floatvector;
typedef vectorbase<double> doublevector;
typedef vectorbase<unsigned int> uintvector;
typedef vectorbase<size_t> sizetvector;
typedef vectorbase<msra::lattices::lrhmmdef> lrhmmdefvector;
typedef vectorbase<msra::lattices::lr3transP> lr3transPvector;
typedef vectorbase<msra::lattices::nodeinfo> nodeinfovector;
typedef vectorbase<msra::lattices::edgeinfowithscores> edgeinfowithscoresvector;
typedef vectorbase<msra::lattices::aligninfo> aligninfovector;
struct latticefunctions : public vectorbase<msra::lattices::empty>
{
virtual void edgealignment(const lrhmmdefvector& hmms, const lr3transPvector& transPs, const size_t spalignunitid,
const size_t silalignunitid, const Microsoft::MSR::CNTK::Matrix<float>& logLLs, const nodeinfovector& nodes,
const edgeinfowithscoresvector& edges, const aligninfovector& aligns,
const uintvector& alignoffsets, ushortvector& backptrstorage, const sizetvector& backptroffsets,
ushortvector& alignresult, floatvector& edgeacscores) = 0; // output
virtual void forwardbackwardlattice(const size_t* batchsizeforward, const size_t* batchsizebackward,
const size_t numlaunchforward, const size_t numlaunchbackward,
const size_t spalignunitid, const size_t silalignunitid,
const floatvector& edgeacscores, const edgeinfowithscoresvector& edges,
const nodeinfovector& nodes, const aligninfovector& aligns,
const ushortvector& alignoutput, const uintvector& alignoffsets,
doublevector& logpps, doublevector& logalphas, doublevector& logbetas,
const float lmf, const float wp, const float amf, const float boostingfactor, const bool returnEframescorrect,
const ushortvector& uids, const ushortvector& senone2classmap,
doublevector& logaccalphas, doublevector& logaccbetas,
doublevector& logframescorrectedge, doublevector& logEframescorrect,
doublevector& Eframescorrectbuf, double& logEframescorrecttotal, double& totalfwscore) = 0;
virtual void sMBRerrorsignal(const ushortvector& alignstateids, const uintvector& alignoffsets,
const edgeinfowithscoresvector& edges, const nodeinfovector& nodes,
const doublevector& logpps, const float amf, const doublevector& logEframescorrect,
const double logEframescorrecttotal, Microsoft::MSR::CNTK::Matrix<float>& dengammas, Microsoft::MSR::CNTK::Matrix<float>& dengammasbuf) = 0;
virtual void mmierrorsignal(const ushortvector& alignstateids, const uintvector& alignoffsets,
const edgeinfowithscoresvector& edges, const nodeinfovector& nodes,
const doublevector& logpps, Microsoft::MSR::CNTK::Matrix<float>& dengammas) = 0;
virtual void stateposteriors(const ushortvector& alignstateids, const uintvector& alignoffsets,
const edgeinfowithscoresvector& edges, const nodeinfovector& nodes,
const doublevector& logqs, Microsoft::MSR::CNTK::Matrix<float>& logacc) = 0;
};
// ---------------------------------------------------------------------------
// factor methods
// ---------------------------------------------------------------------------
somedatavector* newsomedatavector(size_t deviceid);
ushortvector* newushortvector(size_t deviceid);
uintvector* newuintvector(size_t deviceid);
floatvector* newfloatvector(size_t deviceid);
doublevector* newdoublevector(size_t deviceid);
sizetvector* newsizetvector(size_t deviceid);
latticefunctions* newlatticefunctions(size_t deviceid);
lrhmmdefvector* newlrhmmdefvector(size_t deviceid);
lr3transPvector* newlr3transPvector(size_t deviceid);
nodeinfovector* newnodeinfovector(size_t deviceid);
edgeinfowithscoresvector* newedgeinfovector(size_t deviceid);
aligninfovector* newaligninfovector(size_t deviceid);
};
};