Skip to content

GitLab

B
BioEM

There is a maintenance of MPCDF Gitlab on Thursday, April 22st 2020, 9:00 am CEST - Expect some service interruptions during this time

Commit a30096fc authored by David Rohr's avatar David Rohr
Browse files
Options

start work to remove BIOEM_MAX_MAPS constant

parent 81fab4b2
Hide whitespace changes
Inline Side-by-side
Showing with 166 additions and 111 deletions
bioem.cpp
View file @ a30096fc
......@@ -57,8 +57,8 @@ int bioem::configure(int ac, char* av[])
std::string infile,modelfile,mapfile;
Model.readPDB=false;
param.writeAngles=false;
RefMap.dumpMap = false;
RefMap.loadMap = false;
param.dumpMap = false;
param.loadMap = false;
/*************************************************************************************/
cout << " ++++++++++++ FROM COMMAND LINE +++++++++++\n\n";
......@@ -93,12 +93,12 @@ int bioem::configure(int ac, char* av[])
if((ac < 6)) {
std::cout << desc << std::endl;
return 0;
return 1;
}
if (vm.count("help")) {
cout << "Usage: options_description [options]\n";
cout << desc;
return 0;
return 1;
}
if (vm.count("Inputfile"))
......@@ -123,13 +123,13 @@ int bioem::configure(int ac, char* av[])
if (vm.count("DumpMaps"))
{
cout << "Dumping Maps after reading from file.\n";
RefMap.dumpMap = true;
param.dumpMap = true;
}
if (vm.count("LoadMapDump"))
{
cout << "Loading Map dump.\n";
RefMap.loadMap = true;
param.loadMap = true;
}
if (vm.count("Particlesfile"))
......@@ -158,7 +158,7 @@ int bioem::configure(int ac, char* av[])
/********************* Reading Particle Maps Input **********************/
/********* HERE: PROBLEM if maps dont fit on the memory!! ***************/
// copying mapfile to ref map class
RefMap.filemap = mapfile.c_str();
param.filemap = mapfile.c_str();
RefMap.readRefMaps(param);
/****************** Precalculating Necessary Stuff *********************/
......@@ -189,7 +189,7 @@ int bioem::precalculate()
//Precalculating cross-correlations of maps
for (int iRefMap = 0; iRefMap < RefMap.ntotRefMap ; iRefMap++)
{
calcross_cor(RefMap.Ref[iRefMap],sum,sumsquare);
calcross_cor(RefMap.getmap(iRefMap),sum,sumsquare);
//Storing Crosscorrelations in Map class
RefMap.sum_RefMap[iRefMap]=sum;
RefMap.sumsquare_RefMap[iRefMap]=sumsquare;
......@@ -235,7 +235,7 @@ int bioem::run()
/*** Declaring Private variables for each thread *****/
mycomplex_t* proj_mapFFT;
bioem_map conv_map;
myfloat_t* conv_map = new myfloat_t[param.param_device.NumberPixels * param.param_device.NumberPixels];
mycomplex_t* conv_mapFFT;
myfloat_t sumCONV,sumsquareCONV;
......@@ -285,6 +285,7 @@ int bioem::run()
//deallocating fftw_complex vector
myfftw_free(proj_mapFFT);
myfftw_free(conv_mapFFT);
delete[] conv_map;
deviceFinishRun();
......@@ -348,15 +349,11 @@ int bioem::run()
param.refCTF =NULL;
}
if(RefMap.RefMapsFFT)
{
delete[] RefMap.RefMapsFFT;
RefMap.RefMapsFFT = NULL;
}
RefMap.freePointers();
return(0);
}
int bioem::compareRefMaps(int iProjectionOut, int iConv, const bioem_map& conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap)
int bioem::compareRefMaps(int iProjectionOut, int iConv, const myfloat_t* conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap)
{
if (FFTAlgo)
{
......@@ -394,7 +391,7 @@ int bioem::compareRefMaps(int iProjectionOut, int iConv, const bioem_map& conv_m
inline void bioem::calculateCCFFT(int iRefMap, int iOrient, int iConv, myfloat_t sumC,myfloat_t sumsquareC, mycomplex_t* localConvFFT,mycomplex_t* localCCT,myfloat_t* lCC)
{
const mycomplex_t* RefMapFFT = &RefMap.RefMapsFFT[iRefMap * param.RefMapSize];
const mycomplex_t* RefMapFFT = &RefMap.RefMapsFFT[iRefMap * param.FFTMapSize];
for(int i = 0;i < param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D;i++)
{
localCCT[i][0] = localConvFFT[i][0] * RefMapFFT[i][0] + localConvFFT[i][1] * RefMapFFT[i][1];
......@@ -493,7 +490,7 @@ int bioem::createProjection(int iMap,mycomplex_t* mapFFT)
return(0);
}
int bioem::createConvolutedProjectionMap(int iMap,int iConv,mycomplex_t* lproj,bioem_map& Mapconv, mycomplex_t* localmultFFT, myfloat_t& sumC, myfloat_t& sumsquareC)
int bioem::createConvolutedProjectionMap(int iMap,int iConv,mycomplex_t* lproj,myfloat_t* Mapconv, mycomplex_t* localmultFFT, myfloat_t& sumC, myfloat_t& sumsquareC)
{
/**************************************************************************************/
/**** BioEM Create Convoluted Projection Map routine, multiplies in Fourier **********
......@@ -508,7 +505,7 @@ int bioem::createConvolutedProjectionMap(int iMap,int iConv,mycomplex_t* lproj,b
/**** Multiplying FFTmap with corresponding kernel ****/
const mycomplex_t* refCTF = &param.refCTF[iConv * param.RefMapSize];
const mycomplex_t* refCTF = &param.refCTF[iConv * param.FFTMapSize];
for(int i=0;i < param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D;i++)
{
localmultFFT[i][0] = lproj[i][0] * refCTF[i][0] + lproj[i][1] * refCTF[i][1];
......@@ -527,7 +524,7 @@ int bioem::createConvolutedProjectionMap(int iMap,int iConv,mycomplex_t* lproj,b
{
for(int j=0; j < param.param_device.NumberPixels ; j++ )
{
Mapconv.points[i][j]=localconvFFT[i*param.param_device.NumberPixels+j];
Mapconv[i*param.param_device.NumberPixels+j] = localconvFFT[i*param.param_device.NumberPixels+j];
}
}
......@@ -553,7 +550,7 @@ int bioem::createConvolutedProjectionMap(int iMap,int iConv,mycomplex_t* lproj,b
return(0);
}
int bioem::calcross_cor(bioem_map& localmap,myfloat_t& sum,myfloat_t& sumsquare)
int bioem::calcross_cor(myfloat_t* localmap,myfloat_t& sum,myfloat_t& sumsquare)
{
/*********************** Routine to calculate Cross correlations***********************/
......@@ -564,9 +561,9 @@ int bioem::calcross_cor(bioem_map& localmap,myfloat_t& sum,myfloat_t& sumsquare)
for (int j = 0; j < param.param_device.NumberPixels; j++)
{
// Calculate Sum of pixels
sum += localmap.points[i][j];
sum += localmap[i*param.param_device.NumberPixels+j];
// Calculate Sum of pixels squared
sumsquare += localmap.points[i][j]*localmap.points[i][j];
sumsquare += localmap[i*param.param_device.NumberPixels+j] * localmap[i*param.param_device.NumberPixels+j];
}
}
return(0);
......
bioem_algorithm.h
View file @ a30096fc
......@@ -131,16 +131,16 @@ __device__ static inline void doRefMapFFT(const int iRefMap, const int iOrient,
}
template <int GPUAlgo, class RefT>
__device__ static inline void compareRefMap(const int iRefMap, const int iOrient, const int iConv, const bioem_map& Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap,
__device__ static inline void compareRefMap(const int iRefMap, const int iOrient, const int iConv, const myfloat_t* Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap,
const int cent_x, const int cent_y, const int myShift = 0, const int nShifts2 = 0, const int myRef = 0, const bool threadActive = true)
{
/**************************************************************************************/
/********************** Calculating BioEM Probability ********************************/
/************************* Loop of center displacement here ***************************/
// ********************** Calculating BioEM Probability ********************************
// ************************* Loop of center displacement here ***************************
// Taking into account the center displacement
/*** Inizialzing crosscorrelations of calculated projected convolutions ***/
// Inizialzing crosscorrelations of calculated projected convolutions
#ifdef SSECODE
myfloat_t sum, sumsquare, crossproMapConv;
__m128 sum_v = _mm_setzero_ps(), sumsquare_v = _mm_setzero_ps(), cross_v = _mm_setzero_ps(), d1, d2;
......@@ -149,7 +149,7 @@ __device__ static inline void compareRefMap(const int iRefMap, const int iOrient
myfloat_t sumsquare=0.0;
myfloat_t crossproMapConv=0.0;
#endif
/****** Loop over Pixels to calculate dot product and cross-correlations of displaced Ref Conv. Map***/
// Loop over Pixels to calculate dot product and cross-correlations of displaced Ref Conv. Map
myfloat_t logpro;
if (GPUAlgo != 2 || threadActive)
{
......@@ -195,7 +195,7 @@ __device__ static inline void compareRefMap(const int iRefMap, const int iOrient
#else
for (int j = jStart; j < jEnd; j += 1)
{
const myfloat_t pointMap = Mapconv.points[i + cent_x][j + cent_y];
const myfloat_t pointMap = Mapconv[(i + cent_x) * param.NumberPixels + j + cent_y];
const myfloat_t pointRefMap = RefMap.get(iRefMap, i, j);
crossproMapConv += pointMap * pointRefMap;
// Crosscorrelation of calculated displaced map
......@@ -217,7 +217,7 @@ __device__ static inline void compareRefMap(const int iRefMap, const int iOrient
crossproMapConv = _mm_cvtss_f32(cross_v);
#endif
/********** Calculating elements in BioEM Probability formula ********/
// Calculating elements in BioEM Probability formula
logpro = calc_logpro(param, sum, sumsquare, crossproMapConv, RefMap.sum_RefMap[iRefMap], RefMap.sumsquare_RefMap[iRefMap]);
}
else
......@@ -380,14 +380,14 @@ __device__ static inline void compareRefMap(const int iRefMap, const int iOrient
else
#endif
/***** Summing & Storing total/Orientation Probabilites for each map ************/
// Summing & Storing total/Orientation Probabilites for each map
{
update_prob<-1>(logpro, iRefMap, iOrient, iConv, cent_x, cent_y, pProb);
}
}
template <int GPUAlgo, class RefT>
__device__ static inline void compareRefMapShifted(const int iRefMap, const int iOrient, const int iConv, const bioem_map& Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap)
__device__ static inline void compareRefMapShifted(const int iRefMap, const int iOrient, const int iConv, const myfloat_t* Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap)
{
for (int cent_x = -param.maxDisplaceCenter; cent_x <= param.maxDisplaceCenter; cent_x=cent_x+param.GridSpaceCenter)
{
......
bioem_cuda.cu
View file @ a30096fc
......@@ -10,8 +10,6 @@ using namespace std;
#include "bioem_cuda_internal.h"
#include "bioem_algorithm.h"
//__constant__ bioem_map gConvMap;
#define checkCudaErrors(error) \
{ \
if ((error) != cudaSuccess) \
......@@ -34,21 +32,21 @@ bioem_cuda::~bioem_cuda()
deviceExit();
}
__global__ void compareRefMap_kernel(const int iOrient, const int iConv, const bioem_map* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap_Mod* RefMap, const int cent_x, const int cent_y, const int maxRef)
__global__ void compareRefMap_kernel(const int iOrient, const int iConv, const myfloat_t* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap_Mod* RefMap, const int cent_x, const int cent_y, const int maxRef)
{
const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
if (iRefMap < maxRef)
{
compareRefMap<0>(iRefMap,iOrient,iConv,*pMap, pProb, param, *RefMap, cent_x, cent_y);
compareRefMap<0>(iRefMap,iOrient,iConv, pMap, pProb, param, *RefMap, cent_x, cent_y);
}
}
__global__ void compareRefMapShifted_kernel(const int iOrient, const int iConv, const bioem_map* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap_Mod* RefMap, const int maxRef)
__global__ void compareRefMapShifted_kernel(const int iOrient, const int iConv, const myfloat_t* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap_Mod* RefMap, const int maxRef)
{
const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
if (iRefMap < maxRef)
{
compareRefMapShifted<1>(iRefMap,iOrient,iConv,*pMap, pProb, param, *RefMap);
compareRefMapShifted<1>(iRefMap,iOrient,iConv, pMap, pProb, param, *RefMap);
}
}
......@@ -61,7 +59,7 @@ __global__ void cudaZeroMem(void* ptr, size_t size)
for (int i = myid;i < mysize;i += mygrid) myptr[i] = 0;
}
__global__ void compareRefMapLoopShifts_kernel(const int iOrient, const int iConv, const bioem_map* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap* RefMap, const int blockoffset, const int nShifts, const int nShiftBits, const int maxRef)
__global__ void compareRefMapLoopShifts_kernel(const int iOrient, const int iConv, const myfloat_t* pMap, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap RefMap, const int blockoffset, const int nShifts, const int nShiftBits, const int maxRef)
{
const size_t myid = (myBlockIdxX + blockoffset) * myBlockDimX + myThreadIdxX;
const int iRefMap = myid >> (nShiftBits << 1);
......@@ -74,7 +72,7 @@ __global__ void compareRefMapLoopShifts_kernel(const int iOrient, const int iCon
const bool threadActive = myShiftIdx < nShifts && myShiftIdy < nShifts && iRefMap < maxRef;
compareRefMap<2>(iRefMap,iOrient,iConv,*pMap, pProb, param, *RefMap, cent_x, cent_y, myShift, nShifts * nShifts, myRef, threadActive);
compareRefMap<2>(iRefMap,iOrient,iConv, pMap, pProb, param, RefMap, cent_x, cent_y, myShift, nShifts * nShifts, myRef, threadActive);
}
__global__ void multComplexMap(const mycomplex_t* convmap, const mycomplex_t* refmap, mycuComplex_t* out, const int NumberPixelsTotal, const int MapSize, const int NumberMaps, const int Offset)
......@@ -89,12 +87,12 @@ __global__ void multComplexMap(const mycomplex_t* convmap, const mycomplex_t* re
}
}
__global__ void cuDoRefMapsFFT(const int iOrient, const int iConv, const myfloat_t* lCC, const myfloat_t sumC, const myfloat_t sumsquareC, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap* RefMap, const int maxRef, const int Offset)
__global__ void cuDoRefMapsFFT(const int iOrient, const int iConv, const myfloat_t* lCC, const myfloat_t sumC, const myfloat_t sumsquareC, bioem_Probability* pProb, const bioem_param_device param, const bioem_RefMap RefMap, const int maxRef, const int Offset)
{
const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX + Offset;
const myfloat_t* mylCC = &lCC[(myBlockIdxX * myBlockDimX + myThreadIdxX) * param.NumberPixels * param.NumberPixels];
if (iRefMap >= maxRef) return;
doRefMapFFT(iRefMap, iOrient, iConv, mylCC, sumC, sumsquareC, pProb, param, *RefMap);
doRefMapFFT(iRefMap, iOrient, iConv, mylCC, sumC, sumsquareC, pProb, param, RefMap);
}
template <class T> static inline T divup(T num, T divider) {return((num + divider - 1) / divider);}
......@@ -110,14 +108,13 @@ static inline int ilog2 (int value)
static inline int ilog2(int value) {return 31 - __builtin_clz(value);}
#endif
int bioem_cuda::compareRefMaps(int iProjectionOut, int iConv, const bioem_map& conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap)
int bioem_cuda::compareRefMaps(int iProjectionOut, int iConv, const myfloat_t* conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap)
{
if (startMap)
{
cout << "Error startMap not implemented for GPU Code\n";
exit(1);
}
//cudaMemcpyToSymbolAsync(&gConvMap, &conv_map, sizeof(bioem_map), 0, cudaMemcpyHostToDevice, cudaStream);
if (GPUAsync)
{
checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
......@@ -125,23 +122,23 @@ int bioem_cuda::compareRefMaps(int iProjectionOut, int iConv, const bioem_map& c
if (FFTAlgo)
{
checkCudaErrors(cudaMemcpyAsync(&pConvMapFFT[(iConv & 1) * param.RefMapSize], localmultFFT, param.RefMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice, cudaStream));
checkCudaErrors(cudaMemcpyAsync(&pConvMapFFT[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice, cudaStream));
for (int i = 0;i < maxRef;i += CUDA_FFTS_AT_ONCE)
{
const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
multComplexMap<<<num, CUDA_THREAD_COUNT, 0, cudaStream>>>(&pConvMapFFT[(iConv & 1) * param.RefMapSize], pRefMapsFFT, pFFTtmp2, param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D, param.RefMapSize, num, i);
multComplexMap<<<num, CUDA_THREAD_COUNT, 0, cudaStream>>>(&pConvMapFFT[(iConv & 1) * param.FFTMapSize], pRefMapsFFT, pFFTtmp2, param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D, param.FFTMapSize, num, i);
if (mycufftExecC2R(i + CUDA_FFTS_AT_ONCE > maxRef ? plan[1] : plan[0], pFFTtmp2, pFFTtmp) != CUFFT_SUCCESS)
{
cout << "Error running CUFFT\n";
exit(1);
}
cuDoRefMapsFFT<<<divup(num, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream>>>(iProjectionOut, iConv, pFFTtmp, sumC, sumsquareC, pProb_device, param.param_device, pRefMap_device, num, i);
cuDoRefMapsFFT<<<divup(num, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream>>>(iProjectionOut, iConv, pFFTtmp, sumC, sumsquareC, pProb_device, param.param_device, *gpumap, num, i);
}
checkCudaErrors(cudaGetLastError());
}
else
{
checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], &conv_map, sizeof(bioem_map), cudaMemcpyHostToDevice, cudaStream));
checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream));
if (GPUAlgo == 2) //Loop over shifts
{
......@@ -166,7 +163,7 @@ int bioem_cuda::compareRefMaps(int iProjectionOut, int iConv, const bioem_map& c
size_t nBlocks = CUDA_BLOCK_COUNT;
for (size_t i = 0;i < totalBlocks;i += nBlocks)
{
compareRefMapLoopShifts_kernel <<<min(nBlocks, totalBlocks - i), CUDA_THREAD_COUNT, (CUDA_THREAD_COUNT * 2 + CUDA_THREAD_COUNT / (nShifts * nShifts) * 4) * sizeof(myfloat_t), cudaStream>>> (iProjectionOut, iConv, pConvMap_device[iConv & 1], pProb_device, param.param_device, pRefMap_device, i, nShifts, nShiftBits, maxRef);
compareRefMapLoopShifts_kernel <<<min(nBlocks, totalBlocks - i), CUDA_THREAD_COUNT, (CUDA_THREAD_COUNT * 2 + CUDA_THREAD_COUNT / (nShifts * nShifts) * 4) * sizeof(myfloat_t), cudaStream>>> (iProjectionOut, iConv, pConvMap_device[iConv & 1], pProb_device, param.param_device, *gpumap, i, nShifts, nShiftBits, maxRef);
}
}
else if (GPUAlgo == 1) //Split shifts in multiple kernels
......@@ -210,40 +207,45 @@ int bioem_cuda::deviceInit()
if (FFTAlgo) GPUAlgo = 2;
gpumap = new bioem_RefMap;
memcpy(gpumap, &RefMap, sizeof(bioem_RefMap));
if (FFTAlgo == 0)
{
checkCudaErrors(cudaMalloc(&maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize));
checkCudaErrors(cudaMemcpy(maps, RefMap.maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize, cudaMemcpyHostToDevice));
}
checkCudaErrors(cudaMalloc(&sum, sizeof(myfloat_t) * RefMap.ntotRefMap));
checkCudaErrors(cudaMemcpy(sum, RefMap.sum_RefMap, sizeof(myfloat_t) * RefMap.ntotRefMap, cudaMemcpyHostToDevice));
checkCudaErrors(cudaMalloc(&sumsquare, sizeof(myfloat_t) * RefMap.ntotRefMap));
checkCudaErrors(cudaMemcpy(sumsquare, RefMap.sumsquare_RefMap, sizeof(myfloat_t) * RefMap.ntotRefMap, cudaMemcpyHostToDevice));
gpumap->maps = maps;
gpumap->sum_RefMap = sum;
gpumap->sumsquare_RefMap = sumsquare;
checkCudaErrors(cudaStreamCreate(&cudaStream));
cout << "\tSize RefMap\t\t" << sizeof(bioem_RefMap) << "\n";
checkCudaErrors(cudaMalloc(&pRefMap_device, sizeof(bioem_RefMap)));
cout << "\tSize Probability\t" << sizeof(bioem_Probability) * RefMap.ntotRefMap << "\n";
checkCudaErrors(cudaMalloc(&pProb_device, sizeof(bioem_Probability) * RefMap.ntotRefMap));
for (int i = 0;i < 2;i++)
{
checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(bioem_map)));
checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
}
pRefMap_device_Mod = (bioem_RefMap_Mod*) pRefMap_device;
if (FFTAlgo)
{
cout << "\tSize RefMapFFT\t\t" << RefMap.ntotRefMap * param.RefMapSize * sizeof(mycomplex_t) << "\n";
checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.RefMapSize * sizeof(mycomplex_t)));
cout << "\tSize RefMapFFT Copy\t" << CUDA_FFTS_AT_ONCE * param.RefMapSize * sizeof(mycomplex_t) << "\n";
checkCudaErrors(cudaMalloc(&pFFTtmp2, CUDA_FFTS_AT_ONCE * param.RefMapSize * sizeof(mycomplex_t)));
cout << "\tSize RefMapFFT Read\t" << CUDA_FFTS_AT_ONCE * param.param_device.NumberPixels * param.param_device.NumberPixels * sizeof(myfloat_t) << "\n";
checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
checkCudaErrors(cudaMalloc(&pFFTtmp2, CUDA_FFTS_AT_ONCE * param.FFTMapSize * sizeof(mycomplex_t)));
checkCudaErrors(cudaMalloc(&pFFTtmp, CUDA_FFTS_AT_ONCE * param.param_device.NumberPixels * param.param_device.NumberPixels * sizeof(myfloat_t)));
checkCudaErrors(cudaMalloc(&pConvMapFFT, param.RefMapSize * sizeof(mycomplex_t) * 2));
cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.RefMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice);
checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
}
if (GPUAlgo == 0 || GPUAlgo == 1)
{
checkCudaErrors(cudaMalloc(&pRefMap_device_Mod, sizeof(bioem_RefMap_Mod)));
bioem_RefMap_Mod* RefMapGPU = new bioem_RefMap_Mod(RefMap);
cudaMemcpy(pRefMap_device_Mod, RefMapGPU, sizeof(bioem_RefMap), cudaMemcpyHostToDevice);
delete RefMapGPU;
}
else
{
cudaMemcpy(pRefMap_device, &RefMap, sizeof(bioem_RefMap), cudaMemcpyHostToDevice);
}
deviceInitialized = 1;
return(0);
......@@ -254,12 +256,13 @@ int bioem_cuda::deviceExit()
if (deviceInitialized == 0) return(0);
cudaStreamDestroy(cudaStream);
cudaFree(pRefMap_device);
cudaFree(pProb_device);
cudaFree(sum);
cudaFree(sumsquare);
for (int i = 0;i < 2;i++)
{
cudaEventDestroy(cudaEvent[i]);
cudaFree(pConvMap_device);
cudaFree(pConvMap_device[i]);
}
if (FFTAlgo)
{
......@@ -268,6 +271,15 @@ int bioem_cuda::deviceExit()
//cudaFree(pFFTtmp);
cudaFree(pFFTtmp2);
}
else
{
cudaFree(maps);
}
if (GPUAlgo == 0 || GPUAlgo == 1)
{
cudaFree(pRefMap_device_Mod);
}
delete gpumap;
cudaThreadExit();
deviceInitialized = 0;
......
include/bioem.h
View file @ a30096fc
......@@ -21,12 +21,12 @@ public:
int dopreCalCrossCorrelation(int iRefMap, int iRefMapLocal);
int run();
int doProjections(int iMap);
int createConvolutedProjectionMap(int iOreint,int iMap, mycomplex_t* lproj,bioem_map& Mapconv,mycomplex_t* localmultFFT,myfloat_t& sumC,myfloat_t& sumsquareC);
int createConvolutedProjectionMap(int iOreint,int iMap, mycomplex_t* lproj, myfloat_t* Mapconv,mycomplex_t* localmultFFT, myfloat_t& sumC, myfloat_t& sumsquareC);
virtual int compareRefMaps(int iProjectionOut, int iConv, const bioem_map& conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap = 0);
virtual int compareRefMaps(int iProjectionOut, int iConv, const myfloat_t* conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap = 0);
int createProjection(int iMap, mycomplex_t* map);
int calcross_cor(bioem_map& localmap,myfloat_t& sum,myfloat_t& sumsquare);
int calcross_cor(myfloat_t* localmap, myfloat_t& sum, myfloat_t& sumsquare);
void calculateCCFFT(int iMap, int iOrient, int iConv, myfloat_t sumC, myfloat_t sumsquareC, mycomplex_t* localConvFFT,mycomplex_t* localCCT,myfloat_t* lCC);
bioem_Probability* pProb;
......
include/bioem_cuda_internal.h
View file @ a30096fc
......@@ -17,7 +17,7 @@ public:
bioem_cuda();
virtual ~bioem_cuda();
virtual int compareRefMaps(int iProjectionOut, int iConv, const bioem_map& conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap = 0);
virtual int compareRefMaps(int iProjectionOut, int iConv, const myfloat_t* conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap = 0);
protected:
virtual int deviceInit();
......@@ -30,9 +30,9 @@ protected:
cudaStream_t cudaStream;
cudaEvent_t cudaEvent[2];
bioem_RefMap_Mod* pRefMap_device_Mod;
bioem_RefMap* pRefMap_device;
bioem_RefMap* gpumap;
bioem_Probability* pProb_device;
bioem_map* pConvMap_device[2];
myfloat_t* pConvMap_device[2];
mycomplex_t* pRefMapsFFT;
mycomplex_t* pConvMapFFT;
......@@ -40,6 +40,8 @@ protected:
myfloat_t* pFFTtmp;
cufftHandle plan[2];
myfloat_t *maps, *sum, *sumsquare;
int GPUAlgo; //GPU Algorithm to use, 0: parallelize over maps, 1: as 0 but work split in multiple kernels (better), 2: also parallelize over shifts (best)
int GPUAsync; //Run GPU Asynchronously, do the convolutions on the host in parallel.
int GPUWorkload; //Percentage of workload to perform on GPU. Default 100. Rest is done on processor in parallel.
......
include/defs.h
View file @ a30096fc
......@@ -72,4 +72,26 @@ struct myfloat3_t
#define CUDA_MAX_SHIFT_REDUCE 1024
#define CUDA_FFTS_AT_ONCE 1024
static inline void* mallocchk(size_t size)
{
void* ptr = malloc(size);
if (ptr == 0)
{
std::cout << "Memory allocation error\n";
exit(1);
}
return(ptr);
}
static inline void* reallocchk(void* oldptr, size_t size)
{
void* ptr = realloc(oldptr, size);
if (ptr == 0)
{
std::cout << "Memory allocation error\n";
exit(1);
}
return(ptr);
}
#endif
include/map.h
View file @ a30096fc
......@@ -7,17 +7,26 @@
class bioem_param;
class bioem_map
{
public:
myfloat_t points[BIOEM_MAP_SIZE_X][BIOEM_MAP_SIZE_Y];
};
class bioem_RefMap
{
public:
bioem_RefMap()
{
maps = NULL;
RefMapsFFT = NULL;
sum_RefMap = NULL;
sumsquare_RefMap = NULL;
}
void freePointers()
{
if (maps) free(maps);
if (sum_RefMap) free(sum_RefMap);
if (sumsquare_RefMap) free(sumsquare_RefMap);
if (RefMapsFFT) delete[] RefMapsFFT;
maps = NULL;
sum_RefMap = NULL;
sumsquare_RefMap = NULL;
RefMapsFFT = NULL;
}
int readRefMaps(bioem_param& param);
......@@ -25,29 +34,25 @@ public:
mycomplex_t* RefMapsFFT;
const char* filemap;
int ntotRefMap;
bioem_map Ref[BIOEM_MAX_MAPS];
myfloat_t sum_RefMap[BIOEM_MAX_MAPS];
myfloat_t sumsquare_RefMap[BIOEM_MAX_MAPS];
myfloat_t ForLogProbfromRef[BIOEM_MAX_MAPS];
int numPixels;
int refMapSize;
myfloat_t* maps;
myfloat_t* sum_RefMap;
myfloat_t* sumsquare_RefMap;
bool dumpMap, loadMap;
__host__ __device__ inline myfloat_t get(int map, int x, int y) const {return(Ref[map].points[x][y]);}
__host__ __device__ inline const myfloat_t* getp(int map, int x, int y) const {return(&Ref[map].points[x][y]);}
__host__ __device__ inline myfloat_t get(int map, int x, int y) const {return(maps[map * refMapSize + x * numPixels + y]);}
__host__ __device__ inline const myfloat_t* getp(int map, int x, int y) const {return(&maps[map * refMapSize + x * numPixels]);}
__host__ __device__ inline myfloat_t* getmap(int map) {return(&maps[map * refMapSize]);}
};
class bioem_RefMap_Mod
{
public:
const char* filemap;
int ntotRefMap;
myfloat_t Ref[BIOEM_MAP_SIZE_X][BIOEM_MAP_SIZE_Y][BIOEM_MAX_MAPS];
myfloat_t sum_RefMap[BIOEM_MAX_MAPS];
myfloat_t sumsquare_RefMap[BIOEM_MAX_MAPS];
myfloat_t ForLogProbfromRef[BIOEM_MAX_MAPS];
__host__ __device__ inline myfloat_t get(int map, int x, int y) const {return(Ref[x][y][map]);}
......@@ -58,7 +63,6 @@ public:
ntotRefMap = map.ntotRefMap;
memcpy(sum_RefMap, map.sum_RefMap, sizeof(sum_RefMap));
memcpy(sumsquare_RefMap, map.sumsquare_RefMap, sizeof(sumsquare_RefMap));
memcpy(ForLogProbfromRef, map.ForLogProbfromRef, sizeof(ForLogProbfromRef));
#pragma omp parallel for
for (int i = 0;i < ntotRefMap;i++)
{
......@@ -83,4 +87,5 @@ public:
myfloat_t max_prob;
int max_prob_cent_x, max_prob_cent_y, max_prob_orient, max_prob_conv;
};
#endif
include/param.h
View file @ a30096fc
......@@ -32,7 +32,7 @@ public:
bioem_param_device param_device;
int RefMapSize;
int FFTMapSize;
mycomplex_t* refCTF;
myfloat3_t* CtfParam;
......@@ -73,6 +73,9 @@ public:
int fft_plans_created;
myfftw_plan fft_plan_c2c_forward, fft_plan_c2c_backward, fft_plan_r2c_forward, fft_plan_c2r_backward;
const char* filemap;
bool dumpMap, loadMap;
private:
void releaseFFTPlans();