bioem_cuda.cu

/* ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
        < BioEM software for Bayesian inference of Electron Microscopy images>
            Copyright (C) 2014 Pilar Cossio, David Rohr and Gerhard Hummer.
            Max Planck Institute of Biophysics, Frankfurt, Germany.

                See license statement for terms of distribution.

   ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/

#define BIOEM_GPUCODE

#if defined(_WIN32)
#include <windows.h>
#endif

#include <iostream>
using namespace std;

#include "bioem_cuda_internal.h"
#include "bioem_algorithm.h"

#define checkCudaErrors(error) \
{ \
	if ((error) != cudaSuccess) \
	{ \
		printf("CUDA Error %d / %s (%s: %d)\n", error, cudaGetErrorString(error), __FILE__, __LINE__); \
		exit(1); \
	} \
}

static const char *cufftGetErrorStrung(cufftResult error)
{
    switch (error)
    {
        case CUFFT_SUCCESS:
            return "CUFFT_SUCCESS";

        case CUFFT_INVALID_PLAN:
            return "CUFFT_INVALID_PLAN";

        case CUFFT_ALLOC_FAILED:
            return "CUFFT_ALLOC_FAILED";

        case CUFFT_INVALID_TYPE:
            return "CUFFT_INVALID_TYPE";

        case CUFFT_INVALID_VALUE:
            return "CUFFT_INVALID_VALUE";

        case CUFFT_INTERNAL_ERROR:
            return "CUFFT_INTERNAL_ERROR";

        case CUFFT_EXEC_FAILED:
            return "CUFFT_EXEC_FAILED";

        case CUFFT_SETUP_FAILED:
            return "CUFFT_SETUP_FAILED";

        case CUFFT_INVALID_SIZE:
            return "CUFFT_INVALID_SIZE";

        case CUFFT_UNALIGNED_DATA:
            return "CUFFT_UNALIGNED_DATA";
    }
    return "UNKNOWN";
}

bioem_cuda::bioem_cuda()
{
	deviceInitialized = 0;
	GPUAlgo = getenv("GPUALGO") == NULL ? 2 : atoi(getenv("GPUALGO"));
	GPUAsync = getenv("GPUASYNC") == NULL ? 1 : atoi(getenv("GPUASYNC"));
	GPUWorkload = getenv("GPUWORKLOAD") == NULL ? 100 : atoi(getenv("GPUWORKLOAD"));
	GPUDualStream = getenv("GPUDUALSTREAM") == NULL ? 1 : atoi(getenv("GPUDUALSTREAM"));
}

bioem_cuda::~bioem_cuda()
{
	deviceExit();
}

__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);
	}
}

__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);
	}
}

__global__ void cudaZeroMem(void* ptr, size_t size)
{
	int* myptr = (int*) ptr;
	int mysize = size / sizeof(int);
	int myid = myBlockDimX * myBlockIdxX + myThreadIdxX;
	int mygrid = myBlockDimX * myGridDimX;
	for (int i = myid; i < mysize; i += mygrid) myptr[i] = 0;
}

__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);
	const int myRef = myThreadIdxX >> (nShiftBits << 1);
	const int myShiftIdx = (myid >> nShiftBits) & (nShifts - 1);
	const int myShiftIdy = myid & (nShifts - 1);
	const int myShift = myid & (nShifts * nShifts - 1);
	const int cent_x = myShiftIdx * param.GridSpaceCenter - param.maxDisplaceCenter;
	const int cent_y = myShiftIdy * param.GridSpaceCenter - param.maxDisplaceCenter;

	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);
}

__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)
{
	if (myBlockIdxX >= NumberMaps) return;
	const mycuComplex_t* myin = (mycuComplex_t*) &refmap[(myBlockIdxX + Offset) * MapSize];
	const mycuComplex_t* myconv = (mycuComplex_t*) convmap;
	mycuComplex_t* myout = &out[myBlockIdxX * MapSize];
	for(int i = myThreadIdxX; i < NumberPixelsTotal; i += myBlockDimX)
	{
		mycuComplex_t val;
		const mycuComplex_t conv = myconv[i];
		const mycuComplex_t in = myin[i];

		val.x = conv.x * in.x + conv.y * in.y;
		val.y = conv.y * in.x - conv.x * in.y;
		myout[i] = val;
	}
}

__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)
{
	if (myBlockIdxX * myBlockDimX + myThreadIdxX >= maxRef) return;
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX + Offset;
	const myfloat_t* mylCC = &lCC[(myBlockIdxX * myBlockDimX + myThreadIdxX) * param.NumberPixels * param.NumberPixels];
	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);}
static inline bool IsPowerOf2(int x) {return ((x > 0) && ((x & (x - 1)) == 0));}
#if defined(_WIN32)
static inline int ilog2 (int value)
{
	DWORD index;
	_BitScanReverse (&index, value);
	return(value);
}
#else
static inline int ilog2(int value) {return 31 - __builtin_clz(value);}
#endif

int bioem_cuda::compareRefMaps(int iOrient, 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);
	}
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
	}

	if (FFTAlgo)
	{
		memcpy(&pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t));
		checkCudaErrors(cudaMemcpyAsync(&pConvMapFFT[(iConv & 1) * param.FFTMapSize], &pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice, cudaStream[GPUAsync ? 2 : 0]));
		if (GPUAsync)
		{
			checkCudaErrors(cudaEventRecord(cudaEvent[2], cudaStream[2]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaEvent[2], 0));
		}
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[0], cudaStream[0]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[1], cudaFFTEvent[0], 0));
		}
		for (int i = 0, j = 0; i < maxRef; i += CUDA_FFTS_AT_ONCE, j++)
		{
			if (!GPUDualStream) j = 0;
			const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
			multComplexMap<<<num, CUDA_THREAD_COUNT, 0, cudaStream[j & 1]>>>(&pConvMapFFT[(iConv & 1) * param.FFTMapSize], pRefMapsFFT, pFFTtmp2[j & 1], param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D, param.FFTMapSize, num, i);
			cufftResult err = mycufftExecC2R(i + CUDA_FFTS_AT_ONCE > maxRef ? plan[1][j & 1] : plan[0][j & 1], pFFTtmp2[j & 1], pFFTtmp[j & 1]);
			if (err != CUFFT_SUCCESS)
			{
				cout << "Error running CUFFT " << cufftGetErrorStrung(err) << "\n";
				exit(1);
			}
			cuDoRefMapsFFT<<<divup(num, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[j & 1]>>>(iOrient, iConv, pFFTtmp[j & 1], sumC, sumsquareC, pProb_device, param.param_device, *gpumap, num, i);
		}
		checkCudaErrors(cudaGetLastError());
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[1], cudaStream[1]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaFFTEvent[1], 0));
		}
	}
	else
	{
		checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream[0]));

		if (GPUAlgo == 2) //Loop over shifts
		{
			const int nShifts = 2 * param.param_device.maxDisplaceCenter / param.param_device.GridSpaceCenter + 1;
			if (!IsPowerOf2(nShifts))
			{
				cout << "Invalid number of displacements, no power of two\n";
				exit(1);
			}
			if (CUDA_THREAD_COUNT % (nShifts * nShifts))
			{
				cout << "CUDA Thread count (" << CUDA_THREAD_COUNT << ") is no multiple of number of shifts (" << (nShifts * nShifts) << ")\n";
				exit(1);
			}
			if (nShifts > CUDA_MAX_SHIFT_REDUCE)
			{
				cout << "Too many displacements for CUDA reduction\n";
				exit(1);
			}
			const int nShiftBits = ilog2(nShifts);
			size_t totalBlocks = divup((size_t) maxRef * (size_t) nShifts * (size_t) nShifts, (size_t) CUDA_THREAD_COUNT);
			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[0] >>> (iOrient, iConv, pConvMap_device[iConv & 1], pProb_device, param.param_device, *gpumap, i, nShifts, nShiftBits, maxRef);
			}
		}
		else if (GPUAlgo == 1) //Split shifts in multiple kernels
		{
			for (int cent_x = -param.param_device.maxDisplaceCenter; cent_x <= param.param_device.maxDisplaceCenter; cent_x = cent_x + param.param_device.GridSpaceCenter)
			{
				for (int cent_y = -param.param_device.maxDisplaceCenter; cent_y <= param.param_device.maxDisplaceCenter; cent_y = cent_y + param.param_device.GridSpaceCenter)
				{
					compareRefMap_kernel<<<divup(maxRef, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[0]>>> (iOrient, iConv, pConvMap_device[iConv & 1], pProb_device, param.param_device, *pRefMap_device_Mod, cent_x, cent_y, maxRef);
				}
			}
		}
		else if (GPUAlgo == 0) //All shifts in one kernel
		{
			compareRefMapShifted_kernel<<<divup(maxRef, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[0]>>> (iOrient, iConv, pConvMap_device[iConv & 1], pProb_device, param.param_device, *pRefMap_device_Mod, maxRef);
		}
		else
		{
			cout << "Invalid GPU Algorithm selected\n";
			exit(1);
		}
	}
	if (GPUWorkload < 100)
	{
		bioem::compareRefMaps(iOrient, iConv, conv_map, localmultFFT, sumC, sumsquareC, maxRef);
	}
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventRecord(cudaEvent[iConv & 1], cudaStream[0]));
	}
	else
	{
		checkCudaErrors(cudaStreamSynchronize(cudaStream[0]));
	}
	return(0);
}

int bioem_cuda::deviceInit()
{
	deviceExit();

	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));

		if (GPUAlgo == 0 || GPUAlgo == 1)
		{
			pRefMap_device_Mod = (bioem_RefMap_Mod*) gpumap;
			bioem_RefMap_Mod* RefMapGPU = new bioem_RefMap_Mod;
			RefMapGPU->init(RefMap);
			checkCudaErrors(cudaMemcpy(maps, RefMapGPU->maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize, cudaMemcpyHostToDevice));
			delete RefMapGPU;
		}
		else
		{
			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(cudaMalloc(&pProb_memory, pProb_device.get_size(RefMap.ntotRefMap, param.nTotGridAngles)));
	for (int i = 0; i < 2; i++)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[i]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
		checkCudaErrors(cudaEventCreate(&cudaFFTEvent[i]));
		checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
	}
	if (GPUAsync)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[2]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[2]));
	}

	if (FFTAlgo)
	{
		checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
		checkCudaErrors(cudaMalloc(&pFFTtmp2[0], CUDA_FFTS_AT_ONCE * param.FFTMapSize * 2 * sizeof(mycomplex_t)));
		checkCudaErrors(cudaMalloc(&pFFTtmp[0], CUDA_FFTS_AT_ONCE * param.param_device.NumberPixels * param.param_device.NumberPixels * 2 * sizeof(myfloat_t)));
		pFFTtmp2[1] = pFFTtmp2[0] + CUDA_FFTS_AT_ONCE * param.FFTMapSize;
		pFFTtmp[1] = pFFTtmp[0] + CUDA_FFTS_AT_ONCE * param.param_device.NumberPixels * param.param_device.NumberPixels;
		checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
		checkCudaErrors(cudaHostAlloc(&pConvMapFFT_Host, param.FFTMapSize * sizeof(mycomplex_t) * 2, 0));
		checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
	}

	deviceInitialized = 1;
	return(0);
}

int bioem_cuda::deviceExit()
{
	if (deviceInitialized == 0) return(0);


	cudaFree(pProb_memory);
	cudaFree(sum);
	cudaFree(sumsquare);
	for (int i = 0; i < 2; i++)
	{
		cudaStreamDestroy(cudaStream[i]);
		cudaEventDestroy(cudaEvent[i]);
		cudaEventDestroy(cudaFFTEvent[i]);
		cudaFree(pConvMap_device[i]);
	}
	if (FFTAlgo)
	{
		cudaFree(pRefMapsFFT);
		cudaFree(pConvMapFFT);
		cudaFreeHost(pConvMapFFT_Host);
		cudaFree(pFFTtmp[0]);
		cudaFree(pFFTtmp2[0]);
	}
	else
	{
		cudaFree(maps);
	}
	if (GPUAlgo == 0 || GPUAlgo == 1)
	{
		cudaFree(pRefMap_device_Mod);
	}
	if (GPUAsync)
	{
		cudaStreamDestroy(cudaStream[2]);
		cudaEventDestroy(cudaEvent[2]);
	}

	delete gpumap;
	cudaThreadExit();

	deviceInitialized = 0;
	return(0);
}

int bioem_cuda::deviceStartRun()
{
	if (GPUWorkload >= 100)
	{
		maxRef = RefMap.ntotRefMap;
		pProb_host = &pProb;
	}
	else
	{
		maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
		pProb_host = new bioem_Probability;
		pProb_host->init(maxRef, param.nTotGridAngles, *this);
		pProb_host->copyFrom(&pProb, *this);
	}

	pProb_device = *pProb_host;
	pProb_device.ptr = pProb_memory;
	pProb_device.set_pointers();
	checkCudaErrors(cudaMemcpyAsync(pProb_device.ptr, pProb_host->ptr, pProb_host->get_size(maxRef, param.nTotGridAngles), cudaMemcpyHostToDevice, cudaStream[0]));

	if (FFTAlgo)
	{
		for (int j = 0;j < 2;j++)
		{
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				int n[2] = {param.param_device.NumberPixels, param.param_device.NumberPixels};
				if (cufftPlanMany(&plan[i][j], 2, n, NULL, 1, param.FFTMapSize, NULL, 1, 0, MY_CUFFT_C2R, i ? (maxRef % CUDA_FFTS_AT_ONCE) : CUDA_FFTS_AT_ONCE) != CUFFT_SUCCESS)
				{
					cout << "Error planning CUFFT\n";
					exit(1);
				}
				if (cufftSetCompatibilityMode(plan[i][j], CUFFT_COMPATIBILITY_NATIVE) != CUFFT_SUCCESS)
				{
					cout << "Error planning CUFFT compatibility\n";
					exit(1);
				}
				if (cufftSetStream(plan[i][j], cudaStream[j]) != CUFFT_SUCCESS)
				{
					cout << "Error setting CUFFT stream\n";
					exit(1);
				}
			}
			if (!GPUDualStream) break;
		}
	}
	return(0);
}

int bioem_cuda::deviceFinishRun()
{
	if (GPUAsync) cudaStreamSynchronize(cudaStream[0]);
	checkCudaErrors(cudaMemcpyAsync(pProb_host->ptr, pProb_device.ptr, pProb_host->get_size(maxRef, param.nTotGridAngles), cudaMemcpyDeviceToHost, cudaStream[0]));

	if (FFTAlgo)
	{
		for (int j = 0;j < 2;j++)
		{
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				cufftDestroy(plan[i][j]);
			}
			if (!GPUDualStream) break;
		}
	}
	cudaThreadSynchronize();
	if (GPUWorkload < 100)
	{
		pProb.copyFrom(pProb_host, *this);
		delete[] pProb_host;
	}

	return(0);
}

void* bioem_cuda::malloc_device_host(size_t size)
{
	void* ptr;
	checkCudaErrors(cudaHostAlloc(&ptr, size, 0));
	return(ptr);
}

void bioem_cuda::free_device_host(void* ptr)
{
	cudaFreeHost(ptr);
}

bioem* bioem_cuda_create()
{
	return new bioem_cuda;
}