bioem_cuda.cu 20.5 KB
Newer Older
Pilar Cossio's avatar
License  
Pilar Cossio committed
1
/* ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
2 3 4 5
   < BioEM software for Bayesian inference of Electron Microscopy images>
   Copyright (C) 2016 Pilar Cossio, David Rohr, Fabio Baruffa, Markus Rampp, 
        Volker Lindenstruth and Gerhard Hummer.
   Max Planck Institute of Biophysics, Frankfurt, Germany.
6

7
   See license statement for terms of distribution.
Pilar Cossio's avatar
License  
Pilar Cossio committed
8 9 10

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

11 12 13 14 15 16 17 18 19 20 21
#define BIOEM_GPUCODE

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

#include <iostream>
using namespace std;

#include "bioem_cuda_internal.h"
#include "bioem_algorithm.h"
Pilar Cossio's avatar
Pilar Cossio committed
22
//#include "helper_cuda.h"
23

24 25 26 27 28 29 30
#define checkCudaErrors(error) \
{ \
	if ((error) != cudaSuccess) \
	{ \
		printf("CUDA Error %d / %s (%s: %d)\n", error, cudaGetErrorString(error), __FILE__, __LINE__); \
		exit(1); \
	} \
31 32
}

David Rohr's avatar
David Rohr committed
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69
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";
}

70 71 72 73
bioem_cuda::bioem_cuda()
{
	deviceInitialized = 0;
	GPUAlgo = getenv("GPUALGO") == NULL ? 2 : atoi(getenv("GPUALGO"));
74 75
	GPUAsync = getenv("GPUASYNC") == NULL ? 1 : atoi(getenv("GPUASYNC"));
	GPUWorkload = getenv("GPUWORKLOAD") == NULL ? 100 : atoi(getenv("GPUWORKLOAD"));
76
	GPUDualStream = getenv("GPUDUALSTREAM") == NULL ? 1 : atoi(getenv("GPUDUALSTREAM"));
77 78 79 80 81 82 83
}

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

84 85 86
__global__ void compareRefMap_kernel(const int iOrient, const int iConv,  const myfloat_t amp, const myfloat_t pha, const myfloat_t env, const myfloat_t sumC,
                                                const myfloat_t sumsquareC, 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)
87 88
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
89
	if (iRefMap < maxRef)
90
	{
91
		compareRefMap<0>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap, cent_x, cent_y);
92 93 94
	}
}

Pilar Cossio's avatar
Pilar Cossio committed
95
__global__ void compareRefMapShifted_kernel(const int iOrient, const int iConv, const myfloat_t amp, const myfloat_t pha, const myfloat_t env, const myfloat_t sumC, const myfloat_t sumsquareC, const myfloat_t* pMap, bioem_Probability pProb, const bioem_param_device param, const bioem_RefMap_Mod RefMap, const int maxRef)
96 97
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
98
	if (iRefMap < maxRef)
99
	{
100
		compareRefMapShifted<1>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap);
101 102 103
	}
}

104 105 106 107 108 109
__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;
110
	for (int i = myid; i < mysize; i += mygrid) myptr[i] = 0;
111 112
}

Pilar Cossio's avatar
Pilar Cossio committed
113
__global__ void compareRefMapLoopShifts_kernel(const int iOrient, const int iConv, const myfloat_t amp, const myfloat_t pha, const myfloat_t env, const myfloat_t sumC, const myfloat_t sumsquareC, 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)
114 115 116 117 118 119 120 121 122
{
	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;
123

124
	const bool threadActive = myShiftIdx < nShifts && myShiftIdy < nShifts && iRefMap < maxRef;
125

Pilar Cossio's avatar
Pilar Cossio committed
126
	compareRefMap<2>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap, cent_x, cent_y, myShift, nShifts * nShifts, myRef, threadActive);
127 128
}

129
__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)
130 131
{
	if (myBlockIdxX >= NumberMaps) return;
132
	const mycuComplex_t* myin = (mycuComplex_t*) &refmap[(myBlockIdxX + Offset) * MapSize];
133
	const mycuComplex_t* myconv = (mycuComplex_t*) convmap;
134
	mycuComplex_t* myout = &out[myBlockIdxX * MapSize];
135
	for(int i = myThreadIdxX; i < NumberPixelsTotal; i += myBlockDimX)
136
	{
137 138 139 140 141 142 143
		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;
144 145 146
	}
}

147
__global__ void cuDoRefMapsFFT(const int iOrient, const int iConv, const myfloat_t amp, const myfloat_t pha, const myfloat_t env, 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)
148
{
149
	if (myBlockIdxX * myBlockDimX + myThreadIdxX >= maxRef) return;
150 151
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX + Offset;
	const myfloat_t* mylCC = &lCC[(myBlockIdxX * myBlockDimX + myThreadIdxX) * param.NumberPixels * param.NumberPixels];
152
	doRefMapFFT(iRefMap, iOrient, iConv, amp, pha, env, mylCC, sumC, sumsquareC, pProb, param, RefMap);
153 154
}

155 156 157 158 159 160 161 162 163 164 165 166 167
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

168
int bioem_cuda::compareRefMaps(int iOrient, int iConv, myfloat_t amp, myfloat_t pha, myfloat_t env, const myfloat_t* conv_map, mycomplex_t* localmultFFT, myfloat_t sumC, myfloat_t sumsquareC, const int startMap)
169
{
170 171 172 173 174
	if (startMap)
	{
		cout << "Error startMap not implemented for GPU Code\n";
		exit(1);
	}
175 176 177 178
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
	}
179

180
	if (FFTAlgo)
181
	{
182
		memcpy(&pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t));
183 184 185 186 187 188
		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));
		}
189
		if (GPUDualStream)
190
		{
191 192 193 194 195 196
			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;
197
			const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
198 199
			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]);
David Rohr's avatar
David Rohr committed
200
			if (err != CUFFT_SUCCESS)
201
			{
David Rohr's avatar
David Rohr committed
202
				cout << "Error running CUFFT " << cufftGetErrorStrung(err) << "\n";
203 204
				exit(1);
			}
205
			cuDoRefMapsFFT<<<divup(num, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[j & 1]>>>(iOrient, iConv,  amp, pha, env, pFFTtmp[j & 1], sumC, sumsquareC, pProb_device, param.param_device, *gpumap, num, i);
206
		}
207
		checkCudaErrors(cudaGetLastError());
208 209 210 211 212
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[1], cudaStream[1]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaFFTEvent[1], 0));
		}
213 214 215
	}
	else
	{
216
		checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream[0]));
217 218

		if (GPUAlgo == 2) //Loop over shifts
219
		{
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238
			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;
239
			for (size_t i = 0; i < totalBlocks; i += nBlocks)
240
			{
Pilar Cossio's avatar
Pilar Cossio committed
241
				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, amp, pha, env, sumC, sumsquareC, pConvMap_device[iConv & 1], pProb_device, param.param_device, *gpumap, i, nShifts, nShiftBits, maxRef);
242
			}
243
		}
244
		else if (GPUAlgo == 1) //Split shifts in multiple kernels
245
		{
246
			for (int cent_x = -param.param_device.maxDisplaceCenter; cent_x <= param.param_device.maxDisplaceCenter; cent_x = cent_x + param.param_device.GridSpaceCenter)
247
			{
248
				for (int cent_y = -param.param_device.maxDisplaceCenter; cent_y <= param.param_device.maxDisplaceCenter; cent_y = cent_y + param.param_device.GridSpaceCenter)
249
				{
Pilar Cossio's avatar
Pilar Cossio committed
250
					compareRefMap_kernel<<<divup(maxRef, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[0]>>> (iOrient, iConv, amp, pha, env, sumC, sumsquareC, pConvMap_device[iConv & 1], pProb_device, param.param_device, *pRefMap_device_Mod, cent_x, cent_y, maxRef);
251 252
				}
			}
253
		}
254
		else if (GPUAlgo == 0) //All shifts in one kernel
255
		{
256
			compareRefMapShifted_kernel<<<divup(maxRef, CUDA_THREAD_COUNT), CUDA_THREAD_COUNT, 0, cudaStream[0]>>> (iOrient, iConv, amp, pha, env, sumC, sumsquareC, pConvMap_device[iConv & 1], pProb_device, param.param_device, *pRefMap_device_Mod, maxRef);
257
		}
258
		else
259
		{
260 261
			cout << "Invalid GPU Algorithm selected\n";
			exit(1);
262
		}
263
	}
264 265
	if (GPUWorkload < 100)
	{
266
		bioem::compareRefMaps(iOrient, iConv, amp, pha, env, conv_map, localmultFFT, sumC, sumsquareC, maxRef);
267
	}
268 269
	if (GPUAsync)
	{
270
		checkCudaErrors(cudaEventRecord(cudaEvent[iConv & 1], cudaStream[0]));
271
	}
272 273
	else
	{
274
		checkCudaErrors(cudaStreamSynchronize(cudaStream[0]));
275 276 277 278
	}
	return(0);
}

David Rohr's avatar
David Rohr committed
279 280 281 282 283 284 285 286 287 288 289 290 291 292
int bioem_cuda::selectCudaDevice()
{
	int count;
	
	long long int bestDeviceSpeed = -1;
	int bestDevice;
	cudaDeviceProp deviceProp;
	
	checkCudaErrors(cudaGetDeviceCount(&count));
	if (count == 0)
	{
		printf("No CUDA device detected\n");
		return(1);
	}
293 294 295 296 297
	/* The following code, doing search for a fastest GPU, 
	   is temporarily disabled since it causes initialization 
	   errors on dvl machine. It is safe to ignore warnings
	   for "bestDeviceSpeed" */
#if 0	
David Rohr's avatar
David Rohr committed
298 299 300 301 302 303 304
	for (int i = 0;i < count;i++)
	{
#if CUDA_VERSION > 3010
		size_t free, total;
#else
		unsigned int free, total;
#endif
Luka Stanisic's avatar
Luka Stanisic committed
305
		CU_ERROR_CHECK(cuInit(0));
David Rohr's avatar
David Rohr committed
306
		CUdevice tmpDevice;
Luka Stanisic's avatar
Luka Stanisic committed
307
		CU_ERROR_CHECK(cuDeviceGet(&tmpDevice, i));
David Rohr's avatar
David Rohr committed
308
		CUcontext tmpContext;
Luka Stanisic's avatar
Luka Stanisic committed
309
		CU_ERROR_CHECK(cuCtxCreate(&tmpContext, 0, tmpDevice));
David Rohr's avatar
David Rohr committed
310
		if(cuMemGetInfo(&free, &total)) exit(1);
Luka Stanisic's avatar
Luka Stanisic committed
311
		CU_ERROR_CHECK(cuCtxDestroy(tmpContext));
David Rohr's avatar
David Rohr committed
312 313
		checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));

David Rohr's avatar
David Rohr committed
314
		if (DebugOutput >= 2 && mpi_rank == 0) printf("CUDA Device %2d: %s (Rev: %d.%d - Mem Avail %lld / %lld)\n", i, deviceProp.name, deviceProp.major, deviceProp.minor, (long long int) free, (long long int) deviceProp.totalGlobalMem);
David Rohr's avatar
David Rohr committed
315 316 317 318 319 320 321
		long long int deviceSpeed = (long long int) deviceProp.multiProcessorCount * (long long int) deviceProp.clockRate * (long long int) deviceProp.warpSize;
		if (deviceSpeed > bestDeviceSpeed)
		{
			bestDevice = i;
			bestDeviceSpeed = deviceSpeed;
		}
	}
322
#endif	
323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
	if (getenv("GPUDEVICE"))
	{
		int device = atoi(getenv("GPUDEVICE"));
		if (device > count)
		{
			printf("Invalid CUDA device specified, max device number is %d\n", count);
			exit(1);
		}
#ifdef WITH_MPI
		if (device == -1)
		{
			device = mpi_rank % count;
		}
#endif
		if (device < 0)
		{
			printf("Negative CUDA device specified: %d, invalid!\n", device);
		}
		bestDevice = device;
	}
	checkCudaErrors(cudaSetDevice(bestDevice));
David Rohr's avatar
David Rohr committed
344 345 346

	cudaGetDeviceProperties(&deviceProp ,bestDevice); 

David Rohr's avatar
David Rohr committed
347
	if (DebugOutput >= 3)
David Rohr's avatar
David Rohr committed
348
	{
David Rohr's avatar
David Rohr committed
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364
		printf("Using CUDA Device %s with Properties:\n", deviceProp.name);
		printf("totalGlobalMem = %lld\n", (unsigned long long int) deviceProp.totalGlobalMem);
		printf("sharedMemPerBlock = %lld\n", (unsigned long long int) deviceProp.sharedMemPerBlock);
		printf("regsPerBlock = %d\n", deviceProp.regsPerBlock);
		printf("warpSize = %d\n", deviceProp.warpSize);
		printf("memPitch = %lld\n", (unsigned long long int) deviceProp.memPitch);
		printf("maxThreadsPerBlock = %d\n", deviceProp.maxThreadsPerBlock);
		printf("maxThreadsDim = %d %d %d\n", deviceProp.maxThreadsDim[0], deviceProp.maxThreadsDim[1], deviceProp.maxThreadsDim[2]);
		printf("maxGridSize = %d %d %d\n", deviceProp.maxGridSize[0], deviceProp.maxGridSize[1], deviceProp.maxGridSize[2]);
		printf("totalConstMem = %lld\n", (unsigned long long int) deviceProp.totalConstMem);
		printf("major = %d\n", deviceProp.major);
		printf("minor = %d\n", deviceProp.minor);
		printf("clockRate = %d\n", deviceProp.clockRate);
		printf("memoryClockRate = %d\n", deviceProp.memoryClockRate);
		printf("multiProcessorCount = %d\n", deviceProp.multiProcessorCount);
		printf("textureAlignment = %lld\n", (unsigned long long int) deviceProp.textureAlignment);
David Rohr's avatar
David Rohr committed
365 366
	}
	
David Rohr's avatar
David Rohr committed
367 368
	if (DebugOutput >= 1)
	{
David Rohr's avatar
David Rohr committed
369
		printf("BioEM for CUDA initialized (MPI Rank %d), %d GPUs found, using GPU %d\n", mpi_rank, count, bestDevice);
David Rohr's avatar
David Rohr committed
370 371
	}
	
David Rohr's avatar
David Rohr committed
372 373 374
	return(0);
}

375 376 377
int bioem_cuda::deviceInit()
{
	deviceExit();
David Rohr's avatar
David Rohr committed
378
	
379
	selectCudaDevice();
380

381 382
	if (FFTAlgo) GPUAlgo = 2;

383 384 385 386 387
	gpumap = new bioem_RefMap;
	memcpy(gpumap, &RefMap, sizeof(bioem_RefMap));
	if (FFTAlgo == 0)
	{
		checkCudaErrors(cudaMalloc(&maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize));
388 389 390 391 392 393 394 395 396 397 398 399 400

		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));
		}
401 402 403 404 405 406 407 408 409
	}
	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;

410
	checkCudaErrors(cudaMalloc(&pProb_memory, pProb_device.get_size(RefMap.ntotRefMap, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC)));
411

412
	for (int i = 0; i < 2; i++)
413
	{
414
		checkCudaErrors(cudaStreamCreate(&cudaStream[i]));
415
		checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
416
		checkCudaErrors(cudaEventCreate(&cudaFFTEvent[i]));
417
		checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
418
	}
419 420 421 422 423
	if (GPUAsync)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[2]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[2]));
	}
424

425 426
	if (FFTAlgo)
	{
427
		checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
428 429 430 431
		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;
432
		checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
433
		checkCudaErrors(cudaHostAlloc(&pConvMapFFT_Host, param.FFTMapSize * sizeof(mycomplex_t) * 2, 0));
434
		checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
435 436
	}

437 438 439 440 441 442 443
	deviceInitialized = 1;
	return(0);
}

int bioem_cuda::deviceExit()
{
	if (deviceInitialized == 0) return(0);
444

445

David Rohr's avatar
David Rohr committed
446
	cudaFree(pProb_memory);
447 448
	cudaFree(sum);
	cudaFree(sumsquare);
449
	for (int i = 0; i < 2; i++)
450
	{
451
		cudaStreamDestroy(cudaStream[i]);
452
		cudaEventDestroy(cudaEvent[i]);
453
		cudaEventDestroy(cudaFFTEvent[i]);
454
		cudaFree(pConvMap_device[i]);
455
	}
456 457 458 459
	if (FFTAlgo)
	{
		cudaFree(pRefMapsFFT);
		cudaFree(pConvMapFFT);
460
		cudaFreeHost(pConvMapFFT_Host);
461 462
		cudaFree(pFFTtmp[0]);
		cudaFree(pFFTtmp2[0]);
463
	}
464 465 466 467 468 469 470 471
	else
	{
		cudaFree(maps);
	}
	if (GPUAlgo == 0 || GPUAlgo == 1)
	{
		cudaFree(pRefMap_device_Mod);
	}
472 473 474 475 476 477
	if (GPUAsync)
	{
		cudaStreamDestroy(cudaStream[2]);
		cudaEventDestroy(cudaEvent[2]);
	}

478
	delete gpumap;
479
	cudaThreadExit();
480

481 482 483 484 485 486
	deviceInitialized = 0;
	return(0);
}

int bioem_cuda::deviceStartRun()
{
David Rohr's avatar
David Rohr committed
487 488 489 490 491 492 493 494 495
	if (GPUWorkload >= 100)
	{
		maxRef = RefMap.ntotRefMap;
		pProb_host = &pProb;
	}
	else
	{
		maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
		pProb_host = new bioem_Probability;
496
		pProb_host->init(maxRef, param.nTotGridAngles, param.nTotCC, *this);
David Rohr's avatar
David Rohr committed
497 498
		pProb_host->copyFrom(&pProb, *this);
	}
499

David Rohr's avatar
David Rohr committed
500 501 502
	pProb_device = *pProb_host;
	pProb_device.ptr = pProb_memory;
	pProb_device.set_pointers();
503
	checkCudaErrors(cudaMemcpyAsync(pProb_device.ptr, pProb_host->ptr, pProb_host->get_size(maxRef, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC), cudaMemcpyHostToDevice, cudaStream[0]));
504 505 506

	if (FFTAlgo)
	{
507
		for (int j = 0;j < 2;j++)
508
		{
509
			for (int i = 0; i < 2; i++)
510
			{
511
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
512 513 514 515 516 517
				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);
				}
518
			        if (cufftSetCompatibilityMode(plan[i][j], CUFFT_COMPATIBILITY_FFTW_PADDING) != CUFFT_SUCCESS)
519 520 521 522 523 524 525 526 527
				{
					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);
				}
528
			}
529
			if (!GPUDualStream) break;
530 531
		}
	}
532 533 534 535 536
	return(0);
}

int bioem_cuda::deviceFinishRun()
{
537
	if (GPUAsync) cudaStreamSynchronize(cudaStream[0]);
538
	checkCudaErrors(cudaMemcpyAsync(pProb_host->ptr, pProb_device.ptr, pProb_host->get_size(maxRef, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC), cudaMemcpyDeviceToHost, cudaStream[0]));
539

540 541
	if (FFTAlgo)
	{
542 543
		for (int j = 0;j < 2;j++)
		{
544 545 546 547 548
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				cufftDestroy(plan[i][j]);
			}
549 550
			if (!GPUDualStream) break;
		}
551
	}
David Rohr's avatar
David Rohr committed
552 553 554 555
	cudaThreadSynchronize();
	if (GPUWorkload < 100)
	{
		pProb.copyFrom(pProb_host, *this);
556
		free_device_host(pProb_host->ptr);
David Rohr's avatar
David Rohr committed
557 558
		delete[] pProb_host;
	}
559

560 561 562
	return(0);
}

563 564 565 566 567 568 569 570 571 572 573 574
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);
}

575 576
bioem* bioem_cuda_create()
{
David Rohr's avatar
David Rohr committed
577 578 579 580 581 582 583 584 585
	int count;
	
	if (cudaGetDeviceCount(&count) != cudaSuccess) count = 0;
	if (count == 0)
	{
		printf("No CUDA device available, using fallback to CPU version\n");
		return new bioem;
	}

586 587
	return new bioem_cuda;
}