bioem_cuda.cu 19.6 KB
Newer Older
Pilar Cossio's avatar
License  
Pilar Cossio committed
1 2 3 4
/* ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
        < 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.
5

Pilar Cossio's avatar
License  
Pilar Cossio committed
6 7 8 9
                See license statement for terms of distribution.

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

10 11 12 13 14 15 16 17 18 19 20
#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
21
//#include "helper_cuda.h"
22

23 24 25 26 27 28 29
#define checkCudaErrors(error) \
{ \
	if ((error) != cudaSuccess) \
	{ \
		printf("CUDA Error %d / %s (%s: %d)\n", error, cudaGetErrorString(error), __FILE__, __LINE__); \
		exit(1); \
	} \
30 31
}

David Rohr's avatar
David Rohr committed
32 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
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";
}

69 70 71 72
bioem_cuda::bioem_cuda()
{
	deviceInitialized = 0;
	GPUAlgo = getenv("GPUALGO") == NULL ? 2 : atoi(getenv("GPUALGO"));
73 74
	GPUAsync = getenv("GPUASYNC") == NULL ? 1 : atoi(getenv("GPUASYNC"));
	GPUWorkload = getenv("GPUWORKLOAD") == NULL ? 100 : atoi(getenv("GPUWORKLOAD"));
75
	GPUDualStream = getenv("GPUDUALSTREAM") == NULL ? 1 : atoi(getenv("GPUDUALSTREAM"));
76 77 78 79 80 81 82
}

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

83
__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)
84 85
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
86
	if (iRefMap < maxRef)
87
	{
88
		compareRefMap<0>(iRefMap, iOrient, iConv, pMap, pProb, param, RefMap, cent_x, cent_y);
89 90 91
	}
}

92
__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)
93 94
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
95
	if (iRefMap < maxRef)
96
	{
97
		compareRefMapShifted<1>(iRefMap, iOrient, iConv, pMap, pProb, param, RefMap);
98 99 100
	}
}

101 102 103 104 105 106
__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;
107
	for (int i = myid; i < mysize; i += mygrid) myptr[i] = 0;
108 109
}

110
__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)
111 112 113 114 115 116 117 118 119
{
	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;
120

121
	const bool threadActive = myShiftIdx < nShifts && myShiftIdy < nShifts && iRefMap < maxRef;
122

123
	compareRefMap<2>(iRefMap, iOrient, iConv, pMap, pProb, param, RefMap, cent_x, cent_y, myShift, nShifts * nShifts, myRef, threadActive);
124 125
}

126
__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)
127 128
{
	if (myBlockIdxX >= NumberMaps) return;
129
	const mycuComplex_t* myin = (mycuComplex_t*) &refmap[(myBlockIdxX + Offset) * MapSize];
130
	const mycuComplex_t* myconv = (mycuComplex_t*) convmap;
131
	mycuComplex_t* myout = &out[myBlockIdxX * MapSize];
132
	for(int i = myThreadIdxX; i < NumberPixelsTotal; i += myBlockDimX)
133
	{
134 135 136 137 138 139 140
		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;
141 142 143
	}
}

144
__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)
145
{
146
	if (myBlockIdxX * myBlockDimX + myThreadIdxX >= maxRef) return;
147 148
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX + Offset;
	const myfloat_t* mylCC = &lCC[(myBlockIdxX * myBlockDimX + myThreadIdxX) * param.NumberPixels * param.NumberPixels];
149
	doRefMapFFT(iRefMap, iOrient, iConv, amp, pha, env, mylCC, sumC, sumsquareC, pProb, param, RefMap);
150 151
}

152 153 154 155 156 157 158 159 160 161 162 163 164
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

165
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)
166
{
167 168 169 170 171
	if (startMap)
	{
		cout << "Error startMap not implemented for GPU Code\n";
		exit(1);
	}
172 173 174 175
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
	}
176

177
	if (FFTAlgo)
178
	{
179
		memcpy(&pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t));
180 181 182 183 184 185
		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));
		}
186
		if (GPUDualStream)
187
		{
188 189 190 191 192 193
			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;
194
			const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
195 196
			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
197
			if (err != CUFFT_SUCCESS)
198
			{
David Rohr's avatar
David Rohr committed
199
				cout << "Error running CUFFT " << cufftGetErrorStrung(err) << "\n";
200 201
				exit(1);
			}
202
			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);
203
		}
204
		checkCudaErrors(cudaGetLastError());
205 206 207 208 209
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[1], cudaStream[1]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaFFTEvent[1], 0));
		}
210 211 212
	}
	else
	{
213
		checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream[0]));
214 215

		if (GPUAlgo == 2) //Loop over shifts
216
		{
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
			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;
236
			for (size_t i = 0; i < totalBlocks; i += nBlocks)
237
			{
238
				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);
239
			}
240
		}
241
		else if (GPUAlgo == 1) //Split shifts in multiple kernels
242
		{
243
			for (int cent_x = -param.param_device.maxDisplaceCenter; cent_x <= param.param_device.maxDisplaceCenter; cent_x = cent_x + param.param_device.GridSpaceCenter)
244
			{
245
				for (int cent_y = -param.param_device.maxDisplaceCenter; cent_y <= param.param_device.maxDisplaceCenter; cent_y = cent_y + param.param_device.GridSpaceCenter)
246
				{
247
					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);
248 249
				}
			}
250
		}
251
		else if (GPUAlgo == 0) //All shifts in one kernel
252
		{
253
			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);
254
		}
255
		else
256
		{
257 258
			cout << "Invalid GPU Algorithm selected\n";
			exit(1);
259
		}
260
	}
261 262
	if (GPUWorkload < 100)
	{
263
		bioem::compareRefMaps(iOrient, iConv, amp, pha, env, conv_map, localmultFFT, sumC, sumsquareC, maxRef);
264
	}
265 266
	if (GPUAsync)
	{
267
		checkCudaErrors(cudaEventRecord(cudaEvent[iConv & 1], cudaStream[0]));
268
	}
269 270
	else
	{
271
		checkCudaErrors(cudaStreamSynchronize(cudaStream[0]));
272 273 274 275
	}
	return(0);
}

David Rohr's avatar
David Rohr committed
276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
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);
	}
	for (int i = 0;i < count;i++)
	{
#if CUDA_VERSION > 3010
		size_t free, total;
#else
		unsigned int free, total;
#endif
		cuInit(0);
		CUdevice tmpDevice;
		cuDeviceGet(&tmpDevice, i);
		CUcontext tmpContext;
		cuCtxCreate(&tmpContext, 0, tmpDevice);
		if(cuMemGetInfo(&free, &total)) exit(1);
		cuCtxDestroy(tmpContext);
		checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));

David Rohr's avatar
David Rohr committed
306
		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
307 308 309 310 311 312 313
		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;
		}
	}
314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
	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
335 336 337

	cudaGetDeviceProperties(&deviceProp ,bestDevice); 

David Rohr's avatar
David Rohr committed
338
	if (DebugOutput >= 3)
David Rohr's avatar
David Rohr committed
339
	{
David Rohr's avatar
David Rohr committed
340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355
		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
356 357
	}
	
David Rohr's avatar
David Rohr committed
358 359
	if (DebugOutput >= 1)
	{
David Rohr's avatar
David Rohr committed
360
		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
361 362
	}
	
David Rohr's avatar
David Rohr committed
363 364 365
	return(0);
}

366 367 368
int bioem_cuda::deviceInit()
{
	deviceExit();
David Rohr's avatar
David Rohr committed
369
	
370
	selectCudaDevice();
371

372 373
	if (FFTAlgo) GPUAlgo = 2;

374 375 376 377 378
	gpumap = new bioem_RefMap;
	memcpy(gpumap, &RefMap, sizeof(bioem_RefMap));
	if (FFTAlgo == 0)
	{
		checkCudaErrors(cudaMalloc(&maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize));
379 380 381 382 383 384 385 386 387 388 389 390 391

		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));
		}
392 393 394 395 396 397 398 399 400
	}
	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;

401
	checkCudaErrors(cudaMalloc(&pProb_memory, pProb_device.get_size(RefMap.ntotRefMap, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC)));
402

403
	for (int i = 0; i < 2; i++)
404
	{
405
		checkCudaErrors(cudaStreamCreate(&cudaStream[i]));
406
		checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
407
		checkCudaErrors(cudaEventCreate(&cudaFFTEvent[i]));
408
		checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
409
	}
410 411 412 413 414
	if (GPUAsync)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[2]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[2]));
	}
415

416 417
	if (FFTAlgo)
	{
418
		checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
419 420 421 422
		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;
423
		checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
424
		checkCudaErrors(cudaHostAlloc(&pConvMapFFT_Host, param.FFTMapSize * sizeof(mycomplex_t) * 2, 0));
425
		checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
426 427
	}

428 429 430 431 432 433 434
	deviceInitialized = 1;
	return(0);
}

int bioem_cuda::deviceExit()
{
	if (deviceInitialized == 0) return(0);
435

436

David Rohr's avatar
David Rohr committed
437
	cudaFree(pProb_memory);
438 439
	cudaFree(sum);
	cudaFree(sumsquare);
440
	for (int i = 0; i < 2; i++)
441
	{
442
		cudaStreamDestroy(cudaStream[i]);
443
		cudaEventDestroy(cudaEvent[i]);
444
		cudaEventDestroy(cudaFFTEvent[i]);
445
		cudaFree(pConvMap_device[i]);
446
	}
447 448 449 450
	if (FFTAlgo)
	{
		cudaFree(pRefMapsFFT);
		cudaFree(pConvMapFFT);
451
		cudaFreeHost(pConvMapFFT_Host);
452 453
		cudaFree(pFFTtmp[0]);
		cudaFree(pFFTtmp2[0]);
454
	}
455 456 457 458 459 460 461 462
	else
	{
		cudaFree(maps);
	}
	if (GPUAlgo == 0 || GPUAlgo == 1)
	{
		cudaFree(pRefMap_device_Mod);
	}
463 464 465 466 467 468
	if (GPUAsync)
	{
		cudaStreamDestroy(cudaStream[2]);
		cudaEventDestroy(cudaEvent[2]);
	}

469
	delete gpumap;
470
	cudaThreadExit();
471

472 473 474 475 476 477
	deviceInitialized = 0;
	return(0);
}

int bioem_cuda::deviceStartRun()
{
David Rohr's avatar
David Rohr committed
478 479 480 481 482 483 484 485 486
	if (GPUWorkload >= 100)
	{
		maxRef = RefMap.ntotRefMap;
		pProb_host = &pProb;
	}
	else
	{
		maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
		pProb_host = new bioem_Probability;
487
		pProb_host->init(maxRef, param.nTotGridAngles, param.nTotCC, *this);
David Rohr's avatar
David Rohr committed
488 489
		pProb_host->copyFrom(&pProb, *this);
	}
490

David Rohr's avatar
David Rohr committed
491 492 493
	pProb_device = *pProb_host;
	pProb_device.ptr = pProb_memory;
	pProb_device.set_pointers();
494
	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]));
495 496 497

	if (FFTAlgo)
	{
498
		for (int j = 0;j < 2;j++)
499
		{
500
			for (int i = 0; i < 2; i++)
501
			{
502
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
				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);
				}
519
			}
520
			if (!GPUDualStream) break;
521 522
		}
	}
523 524 525 526 527
	return(0);
}

int bioem_cuda::deviceFinishRun()
{
528
	if (GPUAsync) cudaStreamSynchronize(cudaStream[0]);
529
	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]));
530

531 532
	if (FFTAlgo)
	{
533 534
		for (int j = 0;j < 2;j++)
		{
535 536 537 538 539
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				cufftDestroy(plan[i][j]);
			}
540 541
			if (!GPUDualStream) break;
		}
542
	}
David Rohr's avatar
David Rohr committed
543 544 545 546
	cudaThreadSynchronize();
	if (GPUWorkload < 100)
	{
		pProb.copyFrom(pProb_host, *this);
547
		free_device_host(pProb_host->ptr);
David Rohr's avatar
David Rohr committed
548 549
		delete[] pProb_host;
	}
550

551 552 553
	return(0);
}

554 555 556 557 558 559 560 561 562 563 564 565
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);
}

566 567
bioem* bioem_cuda_create()
{
David Rohr's avatar
David Rohr committed
568 569 570 571 572 573 574 575 576
	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;
	}

577 578
	return new bioem_cuda;
}