bioem_cuda.cu 25.1 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 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
/* Handing CUDA Driver errors */

// Expand and stringify argument
#define STRINGx(x) #x
#define STRING(x) STRINGx(x)

#define CU_ERROR_CHECK(call) \
  do { \
    CUresult __error__; \
    if ((__error__ = (call)) != CUDA_SUCCESS) { \
      printf(STRING(call), __func__, __FILE__, __LINE__, __error__, \
                     (const char * (*)(int))cuGetError); \
      return __error__; \
    } \
  } while (false)

static const char * cuGetError(CUresult result) {
  switch (result) {
    case CUDA_SUCCESS:                              return "No errors";
    case CUDA_ERROR_INVALID_VALUE:                  return "Invalid value";
    case CUDA_ERROR_OUT_OF_MEMORY:                  return "Out of memory";
    case CUDA_ERROR_NOT_INITIALIZED:                return "Driver not initialized";
    case CUDA_ERROR_DEINITIALIZED:                  return "Driver deinitialized";
    case CUDA_ERROR_PROFILER_DISABLED:              return "Profiler disabled";
    case CUDA_ERROR_PROFILER_NOT_INITIALIZED:       return "Profiler not initialized";
    case CUDA_ERROR_PROFILER_ALREADY_STARTED:       return "Profiler already started";
    case CUDA_ERROR_PROFILER_ALREADY_STOPPED:       return "Profiler already stopped";
    case CUDA_ERROR_NO_DEVICE:                      return "No CUDA-capable device available";
    case CUDA_ERROR_INVALID_DEVICE:                 return "Invalid device";
    case CUDA_ERROR_INVALID_IMAGE:                  return "Invalid kernel image";
    case CUDA_ERROR_INVALID_CONTEXT:                return "Invalid context";
    case CUDA_ERROR_CONTEXT_ALREADY_CURRENT:        return "Context already current";
    case CUDA_ERROR_MAP_FAILED:                     return "Map failed";
    case CUDA_ERROR_UNMAP_FAILED:                   return "Unmap failed";
    case CUDA_ERROR_ARRAY_IS_MAPPED:                return "Array is mapped";
    case CUDA_ERROR_ALREADY_MAPPED:                 return "Already mapped";
    case CUDA_ERROR_NO_BINARY_FOR_GPU:              return "No binary for GPU";
    case CUDA_ERROR_ALREADY_ACQUIRED:               return "Already acquired";
    case CUDA_ERROR_NOT_MAPPED:                     return "Not mapped";
    case CUDA_ERROR_NOT_MAPPED_AS_ARRAY:            return "Not mapped as array";
    case CUDA_ERROR_NOT_MAPPED_AS_POINTER:          return "Not mapped as pointer";
    case CUDA_ERROR_ECC_UNCORRECTABLE:              return "Uncorrectable ECC error";
    case CUDA_ERROR_UNSUPPORTED_LIMIT:              return "Unsupported CUlimit";
    case CUDA_ERROR_CONTEXT_ALREADY_IN_USE:         return "Context already in use";
    case CUDA_ERROR_INVALID_SOURCE:                 return "Invalid source";
    case CUDA_ERROR_FILE_NOT_FOUND:                 return "File not found";
    case CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND: return "Shared object symbol not found";
    case CUDA_ERROR_SHARED_OBJECT_INIT_FAILED:      return "Shared object initialization failed";
    case CUDA_ERROR_OPERATING_SYSTEM:               return "Operating System call failed";
    case CUDA_ERROR_INVALID_HANDLE:                 return "Invalid handle";
    case CUDA_ERROR_NOT_FOUND:                      return "Not found";
    case CUDA_ERROR_NOT_READY:                      return "CUDA not ready";
    case CUDA_ERROR_LAUNCH_FAILED:                  return "Launch failed";
    case CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES:        return "Launch exceeded resources";
    case CUDA_ERROR_LAUNCH_TIMEOUT:                 return "Launch exceeded timeout";
    case CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING:  return "Launch with incompatible texturing";
    case CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED:    return "Peer access already enabled";
    case CUDA_ERROR_PEER_ACCESS_NOT_ENABLED:        return "Peer access not enabled";
    case CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE:         return "Primary context active";
    case CUDA_ERROR_CONTEXT_IS_DESTROYED:           return "Context is destroyed";
    case CUDA_ERROR_ASSERT:                         return "Device assert failed";
    case CUDA_ERROR_TOO_MANY_PEERS:                 return "Too many peers";
    case CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED: return "Host memory already registered";
    case CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED:     return "Host memory not registered";
    case CUDA_ERROR_UNKNOWN:                        return "Unknown error";
    default:                                        return "Unknown error code";
  }
}

139 140 141 142
bioem_cuda::bioem_cuda()
{
	deviceInitialized = 0;
	GPUAlgo = getenv("GPUALGO") == NULL ? 2 : atoi(getenv("GPUALGO"));
143 144
	GPUAsync = getenv("GPUASYNC") == NULL ? 1 : atoi(getenv("GPUASYNC"));
	GPUWorkload = getenv("GPUWORKLOAD") == NULL ? 100 : atoi(getenv("GPUWORKLOAD"));
145
	GPUDualStream = getenv("GPUDUALSTREAM") == NULL ? 1 : atoi(getenv("GPUDUALSTREAM"));
146 147 148 149 150 151 152
}

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

153 154 155
__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)
156 157
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
158
	if (iRefMap < maxRef)
159
	{
160
		compareRefMap<0>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap, cent_x, cent_y);
161 162 163
	}
}

Pilar Cossio's avatar
Pilar Cossio committed
164
__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)
165 166
{
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX;
167
	if (iRefMap < maxRef)
168
	{
169
		compareRefMapShifted<1>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap);
170 171 172
	}
}

173 174 175 176 177 178
__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;
179
	for (int i = myid; i < mysize; i += mygrid) myptr[i] = 0;
180 181
}

Pilar Cossio's avatar
Pilar Cossio committed
182
__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)
183 184 185 186 187 188 189 190 191
{
	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;
192

193
	const bool threadActive = myShiftIdx < nShifts && myShiftIdy < nShifts && iRefMap < maxRef;
194

Pilar Cossio's avatar
Pilar Cossio committed
195
	compareRefMap<2>(iRefMap, iOrient, iConv, amp, pha, env, sumC, sumsquareC, pMap, pProb, param, RefMap, cent_x, cent_y, myShift, nShifts * nShifts, myRef, threadActive);
196 197
}

198
__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)
199 200
{
	if (myBlockIdxX >= NumberMaps) return;
201
	const mycuComplex_t* myin = (mycuComplex_t*) &refmap[(myBlockIdxX + Offset) * MapSize];
202
	const mycuComplex_t* myconv = (mycuComplex_t*) convmap;
203
	mycuComplex_t* myout = &out[myBlockIdxX * MapSize];
204
	for(int i = myThreadIdxX; i < NumberPixelsTotal; i += myBlockDimX)
205
	{
206 207 208 209 210 211 212
		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;
213 214 215
	}
}

216
__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)
217
{
218
	if (myBlockIdxX * myBlockDimX + myThreadIdxX >= maxRef) return;
219 220
	const int iRefMap = myBlockIdxX * myBlockDimX + myThreadIdxX + Offset;
	const myfloat_t* mylCC = &lCC[(myBlockIdxX * myBlockDimX + myThreadIdxX) * param.NumberPixels * param.NumberPixels];
221
	doRefMapFFT(iRefMap, iOrient, iConv, amp, pha, env, mylCC, sumC, sumsquareC, pProb, param, RefMap);
222 223
}

224 225 226 227 228 229 230 231 232 233 234 235 236
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

237
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)
238
{
239 240 241 242 243
	if (startMap)
	{
		cout << "Error startMap not implemented for GPU Code\n";
		exit(1);
	}
244 245 246 247
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
	}
248

249
	if (FFTAlgo)
250
	{
251
		memcpy(&pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t));
252 253 254 255 256 257
		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));
		}
258
		if (GPUDualStream)
259
		{
260 261 262 263 264 265
			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;
266
			const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
267 268
			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
269
			if (err != CUFFT_SUCCESS)
270
			{
David Rohr's avatar
David Rohr committed
271
				cout << "Error running CUFFT " << cufftGetErrorStrung(err) << "\n";
272 273
				exit(1);
			}
274
			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);
275
		}
276
		checkCudaErrors(cudaGetLastError());
277 278 279 280 281
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[1], cudaStream[1]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaFFTEvent[1], 0));
		}
282 283 284
	}
	else
	{
285
		checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream[0]));
286 287

		if (GPUAlgo == 2) //Loop over shifts
288
		{
289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307
			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;
308
			for (size_t i = 0; i < totalBlocks; i += nBlocks)
309
			{
Pilar Cossio's avatar
Pilar Cossio committed
310
				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);
311
			}
312
		}
313
		else if (GPUAlgo == 1) //Split shifts in multiple kernels
314
		{
315
			for (int cent_x = -param.param_device.maxDisplaceCenter; cent_x <= param.param_device.maxDisplaceCenter; cent_x = cent_x + param.param_device.GridSpaceCenter)
316
			{
317
				for (int cent_y = -param.param_device.maxDisplaceCenter; cent_y <= param.param_device.maxDisplaceCenter; cent_y = cent_y + param.param_device.GridSpaceCenter)
318
				{
Pilar Cossio's avatar
Pilar Cossio committed
319
					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);
320 321
				}
			}
322
		}
323
		else if (GPUAlgo == 0) //All shifts in one kernel
324
		{
325
			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);
326
		}
327
		else
328
		{
329 330
			cout << "Invalid GPU Algorithm selected\n";
			exit(1);
331
		}
332
	}
333 334
	if (GPUWorkload < 100)
	{
335
		bioem::compareRefMaps(iOrient, iConv, amp, pha, env, conv_map, localmultFFT, sumC, sumsquareC, maxRef);
336
	}
337 338
	if (GPUAsync)
	{
339
		checkCudaErrors(cudaEventRecord(cudaEvent[iConv & 1], cudaStream[0]));
340
	}
341 342
	else
	{
343
		checkCudaErrors(cudaStreamSynchronize(cudaStream[0]));
344 345 346 347
	}
	return(0);
}

David Rohr's avatar
David Rohr committed
348 349 350 351 352 353 354 355 356 357 358 359 360 361
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);
	}
362 363 364 365 366
	/* 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
367 368 369 370 371 372 373
	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
374
		CU_ERROR_CHECK(cuInit(0));
David Rohr's avatar
David Rohr committed
375
		CUdevice tmpDevice;
Luka Stanisic's avatar
Luka Stanisic committed
376
		CU_ERROR_CHECK(cuDeviceGet(&tmpDevice, i));
David Rohr's avatar
David Rohr committed
377
		CUcontext tmpContext;
Luka Stanisic's avatar
Luka Stanisic committed
378
		CU_ERROR_CHECK(cuCtxCreate(&tmpContext, 0, tmpDevice));
David Rohr's avatar
David Rohr committed
379
		if(cuMemGetInfo(&free, &total)) exit(1);
Luka Stanisic's avatar
Luka Stanisic committed
380
		CU_ERROR_CHECK(cuCtxDestroy(tmpContext));
David Rohr's avatar
David Rohr committed
381 382
		checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));

David Rohr's avatar
David Rohr committed
383
		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
384 385 386 387 388 389 390
		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;
		}
	}
391
#endif	
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
	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
413 414 415

	cudaGetDeviceProperties(&deviceProp ,bestDevice); 

David Rohr's avatar
David Rohr committed
416
	if (DebugOutput >= 3)
David Rohr's avatar
David Rohr committed
417
	{
David Rohr's avatar
David Rohr committed
418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
		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
434 435
	}
	
David Rohr's avatar
David Rohr committed
436 437
	if (DebugOutput >= 1)
	{
David Rohr's avatar
David Rohr committed
438
		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
439 440
	}
	
David Rohr's avatar
David Rohr committed
441 442 443
	return(0);
}

444 445 446
int bioem_cuda::deviceInit()
{
	deviceExit();
David Rohr's avatar
David Rohr committed
447
	
448
	selectCudaDevice();
449

450 451
	if (FFTAlgo) GPUAlgo = 2;

452 453 454 455 456
	gpumap = new bioem_RefMap;
	memcpy(gpumap, &RefMap, sizeof(bioem_RefMap));
	if (FFTAlgo == 0)
	{
		checkCudaErrors(cudaMalloc(&maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize));
457 458 459 460 461 462 463 464 465 466 467 468 469

		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));
		}
470 471 472 473 474 475 476 477 478
	}
	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;

479
	checkCudaErrors(cudaMalloc(&pProb_memory, pProb_device.get_size(RefMap.ntotRefMap, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC)));
480

481
	for (int i = 0; i < 2; i++)
482
	{
483
		checkCudaErrors(cudaStreamCreate(&cudaStream[i]));
484
		checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
485
		checkCudaErrors(cudaEventCreate(&cudaFFTEvent[i]));
486
		checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
487
	}
488 489 490 491 492
	if (GPUAsync)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[2]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[2]));
	}
493

494 495
	if (FFTAlgo)
	{
496
		checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
497 498 499 500
		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;
501
		checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
502
		checkCudaErrors(cudaHostAlloc(&pConvMapFFT_Host, param.FFTMapSize * sizeof(mycomplex_t) * 2, 0));
503
		checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
504 505
	}

506 507 508 509 510 511 512
	deviceInitialized = 1;
	return(0);
}

int bioem_cuda::deviceExit()
{
	if (deviceInitialized == 0) return(0);
513

514

David Rohr's avatar
David Rohr committed
515
	cudaFree(pProb_memory);
516 517
	cudaFree(sum);
	cudaFree(sumsquare);
518
	for (int i = 0; i < 2; i++)
519
	{
520
		cudaStreamDestroy(cudaStream[i]);
521
		cudaEventDestroy(cudaEvent[i]);
522
		cudaEventDestroy(cudaFFTEvent[i]);
523
		cudaFree(pConvMap_device[i]);
524
	}
525 526 527 528
	if (FFTAlgo)
	{
		cudaFree(pRefMapsFFT);
		cudaFree(pConvMapFFT);
529
		cudaFreeHost(pConvMapFFT_Host);
530 531
		cudaFree(pFFTtmp[0]);
		cudaFree(pFFTtmp2[0]);
532
	}
533 534 535 536 537 538 539 540
	else
	{
		cudaFree(maps);
	}
	if (GPUAlgo == 0 || GPUAlgo == 1)
	{
		cudaFree(pRefMap_device_Mod);
	}
541 542 543 544 545 546
	if (GPUAsync)
	{
		cudaStreamDestroy(cudaStream[2]);
		cudaEventDestroy(cudaEvent[2]);
	}

547
	delete gpumap;
548
	cudaThreadExit();
549

550 551 552 553 554 555
	deviceInitialized = 0;
	return(0);
}

int bioem_cuda::deviceStartRun()
{
David Rohr's avatar
David Rohr committed
556 557 558 559 560 561 562 563 564
	if (GPUWorkload >= 100)
	{
		maxRef = RefMap.ntotRefMap;
		pProb_host = &pProb;
	}
	else
	{
		maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
		pProb_host = new bioem_Probability;
565
		pProb_host->init(maxRef, param.nTotGridAngles, param.nTotCC, *this);
David Rohr's avatar
David Rohr committed
566 567
		pProb_host->copyFrom(&pProb, *this);
	}
568

David Rohr's avatar
David Rohr committed
569 570 571
	pProb_device = *pProb_host;
	pProb_device.ptr = pProb_memory;
	pProb_device.set_pointers();
572
	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]));
573 574 575

	if (FFTAlgo)
	{
576
		for (int j = 0;j < 2;j++)
577
		{
578
			for (int i = 0; i < 2; i++)
579
			{
580
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
581 582 583 584 585 586
				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);
				}
587
			        if (cufftSetCompatibilityMode(plan[i][j], CUFFT_COMPATIBILITY_FFTW_PADDING) != CUFFT_SUCCESS)
588 589 590 591 592 593 594 595 596
				{
					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);
				}
597
			}
598
			if (!GPUDualStream) break;
599 600
		}
	}
601 602 603 604 605
	return(0);
}

int bioem_cuda::deviceFinishRun()
{
606
	if (GPUAsync) cudaStreamSynchronize(cudaStream[0]);
607
	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]));
608

609 610
	if (FFTAlgo)
	{
611 612
		for (int j = 0;j < 2;j++)
		{
613 614 615 616 617
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				cufftDestroy(plan[i][j]);
			}
618 619
			if (!GPUDualStream) break;
		}
620
	}
David Rohr's avatar
David Rohr committed
621 622 623 624
	cudaThreadSynchronize();
	if (GPUWorkload < 100)
	{
		pProb.copyFrom(pProb_host, *this);
625
		free_device_host(pProb_host->ptr);
David Rohr's avatar
David Rohr committed
626 627
		delete[] pProb_host;
	}
628

629 630 631
	return(0);
}

632 633 634 635 636 637 638 639 640 641 642 643
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);
}

644 645 646 647 648 649 650 651 652 653 654 655 656
void bioem_cuda::rebalance(int workload)
{
  	if ((workload < 0) || (workload > 100) || (workload > GPUWorkload)) return;
	
  	if (DebugOutput >= 1)
	  {
	    printf("\t\tSetting GPU workload to %d%%\n", workload);
	  }
	
	GPUWorkload = workload;
	maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
}

657 658
bioem* bioem_cuda_create()
{
David Rohr's avatar
David Rohr committed
659 660 661 662 663 664 665 666 667
	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;
	}

668 669
	return new bioem_cuda;
}