bioem_cuda.cu 26.5 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
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
248
249
250
251
252

	float time;
	cudaEvent_t start, stop;
	if (DebugOutput >= 4)
	{
		checkCudaErrors(cudaEventCreate(&start));
		checkCudaErrors(cudaEventCreate(&stop));
		checkCudaErrors(cudaEventRecord(start, 0));
	}
253
254
255
256
	if (GPUAsync)
	{
		checkCudaErrors(cudaEventSynchronize(cudaEvent[iConv & 1]));
	}
257
258
259
260
261
262
263
264
	if (DebugOutput >= 4)
	{
		checkCudaErrors(cudaEventRecord(stop, 0));
		checkCudaErrors(cudaEventSynchronize(stop));
		checkCudaErrors(cudaEventElapsedTime(&time, start, stop));
		printf("\t\t\tGPU: time to synch projections %1.6f sec\n", time/1000);
		checkCudaErrors(cudaEventRecord(start, 0));
	}
265
	if (FFTAlgo)
266
	{
267
		memcpy(&pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], localmultFFT, param.FFTMapSize * sizeof(mycomplex_t));
268
		checkCudaErrors(cudaMemcpyAsync(&pConvMapFFT[(iConv & 1) * param.FFTMapSize], &pConvMapFFT_Host[(iConv & 1) * param.FFTMapSize], param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice, cudaStream[GPUAsync ? 2 : 0]));
269
270
271
272
273
274
275
276
		if (DebugOutput >= 4)
		{
			checkCudaErrors(cudaEventRecord(stop, 0));
			checkCudaErrors(cudaEventSynchronize(stop));
			checkCudaErrors(cudaEventElapsedTime(&time, start, stop));
			printf("\t\t\tGPU: time for memcpy %1.6f sec\n", time/1000);
			checkCudaErrors(cudaEventRecord(start, 0));
		}
277
278
279
280
281
		if (GPUAsync)
		{
			checkCudaErrors(cudaEventRecord(cudaEvent[2], cudaStream[2]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaEvent[2], 0));
		}
282
		if (GPUDualStream)
283
		{
284
285
286
287
288
289
			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;
290
			const int num = min(CUDA_FFTS_AT_ONCE, maxRef - i);
291
292
			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
293
			if (err != CUFFT_SUCCESS)
294
			{
David Rohr's avatar
David Rohr committed
295
				cout << "Error running CUFFT " << cufftGetErrorStrung(err) << "\n";
296
297
				exit(1);
			}
298
			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);
299
		}
300
		checkCudaErrors(cudaGetLastError());
301
302
303
304
305
		if (GPUDualStream)
		{
			checkCudaErrors(cudaEventRecord(cudaFFTEvent[1], cudaStream[1]));
			checkCudaErrors(cudaStreamWaitEvent(cudaStream[0], cudaFFTEvent[1], 0));
		}
306
307
308
	}
	else
	{
309
		checkCudaErrors(cudaMemcpyAsync(pConvMap_device[iConv & 1], conv_map, sizeof(myfloat_t) * RefMap.refMapSize, cudaMemcpyHostToDevice, cudaStream[0]));
310
311
312
313
314
315
316
317
		if (DebugOutput >= 4)
		{
			checkCudaErrors(cudaEventRecord(stop, 0));
			checkCudaErrors(cudaEventSynchronize(stop));
			checkCudaErrors(cudaEventElapsedTime(&time, start, stop));
			printf("\t\t\tGPU: time for memcpy %1.6f sec\n", time/1000);
			checkCudaErrors(cudaEventRecord(start, 0) );
		}
318
		if (GPUAlgo == 2) //Loop over shifts
319
		{
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
			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;
339
			for (size_t i = 0; i < totalBlocks; i += nBlocks)
340
			{
Pilar Cossio's avatar
Pilar Cossio committed
341
				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);
342
			}
343
		}
344
		else if (GPUAlgo == 1) //Split shifts in multiple kernels
345
		{
346
			for (int cent_x = -param.param_device.maxDisplaceCenter; cent_x <= param.param_device.maxDisplaceCenter; cent_x = cent_x + param.param_device.GridSpaceCenter)
347
			{
348
				for (int cent_y = -param.param_device.maxDisplaceCenter; cent_y <= param.param_device.maxDisplaceCenter; cent_y = cent_y + param.param_device.GridSpaceCenter)
349
				{
Pilar Cossio's avatar
Pilar Cossio committed
350
					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);
351
352
				}
			}
353
		}
354
		else if (GPUAlgo == 0) //All shifts in one kernel
355
		{
356
			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);
357
		}
358
		else
359
		{
360
361
			cout << "Invalid GPU Algorithm selected\n";
			exit(1);
362
		}
363
	}
364
365
366
367
368
369
370
371
	if (DebugOutput >= 4)
	{
		checkCudaErrors(cudaEventRecord(stop, 0));
		checkCudaErrors(cudaEventSynchronize(stop));
		checkCudaErrors(cudaEventElapsedTime(&time, start, stop));
		printf("\t\t\tGPU: time to run CUDA %1.6f sec\n", time/1000);
		checkCudaErrors(cudaEventRecord(start, 0));
	}
372
373
	if (GPUWorkload < 100)
	{
374
		bioem::compareRefMaps(iOrient, iConv, amp, pha, env, conv_map, localmultFFT, sumC, sumsquareC, maxRef);
375
	}
376
377
378
379
380
381
382
	if (DebugOutput >= 4)
	{
		checkCudaErrors(cudaEventRecord(stop, 0));
		checkCudaErrors(cudaEventSynchronize(stop));
		checkCudaErrors(cudaEventElapsedTime(&time, start, stop));
		printf("\t\t\tGPU: time to run OMP %1.6f sec\n", time/1000);
	}
383
384
	if (GPUAsync)
	{
385
		checkCudaErrors(cudaEventRecord(cudaEvent[iConv & 1], cudaStream[0]));
386
	}
387
388
	else
	{
389
		checkCudaErrors(cudaStreamSynchronize(cudaStream[0]));
390
391
392
393
	}
	return(0);
}

David Rohr's avatar
David Rohr committed
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
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
Luka Stanisic's avatar
Luka Stanisic committed
415
		CU_ERROR_CHECK(cuInit(0));
David Rohr's avatar
David Rohr committed
416
		CUdevice tmpDevice;
Luka Stanisic's avatar
Luka Stanisic committed
417
		CU_ERROR_CHECK(cuDeviceGet(&tmpDevice, i));
David Rohr's avatar
David Rohr committed
418
		CUcontext tmpContext;
Luka Stanisic's avatar
Luka Stanisic committed
419
		CU_ERROR_CHECK(cuCtxCreate(&tmpContext, 0, tmpDevice));
David Rohr's avatar
David Rohr committed
420
		if(cuMemGetInfo(&free, &total)) exit(1);
Luka Stanisic's avatar
Luka Stanisic committed
421
		CU_ERROR_CHECK(cuCtxDestroy(tmpContext));
David Rohr's avatar
David Rohr committed
422
423
		checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));

David Rohr's avatar
David Rohr committed
424
		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
425
426
427
428
429
430
431
		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;
		}
	}
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
	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
453
454
455

	cudaGetDeviceProperties(&deviceProp ,bestDevice); 

David Rohr's avatar
David Rohr committed
456
	if (DebugOutput >= 3)
David Rohr's avatar
David Rohr committed
457
	{
David Rohr's avatar
David Rohr committed
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
		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
474
475
	}
	
David Rohr's avatar
David Rohr committed
476
477
	if (DebugOutput >= 1)
	{
David Rohr's avatar
David Rohr committed
478
		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
479
480
	}
	
David Rohr's avatar
David Rohr committed
481
482
483
	return(0);
}

484
485
486
int bioem_cuda::deviceInit()
{
	deviceExit();
David Rohr's avatar
David Rohr committed
487
	
488
	selectCudaDevice();
489

490
491
	if (FFTAlgo) GPUAlgo = 2;

492
493
494
495
496
	gpumap = new bioem_RefMap;
	memcpy(gpumap, &RefMap, sizeof(bioem_RefMap));
	if (FFTAlgo == 0)
	{
		checkCudaErrors(cudaMalloc(&maps, sizeof(myfloat_t) * RefMap.ntotRefMap * RefMap.refMapSize));
497
498
499
500
501
502
503
504
505
506
507
508
509

		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));
		}
510
511
512
513
514
515
516
517
518
	}
	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;

519
	checkCudaErrors(cudaMalloc(&pProb_memory, pProb_device.get_size(RefMap.ntotRefMap, param.nTotGridAngles, param.nTotCC, param.param_device.writeAngles, param.param_device.writeCC)));
520

521
	for (int i = 0; i < 2; i++)
522
	{
523
		checkCudaErrors(cudaStreamCreate(&cudaStream[i]));
524
		checkCudaErrors(cudaEventCreate(&cudaEvent[i]));
525
		checkCudaErrors(cudaEventCreate(&cudaFFTEvent[i]));
526
		checkCudaErrors(cudaMalloc(&pConvMap_device[i], sizeof(myfloat_t) * RefMap.refMapSize));
527
	}
528
529
530
531
532
	if (GPUAsync)
	{
		checkCudaErrors(cudaStreamCreate(&cudaStream[2]));
		checkCudaErrors(cudaEventCreate(&cudaEvent[2]));
	}
533

534
535
	if (FFTAlgo)
	{
536
		checkCudaErrors(cudaMalloc(&pRefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t)));
537
538
539
540
		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;
541
		checkCudaErrors(cudaMalloc(&pConvMapFFT, param.FFTMapSize * sizeof(mycomplex_t) * 2));
542
		checkCudaErrors(cudaHostAlloc(&pConvMapFFT_Host, param.FFTMapSize * sizeof(mycomplex_t) * 2, 0));
543
		checkCudaErrors(cudaMemcpy(pRefMapsFFT, RefMap.RefMapsFFT, RefMap.ntotRefMap * param.FFTMapSize * sizeof(mycomplex_t), cudaMemcpyHostToDevice));
544
545
	}

546
547
548
549
550
551
552
	deviceInitialized = 1;
	return(0);
}

int bioem_cuda::deviceExit()
{
	if (deviceInitialized == 0) return(0);
553

554

David Rohr's avatar
David Rohr committed
555
	cudaFree(pProb_memory);
556
557
	cudaFree(sum);
	cudaFree(sumsquare);
558
	for (int i = 0; i < 2; i++)
559
	{
560
		cudaStreamDestroy(cudaStream[i]);
561
		cudaEventDestroy(cudaEvent[i]);
562
		cudaEventDestroy(cudaFFTEvent[i]);
563
		cudaFree(pConvMap_device[i]);
564
	}
565
566
567
568
	if (FFTAlgo)
	{
		cudaFree(pRefMapsFFT);
		cudaFree(pConvMapFFT);
569
		cudaFreeHost(pConvMapFFT_Host);
570
571
		cudaFree(pFFTtmp[0]);
		cudaFree(pFFTtmp2[0]);
572
	}
573
574
575
576
577
578
579
580
	else
	{
		cudaFree(maps);
	}
	if (GPUAlgo == 0 || GPUAlgo == 1)
	{
		cudaFree(pRefMap_device_Mod);
	}
581
582
583
584
585
586
	if (GPUAsync)
	{
		cudaStreamDestroy(cudaStream[2]);
		cudaEventDestroy(cudaEvent[2]);
	}

587
	delete gpumap;
588
	cudaThreadExit();
589

590
591
592
593
594
595
	deviceInitialized = 0;
	return(0);
}

int bioem_cuda::deviceStartRun()
{
David Rohr's avatar
David Rohr committed
596
597
598
599
600
601
602
603
604
	if (GPUWorkload >= 100)
	{
		maxRef = RefMap.ntotRefMap;
		pProb_host = &pProb;
	}
	else
	{
		maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
		pProb_host = new bioem_Probability;
605
		pProb_host->init(maxRef, param.nTotGridAngles, param.nTotCC, *this);
David Rohr's avatar
David Rohr committed
606
607
		pProb_host->copyFrom(&pProb, *this);
	}
608

David Rohr's avatar
David Rohr committed
609
610
611
	pProb_device = *pProb_host;
	pProb_device.ptr = pProb_memory;
	pProb_device.set_pointers();
612
	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]));
613
614
615

	if (FFTAlgo)
	{
616
		for (int j = 0;j < 2;j++)
617
		{
618
			for (int i = 0; i < 2; i++)
619
			{
620
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
621
622
623
624
625
626
				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);
				}
627
			        if (cufftSetCompatibilityMode(plan[i][j], CUFFT_COMPATIBILITY_FFTW_PADDING) != CUFFT_SUCCESS)
628
629
630
631
632
633
634
635
636
				{
					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);
				}
637
			}
638
			if (!GPUDualStream) break;
639
640
		}
	}
641
642
643
644
645
	return(0);
}

int bioem_cuda::deviceFinishRun()
{
646
	if (GPUAsync) cudaStreamSynchronize(cudaStream[0]);
647
	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]));
648

649
650
	if (FFTAlgo)
	{
651
652
		for (int j = 0;j < 2;j++)
		{
653
654
655
656
657
			for (int i = 0; i < 2; i++)
			{
				if (i && maxRef % CUDA_FFTS_AT_ONCE == 0) continue;
				cufftDestroy(plan[i][j]);
			}
658
659
			if (!GPUDualStream) break;
		}
660
	}
David Rohr's avatar
David Rohr committed
661
662
663
664
	cudaThreadSynchronize();
	if (GPUWorkload < 100)
	{
		pProb.copyFrom(pProb_host, *this);
665
		free_device_host(pProb_host->ptr);
David Rohr's avatar
David Rohr committed
666
667
		delete[] pProb_host;
	}
668

669
670
671
	return(0);
}

672
673
674
675
676
677
678
679
680
681
682
683
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);
}

684
685
void bioem_cuda::rebalance(int workload)
{
686
687
688
689
690
691
692
693
694
	if ((workload < 0) || (workload > 100) || (workload == GPUWorkload)) return;

	deviceFinishRun();

	if (DebugOutput >= 2)
	{
	  printf("\t\tSetting GPU workload to %d%%\n", workload);
	}

695
696
	GPUWorkload = workload;
	maxRef = (size_t) RefMap.ntotRefMap * (size_t) GPUWorkload / 100;
697
698

	deviceStartRun();
699
700
}

701
702
bioem* bioem_cuda_create()
{
David Rohr's avatar
David Rohr committed
703
704
705
706
707
708
709
710
711
	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;
	}

712
713
	return new bioem_cuda;
}