CudaContext.cpp 36.8 KB
Newer Older
1
2
3
4
5
6
7
8
/* -------------------------------------------------------------------------- *
 *                                   OpenMM                                   *
 * -------------------------------------------------------------------------- *
 * This is part of the OpenMM molecular simulation toolkit originating from   *
 * Simbios, the NIH National Center for Physics-Based Simulation of           *
 * Biological Structures at Stanford, funded under the NIH Roadmap for        *
 * Medical Research, grant U54 GM072970. See https://simtk.org.               *
 *                                                                            *
9
 * Portions copyright (c) 2009-2025 Stanford University and the Authors.      *
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
 * Authors: Peter Eastman                                                     *
 * Contributors:                                                              *
 *                                                                            *
 * This program is free software: you can redistribute it and/or modify       *
 * it under the terms of the GNU Lesser General Public License as published   *
 * by the Free Software Foundation, either version 3 of the License, or       *
 * (at your option) any later version.                                        *
 *                                                                            *
 * This program is distributed in the hope that it will be useful,            *
 * but WITHOUT ANY WARRANTY; without even the implied warranty of             *
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the              *
 * GNU Lesser General Public License for more details.                        *
 *                                                                            *
 * You should have received a copy of the GNU Lesser General Public License   *
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.      *
 * -------------------------------------------------------------------------- */

#ifdef WIN32
  #define _USE_MATH_DEFINES // Needed to get M_PI
#endif
#include <cmath>
#include "CudaContext.h"
#include "CudaArray.h"
33
#include "CudaBondedUtilities.h"
34
#include "CudaEvent.h"
35
#include "CudaFFT3D.h"
36
#include "CudaIntegrationUtilities.h"
37
#include "CudaKernels.h"
38
#include "CudaKernelSources.h"
39
#include "CudaNonbondedUtilities.h"
40
#include "CudaProgram.h"
41
#include "CudaSort.h"
42
#include "openmm/common/ComputeArray.h"
43
#include "openmm/common/ContextSelector.h"
44
#include "SHA1.h"
45
#include "openmm/MonteCarloFlexibleBarostat.h"
46
47
48
#include "openmm/Platform.h"
#include "openmm/System.h"
#include "openmm/VirtualSite.h"
49
#include "CudaExpressionUtilities.h"
50
#include "openmm/internal/ContextImpl.h"
51
52
53
#include <algorithm>
#include <cstdlib>
#include <fstream>
54
#include <iomanip>
55
#include <iostream>
56
#include <set>
57
58
#include <sstream>
#include <typeinfo>
59
#include <sys/stat.h>
60
#include <cudaProfiler.h>
61
#include <nvrtc.h>
62
63
64
#ifndef WIN32
  #include <unistd.h>
#endif
peastman's avatar
peastman committed
65

66
67
68
69
70

#define CHECK_RESULT(result) CHECK_RESULT2(result, errorMessage);
#define CHECK_RESULT2(result, prefix) \
    if (result != CUDA_SUCCESS) { \
        std::stringstream m; \
71
        m<<prefix<<": "<<getErrorString(result)<<" ("<<result<<")"<<" at "<<__FILE__<<":"<<__LINE__; \
72
73
        throw OpenMMException(m.str());\
    }
74
75
76
77
78
79
#define CHECK_NVRTC_RESULT(result, prefix) \
    if (result != NVRTC_SUCCESS) { \
        stringstream m; \
        m<<prefix<<": "<<nvrtcGetErrorString(result)<<" ("<<result<<")"<<" at "<<__FILE__<<":"<<__LINE__; \
        throw OpenMMException(m.str());\
    }
80
81
82
83
84

using namespace OpenMM;
using namespace std;

const int CudaContext::ThreadBlockSize = 64;
85
const int CudaContext::TileSize = sizeof(tileflags)*8;
86
87
bool CudaContext::hasInitializedCuda = false;

88
CudaContext::CudaContext(const System& system, int deviceIndex, bool useBlockingSync, const string& precision, const string& tempDir, CudaPlatform::PlatformData& platformData,
89
        CudaContext* originalContext) : ComputeContext(system), platformData(platformData), contextIsValid(false), hasAssignedPosqCharges(false),
90
        pinnedBuffer(NULL), integration(NULL), expression(NULL), bonded(NULL), nonbonded(NULL), useBlockingSync(useBlockingSync) {
91
92
    int cudaDriverVersion;
    cuDriverGetVersion(&cudaDriverVersion);
93
94
95
96
97
98
    if (!hasInitializedCuda) {
        CHECK_RESULT2(cuInit(0), "Error initializing CUDA");
        hasInitializedCuda = true;
    }
    if (precision == "single") {
        useDoublePrecision = false;
99
        useMixedPrecision = false;
100
101
102
    }
    else if (precision == "mixed") {
        useDoublePrecision = false;
103
        useMixedPrecision = true;
104
105
106
    }
    else if (precision == "double") {
        useDoublePrecision = true;
107
        useMixedPrecision = false;
108
109
    }
    else
110
        throw OpenMMException("Illegal value for Precision: "+precision);
111
112
    char* cacheVariable = getenv("OPENMM_CACHE_DIR");
    cacheDir = (cacheVariable == NULL ? tempDir : string(cacheVariable));
113
#ifdef WIN32
114
    this->tempDir = tempDir+"\\";
115
    cacheDir = cacheDir+"\\";
116
117
#else
    this->tempDir = tempDir+"/";
118
    cacheDir = cacheDir+"/";
119
120
121
#endif
    contextIndex = platformData.contexts.size();
    string errorMessage = "Error initializing Context";
122
123
124
125
126
127
    if (originalContext == NULL) {
        isLinkedContext = false;
        int numDevices;
        CHECK_RESULT(cuDeviceGetCount(&numDevices));
        if (deviceIndex < -1 || deviceIndex >= numDevices)
            throw OpenMMException("Illegal value for DeviceIndex: "+intToString(deviceIndex));
128

129
130
131
132
133
134
        vector<int> devicePrecedence;
        if (deviceIndex == -1) {
            devicePrecedence = getDevicePrecedence();
        } else {
            devicePrecedence.push_back(deviceIndex);
        }
135

136
137
138
139
140
141
142
143
144
145
        this->deviceIndex = -1;
        for (int i = 0; i < static_cast<int>(devicePrecedence.size()); i++) {
            int trialDeviceIndex = devicePrecedence[i];
            CHECK_RESULT(cuDeviceGet(&device, trialDeviceIndex));
            defaultOptimizationOptions = "--use_fast_math";
            unsigned int flags = CU_CTX_MAP_HOST;
            if (useBlockingSync)
                flags += CU_CTX_SCHED_BLOCKING_SYNC;
            else
                flags += CU_CTX_SCHED_SPIN;
146

147
148
            if (cuCtxCreate(&context, flags, device) == CUDA_SUCCESS) {
                this->deviceIndex = trialDeviceIndex;
149
150
                CUcontext popped;
                cuCtxPopCurrent(&popped);
151
152
                break;
            }
153
        }
154
        if (this->deviceIndex == -1) {
155
156
157
158
            if (deviceIndex != -1)
                throw OpenMMException("The requested CUDA device could not be loaded");
            else
                throw OpenMMException("No compatible CUDA device is available");
159
        }
160
161
162
163
164
165
    }
    else {
        isLinkedContext = true;
        context = originalContext->context;
        this->deviceIndex = originalContext->deviceIndex;
        this->device = originalContext->device;
166
    }
167

168
    int major, minor;
169
170
    CHECK_RESULT(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device));
    CHECK_RESULT(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device));
peastman's avatar
peastman committed
171
    int numThreadBlocksPerComputeUnit = (major == 6 ? 4 : 6);
172
    if (cudaDriverVersion < 7000) {
173
174
175
176
177
        // This is a workaround to support GTX 980 with CUDA 6.5.  It reports
        // its compute capability as 5.2, but the compiler doesn't support
        // anything beyond 5.0.
        if (major == 5)
            minor = 0;
178
179
    }
    if (cudaDriverVersion < 8000) {
180
181
182
183
184
185
186
        // This is a workaround to support Pascal with CUDA 7.5.  It reports
        // its compute capability as 6.x, but the compiler doesn't support
        // anything beyond 5.3.
        if (major == 6) {
            major = 5;
            minor = 3;
        }
187
    }
188
    gpuArchitecture = 10*major+minor;
189
    computeCapability = major+0.1*minor;
190

191
    contextIsValid = true;
192
    ContextSelector selector(*this);
193
    CHECK_RESULT(cuCtxSetCacheConfig(CU_FUNC_CACHE_PREFER_SHARED));
194
    if (contextIndex > 0 && originalContext == NULL) {
root's avatar
root committed
195
196
197
        int canAccess;
        cuDeviceCanAccessPeer(&canAccess, getDevice(), platformData.contexts[0]->getDevice());
        if (canAccess) {
198
199
200
201
            {
                ContextSelector selector2(*platformData.contexts[0]);
                CHECK_RESULT(cuCtxEnablePeerAccess(getContext(), 0));
            }
root's avatar
root committed
202
203
204
            CHECK_RESULT(cuCtxEnablePeerAccess(platformData.contexts[0]->getContext(), 0));
        }
    }
205
206
    defaultQueue = shared_ptr<ComputeQueueImpl>(new CudaQueue(0));
    currentQueue = defaultQueue;
207
208
209
210
211
212
    numAtoms = system.getNumParticles();
    paddedNumAtoms = TileSize*((numAtoms+TileSize-1)/TileSize);
    numAtomBlocks = (paddedNumAtoms+(TileSize-1))/TileSize;
    int multiprocessors;
    CHECK_RESULT(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device));
    numThreadBlocks = numThreadBlocksPerComputeUnit*multiprocessors;
213
    if (cudaDriverVersion >= 9000) {
Peter Eastman's avatar
Peter Eastman committed
214
215
        compilationDefines["SYNC_WARPS"] = "__syncwarp();";
        compilationDefines["SHFL(var, srcLane)"] = "__shfl_sync(0xffffffff, var, srcLane);";
216
        compilationDefines["BALLOT(var)"] = "__ballot_sync(0xffffffff, var);";
Peter Eastman's avatar
Peter Eastman committed
217
218
219
220
    }
    else {
        compilationDefines["SYNC_WARPS"] = "";
        compilationDefines["SHFL(var, srcLane)"] = "__shfl(var, srcLane);";
221
        compilationDefines["BALLOT(var)"] = "__ballot(var);";
Peter Eastman's avatar
Peter Eastman committed
222
    }
223
    if (useDoublePrecision) {
Peter Eastman's avatar
Peter Eastman committed
224
225
        posq.initialize<double4>(*this, paddedNumAtoms, "posq");
        velm.initialize<double4>(*this, paddedNumAtoms, "velm");
226
        compilationDefines["USE_DOUBLE_PRECISION"] = "1";
227
228
229
        compilationDefines["make_real2"] = "make_double2";
        compilationDefines["make_real3"] = "make_double3";
        compilationDefines["make_real4"] = "make_double4";
230
231
232
        compilationDefines["make_mixed2"] = "make_double2";
        compilationDefines["make_mixed3"] = "make_double3";
        compilationDefines["make_mixed4"] = "make_double4";
233
    }
234
    else if (useMixedPrecision) {
Peter Eastman's avatar
Peter Eastman committed
235
236
237
        posq.initialize<float4>(*this, paddedNumAtoms, "posq");
        posqCorrection.initialize<float4>(*this, paddedNumAtoms, "posqCorrection");
        velm.initialize<double4>(*this, paddedNumAtoms, "velm");
238
239
240
241
242
243
244
245
        compilationDefines["USE_MIXED_PRECISION"] = "1";
        compilationDefines["make_real2"] = "make_float2";
        compilationDefines["make_real3"] = "make_float3";
        compilationDefines["make_real4"] = "make_float4";
        compilationDefines["make_mixed2"] = "make_double2";
        compilationDefines["make_mixed3"] = "make_double3";
        compilationDefines["make_mixed4"] = "make_double4";
    }
246
    else {
Peter Eastman's avatar
Peter Eastman committed
247
248
        posq.initialize<float4>(*this, paddedNumAtoms, "posq");
        velm.initialize<float4>(*this, paddedNumAtoms, "velm");
249
250
251
        compilationDefines["make_real2"] = "make_float2";
        compilationDefines["make_real3"] = "make_float3";
        compilationDefines["make_real4"] = "make_float4";
252
253
254
        compilationDefines["make_mixed2"] = "make_float2";
        compilationDefines["make_mixed3"] = "make_float3";
        compilationDefines["make_mixed4"] = "make_float4";
255
    }
256
257
    force.initialize<long long>(*this, paddedNumAtoms*3, "force");
    posCellOffsets.resize(paddedNumAtoms, mm_int4(0, 0, 0, 0));
258
259
260
261
262
    atomIndexDevice.initialize<int>(*this, paddedNumAtoms, "atomIndex");
    atomIndex.resize(paddedNumAtoms);
    for (int i = 0; i < paddedNumAtoms; ++i)
        atomIndex[i] = i;
    atomIndexDevice.upload(atomIndex);
263
264
265
266

    // Create utility kernels that are used in multiple places.

    CUmodule utilities = createModule(CudaKernelSources::vectorOps+CudaKernelSources::utilities);
267
268
269
270
271
272
    clearBufferKernel = getKernel(utilities, "clearBuffer");
    clearTwoBuffersKernel = getKernel(utilities, "clearTwoBuffers");
    clearThreeBuffersKernel = getKernel(utilities, "clearThreeBuffers");
    clearFourBuffersKernel = getKernel(utilities, "clearFourBuffers");
    clearFiveBuffersKernel = getKernel(utilities, "clearFiveBuffers");
    clearSixBuffersKernel = getKernel(utilities, "clearSixBuffers");
Peter Eastman's avatar
Peter Eastman committed
273
    reduceEnergyKernel = getKernel(utilities, "reduceEnergy");
274
    setChargesKernel = getKernel(utilities, "setCharges");
275
276
277
278
279
280
281
282

    // Set defines based on the requested precision.

    compilationDefines["SQRT"] = useDoublePrecision ? "sqrt" : "sqrtf";
    compilationDefines["RSQRT"] = useDoublePrecision ? "rsqrt" : "rsqrtf";
    compilationDefines["RECIP"] = useDoublePrecision ? "1.0/" : "1.0f/";
    compilationDefines["EXP"] = useDoublePrecision ? "exp" : "expf";
    compilationDefines["LOG"] = useDoublePrecision ? "log" : "logf";
283
    compilationDefines["POW"] = useDoublePrecision ? "pow" : "powf";
284
285
286
287
288
289
    compilationDefines["COS"] = useDoublePrecision ? "cos" : "cosf";
    compilationDefines["SIN"] = useDoublePrecision ? "sin" : "sinf";
    compilationDefines["TAN"] = useDoublePrecision ? "tan" : "tanf";
    compilationDefines["ACOS"] = useDoublePrecision ? "acos" : "acosf";
    compilationDefines["ASIN"] = useDoublePrecision ? "asin" : "asinf";
    compilationDefines["ATAN"] = useDoublePrecision ? "atan" : "atanf";
290
291
    compilationDefines["ERF"] = useDoublePrecision ? "erf" : "erff";
    compilationDefines["ERFC"] = useDoublePrecision ? "erfc" : "erfcf";
292

293
    // Set defines for applying periodic boundary conditions.
294

295
296
297
298
299
    Vec3 boxVectors[3];
    system.getDefaultPeriodicBoxVectors(boxVectors[0], boxVectors[1], boxVectors[2]);
    boxIsTriclinic = (boxVectors[0][1] != 0.0 || boxVectors[0][2] != 0.0 ||
                      boxVectors[1][0] != 0.0 || boxVectors[1][2] != 0.0 ||
                      boxVectors[2][0] != 0.0 || boxVectors[2][1] != 0.0);
300
301
302
    for (int i = 0; i < system.getNumForces(); i++)
        if (dynamic_cast<const MonteCarloFlexibleBarostat*>(&system.getForce(i)) != NULL)
            boxIsTriclinic = true;
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
    if (boxIsTriclinic) {
        compilationDefines["APPLY_PERIODIC_TO_DELTA(delta)"] =
            "{"
            "real scale3 = floor(delta.z*invPeriodicBoxSize.z+0.5f); \\\n"
            "delta.x -= scale3*periodicBoxVecZ.x; \\\n"
            "delta.y -= scale3*periodicBoxVecZ.y; \\\n"
            "delta.z -= scale3*periodicBoxVecZ.z; \\\n"
            "real scale2 = floor(delta.y*invPeriodicBoxSize.y+0.5f); \\\n"
            "delta.x -= scale2*periodicBoxVecY.x; \\\n"
            "delta.y -= scale2*periodicBoxVecY.y; \\\n"
            "real scale1 = floor(delta.x*invPeriodicBoxSize.x+0.5f); \\\n"
            "delta.x -= scale1*periodicBoxVecX.x;}";
        compilationDefines["APPLY_PERIODIC_TO_POS(pos)"] =
            "{"
            "real scale3 = floor(pos.z*invPeriodicBoxSize.z); \\\n"
            "pos.x -= scale3*periodicBoxVecZ.x; \\\n"
            "pos.y -= scale3*periodicBoxVecZ.y; \\\n"
            "pos.z -= scale3*periodicBoxVecZ.z; \\\n"
            "real scale2 = floor(pos.y*invPeriodicBoxSize.y); \\\n"
            "pos.x -= scale2*periodicBoxVecY.x; \\\n"
            "pos.y -= scale2*periodicBoxVecY.y; \\\n"
            "real scale1 = floor(pos.x*invPeriodicBoxSize.x); \\\n"
            "pos.x -= scale1*periodicBoxVecX.x;}";
        compilationDefines["APPLY_PERIODIC_TO_POS_WITH_CENTER(pos, center)"] =
            "{"
            "real scale3 = floor((pos.z-center.z)*invPeriodicBoxSize.z+0.5f); \\\n"
            "pos.x -= scale3*periodicBoxVecZ.x; \\\n"
            "pos.y -= scale3*periodicBoxVecZ.y; \\\n"
            "pos.z -= scale3*periodicBoxVecZ.z; \\\n"
            "real scale2 = floor((pos.y-center.y)*invPeriodicBoxSize.y+0.5f); \\\n"
            "pos.x -= scale2*periodicBoxVecY.x; \\\n"
            "pos.y -= scale2*periodicBoxVecY.y; \\\n"
            "real scale1 = floor((pos.x-center.x)*invPeriodicBoxSize.x+0.5f); \\\n"
            "pos.x -= scale1*periodicBoxVecX.x;}";
    }
    else {
        compilationDefines["APPLY_PERIODIC_TO_DELTA(delta)"] =
            "{"
            "delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x; \\\n"
            "delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y; \\\n"
            "delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;}";
        compilationDefines["APPLY_PERIODIC_TO_POS(pos)"] =
            "{"
            "pos.x -= floor(pos.x*invPeriodicBoxSize.x)*periodicBoxSize.x; \\\n"
            "pos.y -= floor(pos.y*invPeriodicBoxSize.y)*periodicBoxSize.y; \\\n"
            "pos.z -= floor(pos.z*invPeriodicBoxSize.z)*periodicBoxSize.z;}";
        compilationDefines["APPLY_PERIODIC_TO_POS_WITH_CENTER(pos, center)"] =
            "{"
            "pos.x -= floor((pos.x-center.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x; \\\n"
            "pos.y -= floor((pos.y-center.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y; \\\n"
            "pos.z -= floor((pos.z-center.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;}";
    }

356
    // Create utilities objects.
357

358
359
    bonded = new CudaBondedUtilities(*this);
    nonbonded = new CudaNonbondedUtilities(*this);
360
361
    integration = new CudaIntegrationUtilities(*this, system);
    expression = new CudaExpressionUtilities(*this);
362
    clearBuffer(posq);
363
364
365
}

CudaContext::~CudaContext() {
366
    pushAsCurrent();
peastman's avatar
peastman committed
367
368
369
370
371
372
373
374
    for (auto force : forces)
        delete force;
    for (auto listener : reorderListeners)
        delete listener;
    for (auto computation : preComputations)
        delete computation;
    for (auto computation : postComputations)
        delete computation;
375
376
    if (pinnedBuffer != NULL)
        cuMemFreeHost(pinnedBuffer);
377
378
379
380
    if (integration != NULL)
        delete integration;
    if (expression != NULL)
        delete expression;
381
382
    if (bonded != NULL)
        delete bonded;
383
384
    if (nonbonded != NULL)
        delete nonbonded;
385
386
    if (contextIsValid && !isLinkedContext)
        cuProfilerStop();
387
    popAsCurrent();
388
    string errorMessage = "Error deleting Context";
389
    if (contextIsValid && !isLinkedContext)
390
        cuCtxDestroy(context);
391
    contextIsValid = false;
392
393
}

394
void CudaContext::initialize() {
395
    ContextSelector selector(*this);
396
    string errorMessage = "Error initializing Context";
397
    int numEnergyBuffers = max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers());
398
399
    int multiprocessors;
    CHECK_RESULT2(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device), "Error checking GPU properties");
400
    if (useDoublePrecision) {
Peter Eastman's avatar
Peter Eastman committed
401
        energyBuffer.initialize<double>(*this, numEnergyBuffers, "energyBuffer");
402
        energySum.initialize<double>(*this, multiprocessors, "energySum");
403
404
405
406
        int pinnedBufferSize = max(paddedNumAtoms*4, numEnergyBuffers);
        CHECK_RESULT(cuMemHostAlloc(&pinnedBuffer, pinnedBufferSize*sizeof(double), 0));
    }
    else if (useMixedPrecision) {
Peter Eastman's avatar
Peter Eastman committed
407
        energyBuffer.initialize<double>(*this, numEnergyBuffers, "energyBuffer");
408
        energySum.initialize<double>(*this, multiprocessors, "energySum");
409
410
411
412
        int pinnedBufferSize = max(paddedNumAtoms*4, numEnergyBuffers);
        CHECK_RESULT(cuMemHostAlloc(&pinnedBuffer, pinnedBufferSize*sizeof(double), 0));
    }
    else {
Peter Eastman's avatar
Peter Eastman committed
413
        energyBuffer.initialize<float>(*this, numEnergyBuffers, "energyBuffer");
414
        energySum.initialize<float>(*this, multiprocessors, "energySum");
415
416
417
        int pinnedBufferSize = max(paddedNumAtoms*6, numEnergyBuffers);
        CHECK_RESULT(cuMemHostAlloc(&pinnedBuffer, pinnedBufferSize*sizeof(float), 0));
    }
418
419
    for (int i = 0; i < numAtoms; i++) {
        double mass = system.getParticleMass(i);
420
        if (useDoublePrecision || useMixedPrecision)
421
422
423
424
            ((double4*) pinnedBuffer)[i] = make_double4(0.0, 0.0, 0.0, mass == 0.0 ? 0.0 : 1.0/mass);
        else
            ((float4*) pinnedBuffer)[i] = make_float4(0.0f, 0.0f, 0.0f, mass == 0.0 ? 0.0f : (float) (1.0/mass));
    }
Peter Eastman's avatar
Peter Eastman committed
425
    velm.upload(pinnedBuffer);
426
    bonded->initialize(system);
Peter Eastman's avatar
Peter Eastman committed
427
428
    addAutoclearBuffer(force.getDevicePointer(), force.getSize()*force.getElementSize());
    addAutoclearBuffer(energyBuffer.getDevicePointer(), energyBuffer.getSize()*energyBuffer.getElementSize());
429
430
431
    int numEnergyParamDerivs = energyParamDerivNames.size();
    if (numEnergyParamDerivs > 0) {
        if (useDoublePrecision || useMixedPrecision)
Peter Eastman's avatar
Peter Eastman committed
432
            energyParamDerivBuffer.initialize<double>(*this, numEnergyParamDerivs*numEnergyBuffers, "energyParamDerivBuffer");
433
        else
Peter Eastman's avatar
Peter Eastman committed
434
435
            energyParamDerivBuffer.initialize<float>(*this, numEnergyParamDerivs*numEnergyBuffers, "energyParamDerivBuffer");
        addAutoclearBuffer(energyParamDerivBuffer);
436
    }
437
    findMoleculeGroups();
438
    nonbonded->initialize(system);
439
}
440

441
442
void CudaContext::initializeContexts() {
    getPlatformData().initializeContexts(system);
443
444
}

445
446
FFT3D CudaContext::createFFT(int xsize, int ysize, int zsize, bool realToComplex) {
    return FFT3D(new CudaFFT3D(*this, xsize, ysize, zsize, realToComplex));
447
448
}

449
450
451
452
453
void CudaContext::setAsCurrent() {
    if (contextIsValid)
        cuCtxSetCurrent(context);
}

454
455
456
457
458
459
460
461
462
463
464
void CudaContext::pushAsCurrent() {
    if (contextIsValid)
        cuCtxPushCurrent(context);
}

void CudaContext::popAsCurrent() {
    CUcontext popped;
    if (contextIsValid)
        cuCtxPopCurrent(&popped);
}

465
466
467
468
469
CUmodule CudaContext::createModule(const string source, const char* optimizationFlags) {
    return createModule(source, map<string, string>(), optimizationFlags);
}

CUmodule CudaContext::createModule(const string source, const map<string, string>& defines, const char* optimizationFlags) {
470
    string bits = intToString(8*sizeof(void*));
471
472
473
474
    string options = (optimizationFlags == NULL ? defaultOptimizationOptions : string(optimizationFlags));
    stringstream src;
    if (!options.empty())
        src << "// Compilation Options: " << options << endl << endl;
peastman's avatar
peastman committed
475
    for (auto& pair : compilationDefines) {
476
477
478
479
480
481
482
        // Query defines to avoid duplicate variables
        if (defines.find(pair.first) == defines.end()) {
            src << "#define " << pair.first;
            if (!pair.second.empty())
                src << " " << pair.second;
            src << endl;
        }
483
484
485
    }
    if (!compilationDefines.empty())
        src << endl;
486
487
488
489
490
491
492
493
494
495
496
497
    if (useDoublePrecision) {
        src << "typedef double real;\n";
        src << "typedef double2 real2;\n";
        src << "typedef double3 real3;\n";
        src << "typedef double4 real4;\n";
    }
    else {
        src << "typedef float real;\n";
        src << "typedef float2 real2;\n";
        src << "typedef float3 real3;\n";
        src << "typedef float4 real4;\n";
    }
498
499
500
501
502
503
504
505
506
507
508
509
    if (useDoublePrecision || useMixedPrecision) {
        src << "typedef double mixed;\n";
        src << "typedef double2 mixed2;\n";
        src << "typedef double3 mixed3;\n";
        src << "typedef double4 mixed4;\n";
    }
    else {
        src << "typedef float mixed;\n";
        src << "typedef float2 mixed2;\n";
        src << "typedef float3 mixed3;\n";
        src << "typedef float4 mixed4;\n";
    }
510
    src << "typedef unsigned int tileflags;\n";
511
    src << CudaKernelSources::common << endl;
peastman's avatar
peastman committed
512
513
514
515
    for (auto& pair : defines) {
        src << "#define " << pair.first;
        if (!pair.second.empty())
            src << " " << pair.second;
516
517
518
519
520
        src << endl;
    }
    if (!defines.empty())
        src << endl;
    src << source << endl;
521
522
    
    // Determine what architecture to compile for.
523
524
525
526

    int maxCompilerArchitecture;
#if CUDA_VERSION < 11020
    // CUDA versions before 11.2 can't query the compiler to see what it supports.
527
    
528
529
530
531
532
533
534
535
536
    maxCompilerArchitecture = 75;
#else
    int numArchs;
    CHECK_NVRTC_RESULT(nvrtcGetNumSupportedArchs(&numArchs), "Error querying supported architectures");
    vector<int> archs(numArchs);
    CHECK_NVRTC_RESULT(nvrtcGetSupportedArchs(archs.data()), "Error querying supported architectures");
    maxCompilerArchitecture = archs.back();
#endif
    string compileArchitecture = intToString(min(gpuArchitecture, maxCompilerArchitecture));
537

538
    // See whether we already have PTX for this kernel cached.
539

540
541
542
543
544
545
    CSHA1 sha1;
    sha1.Update((const UINT_8*) src.str().c_str(), src.str().size());
    sha1.Final();
    UINT_8 hash[20];
    sha1.GetHash(hash);
    stringstream cacheFile;
546
    cacheFile << cacheDir;
547
548
549
    cacheFile.flags(ios::hex);
    for (int i = 0; i < 20; i++)
        cacheFile << setw(2) << setfill('0') << (int) hash[i];
550
    cacheFile << '_' << compileArchitecture << '_' << bits;
551
552
553
    CUmodule module;
    if (cuModuleLoad(&module, cacheFile.str().c_str()) == CUDA_SUCCESS)
        return module;
554

555
    // Select a name for the output file.
556

557
558
    stringstream tempFileName;
    tempFileName << "openmmTempKernel" << this; // Include a pointer to this context as part of the filename to avoid collisions.
559
    tempFileName << "_" << this_thread::get_id();
560
    string outputFile = (tempDir+tempFileName.str()+".ptx");
561

562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    // Split the command line flags into an array of options.
    
    string flags = "-arch=compute_"+compileArchitecture+" "+options;
    stringstream flagsStream(flags);
    string flag;
    vector<string> splitFlags;
    while (flagsStream >> flag)
        splitFlags.push_back(flag);
    int numOptions = splitFlags.size();
    vector<const char*> optionsVec(numOptions);
    for (int i = 0; i < numOptions; i++)
        optionsVec[i] = &splitFlags[i][0];
    
    // Compile the program to PTX.
    
    nvrtcProgram program;
    CHECK_NVRTC_RESULT(nvrtcCreateProgram(&program, src.str().c_str(), NULL, 0, NULL, NULL), "Error creating program");
    try {
        nvrtcResult result = nvrtcCompileProgram(program, optionsVec.size(), &optionsVec[0]);
        if (result != NVRTC_SUCCESS) {
            size_t logSize;
            nvrtcGetProgramLogSize(program, &logSize);
            vector<char> log(logSize);
            nvrtcGetProgramLog(program, &log[0]);
            throw OpenMMException("Error compiling program: "+string(&log[0]));
        }
        size_t ptxSize;
        nvrtcGetPTXSize(program, &ptxSize);
        vector<char> ptx(ptxSize);
        nvrtcGetPTX(program, &ptx[0]);
        nvrtcDestroyProgram(&program);
593

594
        // If possible, write the PTX out to a temporary file so we can cache it for later use.
595

596
        bool wroteCache = false;
597
598
        try {
            ofstream out(outputFile.c_str());
599
            out << string(&ptx[0]);
600
            out.close();
601
602
            if (!out.fail())
                wroteCache = true;
603
604
        }
        catch (...) {
605
606
607
608
            // Ignore.
        }
        if (!wroteCache) {
            // An error occurred.  Possibly we don't have permission to write to the temp directory.  Just try to load the module directly.
609

610
611
612
613
            CHECK_RESULT2(cuModuleLoadDataEx(&module, &ptx[0], 0, NULL, NULL), "Error loading CUDA module");
            return module;
        }
    }
614
615
616
    catch (...) {
        nvrtcDestroyProgram(&program);
        throw;
617
    }
618
619
620
621
    try {
        CUresult result = cuModuleLoad(&module, outputFile.c_str());
        if (result != CUDA_SUCCESS) {
            std::stringstream m;
622
            m<<"Error loading CUDA module: "<<getErrorString(result)<<" ("<<result<<")";
623
624
            throw OpenMMException(m.str());
        }
625
626
        if (rename(outputFile.c_str(), cacheFile.str().c_str()) != 0)
            remove(outputFile.c_str());
627
628
629
630
631
632
633
        return module;
    }
    catch (...) {
        remove(outputFile.c_str());
        throw;
    }
}
634
635
636
637
638
639
640
641
642
643
644
645

CUfunction CudaContext::getKernel(CUmodule& module, const string& name) {
    CUfunction function;
    CUresult result = cuModuleGetFunction(&function, module, name.c_str());
    if (result != CUDA_SUCCESS) {
        std::stringstream m;
        m<<"Error creating kernel "<<name<<": "<<getErrorString(result)<<" ("<<result<<")";
        throw OpenMMException(m.str());
    }
    return function;
}

646
647
648
649
650
651
652
vector<ComputeContext*> CudaContext::getAllContexts() {
    vector<ComputeContext*> result;
    for (CudaContext* c : platformData.contexts)
        result.push_back(c);
    return result;
}

Peter Eastman's avatar
Peter Eastman committed
653
654
655
656
double& CudaContext::getEnergyWorkspace() {
    return platformData.contextEnergy[contextIndex];
}

657
658
ComputeQueue CudaContext::createQueue() {
    return shared_ptr<ComputeQueueImpl>(new CudaQueue());
659
660
}

661
662
CUstream CudaContext::getCurrentStream() {
    return dynamic_cast<CudaQueue*>(currentQueue.get())->getStream();
663
664
}

665
666
CudaArray* CudaContext::createArray() {
    return new CudaArray();
667
668
}

669
670
671
672
ComputeEvent CudaContext::createEvent() {
    return shared_ptr<ComputeEventImpl>(new CudaEvent(*this));
}

673
674
675
676
ComputeSort CudaContext::createSort(ComputeSortImpl::SortTrait* trait, unsigned int length, bool uniform) {
    return shared_ptr<ComputeSortImpl>(new CudaSort(*this, trait, length, uniform));
}

677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
ComputeProgram CudaContext::compileProgram(const std::string source, const std::map<std::string, std::string>& defines) {
    CUmodule module = createModule(CudaKernelSources::vectorOps+source, defines);
    return shared_ptr<ComputeProgramImpl>(new CudaProgram(*this, module));
}

CudaArray& CudaContext::unwrap(ArrayInterface& array) const {
    CudaArray* cuarray;
    ComputeArray* wrapper = dynamic_cast<ComputeArray*>(&array);
    if (wrapper != NULL)
        cuarray = dynamic_cast<CudaArray*>(&wrapper->getArray());
    else
        cuarray = dynamic_cast<CudaArray*>(&array);
    if (cuarray == NULL)
        throw OpenMMException("Array argument is not an CudaArray");
    return *cuarray;
692
693
694
}

std::string CudaContext::getErrorString(CUresult result) {
Peter Eastman's avatar
Peter Eastman committed
695
696
697
    const char* message;
    if (cuGetErrorName(result, &message) == CUDA_SUCCESS)
        return string(message);
peastman's avatar
peastman committed
698
    return "CUDA error";
699
700
701
702
703
704
}

void CudaContext::executeKernel(CUfunction kernel, void** arguments, int threads, int blockSize, unsigned int sharedSize) {
    if (blockSize == -1)
        blockSize = ThreadBlockSize;
    int gridSize = std::min((threads+blockSize-1)/blockSize, numThreadBlocks);
705
    CUresult result = cuLaunchKernel(kernel, gridSize, 1, 1, blockSize, 1, 1, sharedSize, getCurrentStream(), arguments, NULL);
706
707
708
709
710
711
712
    if (result != CUDA_SUCCESS) {
        stringstream str;
        str<<"Error invoking kernel: "<<getErrorString(result)<<" ("<<result<<")";
        throw OpenMMException(str.str());
    }
}

713
714
715
int CudaContext::computeThreadBlockSize(double memory) const {
    int maxShared;
    CHECK_RESULT2(cuDeviceGetAttribute(&maxShared, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK, device), "Error querying device property");
716
717
718
719
720
721
722
723
724
    int max = (int) (maxShared/memory);
    if (max < 64)
        return 32;
    int threads = 64;
    while (threads+64 < max)
        threads += 64;
    return threads;
}

725
726
void CudaContext::clearBuffer(ArrayInterface& array) {
    clearBuffer(unwrap(array).getDevicePointer(), array.getSize()*array.getElementSize());
727
728
729
}

void CudaContext::clearBuffer(CUdeviceptr memory, int size) {
730
731
    int words = size/4;
    void* args[] = {&memory, &words};
Peter Eastman's avatar
Peter Eastman committed
732
    executeKernel(clearBufferKernel, args, words, 128);
733
734
}

735
736
void CudaContext::addAutoclearBuffer(ArrayInterface& array) {
    addAutoclearBuffer(unwrap(array).getDevicePointer(), array.getSize()*array.getElementSize());
737
738
}

739
740
void CudaContext::addAutoclearBuffer(CUdeviceptr memory, int size) {
    autoclearBuffers.push_back(memory);
741
    autoclearBufferSizes.push_back(size/4);
742
743
}

744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
void CudaContext::clearAutoclearBuffers() {
    int base = 0;
    int total = autoclearBufferSizes.size();
    while (total-base >= 6) {
        void* args[] = {&autoclearBuffers[base], &autoclearBufferSizes[base],
                        &autoclearBuffers[base+1], &autoclearBufferSizes[base+1],
                        &autoclearBuffers[base+2], &autoclearBufferSizes[base+2],
                        &autoclearBuffers[base+3], &autoclearBufferSizes[base+3],
                        &autoclearBuffers[base+4], &autoclearBufferSizes[base+4],
                        &autoclearBuffers[base+5], &autoclearBufferSizes[base+5]};
        executeKernel(clearSixBuffersKernel, args, max(max(max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), autoclearBufferSizes[base+4]), autoclearBufferSizes[base+5]), 128);
        base += 6;
    }
    if (total-base == 5) {
        void* args[] = {&autoclearBuffers[base], &autoclearBufferSizes[base],
                        &autoclearBuffers[base+1], &autoclearBufferSizes[base+1],
                        &autoclearBuffers[base+2], &autoclearBufferSizes[base+2],
                        &autoclearBuffers[base+3], &autoclearBufferSizes[base+3],
                        &autoclearBuffers[base+4], &autoclearBufferSizes[base+4]};
        executeKernel(clearFiveBuffersKernel, args, max(max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), autoclearBufferSizes[base+4]), 128);
    }
    else if (total-base == 4) {
        void* args[] = {&autoclearBuffers[base], &autoclearBufferSizes[base],
                        &autoclearBuffers[base+1], &autoclearBufferSizes[base+1],
                        &autoclearBuffers[base+2], &autoclearBufferSizes[base+2],
                        &autoclearBuffers[base+3], &autoclearBufferSizes[base+3]};
        executeKernel(clearFourBuffersKernel, args, max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), 128);
    }
    else if (total-base == 3) {
        void* args[] = {&autoclearBuffers[base], &autoclearBufferSizes[base],
                        &autoclearBuffers[base+1], &autoclearBufferSizes[base+1],
                        &autoclearBuffers[base+2], &autoclearBufferSizes[base+2]};
        executeKernel(clearThreeBuffersKernel, args, max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), 128);
    }
    else if (total-base == 2) {
        void* args[] = {&autoclearBuffers[base], &autoclearBufferSizes[base],
                        &autoclearBuffers[base+1], &autoclearBufferSizes[base+1]};
        executeKernel(clearTwoBuffersKernel, args, max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), 128);
    }
    else if (total-base == 1) {
        clearBuffer(autoclearBuffers[base], autoclearBufferSizes[base]*4);
    }
}
787

Peter Eastman's avatar
Peter Eastman committed
788
double CudaContext::reduceEnergy() {
Peter Eastman's avatar
Peter Eastman committed
789
    int bufferSize = energyBuffer.getSize();
Peter Eastman's avatar
Peter Eastman committed
790
    int workGroupSize  = 512;
Peter Eastman's avatar
Peter Eastman committed
791
    void* args[] = {&energyBuffer.getDevicePointer(), &energySum.getDevicePointer(), &bufferSize, &workGroupSize};
792
    executeKernel(reduceEnergyKernel, args, workGroupSize*energySum.getSize(), workGroupSize, workGroupSize*energyBuffer.getElementSize());
Peter Eastman's avatar
Peter Eastman committed
793
    energySum.download(pinnedBuffer);
794
795
796
797
798
799
800
801
802
803
    double result = 0;
    if (getUseDoublePrecision() || getUseMixedPrecision()) {
        for (int i = 0; i < energySum.getSize(); i++)
            result += ((double*) pinnedBuffer)[i];
    }
    else {
        for (int i = 0; i < energySum.getSize(); i++)
            result += ((float*) pinnedBuffer)[i];
    }
    return result;
Peter Eastman's avatar
Peter Eastman committed
804
805
}

806
void CudaContext::setCharges(const vector<double>& charges) {
Peter Eastman's avatar
Peter Eastman committed
807
808
    if (!chargeBuffer.isInitialized())
        chargeBuffer.initialize(*this, numAtoms, useDoublePrecision ? sizeof(double) : sizeof(float), "chargeBuffer");
Peter Eastman's avatar
Peter Eastman committed
809
810
811
812
    vector<double> c(numAtoms);
    for (int i = 0; i < numAtoms; i++)
        c[i] = charges[i];
    chargeBuffer.upload(c, true);
Peter Eastman's avatar
Peter Eastman committed
813
    void* args[] = {&chargeBuffer.getDevicePointer(), &posq.getDevicePointer(), &atomIndexDevice.getDevicePointer(), &numAtoms};
814
815
816
    executeKernel(setChargesKernel, args, numAtoms);
}

817
818
819
820
821
822
bool CudaContext::requestPosqCharges() {
    bool allow = !hasAssignedPosqCharges;
    hasAssignedPosqCharges = true;
    return allow;
}

823
824
825
826
827
828
829
830
831
void CudaContext::addEnergyParameterDerivative(const string& param) {
    // See if this parameter has already been registered.
    
    for (int i = 0; i < energyParamDerivNames.size(); i++)
        if (param == energyParamDerivNames[i])
            return;
    energyParamDerivNames.push_back(param);
}

832
833
void CudaContext::flushQueue() {
    cuStreamSynchronize(getCurrentStream());
834
835
}

836
837
838
839
840
841
842
843
844
845
vector<int> CudaContext::getDevicePrecedence() {
    int numDevices;
    CUdevice thisDevice;
    string errorMessage = "Error initializing Context";
    vector<pair<pair<int, int>, int> > devices;

    CHECK_RESULT(cuDeviceGetCount(&numDevices));
    for (int i = 0; i < numDevices; i++) {
        CHECK_RESULT(cuDeviceGet(&thisDevice, i));
        int major, minor, clock, multiprocessors, speed;
846
847
        CHECK_RESULT(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, thisDevice));
        CHECK_RESULT(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, thisDevice));
848
849
850
        if (major == 1 && minor < 2)
            continue;

Robert T. McGibbon's avatar
Robert T. McGibbon committed
851
        if ((useDoublePrecision || useMixedPrecision) && (major+0.1*minor < 1.3))
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
            continue;

        CHECK_RESULT(cuDeviceGetAttribute(&clock, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, thisDevice));
        CHECK_RESULT(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, thisDevice));
        speed = clock*multiprocessors;
        pair<int, int> deviceProperties = std::make_pair(major, speed);
        devices.push_back(std::make_pair(deviceProperties, -i));
    }

    // sort first by compute capability (higher is better), then speed
    // (higher is better), and finally device index (lower is better)
    std::sort(devices.begin(), devices.end());
    std::reverse(devices.begin(), devices.end());

    vector<int> precedence;
    for (int i = 0; i < static_cast<int>(devices.size()); i++) {
        precedence.push_back(-devices[i].second);
    }

    return precedence;
}
873
874
875
876
877
878
879

unsigned int CudaContext::getEventFlags() {
    unsigned int flags = CU_EVENT_DISABLE_TIMING;
    if (useBlockingSync)
        flags += CU_EVENT_BLOCKING_SYNC;
    return flags;
}