OpenCLContext.cpp 54.2 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) 2009-2012 Stanford University and the Authors.      *
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
 * 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/>.      *
 * -------------------------------------------------------------------------- */

27
28
29
30
#ifdef WIN32
  #define _USE_MATH_DEFINES // Needed to get M_PI
#endif
#include <cmath>
31
32
#include "OpenCLContext.h"
#include "OpenCLArray.h"
Peter Eastman's avatar
Peter Eastman committed
33
#include "OpenCLBondedUtilities.h"
34
#include "OpenCLForceInfo.h"
35
#include "OpenCLIntegrationUtilities.h"
36
#include "OpenCLKernelSources.h"
37
#include "OpenCLNonbondedUtilities.h"
38
#include "hilbert.h"
39
#include "openmm/Platform.h"
40
#include "openmm/System.h"
41
#include "openmm/VirtualSite.h"
Peter Eastman's avatar
Peter Eastman committed
42
#include <algorithm>
43
44
#include <fstream>
#include <iostream>
45
#include <sstream>
46
#include <typeinfo>
47
48

using namespace OpenMM;
49
using namespace std;
50

51
52
53
#ifndef CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV
  #define CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV 0x4000
#endif
54
55
56
#ifndef CL_DEVICE_COMPUTE_CAPABILITY_MINOR_NV
  #define CL_DEVICE_COMPUTE_CAPABILITY_MINOR_NV 0x4001
#endif
57

58
59
60
const int OpenCLContext::ThreadBlockSize = 64;
const int OpenCLContext::TileSize = 32;

61
static void CL_CALLBACK errorCallback(const char* errinfo, const void* private_info, size_t cb, void* user_data) {
62
63
64
    string skip = "OpenCL Build Warning : Compiler build log:";
    if (strncmp(errinfo, skip.c_str(), skip.length()) == 0)
        return; // OS X Lion insists on calling this for every build warning, even though they aren't errors.
65
66
67
    std::cerr << "OpenCL internal error: " << errinfo << std::endl;
}

68
OpenCLContext::OpenCLContext(const System& system, int platformIndex, int deviceIndex, const string& precision, OpenCLPlatform::PlatformData& platformData) :
69
        system(system), time(0.0), platformData(platformData), stepCount(0), computeForceCount(0), stepsSinceReorder(99999), atomsWereReordered(false), posq(NULL),
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
70
        posqCorrection(NULL), velm(NULL), forceBuffers(NULL), longForceBuffer(NULL), energyBuffer(NULL), atomIndexDevice(NULL), integration(NULL),
71
        expression(NULL), bonded(NULL), nonbonded(NULL), thread(NULL) {
72
73
74
75
76
77
78
79
80
81
82
83
84
85
    if (precision == "single") {
        useDoublePrecision = false;
        useMixedPrecision = false;
    }
    else if (precision == "mixed") {
        useDoublePrecision = false;
        useMixedPrecision = true;
    }
    else if (precision == "double") {
        useDoublePrecision = true;
        useMixedPrecision = false;
    }
    else
        throw OpenMMException("Illegal value for OpenCLPrecision: "+precision);
86
    try {
87
        contextIndex = platformData.contexts.size();
88
89
        std::vector<cl::Platform> platforms;
        cl::Platform::get(&platforms);
90
        if (platformIndex < 0 || platformIndex >= (int) platforms.size())
91
            throw OpenMMException("Illegal value for OpenCL platform index");
Peter Eastman's avatar
Peter Eastman committed
92
        string platformVendor = platforms[platformIndex].getInfo<CL_PLATFORM_VENDOR>();
93
94
        vector<cl::Device> devices;
        platforms[platformIndex].getDevices(CL_DEVICE_TYPE_ALL, &devices);
95
        const int minThreadBlockSize = 32;
96
        if (deviceIndex < 0 || deviceIndex >= (int) devices.size()) {
97
            // Try to figure out which device is the fastest.
98

99
            int bestSpeed = -1;
100
            for (int i = 0; i < (int) devices.size(); i++) {
Peter Eastman's avatar
Peter Eastman committed
101
                if (platformVendor == "Apple" && devices[i].getInfo<CL_DEVICE_VENDOR>() == "AMD")
Peter Eastman's avatar
Peter Eastman committed
102
                    continue; // Don't use AMD GPUs on OS X due to serious bugs.
103
                int maxSize = devices[i].getInfo<CL_DEVICE_MAX_WORK_ITEM_SIZES>()[0];
104
105
106
107
108
                int processingElementsPerComputeUnit = 8;
                if (devices[i].getInfo<CL_DEVICE_TYPE>() != CL_DEVICE_TYPE_GPU) {
                    processingElementsPerComputeUnit = 1;
                }
                else if (devices[i].getInfo<CL_DEVICE_EXTENSIONS>().find("cl_nv_device_attribute_query") != string::npos) {
109
110
111
112
                    cl_uint computeCapabilityMajor;
                    clGetDeviceInfo(devices[i](), CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV, sizeof(cl_uint), &computeCapabilityMajor, NULL);
                    processingElementsPerComputeUnit = (computeCapabilityMajor < 2 ? 8 : 32);
                }
113
114
115
116
                else if (devices[i].getInfo<CL_DEVICE_EXTENSIONS>().find("cl_amd_device_attribute_query") != string::npos) {
                    // This attribute does not ensure that all queries are supported by the runtime (it may be an older runtime,
                    // or the CPU device) so still have to check for errors.
                    try {
117
#ifdef CL_DEVICE_SIMD_WIDTH_AMD
118
119
120
121
122
123
124
125
126
127
128
129
130
                        processingElementsPerComputeUnit =
                            // AMD GPUs either have a single VLIW SIMD or multiple scalar SIMDs.
                            // The SIMD width is the number of threads the SIMD executes per cycle.
                            // This will be less than the wavefront width since it takes several
                            // cycles to execute the full wavefront.
                            // The SIMD instruction width is the VLIW instruction width (or 1 for scalar),
                            // this is the number of ALUs that can be executing per instruction per thread. 
                            devices[i].getInfo<CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD>() *
                            devices[i].getInfo<CL_DEVICE_SIMD_WIDTH_AMD>() *
                            devices[i].getInfo<CL_DEVICE_SIMD_INSTRUCTION_WIDTH_AMD>();
                        // Just in case any of the queries return 0.
                        if (processingElementsPerComputeUnit <= 0)
                            processingElementsPerComputeUnit = 1;
131
#endif
132
133
134
135
136
                    }
                    catch (cl::Error err) {
                        // Runtime does not support the queries so use default.
                    }
                }
137
                int speed = devices[i].getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>()*processingElementsPerComputeUnit*devices[i].getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>();
138
                if (maxSize >= minThreadBlockSize && speed > bestSpeed) {
139
                    deviceIndex = i;
140
141
                    bestSpeed = speed;
                }
142
            }
143
        }
144
145
146
        if (deviceIndex == -1)
            throw OpenMMException("No compatible OpenCL device is available");
        device = devices[deviceIndex];
Peter Eastman's avatar
Peter Eastman committed
147
        this->deviceIndex = deviceIndex;
148
        if (device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>() < minThreadBlockSize)
149
            throw OpenMMException("The specified OpenCL device is not compatible with OpenMM");
150
        compilationDefines["WORK_GROUP_SIZE"] = intToString(ThreadBlockSize);
Peter Eastman's avatar
Peter Eastman committed
151
        if (platformVendor.size() >= 5 && platformVendor.substr(0, 5) == "Intel")
152
153
154
            defaultOptimizationOptions = "";
        else
            defaultOptimizationOptions = "-cl-fast-relaxed-math";
155
        supports64BitGlobalAtomics = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_int64_base_atomics") != string::npos);
156
        supportsDoublePrecision = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp64") != string::npos);
157
158
        if ((useDoublePrecision || useMixedPrecision) && !supportsDoublePrecision)
            throw OpenMMException("This device does not support double precision");
159
        string vendor = device.getInfo<CL_DEVICE_VENDOR>();
160
        int numThreadBlocksPerComputeUnit = 6;
161
        if (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA") {
162
            compilationDefines["WARPS_ARE_ATOMIC"] = "";
163
            simdWidth = 32;
164
165
            if (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_nv_device_attribute_query") != string::npos) {
                // Compute level 1.2 and later Nvidia GPUs support 64 bit atomics, even though they don't list the
166
167
                // proper extension as supported.  We only use them on compute level 2.0 or later, since they're very
                // slow on earlier GPUs.
168

169
                cl_uint computeCapabilityMajor;
170
                clGetDeviceInfo(device(), CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV, sizeof(cl_uint), &computeCapabilityMajor, NULL);
171
                if (computeCapabilityMajor > 1)
172
173
                    supports64BitGlobalAtomics = true;
            }
174
        }
175
        else if (vendor.size() >= 28 && vendor.substr(0, 28) == "Advanced Micro Devices, Inc.") {
176
177
178
179
180
181
182
183
184
185
186
187
188
            if (device.getInfo<CL_DEVICE_TYPE>() != CL_DEVICE_TYPE_GPU) {
                /// \todo Is 6 a good value for the OpenCL CPU device?
                // numThreadBlocksPerComputeUnit = ?;
                simdWidth = 1;
            }
            else {
                bool amdPostSdk2_4 = false;
                // Default to 1 which will use the default kernels.
                simdWidth = 1;
                if (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_amd_device_attribute_query") != string::npos) {
                    // This attribute does not ensure that all queries are supported by the runtime so still have to
                    // check for errors.
                    try {
Peter Eastman's avatar
Peter Eastman committed
189
#ifdef CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD
190
191
192
                        // Must catch cl:Error as will fail if runtime does not support queries.

                        cl_uint simdPerComputeUnit = device.getInfo<CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD>();
193
194
                        simdWidth = device.getInfo<CL_DEVICE_WAVEFRONT_WIDTH_AMD>();

195
196
197
198
199
200
201
202
                        // If the GPU has multiple SIMDs per compute unit then it is uses the scalar instruction
                        // set instead of the VLIW instruction set. It therefore needs more thread blocks per
                        // compute unit to hide memory latency.
                        if (simdPerComputeUnit > 1)
                            numThreadBlocksPerComputeUnit = 4 * simdPerComputeUnit;

                        // If the queries are supported then must be newer than SDK 2.4.
                        amdPostSdk2_4 = true;
Peter Eastman's avatar
Peter Eastman committed
203
#endif
204
205
206
207
208
209
210
211
212
213
                    }
                    catch (cl::Error err) {
                        // Runtime does not support the query so is unlikely to be the newer scalar GPU.
                        // Stay with the default simdWidth and numThreadBlocksPerComputeUnit.
                    }
                }
                // AMD APP SDK 2.4 has a performance problem with atomics. Enable the work around. This is fixed after SDK 2.4.
                if (!amdPostSdk2_4)
                    compilationDefines["AMD_ATOMIC_WORK_AROUND"] = "";
            }
214
        }
215
216
        else
            simdWidth = 1;
Peter Eastman's avatar
Peter Eastman committed
217
        if (platformVendor == "Apple" && vendor == "AMD")
218
            compilationDefines["MAC_AMD_WORKAROUND"] = "";
219
        if (supports64BitGlobalAtomics)
220
            compilationDefines["SUPPORTS_64_BIT_ATOMICS"] = "";
221
222
        if (supportsDoublePrecision)
            compilationDefines["SUPPORTS_DOUBLE_PRECISION"] = "";
223
224
225
226
        if (simdWidth >= 32)
            compilationDefines["SYNC_WARPS"] = "";
        else
            compilationDefines["SYNC_WARPS"] = "barrier(CLK_LOCAL_MEM_FENCE)";
227
228
229
230
        vector<cl::Device> contextDevices;
        contextDevices.push_back(device);
        cl_context_properties cprops[] = {CL_CONTEXT_PLATFORM, (cl_context_properties) platforms[platformIndex](), 0};
        context = cl::Context(contextDevices, cprops, errorCallback);
231
        queue = cl::CommandQueue(context, device);
Peter Eastman's avatar
Peter Eastman committed
232
233
        numAtoms = system.getNumParticles();
        paddedNumAtoms = TileSize*((numAtoms+TileSize-1)/TileSize);
234
        numAtomBlocks = (paddedNumAtoms+(TileSize-1))/TileSize;
235
        numThreadBlocks = numThreadBlocksPerComputeUnit*device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
Peter Eastman's avatar
Peter Eastman committed
236
        bonded = new OpenCLBondedUtilities(*this);
237
        nonbonded = new OpenCLNonbondedUtilities(*this);
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
        if (useDoublePrecision) {
            posq = OpenCLArray::create<mm_double4>(*this, paddedNumAtoms, "posq");
            velm = OpenCLArray::create<mm_double4>(*this, paddedNumAtoms, "velm");
            compilationDefines["USE_DOUBLE_PRECISION"] = "1";
            compilationDefines["convert_real4"] = "convert_double4";
            compilationDefines["convert_mixed4"] = "convert_double4";
        }
        else if (useMixedPrecision) {
            posq = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms, "posq");
            posqCorrection = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms, "posq");
            velm = OpenCLArray::create<mm_double4>(*this, paddedNumAtoms, "velm");
            compilationDefines["USE_MIXED_PRECISION"] = "1";
            compilationDefines["convert_real4"] = "convert_float4";
            compilationDefines["convert_mixed4"] = "convert_double4";
        }
        else {
            posq = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms, "posq");
            velm = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms, "velm");
            compilationDefines["convert_real4"] = "convert_float4";
            compilationDefines["convert_mixed4"] = "convert_float4";
        }
259
        posCellOffsets.resize(paddedNumAtoms, mm_int4(0, 0, 0, 0));
260
261
262
263
264
    }
    catch (cl::Error err) {
        std::stringstream str;
        str<<"Error initializing context: "<<err.what()<<" ("<<err.err()<<")";
        throw OpenMMException(str.str());
265
    }
266
267
268

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

Peter Eastman's avatar
Peter Eastman committed
269
    cl::Program utilities = createProgram(OpenCLKernelSources::utilities);
270
    clearBufferKernel = cl::Kernel(utilities, "clearBuffer");
271
272
273
    clearTwoBuffersKernel = cl::Kernel(utilities, "clearTwoBuffers");
    clearThreeBuffersKernel = cl::Kernel(utilities, "clearThreeBuffers");
    clearFourBuffersKernel = cl::Kernel(utilities, "clearFourBuffers");
274
275
    clearFiveBuffersKernel = cl::Kernel(utilities, "clearFiveBuffers");
    clearSixBuffersKernel = cl::Kernel(utilities, "clearSixBuffers");
276
    reduceReal4Kernel = cl::Kernel(utilities, "reduceReal4Buffer");
277
278
    if (supports64BitGlobalAtomics)
        reduceForcesKernel = cl::Kernel(utilities, "reduceForces");
279
280
281

    // Decide whether native_sqrt(), native_rsqrt(), and native_recip() are sufficiently accurate to use.

282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    if (!useDoublePrecision) {
        cl::Kernel accuracyKernel(utilities, "determineNativeAccuracy");
        OpenCLArray valuesArray(*this, 20, sizeof(mm_float8), "values");
        vector<mm_float8> values(valuesArray.getSize());
        float nextValue = 1e-4f;
        for (int i = 0; i < (int) values.size(); ++i) {
            values[i].s0 = nextValue;
            nextValue *= (float) M_PI;
        }
        valuesArray.upload(values);
        accuracyKernel.setArg<cl::Buffer>(0, valuesArray.getDeviceBuffer());
        accuracyKernel.setArg<cl_int>(1, values.size());
        executeKernel(accuracyKernel, values.size());
        valuesArray.download(values);
        double maxSqrtError = 0.0, maxRsqrtError = 0.0, maxRecipError = 0.0, maxExpError = 0.0, maxLogError = 0.0;
        for (int i = 0; i < (int) values.size(); ++i) {
            double v = values[i].s0;
            double correctSqrt = sqrt(v);
            maxSqrtError = max(maxSqrtError, fabs(correctSqrt-values[i].s1)/correctSqrt);
            maxRsqrtError = max(maxRsqrtError, fabs(1.0/correctSqrt-values[i].s2)*correctSqrt);
            maxRecipError = max(maxRecipError, fabs(1.0/v-values[i].s3)/values[i].s3);
            maxExpError = max(maxExpError, fabs(exp(v)-values[i].s4)/values[i].s4);
            maxLogError = max(maxLogError, fabs(log(v)-values[i].s5)/values[i].s5);
        }
        compilationDefines["SQRT"] = (maxSqrtError < 1e-6) ? "native_sqrt" : "sqrt";
        compilationDefines["RSQRT"] = (maxRsqrtError < 1e-6) ? "native_rsqrt" : "rsqrt";
        compilationDefines["RECIP"] = (maxRecipError < 1e-6) ? "native_recip" : "1.0f/";
        compilationDefines["EXP"] = (maxExpError < 1e-6) ? "native_exp" : "exp";
        compilationDefines["LOG"] = (maxLogError < 1e-6) ? "native_log" : "log";
    }
    else {
        compilationDefines["SQRT"] = "sqrt";
        compilationDefines["RSQRT"] = "rsqrt";
        compilationDefines["RECIP"] = "1.0/";
        compilationDefines["EXP"] = "exp";
        compilationDefines["LOG"] = "log";
    }
319
320
321
322
    
    // Create the work thread used for parallelization when running on multiple devices.
    
    thread = new WorkThread();
Peter Eastman's avatar
Peter Eastman committed
323
    
324
    // Create utilities objects.
Peter Eastman's avatar
Peter Eastman committed
325
326
    
    integration = new OpenCLIntegrationUtilities(*this, system);
327
    expression = new OpenCLExpressionUtilities(*this);
328
329
330
}

OpenCLContext::~OpenCLContext() {
331
332
    for (int i = 0; i < (int) forces.size(); i++)
        delete forces[i];
333
334
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        delete reorderListeners[i];
335
336
337
338
    for (int i = 0; i < (int) preComputations.size(); i++)
        delete preComputations[i];
    for (int i = 0; i < (int) postComputations.size(); i++)
        delete postComputations[i];
339
340
    if (pinnedBuffer != NULL)
        delete pinnedBuffer;
341
342
    if (posq != NULL)
        delete posq;
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
343
344
    if (posqCorrection != NULL)
        delete posqCorrection;
345
346
347
348
349
350
    if (velm != NULL)
        delete velm;
    if (force != NULL)
        delete force;
    if (forceBuffers != NULL)
        delete forceBuffers;
351
352
    if (longForceBuffer != NULL)
        delete longForceBuffer;
353
354
    if (energyBuffer != NULL)
        delete energyBuffer;
355
356
    if (atomIndexDevice != NULL)
        delete atomIndexDevice;
357
358
    if (integration != NULL)
        delete integration;
359
360
    if (expression != NULL)
        delete expression;
Peter Eastman's avatar
Peter Eastman committed
361
362
    if (bonded != NULL)
        delete bonded;
363
364
    if (nonbonded != NULL)
        delete nonbonded;
365
366
    if (thread != NULL)
        delete thread;
367
368
}

369
void OpenCLContext::initialize() {
Peter Eastman's avatar
Peter Eastman committed
370
    bonded->initialize(system);
371
    numForceBuffers = platformData.contexts.size();
Peter Eastman's avatar
Peter Eastman committed
372
    numForceBuffers = std::max(numForceBuffers, bonded->getNumForceBuffers());
373
374
    for (int i = 0; i < (int) forces.size(); i++)
        numForceBuffers = std::max(numForceBuffers, forces[i]->getRequiredForceBuffers());
375
376
377
378
379
380
381
382
383
384
    if (useDoublePrecision) {
        forceBuffers = OpenCLArray::create<mm_double4>(*this, paddedNumAtoms*numForceBuffers, "forceBuffers");
        force = OpenCLArray::create<mm_double4>(*this, &forceBuffers->getDeviceBuffer(), paddedNumAtoms, "force");
        energyBuffer = OpenCLArray::create<cl_double>(*this, max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers()), "energyBuffer");
    }
    else {
        forceBuffers = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms*numForceBuffers, "forceBuffers");
        force = OpenCLArray::create<mm_float4>(*this, &forceBuffers->getDeviceBuffer(), paddedNumAtoms, "force");
        energyBuffer = OpenCLArray::create<cl_float>(*this, max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers()), "energyBuffer");
    }
385
    if (supports64BitGlobalAtomics) {
386
        longForceBuffer = OpenCLArray::create<cl_long>(*this, 3*paddedNumAtoms, "longForceBuffer");
387
388
389
390
        reduceForcesKernel.setArg<cl::Buffer>(0, longForceBuffer->getDeviceBuffer());
        reduceForcesKernel.setArg<cl::Buffer>(1, forceBuffers->getDeviceBuffer());
        reduceForcesKernel.setArg<cl_int>(2, paddedNumAtoms);
        reduceForcesKernel.setArg<cl_int>(3, numForceBuffers);
391
        addAutoclearBuffer(*longForceBuffer);
392
    }
393
394
    addAutoclearBuffer(*forceBuffers);
    addAutoclearBuffer(*energyBuffer);
395
    int bufferBytes = max(velm->getSize()*velm->getElementSize(), energyBuffer->getSize()*energyBuffer->getElementSize());
396
397
    pinnedBuffer = new cl::Buffer(context, CL_MEM_ALLOC_HOST_PTR, bufferBytes);
    pinnedMemory = queue.enqueueMapBuffer(*pinnedBuffer, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, bufferBytes);
398
399
400
401
402
403
404
405
    for (int i = 0; i < numAtoms; i++) {
        double mass = system.getParticleMass(i);
        if (useDoublePrecision || useMixedPrecision)
            ((mm_double4*) pinnedMemory)[i] = mm_double4(0.0, 0.0, 0.0, mass == 0.0 ? 0.0 : 1.0/mass);
        else
            ((mm_float4*) pinnedMemory)[i] = mm_float4(0.0f, 0.0f, 0.0f, mass == 0.0 ? 0.0f : (cl_float) (1.0/mass));
    }
    velm->upload(pinnedMemory);
406
407
    atomIndexDevice = OpenCLArray::create<cl_int>(*this, paddedNumAtoms, "atomIndexDevice");
    atomIndex.resize(paddedNumAtoms);
408
    for (int i = 0; i < paddedNumAtoms; ++i)
409
410
        atomIndex[i] = i;
    atomIndexDevice->upload(atomIndex);
411
    findMoleculeGroups();
412
    nonbonded->initialize(system);
413
414
415
416
417
418
}

void OpenCLContext::addForce(OpenCLForceInfo* force) {
    forces.push_back(force);
}

419
420
string OpenCLContext::replaceStrings(const string& input, const std::map<std::string, std::string>& replacements) const {
    string result = input;
421
422
423
    for (map<string, string>::const_iterator iter = replacements.begin(); iter != replacements.end(); iter++) {
        int index = -1;
        do {
424
425
426
427
            index = result.find(iter->first);
            if (index != result.npos)
                result.replace(index, iter->first.size(), iter->second);
        } while (index != result.npos);
428
    }
429
    return result;
430
431
}

432
433
cl::Program OpenCLContext::createProgram(const string source, const char* optimizationFlags) {
    return createProgram(source, map<string, string>(), optimizationFlags);
434
435
}

436
cl::Program OpenCLContext::createProgram(const string source, const map<string, string>& defines, const char* optimizationFlags) {
Peter Eastman's avatar
Peter Eastman committed
437
    string options = (optimizationFlags == NULL ? defaultOptimizationOptions : string(optimizationFlags));
438
439
440
441
442
443
444
445
446
447
448
    stringstream src;
    if (!options.empty())
        src << "// Compilation Options: " << options << endl << endl;
    for (map<string, string>::const_iterator iter = compilationDefines.begin(); iter != compilationDefines.end(); ++iter) {
        src << "#define " << iter->first;
        if (!iter->second.empty())
            src << " " << iter->second;
        src << endl;
    }
    if (!compilationDefines.empty())
        src << endl;
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    if (supportsDoublePrecision)
        src << "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n";
    if (useDoublePrecision) {
        src << "typedef double real;\n";
        src << "typedef double2 real2;\n";
        src << "typedef double4 real4;\n";
    }
    else {
        src << "typedef float real;\n";
        src << "typedef float2 real2;\n";
        src << "typedef float4 real4;\n";
    }
    if (useDoublePrecision || useMixedPrecision) {
        src << "typedef double mixed;\n";
        src << "typedef double2 mixed2;\n";
        src << "typedef double4 mixed4;\n";
    }
    else {
        src << "typedef float mixed;\n";
        src << "typedef float2 mixed2;\n";
        src << "typedef float4 mixed4;\n";
    }
471
472
473
474
475
476
477
478
479
480
481
482
483
    for (map<string, string>::const_iterator iter = defines.begin(); iter != defines.end(); ++iter) {
        src << "#define " << iter->first;
        if (!iter->second.empty())
            src << " " << iter->second;
        src << endl;
    }
    if (!defines.empty())
        src << endl;
    src << source << endl;
    // Get length before using c_str() to avoid length() call invalidating the c_str() value.
    string src_string = src.str();
    ::size_t src_length = src_string.length();
    cl::Program::Sources sources(1, make_pair(src_string.c_str(), src_length));
484
485
    cl::Program program(context, sources);
    try {
486
        program.build(vector<cl::Device>(1, device), options.c_str());
487
488
489
490
491
492
    } catch (cl::Error err) {
        throw OpenMMException("Error compiling kernel: "+program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device));
    }
    return program;
}

493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
string OpenCLContext::doubleToString(double value) {
    stringstream s;
    s.precision(useDoublePrecision ? 16 : 8);
    s << scientific << value;
    if (!useDoublePrecision)
        s << "f";
    return s.str();
}

string OpenCLContext::intToString(int value) {
    stringstream s;
    s << value;
    return s.str();
}

508
509
510
511
void OpenCLContext::executeKernel(cl::Kernel& kernel, int workUnits, int blockSize) {
    if (blockSize == -1)
        blockSize = ThreadBlockSize;
    int size = std::min((workUnits+blockSize-1)/blockSize, numThreadBlocks)*blockSize;
512
    try {
513
        queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(size), cl::NDRange(blockSize));
514
515
516
    }
    catch (cl::Error err) {
        stringstream str;
517
        str<<"Error invoking kernel "<<kernel.getInfo<CL_KERNEL_FUNCTION_NAME>()<<": "<<err.what()<<" ("<<err.err()<<")";
518
519
520
521
        throw OpenMMException(str.str());
    }
}

522
void OpenCLContext::clearBuffer(OpenCLArray& array) {
523
    clearBuffer(array.getDeviceBuffer(), array.getSize()*array.getElementSize());
524
525
}

526
void OpenCLContext::clearBuffer(cl::Memory& memory, int size) {
527
    int words = size/4;
528
    clearBufferKernel.setArg<cl::Memory>(0, memory);
529
530
531
532
533
534
    clearBufferKernel.setArg<cl_int>(1, words);
    executeKernel(clearBufferKernel, words, 128);
}

void OpenCLContext::addAutoclearBuffer(OpenCLArray& array) {
    addAutoclearBuffer(array.getDeviceBuffer(), array.getSize()*array.getElementSize());
535
536
}

537
538
void OpenCLContext::addAutoclearBuffer(cl::Memory& memory, int size) {
    autoclearBuffers.push_back(&memory);
539
    autoclearBufferSizes.push_back(size/4);
540
541
542
543
544
}

void OpenCLContext::clearAutoclearBuffers() {
    int base = 0;
    int total = autoclearBufferSizes.size();
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
    while (total-base >= 6) {
        clearSixBuffersKernel.setArg<cl::Memory>(0, *autoclearBuffers[base]);
        clearSixBuffersKernel.setArg<cl_int>(1, autoclearBufferSizes[base]);
        clearSixBuffersKernel.setArg<cl::Memory>(2, *autoclearBuffers[base+1]);
        clearSixBuffersKernel.setArg<cl_int>(3, autoclearBufferSizes[base+1]);
        clearSixBuffersKernel.setArg<cl::Memory>(4, *autoclearBuffers[base+2]);
        clearSixBuffersKernel.setArg<cl_int>(5, autoclearBufferSizes[base+2]);
        clearSixBuffersKernel.setArg<cl::Memory>(6, *autoclearBuffers[base+3]);
        clearSixBuffersKernel.setArg<cl_int>(7, autoclearBufferSizes[base+3]);
        clearSixBuffersKernel.setArg<cl::Memory>(8, *autoclearBuffers[base+4]);
        clearSixBuffersKernel.setArg<cl_int>(9, autoclearBufferSizes[base+4]);
        clearSixBuffersKernel.setArg<cl::Memory>(10, *autoclearBuffers[base+5]);
        clearSixBuffersKernel.setArg<cl_int>(11, autoclearBufferSizes[base+5]);
        executeKernel(clearSixBuffersKernel, 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) {
        clearFiveBuffersKernel.setArg<cl::Memory>(0, *autoclearBuffers[base]);
        clearFiveBuffersKernel.setArg<cl_int>(1, autoclearBufferSizes[base]);
        clearFiveBuffersKernel.setArg<cl::Memory>(2, *autoclearBuffers[base+1]);
        clearFiveBuffersKernel.setArg<cl_int>(3, autoclearBufferSizes[base+1]);
        clearFiveBuffersKernel.setArg<cl::Memory>(4, *autoclearBuffers[base+2]);
        clearFiveBuffersKernel.setArg<cl_int>(5, autoclearBufferSizes[base+2]);
        clearFiveBuffersKernel.setArg<cl::Memory>(6, *autoclearBuffers[base+3]);
        clearFiveBuffersKernel.setArg<cl_int>(7, autoclearBufferSizes[base+3]);
        clearFiveBuffersKernel.setArg<cl::Memory>(8, *autoclearBuffers[base+4]);
        clearFiveBuffersKernel.setArg<cl_int>(9, autoclearBufferSizes[base+4]);
        executeKernel(clearFiveBuffersKernel, max(max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), autoclearBufferSizes[base+4]), 128);
    }
    else if (total-base == 4) {
575
576
577
578
579
580
581
582
        clearFourBuffersKernel.setArg<cl::Memory>(0, *autoclearBuffers[base]);
        clearFourBuffersKernel.setArg<cl_int>(1, autoclearBufferSizes[base]);
        clearFourBuffersKernel.setArg<cl::Memory>(2, *autoclearBuffers[base+1]);
        clearFourBuffersKernel.setArg<cl_int>(3, autoclearBufferSizes[base+1]);
        clearFourBuffersKernel.setArg<cl::Memory>(4, *autoclearBuffers[base+2]);
        clearFourBuffersKernel.setArg<cl_int>(5, autoclearBufferSizes[base+2]);
        clearFourBuffersKernel.setArg<cl::Memory>(6, *autoclearBuffers[base+3]);
        clearFourBuffersKernel.setArg<cl_int>(7, autoclearBufferSizes[base+3]);
583
        executeKernel(clearFourBuffersKernel, max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), 128);
584
    }
585
    else if (total-base == 3) {
586
587
588
589
590
591
        clearThreeBuffersKernel.setArg<cl::Memory>(0, *autoclearBuffers[base]);
        clearThreeBuffersKernel.setArg<cl_int>(1, autoclearBufferSizes[base]);
        clearThreeBuffersKernel.setArg<cl::Memory>(2, *autoclearBuffers[base+1]);
        clearThreeBuffersKernel.setArg<cl_int>(3, autoclearBufferSizes[base+1]);
        clearThreeBuffersKernel.setArg<cl::Memory>(4, *autoclearBuffers[base+2]);
        clearThreeBuffersKernel.setArg<cl_int>(5, autoclearBufferSizes[base+2]);
592
        executeKernel(clearThreeBuffersKernel, max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), 128);
593
594
595
596
597
598
    }
    else if (total-base == 2) {
        clearTwoBuffersKernel.setArg<cl::Memory>(0, *autoclearBuffers[base]);
        clearTwoBuffersKernel.setArg<cl_int>(1, autoclearBufferSizes[base]);
        clearTwoBuffersKernel.setArg<cl::Memory>(2, *autoclearBuffers[base+1]);
        clearTwoBuffersKernel.setArg<cl_int>(3, autoclearBufferSizes[base+1]);
599
        executeKernel(clearTwoBuffersKernel, max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), 128);
600
601
602
603
604
605
    }
    else if (total-base == 1) {
        clearBuffer(*autoclearBuffers[base], autoclearBufferSizes[base]);
    }
}

606
607
608
609
610
611
612
void OpenCLContext::reduceForces() {
    if (supports64BitGlobalAtomics)
        executeKernel(reduceForcesKernel, paddedNumAtoms, 128);
    else
        reduceBuffer(*forceBuffers, numForceBuffers);
}

613
void OpenCLContext::reduceBuffer(OpenCLArray& array, int numBuffers) {
614
    int bufferSize = array.getSize()/numBuffers;
615
616
617
618
    reduceReal4Kernel.setArg<cl::Buffer>(0, array.getDeviceBuffer());
    reduceReal4Kernel.setArg<cl_int>(1, bufferSize);
    reduceReal4Kernel.setArg<cl_int>(2, numBuffers);
    executeKernel(reduceReal4Kernel, bufferSize, 128);
619
}
620
621
622
623
624
625
626
627
628
629

void OpenCLContext::tagAtomsInMolecule(int atom, int molecule, vector<int>& atomMolecule, vector<vector<int> >& atomBonds) {
    // Recursively tag atoms as belonging to a particular molecule.

    atomMolecule[atom] = molecule;
    for (int i = 0; i < (int) atomBonds[atom].size(); i++)
        if (atomMolecule[atomBonds[atom][i]] == -1)
            tagAtomsInMolecule(atomBonds[atom][i], molecule, atomMolecule, atomBonds);
}

630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
/**
 * This class ensures that atom reordering doesn't break virtual sites.
 */
class OpenCLContext::VirtualSiteInfo : public OpenCLForceInfo {
public:
    VirtualSiteInfo(const System& system) : OpenCLForceInfo(0) {
        for (int i = 0; i < system.getNumParticles(); i++) {
            if (system.isVirtualSite(i)) {
                siteTypes.push_back(&typeid(system.getVirtualSite(i)));
                vector<int> particles;
                particles.push_back(i);
                for (int j = 0; j < system.getVirtualSite(i).getNumParticles(); j++)
                    particles.push_back(system.getVirtualSite(i).getParticle(j));
                siteParticles.push_back(particles);
                vector<double> weights;
                if (dynamic_cast<const TwoParticleAverageSite*>(&system.getVirtualSite(i)) != NULL) {
                    // A two particle average.

                    const TwoParticleAverageSite& site = dynamic_cast<const TwoParticleAverageSite&>(system.getVirtualSite(i));
                    weights.push_back(site.getWeight(0));
                    weights.push_back(site.getWeight(1));
                }
                else if (dynamic_cast<const ThreeParticleAverageSite*>(&system.getVirtualSite(i)) != NULL) {
                    // A three particle average.

                    const ThreeParticleAverageSite& site = dynamic_cast<const ThreeParticleAverageSite&>(system.getVirtualSite(i));
                    weights.push_back(site.getWeight(0));
                    weights.push_back(site.getWeight(1));
                    weights.push_back(site.getWeight(2));
                }
                else if (dynamic_cast<const OutOfPlaneSite*>(&system.getVirtualSite(i)) != NULL) {
                    // An out of plane site.

                    const OutOfPlaneSite& site = dynamic_cast<const OutOfPlaneSite&>(system.getVirtualSite(i));
                    weights.push_back(site.getWeight12());
                    weights.push_back(site.getWeight13());
                    weights.push_back(site.getWeightCross());
                }
                siteWeights.push_back(weights);
            }
        }
    }
    int getNumParticleGroups() {
        return siteTypes.size();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        particles = siteParticles[index];
    }
    bool areGroupsIdentical(int group1, int group2) {
        if (siteTypes[group1] != siteTypes[group2])
            return false;
        int numParticles = siteWeights[group1].size();
        if (siteWeights[group2].size() != numParticles)
            return false;
        for (int i = 0; i < numParticles; i++)
            if (siteWeights[group1][i] != siteWeights[group2][i])
                return false;
        return true;
    }
private:
    vector<const type_info*> siteTypes;
    vector<vector<int> > siteParticles;
    vector<vector<double> > siteWeights;
};


696
697
void OpenCLContext::findMoleculeGroups() {
    // The first time this is called, we need to identify all the molecules in the system.
698
    
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
    if (moleculeGroups.size() == 0) {
        // Add a ForceInfo that makes sure reordering doesn't break virtual sites.

        addForce(new VirtualSiteInfo(system));

        // First make a list of every other atom to which each atom is connect by a constraint or force group.

        vector<vector<int> > atomBonds(system.getNumParticles());
        for (int i = 0; i < system.getNumConstraints(); i++) {
            int particle1, particle2;
            double distance;
            system.getConstraintParameters(i, particle1, particle2, distance);
            atomBonds[particle1].push_back(particle2);
            atomBonds[particle2].push_back(particle1);
        }
        for (int i = 0; i < (int) forces.size(); i++) {
            for (int j = 0; j < forces[i]->getNumParticleGroups(); j++) {
                vector<int> particles;
                forces[i]->getParticlesInGroup(j, particles);
                for (int k = 0; k < (int) particles.size(); k++)
                    for (int m = 0; m < (int) particles.size(); m++)
                        if (k != m)
                            atomBonds[particles[k]].push_back(particles[m]);
            }
723
724
        }

725
        // Now tag atoms by which molecule they belong to.
726

727
728
729
730
731
732
733
734
        vector<int> atomMolecule(numAtoms, -1);
        int numMolecules = 0;
        for (int i = 0; i < numAtoms; i++)
            if (atomMolecule[i] == -1)
                tagAtomsInMolecule(i, numMolecules++, atomMolecule, atomBonds);
        vector<vector<int> > atomIndices(numMolecules);
        for (int i = 0; i < numAtoms; i++)
            atomIndices[atomMolecule[i]].push_back(i);
735

736
        // Construct a description of each molecule.
737

738
739
740
741
        molecules.resize(numMolecules);
        for (int i = 0; i < numMolecules; i++) {
            molecules[i].atoms = atomIndices[i];
            molecules[i].groups.resize(forces.size());
742
        }
743
744
745
746
747
748
749
750
751
752
753
754
755
        for (int i = 0; i < system.getNumConstraints(); i++) {
            int particle1, particle2;
            double distance;
            system.getConstraintParameters(i, particle1, particle2, distance);
            molecules[atomMolecule[particle1]].constraints.push_back(i);
        }
        for (int i = 0; i < (int) forces.size(); i++)
            for (int j = 0; j < forces[i]->getNumParticleGroups(); j++) {
                vector<int> particles;
                forces[i]->getParticlesInGroup(j, particles);
                molecules[atomMolecule[particles[0]]].groups[i].push_back(j);
            }
    }
756
757
758
759
760

    // Sort them into groups of identical molecules.

    vector<Molecule> uniqueMolecules;
    vector<vector<int> > moleculeInstances;
761
    vector<vector<int> > moleculeOffsets;
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
    for (int molIndex = 0; molIndex < (int) molecules.size(); molIndex++) {
        Molecule& mol = molecules[molIndex];

        // See if it is identical to another molecule.

        bool isNew = true;
        for (int j = 0; j < (int) uniqueMolecules.size() && isNew; j++) {
            Molecule& mol2 = uniqueMolecules[j];
            bool identical = (mol.atoms.size() == mol2.atoms.size() && mol.constraints.size() == mol2.constraints.size());

            // See if the atoms are identical.

            int atomOffset = mol2.atoms[0]-mol.atoms[0];
            for (int i = 0; i < (int) mol.atoms.size() && identical; i++) {
                if (mol.atoms[i] != mol2.atoms[i]-atomOffset || system.getParticleMass(mol.atoms[i]) != system.getParticleMass(mol2.atoms[i]))
                    identical = false;
778
                for (int k = 0; k < (int) forces.size(); k++)
779
780
781
782
783
784
785
786
787
788
789
                    if (!forces[k]->areParticlesIdentical(mol.atoms[i], mol2.atoms[i]))
                        identical = false;
            }
            
            // See if the constraints are identical.

            for (int i = 0; i < (int) mol.constraints.size() && identical; i++) {
                int c1particle1, c1particle2, c2particle1, c2particle2;
                double distance1, distance2;
                system.getConstraintParameters(mol.constraints[i], c1particle1, c1particle2, distance1);
                system.getConstraintParameters(mol2.constraints[i], c2particle1, c2particle2, distance2);
790
                if (c1particle1 != c2particle1-atomOffset || c1particle2 != c2particle2-atomOffset || distance1 != distance2)
791
792
793
794
795
                    identical = false;
            }

            // See if the force groups are identical.

796
            for (int i = 0; i < (int) forces.size() && identical; i++) {
797
798
                if (mol.groups[i].size() != mol2.groups[i].size())
                    identical = false;
799
                for (int k = 0; k < (int) mol.groups[i].size() && identical; k++)
800
801
802
803
                    if (!forces[i]->areGroupsIdentical(mol.groups[i][k], mol2.groups[i][k]))
                        identical = false;
            }
            if (identical) {
804
805
                moleculeInstances[j].push_back(molIndex);
                moleculeOffsets[j].push_back(mol.atoms[0]);
806
807
808
809
810
811
                isNew = false;
            }
        }
        if (isNew) {
            uniqueMolecules.push_back(mol);
            moleculeInstances.push_back(vector<int>());
812
813
814
            moleculeInstances[moleculeInstances.size()-1].push_back(molIndex);
            moleculeOffsets.push_back(vector<int>());
            moleculeOffsets[moleculeOffsets.size()-1].push_back(mol.atoms[0]);
815
816
817
818
819
820
        }
    }
    moleculeGroups.resize(moleculeInstances.size());
    for (int i = 0; i < (int) moleculeInstances.size(); i++)
    {
        moleculeGroups[i].instances = moleculeInstances[i];
821
        moleculeGroups[i].offsets = moleculeOffsets[i];
822
823
824
825
826
827
828
        vector<int>& atoms = uniqueMolecules[i].atoms;
        moleculeGroups[i].atoms.resize(atoms.size());
        for (int j = 0; j < (int) atoms.size(); j++)
            moleculeGroups[i].atoms[j] = atoms[j]-atoms[0];
    }
}

829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
void OpenCLContext::invalidateMolecules() {
    if (numAtoms == 0 || nonbonded == NULL || !nonbonded->getUseCutoff())
        return;
    bool valid = true;
    for (int group = 0; valid && group < (int) moleculeGroups.size(); group++) {
        MoleculeGroup& mol = moleculeGroups[group];
        vector<int>& instances = mol.instances;
        vector<int>& offsets = mol.offsets;
        vector<int>& atoms = mol.atoms;
        int numMolecules = instances.size();
        Molecule& m1 = molecules[instances[0]];
        int offset1 = offsets[0];
        for (int j = 1; valid && j < numMolecules; j++) {
            // See if the atoms are identical.

            Molecule& m2 = molecules[instances[j]];
            int offset2 = offsets[j];
            for (int i = 0; i < (int) atoms.size() && valid; i++) {
                for (int k = 0; k < (int) forces.size(); k++)
                    if (!forces[k]->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
                        valid = false;
            }

            // See if the force groups are identical.

            for (int i = 0; i < (int) forces.size() && valid; i++) {
                for (int k = 0; k < (int) m1.groups[i].size() && valid; k++)
                    if (!forces[i]->areGroupsIdentical(m1.groups[i][k], m2.groups[i][k]))
                        valid = false;
            }
        }
    }
    if (valid)
        return;
    
    // The list of which molecules are identical is no longer valid.  We need to restore the
    // atoms to their original order, rebuild the list of identical molecules, and sort them
    // again.
    
    vector<mm_int4> newCellOffsets(numAtoms);
869
870
    if (useDoublePrecision) {
        vector<mm_double4> oldPosq(paddedNumAtoms);
871
        vector<mm_double4> newPosq(paddedNumAtoms, mm_double4(0,0,0,0));
872
        vector<mm_double4> oldVelm(paddedNumAtoms);
873
        vector<mm_double4> newVelm(paddedNumAtoms, mm_double4(0,0,0,0));
874
875
876
877
878
879
880
881
882
883
884
885
886
        posq->download(oldPosq);
        velm->download(oldVelm);
        for (int i = 0; i < numAtoms; i++) {
            int index = atomIndex[i];
            newPosq[index] = oldPosq[i];
            newVelm[index] = oldVelm[i];
            newCellOffsets[index] = posCellOffsets[i];
        }
        posq->upload(newPosq);
        velm->upload(newVelm);
    }
    else if (useMixedPrecision) {
        vector<mm_float4> oldPosq(paddedNumAtoms);
887
        vector<mm_float4> newPosq(paddedNumAtoms, mm_float4(0,0,0,0));
888
        vector<mm_float4> oldPosqCorrection(paddedNumAtoms);
889
        vector<mm_float4> newPosqCorrection(paddedNumAtoms, mm_float4(0,0,0,0));
890
        vector<mm_double4> oldVelm(paddedNumAtoms);
891
        vector<mm_double4> newVelm(paddedNumAtoms, mm_double4(0,0,0,0));
892
893
894
895
896
897
898
899
900
901
        posq->download(oldPosq);
        velm->download(oldVelm);
        for (int i = 0; i < numAtoms; i++) {
            int index = atomIndex[i];
            newPosq[index] = oldPosq[i];
            newPosqCorrection[index] = oldPosqCorrection[i];
            newVelm[index] = oldVelm[i];
            newCellOffsets[index] = posCellOffsets[i];
        }
        posq->upload(newPosq);
Peter Eastman's avatar
Peter Eastman committed
902
        posqCorrection->upload(newPosqCorrection);
903
904
905
906
        velm->upload(newVelm);
    }
    else {
        vector<mm_float4> oldPosq(paddedNumAtoms);
907
        vector<mm_float4> newPosq(paddedNumAtoms, mm_float4(0,0,0,0));
908
        vector<mm_float4> oldVelm(paddedNumAtoms);
909
        vector<mm_float4> newVelm(paddedNumAtoms, mm_float4(0,0,0,0));
910
911
912
913
914
915
916
917
918
919
        posq->download(oldPosq);
        velm->download(oldVelm);
        for (int i = 0; i < numAtoms; i++) {
            int index = atomIndex[i];
            newPosq[index] = oldPosq[i];
            newVelm[index] = oldVelm[i];
            newCellOffsets[index] = posCellOffsets[i];
        }
        posq->upload(newPosq);
        velm->upload(newVelm);
920
921
    }
    for (int i = 0; i < numAtoms; i++) {
922
        atomIndex[i] = i;
923
924
        posCellOffsets[i] = newCellOffsets[i];
    }
925
    atomIndexDevice->upload(atomIndex);
926
    findMoleculeGroups();
927
928
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        reorderListeners[i]->execute();
929
    reorderAtoms();
930
931
}

932
933
934
935
void OpenCLContext::reorderAtoms() {
    atomsWereReordered = false;
    if (numAtoms == 0 || nonbonded == NULL || !nonbonded->getUseCutoff() || stepsSinceReorder < 100) {
        stepsSinceReorder++;
936
        return;
937
    }
Peter Eastman's avatar
Peter Eastman committed
938
    atomsWereReordered = true;
939
    stepsSinceReorder = 0;
940
    if (useDoublePrecision)
941
        reorderAtomsImpl<cl_double, mm_double4, cl_double, mm_double4>();
942
    else if (useMixedPrecision)
943
        reorderAtomsImpl<cl_float, mm_float4, cl_double, mm_double4>();
944
    else
945
946
        reorderAtomsImpl<cl_float, mm_float4, cl_float, mm_float4>();
    nonbonded->updateNeighborListSize();
947
948
949
}

template <class Real, class Real4, class Mixed, class Mixed4>
950
void OpenCLContext::reorderAtomsImpl() {
951
952
953

    // Find the range of positions and the number of bins along each axis.

954
955
956
    vector<Real4> oldPosq(paddedNumAtoms);
    vector<Real4> oldPosqCorrection(paddedNumAtoms);
    vector<Mixed4> oldVelm(paddedNumAtoms);
957
958
    posq->download(oldPosq);
    velm->download(oldVelm);
959
960
961
962
963
    if (useMixedPrecision)
        posqCorrection->download(oldPosqCorrection);
    Real minx = oldPosq[0].x, maxx = oldPosq[0].x;
    Real miny = oldPosq[0].y, maxy = oldPosq[0].y;
    Real minz = oldPosq[0].z, maxz = oldPosq[0].z;
964
965
    if (nonbonded->getUsePeriodic()) {
        minx = miny = minz = 0.0;
966
967
968
        maxx = periodicBoxSizeDouble.x;
        maxy = periodicBoxSizeDouble.y;
        maxz = periodicBoxSizeDouble.z;
969
970
971
    }
    else {
        for (int i = 1; i < numAtoms; i++) {
972
            const Real4& pos = oldPosq[i];
973
974
975
976
977
978
            minx = min(minx, pos.x);
            maxx = max(maxx, pos.x);
            miny = min(miny, pos.y);
            maxy = max(maxy, pos.y);
            minz = min(minz, pos.z);
            maxz = max(maxz, pos.z);
979
980
981
982
983
984
        }
    }

    // Loop over each group of identical molecules and reorder them.

    vector<int> originalIndex(numAtoms);
985
986
987
    vector<Real4> newPosq(paddedNumAtoms, Real4(0,0,0,0));
    vector<Real4> newPosqCorrection(paddedNumAtoms, Real4(0,0,0,0));
    vector<Mixed4> newVelm(paddedNumAtoms, Mixed4(0,0,0,0));
988
989
990
991
992
    vector<mm_int4> newCellOffsets(numAtoms);
    for (int group = 0; group < (int) moleculeGroups.size(); group++) {
        // Find the center of each molecule.

        MoleculeGroup& mol = moleculeGroups[group];
993
        int numMolecules = mol.offsets.size();
994
        vector<int>& atoms = mol.atoms;
995
996
        vector<Real4> molPos(numMolecules);
        Real invNumAtoms = (Real) (1.0/atoms.size());
997
998
999
1000
1001
        for (int i = 0; i < numMolecules; i++) {
            molPos[i].x = 0.0f;
            molPos[i].y = 0.0f;
            molPos[i].z = 0.0f;
            for (int j = 0; j < (int)atoms.size(); j++) {
1002
                int atom = atoms[j]+mol.offsets[i];
1003
                const Real4& pos = oldPosq[atom];
1004
1005
1006
                molPos[i].x += pos.x;
                molPos[i].y += pos.y;
                molPos[i].z += pos.z;
1007
            }
1008
1009
1010
            molPos[i].x *= invNumAtoms;
            molPos[i].y *= invNumAtoms;
            molPos[i].z *= invNumAtoms;
1011
1012
1013
1014
1015
        }
        if (nonbonded->getUsePeriodic()) {
            // Move each molecule position into the same box.

            for (int i = 0; i < numMolecules; i++) {
1016
1017
1018
1019
1020
1021
                int xcell = (int) floor(molPos[i].x*invPeriodicBoxSizeDouble.x);
                int ycell = (int) floor(molPos[i].y*invPeriodicBoxSizeDouble.y);
                int zcell = (int) floor(molPos[i].z*invPeriodicBoxSizeDouble.z);
                Real dx = xcell*periodicBoxSizeDouble.x;
                Real dy = ycell*periodicBoxSizeDouble.y;
                Real dz = zcell*periodicBoxSizeDouble.z;
1022
1023
1024
1025
                if (dx != 0.0f || dy != 0.0f || dz != 0.0f) {
                    molPos[i].x -= dx;
                    molPos[i].y -= dy;
                    molPos[i].z -= dz;
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
                    for (int j = 0; j < (int) atoms.size(); j++) {
                        int atom = atoms[j]+mol.offsets[i];
                        Real4 p = oldPosq[atom];
                        p.x -= dx;
                        p.y -= dy;
                        p.z -= dz;
                        oldPosq[atom] = p;
                        posCellOffsets[atom].x -= xcell;
                        posCellOffsets[atom].y -= ycell;
                        posCellOffsets[atom].z -= zcell;
1036
1037
1038
1039
1040
1041
1042
1043
                    }
                }
            }
        }

        // Select a bin for each molecule, then sort them by bin.

        bool useHilbert = (numMolecules > 5000 || atoms.size() > 8); // For small systems, a simple zigzag curve works better than a Hilbert curve.
1044
        Real binWidth;
1045
        if (useHilbert)
1046
            binWidth = (Real) (max(max(maxx-minx, maxy-miny), maxz-minz)/255.0);
1047
        else
1048
1049
            binWidth = (Real) (0.2*nonbonded->getCutoffDistance());
        Real invBinWidth = (Real) (1.0/binWidth);
1050
1051
        int xbins = 1 + (int) ((maxx-minx)*invBinWidth);
        int ybins = 1 + (int) ((maxy-miny)*invBinWidth);
1052
1053
1054
        vector<pair<int, int> > molBins(numMolecules);
        bitmask_t coords[3];
        for (int i = 0; i < numMolecules; i++) {
1055
1056
1057
            int x = (int) ((molPos[i].x-minx)*invBinWidth);
            int y = (int) ((molPos[i].y-miny)*invBinWidth);
            int z = (int) ((molPos[i].z-minz)*invBinWidth);
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
            int bin;
            if (useHilbert) {
                coords[0] = x;
                coords[1] = y;
                coords[2] = z;
                bin = (int) hilbert_c2i(3, 8, coords);
            }
            else {
                int yodd = y&1;
                int zodd = z&1;
                bin = z*xbins*ybins;
                bin += (zodd ? ybins-y : y)*xbins;
                bin += (yodd ? xbins-x : x);
            }
            molBins[i] = pair<int, int>(bin, i);
        }
        sort(molBins.begin(), molBins.end());

        // Reorder the atoms.

        for (int i = 0; i < numMolecules; i++) {
            for (int j = 0; j < (int)atoms.size(); j++) {
1080
1081
                int oldIndex = mol.offsets[molBins[i].second]+atoms[j];
                int newIndex = mol.offsets[i]+atoms[j];
1082
1083
                originalIndex[newIndex] = atomIndex[oldIndex];
                newPosq[newIndex] = oldPosq[oldIndex];
1084
1085
                if (useMixedPrecision)
                    newPosqCorrection[newIndex] = oldPosqCorrection[oldIndex];
1086
                newVelm[newIndex] = oldVelm[oldIndex];
1087
1088
1089
1090
1091
1092
1093
1094
                newCellOffsets[newIndex] = posCellOffsets[oldIndex];
            }
        }
    }

    // Update the streams.

    for (int i = 0; i < numAtoms; i++) {
1095
        atomIndex[i] = originalIndex[i];
1096
1097
        posCellOffsets[i] = newCellOffsets[i];
    }
1098
    posq->upload(newPosq);
1099
1100
    if (useMixedPrecision)
        posqCorrection->upload(newPosqCorrection);
1101
1102
    velm->upload(newVelm);
    atomIndexDevice->upload(atomIndex);
1103
1104
1105
1106
1107
1108
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        reorderListeners[i]->execute();
}

void OpenCLContext::addReorderListener(ReorderListener* listener) {
    reorderListeners.push_back(listener);
1109
}
1110

1111
1112
1113
1114
1115
1116
1117
1118
void OpenCLContext::addPreComputation(ForcePreComputation* computation) {
    preComputations.push_back(computation);
}

void OpenCLContext::addPostComputation(ForcePostComputation* computation) {
    postComputations.push_back(computation);
}

1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
struct OpenCLContext::WorkThread::ThreadData {
    ThreadData(std::queue<OpenCLContext::WorkTask*>& tasks, bool& waiting,  bool& finished,
            pthread_mutex_t& queueLock, pthread_cond_t& waitForTaskCondition, pthread_cond_t& queueEmptyCondition) :
        tasks(tasks), waiting(waiting), finished(finished), queueLock(queueLock),
        waitForTaskCondition(waitForTaskCondition), queueEmptyCondition(queueEmptyCondition) {
    }
    std::queue<OpenCLContext::WorkTask*>& tasks;
    bool& waiting;
    bool& finished;
    pthread_mutex_t& queueLock;
    pthread_cond_t& waitForTaskCondition;
    pthread_cond_t& queueEmptyCondition;
};

static void* threadBody(void* args) {
    OpenCLContext::WorkThread::ThreadData& data = *reinterpret_cast<OpenCLContext::WorkThread::ThreadData*>(args);
    while (!data.finished || data.tasks.size() > 0) {
        pthread_mutex_lock(&data.queueLock);
        while (data.tasks.empty() && !data.finished) {
            data.waiting = true;
            pthread_cond_signal(&data.queueEmptyCondition);
            pthread_cond_wait(&data.waitForTaskCondition, &data.queueLock);
        }
        OpenCLContext::WorkTask* task = NULL;
        if (!data.tasks.empty()) {
            data.waiting = false;
            task = data.tasks.front();
            data.tasks.pop();
        }
        pthread_mutex_unlock(&data.queueLock);
        if (task != NULL) {
            task->execute();
            delete task;
        }
    }
    data.waiting = true;
    pthread_cond_signal(&data.queueEmptyCondition);
    delete &data;
    return 0;
}

OpenCLContext::WorkThread::WorkThread() : waiting(true), finished(false) {
    pthread_mutex_init(&queueLock, NULL);
    pthread_cond_init(&waitForTaskCondition, NULL);
    pthread_cond_init(&queueEmptyCondition, NULL);
    ThreadData* data = new ThreadData(tasks, waiting, finished, queueLock, waitForTaskCondition, queueEmptyCondition);
    pthread_create(&thread, NULL, threadBody, data);
}

OpenCLContext::WorkThread::~WorkThread() {
    pthread_mutex_lock(&queueLock);
    finished = true;
    pthread_cond_broadcast(&waitForTaskCondition);
    pthread_mutex_unlock(&queueLock);
    pthread_join(thread, NULL);
    pthread_mutex_destroy(&queueLock);
    pthread_cond_destroy(&waitForTaskCondition);
    pthread_cond_destroy(&queueEmptyCondition);
}

void OpenCLContext::WorkThread::addTask(OpenCLContext::WorkTask* task) {
    pthread_mutex_lock(&queueLock);
    tasks.push(task);
    waiting = false;
    pthread_cond_signal(&waitForTaskCondition);
    pthread_mutex_unlock(&queueLock);
}

bool OpenCLContext::WorkThread::isWaiting() {
    return waiting;
}

bool OpenCLContext::WorkThread::isFinished() {
    return finished;
}

void OpenCLContext::WorkThread::flush() {
    pthread_mutex_lock(&queueLock);
    while (!waiting)
       pthread_cond_wait(&queueEmptyCondition, &queueLock);
    pthread_mutex_unlock(&queueLock);
}