OpenCLContext.cpp 54.4 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
            bool bestSupportsDouble = false;
101
            for (int i = 0; i < (int) devices.size(); i++) {
Peter Eastman's avatar
Peter Eastman committed
102
                if (platformVendor == "Apple" && devices[i].getInfo<CL_DEVICE_VENDOR>() == "AMD")
Peter Eastman's avatar
Peter Eastman committed
103
                    continue; // Don't use AMD GPUs on OS X due to serious bugs.
104
                int maxSize = devices[i].getInfo<CL_DEVICE_MAX_WORK_ITEM_SIZES>()[0];
105
106
107
108
109
                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) {
110
111
112
113
                    cl_uint computeCapabilityMajor;
                    clGetDeviceInfo(devices[i](), CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV, sizeof(cl_uint), &computeCapabilityMajor, NULL);
                    processingElementsPerComputeUnit = (computeCapabilityMajor < 2 ? 8 : 32);
                }
114
115
116
117
                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 {
118
#ifdef CL_DEVICE_SIMD_WIDTH_AMD
119
120
121
122
123
124
125
126
127
128
129
130
131
                        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;
132
#endif
133
134
135
136
137
                    }
                    catch (cl::Error err) {
                        // Runtime does not support the queries so use default.
                    }
                }
138
                int speed = devices[i].getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>()*processingElementsPerComputeUnit*devices[i].getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>();
139
140
                bool supportsDouble = (devices[i].getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp64") != string::npos);
                if (maxSize >= minThreadBlockSize && speed > bestSpeed && (supportsDouble || !bestSupportsDouble)) {
141
                    deviceIndex = i;
142
                    bestSpeed = speed;
143
                    bestSupportsDouble = supportsDouble;
144
                }
145
            }
146
        }
147
148
149
        if (deviceIndex == -1)
            throw OpenMMException("No compatible OpenCL device is available");
        device = devices[deviceIndex];
Peter Eastman's avatar
Peter Eastman committed
150
        this->deviceIndex = deviceIndex;
151
        if (device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>() < minThreadBlockSize)
152
            throw OpenMMException("The specified OpenCL device is not compatible with OpenMM");
153
        compilationDefines["WORK_GROUP_SIZE"] = intToString(ThreadBlockSize);
Peter Eastman's avatar
Peter Eastman committed
154
        if (platformVendor.size() >= 5 && platformVendor.substr(0, 5) == "Intel")
155
156
157
            defaultOptimizationOptions = "";
        else
            defaultOptimizationOptions = "-cl-fast-relaxed-math";
158
        supports64BitGlobalAtomics = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_int64_base_atomics") != string::npos);
159
        supportsDoublePrecision = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp64") != string::npos);
160
161
        if ((useDoublePrecision || useMixedPrecision) && !supportsDoublePrecision)
            throw OpenMMException("This device does not support double precision");
162
        string vendor = device.getInfo<CL_DEVICE_VENDOR>();
163
        int numThreadBlocksPerComputeUnit = 6;
164
        if (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA") {
165
            compilationDefines["WARPS_ARE_ATOMIC"] = "";
166
            simdWidth = 32;
167
168
            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
169
170
                // proper extension as supported.  We only use them on compute level 2.0 or later, since they're very
                // slow on earlier GPUs.
171

172
                cl_uint computeCapabilityMajor;
173
                clGetDeviceInfo(device(), CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV, sizeof(cl_uint), &computeCapabilityMajor, NULL);
174
                if (computeCapabilityMajor > 1)
175
176
                    supports64BitGlobalAtomics = true;
            }
177
        }
178
        else if (vendor.size() >= 28 && vendor.substr(0, 28) == "Advanced Micro Devices, Inc.") {
179
180
181
182
183
184
185
186
187
188
189
190
191
            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
192
#ifdef CL_DEVICE_SIMD_PER_COMPUTE_UNIT_AMD
193
194
195
                        // 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>();
196
197
                        simdWidth = device.getInfo<CL_DEVICE_WAVEFRONT_WIDTH_AMD>();

198
199
200
201
202
203
204
205
                        // 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
206
#endif
207
208
209
210
211
212
213
214
215
216
                    }
                    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"] = "";
            }
217
        }
218
219
        else
            simdWidth = 1;
Peter Eastman's avatar
Peter Eastman committed
220
        if (platformVendor == "Apple" && vendor == "AMD")
221
            compilationDefines["MAC_AMD_WORKAROUND"] = "";
222
        if (supports64BitGlobalAtomics)
223
            compilationDefines["SUPPORTS_64_BIT_ATOMICS"] = "";
224
225
        if (supportsDoublePrecision)
            compilationDefines["SUPPORTS_DOUBLE_PRECISION"] = "";
226
227
228
229
        if (simdWidth >= 32)
            compilationDefines["SYNC_WARPS"] = "";
        else
            compilationDefines["SYNC_WARPS"] = "barrier(CLK_LOCAL_MEM_FENCE)";
230
231
232
233
        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);
234
        queue = cl::CommandQueue(context, device);
Peter Eastman's avatar
Peter Eastman committed
235
236
        numAtoms = system.getNumParticles();
        paddedNumAtoms = TileSize*((numAtoms+TileSize-1)/TileSize);
237
        numAtomBlocks = (paddedNumAtoms+(TileSize-1))/TileSize;
238
        numThreadBlocks = numThreadBlocksPerComputeUnit*device.getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
Peter Eastman's avatar
Peter Eastman committed
239
        bonded = new OpenCLBondedUtilities(*this);
240
        nonbonded = new OpenCLNonbondedUtilities(*this);
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
        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";
        }
262
        posCellOffsets.resize(paddedNumAtoms, mm_int4(0, 0, 0, 0));
263
264
265
266
267
    }
    catch (cl::Error err) {
        std::stringstream str;
        str<<"Error initializing context: "<<err.what()<<" ("<<err.err()<<")";
        throw OpenMMException(str.str());
268
    }
269
270
271

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

Peter Eastman's avatar
Peter Eastman committed
272
    cl::Program utilities = createProgram(OpenCLKernelSources::utilities);
273
    clearBufferKernel = cl::Kernel(utilities, "clearBuffer");
274
275
276
    clearTwoBuffersKernel = cl::Kernel(utilities, "clearTwoBuffers");
    clearThreeBuffersKernel = cl::Kernel(utilities, "clearThreeBuffers");
    clearFourBuffersKernel = cl::Kernel(utilities, "clearFourBuffers");
277
278
    clearFiveBuffersKernel = cl::Kernel(utilities, "clearFiveBuffers");
    clearSixBuffersKernel = cl::Kernel(utilities, "clearSixBuffers");
279
    reduceReal4Kernel = cl::Kernel(utilities, "reduceReal4Buffer");
280
281
    if (supports64BitGlobalAtomics)
        reduceForcesKernel = cl::Kernel(utilities, "reduceForces");
282
283
284

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

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
319
320
321
    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";
    }
322
323
324
325
    
    // Create the work thread used for parallelization when running on multiple devices.
    
    thread = new WorkThread();
Peter Eastman's avatar
Peter Eastman committed
326
    
327
    // Create utilities objects.
Peter Eastman's avatar
Peter Eastman committed
328
329
    
    integration = new OpenCLIntegrationUtilities(*this, system);
330
    expression = new OpenCLExpressionUtilities(*this);
331
332
333
}

OpenCLContext::~OpenCLContext() {
334
335
    for (int i = 0; i < (int) forces.size(); i++)
        delete forces[i];
336
337
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        delete reorderListeners[i];
338
339
340
341
    for (int i = 0; i < (int) preComputations.size(); i++)
        delete preComputations[i];
    for (int i = 0; i < (int) postComputations.size(); i++)
        delete postComputations[i];
342
343
    if (pinnedBuffer != NULL)
        delete pinnedBuffer;
344
345
    if (posq != NULL)
        delete posq;
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
346
347
    if (posqCorrection != NULL)
        delete posqCorrection;
348
349
350
351
352
353
    if (velm != NULL)
        delete velm;
    if (force != NULL)
        delete force;
    if (forceBuffers != NULL)
        delete forceBuffers;
354
355
    if (longForceBuffer != NULL)
        delete longForceBuffer;
356
357
    if (energyBuffer != NULL)
        delete energyBuffer;
358
359
    if (atomIndexDevice != NULL)
        delete atomIndexDevice;
360
361
    if (integration != NULL)
        delete integration;
362
363
    if (expression != NULL)
        delete expression;
Peter Eastman's avatar
Peter Eastman committed
364
365
    if (bonded != NULL)
        delete bonded;
366
367
    if (nonbonded != NULL)
        delete nonbonded;
368
369
    if (thread != NULL)
        delete thread;
370
371
}

372
void OpenCLContext::initialize() {
Peter Eastman's avatar
Peter Eastman committed
373
    bonded->initialize(system);
374
    numForceBuffers = platformData.contexts.size();
Peter Eastman's avatar
Peter Eastman committed
375
    numForceBuffers = std::max(numForceBuffers, bonded->getNumForceBuffers());
376
377
    for (int i = 0; i < (int) forces.size(); i++)
        numForceBuffers = std::max(numForceBuffers, forces[i]->getRequiredForceBuffers());
378
379
380
381
382
383
384
385
386
387
    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");
    }
388
    if (supports64BitGlobalAtomics) {
389
        longForceBuffer = OpenCLArray::create<cl_long>(*this, 3*paddedNumAtoms, "longForceBuffer");
390
391
392
393
        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);
394
        addAutoclearBuffer(*longForceBuffer);
395
    }
396
397
    addAutoclearBuffer(*forceBuffers);
    addAutoclearBuffer(*energyBuffer);
398
    int bufferBytes = max(velm->getSize()*velm->getElementSize(), energyBuffer->getSize()*energyBuffer->getElementSize());
399
400
    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);
401
402
403
404
405
406
407
408
    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);
409
410
    atomIndexDevice = OpenCLArray::create<cl_int>(*this, paddedNumAtoms, "atomIndexDevice");
    atomIndex.resize(paddedNumAtoms);
411
    for (int i = 0; i < paddedNumAtoms; ++i)
412
413
        atomIndex[i] = i;
    atomIndexDevice->upload(atomIndex);
414
    findMoleculeGroups();
415
    nonbonded->initialize(system);
416
417
418
419
420
421
}

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

422
423
string OpenCLContext::replaceStrings(const string& input, const std::map<std::string, std::string>& replacements) const {
    string result = input;
424
425
426
    for (map<string, string>::const_iterator iter = replacements.begin(); iter != replacements.end(); iter++) {
        int index = -1;
        do {
427
428
429
430
            index = result.find(iter->first);
            if (index != result.npos)
                result.replace(index, iter->first.size(), iter->second);
        } while (index != result.npos);
431
    }
432
    return result;
433
434
}

435
436
cl::Program OpenCLContext::createProgram(const string source, const char* optimizationFlags) {
    return createProgram(source, map<string, string>(), optimizationFlags);
437
438
}

439
cl::Program OpenCLContext::createProgram(const string source, const map<string, string>& defines, const char* optimizationFlags) {
Peter Eastman's avatar
Peter Eastman committed
440
    string options = (optimizationFlags == NULL ? defaultOptimizationOptions : string(optimizationFlags));
441
442
443
444
445
446
447
448
449
450
451
    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;
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
    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";
    }
474
475
476
477
478
479
480
481
482
483
484
485
486
    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));
487
488
    cl::Program program(context, sources);
    try {
489
        program.build(vector<cl::Device>(1, device), options.c_str());
490
491
492
493
494
495
    } catch (cl::Error err) {
        throw OpenMMException("Error compiling kernel: "+program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device));
    }
    return program;
}

496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
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();
}

511
512
513
514
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;
515
    try {
516
        queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(size), cl::NDRange(blockSize));
517
518
519
    }
    catch (cl::Error err) {
        stringstream str;
520
        str<<"Error invoking kernel "<<kernel.getInfo<CL_KERNEL_FUNCTION_NAME>()<<": "<<err.what()<<" ("<<err.err()<<")";
521
522
523
524
        throw OpenMMException(str.str());
    }
}

525
void OpenCLContext::clearBuffer(OpenCLArray& array) {
526
    clearBuffer(array.getDeviceBuffer(), array.getSize()*array.getElementSize());
527
528
}

529
void OpenCLContext::clearBuffer(cl::Memory& memory, int size) {
530
    int words = size/4;
531
    clearBufferKernel.setArg<cl::Memory>(0, memory);
532
533
534
535
536
537
    clearBufferKernel.setArg<cl_int>(1, words);
    executeKernel(clearBufferKernel, words, 128);
}

void OpenCLContext::addAutoclearBuffer(OpenCLArray& array) {
    addAutoclearBuffer(array.getDeviceBuffer(), array.getSize()*array.getElementSize());
538
539
}

540
541
void OpenCLContext::addAutoclearBuffer(cl::Memory& memory, int size) {
    autoclearBuffers.push_back(&memory);
542
    autoclearBufferSizes.push_back(size/4);
543
544
545
546
547
}

void OpenCLContext::clearAutoclearBuffers() {
    int base = 0;
    int total = autoclearBufferSizes.size();
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
575
576
577
    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) {
578
579
580
581
582
583
584
585
        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]);
586
        executeKernel(clearFourBuffersKernel, max(max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), autoclearBufferSizes[base+3]), 128);
587
    }
588
    else if (total-base == 3) {
589
590
591
592
593
594
        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]);
595
        executeKernel(clearThreeBuffersKernel, max(max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), autoclearBufferSizes[base+2]), 128);
596
597
598
599
600
601
    }
    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]);
602
        executeKernel(clearTwoBuffersKernel, max(autoclearBufferSizes[base], autoclearBufferSizes[base+1]), 128);
603
604
605
606
607
608
    }
    else if (total-base == 1) {
        clearBuffer(*autoclearBuffers[base], autoclearBufferSizes[base]);
    }
}

609
610
611
612
613
614
615
void OpenCLContext::reduceForces() {
    if (supports64BitGlobalAtomics)
        executeKernel(reduceForcesKernel, paddedNumAtoms, 128);
    else
        reduceBuffer(*forceBuffers, numForceBuffers);
}

616
void OpenCLContext::reduceBuffer(OpenCLArray& array, int numBuffers) {
617
    int bufferSize = array.getSize()/numBuffers;
618
619
620
621
    reduceReal4Kernel.setArg<cl::Buffer>(0, array.getDeviceBuffer());
    reduceReal4Kernel.setArg<cl_int>(1, bufferSize);
    reduceReal4Kernel.setArg<cl_int>(2, numBuffers);
    executeKernel(reduceReal4Kernel, bufferSize, 128);
622
}
623
624
625
626
627
628
629
630
631
632

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);
}

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
696
697
698
/**
 * 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;
};


699
700
void OpenCLContext::findMoleculeGroups() {
    // The first time this is called, we need to identify all the molecules in the system.
701
    
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
    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]);
            }
726
727
        }

728
        // Now tag atoms by which molecule they belong to.
729

730
731
732
733
734
735
736
737
        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);
738

739
        // Construct a description of each molecule.
740

741
742
743
744
        molecules.resize(numMolecules);
        for (int i = 0; i < numMolecules; i++) {
            molecules[i].atoms = atomIndices[i];
            molecules[i].groups.resize(forces.size());
745
        }
746
747
748
749
750
751
752
753
754
755
756
757
758
        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);
            }
    }
759
760
761
762
763

    // Sort them into groups of identical molecules.

    vector<Molecule> uniqueMolecules;
    vector<vector<int> > moleculeInstances;
764
    vector<vector<int> > moleculeOffsets;
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
    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;
781
                for (int k = 0; k < (int) forces.size(); k++)
782
783
784
785
786
787
788
789
790
791
792
                    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);
793
                if (c1particle1 != c2particle1-atomOffset || c1particle2 != c2particle2-atomOffset || distance1 != distance2)
794
795
796
797
798
                    identical = false;
            }

            // See if the force groups are identical.

799
            for (int i = 0; i < (int) forces.size() && identical; i++) {
800
801
                if (mol.groups[i].size() != mol2.groups[i].size())
                    identical = false;
802
                for (int k = 0; k < (int) mol.groups[i].size() && identical; k++)
803
804
805
806
                    if (!forces[i]->areGroupsIdentical(mol.groups[i][k], mol2.groups[i][k]))
                        identical = false;
            }
            if (identical) {
807
808
                moleculeInstances[j].push_back(molIndex);
                moleculeOffsets[j].push_back(mol.atoms[0]);
809
810
811
812
813
814
                isNew = false;
            }
        }
        if (isNew) {
            uniqueMolecules.push_back(mol);
            moleculeInstances.push_back(vector<int>());
815
816
817
            moleculeInstances[moleculeInstances.size()-1].push_back(molIndex);
            moleculeOffsets.push_back(vector<int>());
            moleculeOffsets[moleculeOffsets.size()-1].push_back(mol.atoms[0]);
818
819
820
821
822
823
        }
    }
    moleculeGroups.resize(moleculeInstances.size());
    for (int i = 0; i < (int) moleculeInstances.size(); i++)
    {
        moleculeGroups[i].instances = moleculeInstances[i];
824
        moleculeGroups[i].offsets = moleculeOffsets[i];
825
826
827
828
829
830
831
        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];
    }
}

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
869
870
871
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);
872
873
    if (useDoublePrecision) {
        vector<mm_double4> oldPosq(paddedNumAtoms);
874
        vector<mm_double4> newPosq(paddedNumAtoms, mm_double4(0,0,0,0));
875
        vector<mm_double4> oldVelm(paddedNumAtoms);
876
        vector<mm_double4> newVelm(paddedNumAtoms, mm_double4(0,0,0,0));
877
878
879
880
881
882
883
884
885
886
887
888
889
        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);
890
        vector<mm_float4> newPosq(paddedNumAtoms, mm_float4(0,0,0,0));
891
        vector<mm_float4> oldPosqCorrection(paddedNumAtoms);
892
        vector<mm_float4> newPosqCorrection(paddedNumAtoms, mm_float4(0,0,0,0));
893
        vector<mm_double4> oldVelm(paddedNumAtoms);
894
        vector<mm_double4> newVelm(paddedNumAtoms, mm_double4(0,0,0,0));
895
896
897
898
899
900
901
902
903
904
        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
905
        posqCorrection->upload(newPosqCorrection);
906
907
908
909
        velm->upload(newVelm);
    }
    else {
        vector<mm_float4> oldPosq(paddedNumAtoms);
910
        vector<mm_float4> newPosq(paddedNumAtoms, mm_float4(0,0,0,0));
911
        vector<mm_float4> oldVelm(paddedNumAtoms);
912
        vector<mm_float4> newVelm(paddedNumAtoms, mm_float4(0,0,0,0));
913
914
915
916
917
918
919
920
921
922
        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);
923
924
    }
    for (int i = 0; i < numAtoms; i++) {
925
        atomIndex[i] = i;
926
927
        posCellOffsets[i] = newCellOffsets[i];
    }
928
    atomIndexDevice->upload(atomIndex);
929
    findMoleculeGroups();
930
931
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        reorderListeners[i]->execute();
932
    reorderAtoms();
933
934
}

935
936
937
938
void OpenCLContext::reorderAtoms() {
    atomsWereReordered = false;
    if (numAtoms == 0 || nonbonded == NULL || !nonbonded->getUseCutoff() || stepsSinceReorder < 100) {
        stepsSinceReorder++;
939
        return;
940
    }
Peter Eastman's avatar
Peter Eastman committed
941
    atomsWereReordered = true;
942
    stepsSinceReorder = 0;
943
    if (useDoublePrecision)
944
        reorderAtomsImpl<cl_double, mm_double4, cl_double, mm_double4>();
945
    else if (useMixedPrecision)
946
        reorderAtomsImpl<cl_float, mm_float4, cl_double, mm_double4>();
947
    else
948
949
        reorderAtomsImpl<cl_float, mm_float4, cl_float, mm_float4>();
    nonbonded->updateNeighborListSize();
950
951
952
}

template <class Real, class Real4, class Mixed, class Mixed4>
953
void OpenCLContext::reorderAtomsImpl() {
954
955
956

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

957
958
959
    vector<Real4> oldPosq(paddedNumAtoms);
    vector<Real4> oldPosqCorrection(paddedNumAtoms);
    vector<Mixed4> oldVelm(paddedNumAtoms);
960
961
    posq->download(oldPosq);
    velm->download(oldVelm);
962
963
964
965
966
    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;
967
968
    if (nonbonded->getUsePeriodic()) {
        minx = miny = minz = 0.0;
969
970
971
        maxx = periodicBoxSizeDouble.x;
        maxy = periodicBoxSizeDouble.y;
        maxz = periodicBoxSizeDouble.z;
972
973
974
    }
    else {
        for (int i = 1; i < numAtoms; i++) {
975
            const Real4& pos = oldPosq[i];
976
977
978
979
980
981
            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);
982
983
984
985
986
987
        }
    }

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

    vector<int> originalIndex(numAtoms);
988
989
990
    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));
991
992
993
994
995
    vector<mm_int4> newCellOffsets(numAtoms);
    for (int group = 0; group < (int) moleculeGroups.size(); group++) {
        // Find the center of each molecule.

        MoleculeGroup& mol = moleculeGroups[group];
996
        int numMolecules = mol.offsets.size();
997
        vector<int>& atoms = mol.atoms;
998
999
        vector<Real4> molPos(numMolecules);
        Real invNumAtoms = (Real) (1.0/atoms.size());
1000
1001
1002
1003
1004
        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++) {
1005
                int atom = atoms[j]+mol.offsets[i];
1006
                const Real4& pos = oldPosq[atom];
1007
1008
1009
                molPos[i].x += pos.x;
                molPos[i].y += pos.y;
                molPos[i].z += pos.z;
1010
            }
1011
1012
1013
            molPos[i].x *= invNumAtoms;
            molPos[i].y *= invNumAtoms;
            molPos[i].z *= invNumAtoms;
1014
1015
1016
1017
1018
        }
        if (nonbonded->getUsePeriodic()) {
            // Move each molecule position into the same box.

            for (int i = 0; i < numMolecules; i++) {
1019
1020
1021
1022
1023
1024
                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;
1025
1026
1027
1028
                if (dx != 0.0f || dy != 0.0f || dz != 0.0f) {
                    molPos[i].x -= dx;
                    molPos[i].y -= dy;
                    molPos[i].z -= dz;
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
                    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;
1039
1040
1041
1042
1043
1044
1045
1046
                    }
                }
            }
        }

        // 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.
1047
        Real binWidth;
1048
        if (useHilbert)
1049
            binWidth = (Real) (max(max(maxx-minx, maxy-miny), maxz-minz)/255.0);
1050
        else
1051
1052
            binWidth = (Real) (0.2*nonbonded->getCutoffDistance());
        Real invBinWidth = (Real) (1.0/binWidth);
1053
1054
        int xbins = 1 + (int) ((maxx-minx)*invBinWidth);
        int ybins = 1 + (int) ((maxy-miny)*invBinWidth);
1055
1056
1057
        vector<pair<int, int> > molBins(numMolecules);
        bitmask_t coords[3];
        for (int i = 0; i < numMolecules; i++) {
1058
1059
1060
            int x = (int) ((molPos[i].x-minx)*invBinWidth);
            int y = (int) ((molPos[i].y-miny)*invBinWidth);
            int z = (int) ((molPos[i].z-minz)*invBinWidth);
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
            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++) {
1083
1084
                int oldIndex = mol.offsets[molBins[i].second]+atoms[j];
                int newIndex = mol.offsets[i]+atoms[j];
1085
1086
                originalIndex[newIndex] = atomIndex[oldIndex];
                newPosq[newIndex] = oldPosq[oldIndex];
1087
1088
                if (useMixedPrecision)
                    newPosqCorrection[newIndex] = oldPosqCorrection[oldIndex];
1089
                newVelm[newIndex] = oldVelm[oldIndex];
1090
1091
1092
1093
1094
1095
1096
1097
                newCellOffsets[newIndex] = posCellOffsets[oldIndex];
            }
        }
    }

    // Update the streams.

    for (int i = 0; i < numAtoms; i++) {
1098
        atomIndex[i] = originalIndex[i];
1099
1100
        posCellOffsets[i] = newCellOffsets[i];
    }
1101
    posq->upload(newPosq);
1102
1103
    if (useMixedPrecision)
        posqCorrection->upload(newPosqCorrection);
1104
1105
    velm->upload(newVelm);
    atomIndexDevice->upload(atomIndex);
1106
1107
1108
1109
1110
1111
    for (int i = 0; i < (int) reorderListeners.size(); i++)
        reorderListeners[i]->execute();
}

void OpenCLContext::addReorderListener(ReorderListener* listener) {
    reorderListeners.push_back(listener);
1112
}
1113

1114
1115
1116
1117
1118
1119
1120
1121
void OpenCLContext::addPreComputation(ForcePreComputation* computation) {
    preComputations.push_back(computation);
}

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

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
1201
1202
1203
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);
}