Merge github.com:leeping/openmm

platforms/opencl/CMakeLists.txt

@@ -16,7 +16,6 @@
 
 IF (APPLE)
     SET (CMAKE_OSX_DEPLOYMENT_TARGET "10.6")
-    SET (CMAKE_OSX_SYSROOT "/Developer/SDKs/MacOSX10.6.sdk")
 ENDIF (APPLE)
 
 set(OPENMM_BUILD_OPENCL_TESTS TRUE CACHE BOOL "Whether to build OpenCL test cases")
@@ -103,4 +102,9 @@ SET(CL_KERNELS_H ${CMAKE_CURRENT_BINARY_DIR}/src/${CL_SOURCE_CLASS}.h)
 SET(SOURCE_FILES ${SOURCE_FILES} ${CL_KERNELS_CPP} ${CL_KERNELS_H})
 INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_BINARY_DIR}/src)
 
+# Install headers
+
+FILE(GLOB CORE_HEADERS include/*.h)
+INSTALL_FILES(/include/openmm/opencl FILES ${CORE_HEADERS})
+
 SUBDIRS (sharedTarget)

platforms/opencl/src/OpenCLContext.h → platforms/opencl/include/OpenCLContext.h

@@ -158,6 +158,8 @@ public:
     class WorkTask;
     class WorkThread;
     class ReorderListener;
+    class ForcePreComputation;
+    class ForcePostComputation;
     static const int ThreadBlockSize;
     static const int TileSize;
     OpenCLContext(const System& system, int platformIndex, int deviceIndex, const std::string& precision, OpenCLPlatform::PlatformData& platformData);
@@ -554,6 +556,28 @@ public:
     std::vector<ReorderListener*>& getReorderListeners() {
         return reorderListeners;
     }
+    /**
+     * Add a pre-computation that should be called at the very start of force and energy evaluations.
+     * The OpenCLContext assumes ownership of the object, and deletes it when the context itself is deleted.
+     */
+    void addPreComputation(ForcePreComputation* computation);
+    /**
+     * Get the list of ForcePreComputations.
+     */
+    std::vector<ForcePreComputation*>& getPreComputations() {
+        return preComputations;
+    }
+    /**
+     * Add a post-computation that should be called at the very end of force and energy evaluations.
+     * The OpenCLContext assumes ownership of the object, and deletes it when the context itself is deleted.
+     */
+    void addPostComputation(ForcePostComputation* computation);
+    /**
+     * Get the list of ForcePostComputations.
+     */
+    std::vector<ForcePostComputation*>& getPostComputations() {
+        return postComputations;
+    }
     /**
      * Mark that the current molecule definitions (and hence the atom order) may be invalid.
      * This should be called whenever force field parameters change.  It will cause the definitions
@@ -625,6 +649,8 @@ private:
     std::vector<cl::Memory*> autoclearBuffers;
     std::vector<int> autoclearBufferSizes;
     std::vector<ReorderListener*> reorderListeners;
+    std::vector<ForcePreComputation*> preComputations;
+    std::vector<ForcePostComputation*> postComputations;
     OpenCLIntegrationUtilities* integration;
     OpenCLExpressionUtilities* expression;
     OpenCLBondedUtilities* bonded;
@@ -686,7 +712,7 @@ private:
 
 /**
  * This abstract class defines a function to be executed whenever atoms get reordered.
- * Objects that need to know when reordering happens should create a reorderListener
+ * Objects that need to know when reordering happens should create a ReorderListener
  * and register it by calling addReorderListener().
  */
 class OpenCLContext::ReorderListener {
@@ -696,6 +722,39 @@ public:
     }
 };
 
+/**
+ * This abstract class defines a function to be executed at the very beginning of force and
+ * energy evaluation, before any other calculation has been done.  It is useful for operations
+ * that need to be performed at a nonstandard point in the process.  After creating a
+ * ForcePreComputation, register it by calling addForcePreComputation().
+ */
+class OpenCLContext::ForcePreComputation {
+public:
+    /**
+     * @param includeForce  true if forces should be computed
+     * @param includeEnergy true if potential energy should be computed
+     * @param groups        a set of bit flags for which force groups to include
+     */
+    virtual void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
+};
+
+/**
+ * This abstract class defines a function to be executed at the very end of force and
+ * energy evaluation, after all other calculations have been done.  It is useful for operations
+ * that need to be performed at a nonstandard point in the process.  After creating a
+ * ForcePostComputation, register it by calling addForcePostComputation().
+ */
+class OpenCLContext::ForcePostComputation {
+public:
+    /**
+     * @param includeForce  true if forces should be computed
+     * @param includeEnergy true if potential energy should be computed
+     * @param groups        a set of bit flags for which force groups to include
+     * @return an optional contribution to add to the potential energy.
+     */
+    virtual double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
+};
+
 } // namespace OpenMM
 
 #endif /*OPENMM_OPENCLCONTEXT_H_*/

platforms/opencl/src/OpenCLKernels.h → platforms/opencl/include/OpenCLKernels.h

@@ -557,7 +557,7 @@ public:
     OpenCLCalcNonbondedForceKernel(std::string name, const Platform& platform, OpenCLContext& cl, const System& system) : CalcNonbondedForceKernel(name, platform),
             hasInitializedKernel(false), cl(cl), sigmaEpsilon(NULL), exceptionParams(NULL), cosSinSums(NULL), pmeGrid(NULL),
             pmeGrid2(NULL), pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeBsplineTheta(NULL),
-            pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL), fft(NULL) {
+            pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL), fft(NULL), pmeio(NULL) {
     }
     ~OpenCLCalcNonbondedForceKernel();
     /**
@@ -596,6 +596,9 @@ private:
         const char* getMaxValue() const {return "(int2) (INT_MAX, INT_MAX)";}
         const char* getSortKey() const {return "value.y";}
     };
+    class PmeIO;
+    class PmePreComputation;
+    class PmePostComputation;
     OpenCLContext& cl;
     bool hasInitializedKernel;
     OpenCLArray* sigmaEpsilon;
@@ -611,6 +614,8 @@ private:
     OpenCLArray* pmeAtomGridIndex;
     OpenCLSort* sort;
     OpenCLFFT3D* fft;
+    Kernel cpuPme;
+    PmeIO* pmeio;
     cl::Kernel ewaldSumsKernel;
     cl::Kernel ewaldForcesKernel;
     cl::Kernel pmeGridIndexKernel;

platforms/opencl/include/OpenCLPlatform.h

@@ -87,17 +87,25 @@ public:
         static const std::string key = "OpenCLPrecision";
         return key;
     }
+    /**
+     * This is the name of the parameter for selecting whether to use the CPU based PME calculation.
+     */
+    static const std::string& OpenCLUseCpuPme() {
+        static const std::string key = "OpenCLUseCpuPme";
+        return key;
+    }
 };
 
 class OPENMM_EXPORT_OPENCL OpenCLPlatform::PlatformData {
 public:
-    PlatformData(const System& system, const std::string& platformPropValue, const std::string& deviceIndexProperty, const std::string& precisionProperty);
+    PlatformData(const System& system, const std::string& platformPropValue, const std::string& deviceIndexProperty, const std::string& precisionProperty, const std::string& cpuPmeProperty);
     ~PlatformData();
     void initializeContexts(const System& system);
     void syncContexts();
+    ContextImpl* context;
     std::vector<OpenCLContext*> contexts;
     std::vector<double> contextEnergy;
-    bool removeCM;
+    bool removeCM, useCpuPme;
     int cmMotionFrequency;
     int stepCount, computeForceCount;
     double time;

platforms/opencl/src/OpenCLContext.cpp

@@ -334,6 +334,10 @@ OpenCLContext::~OpenCLContext() {
         delete forces[i];
     for (int i = 0; i < (int) reorderListeners.size(); i++)
         delete reorderListeners[i];
+    for (int i = 0; i < (int) preComputations.size(); i++)
+        delete preComputations[i];
+    for (int i = 0; i < (int) postComputations.size(); i++)
+        delete postComputations[i];
     if (pinnedBuffer != NULL)
         delete pinnedBuffer;
     if (posq != NULL)
@@ -1106,6 +1110,14 @@ void OpenCLContext::addReorderListener(ReorderListener* listener) {
     reorderListeners.push_back(listener);
 }
 
+void OpenCLContext::addPreComputation(ForcePreComputation* computation) {
+    preComputations.push_back(computation);
+}
+
+void OpenCLContext::addPostComputation(ForcePostComputation* computation) {
+    postComputations.push_back(computation);
+}
+
 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) :

platforms/opencl/src/OpenCLFFT3D.cpp

@@ -27,7 +27,7 @@
 #include "OpenCLFFT3D.h"
 #include "OpenCLExpressionUtilities.h"
 #include "OpenCLKernelSources.h"
-#include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
+#include "SimTKOpenMMRealType.h"
 #include <map>
 #include <sstream>
 #include <string>

platforms/opencl/src/OpenCLPlatform.cpp

@@ -31,6 +31,7 @@
 #include "openmm/internal/ContextImpl.h"
 #include "openmm/Context.h"
 #include "openmm/System.h"
+#include <algorithm>
 #include <sstream>
 
 using namespace OpenMM;
@@ -78,11 +79,13 @@ OpenCLPlatform::OpenCLPlatform() {
     platformProperties.push_back(OpenCLPlatformIndex());
     platformProperties.push_back(OpenCLPlatformName());
     platformProperties.push_back(OpenCLPrecision());
+    platformProperties.push_back(OpenCLUseCpuPme());
     setPropertyDefaultValue(OpenCLDeviceIndex(), "");
     setPropertyDefaultValue(OpenCLDeviceName(), "");
     setPropertyDefaultValue(OpenCLPlatformIndex(), "");
     setPropertyDefaultValue(OpenCLPlatformName(), "");
     setPropertyDefaultValue(OpenCLPrecision(), "single");
+    setPropertyDefaultValue(OpenCLUseCpuPme(), "false");
 }
 
 double OpenCLPlatform::getSpeed() const {
@@ -112,7 +115,15 @@ void OpenCLPlatform::contextCreated(ContextImpl& context, const map<string, stri
             getPropertyDefaultValue(OpenCLDeviceIndex()) : properties.find(OpenCLDeviceIndex())->second);
     string precisionPropValue = (properties.find(OpenCLPrecision()) == properties.end() ?
             getPropertyDefaultValue(OpenCLPrecision()) : properties.find(OpenCLPrecision())->second);
-    context.setPlatformData(new PlatformData(context.getSystem(), platformPropValue, devicePropValue, precisionPropValue));
+    string cpuPmePropValue = (properties.find(OpenCLUseCpuPme()) == properties.end() ?
+            getPropertyDefaultValue(OpenCLUseCpuPme()) : properties.find(OpenCLUseCpuPme())->second);
+    transform(precisionPropValue.begin(), precisionPropValue.end(), precisionPropValue.begin(), ::tolower);
+    transform(cpuPmePropValue.begin(), cpuPmePropValue.end(), cpuPmePropValue.begin(), ::tolower);
+    vector<string> pmeKernelName;
+    pmeKernelName.push_back(CalcPmeReciprocalForceKernel::Name());
+    if (!supportsKernels(pmeKernelName))
+        cpuPmePropValue = "false";
+    context.setPlatformData(new PlatformData(context.getSystem(), platformPropValue, devicePropValue, precisionPropValue, cpuPmePropValue));
 }
 
 void OpenCLPlatform::contextDestroyed(ContextImpl& context) const {
@@ -121,7 +132,7 @@ void OpenCLPlatform::contextDestroyed(ContextImpl& context) const {
 }
 
 OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& platformPropValue, const string& deviceIndexProperty,
-        const string& precisionProperty) : removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
+        const string& precisionProperty, const string& cpuPmeProperty) : removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
     int platformIndex = 0;
     if (platformPropValue.length() > 0)
         stringstream(platformPropValue) >> platformIndex;
@@ -150,6 +161,7 @@ OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& p
         deviceIndex << contexts[i]->getDeviceIndex();
         deviceName << contexts[i]->getDevice().getInfo<CL_DEVICE_NAME>();
     }
+    useCpuPme = (cpuPmeProperty == "true" && !contexts[0]->getUseDoublePrecision());
     propertyValues[OpenCLPlatform::OpenCLDeviceIndex()] = deviceIndex.str();
     propertyValues[OpenCLPlatform::OpenCLDeviceName()] = deviceName.str();
     propertyValues[OpenCLPlatform::OpenCLPlatformIndex()] = contexts[0]->intToString(platformIndex);
@@ -157,6 +169,7 @@ OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& p
     cl::Platform::get(&platforms);
     propertyValues[OpenCLPlatform::OpenCLPlatformName()] = platforms[platformIndex].getInfo<CL_PLATFORM_NAME>();
     propertyValues[OpenCLPlatform::OpenCLPrecision()] = precisionProperty;
+    propertyValues[OpenCLPlatform::OpenCLUseCpuPme()] = useCpuPme ? "true" : "false";
     contextEnergy.resize(contexts.size());
 }
 

platforms/opencl/src/kernels/customIntegrator.cl

@@ -52,10 +52,10 @@ __kernel void applyPositionDeltas(__global real4* restrict posq, __global real4*
     }
 }
 
-__kernel void generateRandomNumbers(__global float4* restrict random, __global uint4* restrict seed) {
+__kernel void generateRandomNumbers(int numValues, __global float4* restrict random, __global uint4* restrict seed) {
     uint4 state = seed[get_global_id(0)];
     unsigned int carry = 0;
-    for (int index = get_global_id(0); index < NUM_ATOMS; index += get_global_size(0)) {
+    for (int index = get_global_id(0); index < numValues; index += get_global_size(0)) {
         // Generate three uniform random numbers.
 
         state.x = state.x * 69069 + 1;