Disabled compiler optimizations for integrator kernels, since they caused excessive error on Fermi

platforms/opencl/src/OpenCLContext.cpp

@@ -86,7 +86,8 @@ OpenCLContext::OpenCLContext(int numParticles, int deviceIndex) : time(0.0), ste
         this->deviceIndex = deviceIndex;
         if (device.getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>() < minThreadBlockSize)
             throw OpenMMException("The specified OpenCL device is not compatible with OpenMM");
-        compilationOptions = "-cl-fast-relaxed-math -DWORK_GROUP_SIZE="+OpenCLExpressionUtilities::intToString(ThreadBlockSize);
+        compilationOptions = "-DWORK_GROUP_SIZE="+OpenCLExpressionUtilities::intToString(ThreadBlockSize);
+        defaultOptimizationOptions = "-cl-fast-relaxed-math";
         string vendor = device.getInfo<CL_DEVICE_VENDOR>();
         if (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA") {
             compilationOptions += " -DWARPS_ARE_ATOMIC";
@@ -243,18 +244,24 @@ string OpenCLContext::replaceStrings(const string& input, const std::map<std::st
     return result;
 }
 
-cl::Program OpenCLContext::createProgram(const string source) {
-    return createProgram(source, map<string, string>());
+cl::Program OpenCLContext::createProgram(const string source, const char* optimizationFlags) {
+    return createProgram(source, map<string, string>(), optimizationFlags);
 }
 
-cl::Program OpenCLContext::createProgram(const string source, const map<string, string>& defines) {
+cl::Program OpenCLContext::createProgram(const string source, const map<string, string>& defines, const char* optimizationFlags) {
     cl::Program::Sources sources(1, make_pair(source.c_str(), source.size()));
     cl::Program program(context, sources);
-    string options = compilationOptions;
+    stringstream options;
+    options << compilationOptions;
+    if (optimizationFlags == NULL)
+        options << " " << defaultOptimizationOptions;
+    else
+        options << " " << optimizationFlags;
     for (map<string, string>::const_iterator iter = defines.begin(); iter != defines.end(); ++iter)
-        options += " -D"+iter->first+"="+iter->second;
+        options << " -D" << iter->first << "=" << iter->second;
     try {
-        program.build(vector<cl::Device>(1, device), options.c_str());
+        std::cout << options.str()<< std::endl;
+        program.build(vector<cl::Device>(1, device), options.str().c_str());
     } catch (cl::Error err) {
         throw OpenMMException("Error compiling kernel: "+program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device));
     }

platforms/opencl/src/OpenCLContext.h

@@ -230,14 +230,21 @@ public:
     std::string replaceStrings(const std::string& input, const std::map<std::string, std::string>& replacements) const;
     /**
      * Create an OpenCL Program from source code.
+     *
+     * @param source             the source code of the program
+     * @param optimizationFlags  the optimization flags to pass to the OpenCL compiler.  If this is
+     *                           omitted, a default set of options will be used
      */
-    cl::Program createProgram(const std::string source);
+    cl::Program createProgram(const std::string source, const char* optimizationFlags = NULL);
     /**
      * Create an OpenCL Program from source code.
      *
-     * @param defines    a set of preprocessor definitions (name, value) to define when compiling the program
+     * @param source             the source code of the program
+     * @param defines            a set of preprocessor definitions (name, value) to define when compiling the program
+     * @param optimizationFlags  the optimization flags to pass to the OpenCL compiler.  If this is
+     *                           omitted, a default set of options will be used
      */
-    cl::Program createProgram(const std::string source, const std::map<std::string, std::string>& defines);
+    cl::Program createProgram(const std::string source, const std::map<std::string, std::string>& defines, const char* optimizationFlags = NULL);
     /**
      * Execute a kernel.
      *
@@ -406,7 +413,7 @@ private:
     int simdWidth;
     mm_float4 periodicBoxSize;
     mm_float4 invPeriodicBoxSize;
-    std::string compilationOptions;
+    std::string compilationOptions, defaultOptimizationOptions;
     cl::Context context;
     cl::Device device;
     cl::CommandQueue queue;

platforms/opencl/src/OpenCLKernels.cpp

@@ -3236,7 +3236,7 @@ OpenCLIntegrateVerletStepKernel::~OpenCLIntegrateVerletStepKernel() {
 
 void OpenCLIntegrateVerletStepKernel::initialize(const System& system, const VerletIntegrator& integrator) {
     cl.initialize(system);
-    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet);
+    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
     kernel1 = cl::Kernel(program, "integrateVerletPart1");
     kernel2 = cl::Kernel(program, "integrateVerletPart2");
     prevStepSize = -1.0;
@@ -3296,7 +3296,7 @@ void OpenCLIntegrateLangevinStepKernel::initialize(const System& system, const L
     map<string, string> defines;
     defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
     defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
-    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines);
+    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
     kernel1 = cl::Kernel(program, "integrateLangevinPart1");
     kernel2 = cl::Kernel(program, "integrateLangevinPart2");
     params = new OpenCLArray<cl_float>(cl, 3, "langevinParams");
@@ -3369,7 +3369,7 @@ void OpenCLIntegrateBrownianStepKernel::initialize(const System& system, const B
     cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
     map<string, string> defines;
     defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
-    cl::Program program = cl.createProgram(OpenCLKernelSources::brownian, defines);
+    cl::Program program = cl.createProgram(OpenCLKernelSources::brownian, defines, "");
     kernel1 = cl::Kernel(program, "integrateBrownianPart1");
     kernel2 = cl::Kernel(program, "integrateBrownianPart2");
     prevStepSize = -1.0;
@@ -3425,7 +3425,7 @@ OpenCLIntegrateVariableVerletStepKernel::~OpenCLIntegrateVariableVerletStepKerne
 
 void OpenCLIntegrateVariableVerletStepKernel::initialize(const System& system, const VariableVerletIntegrator& integrator) {
     cl.initialize(system);
-    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet);
+    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
     kernel1 = cl::Kernel(program, "integrateVerletPart1");
     kernel2 = cl::Kernel(program, "integrateVerletPart2");
     selectSizeKernel = cl::Kernel(program, "selectVerletStepSize");
@@ -3496,7 +3496,7 @@ void OpenCLIntegrateVariableLangevinStepKernel::initialize(const System& system,
     map<string, string> defines;
     defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
     defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
-    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines);
+    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
     kernel1 = cl::Kernel(program, "integrateLangevinPart1");
     kernel2 = cl::Kernel(program, "integrateLangevinPart2");
     selectSizeKernel = cl::Kernel(program, "selectLangevinStepSize");