Merge pull request #1679 from peastman/customintegrator

Fixes to CustomIntegrator

platforms/cuda/src/CudaKernels.cpp

@@ -7109,6 +7109,7 @@ void CudaIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context,
                 needsForces[step-1] = (needsForces[step] || needsForces[step-1]);
                 needsEnergy[step-1] = (needsEnergy[step] || needsEnergy[step-1]);
                 needsGlobals[step-1] = (needsGlobals[step] || needsGlobals[step-1]);
+                computeBothForceAndEnergy[step-1] = (computeBothForceAndEnergy[step] || computeBothForceAndEnergy[step-1]);
             }
         
         // Loop over all steps and create the kernels for them.
@@ -7187,7 +7188,7 @@ void CudaIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context,
                 args1.push_back(NULL);
                 args1.push_back(NULL);
                 args1.push_back(NULL);
-                if (cu.getUseDoublePrecision())
+                if (cu.getUseDoublePrecision() || cu.getUseMixedPrecision())
                     args1.push_back(&energy);
                 else
                     args1.push_back(&energyFloat);
@@ -7282,7 +7283,7 @@ void CudaIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context,
         kineticEnergyArgs.push_back(NULL);
         kineticEnergyArgs.push_back(NULL);
         kineticEnergyArgs.push_back(&uniformRandoms->getDevicePointer());
-        if (cu.getUseDoublePrecision())
+        if (cu.getUseDoublePrecision() || cu.getUseMixedPrecision())
             kineticEnergyArgs.push_back(&energy);
         else
             kineticEnergyArgs.push_back(&energyFloat);

platforms/cuda/src/kernels/customIntegratorPerDof.cu

@@ -34,7 +34,7 @@ inline __device__ mixed4 convertFromDouble4(double4 a) {
 extern "C" __global__ void computePerDof(real4* __restrict__ posq, real4* __restrict__ posqCorrection, mixed4* __restrict__ posDelta,
         mixed4* __restrict__ velm, const long long* __restrict__ force, const mixed2* __restrict__ dt, const mixed* __restrict__ globals,
         mixed* __restrict__ sum, const float4* __restrict__ gaussianValues, unsigned int gaussianBaseIndex, const float4* __restrict__ uniformValues,
-        const real energy, mixed* __restrict__ energyParamDerivs
+        const mixed energy, mixed* __restrict__ energyParamDerivs
         PARAMETER_ARGUMENTS) {
     mixed stepSize = dt[0].y;
     int index = blockIdx.x*blockDim.x+threadIdx.x;

platforms/opencl/src/OpenCLKernels.cpp

@@ -7398,6 +7398,7 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
                 needsForces[step-1] = (needsForces[step] || needsForces[step-1]);
                 needsEnergy[step-1] = (needsEnergy[step] || needsEnergy[step-1]);
                 needsGlobals[step-1] = (needsGlobals[step] || needsGlobals[step-1]);
+                computeBothForceAndEnergy[step-1] = (computeBothForceAndEnergy[step] || computeBothForceAndEnergy[step-1]);
             }
         
         // Loop over all steps and create the kernels for them.
@@ -7568,7 +7569,7 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
         kineticEnergyKernel.setArg<cl::Buffer>(index++, sumBuffer->getDeviceBuffer());
         index += 2;
         kineticEnergyKernel.setArg<cl::Buffer>(index++, uniformRandoms->getDeviceBuffer());
-        if (cl.getUseDoublePrecision())
+        if (cl.getUseDoublePrecision() || cl.getUseMixedPrecision())
             kineticEnergyKernel.setArg<cl_double>(index++, 0.0);
         else
             kineticEnergyKernel.setArg<cl_float>(index++, 0.0f);
@@ -7717,7 +7718,7 @@ void OpenCLIntegrateCustomStepKernel::execute(ContextImpl& context, CustomIntegr
             kernels[step][0].setArg<cl_uint>(9, integration.prepareRandomNumbers(requiredGaussian[step]));
             kernels[step][0].setArg<cl::Buffer>(8, integration.getRandom().getDeviceBuffer());
             kernels[step][0].setArg<cl::Buffer>(10, uniformRandoms->getDeviceBuffer());
-            if (cl.getUseDoublePrecision())
+            if (cl.getUseDoublePrecision() || cl.getUseMixedPrecision())
                 kernels[step][0].setArg<cl_double>(11, energy);
             else
                 kernels[step][0].setArg<cl_float>(11, (cl_float) energy);
@@ -7735,7 +7736,7 @@ void OpenCLIntegrateCustomStepKernel::execute(ContextImpl& context, CustomIntegr
             kernels[step][0].setArg<cl_uint>(9, integration.prepareRandomNumbers(requiredGaussian[step]));
             kernels[step][0].setArg<cl::Buffer>(8, integration.getRandom().getDeviceBuffer());
             kernels[step][0].setArg<cl::Buffer>(10, uniformRandoms->getDeviceBuffer());
-            if (cl.getUseDoublePrecision())
+            if (cl.getUseDoublePrecision() || cl.getUseMixedPrecision())
                 kernels[step][0].setArg<cl_double>(11, energy);
             else
                 kernels[step][0].setArg<cl_float>(11, (cl_float) energy);

platforms/opencl/src/kernels/customIntegratorPerDof.cl

@@ -26,7 +26,7 @@ void storePos(__global real4* restrict posq, __global real4* restrict posqCorrec
 __kernel void computePerDof(__global real4* restrict posq, __global real4* restrict posqCorrection, __global mixed4* restrict posDelta,
         __global mixed4* restrict velm, __global const real4* restrict force, __global const mixed2* restrict dt, __global const mixed* restrict globals,
         __global mixed* restrict sum, __global const float4* restrict gaussianValues, unsigned int gaussianBaseIndex, __global const float4* restrict uniformValues,
-        const real energy, __global mixed* restrict energyParamDerivs
+        const mixed energy, __global mixed* restrict energyParamDerivs
         PARAMETER_ARGUMENTS) {
     mixed stepSize = dt[0].y;
     int index = get_global_id(0);