Optimization (clear the force and Born sum buffers in a single kernel)

platforms/cuda/src/CudaKernels.cpp

@@ -47,6 +47,9 @@ void CudaCalcForcesAndEnergyKernel::beginForceComputation(ContextImpl& context)
     if (data.nonbondedMethod != NO_CUTOFF && data.computeForceCount%100 == 0)
         gpuReorderAtoms(gpu);
     data.computeForceCount++;
+    if (gpu->bIncludeGBSA || gpu->bIncludeGBVI)
+        kClearBornSumAndForces(gpu);
+    else
         kClearForces(gpu);
 }
 
@@ -77,6 +80,8 @@ void CudaCalcForcesAndEnergyKernel::beginEnergyComputation(ContextImpl& context)
         gpuReorderAtoms(gpu);
     data.stepCount++;
     kClearEnergy(gpu);
+    if (gpu->bIncludeGBSA || gpu->bIncludeGBVI)
+        kClearBornSumAndForces(gpu);
 }
 
 double CudaCalcForcesAndEnergyKernel::finishEnergyComputation(ContextImpl& context) {
@@ -790,7 +795,9 @@ void OPENMMCUDA_EXPORT OpenMM::cudaOpenMMInitializeIntegration(const System& sys
     gpuBuildExclusionList(gpu);
     gpuBuildOutputBuffers(gpu);
     gpuSetConstants(gpu);
-    kClearBornForces(gpu);
+    if (gpu->bIncludeGBSA || gpu->bIncludeGBVI)
+        kClearBornSumAndForces(gpu);
+    else
         kClearForces(gpu);
     cudaThreadSynchronize();
 }

platforms/cuda/src/kernels/cudaKernels.h

@@ -29,7 +29,7 @@
 // Initialization
 extern void kClearForces(gpuContext gpu);
 extern void kClearEnergy(gpuContext gpu);
-extern void kClearBornForces(gpuContext gpu);
+extern void kClearBornSumAndForces(gpuContext gpu);
 extern void kClearObcGbsaBornSum(gpuContext gpu);
 extern void kCalculateObcGbsaBornSum(gpuContext gpu);
 extern void kReduceObcGbsaBornSum(gpuContext gpu);

platforms/cuda/src/kernels/kCalculateCDLJObcGbsaForces1.cu

@@ -134,10 +134,6 @@ extern void kCalculatePME(gpuContext gpu);
 void kCalculateCDLJObcGbsaForces1(gpuContext gpu)
 {
 //    printf("kCalculateCDLJObcGbsaForces1\n");
-
-    // check if Born radii need to be calculated
-
-    kClearBornForces(gpu);
     switch (gpu->sim.nonbondedMethod)
     {
         case NO_CUTOFF:

platforms/cuda/src/kernels/kCalculateGBVIBornSum.cu

@@ -102,21 +102,6 @@ void GetCalculateGBVIBornSumSim(gpuContext gpu)
 #define METHOD_NAME(a, b) a##PeriodicByWarp##b
 #include "kCalculateGBVIBornSum.h"
 
-
-__global__ void kClearGBVIBornSum_kernel()
-{
-    unsigned int pos = blockIdx.x * blockDim.x + threadIdx.x;
-    while (pos < cSim.stride * cSim.nonbondOutputBuffers)
-    {
-        ((float*)cSim.pBornSum)[pos] = 0.0f;
-        pos += gridDim.x * blockDim.x;
-    }
-}
-
-void kClearGBVIBornSum(gpuContext gpu) {
-    kClearGBVIBornSum_kernel<<<gpu->sim.blocks, 384>>>();
-}
-
 __global__ void kReduceGBVIBornSum_kernel()
 {
     unsigned int pos = (blockIdx.x * blockDim.x + threadIdx.x);
@@ -177,8 +162,6 @@ void kReduceGBVIBornSum(gpuContext gpu)
 void kCalculateGBVIBornSum(gpuContext gpu)
 {
     //printf("kCalculateGBVIBornSum\n");
-    kClearGBVIBornSum( gpu );
-    LAUNCHERROR("kClearBornSum");
     //size_t numWithInteractions;
     switch (gpu->sim.nonbondedMethod)
     {

platforms/cuda/src/kernels/kCalculateObcGbsaBornSum.cu

@@ -97,19 +97,6 @@ void GetCalculateObcGbsaBornSumSim(gpuContext gpu)
 #define METHOD_NAME(a, b) a##PeriodicByWarp##b
 #include "kCalculateObcGbsaBornSum.h"
 
-
-__global__ 
-__launch_bounds__(384, 1)
-void kClearObcGbsaBornSum_kernel()
-{
-    unsigned int pos = blockIdx.x * blockDim.x + threadIdx.x;
-    while (pos < cSim.stride * cSim.nonbondOutputBuffers)
-    {
-        ((float*)cSim.pBornSum)[pos] = 0.0f;
-        pos += gridDim.x * blockDim.x;
-    }
-}
-
 __global__ 
 __launch_bounds__(384, 1)
 void kReduceObcGbsaBornSum_kernel()
@@ -154,17 +141,9 @@ void kReduceObcGbsaBornSum(gpuContext gpu)
     LAUNCHERROR("kReduceObcGbsaBornSum");
 }
 
-extern void kClearObcGbsaBornSum(gpuContext gpu)
-{
-  //  printf("kClearObcGbsaBornSum\n");
-    kClearObcGbsaBornSum_kernel<<<gpu->sim.blocks, 384>>>();
-}
-
 void kCalculateObcGbsaBornSum(gpuContext gpu)
 {
   //  printf("kCalculateObcgbsaBornSum\n");
-    kClearObcGbsaBornSum(gpu);
-    LAUNCHERROR("kClearBornSum");
     switch (gpu->sim.nonbondedMethod)
     {
         case NO_CUTOFF:

platforms/cuda/src/kernels/kForces.cu

@@ -74,21 +74,28 @@ void kClearForces(gpuContext gpu)
 
 __global__ 
 __launch_bounds__(384, 1)
-void kClearBornForces_kernel()
+void kClearBornSumAndForces_kernel()
 {
     unsigned int pos = blockIdx.x * blockDim.x + threadIdx.x;
     while (pos < cSim.stride * cSim.nonbondOutputBuffers)
     {
-        ((float*)cSim.pBornForce)[pos] = 0.0f;
+        cSim.pBornSum[pos] = 0.0f;
+        cSim.pBornForce[pos] = 0.0f;
+        cSim.pForce4[pos] = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+        pos += gridDim.x * blockDim.x;
+    }
+    while (pos < cSim.stride * cSim.outputBuffers)
+    {
+        cSim.pForce4[pos] = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
         pos += gridDim.x * blockDim.x;
     }
 }
 
-void kClearBornForces(gpuContext gpu)
+void kClearBornSumAndForces(gpuContext gpu)
 {
-  //  printf("kClearBornForces\n");
-    kClearBornForces_kernel<<<gpu->sim.blocks, 384>>>();
-    LAUNCHERROR("kClearBornForces");
+  //  printf("kClearBornSumAndForces\n");
+    kClearBornSumAndForces_kernel<<<gpu->sim.blocks, 384>>>();
+    LAUNCHERROR("kClearBornSumAndForces");
 }
 
 __global__