Prevent custom forces from trying to use more shared memory than is available

platforms/cuda/src/kernels/gpu.cpp

@@ -1871,6 +1871,7 @@ void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync)
     gpu->sim.nonbond_blocks = deviceProp.multiProcessorCount*gpu->blocksPerSM;
     gpu->sim.bornForce2_blocks = deviceProp.multiProcessorCount*gpu->blocksPerSM;
     gpu->sim.blocks = deviceProp.multiProcessorCount;
+    gpu->sharedMemoryPerBlock = deviceProp.sharedMemPerBlock;
 
     gpu->sim.shake_threads_per_block                = gpu->sim.max_shake_threads_per_block;
     gpu->sim.localForces_threads_per_block          = gpu->sim.max_localForces_threads_per_block;

platforms/cuda/src/kernels/gputypes.h

@@ -72,6 +72,7 @@ struct _gpuContext {
     gpuAtomType* gpAtomTable;
     int gAtomTypes;
     unsigned int blocksPerSM;
+    unsigned int sharedMemoryPerBlock;
     cudaGmxSimulation sim;
     unsigned int* pOutputBufferCounter;
     std::vector<std::vector<int> > exclusions;

platforms/cuda/src/kernels/kCalculateCustomAngleForces.cu

@@ -156,7 +156,9 @@ void kCalculateCustomAngleForces_kernel()
 void kCalculateCustomAngleForces(gpuContext gpu)
 {
 //    printf("kCalculateCustomAngleForces\n");
-    kCalculateCustomAngleForces_kernel<<<gpu->sim.blocks, gpu->sim.localForces_threads_per_block,
-            (gpu->sim.customExpressionStackSize+9)*sizeof(float)*gpu->sim.localForces_threads_per_block>>>();
+    int memoryPerThread = (gpu->sim.customExpressionStackSize+9)*sizeof(float);
+    int maxThreads = (gpu->sharedMemoryPerBlock-16)/memoryPerThread;
+    int threads = min(gpu->sim.localForces_threads_per_block, (maxThreads/64)*64);
+    kCalculateCustomAngleForces_kernel<<<gpu->sim.blocks, threads, memoryPerThread*threads>>>();
     LAUNCHERROR("kCalculateCustomAngleForces");
 }

platforms/cuda/src/kernels/kCalculateCustomBondForces.cu

@@ -135,7 +135,9 @@ void kCalculateCustomBondForces_kernel()
 void kCalculateCustomBondForces(gpuContext gpu)
 {
 //    printf("kCalculateCustomBondForces\n");
-    kCalculateCustomBondForces_kernel<<<gpu->sim.blocks, gpu->sim.localForces_threads_per_block,
-            (gpu->sim.customExpressionStackSize+9)*sizeof(float)*gpu->sim.localForces_threads_per_block>>>();
+    int memoryPerThread = (gpu->sim.customExpressionStackSize+9)*sizeof(float);
+    int maxThreads = (gpu->sharedMemoryPerBlock-16)/memoryPerThread;
+    int threads = min(gpu->sim.localForces_threads_per_block, (maxThreads/64)*64);
+    kCalculateCustomBondForces_kernel<<<gpu->sim.blocks, threads, memoryPerThread*threads>>>();
     LAUNCHERROR("kCalculateCustomBondForces");
 }

platforms/cuda/src/kernels/kCalculateCustomExternalForces.cu

@@ -123,7 +123,9 @@ void kCalculateCustomExternalForces_kernel()
 void kCalculateCustomExternalForces(gpuContext gpu)
 {
 //    printf("kCalculateCustomExternalForces\n");
-    kCalculateCustomExternalForces_kernel<<<gpu->sim.blocks, gpu->sim.localForces_threads_per_block,
-            (gpu->sim.customExpressionStackSize+9)*sizeof(float)*gpu->sim.localForces_threads_per_block>>>();
+    int memoryPerThread = (gpu->sim.customExpressionStackSize+9)*sizeof(float);
+    int maxThreads = (gpu->sharedMemoryPerBlock-16)/memoryPerThread;
+    int threads = min(gpu->sim.localForces_threads_per_block, (maxThreads/64)*64);
+    kCalculateCustomExternalForces_kernel<<<gpu->sim.blocks, threads, memoryPerThread*threads>>>();
     LAUNCHERROR("kCalculateCustomExternalForces");
 }

platforms/cuda/src/kernels/kCalculateCustomNonbondedForces.cu

@@ -150,7 +150,7 @@ void kCalculateCustomNonbondedForces(gpuContext gpu, bool neighborListValid)
     if (gpu->sim.customNonbondedMethod != NO_CUTOFF)
         sharedPerThread += sizeof(float3);
     int threads = gpu->sim.nonbond_threads_per_block;
-    int maxThreads = 16380/sharedPerThread;
+    int maxThreads = (gpu->sharedMemoryPerBlock-16)/sharedPerThread;
     if (threads > maxThreads)
         threads = (maxThreads/32)*32;
     switch (gpu->sim.customNonbondedMethod)

platforms/cuda/src/kernels/kCalculateCustomTorsionForces.cu

@@ -195,7 +195,9 @@ void kCalculateCustomTorsionForces_kernel()
 void kCalculateCustomTorsionForces(gpuContext gpu)
 {
 //    printf("kCalculateCustomTorsionForces\n");
-    kCalculateCustomTorsionForces_kernel<<<gpu->sim.blocks, gpu->sim.localForces_threads_per_block,
-            (gpu->sim.customExpressionStackSize+9)*sizeof(float)*gpu->sim.localForces_threads_per_block>>>();
+    int memoryPerThread = (gpu->sim.customExpressionStackSize+9)*sizeof(float);
+    int maxThreads = (gpu->sharedMemoryPerBlock-16)/memoryPerThread;
+    int threads = min(gpu->sim.localForces_threads_per_block, (maxThreads/64)*64);
+    kCalculateCustomTorsionForces_kernel<<<gpu->sim.blocks, threads, memoryPerThread*threads>>>();
     LAUNCHERROR("kCalculateCustomTorsionForces");
 }