sharedMemoryPerBlock was not set correctly for 480

plugins/amoeba/platforms/cuda/src/kernels/AmoebaGpu.cpp

@@ -62,7 +62,15 @@ amoebaGpuContext amoebaGpuInit( _gpuContext* gpu )
     amoebaGpu->log                             = stderr; 
 #endif
     amoebaGpu->numberOfSorWorkVectors          = 4; 
-    amoebaGpu->sharedMemoryPerBlock            = 16384; 
+    
+    if( gpu->sm_version >= SM_20 ){
+        amoebaGpu->sharedMemoryPerBlock = 49152;
+    } else if( gpu->sm_version >= SM_12 ){
+        amoebaGpu->sharedMemoryPerBlock = 16384;
+    } else {
+        amoebaGpu->sharedMemoryPerBlock = 8192;
+    }
+
     amoebaGpu->paddedNumberOfAtoms             = gpu->sim.paddedNumberOfAtoms; 
     amoebaGpu->amoebaSim.numberOfAtoms         = gpu->natoms;
     amoebaGpu->amoebaSim.paddedNumberOfAtoms   = gpu->sim.paddedNumberOfAtoms;

plugins/amoeba/platforms/cuda/src/kernels/kCalculateAmoebaCudaKirkwoodEDiff.cu

@@ -1203,10 +1203,11 @@ void kCalculateAmoebaKirkwoodEDiff( amoebaGpuContext amoebaGpu )
     }   
     
     if( amoebaGpu->log && timestep == 1 ){
-        (void) fprintf( amoebaGpu->log, "kCalculateAmoebaCudaKirkwoodEDiffN2Forces:  numBlocks=%u numThreads=%u bufferPerWarp=%u atm=%lu shrd=%lu Ebuf=%u ixnCt=%lu workUnits=%u\n",
+        (void) fprintf( amoebaGpu->log, "kCalculateAmoebaCudaKirkwoodEDiffN2Forces: blocks=%u threads=%u bffr/Warp=%u atm=%lu shrd=%lu"
+                                        " Ebuf=%u ixnCt=%lu workUnits=%u sm=%d device=%d sharedMemoryPerBlock=%u\n",
                         amoebaGpu->nonbondBlocks, threadsPerBlock, amoebaGpu->bOutputBufferPerWarp,
                         sizeof(KirkwoodEDiffParticle), sizeof(KirkwoodEDiffParticle)*threadsPerBlock,
-                        amoebaGpu->energyOutputBuffers, (*gpu->psInteractionCount)[0], gpu->sim.workUnits );
+                        amoebaGpu->energyOutputBuffers, (*gpu->psInteractionCount)[0], gpu->sim.workUnits, gpu->sm_version, gpu->device, amoebaGpu->sharedMemoryPerBlock );
         //gpuPrintCudaAmoebaGmxSimulation(amoebaGpu, amoebaGpu->log );
         (void) fflush( amoebaGpu->log );
     }   
@@ -1238,14 +1239,6 @@ void kCalculateAmoebaKirkwoodEDiff( amoebaGpuContext amoebaGpu )
 
     } else {
 
-#ifdef AMOEBA_DEBUG
-        (void) fprintf( amoebaGpu->log, "kCalculateAmoebaCudaKirkwoodEDiffN2Forces no warp:  numBlocks=%u numThreads=%u bufferPerWarp=%u atm=%u shrd=%u Ebuf=%u ixnCt=%u workUnits=%u\n",
-                        amoebaGpu->nonbondBlocks, threadsPerBlock, amoebaGpu->bOutputBufferPerWarp,
-                        sizeof(KirkwoodEDiffParticle), sizeof(KirkwoodEDiffParticle)*threadsPerBlock,
-                        amoebaGpu->energyOutputBuffers, (*gpu->psInteractionCount)[0], gpu->sim.workUnits );
-        (void) fflush( amoebaGpu->log );
-#endif
-
         kCalculateAmoebaCudaKirkwoodEDiffN2Forces_kernel<<<amoebaGpu->nonbondBlocks, threadsPerBlock, sizeof(KirkwoodEDiffParticle)*threadsPerBlock>>>(
                                                                            amoebaGpu->psWorkUnit->_pDevStream[0],
                                                                            gpu->psPosq4->_pDevStream[0],