Fixed errors in GB on AMD GPUs

platforms/opencl/src/OpenCLKernels.cpp

@@ -1532,7 +1532,7 @@ double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeF
         computeBornSumKernel.setArg<cl::Buffer>(index++, (useLong ? longBornSum->getDeviceBuffer() : bornSum->getDeviceBuffer()));
         computeBornSumKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
         computeBornSumKernel.setArg<cl::Buffer>(index++, params->getDeviceBuffer());
-        computeBornSumKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*13*sizeof(cl_float), NULL);
+        computeBornSumKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*7*sizeof(cl_float), NULL);
         computeBornSumKernel.setArg(index++, (deviceIsCpu ? 1 : nb.getForceThreadBlockSize())*sizeof(cl_float), NULL);
         if (nb.getUseCutoff()) {
             computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
@@ -1544,8 +1544,10 @@ double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeF
         }
         else
             computeBornSumKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
-        computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getExclusionIndices().getDeviceBuffer());
-        computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getExclusionRowIndices().getDeviceBuffer());
+        if (cl.getSIMDWidth() == 32) {
+            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getExclusionIndices().getDeviceBuffer());
+            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getExclusionRowIndices().getDeviceBuffer());
+        }
         force1Kernel = cl::Kernel(program, "computeGBSAForce1");
         index = 0;
         force1Kernel.setArg<cl::Buffer>(index++, (useLong ? cl.getLongForceBuffer().getDeviceBuffer() : cl.getForceBuffers().getDeviceBuffer()));
@@ -1553,7 +1555,7 @@ double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeF
         force1Kernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
         force1Kernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
         force1Kernel.setArg<cl::Buffer>(index++, bornRadii->getDeviceBuffer());
-        force1Kernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*13*sizeof(cl_float), NULL);
+        force1Kernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*9*sizeof(cl_float), NULL);
         force1Kernel.setArg(index++, (deviceIsCpu ? 1 : nb.getForceThreadBlockSize())*sizeof(mm_float4), NULL);
         if (nb.getUseCutoff()) {
             force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
@@ -1565,8 +1567,10 @@ double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeF
         }
         else
             force1Kernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
-        force1Kernel.setArg<cl::Buffer>(index++, nb.getExclusionIndices().getDeviceBuffer());
-        force1Kernel.setArg<cl::Buffer>(index++, nb.getExclusionRowIndices().getDeviceBuffer());
+        if (cl.getSIMDWidth() == 32) {
+            force1Kernel.setArg<cl::Buffer>(index++, nb.getExclusionIndices().getDeviceBuffer());
+            force1Kernel.setArg<cl::Buffer>(index++, nb.getExclusionRowIndices().getDeviceBuffer());
+        }
         program = cl.createProgram(OpenCLKernelSources::gbsaObcReductions, defines);
         reduceBornSumKernel = cl::Kernel(program, "reduceBornSum");
         reduceBornSumKernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());

platforms/opencl/src/kernels/gbsaObc_default.cl

@@ -3,12 +3,9 @@
 typedef struct {
     float x, y, z;
     float q;
-    float fx, fy, fz, fw;
     float radius, scaledRadius;
     float bornSum;
-    float bornRadius;
-    float bornForce;
-} AtomData;
+} AtomData1;
 
 /**
  * Compute the Born sum.
@@ -16,7 +13,7 @@ typedef struct {
 
 __kernel __attribute__((reqd_work_group_size(WORK_GROUP_SIZE, 1, 1)))
 void computeBornSum(__global float* global_bornSum, __global float4* posq, __global float2* global_params,
-        __local AtomData* localData, __local float* tempBuffer,
+        __local AtomData1* localData, __local float* tempBuffer,
 #ifdef USE_CUTOFF
         __global ushort2* tiles, __global unsigned int* interactionCount, float4 periodicBoxSize, float4 invPeriodicBoxSize, unsigned int maxTiles) {
 #else
@@ -196,14 +193,21 @@ void computeBornSum(__global float* global_bornSum, __global float4* posq, __glo
     }
 }
 
+typedef struct {
+    float x, y, z;
+    float q;
+    float fx, fy, fz, fw;
+    float bornRadius;
+} AtomData2;
+
 /**
  * First part of computing the GBSA interaction.
  */
 
 __kernel __attribute__((reqd_work_group_size(WORK_GROUP_SIZE, 1, 1)))
-void computeGBSAForce1(__global float4* forceBuffers, __global float* energyBuffer,
-        __global float4* posq, __global float* global_bornRadii, __global float* global_bornForce,
-        __local AtomData* localData, __local float4* tempBuffer,
+void computeGBSAForce1(__global float4* forceBuffers, __global float* global_bornForce,
+        __global float* energyBuffer, __global float4* posq, __global float* global_bornRadii,
+        __local AtomData2* localData, __local float4* tempBuffer,
 #ifdef USE_CUTOFF
         __global ushort2* tiles, __global unsigned int* interactionCount, float4 periodicBoxSize, float4 invPeriodicBoxSize, unsigned int maxTiles) {
 #else