Improved performance on AMD GPUs

platforms/opencl/include/OpenCLSort.h

@@ -94,7 +94,7 @@ private:
     OpenCLArray buckets;
     cl::Kernel shortListKernel, shortList2Kernel, computeRangeKernel, assignElementsKernel, computeBucketPositionsKernel, copyToBucketsKernel, sortBucketsKernel;
     unsigned int dataLength, rangeKernelSize, positionsKernelSize, sortKernelSize;
-    bool isShortList;
+    bool isShortList, useShortList2;
 };
 
 /**

platforms/opencl/src/OpenCLSort.cpp

@@ -32,6 +32,7 @@
 #include "OpenCLKernelSources.h"
 #include <algorithm>
 #include <map>
+#include <string>
 
 using namespace OpenMM;
 using namespace std;
@@ -91,6 +92,11 @@ OpenCLSort::OpenCLSort(OpenCLContext& context, SortTrait* trait, unsigned int le
         bucketOffset.initialize<cl_uint>(context, numBuckets, "bucketOffset");
         bucketOfElement.initialize<cl_uint>(context, length, "bucketOfElement");
         offsetInBucket.initialize<cl_uint>(context, length, "offsetInBucket");
+        string vendor = context.getDevice().getInfo<CL_DEVICE_VENDOR>();
+        if (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA")
+            useShortList2 = (dataLength <= OpenCLContext::ThreadBlockSize*context.getNumThreadBlocks());
+        else
+            useShortList2 = false;
     }
     buckets.initialize(context, length, trait->getDataSize(), "buckets");
 }
@@ -107,7 +113,7 @@ void OpenCLSort::sort(OpenCLArray& data) {
     if (isShortList) {
         // We can use a simpler sort kernel that does the entire operation in one kernel.
         
-        if (dataLength <= OpenCLContext::ThreadBlockSize*context.getNumThreadBlocks()) {
+        if (useShortList2) {
             shortList2Kernel.setArg<cl::Buffer>(0, data.getDeviceBuffer());
             shortList2Kernel.setArg<cl::Buffer>(1, buckets.getDeviceBuffer());
             shortList2Kernel.setArg<cl_int>(2, dataLength);