Select fastest device.

src/nvtt/Compressor.cpp

@@ -221,8 +221,12 @@ Compressor::Compressor() : m(*new Compressor::Private())
 
 	if (m.cudaEnabled)
 	{
+		// Select fastest CUDA device.
+		int device = cuda::getFastestDevice();
+		cuda::setDevice(device);
+		
 		m.cuda = new CudaCompressor();
-
+		
 		if (!m.cuda->isValid())
 		{
 			m.cudaEnabled = false;
@@ -247,6 +251,10 @@ void Compressor::enableCudaAcceleration(bool enable)
 
 	if (m.cudaEnabled && m.cuda == NULL)
 	{
+		// Select fastest CUDA device.
+		int device = cuda::getFastestDevice();
+		cuda::setDevice(device);
+		
 		m.cuda = new CudaCompressor();
 		
 		if (!m.cuda->isValid())

src/nvtt/cuda/CudaUtils.cpp

@@ -149,7 +149,37 @@ int nv::cuda::deviceCount()
 	}
 #endif
 	return 0;
-}
+}
+
+int nv::cuda::getFastestDevice()
+{
+	const int device_count = deviceCount();
+
+	int max_gflops_device = 0;
+	int max_gflops = 0;
+
+	int current_device = 0;
+	while (current_device < device_count)
+	{
+		cudaDeviceProp device_properties;
+		cudaGetDeviceProperties(&device_properties, current_device);
+		int gflops = device_properties.multiProcessorCount * device_properties.clockRate;
+
+		if (device_properties.major != -1 && device_properties.minor != -1)
+		{
+			if( gflops > max_gflops )
+			{
+				max_gflops = gflops;
+				max_gflops_device = current_device;
+			}
+		}
+		
+		current_device++;
+	}
+
+	return max_gflops_device;
+}
+
 
 /// Activate the given devices.
 bool nv::cuda::setDevice(int i)

src/nvtt/cuda/CudaUtils.h

@@ -30,7 +30,8 @@ namespace nv
 	namespace cuda
 	{
 		bool isHardwarePresent();
-		int deviceCount();
+		int deviceCount();
+		int getFastestDevice();
 		bool setDevice(int i);
 	};