From d787b30379c9b251ba749fdc1cea3dff63766c96 Mon Sep 17 00:00:00 2001
From: castano <castano@95f4ed2b-212e-0410-8b90-d31948207fce>
Date: Mon, 1 Dec 2008 07:56:00 +0000
Subject: [PATCH] Select fastest CUDA device. Merged from trunk.

---
 ChangeLog                   |  2 +-
 src/nvtt/Compressor.cpp     |  8 ++++++++
 src/nvtt/cuda/CudaUtils.cpp | 29 +++++++++++++++++++++++++++++
 src/nvtt/cuda/CudaUtils.h   |  1 +
 4 files changed, 39 insertions(+), 1 deletion(-)

diff --git a/ChangeLog b/ChangeLog
index f09e57d..fc19c80 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -3,7 +3,7 @@ NVIDIA Texture Tools version 2.0.5
  * Detect mismatch between CUDA runtime and driver, and disable CUDA in that case.
  * Fix cmake files when compiling NVTT as a shared library.
  * When linking nvtt dynamically on unix, link all libraries dynamically.
- * Merge fixes from trunk.
+ * Select fastest CUDA device.
 
 NVIDIA Texture Tools version 2.0.4
  * Fix error in RGB format output; reported by jonsoh. See issue 49.
diff --git a/src/nvtt/Compressor.cpp b/src/nvtt/Compressor.cpp
index d070d77..d497e8a 100644
--- a/src/nvtt/Compressor.cpp
+++ b/src/nvtt/Compressor.cpp
@@ -211,6 +211,10 @@ 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())
@@ -237,6 +241,10 @@ 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())
diff --git a/src/nvtt/cuda/CudaUtils.cpp b/src/nvtt/cuda/CudaUtils.cpp
index 6cbdb74..784527f 100644
--- a/src/nvtt/cuda/CudaUtils.cpp
+++ b/src/nvtt/cuda/CudaUtils.cpp
@@ -151,6 +151,35 @@ int nv::cuda::deviceCount()
 	return 0;
 }
 
+int nv::cuda::getFastestDevice()
+{
+	int max_gflops_device = 0;
+#if defined HAVE_CUDA	
+	int max_gflops = 0;
+
+	const int device_count = deviceCount();
+	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++;
+	}
+#endif
+	return max_gflops_device;
+}
+
 /// Activate the given devices.
 bool nv::cuda::setDevice(int i)
 {
diff --git a/src/nvtt/cuda/CudaUtils.h b/src/nvtt/cuda/CudaUtils.h
index 5a5bdab..6965960 100644
--- a/src/nvtt/cuda/CudaUtils.h
+++ b/src/nvtt/cuda/CudaUtils.h
@@ -31,6 +31,7 @@ namespace nv
 	{
 		bool isHardwarePresent();
 		int deviceCount();
+		int getFastestDevice();
 		bool setDevice(int i);
 	};