Fix memory leaks.

Shutdown CUDA properly when nvtt context is destroyed.
Fixes issue 83.

src/nvtt/Context.cpp

@@ -224,8 +224,13 @@ Compressor::Compressor() : m(*new Compressor::Private())
 	{
 		// Select fastest CUDA device.
 		int device = cuda::getFastestDevice();
-		cuda::setDevice(device);
-		
+		if (!cuda::setDevice(device))
+		{
+			m.cudaEnabled = false;
+			m.cuda = NULL;
+		}
+		else
+		{
 			m.cuda = new CudaCompressor();
 
 			if (!m.cuda->isValid())
@@ -235,10 +240,12 @@ Compressor::Compressor() : m(*new Compressor::Private())
 			}
 		}
 	}
+}
 
 Compressor::~Compressor()
 {
 	delete &m;
+	cuda::exit();
 }
 
 
@@ -254,8 +261,13 @@ void Compressor::enableCudaAcceleration(bool enable)
 	{
 		// Select fastest CUDA device.
 		int device = cuda::getFastestDevice();
-		cuda::setDevice(device);
-		
+		if (!cuda::setDevice(device))
+		{
+			m.cudaEnabled = false;
+			m.cuda = NULL;
+		}
+		else
+		{
 			m.cuda = new CudaCompressor();
 
 			if (!m.cuda->isValid())
@@ -265,6 +277,7 @@ void Compressor::enableCudaAcceleration(bool enable)
 			}
 		}
 	}
+}
 
 /// Check if CUDA acceleration is enabled.
 bool Compressor::isCudaAccelerationEnabled() const

src/nvtt/cuda/CudaCompressDXT.cpp

@@ -90,20 +90,34 @@ CudaCompressor::CudaCompressor() : m_bitmapTable(NULL), m_bitmapTableCTX(NULL),
 #if defined HAVE_CUDA
     // Allocate and upload bitmaps.
     cudaMalloc((void**) &m_bitmapTable, 992 * sizeof(uint));
+	cudaError_t err = cudaGetLastError();
+	if (err != cudaSuccess) {
+		fprintf(stderr, "CUDA Error: %s\n", cudaGetErrorString(err));
+		fflush(stderr);
+		nvDebugBreak();
+	}
+
 	if (m_bitmapTable != NULL)
 	{
 		cudaMemcpy(m_bitmapTable, s_bitmapTable, 992 * sizeof(uint), cudaMemcpyHostToDevice);
+		if (cudaGetLastError() != cudaSuccess) nvDebugBreak();
 	}
 
     cudaMalloc((void**) &m_bitmapTableCTX, 704 * sizeof(uint));
+	if (cudaGetLastError() != cudaSuccess) nvDebugBreak();
+
 	if (m_bitmapTableCTX != NULL)
 	{
 		cudaMemcpy(m_bitmapTableCTX, s_bitmapTableCTX, 704 * sizeof(uint), cudaMemcpyHostToDevice);
+		if (cudaGetLastError() != cudaSuccess) nvDebugBreak();
 	}
 
 	// Allocate scratch buffers.
     cudaMalloc((void**) &m_data, MAX_BLOCKS * 64U);
+	if (cudaGetLastError() != cudaSuccess) nvDebugBreak();
+
     cudaMalloc((void**) &m_result, MAX_BLOCKS * 8U);
+	if (cudaGetLastError() != cudaSuccess) nvDebugBreak();
 #endif
 }
 
@@ -114,14 +128,17 @@ CudaCompressor::~CudaCompressor()
 	cudaFree(m_data);
 	cudaFree(m_result);
 	cudaFree(m_bitmapTable);
+	cudaFree(m_bitmapTableCTX);
 #endif
 }
 
 bool CudaCompressor::isValid() const
 {
 #if defined HAVE_CUDA
-	if (cudaGetLastError() != cudaSuccess)
+	cudaError_t err = cudaGetLastError();
+	if (err != cudaSuccess)
    	{
+		nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(err));
 		return false;
 	}
 #endif

src/nvtt/cuda/CudaUtils.cpp

@@ -84,19 +84,26 @@ static bool isCudaDriverAvailable(int version)
 	
 	if (!nvcuda.isValid())
 	{
+		nvDebug("*** CUDA driver not found.\n");
 		return false;
 	}
 	
 	if (version >= 2000)
 	{
 		void * address = nvcuda.bindSymbol("cuStreamCreate");
-		if (address == NULL) return false;
+		if (address == NULL) {
+			nvDebug("*** CUDA driver version < 2.0.\n");
+			return false;
+		}
 	}
 
 	if (version >= 2010)
 	{
 		void * address = nvcuda.bindSymbol("cuModuleLoadDataEx");
-		if (address == NULL) return false;
+		if (address == NULL) {
+			nvDebug("*** CUDA driver version < 2.1.\n");
+			return false;
+		}
 	}
 	
 	if (version >= 2020)
@@ -104,16 +111,23 @@ static bool isCudaDriverAvailable(int version)
 		typedef CUresult (CUDAAPI * PFCU_DRIVERGETVERSION)(int * version);
 
 		PFCU_DRIVERGETVERSION driverGetVersion = (PFCU_DRIVERGETVERSION)nvcuda.bindSymbol("cuDriverGetVersion");
-		if (driverGetVersion == NULL) return false;
+		if (driverGetVersion == NULL) {
+			nvDebug("*** CUDA driver version < 2.2.\n");
+			return false;
+		}
 
 		int driverVersion;
-		if (driverGetVersion(&driverVersion) != CUDA_SUCCESS) return false;
+		CUresult err = driverGetVersion(&driverVersion);
+		if (err != CUDA_SUCCESS) {
+			nvDebug("*** Error querying driver version: '%s'.\n", cudaGetErrorString((cudaError_t)err));
+			return false;
+		}
 
 		return driverVersion >= version;
 	}
 #endif // HAVE_CUDA
 
-	return false;
+	return true;
 }
 
 
@@ -121,10 +135,13 @@ static bool isCudaDriverAvailable(int version)
 bool nv::cuda::isHardwarePresent()
 {
 #if defined HAVE_CUDA
-#if NV_OS_WIN32
-	//if (isWindowsVista()) return false;
-	//if (isWindowsVista() || !isWow32()) return false;
-#endif
+	// Make sure that CUDA driver matches CUDA runtime.
+	if (!isCudaDriverAvailable(CUDART_VERSION))
+	{
+		nvDebug("CUDA driver not available for CUDA runtime %d\n", CUDART_VERSION);
+		return false;
+	}
+
 	int count = deviceCount();
 	if (count == 1)
 	{
@@ -137,15 +154,9 @@ bool nv::cuda::isHardwarePresent()
 		{
 			return false;
 		}
-
-		// Make sure that CUDA driver matches CUDA runtime.
-		if (!isCudaDriverAvailable(CUDART_VERSION))
-		{
-			return false;
 	}
 
 	// @@ Make sure that warp size == 32
-	}
 
 	return count > 0;
 #else
@@ -205,8 +216,24 @@ bool nv::cuda::setDevice(int i)
 	nvCheck(i < deviceCount());
 #if defined HAVE_CUDA
 	cudaError_t result = cudaSetDevice(i);
+
+	if (result != cudaSuccess) {
+		nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
+	}
+
 	return result == cudaSuccess;
 #else
 	return false;
 #endif
 }
+
+void nv::cuda::exit()
+{
+#if defined HAVE_CUDA
+	cudaError_t result = cudaThreadExit();
+
+	if (result != cudaSuccess) {
+		nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
+	}
+#endif
+}

src/nvtt/cuda/CudaUtils.h

@@ -33,6 +33,7 @@ namespace nv
 		int deviceCount();
 		int getFastestDevice();
 		bool setDevice(int i);
+		void exit();
 	};
 	
 } // nv namespace