fixed usage of CPUONLY (GPUDummy.cpp still does not compile due to missing functions, and PTasks ignore CPUONLY as well currently)

Math/Math/GPUDummy.cpp

@@ -8,8 +8,8 @@
 
 #ifdef CPUONLY
 
-#include "GPUMatrix.cuh"
+#include "GPUMatrix.h"
-#include "GPUSparseMatrix.cuh"
+#include "GPUSparseMatrix.h"
 
 namespace Microsoft { namespace MSR { namespace CNTK {
     void PrepareDevice(short deviceId);
@@ -1646,7 +1646,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 }}}
 
 // define a dummy GPUWatcher class too
-#include "GPUWatcher.cuh"
+#include "GPUWatcher.h"
 
 int GPUWatcher::GetGPUIdWithTheMostFreeMemory()
 {

Math/Math/GPUMatrix.cu

@@ -3,7 +3,9 @@
 //     Copyright (c) Microsoft Corporation.  All rights reserved.
 // </copyright>
 //
-#pragma once
+
+#ifndef CPUONLY
+
 #include "stdafx.h"
 #include "cublas_v2.h"
 #include <assert.h>
@@ -3463,4 +3465,4 @@ int _ConvertSMVer2Cores(int major, int minor)
 //        return (long)free;
 //}
 
-
+#endif // CPUONLY

Math/Math/GPUMatrixCUDAKernels.cu

@@ -4,7 +4,7 @@
 // </copyright>
 //
 
-#ifndef CPU_ONLY
+#ifndef CPUONLY
 
 #include <float.h>
 #include <cuda_runtime.h>
@@ -3231,4 +3231,4 @@ d_tmp[0] = max((ElemType)0, d_tmp[0]/max((ElemType)1.0e-10,sqrt(d_tmp[1]))/max((
 }
 */
 
-#endif // !CPU_ONLY
+#endif // !CPUONLY

Math/Math/GPUSparseMatrix.cu

@@ -3,7 +3,9 @@
 //     Copyright (c) Microsoft Corporation.  All rights reserved.
 // </copyright>
 //
-#pragma once
+
+#ifndef CPUONLY
+
 #include "GPUSparseMatrix.h"
 #include <cuda_runtime.h>
 #include <cusparse_v2.h>
@@ -2190,3 +2192,5 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     template MATH_API File& operator<<(File& stream, const GPUSparseMatrix<double>& us);
 
 }}}
+
+#endif  // CPUONLY

Math/Math/GPUWatcher.cu

@@ -3,12 +3,14 @@
 //     Copyright (c) Microsoft Corporation.  All rights reserved.
 // </copyright>
 //
+
+#ifndef CPUONLY
+
 #include "stdafx.h"
 #include "GPUWatcher.h"
 #include <cuda.h>
 #include <cuda_runtime.h>
 
-
 int GPUWatcher::GetGPUIdWithTheMostFreeMemory()
 {
     int deviceCount = 0;
@@ -59,3 +61,4 @@ GPUWatcher::~GPUWatcher(void)
 {
 }
 
+#endif  // CPUONLY