moved DeviceFromConfig() for CPUONLY mode back to BestGpu.h since it otherwise caused link errors for reader projects that did not include BestGpu.cpp;

.CU files now include BestGpu.h to get access to CPUONLY flag;
CPUONLY mode set from BestGpu.h now compiles in VS with SDK installed

Common/Include/BestGpu.h

@@ -6,8 +6,13 @@
 
 #pragma once
 
-// #define CPUONLY      // #define this to build without GPU support nor needing the SDK installed
+#define CPUONLY      // #define this to build without GPU support nor needing the SDK installed
 
 namespace Microsoft { namespace MSR { namespace CNTK {
+#ifndef CPUONLY
     short DeviceFromConfig(const class ConfigParameters& config);
+#else
+    static inline short DeviceFromConfig(const class ConfigParameters& /*config*/) { return -1/*CPUDEVICE*/; }    // tells runtime system to not try to use GPUs
+    // TODO: find a way to use CPUDEVICE without a huge include overhead; OK so far since CPUONLY mode is sorta special...
+#endif
 }}}

Math/Math/GPUMatrix.cu

@@ -4,9 +4,11 @@
 // </copyright>
 //
 
+#include "stdafx.h"
+#include "BestGpu.h"
+
 #ifndef CPUONLY
 
-#include "stdafx.h"
 #include "cublas_v2.h"
 #include <assert.h>
 #include <cuda.h>

Math/Math/GPUMatrixCUDAKernels.cu

@@ -4,6 +4,8 @@
 // </copyright>
 //
 
+#include "BestGpu.h"
+
 #ifndef CPUONLY
 
 #include <float.h>

Math/Math/GPUSparseMatrix.cu

@@ -4,6 +4,8 @@
 // </copyright>
 //
 
+#include "BestGpu.h"
+
 #ifndef CPUONLY
 
 #include "GPUSparseMatrix.h"

Math/Math/GPUWatcher.cu

@@ -4,9 +4,11 @@
 // </copyright>
 //
 
+#include "stdafx.h"
+#include "BestGpu.h"
+
 #ifndef CPUONLY
 
-#include "stdafx.h"
 #include "GPUWatcher.h"
 #include <cuda.h>
 #include <cuda_runtime.h>

Math/Math/NoGPU.cpp

@@ -16,8 +16,6 @@
 #pragma warning (disable: 4702) // unreachable code, which we get from the NOT_IMPLEMENTED macro which is OK
 
 namespace Microsoft { namespace MSR { namespace CNTK {
-    short DeviceFromConfig(const ConfigParameters& /*config*/) { return CPUDEVICE; }    // tells runtime system to not try to use GPUs
-
     // the reset below are dummy implementations
 
     void PrepareDevice(short deviceId);
@@ -39,8 +37,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     // PrepareDevice - Setup the correct cuda context for an operation
     // deviceId - the device on which the operation will take place
     //            defaults to -1, which means use matrices current device
-    template<class ElemType> void GPUSparseMatrix<ElemType>::PrepareDevice(short deviceId /*=-1*/) const
+    template<class ElemType> void GPUSparseMatrix<ElemType>::PrepareDevice(short deviceId /*=-1*/) const {}
-    {}
 
     template<class ElemType> void GPUSparseMatrix<ElemType>::DeepCopy(const GPUSparseMatrix<ElemType>& deepCopy) { }