diff --git a/Source/Math/CntkBatchNormalization.cuh b/Source/Math/CntkBatchNormalization.cuh index 5ca447537..39544dde5 100644 --- a/Source/Math/CntkBatchNormalization.cuh +++ b/Source/Math/CntkBatchNormalization.cuh @@ -108,12 +108,13 @@ __device__ __forceinline__ void StoreValues<4, float>(const float src[4], float* template __device__ __forceinline__ T Shuffle(T input, int srcLane) { -#if __CUDA_ARCH__ >= 300 +#ifdef __CUDA_ARCH__ // shfl is supported only on Kepler+ + static_assert(__CUDA_ARCH__ >= 300, "CNTK only supports only Kepler GPU architecture or newer"); return cub::ShuffleIndex(input, srcLane); #else assert(false); - return input; // keep compiler happy + return input; #endif } diff --git a/Source/Math/MathCUDA.vcxproj b/Source/Math/MathCUDA.vcxproj index 9edfef794..ae9823119 100644 --- a/Source/Math/MathCUDA.vcxproj +++ b/Source/Math/MathCUDA.vcxproj @@ -52,7 +52,7 @@ $(CNTK_CUDA_CODEGEN_DEBUG) - compute_20,sm_20;compute_30,sm_30 + compute_30,sm_30 $(CNTK_CUDA_CODEGEN_RELEASE)