diff --git a/lyra2/cuda_lyra2v2.cu b/lyra2/cuda_lyra2v2.cu index c3f89e6..df3291c 100644 --- a/lyra2/cuda_lyra2v2.cu +++ b/lyra2/cuda_lyra2v2.cu @@ -14,6 +14,8 @@ #define __CUDA_ARCH__ 500 #endif +#define TPB 32 + #if __CUDA_ARCH__ >= 500 #include "cuda_lyra2_vectors.h" @@ -22,8 +24,6 @@ #define Ncol 4 #define memshift 3 -#define TPB 32 - __device__ uint2x4 *DMatrix; __device__ __forceinline__ uint2 LD4S(const int index)