95 static __device__ __inline__ void read_V_2d(const T *dV, const int compstride, T rV[DIM_V][NUM_COMP…  in read_V_2d()
99         rV[i_DIM][comp][j] = dV[comp * compstride + j * Q + tx];  in read_V_2d()
112 static __device__ __inline__ void write_V_2d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_…  in write_V_2d()
116         dV[comp * compstride + j * Q + tx] = rV[i_DIM][comp][j];  in write_V_2d()
129 static __device__ __inline__ void sum_V_2d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_SI…  in sum_V_2d()
133         dV[comp * compstride + j * Q + tx] += rV[i_DIM][comp][j];  in sum_V_2d()
183 static __device__ __inline__ void read_V_3d(const T *dV, const int compstride, T rV[DIM_V][NUM_COMP…  in read_V_3d()
187         rV[i_DIM][comp][j] = dV[comp * compstride + j * (Q * Q) + tx];  in read_V_3d()
200 static __device__ __inline__ void write_V_3d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_…  in write_V_3d()
204         dV[comp * compstride + j * (Q * Q) + tx] = rV[i_DIM][comp][j];  in write_V_3d()
217 static __device__ __inline__ void sum_V_3d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_SI…  in sum_V_3d()
221         dV[comp * compstride + j * (Q * Q) + tx] += rV[i_DIM][comp][j];  in sum_V_3d()