18 template <typename T, int LENGTH, int NUM_COMP>
19 …_inline__ void read_1d(const T *devptr, const int compstride, T *sBuffer[NUM_COMP], const int tx) {  in read_1d()  argument
21     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_1d()
30 template <typename T, int LENGTH, int NUM_COMP>
31 static __device__ __inline__ void write_1d(T *sBuffer[NUM_COMP], T *devptr, const int compstride, c…  in write_1d()  argument
33     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_1d()
42 template <typename T, int LENGTH, int NUM_COMP>
43 static __device__ __inline__ void sum_1d(T *sBuffer[NUM_COMP], T *devptr, const int compstride, con…  in sum_1d()  argument
45     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_1d()
58 template <typename T, int P, int DIM_U, int NUM_COMP, int rU_SIZE, int i_DIM>
59 static __device__ __inline__ void read_U_2d(const T *dU, const int compstride, T rU[DIM_U][NUM_COMP…  in read_U_2d()
70   for (int comp = 0; comp < NUM_COMP; comp++) {  in read_U_2d()
94 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
95 static __device__ __inline__ void read_V_2d(const T *dV, const int compstride, T rV[DIM_V][NUM_COMP…  in read_V_2d()
97     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_V_2d()
111 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
112 static __device__ __inline__ void write_V_2d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_…  in write_V_2d()
114     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_V_2d()
128 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
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()
131     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_V_2d()
146 template <typename T, int P, int DIM_U, int NUM_COMP, int rU_SIZE, int i_DIM>
147 static __device__ __inline__ void read_U_3d(const T *dU, const int compstride, T rU[DIM_U][NUM_COMP…  in read_U_3d()
158   for (int comp = 0; comp < NUM_COMP; comp++) {  in read_U_3d()
182 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
183 static __device__ __inline__ void read_V_3d(const T *dV, const int compstride, T rV[DIM_V][NUM_COMP…  in read_V_3d()
185     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_V_3d()
199 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
200 static __device__ __inline__ void write_V_3d(T *dV, const int compstride, T rV[DIM_V][NUM_COMP][rV_…  in write_V_3d()
202     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_V_3d()
216 template <typename T, int Q, int DIM_V, int NUM_COMP, int rV_SIZE, int i_DIM>
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()
219     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_V_3d()