19 template <typename T, int P, int Q, int BY>
20 …line__ void read_A_notrans_g2r_1D_nosync(const int tx, const int ty, const T *dA, T *sA, T rA[Q]) {  in read_A_notrans_g2r_1D_nosync()  argument
25   for (i = 0; i < P * Q - P * BY; i += P * BY) {  in read_A_notrans_g2r_1D_nosync()
28   if (i + tid < P * Q) {  in read_A_notrans_g2r_1D_nosync()
34   for (int j = 0; j < Q; j++) {  in read_A_notrans_g2r_1D_nosync()
44 template <typename T, int P, int Q, int BY>
45 …inline__ void read_A_trans_g2r_1D_nosync(const int tx, const int ty, const T *dA, T *sA, T rA[Q]) {  in read_A_trans_g2r_1D_nosync()  argument
50   for (i = 0; i < P * Q - P * BY; i += P * BY) {  in read_A_trans_g2r_1D_nosync()
53   if (i + tid < P * Q) {  in read_A_trans_g2r_1D_nosync()
59   for (int j = 0; j < Q; j++) {  in read_A_trans_g2r_1D_nosync()
60     rA[j] = sA[tx * Q + j];  in read_A_trans_g2r_1D_nosync()
69 template <typename T, int P, int Q, int NB>
74     for (i = 0; i < Q * n - P; i += P) {  in read_B_g2s_1D_nosync()
79     for (i = 0; i < Q * NB - P; i += P) {  in read_B_g2s_1D_nosync()
83   if (i + tx < Q * n) {  in read_B_g2s_1D_nosync()
93 template <typename T, int P, int Q, int NB>
112 template <typename T, int P, int Q, int NB>
132 template <typename T, int P, int Q, int NB>
133 static __device__ __inline__ void mul_rAsBrC_1D_nosync(T rA[Q], T *sB, T rC[NB]) {  in mul_rAsBrC_1D_nosync()  argument
134   T rB[Q];  in mul_rAsBrC_1D_nosync()
139     for (int j = 0; j < Q; j++) {  in mul_rAsBrC_1D_nosync()
140       rB[j] = sB[i * Q + j];  in mul_rAsBrC_1D_nosync()
144     for (int j = 0; j < Q; j++) {  in mul_rAsBrC_1D_nosync()
156 template <typename T, int P, int Q, int NB>
157 static __device__ __inline__ void addmul_rAsBrC_1D_nosync(T rA[Q], T *sB, T rC[NB]) {  in addmul_rAsBrC_1D_nosync()  argument
158   T rB[Q];  in addmul_rAsBrC_1D_nosync()
163     for (int j = 0; j < Q; j++) {  in addmul_rAsBrC_1D_nosync()
164       rB[j] = sB[i * Q + j];  in addmul_rAsBrC_1D_nosync()
167     for (int j = 0; j < Q; j++) {  in addmul_rAsBrC_1D_nosync()