20 static __device__ __inline__ void read_A_notrans_g2r_1D_nosync(const int tx, const int ty, const T …  in read_A_notrans_g2r_1D_nosync()  argument
21   const int tid = ty * P + tx;  in read_A_notrans_g2r_1D_nosync()
35     rA[j] = sA[j * P + tx];  in read_A_notrans_g2r_1D_nosync()
45 static __device__ __inline__ void read_A_trans_g2r_1D_nosync(const int tx, const int ty, const T *d…  in read_A_trans_g2r_1D_nosync()  argument
46   const int tid = ty * P + tx;  in read_A_trans_g2r_1D_nosync()
60     rA[j] = sA[tx * Q + j];  in read_A_trans_g2r_1D_nosync()
70 static __device__ __inline__ void read_B_g2s_1D_nosync(const int tx, const int n, const T *dB, T *s…  in read_B_g2s_1D_nosync()  argument
75       sB[i + tx] = dB[i + tx];  in read_B_g2s_1D_nosync()
80       sB[i + tx] = dB[i + tx];  in read_B_g2s_1D_nosync()
83   if (i + tx < Q * n) {  in read_B_g2s_1D_nosync()
84     sB[i + tx] = dB[i + tx];  in read_B_g2s_1D_nosync()
94 static __device__ __inline__ void write_C_r2g_1D_nosync(const int tx, const int n, T rC[NB], T *dC)…  in write_C_r2g_1D_nosync()  argument
97       dC[i * P + tx] = rC[i];  in write_C_r2g_1D_nosync()
102       dC[i * P + tx] = rC[i];  in write_C_r2g_1D_nosync()
113 static __device__ __inline__ void sum_C_r2g_1D_nosync(const int tx, const int n, T rC[NB], T *dC) {  in sum_C_r2g_1D_nosync()  argument
116       dC[i * P + tx] += rC[i];  in sum_C_r2g_1D_nosync()
121       dC[i * P + tx] += rC[i];  in sum_C_r2g_1D_nosync()