21     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_1d()  local
22       sBuffer[comp][tx] = devptr[comp * compstride + tx];  in read_1d()
33     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_1d()  local
34       devptr[comp * compstride + tx] = sBuffer[comp][tx];  in write_1d()
45     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_1d()  local
46       devptr[comp * compstride + tx] += sBuffer[comp][tx];  in sum_1d()
70   for (int comp = 0; comp < NUM_COMP; comp++) {  in read_U_2d()  local
74         sTmp[i * P + tx] = dU[comp * compstride + i * P + tx];  in read_U_2d()
81         rU[i_DIM][comp][i] = sTmp[tx * P + i];  in read_U_2d()
97     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_V_2d()  local
99         rV[i_DIM][comp][j] = dV[comp * compstride + j * Q + tx];  in read_V_2d()
114     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_V_2d()  local
116         dV[comp * compstride + j * Q + tx] = rV[i_DIM][comp][j];  in write_V_2d()
131     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_V_2d()  local
133         dV[comp * compstride + j * Q + tx] += rV[i_DIM][comp][j];  in sum_V_2d()
158   for (int comp = 0; comp < NUM_COMP; comp++) {  in read_U_3d()  local
162         sTmp[i * P * P + tx] = dU[comp * compstride + i * P * P + tx];  in read_U_3d()
169         rU[i_DIM][comp][i] = sTmp[tx * P + i];  in read_U_3d()
185     for (int comp = 0; comp < NUM_COMP; comp++) {  in read_V_3d()  local
187         rV[i_DIM][comp][j] = dV[comp * compstride + j * (Q * Q) + tx];  in read_V_3d()
202     for (int comp = 0; comp < NUM_COMP; comp++) {  in write_V_3d()  local
204         dV[comp * compstride + j * (Q * Q) + tx] = rV[i_DIM][comp][j];  in write_V_3d()
219     for (int comp = 0; comp < NUM_COMP; comp++) {  in sum_V_3d()  local
221         dV[comp * compstride + j * (Q * Q) + tx] += rV[i_DIM][comp][j];  in sum_V_3d()