77     int i,j,k,n,cols;  in arm_mat_solve_upper_triangular_f32()  local
96       for(j=0; j+3 < cols; j +=4)  in arm_mat_solve_upper_triangular_f32()
98             vecA = vld1q_f32(&pA[i * cols + j]);  in arm_mat_solve_upper_triangular_f32()
102                 vecX = vld1q_f32(&pX[cols*k+j]);            in arm_mat_solve_upper_triangular_f32()
115             vst1q(&pX[i*cols+j],vecA);  in arm_mat_solve_upper_triangular_f32()
118       for(; j < cols; j ++)  in arm_mat_solve_upper_triangular_f32()
120             a_col = &pA[j];  in arm_mat_solve_upper_triangular_f32()
128                 tmp -= ut_row[k] * pX[cols*k+j];  in arm_mat_solve_upper_triangular_f32()
136             pX[i*cols+j] = tmp;  in arm_mat_solve_upper_triangular_f32()
174     int i,j,k,n,cols;  in arm_mat_solve_upper_triangular_f32()  local
193       for(j=0; j+3 < cols; j +=4)  in arm_mat_solve_upper_triangular_f32()
195             vecA = vld1q_f32(&pA[i * cols + j]);  in arm_mat_solve_upper_triangular_f32()
199                 vecX = vld1q_f32(&pX[cols*k+j]);            in arm_mat_solve_upper_triangular_f32()
212             vst1q_f32(&pX[i*cols+j],vecA);  in arm_mat_solve_upper_triangular_f32()
215       for(; j < cols; j ++)  in arm_mat_solve_upper_triangular_f32()
217             a_col = &pA[j];  in arm_mat_solve_upper_triangular_f32()
225                 tmp -= ut_row[k] * pX[cols*k+j];  in arm_mat_solve_upper_triangular_f32()
233             pX[i*cols+j] = tmp;  in arm_mat_solve_upper_triangular_f32()
270     int i,j,k,n,cols;  in arm_mat_solve_upper_triangular_f32()  local
282     for(j=0; j < cols; j ++)  in arm_mat_solve_upper_triangular_f32()
284        a_col = &pA[j];  in arm_mat_solve_upper_triangular_f32()
294                 tmp -= ut_row[k] * pX[cols*k+j];  in arm_mat_solve_upper_triangular_f32()
302             pX[i*cols+j] = tmp;  in arm_mat_solve_upper_triangular_f32()