781 #define DST(x, y) dst[(x) + (y) * stride]  macro
792     DST(0,0) = (a0 + a1 * 2 + a2 + 2) >> 2;  in diag_downleft_4x4_c()
793     DST(1,0) = DST(0,1) = (a1 + a2 * 2 + a3 + 2) >> 2;  in diag_downleft_4x4_c()
794     DST(2,0) = DST(1,1) = DST(0,2) = (a2 + a3 * 2 + a4 + 2) >> 2;  in diag_downleft_4x4_c()
795     DST(3,0) = DST(2,1) = DST(1,2) = DST(0,3) = (a3 + a4 * 2 + a5 + 2) >> 2;  in diag_downleft_4x4_c()
796     DST(3,1) = DST(2,2) = DST(1,3) = (a4 + a5 * 2 + a6 + 2) >> 2;  in diag_downleft_4x4_c()
797     DST(3,2) = DST(2,3) = (a5 + a6 * 2 + a7 + 2) >> 2;  in diag_downleft_4x4_c()
798     DST(3,3) = a7;  // note: this is different from vp8 and such  in diag_downleft_4x4_c()
835     DST(0,3) = (l1 + l2 * 2 + l3 + 2) >> 2;  in diag_downright_4x4_c()
836     DST(0,2) = DST(1,3) = (l0 + l1 * 2 + l2 + 2) >> 2;  in diag_downright_4x4_c()
837     DST(0,1) = DST(1,2) = DST(2,3) = (tl + l0 * 2 + l1 + 2) >> 2;  in diag_downright_4x4_c()
838     DST(0,0) = DST(1,1) = DST(2,2) = DST(3,3) = (l0 + tl * 2 + a0 + 2) >> 2;  in diag_downright_4x4_c()
839     DST(1,0) = DST(2,1) = DST(3,2) = (tl + a0 * 2 + a1 + 2) >> 2;  in diag_downright_4x4_c()
840     DST(2,0) = DST(3,1) = (a0 + a1 * 2 + a2 + 2) >> 2;  in diag_downright_4x4_c()
841     DST(3,0) = (a1 + a2 * 2 + a3 + 2) >> 2;  in diag_downright_4x4_c()
881     DST(0,3) = (l0 + l1 * 2 + l2 + 2) >> 2;  in vert_right_4x4_c()
882     DST(0,2) = (tl + l0 * 2 + l1 + 2) >> 2;  in vert_right_4x4_c()
883     DST(0,0) = DST(1,2) = (tl + a0 + 1) >> 1;  in vert_right_4x4_c()
884     DST(0,1) = DST(1,3) = (l0 + tl * 2 + a0 + 2) >> 2;  in vert_right_4x4_c()
885     DST(1,0) = DST(2,2) = (a0 + a1 + 1) >> 1;  in vert_right_4x4_c()
886     DST(1,1) = DST(2,3) = (tl + a0 * 2 + a1 + 2) >> 2;  in vert_right_4x4_c()
887     DST(2,0) = DST(3,2) = (a1 + a2 + 1) >> 1;  in vert_right_4x4_c()
888     DST(2,1) = DST(3,3) = (a0 + a1 * 2 + a2 + 2) >> 2;  in vert_right_4x4_c()
889     DST(3,0) = (a2 + a3 + 1) >> 1;  in vert_right_4x4_c()
890     DST(3,1) = (a1 + a2 * 2 + a3 + 2) >> 2;  in vert_right_4x4_c()
938     DST(2,0) = (tl + a0 * 2 + a1 + 2) >> 2;  in hor_down_4x4_c()
939     DST(3,0) = (a0 + a1 * 2 + a2 + 2) >> 2;  in hor_down_4x4_c()
940     DST(0,0) = DST(2,1) = (tl + l0 + 1) >> 1;  in hor_down_4x4_c()
941     DST(1,0) = DST(3,1) = (a0 + tl * 2 + l0 + 2) >> 2;  in hor_down_4x4_c()
942     DST(0,1) = DST(2,2) = (l0 + l1 + 1) >> 1;  in hor_down_4x4_c()
943     DST(1,1) = DST(3,2) = (tl + l0 * 2 + l1 + 2) >> 2;  in hor_down_4x4_c()
944     DST(0,2) = DST(2,3) = (l1 + l2 + 1) >> 1;  in hor_down_4x4_c()
945     DST(1,2) = DST(3,3) = (l0 + l1 * 2 + l2 + 2) >> 2;  in hor_down_4x4_c()
946     DST(0,3) = (l2 + l3 + 1) >> 1;  in hor_down_4x4_c()
947     DST(1,3) = (l1 + l2 * 2 + l3 + 2) >> 2;  in hor_down_4x4_c()
988     DST(0,0) = (a0 + a1 + 1) >> 1;  in vert_left_4x4_c()
989     DST(0,1) = (a0 + a1 * 2 + a2 + 2) >> 2;  in vert_left_4x4_c()
990     DST(1,0) = DST(0,2) = (a1 + a2 + 1) >> 1;  in vert_left_4x4_c()
991     DST(1,1) = DST(0,3) = (a1 + a2 * 2 + a3 + 2) >> 2;  in vert_left_4x4_c()
992     DST(2,0) = DST(1,2) = (a2 + a3 + 1) >> 1;  in vert_left_4x4_c()
993     DST(2,1) = DST(1,3) = (a2 + a3 * 2 + a4 + 2) >> 2;  in vert_left_4x4_c()
994     DST(3,0) = DST(2,2) = (a3 + a4 + 1) >> 1;  in vert_left_4x4_c()
995     DST(3,1) = DST(2,3) = (a3 + a4 * 2 + a5 + 2) >> 2;  in vert_left_4x4_c()
996     DST(3,2) = (a4 + a5 + 1) >> 1;  in vert_left_4x4_c()
997     DST(3,3) = (a4 + a5 * 2 + a6 + 2) >> 2;  in vert_left_4x4_c()
1037     DST(0,0) = (l0 + l1 + 1) >> 1;  in hor_up_4x4_c()
1038     DST(1,0) = (l0 + l1 * 2 + l2 + 2) >> 2;  in hor_up_4x4_c()
1039     DST(0,1) = DST(2,0) = (l1 + l2 + 1) >> 1;  in hor_up_4x4_c()
1040     DST(1,1) = DST(3,0) = (l1 + l2 * 2 + l3 + 2) >> 2;  in hor_up_4x4_c()
1041     DST(0,2) = DST(2,1) = (l2 + l3 + 1) >> 1;  in hor_up_4x4_c()
1042     DST(1,2) = DST(3,1) = (l2 + l3 * 3 + 2) >> 2;  in hor_up_4x4_c()
1043     DST(0,3) = DST(1,3) = DST(2,2) = DST(2,3) = DST(3,2) = DST(3,3) = l3;  in hor_up_4x4_c()
1076 #undef DST