177 #define OUT(x, y) dst[(x) + (y) * BPS]  macro
302   OUT(0, 3)                                     = AVG3(J, K, L);  in RD4()
303   OUT(0, 2) = OUT(1, 3)                         = AVG3(I, J, K);  in RD4()
304   OUT(0, 1) = OUT(1, 2) = OUT(2, 3)             = AVG3(X, I, J);  in RD4()
305   OUT(0, 0) = OUT(1, 1) = OUT(2, 2) = OUT(3, 3) = AVG3(A, X, I);  in RD4()
306   OUT(1, 0) = OUT(2, 1) = OUT(3, 2)             = AVG3(B, A, X);  in RD4()
307   OUT(2, 0) = OUT(3, 1)                         = AVG3(C, B, A);  in RD4()
308   OUT(3, 0)                                     = AVG3(D, C, B);  in RD4()
320   OUT(0, 0)                                     = AVG3(A, B, C);  in LD4()
321   OUT(1, 0) = OUT(0, 1)                         = AVG3(B, C, D);  in LD4()
322   OUT(2, 0) = OUT(1, 1) = OUT(0, 2)             = AVG3(C, D, E);  in LD4()
323   OUT(3, 0) = OUT(2, 1) = OUT(1, 2) = OUT(0, 3) = AVG3(D, E, F);  in LD4()
324   OUT(3, 1) = OUT(2, 2) = OUT(1, 3)             = AVG3(E, F, G);  in LD4()
325   OUT(3, 2) = OUT(2, 3)                         = AVG3(F, G, H);  in LD4()
326   OUT(3, 3)                                     = AVG3(G, H, H);  in LD4()
338   OUT(0, 0) = OUT(1, 2) = AVG2(X, A);  in VR4()
339   OUT(1, 0) = OUT(2, 2) = AVG2(A, B);  in VR4()
340   OUT(2, 0) = OUT(3, 2) = AVG2(B, C);  in VR4()
341   OUT(3, 0)             = AVG2(C, D);  in VR4()
343   OUT(0, 3) =             AVG3(K, J, I);  in VR4()
344   OUT(0, 2) =             AVG3(J, I, X);  in VR4()
345   OUT(0, 1) = OUT(1, 3) = AVG3(I, X, A);  in VR4()
346   OUT(1, 1) = OUT(2, 3) = AVG3(X, A, B);  in VR4()
347   OUT(2, 1) = OUT(3, 3) = AVG3(A, B, C);  in VR4()
348   OUT(3, 1) =             AVG3(B, C, D);  in VR4()
360   OUT(0, 0) =             AVG2(A, B);  in VL4()
361   OUT(1, 0) = OUT(0, 2) = AVG2(B, C);  in VL4()
362   OUT(2, 0) = OUT(1, 2) = AVG2(C, D);  in VL4()
363   OUT(3, 0) = OUT(2, 2) = AVG2(D, E);  in VL4()
365   OUT(0, 1) =             AVG3(A, B, C);  in VL4()
366   OUT(1, 1) = OUT(0, 3) = AVG3(B, C, D);  in VL4()
367   OUT(2, 1) = OUT(1, 3) = AVG3(C, D, E);  in VL4()
368   OUT(3, 1) = OUT(2, 3) = AVG3(D, E, F);  in VL4()
369               OUT(3, 2) = AVG3(E, F, G);  in VL4()
370               OUT(3, 3) = AVG3(F, G, H);  in VL4()
378   OUT(0, 0) =             AVG2(I, J);  in HU4()
379   OUT(2, 0) = OUT(0, 1) = AVG2(J, K);  in HU4()
380   OUT(2, 1) = OUT(0, 2) = AVG2(K, L);  in HU4()
381   OUT(1, 0) =             AVG3(I, J, K);  in HU4()
382   OUT(3, 0) = OUT(1, 1) = AVG3(J, K, L);  in HU4()
383   OUT(3, 1) = OUT(1, 2) = AVG3(K, L, L);  in HU4()
384   OUT(3, 2) = OUT(2, 2) =  in HU4()
385     OUT(0, 3) = OUT(1, 3) = OUT(2, 3) = OUT(3, 3) = L;  in HU4()
398   OUT(0, 0) = OUT(2, 1) = AVG2(I, X);  in HD4()
399   OUT(0, 1) = OUT(2, 2) = AVG2(J, I);  in HD4()
400   OUT(0, 2) = OUT(2, 3) = AVG2(K, J);  in HD4()
401   OUT(0, 3)             = AVG2(L, K);  in HD4()
403   OUT(3, 0)             = AVG3(A, B, C);  in HD4()
404   OUT(2, 0)             = AVG3(X, A, B);  in HD4()
405   OUT(1, 0) = OUT(3, 1) = AVG3(I, X, A);  in HD4()
406   OUT(1, 1) = OUT(3, 2) = AVG3(J, I, X);  in HD4()
407   OUT(1, 2) = OUT(3, 3) = AVG3(K, J, I);  in HD4()
408   OUT(1, 3)             = AVG3(L, K, J);  in HD4()