+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_12x16, 4, 7, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ pmaddubsw m8, [r5]
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ pmaddubsw m9, [r5]
+ movu xm11, [r0 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ pmaddubsw m12, m10, [r5 + 1 * mmsize]
+ paddw m8, m12
+ pmaddubsw m10, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm12, [r0] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+ pmaddubsw m13, m11, [r5 + 1 * mmsize]
+ paddw m9, m13
+ pmaddubsw m11, [r5]
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movq [r2], xm0
+ pextrd [r2 + 8], xm0, 2
+ movq [r2 + r3], xm1
+ pextrd [r2 + r3 + 8], xm1, 2
+ movq [r2 + r3 * 2], xm2
+ pextrd [r2 + r3 * 2 + 8], xm2, 2
+ movq [r2 + r6], xm3
+ pextrd [r2 + r6 + 8], xm3, 2
+ lea r2, [r2 + r3 * 4]
+ movq [r2], xm4
+ pextrd [r2 + 8], xm4, 2
+ movq [r2 + r3], xm5
+ pextrd [r2 + r3 + 8], xm5, 2
+
+ movu xm13, [r0 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ pmaddubsw m0, m12, [r5 + 2 * mmsize]
+ paddw m8, m0
+ pmaddubsw m0, m12, [r5 + 1 * mmsize]
+ paddw m10, m0
+ pmaddubsw m12, [r5]
+ movu xm0, [r0 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+ pmaddubsw m1, m13, [r5 + 2 * mmsize]
+ paddw m9, m1
+ pmaddubsw m1, m13, [r5 + 1 * mmsize]
+ paddw m11, m1
+ pmaddubsw m13, [r5]
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movq [r2 + r3 * 2], xm6
+ pextrd [r2 + r3 * 2 + 8], xm6, 2
+ movq [r2 + r6], xm7
+ pextrd [r2 + r6 + 8], xm7, 2
+ lea r2, [r2 + r3 * 4]
+
+ movu xm1, [r0 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, [r5 + 3 * mmsize]
+ paddw m8, m2
+ pmaddubsw m2, m0, [r5 + 2 * mmsize]
+ paddw m10, m2
+ pmaddubsw m2, m0, [r5 + 1 * mmsize]
+ paddw m12, m2
+ pmaddubsw m0, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm2, [r0] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 3 * mmsize]
+ paddw m9, m3
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m11, m3
+ pmaddubsw m3, m1, [r5 + 1 * mmsize]
+ paddw m13, m3
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 3 * mmsize]
+ paddw m10, m4
+ pmaddubsw m4, m2, [r5 + 2 * mmsize]
+ paddw m12, m4
+ pmaddubsw m2, [r5 + 1 * mmsize]
+ paddw m0, m2
+ movu xm4, [r0 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 3 * mmsize]
+ paddw m11, m5
+ pmaddubsw m5, m3, [r5 + 2 * mmsize]
+ paddw m13, m5
+ pmaddubsw m3, [r5 + 1 * mmsize]
+ paddw m1, m3
+ movu xm5, [r0 + r4] ; m5 = row 19
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 3 * mmsize]
+ paddw m12, m6
+ pmaddubsw m4, [r5 + 2 * mmsize]
+ paddw m0, m4
+ lea r0, [r0 + r1 * 4]
+ movu xm6, [r0] ; m6 = row 20
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 3 * mmsize]
+ paddw m13, m7
+ pmaddubsw m5, [r5 + 2 * mmsize]
+ paddw m1, m5
+ movu xm7, [r0 + r1] ; m7 = row 21
+ punpckhbw xm2, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm2, 1
+ pmaddubsw m6, [r5 + 3 * mmsize]
+ paddw m0, m6
+ movu xm2, [r0 + r1 * 2] ; m2 = row 22
+ punpckhbw xm3, xm7, xm2
+ punpcklbw xm7, xm2
+ vinserti128 m7, m7, xm3, 1
+ pmaddubsw m7, [r5 + 3 * mmsize]
+ paddw m1, m7
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ pmulhrsw m12, m14 ; m12 = word: row 12
+ pmulhrsw m13, m14 ; m13 = word: row 13
+ pmulhrsw m0, m14 ; m0 = word: row 14
+ pmulhrsw m1, m14 ; m1 = word: row 15
+ packuswb m8, m9
+ packuswb m10, m11
+ packuswb m12, m13
+ packuswb m0, m1
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vpermq m12, m12, 11011000b
+ vpermq m0, m0, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ vextracti128 xm13, m12, 1
+ vextracti128 xm1, m0, 1
+ movq [r2], xm8
+ pextrd [r2 + 8], xm8, 2
+ movq [r2 + r3], xm9
+ pextrd [r2 + r3 + 8], xm9, 2
+ movq [r2 + r3 * 2], xm10
+ pextrd [r2 + r3 * 2 + 8], xm10, 2
+ movq [r2 + r6], xm11
+ pextrd [r2 + r6 + 8], xm11, 2
+ lea r2, [r2 + r3 * 4]
+ movq [r2], xm12
+ pextrd [r2 + 8], xm12, 2
+ movq [r2 + r3], xm13
+ pextrd [r2 + r3 + 8], xm13, 2
+ movq [r2 + r3 * 2], xm0
+ pextrd [r2 + r3 * 2 + 8], xm0, 2
+ movq [r2 + r6], xm1
+ pextrd [r2 + r6 + 8], xm1, 2
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_16x16, 4, 7, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ pmaddubsw m8, [r5]
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ pmaddubsw m9, [r5]
+ movu xm11, [r0 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ pmaddubsw m12, m10, [r5 + 1 * mmsize]
+ paddw m8, m12
+ pmaddubsw m10, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm12, [r0] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+ pmaddubsw m13, m11, [r5 + 1 * mmsize]
+ paddw m9, m13
+ pmaddubsw m11, [r5]
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm4
+ movu [r2 + r3], xm5
+
+ movu xm13, [r0 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ pmaddubsw m0, m12, [r5 + 2 * mmsize]
+ paddw m8, m0
+ pmaddubsw m0, m12, [r5 + 1 * mmsize]
+ paddw m10, m0
+ pmaddubsw m12, [r5]
+ movu xm0, [r0 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+ pmaddubsw m1, m13, [r5 + 2 * mmsize]
+ paddw m9, m1
+ pmaddubsw m1, m13, [r5 + 1 * mmsize]
+ paddw m11, m1
+ pmaddubsw m13, [r5]
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r2 + r3 * 2], xm6
+ movu [r2 + r6], xm7
+ lea r2, [r2 + r3 * 4]
+
+ movu xm1, [r0 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, [r5 + 3 * mmsize]
+ paddw m8, m2
+ pmaddubsw m2, m0, [r5 + 2 * mmsize]
+ paddw m10, m2
+ pmaddubsw m2, m0, [r5 + 1 * mmsize]
+ paddw m12, m2
+ pmaddubsw m0, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm2, [r0] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 3 * mmsize]
+ paddw m9, m3
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m11, m3
+ pmaddubsw m3, m1, [r5 + 1 * mmsize]
+ paddw m13, m3
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 3 * mmsize]
+ paddw m10, m4
+ pmaddubsw m4, m2, [r5 + 2 * mmsize]
+ paddw m12, m4
+ pmaddubsw m2, [r5 + 1 * mmsize]
+ paddw m0, m2
+ movu xm4, [r0 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 3 * mmsize]
+ paddw m11, m5
+ pmaddubsw m5, m3, [r5 + 2 * mmsize]
+ paddw m13, m5
+ pmaddubsw m3, [r5 + 1 * mmsize]
+ paddw m1, m3
+ movu xm5, [r0 + r4] ; m5 = row 19
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 3 * mmsize]
+ paddw m12, m6
+ pmaddubsw m4, [r5 + 2 * mmsize]
+ paddw m0, m4
+ lea r0, [r0 + r1 * 4]
+ movu xm6, [r0] ; m6 = row 20
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 3 * mmsize]
+ paddw m13, m7
+ pmaddubsw m5, [r5 + 2 * mmsize]
+ paddw m1, m5
+ movu xm7, [r0 + r1] ; m7 = row 21
+ punpckhbw xm2, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm2, 1
+ pmaddubsw m6, [r5 + 3 * mmsize]
+ paddw m0, m6
+ movu xm2, [r0 + r1 * 2] ; m2 = row 22
+ punpckhbw xm3, xm7, xm2
+ punpcklbw xm7, xm2
+ vinserti128 m7, m7, xm3, 1
+ pmaddubsw m7, [r5 + 3 * mmsize]
+ paddw m1, m7
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ pmulhrsw m12, m14 ; m12 = word: row 12
+ pmulhrsw m13, m14 ; m13 = word: row 13
+ pmulhrsw m0, m14 ; m0 = word: row 14
+ pmulhrsw m1, m14 ; m1 = word: row 15
+ packuswb m8, m9
+ packuswb m10, m11
+ packuswb m12, m13
+ packuswb m0, m1
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vpermq m12, m12, 11011000b
+ vpermq m0, m0, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ vextracti128 xm13, m12, 1
+ vextracti128 xm1, m0, 1
+ movu [r2], xm8
+ movu [r2 + r3], xm9
+ movu [r2 + r3 * 2], xm10
+ movu [r2 + r6], xm11
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm12
+ movu [r2 + r3], xm13
+ movu [r2 + r3 * 2], xm0
+ movu [r2 + r6], xm1
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_16x12, 4, 7, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ pmaddubsw m8, [r5]
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ pmaddubsw m9, [r5]
+ movu xm11, [r0 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ pmaddubsw m12, m10, [r5 + 1 * mmsize]
+ paddw m8, m12
+ pmaddubsw m10, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm12, [r0] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+ pmaddubsw m13, m11, [r5 + 1 * mmsize]
+ paddw m9, m13
+ pmaddubsw m11, [r5]
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm4
+ movu [r2 + r3], xm5
+
+ movu xm13, [r0 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ pmaddubsw m0, m12, [r5 + 2 * mmsize]
+ paddw m8, m0
+ pmaddubsw m0, m12, [r5 + 1 * mmsize]
+ paddw m10, m0
+ movu xm0, [r0 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+ pmaddubsw m1, m13, [r5 + 2 * mmsize]
+ paddw m9, m1
+ pmaddubsw m1, m13, [r5 + 1 * mmsize]
+ paddw m11, m1
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r2 + r3 * 2], xm6
+ movu [r2 + r6], xm7
+ lea r2, [r2 + r3 * 4]
+
+ movu xm1, [r0 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, [r5 + 3 * mmsize]
+ paddw m8, m2
+ pmaddubsw m2, m0, [r5 + 2 * mmsize]
+ paddw m10, m2
+ lea r0, [r0 + r1 * 4]
+ movu xm2, [r0] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 3 * mmsize]
+ paddw m9, m3
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m11, m3
+ movu xm3, [r0 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 3 * mmsize]
+ paddw m10, m4
+ movu xm4, [r0 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 3 * mmsize]
+ paddw m11, m5
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ packuswb m8, m9
+ packuswb m10, m11
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ movu [r2], xm8
+ movu [r2 + r3], xm9
+ movu [r2 + r3 * 2], xm10
+ movu [r2 + r6], xm11
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_16x8, 4, 7, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ movu xm11, [r0 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ lea r0, [r0 + r1 * 4]
+ movu xm12, [r0] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm4
+ movu [r2 + r3], xm5
+
+ movu xm13, [r0 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ movu xm0, [r0 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r2 + r3 * 2], xm6
+ movu [r2 + r6], xm7
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_16x4, 4, 7, 13
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m12, [pw_512]
+
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+
+ pmulhrsw m0, m12 ; m0 = word: row 0
+ pmulhrsw m1, m12 ; m1 = word: row 1
+ pmulhrsw m2, m12 ; m2 = word: row 2
+ pmulhrsw m3, m12 ; m3 = word: row 3
+ packuswb m0, m1
+ packuswb m2, m3
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ RET
+%endif
+
+%macro FILTER_VER_LUMA_AVX2_16xN 2
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_%1x%2, 4, 9, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ lea r7, [r1 * 4]
+ mova m14, [pw_512]
+ mov r8d, %2 / 16
+
+.loop:
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm4, [r0] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r0 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r0 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r0 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm8, [r0] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r0 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ pmaddubsw m8, [r5]
+ movu xm10, [r0 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ pmaddubsw m9, [r5]
+ movu xm11, [r0 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ pmaddubsw m12, m10, [r5 + 1 * mmsize]
+ paddw m8, m12
+ pmaddubsw m10, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm12, [r0] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+ pmaddubsw m13, m11, [r5 + 1 * mmsize]
+ paddw m9, m13
+ pmaddubsw m11, [r5]
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm4
+ movu [r2 + r3], xm5
+
+ movu xm13, [r0 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ pmaddubsw m0, m12, [r5 + 2 * mmsize]
+ paddw m8, m0
+ pmaddubsw m0, m12, [r5 + 1 * mmsize]
+ paddw m10, m0
+ pmaddubsw m12, [r5]
+ movu xm0, [r0 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+ pmaddubsw m1, m13, [r5 + 2 * mmsize]
+ paddw m9, m1
+ pmaddubsw m1, m13, [r5 + 1 * mmsize]
+ paddw m11, m1
+ pmaddubsw m13, [r5]
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r2 + r3 * 2], xm6
+ movu [r2 + r6], xm7
+ lea r2, [r2 + r3 * 4]
+
+ movu xm1, [r0 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, [r5 + 3 * mmsize]
+ paddw m8, m2
+ pmaddubsw m2, m0, [r5 + 2 * mmsize]
+ paddw m10, m2
+ pmaddubsw m2, m0, [r5 + 1 * mmsize]
+ paddw m12, m2
+ pmaddubsw m0, [r5]
+ lea r0, [r0 + r1 * 4]
+ movu xm2, [r0] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 3 * mmsize]
+ paddw m9, m3
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m11, m3
+ pmaddubsw m3, m1, [r5 + 1 * mmsize]
+ paddw m13, m3
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 3 * mmsize]
+ paddw m10, m4
+ pmaddubsw m4, m2, [r5 + 2 * mmsize]
+ paddw m12, m4
+ pmaddubsw m2, [r5 + 1 * mmsize]
+ paddw m0, m2
+ movu xm4, [r0 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 3 * mmsize]
+ paddw m11, m5
+ pmaddubsw m5, m3, [r5 + 2 * mmsize]
+ paddw m13, m5
+ pmaddubsw m3, [r5 + 1 * mmsize]
+ paddw m1, m3
+ movu xm5, [r0 + r4] ; m5 = row 19
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 3 * mmsize]
+ paddw m12, m6
+ pmaddubsw m4, [r5 + 2 * mmsize]
+ paddw m0, m4
+ lea r0, [r0 + r1 * 4]
+ movu xm6, [r0] ; m6 = row 20
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 3 * mmsize]
+ paddw m13, m7
+ pmaddubsw m5, [r5 + 2 * mmsize]
+ paddw m1, m5
+ movu xm7, [r0 + r1] ; m7 = row 21
+ punpckhbw xm2, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm2, 1
+ pmaddubsw m6, [r5 + 3 * mmsize]
+ paddw m0, m6
+ movu xm2, [r0 + r1 * 2] ; m2 = row 22
+ punpckhbw xm3, xm7, xm2
+ punpcklbw xm7, xm2
+ vinserti128 m7, m7, xm3, 1
+ pmaddubsw m7, [r5 + 3 * mmsize]
+ paddw m1, m7
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ pmulhrsw m12, m14 ; m12 = word: row 12
+ pmulhrsw m13, m14 ; m13 = word: row 13
+ pmulhrsw m0, m14 ; m0 = word: row 14
+ pmulhrsw m1, m14 ; m1 = word: row 15
+ packuswb m8, m9
+ packuswb m10, m11
+ packuswb m12, m13
+ packuswb m0, m1
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vpermq m12, m12, 11011000b
+ vpermq m0, m0, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ vextracti128 xm13, m12, 1
+ vextracti128 xm1, m0, 1
+ movu [r2], xm8
+ movu [r2 + r3], xm9
+ movu [r2 + r3 * 2], xm10
+ movu [r2 + r6], xm11
+ lea r2, [r2 + r3 * 4]
+ movu [r2], xm12
+ movu [r2 + r3], xm13
+ movu [r2 + r3 * 2], xm0
+ movu [r2 + r6], xm1
+ lea r2, [r2 + r3 * 4]
+ sub r0, r7
+ dec r8d
+ jnz .loop
+ RET
+%endif
+%endmacro
+
+FILTER_VER_LUMA_AVX2_16xN 16, 32
+FILTER_VER_LUMA_AVX2_16xN 16, 64
+
+%macro PROCESS_LUMA_AVX2_W16_16R 0
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r7, [r0 + r1 * 4]
+ movu xm4, [r7] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r7 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r7 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r7 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r7, [r7 + r1 * 4]
+ movu xm8, [r7] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r7 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ pmaddubsw m8, [r5]
+ movu xm10, [r7 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ pmaddubsw m9, [r5]
+ movu xm11, [r7 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ pmaddubsw m12, m10, [r5 + 1 * mmsize]
+ paddw m8, m12
+ pmaddubsw m10, [r5]
+ lea r7, [r7 + r1 * 4]
+ movu xm12, [r7] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+ pmaddubsw m13, m11, [r5 + 1 * mmsize]
+ paddw m9, m13
+ pmaddubsw m11, [r5]
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r8, [r2 + r3 * 4]
+ movu [r8], xm4
+ movu [r8 + r3], xm5
+
+ movu xm13, [r7 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ pmaddubsw m0, m12, [r5 + 2 * mmsize]
+ paddw m8, m0
+ pmaddubsw m0, m12, [r5 + 1 * mmsize]
+ paddw m10, m0
+ pmaddubsw m12, [r5]
+ movu xm0, [r7 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+ pmaddubsw m1, m13, [r5 + 2 * mmsize]
+ paddw m9, m1
+ pmaddubsw m1, m13, [r5 + 1 * mmsize]
+ paddw m11, m1
+ pmaddubsw m13, [r5]
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r8 + r3 * 2], xm6
+ movu [r8 + r6], xm7
+ lea r8, [r8 + r3 * 4]
+
+ movu xm1, [r7 + r4] ; m1 = row 15
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m2, m0, [r5 + 3 * mmsize]
+ paddw m8, m2
+ pmaddubsw m2, m0, [r5 + 2 * mmsize]
+ paddw m10, m2
+ pmaddubsw m2, m0, [r5 + 1 * mmsize]
+ paddw m12, m2
+ pmaddubsw m0, [r5]
+ lea r7, [r7 + r1 * 4]
+ movu xm2, [r7] ; m2 = row 16
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 3 * mmsize]
+ paddw m9, m3
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m11, m3
+ pmaddubsw m3, m1, [r5 + 1 * mmsize]
+ paddw m13, m3
+ pmaddubsw m1, [r5]
+ movu xm3, [r7 + r1] ; m3 = row 17
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 3 * mmsize]
+ paddw m10, m4
+ pmaddubsw m4, m2, [r5 + 2 * mmsize]
+ paddw m12, m4
+ pmaddubsw m2, [r5 + 1 * mmsize]
+ paddw m0, m2
+ movu xm4, [r7 + r1 * 2] ; m4 = row 18
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 3 * mmsize]
+ paddw m11, m5
+ pmaddubsw m5, m3, [r5 + 2 * mmsize]
+ paddw m13, m5
+ pmaddubsw m3, [r5 + 1 * mmsize]
+ paddw m1, m3
+ movu xm5, [r7 + r4] ; m5 = row 19
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 3 * mmsize]
+ paddw m12, m6
+ pmaddubsw m4, [r5 + 2 * mmsize]
+ paddw m0, m4
+ lea r7, [r7 + r1 * 4]
+ movu xm6, [r7] ; m6 = row 20
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 3 * mmsize]
+ paddw m13, m7
+ pmaddubsw m5, [r5 + 2 * mmsize]
+ paddw m1, m5
+ movu xm7, [r7 + r1] ; m7 = row 21
+ punpckhbw xm2, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm2, 1
+ pmaddubsw m6, [r5 + 3 * mmsize]
+ paddw m0, m6
+ movu xm2, [r7 + r1 * 2] ; m2 = row 22
+ punpckhbw xm3, xm7, xm2
+ punpcklbw xm7, xm2
+ vinserti128 m7, m7, xm3, 1
+ pmaddubsw m7, [r5 + 3 * mmsize]
+ paddw m1, m7
+
+ pmulhrsw m8, m14 ; m8 = word: row 8
+ pmulhrsw m9, m14 ; m9 = word: row 9
+ pmulhrsw m10, m14 ; m10 = word: row 10
+ pmulhrsw m11, m14 ; m11 = word: row 11
+ pmulhrsw m12, m14 ; m12 = word: row 12
+ pmulhrsw m13, m14 ; m13 = word: row 13
+ pmulhrsw m0, m14 ; m0 = word: row 14
+ pmulhrsw m1, m14 ; m1 = word: row 15
+ packuswb m8, m9
+ packuswb m10, m11
+ packuswb m12, m13
+ packuswb m0, m1
+ vpermq m8, m8, 11011000b
+ vpermq m10, m10, 11011000b
+ vpermq m12, m12, 11011000b
+ vpermq m0, m0, 11011000b
+ vextracti128 xm9, m8, 1
+ vextracti128 xm11, m10, 1
+ vextracti128 xm13, m12, 1
+ vextracti128 xm1, m0, 1
+ movu [r8], xm8
+ movu [r8 + r3], xm9
+ movu [r8 + r3 * 2], xm10
+ movu [r8 + r6], xm11
+ lea r8, [r8 + r3 * 4]
+ movu [r8], xm12
+ movu [r8 + r3], xm13
+ movu [r8 + r3 * 2], xm0
+ movu [r8 + r6], xm1
+%endmacro
+
+%macro PROCESS_LUMA_AVX2_W16_8R 0
+ movu xm0, [r0] ; m0 = row 0
+ movu xm1, [r0 + r1] ; m1 = row 1
+ punpckhbw xm2, xm0, xm1
+ punpcklbw xm0, xm1
+ vinserti128 m0, m0, xm2, 1
+ pmaddubsw m0, [r5]
+ movu xm2, [r0 + r1 * 2] ; m2 = row 2
+ punpckhbw xm3, xm1, xm2
+ punpcklbw xm1, xm2
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m1, [r5]
+ movu xm3, [r0 + r4] ; m3 = row 3
+ punpckhbw xm4, xm2, xm3
+ punpcklbw xm2, xm3
+ vinserti128 m2, m2, xm4, 1
+ pmaddubsw m4, m2, [r5 + 1 * mmsize]
+ paddw m0, m4
+ pmaddubsw m2, [r5]
+ lea r7, [r0 + r1 * 4]
+ movu xm4, [r7] ; m4 = row 4
+ punpckhbw xm5, xm3, xm4
+ punpcklbw xm3, xm4
+ vinserti128 m3, m3, xm5, 1
+ pmaddubsw m5, m3, [r5 + 1 * mmsize]
+ paddw m1, m5
+ pmaddubsw m3, [r5]
+ movu xm5, [r7 + r1] ; m5 = row 5
+ punpckhbw xm6, xm4, xm5
+ punpcklbw xm4, xm5
+ vinserti128 m4, m4, xm6, 1
+ pmaddubsw m6, m4, [r5 + 2 * mmsize]
+ paddw m0, m6
+ pmaddubsw m6, m4, [r5 + 1 * mmsize]
+ paddw m2, m6
+ pmaddubsw m4, [r5]
+ movu xm6, [r7 + r1 * 2] ; m6 = row 6
+ punpckhbw xm7, xm5, xm6
+ punpcklbw xm5, xm6
+ vinserti128 m5, m5, xm7, 1
+ pmaddubsw m7, m5, [r5 + 2 * mmsize]
+ paddw m1, m7
+ pmaddubsw m7, m5, [r5 + 1 * mmsize]
+ paddw m3, m7
+ pmaddubsw m5, [r5]
+ movu xm7, [r7 + r4] ; m7 = row 7
+ punpckhbw xm8, xm6, xm7
+ punpcklbw xm6, xm7
+ vinserti128 m6, m6, xm8, 1
+ pmaddubsw m8, m6, [r5 + 3 * mmsize]
+ paddw m0, m8
+ pmaddubsw m8, m6, [r5 + 2 * mmsize]
+ paddw m2, m8
+ pmaddubsw m8, m6, [r5 + 1 * mmsize]
+ paddw m4, m8
+ pmaddubsw m6, [r5]
+ lea r7, [r7 + r1 * 4]
+ movu xm8, [r7] ; m8 = row 8
+ punpckhbw xm9, xm7, xm8
+ punpcklbw xm7, xm8
+ vinserti128 m7, m7, xm9, 1
+ pmaddubsw m9, m7, [r5 + 3 * mmsize]
+ paddw m1, m9
+ pmaddubsw m9, m7, [r5 + 2 * mmsize]
+ paddw m3, m9
+ pmaddubsw m9, m7, [r5 + 1 * mmsize]
+ paddw m5, m9
+ pmaddubsw m7, [r5]
+ movu xm9, [r7 + r1] ; m9 = row 9
+ punpckhbw xm10, xm8, xm9
+ punpcklbw xm8, xm9
+ vinserti128 m8, m8, xm10, 1
+ pmaddubsw m10, m8, [r5 + 3 * mmsize]
+ paddw m2, m10
+ pmaddubsw m10, m8, [r5 + 2 * mmsize]
+ paddw m4, m10
+ pmaddubsw m10, m8, [r5 + 1 * mmsize]
+ paddw m6, m10
+ movu xm10, [r7 + r1 * 2] ; m10 = row 10
+ punpckhbw xm11, xm9, xm10
+ punpcklbw xm9, xm10
+ vinserti128 m9, m9, xm11, 1
+ pmaddubsw m11, m9, [r5 + 3 * mmsize]
+ paddw m3, m11
+ pmaddubsw m11, m9, [r5 + 2 * mmsize]
+ paddw m5, m11
+ pmaddubsw m11, m9, [r5 + 1 * mmsize]
+ paddw m7, m11
+ movu xm11, [r7 + r4] ; m11 = row 11
+ punpckhbw xm12, xm10, xm11
+ punpcklbw xm10, xm11
+ vinserti128 m10, m10, xm12, 1
+ pmaddubsw m12, m10, [r5 + 3 * mmsize]
+ paddw m4, m12
+ pmaddubsw m12, m10, [r5 + 2 * mmsize]
+ paddw m6, m12
+ lea r7, [r7 + r1 * 4]
+ movu xm12, [r7] ; m12 = row 12
+ punpckhbw xm13, xm11, xm12
+ punpcklbw xm11, xm12
+ vinserti128 m11, m11, xm13, 1
+ pmaddubsw m13, m11, [r5 + 3 * mmsize]
+ paddw m5, m13
+ pmaddubsw m13, m11, [r5 + 2 * mmsize]
+ paddw m7, m13
+
+ pmulhrsw m0, m14 ; m0 = word: row 0
+ pmulhrsw m1, m14 ; m1 = word: row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2
+ pmulhrsw m3, m14 ; m3 = word: row 3
+ pmulhrsw m4, m14 ; m4 = word: row 4
+ pmulhrsw m5, m14 ; m5 = word: row 5
+ packuswb m0, m1
+ packuswb m2, m3
+ packuswb m4, m5
+ vpermq m0, m0, 11011000b
+ vpermq m2, m2, 11011000b
+ vpermq m4, m4, 11011000b
+ vextracti128 xm1, m0, 1
+ vextracti128 xm3, m2, 1
+ vextracti128 xm5, m4, 1
+ movu [r2], xm0
+ movu [r2 + r3], xm1
+ movu [r2 + r3 * 2], xm2
+ movu [r2 + r6], xm3
+ lea r8, [r2 + r3 * 4]
+ movu [r8], xm4
+ movu [r8 + r3], xm5
+
+ movu xm13, [r7 + r1] ; m13 = row 13
+ punpckhbw xm0, xm12, xm13
+ punpcklbw xm12, xm13
+ vinserti128 m12, m12, xm0, 1
+ pmaddubsw m0, m12, [r5 + 3 * mmsize]
+ paddw m6, m0
+ movu xm0, [r7 + r1 * 2] ; m0 = row 14
+ punpckhbw xm1, xm13, xm0
+ punpcklbw xm13, xm0
+ vinserti128 m13, m13, xm1, 1
+ pmaddubsw m1, m13, [r5 + 3 * mmsize]
+ paddw m7, m1
+
+ pmulhrsw m6, m14 ; m6 = word: row 6
+ pmulhrsw m7, m14 ; m7 = word: row 7
+ packuswb m6, m7
+ vpermq m6, m6, 11011000b
+ vextracti128 xm7, m6, 1
+ movu [r8 + r3 * 2], xm6
+ movu [r8 + r6], xm7
+%endmacro
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_24x32, 4, 11, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ lea r10, [r1 * 4]
+ mova m14, [pw_512]
+ mov r9d, 2
+.loopH:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+
+ movq xm1, [r0] ; m1 = row 0
+ movq xm2, [r0 + r1] ; m2 = row 1
+ punpcklbw xm1, xm2
+ movq xm3, [r0 + r1 * 2] ; m3 = row 2
+ punpcklbw xm2, xm3
+ vinserti128 m5, m1, xm2, 1
+ pmaddubsw m5, [r5]
+ movq xm4, [r0 + r4] ; m4 = row 3
+ punpcklbw xm3, xm4
+ lea r7, [r0 + r1 * 4]
+ movq xm1, [r7] ; m1 = row 4
+ punpcklbw xm4, xm1
+ vinserti128 m2, m3, xm4, 1
+ pmaddubsw m0, m2, [r5 + 1 * mmsize]
+ paddw m5, m0
+ pmaddubsw m2, [r5]
+ movq xm3, [r7 + r1] ; m3 = row 5
+ punpcklbw xm1, xm3
+ movq xm4, [r7 + r1 * 2] ; m4 = row 6
+ punpcklbw xm3, xm4
+ vinserti128 m1, m1, xm3, 1
+ pmaddubsw m3, m1, [r5 + 2 * mmsize]
+ paddw m5, m3
+ pmaddubsw m0, m1, [r5 + 1 * mmsize]
+ paddw m2, m0
+ pmaddubsw m1, [r5]
+ movq xm3, [r7 + r4] ; m3 = row 7
+ punpcklbw xm4, xm3
+ lea r7, [r7 + r1 * 4]
+ movq xm0, [r7] ; m0 = row 8
+ punpcklbw xm3, xm0
+ vinserti128 m4, m4, xm3, 1
+ pmaddubsw m3, m4, [r5 + 3 * mmsize]
+ paddw m5, m3
+ pmaddubsw m3, m4, [r5 + 2 * mmsize]
+ paddw m2, m3
+ pmaddubsw m3, m4, [r5 + 1 * mmsize]
+ paddw m1, m3
+ pmaddubsw m4, [r5]
+ movq xm3, [r7 + r1] ; m3 = row 9
+ punpcklbw xm0, xm3
+ movq xm6, [r7 + r1 * 2] ; m6 = row 10
+ punpcklbw xm3, xm6
+ vinserti128 m0, m0, xm3, 1
+ pmaddubsw m3, m0, [r5 + 3 * mmsize]
+ paddw m2, m3
+ pmaddubsw m3, m0, [r5 + 2 * mmsize]
+ paddw m1, m3
+ pmaddubsw m3, m0, [r5 + 1 * mmsize]
+ paddw m4, m3
+ pmaddubsw m0, [r5]
+
+ movq xm3, [r7 + r4] ; m3 = row 11
+ punpcklbw xm6, xm3
+ lea r7, [r7 + r1 * 4]
+ movq xm7, [r7] ; m7 = row 12
+ punpcklbw xm3, xm7
+ vinserti128 m6, m6, xm3, 1
+ pmaddubsw m3, m6, [r5 + 3 * mmsize]
+ paddw m1, m3
+ pmaddubsw m3, m6, [r5 + 2 * mmsize]
+ paddw m4, m3
+ pmaddubsw m3, m6, [r5 + 1 * mmsize]
+ paddw m0, m3
+ pmaddubsw m6, [r5]
+ movq xm3, [r7 + r1] ; m3 = row 13
+ punpcklbw xm7, xm3
+ movq xm8, [r7 + r1 * 2] ; m8 = row 14
+ punpcklbw xm3, xm8
+ vinserti128 m7, m7, xm3, 1
+ pmaddubsw m3, m7, [r5 + 3 * mmsize]
+ paddw m4, m3
+ pmaddubsw m3, m7, [r5 + 2 * mmsize]
+ paddw m0, m3
+ pmaddubsw m3, m7, [r5 + 1 * mmsize]
+ paddw m6, m3
+ pmaddubsw m7, [r5]
+ movq xm3, [r7 + r4] ; m3 = row 15
+ punpcklbw xm8, xm3
+ lea r7, [r7 + r1 * 4]
+ movq xm9, [r7] ; m9 = row 16
+ punpcklbw xm3, xm9
+ vinserti128 m8, m8, xm3, 1
+ pmaddubsw m3, m8, [r5 + 3 * mmsize]
+ paddw m0, m3
+ pmaddubsw m3, m8, [r5 + 2 * mmsize]
+ paddw m6, m3
+ pmaddubsw m3, m8, [r5 + 1 * mmsize]
+ paddw m7, m3
+ pmaddubsw m8, [r5]
+ movq xm3, [r7 + r1] ; m3 = row 17
+ punpcklbw xm9, xm3
+ movq xm10, [r7 + r1 * 2] ; m10 = row 18
+ punpcklbw xm3, xm10
+ vinserti128 m9, m9, xm3, 1
+ pmaddubsw m3, m9, [r5 + 3 * mmsize]
+ paddw m6, m3
+ pmaddubsw m3, m9, [r5 + 2 * mmsize]
+ paddw m7, m3
+ pmaddubsw m3, m9, [r5 + 1 * mmsize]
+ paddw m8, m3
+ movq xm3, [r7 + r4] ; m3 = row 19
+ punpcklbw xm10, xm3
+ lea r7, [r7 + r1 * 4]
+ movq xm9, [r7] ; m9 = row 20
+ punpcklbw xm3, xm9
+ vinserti128 m10, m10, xm3, 1
+ pmaddubsw m3, m10, [r5 + 3 * mmsize]
+ paddw m7, m3
+ pmaddubsw m3, m10, [r5 + 2 * mmsize]
+ paddw m8, m3
+ movq xm3, [r7 + r1] ; m3 = row 21
+ punpcklbw xm9, xm3
+ movq xm10, [r7 + r1 * 2] ; m10 = row 22
+ punpcklbw xm3, xm10
+ vinserti128 m9, m9, xm3, 1
+ pmaddubsw m3, m9, [r5 + 3 * mmsize]
+ paddw m8, m3
+
+ pmulhrsw m5, m14 ; m5 = word: row 0, row 1
+ pmulhrsw m2, m14 ; m2 = word: row 2, row 3
+ pmulhrsw m1, m14 ; m1 = word: row 4, row 5
+ pmulhrsw m4, m14 ; m4 = word: row 6, row 7
+ pmulhrsw m0, m14 ; m0 = word: row 8, row 9
+ pmulhrsw m6, m14 ; m6 = word: row 10, row 11
+ pmulhrsw m7, m14 ; m7 = word: row 12, row 13
+ pmulhrsw m8, m14 ; m8 = word: row 14, row 15
+ packuswb m5, m2
+ packuswb m1, m4
+ packuswb m0, m6
+ packuswb m7, m8
+ vextracti128 xm2, m5, 1
+ vextracti128 xm4, m1, 1
+ vextracti128 xm6, m0, 1
+ vextracti128 xm8, m7, 1
+ movq [r2], xm5
+ movq [r2 + r3], xm2
+ movhps [r2 + r3 * 2], xm5
+ movhps [r2 + r6], xm2
+ lea r8, [r2 + r3 * 4]
+ movq [r8], xm1
+ movq [r8 + r3], xm4
+ movhps [r8 + r3 * 2], xm1
+ movhps [r8 + r6], xm4
+ lea r8, [r8 + r3 * 4]
+ movq [r8], xm0
+ movq [r8 + r3], xm6
+ movhps [r8 + r3 * 2], xm0
+ movhps [r8 + r6], xm6
+ lea r8, [r8 + r3 * 4]
+ movq [r8], xm7
+ movq [r8 + r3], xm8
+ movhps [r8 + r3 * 2], xm7
+ movhps [r8 + r6], xm8
+
+ sub r7, r10
+ lea r0, [r7 - 16]
+ lea r2, [r8 + r3 * 4 - 16]
+ dec r9d
+ jnz .loopH
+ RET
+%endif
+
+%macro FILTER_VER_LUMA_AVX2_32xN 2
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_%1x%2, 4, 12, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ lea r11, [r1 * 4]
+ mova m14, [pw_512]
+ mov r9d, %2 / 16
+.loopH:
+ mov r10d, %1 / 16
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r10d
+ jnz .loopW
+ sub r7, r11
+ lea r0, [r7 - 16]
+ lea r2, [r8 + r3 * 4 - 16]
+ dec r9d
+ jnz .loopH
+ RET
+%endif
+%endmacro
+
+FILTER_VER_LUMA_AVX2_32xN 32, 32
+FILTER_VER_LUMA_AVX2_32xN 32, 64
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_32x16, 4, 10, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+ mov r9d, 2
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r9d
+ jnz .loopW
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_32x24, 4, 10, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+ mov r9d, 2
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r9d
+ jnz .loopW
+ lea r9, [r1 * 4]
+ sub r7, r9
+ lea r0, [r7 - 16]
+ lea r2, [r8 + r3 * 4 - 16]
+ mov r9d, 2
+.loop:
+ PROCESS_LUMA_AVX2_W16_8R
+ add r2, 16
+ add r0, 16
+ dec r9d
+ jnz .loop
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_32x8, 4, 10, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+ mov r9d, 2
+.loopW:
+ PROCESS_LUMA_AVX2_W16_8R
+ add r2, 16
+ add r0, 16
+ dec r9d
+ jnz .loopW
+ RET
+%endif
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_48x64, 4, 12, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ lea r11, [r1 * 4]
+ mova m14, [pw_512]
+ mov r9d, 4
+.loopH:
+ mov r10d, 3
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r10d
+ jnz .loopW
+ sub r7, r11
+ lea r0, [r7 - 32]
+ lea r2, [r8 + r3 * 4 - 32]
+ dec r9d
+ jnz .loopH
+ RET
+%endif
+
+%macro FILTER_VER_LUMA_AVX2_64xN 2
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_%1x%2, 4, 12, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ lea r11, [r1 * 4]
+ mova m14, [pw_512]
+ mov r9d, %2 / 16
+.loopH:
+ mov r10d, %1 / 16
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r10d
+ jnz .loopW
+ sub r7, r11
+ lea r0, [r7 - 48]
+ lea r2, [r8 + r3 * 4 - 48]
+ dec r9d
+ jnz .loopH
+ RET
+%endif
+%endmacro
+
+FILTER_VER_LUMA_AVX2_64xN 64, 32
+FILTER_VER_LUMA_AVX2_64xN 64, 48
+FILTER_VER_LUMA_AVX2_64xN 64, 64
+
+INIT_YMM avx2
+%if ARCH_X86_64 == 1
+cglobal interp_8tap_vert_pp_64x16, 4, 10, 15
+ mov r4d, r4m
+ shl r4d, 7
+
+%ifdef PIC
+ lea r5, [tab_LumaCoeffVer_32]
+ add r5, r4
+%else
+ lea r5, [tab_LumaCoeffVer_32 + r4]
+%endif
+
+ lea r4, [r1 * 3]
+ sub r0, r4
+ lea r6, [r3 * 3]
+ mova m14, [pw_512]
+ mov r9d, 4
+.loopW:
+ PROCESS_LUMA_AVX2_W16_16R
+ add r2, 16
+ add r0, 16
+ dec r9d
+ jnz .loopW
+ RET
+%endif
+