Skip to content

GitLab

Commit a755b6e3 authored by Henrik Gramner's avatar Henrik Gramner
Browse files
Options

Clip coefficients in SSSE3/AVX2 inverse transform asm

parent eb01bdb9
Pipeline #3601 passed with stages
      in 4 minutes and 56 seconds
      Hide whitespace changes
      Inline Side-by-side
      Showing with 358 additions and 358 deletions
      src/x86/itx.asm
      View file @ a755b6e3
      ......@@ -231,20 +231,20 @@ SECTION .text
      psubw m%1, m%3
      pmulhrsw m%1, m%6 ; t1
      pmulhrsw m%5, m%6 ; t0
      psubw m%3, m%1, m%2
      paddw m%2, m%1
      paddw m%1, m%5, m%4
      psubw m%4, m%5, m%4
      psubsw m%3, m%1, m%2
      paddsw m%2, m%1
      paddsw m%1, m%5, m%4
      psubsw m%4, m%5, m%4
      %endmacro
      %macro IDCT8_1D 11 ; src[1-8], tmp[1-2], pd_2048
      ITX_MULSUB_2W %6, %4, %9, %10, %11, 3406, 2276 ; t5a, t6a
      ITX_MULSUB_2W %2, %8, %9, %10, %11, 799, 4017 ; t4a, t7a
      ITX_MULSUB_2W %3, %7, %9, %10, %11, 1567, 3784 ; t2, t3
      paddw m%9, m%2, m%6 ; t4
      psubw m%2, m%6 ; t5a
      paddw m%10, m%8, m%4 ; t7
      psubw m%8, m%4 ; t6a
      paddsw m%9, m%2, m%6 ; t4
      psubsw m%2, m%6 ; t5a
      paddsw m%10, m%8, m%4 ; t7
      psubsw m%8, m%4 ; t6a
      vpbroadcastd m%4, [o(pw_2896x8)]
      psubw m%6, m%1, m%5
      paddw m%1, m%5
      ......@@ -254,18 +254,18 @@ SECTION .text
      pmulhrsw m%6, m%4 ; t1
      pmulhrsw m%8, m%4 ; t6
      pmulhrsw m%5, m%4 ; t5
      psubw m%4, m%1, m%7 ; dct4 out3
      paddw m%1, m%7 ; dct4 out0
      paddw m%7, m%6, m%3 ; dct4 out1
      psubw m%6, m%3 ; dct4 out2
      paddw m%2, m%7, m%8 ; out1
      psubw m%7, m%8 ; out6
      psubw m%8, m%1, m%10 ; out7
      paddw m%1, m%10 ; out0
      paddw m%3, m%6, m%5 ; out2
      psubw m%6, m%5 ; out5
      psubw m%5, m%4, m%9 ; out4
      paddw m%4, m%9 ; out3
      psubsw m%4, m%1, m%7 ; dct4 out3
      paddsw m%1, m%7 ; dct4 out0
      paddsw m%7, m%6, m%3 ; dct4 out1
      psubsw m%6, m%3 ; dct4 out2
      paddsw m%2, m%7, m%8 ; out1
      psubsw m%7, m%8 ; out6
      psubsw m%8, m%1, m%10 ; out7
      paddsw m%1, m%10 ; out0
      paddsw m%3, m%6, m%5 ; out2
      psubsw m%6, m%5 ; out5
      psubsw m%5, m%4, m%9 ; out4
      paddsw m%4, m%9 ; out3
      %endmacro
      ; in1 = %1, in3 = %2, in5 = %3, in7 = %4
      ......@@ -275,25 +275,25 @@ SECTION .text
      ITX_MULSUB_2W %5, %4, %9, %10, %11, 3166, 2598 ; t9a, t14a
      ITX_MULSUB_2W %3, %6, %9, %10, %11, 1931, 3612 ; t10a, t13a
      ITX_MULSUB_2W %7, %2, %9, %10, %11, 3920, 1189 ; t11a, t12a
      psubw m%9, m%2, m%6 ; t13
      paddw m%6, m%2 ; t12
      psubw m%2, m%8, m%4 ; t14
      paddw m%8, m%4 ; t15
      psubw m%4, m%7, m%3 ; t10
      paddw m%3, m%7 ; t11
      psubw m%7, m%1, m%5 ; t9
      paddw m%1, m%5 ; t8
      psubsw m%9, m%2, m%6 ; t13
      paddsw m%6, m%2 ; t12
      psubsw m%2, m%8, m%4 ; t14
      paddsw m%8, m%4 ; t15
      psubsw m%4, m%7, m%3 ; t10
      paddsw m%3, m%7 ; t11
      psubsw m%7, m%1, m%5 ; t9
      paddsw m%1, m%5 ; t8
      ITX_MULSUB_2W %2, %7, %5, %10, %11, 1567, 3784 ; t9a, t14a
      ITX_MULSUB_2W %9, %4, %5, %10, %11, m3784, 1567 ; t10a, t13a
      vpbroadcastd m%10, [o(pw_2896x8)]
      psubw m%5, m%2, m%9 ; t10
      paddw m%2, m%9 ; t9
      psubw m%9, m%1, m%3 ; t11a
      paddw m%1, m%3 ; t8a
      psubw m%3, m%7, m%4 ; t13
      paddw m%7, m%4 ; t14
      psubw m%4, m%8, m%6 ; t12a
      paddw m%8, m%6 ; t15a
      psubsw m%5, m%2, m%9 ; t10
      paddsw m%2, m%9 ; t9
      psubsw m%9, m%1, m%3 ; t11a
      paddsw m%1, m%3 ; t8a
      psubsw m%3, m%7, m%4 ; t13
      paddsw m%7, m%4 ; t14
      psubsw m%4, m%8, m%6 ; t12a
      paddsw m%8, m%6 ; t15a
      paddw m%6, m%3, m%5 ; t13a
      psubw m%3, m%5 ; t10a
      paddw m%5, m%4, m%9 ; t12
      ......@@ -458,8 +458,8 @@ ALIGN function_align
      vpbroadcastd m4, [o(pw_2896x8)]
      pmulhrsw m0, m4 ; t0 t1
      %endif
      psubw m1, m0, m2 ; out3 out2
      paddw m0, m2 ; out0 out1
      psubsw m1, m0, m2 ; out3 out2
      paddsw m0, m2 ; out0 out1
      %endmacro
      %macro IADST4_1D_PACKED 0
      ......@@ -693,22 +693,22 @@ cglobal iidentity_4x4_internal, 0, 5, 6, dst, stride, c, eob, tx2
      ITX_MUL2X_PACK 4, 1, 2, 6, 3406, 2276, 1 ; t5a t6a
      ITX_MUL2X_PACK 3, 1, 2, 6, 1567, 3784 ; t3 t2
      vpbroadcastd m6, [o(pw_2896x8)]
      psubw m2, m5, m4 ; t4 t7
      paddw m5, m4 ; t5a t6a
      psubsw m2, m5, m4 ; t4 t7
      paddsw m5, m4 ; t5a t6a
      pshufd m4, m2, q1032
      psubw m1, m2, m4
      paddw m4, m2
      vpblendd m4, m4, m1, 0xcc
      pmulhrsw m0, m6 ; t0 t1
      pmulhrsw m4, m6 ; t6 t5
      psubw m1, m0, m3 ; tmp3 tmp2
      paddw m0, m3 ; tmp0 tmp1
      psubsw m1, m0, m3 ; tmp3 tmp2
      paddsw m0, m3 ; tmp0 tmp1
      shufps m2, m5, m4, q1032 ; t7 t6
      vpblendd m5, m5, m4, 0xcc ; t4 t5
      psubw m3, m0, m2 ; out7 out6
      paddw m0, m2 ; out0 out1
      psubw m2, m1, m5 ; out4 out5
      paddw m1, m5 ; out3 out2
      psubsw m3, m0, m2 ; out7 out6
      paddsw m0, m2 ; out0 out1
      psubsw m2, m1, m5 ; out4 out5
      paddsw m1, m5 ; out3 out2
      %endmacro
      %macro IADST8_1D_PACKED 0
      ......@@ -721,19 +721,19 @@ cglobal iidentity_4x4_internal, 0, 5, 6, dst, stride, c, eob, tx2
      ITX_MUL2X_PACK 1, 4, 5, 6, 1931, 3612 ; t2a t3a
      ITX_MUL2X_PACK 2, 4, 5, 6, 3166, 2598 ; t4a t5a
      ITX_MUL2X_PACK 3, 4, 5, 6, 3920, 1189 ; t6a t7a
      psubw m4, m0, m2 ; t4 t5
      paddw m0, m2 ; t0 t1
      psubw m5, m1, m3 ; t6 t7
      paddw m1, m3 ; t2 t3
      psubsw m4, m0, m2 ; t4 t5
      paddsw m0, m2 ; t0 t1
      psubsw m5, m1, m3 ; t6 t7
      paddsw m1, m3 ; t2 t3
      shufps m2, m5, m4, q1032
      punpckhwd m4, m2
      punpcklwd m5, m2
      ITX_MUL2X_PACK 4, 2, 3, 6, 1567, 3784, 1 ; t5a t4a
      ITX_MUL2X_PACK 5, 2, 3, 6, 3784, 1567 ; t7a t6a
      psubw m2, m0, m1 ; t2 t3
      paddw m0, m1 ; out0 -out7
      psubw m1, m4, m5 ; t7 t6
      paddw m4, m5 ; out6 -out1
      psubsw m2, m0, m1 ; t2 t3
      paddsw m0, m1 ; out0 -out7
      psubsw m1, m4, m5 ; t7 t6
      paddsw m4, m5 ; out6 -out1
      vpbroadcastd m5, [o(pw_2896x8)]
      vpblendd m3, m0, m4, 0x33 ; out6 -out7
      vpblendd m0, m0, m4, 0xcc ; out0 -out1
      ......@@ -981,10 +981,10 @@ cglobal iidentity_4x8_internal, 0, 5, 7, dst, stride, c, eob, tx2
      ITX_MUL2X_PACK 7, 2, 4, 10, 799, 4017, 1 ; t4a t7a
      ITX_MUL2X_PACK 3, 2, 4, 10, 3406, 2276, 1 ; t5a t6a
      ITX_MUL2X_PACK 6, 2, 4, 10, 1567, 3784 ; t3 t2
      psubw m2, m8, m0 ; t9 t14
      paddw m8, m0 ; t8 t15
      psubw m0, m1, m5 ; t10 t13
      paddw m1, m5 ; t11 t12
      psubsw m2, m8, m0 ; t9 t14
      paddsw m8, m0 ; t8 t15
      psubsw m0, m1, m5 ; t10 t13
      paddsw m1, m5 ; t11 t12
      %if mmsize > 16
      vbroadcasti128 m5, [o(deint_shuf)]
      %else
      ......@@ -996,12 +996,12 @@ cglobal iidentity_4x8_internal, 0, 5, 7, dst, stride, c, eob, tx2
      ITX_MUL2X_PACK 2, 4, _, 10, 4, 5, 4 ; t9a t14a
      vpbroadcastd m4, [o(pw_m1567_m3784)] ; reuse pw_m3784_1567
      ITX_MUL2X_PACK 0, 5, _, 10, 5, 4, 4 ; t10a t13a
      psubw m5, m7, m3 ; t5a t6a
      paddw m7, m3 ; t4 t7
      psubw m4, m8, m1 ; t11a t12a
      paddw m8, m1 ; t8a t15a
      paddw m1, m2, m0 ; t9 t14
      psubw m2, m0 ; t10 t13
      psubsw m5, m7, m3 ; t5a t6a
      paddsw m7, m3 ; t4 t7
      psubsw m4, m8, m1 ; t11a t12a
      paddsw m8, m1 ; t8a t15a
      paddsw m1, m2, m0 ; t9 t14
      psubsw m2, m0 ; t10 t13
      punpckhqdq m0, m8, m1 ; t15a t14
      punpcklqdq m8, m1 ; t8a t9
      pshufd m3, m5, q1032
      ......@@ -1019,20 +1019,20 @@ cglobal iidentity_4x8_internal, 0, 5, 7, dst, stride, c, eob, tx2
      pmulhrsw m5, m1 ; t12 t13a
      shufps m2, m7, m3, q1032 ; t7 t6
      vpblendd m7, m7, m3, 0xcc ; t4 t5
      psubw m1, m9, m6 ; dct4 out3 out2
      paddw m9, m6 ; dct4 out0 out1
      psubw m3, m9, m2 ; dct8 out7 out6
      paddw m9, m2 ; dct8 out0 out1
      psubw m2, m1, m7 ; dct8 out4 out5
      paddw m1, m7 ; dct8 out3 out2
      psubw m7, m9, m0 ; out15 out14
      paddw m0, m9 ; out0 out1
      psubw m6, m1, m5 ; out12 out13
      paddw m1, m5 ; out3 out2
      psubw m5, m2, m4 ; out11 out10
      paddw m2, m4 ; out4 out5
      psubw m4, m3, m8 ; out8 out9
      paddw m3, m8 ; out7 out6
      psubsw m1, m9, m6 ; dct4 out3 out2
      paddsw m9, m6 ; dct4 out0 out1
      psubsw m3, m9, m2 ; dct8 out7 out6
      paddsw m9, m2 ; dct8 out0 out1
      psubsw m2, m1, m7 ; dct8 out4 out5
      paddsw m1, m7 ; dct8 out3 out2
      psubsw m7, m9, m0 ; out15 out14
      paddsw m0, m9 ; out0 out1
      psubsw m6, m1, m5 ; out12 out13
      paddsw m1, m5 ; out3 out2
      psubsw m5, m2, m4 ; out11 out10
      paddsw m2, m4 ; out4 out5
      psubsw m4, m3, m8 ; out8 out9
      paddsw m3, m8 ; out7 out6
      %endmacro
      INV_TXFM_4X16_FN dct, dct, 0
      ......@@ -1153,20 +1153,20 @@ ALIGN function_align
      ITX_MUL4X_PACK 1, 2, 5, 6, 8, 1751, 3703, 2440, 3290, 3
      ITX_MUL4X_PACK 3, 2, 5, 6, 8, 3035, 2751, 3513, 2106, 3
      ITX_MUL4X_PACK 4, 2, 5, 6, 8, 3857, 1380, 4052, 601, 3
      psubw m2, m0, m3 ; t9a t8a t11a t10a
      paddw m0, m3 ; t1a t0a t3a t2a
      psubw m3, m1, m4 ; t13a t12a t15a t14a
      paddw m1, m4 ; t5a t4a t7a t6a
      psubsw m2, m0, m3 ; t9a t8a t11a t10a
      paddsw m0, m3 ; t1a t0a t3a t2a
      psubsw m3, m1, m4 ; t13a t12a t15a t14a
      paddsw m1, m4 ; t5a t4a t7a t6a
      ITX_MUL4X_PACK 2, 4, 5, 6, 8, 799, 4017, 3406, 2276, 3
      psubw m6, m7, m5
      ITX_MUL2X_PACK 3, 5, _, 8, 6, 4, 6
      vpbroadcastd m6, [o(pw_m3784_1567)]
      vpbroadcastd m5, [o(pw_1567_3784)]
      psubw m4, m0, m1 ; t5 t4 t7 t6
      paddw m0, m1 ; t1 t0 t3 t2
      psubw m1, m2, m3 ; t13a t12a t15a t14a
      paddw m2, m3 ; t9a t8a t11a t10a
      psubw m3, m7, m6
      psubsw m4, m0, m1 ; t5 t4 t7 t6
      paddsw m0, m1 ; t1 t0 t3 t2
      psubsw m1, m2, m3 ; t13a t12a t15a t14a
      paddsw m2, m3 ; t9a t8a t11a t10a
      psubw m3, m7, m6 ; pw_3784_m1567
      vpblendd m6, m6, m3, 0xf0
      ITX_MUL2X_PACK 4, 3, _, 8, 6, 5, 4 ; t4a t5a t7a t6a
      ITX_MUL2X_PACK 1, 3, _, 8, 6, 5, 4 ; t12 t13 t15 t14
      ......@@ -1179,10 +1179,10 @@ ALIGN function_align
      vinserti128 m4, m4, xm1, 1 ; t4a t5a t12 t13
      vpbroadcastd m5, [o(pw_2896x8)]
      pshufd m2, m2, q1032 ; t6a t7a t14 t15
      psubw m1, m0, m3 ; t3a t2a t11 t10
      paddw m0, m3 ; -out15 out0 out14 -out1
      paddw m3, m4, m2 ; -out3 out12 out2 -out13
      psubw m4, m2 ; t6 t7 t14a t15a
      psubsw m1, m0, m3 ; t3a t2a t11 t10
      paddsw m0, m3 ; -out15 out0 out14 -out1
      paddsw m3, m4, m2 ; -out3 out12 out2 -out13
      psubsw m4, m2 ; t6 t7 t14a t15a
      shufps m2, m1, m4, q1032 ; t2a t6 t10 t14a
      vpblendd m4, m4, m1, 0x33 ; t3a t7 t11 t15a
      paddw m1, m2, m4
      ......@@ -1902,53 +1902,53 @@ ALIGN function_align
      ITX_MUL2X_PACK 6, 4, 9, 10, 3513, 2106, 3 ; t10 t11
      ITX_MUL2X_PACK 7, 4, 9, 10, 3857, 1380, 3 ; t12 t13
      ITX_MUL2X_PACK 8, 4, 9, 10, 4052, 601, 3 ; t14 t15
      psubw m4, m0, m5 ; t9a t8a
      paddw m0, m5 ; t1a t0a
      psubw m5, m1, m6 ; t11a t10a
      paddw m1, m6 ; t3a t2a
      psubw m6, m2, m7 ; t13a t12a
      paddw m2, m7 ; t5a t4a
      psubw m7, m3, m8 ; t15a t14a
      paddw m3, m8 ; t7a t6a
      psubsw m4, m0, m5 ; t9a t8a
      paddsw m0, m5 ; t1a t0a
      psubsw m5, m1, m6 ; t11a t10a
      paddsw m1, m6 ; t3a t2a
      psubsw m6, m2, m7 ; t13a t12a
      paddsw m2, m7 ; t5a t4a
      psubsw m7, m3, m8 ; t15a t14a
      paddsw m3, m8 ; t7a t6a
      vpbroadcastd m11, [o(pw_m4017_799)]
      vpbroadcastd m12, [o(pw_799_4017)]
      pxor m9, m9
      ITX_MUL2X_PACK 4, 8, _, 10, 11, 12, 6 ; t8 t9
      psubw m8, m9, m11
      psubw m8, m9, m11 ; pw_4017_m799
      ITX_MUL2X_PACK 6, 12, _, 10, 12, 8, 6 ; t12 t13
      vpbroadcastd m11, [o(pw_m2276_3406)]
      vpbroadcastd m12, [o(pw_3406_2276)]
      ITX_MUL2X_PACK 5, 8, _, 10, 11, 12, 6 ; t10 t11
      psubw m8, m9, m11
      psubw m8, m9, m11 ; pw_2276_m3406
      ITX_MUL2X_PACK 7, 12, _, 10, 12, 8, 6 ; t14 t15
      psubw m8, m1, m3 ; t7 t6
      paddw m1, m3 ; t3 t2
      psubw m3, m0, m2 ; t5 t4
      paddw m0, m2 ; t1 t0
      psubw m2, m5, m7 ; t14a t15a
      paddw m7, m5 ; t10a t11a
      psubw m5, m4, m6 ; t12a t13a
      paddw m4, m6 ; t8a t9a
      psubsw m8, m1, m3 ; t7 t6
      paddsw m1, m3 ; t3 t2
      psubsw m3, m0, m2 ; t5 t4
      paddsw m0, m2 ; t1 t0
      psubsw m2, m5, m7 ; t14a t15a
      paddsw m7, m5 ; t10a t11a
      psubsw m5, m4, m6 ; t12a t13a
      paddsw m4, m6 ; t8a t9a
      vpbroadcastd m11, [o(pw_m3784_1567)]
      vpbroadcastd m12, [o(pw_1567_3784)]
      ITX_MUL2X_PACK 3, 6, _, 10, 11, 12, 4 ; t4a t5a
      psubw m6, m9, m11
      psubw m6, m9, m11 ; pw_3784_m1567
      ITX_MUL2X_PACK 8, 12, _, 10, 12, 6, 4 ; t6a t7a
      vpbroadcastd m11, [o(pw_m1567_3784)]
      vpbroadcastd m12, [o(pw_3784_1567)]
      ITX_MUL2X_PACK 2, 6, _, 10, 11, 12, 4 ; t15 t14
      psubw m6, m9, m11
      psubw m6, m9, m11 ; pw_1567_m3784
      ITX_MUL2X_PACK 5, 12, _, 10, 12, 6, 4 ; t13 t12
      vbroadcasti128 m11, [o(deint_shuf)]
      vpbroadcastd m12, [o(pw_2896x8)]
      psubw m6, m0, m1 ; t3a t2a
      paddw m0, m1 ; -out15 out0
      paddw m1, m2, m5 ; -out13 out2
      psubw m5, m2 ; t15a t14a
      paddw m2, m4, m7 ; -out1 out14
      psubw m4, m7 ; t10 t11
      psubw m7, m3, m8 ; t6 t7
      paddw m8, m3 ; -out3 out12
      psubsw m6, m0, m1 ; t3a t2a
      paddsw m0, m1 ; -out15 out0
      paddsw m1, m2, m5 ; -out13 out2
      psubsw m5, m2 ; t15a t14a
      paddsw m2, m4, m7 ; -out1 out14
      psubsw m4, m7 ; t10 t11
      psubsw m7, m3, m8 ; t6 t7
      paddsw m8, m3 ; -out3 out12
      REPX {pshufb x, m11}, m6, m4, m0, m2
      vpblendd m3, m6, m4, 0xcc ; t3a t11
      shufps m6, m6, m4, q1032 ; t2a t10
      ......@@ -2580,25 +2580,25 @@ ALIGN function_align
      ITX_MULSUB_2W 3, 4, 8, 9, 10, 3166, 2598 ; t5a, t4a
      ITX_MULSUB_2W 1, 6, 8, 9, 10, 3920, 1189 ; t7a, t6a
      ITX_MULSUB_2W 5, 2, 8, 9, 10, 1931, 3612 ; t3a, t2a
      psubw m8, m2, m6 ; t6
      paddw m2, m6 ; t2
      psubw m6, m0, m4 ; t4
      paddw m0, m4 ; t0
      psubw m4, m5, m1 ; t7
      paddw m5, m1 ; t3
      psubw m1, m7, m3 ; t5
      paddw m7, m3 ; t1
      psubsw m8, m2, m6 ; t6
      paddsw m2, m6 ; t2
      psubsw m6, m0, m4 ; t4
      paddsw m0, m4 ; t0
      psubsw m4, m5, m1 ; t7
      paddsw m5, m1 ; t3
      psubsw m1, m7, m3 ; t5
      paddsw m7, m3 ; t1
      ITX_MULSUB_2W 6, 1, 3, 9, 10, 1567, 3784 ; t5a, t4a
      ITX_MULSUB_2W 4, 8, 3, 9, 10, 3784, 1567 ; t6a, t7a
      psubw m9, m6, m8 ; t7
      paddw m6, m8 ; out6
      psubsw m9, m6, m8 ; t7
      paddsw m6, m8 ; out6
      vpbroadcastd m8, [o(pw_2896x8)]
      psubw m3, m7, m5 ; t3
      paddw m7, m5 ; -out7
      psubw m5, m0, m2 ; t2
      paddw m0, m2 ; out0
      psubw m2, m1, m4 ; t6
      paddw m1, m4 ; -out1
      psubsw m3, m7, m5 ; t3
      paddsw m7, m5 ; -out7
      psubsw m5, m0, m2 ; t2
      paddsw m0, m2 ; out0
      psubsw m2, m1, m4 ; t6
      paddsw m1, m4 ; -out1
      psubw m4, m5, m3
      paddw m3, m5
      psubw m5, m2, m9
      ......@@ -2959,25 +2959,25 @@ ALIGN function_align
      mova [rsp+gprsize+32*0], m6 ; tmp3
      IDCT16_1D_ODDHALF 9, 3, 5, 7, 1, 11, 13, 14, 6, 10, 15
      mova m6, [rsp+gprsize+32*1] ; tmp5
      psubw m15, m0, m14 ; out15
      paddw m0, m14 ; out0
      psubw m14, m2, m13 ; out14
      paddw m2, m13 ; out1
      psubsw m15, m0, m14 ; out15
      paddsw m0, m14 ; out0
      psubsw m14, m2, m13 ; out14
      paddsw m2, m13 ; out1
      mova [rsp+gprsize+32*1], m2
      psubw m13, m4, m11 ; out13
      paddw m2, m4, m11 ; out2
      psubw m11, m8, m7 ; out11
      paddw m4, m8, m7 ; out4
      psubsw m13, m4, m11 ; out13
      paddsw m2, m4, m11 ; out2
      psubsw m11, m8, m7 ; out11
      paddsw m4, m8, m7 ; out4
      mova m7, [rsp+gprsize+32*2] ; tmp7
      psubw m10, m6, m5 ; out10
      paddw m5, m6 ; out5
      psubw m8, m7, m9 ; out8
      paddw m7, m9 ; out7
      psubw m9, m12, m3 ; out9
      paddw m6, m12, m3 ; out6
      psubsw m10, m6, m5 ; out10
      paddsw m5, m6 ; out5
      psubsw m8, m7, m9 ; out8
      paddsw m7, m9 ; out7
      psubsw m9, m12, m3 ; out9
      paddsw m6, m12, m3 ; out6
      mova m3, [rsp+gprsize+32*0] ; tmp3
      psubw m12, m3, m1 ; out12
      paddw m3, m1 ; out3
      psubsw m12, m3, m1 ; out12
      paddsw m3, m1 ; out3
      ret
      INV_TXFM_16X16_FN adst, dct
      ......@@ -3012,24 +3012,24 @@ ALIGN function_align
      ITX_MULSUB_2W 9, 6, 0, 4, 15, 2440, 3290 ; t7, t6
      ITX_MULSUB_2W 5, 10, 0, 4, 15, 3513, 2106 ; t11, t10
      ITX_MULSUB_2W 1, 14, 0, 4, 15, 4052, 601 ; t15, t14
      psubw m0, m2, m10 ; t10a
      paddw m2, m10 ; t2a
      psubw m10, m13, m5 ; t11a
      paddw m13, m5 ; t3a
      psubw m5, m6, m14 ; t14a
      paddw m6, m14 ; t6a
      psubw m14, m9, m1 ; t15a
      paddw m9, m1 ; t7a
      psubsw m0, m2, m10 ; t10a
      paddsw m2, m10 ; t2a
      psubsw m10, m13, m5 ; t11a
      paddsw m13, m5 ; t3a
      psubsw m5, m6, m14 ; t14a
      paddsw m6, m14 ; t6a
      psubsw m14, m9, m1 ; t15a
      paddsw m9, m1 ; t7a
      ITX_MULSUB_2W 0, 10, 1, 4, 15, 3406, 2276 ; t11, t10
      ITX_MULSUB_2W 14, 5, 1, 4, 15, 2276, 3406 ; t14, t15
      psubw m1, m10, m14 ; t14a
      paddw m10, m14 ; t10a
      psubw m14, m0, m5 ; t15a
      paddw m0, m5 ; t11a
      psubw m5, m2, m6 ; t6
      paddw m2, m6 ; t2
      psubw m6, m13, m9 ; t7
      paddw m13, m9 ; t3
      psubsw m1, m10, m14 ; t14a
      paddsw m10, m14 ; t10a
      psubsw m14, m0, m5 ; t15a
      paddsw m0, m5 ; t11a
      psubsw m5, m2, m6 ; t6
      paddsw m2, m6 ; t2
      psubsw m6, m13, m9 ; t7
      paddsw m13, m9 ; t3
      ITX_MULSUB_2W 6, 5, 4, 9, 15, 3784, 1567 ; t6a, t7a
      ITX_MULSUB_2W 14, 1, 4, 9, 15, 3784, 1567 ; t14, t15
      mova m9, [rsp+gprsize+32*0] ; in15
      ......@@ -3042,46 +3042,46 @@ ALIGN function_align
      ITX_MULSUB_2W 11, 6, 2, 10, 15, 1751, 3703 ; t5, t4
      ITX_MULSUB_2W 7, 8, 2, 10, 15, 3035, 2751 ; t9, t8
      ITX_MULSUB_2W 3, 12, 2, 10, 15, 3857, 1380 ; t13, t12
      psubw m10, m4, m8 ; t8a
      paddw m8, m4 ; t0a
      psubw m4, m9, m7 ; t9a
      paddw m9, m7 ; t1a
      psubw m7, m6, m12 ; t12a
      paddw m6, m12 ; t4a
      psubw m12, m11, m3 ; t13a
      paddw m11, m3 ; t5a
      psubsw m10, m4, m8 ; t8a
      paddsw m8, m4 ; t0a
      psubsw m4, m9, m7 ; t9a
      paddsw m9, m7 ; t1a
      psubsw m7, m6, m12 ; t12a
      paddsw m6, m12 ; t4a
      psubsw m12, m11, m3 ; t13a
      paddsw m11, m3 ; t5a
      ITX_MULSUB_2W 10, 4, 2, 3, 15, 799, 4017 ; t9, t8
      ITX_MULSUB_2W 12, 7, 2, 3, 15, 4017, 799 ; t12, t13
      psubw m3, m9, m11 ; t5
      paddw m9, m11 ; t1
      psubw m11, m4, m12 ; t12a
      paddw m4, m12 ; t8a
      paddw m12, m8, m6 ; t0
      psubw m8, m6 ; t4
      paddw m6, m10, m7 ; t9a
      psubw m10, m7 ; t13a
      psubsw m3, m9, m11 ; t5
      paddsw m9, m11 ; t1
      psubsw m11, m4, m12 ; t12a
      paddsw m4, m12 ; t8a
      paddsw m12, m8, m6 ; t0
      psubsw m8, m6 ; t4
      paddsw m6, m10, m7 ; t9a
      psubsw m10, m7 ; t13a
      ITX_MULSUB_2W 8, 3, 2, 7, 15, 1567, 3784 ; t5a, t4a
      ITX_MULSUB_2W 11, 10, 2, 7, 15, 1567, 3784 ; t13, t12
      mova m7, [rsp+gprsize+32*0] ; t10a
      mova m2, [rsp+gprsize+32*1] ; t6a
      paddw m15, m9, m13 ; -out15
      psubw m9, m13 ; t3a
      paddw m13, m11, m1 ; -out13
      psubw m11, m1 ; t15a
      psubw m1, m4, m7 ; t10
      paddw m7, m4 ; -out1
      psubw m4, m3, m2 ; t6
      paddw m3, m2 ; -out3
      paddw m2, m10, m14 ; out2
      psubw m10, m14 ; t14a
      paddw m14, m6, m0 ; out14
      psubw m6, m0 ; t11
      paddsw m15, m9, m13 ; -out15
      psubsw m9, m13 ; t3a
      paddsw m13, m11, m1 ; -out13
      psubsw m11, m1 ; t15a
      psubsw m1, m4, m7 ; t10
      paddsw m7, m4 ; -out1
      psubsw m4, m3, m2 ; t6
      paddsw m3, m2 ; -out3
      paddsw m2, m10, m14 ; out2
      psubsw m10, m14 ; t14a
      paddsw m14, m6, m0 ; out14
      psubsw m6, m0 ; t11
      mova m0, [rsp+gprsize+32*2] ; t2
      mova [rsp+gprsize+32*1], m7
      psubw m7, m12, m0 ; t2a
      paddw m0, m12 ; out0
      paddw m12, m8, m5 ; out12
      psubw m8, m5 ; t7
      psubsw m7, m12, m0 ; t2a
      paddsw m0, m12 ; out0
      paddsw m12, m8, m5 ; out12
      psubsw m8, m5 ; t7
      paddw m5, m10, m11 ; -out5
      psubw m10, m11 ; out10
      psubw m11, m4, m8 ; -out11
      ......@@ -3475,74 +3475,74 @@ ALIGN function_align
      ITX_MUL2X_PACK 13, 6, 12, 10, 2440, 3290, 3 ; t22a, t25a
      ITX_MUL2X_PACK 11, 6, 12, 10, 3035, 2751, 3 ; t17a, t30a
      .main2:
      psubw m6, m1, m11 ; t17 t30
      paddw m1, m11 ; t16 t31
      psubw m11, m9, m14 ; t18 t29
      paddw m9, m14 ; t19 t28
      psubw m14, m15, m0 ; t21 t26
      paddw m15, m0 ; t20 t27
      psubw m0, m8, m13 ; t22 t25
      paddw m8, m13 ; t23 t24
      psubsw m6, m1, m11 ; t17 t30
      paddsw m1, m11 ; t16 t31
      psubsw m11, m9, m14 ; t18 t29
      paddsw m9, m14 ; t19 t28
      psubsw m14, m15, m0 ; t21 t26
      paddsw m15, m0 ; t20 t27