[x265] [PATCH] asm: idct16 sse2 28900->25000 improvement over intrinsic
dtyx265 at gmail.com
dtyx265 at gmail.com
Fri Jan 16 20:10:39 CET 2015
# HG changeset patch
# User David T Yuen <dtyx265 at gmail.com>
# Date 1421435329 28800
# Node ID 592f6cd42f7ff15c9002db32b39b411d0adcd616
# Parent f9a54da187139a7331f8048365673787cbf14711
asm: idct16 sse2 28900->25000 improvement over intrinsic
This is based on the gcc assembler generated from a tweaked version of the idct16 intrinsic.
I can submit a patch of the intrinsic if anyone wants it.
diff -r f9a54da18713 -r 592f6cd42f7f source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Thu Jan 15 11:49:26 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp Fri Jan 16 11:08:49 2015 -0800
@@ -1337,6 +1337,7 @@
p.cu[BLOCK_4x4].idct = x265_idct4_sse2;
#if X86_64
p.cu[BLOCK_8x8].idct = x265_idct8_sse2;
+ p.cu[BLOCK_16x16].idct = x265_idct16_sse2;
#endif
p.idst4x4 = x265_idst4_sse2;
@@ -1555,6 +1556,7 @@
p.cu[BLOCK_4x4].idct = x265_idct4_sse2;
#if X86_64
p.cu[BLOCK_8x8].idct = x265_idct8_sse2;
+ p.cu[BLOCK_16x16].idct = x265_idct16_sse2;
#endif
p.idst4x4 = x265_idst4_sse2;
diff -r f9a54da18713 -r 592f6cd42f7f source/common/x86/dct8.asm
--- a/source/common/x86/dct8.asm Thu Jan 15 11:49:26 2015 +0530
+++ b/source/common/x86/dct8.asm Fri Jan 16 11:08:49 2015 -0800
@@ -179,6 +179,46 @@
dw 64, -75, 36, 18, -64, 89, -83, 50
dw 64, -89, 83, -75, 64, -50, 36, -18
+sse2_idct16_0: times 4 dw 90, 87
+ times 4 dw 80, 70
+ times 4 dw 57, 43
+ times 4 dw 25, 9
+
+sse2_idct16_1: times 4 dw 87, 57
+ times 4 dw 9, -43
+ times 4 dw -80, -90
+ times 4 dw -70, -25
+
+sse2_idct16_2: times 4 dw 80, 9
+ times 4 dw -70, -87
+ times 4 dw -25, 57
+ times 4 dw 90, 43
+
+sse2_idct16_3: times 4 dw 70, -43
+ times 4 dw -87, 9
+ times 4 dw 90, 25
+ times 4 dw -80, -57
+
+sse2_idct16_4: times 4 dw 57, -80
+ times 4 dw -25, 90
+ times 4 dw -9, -87
+ times 4 dw 43, 70
+
+sse2_idct16_5: times 4 dw 43, -90
+ times 4 dw 57, 25
+ times 4 dw -87, 70
+ times 4 dw 9, -80
+
+sse2_idct16_6: times 4 dw 25, -70
+ times 4 dw 90, -80
+ times 4 dw 43, 9
+ times 4 dw -57, 87
+
+sse2_idct16_7: times 4 dw 9, -25
+ times 4 dw 43, -57
+ times 4 dw 70, -80
+ times 4 dw 87, -90
+
idct16_shuff: dd 0, 4, 2, 6, 1, 5, 3, 7
idct16_shuff1: dd 2, 6, 0, 4, 3, 7, 1, 5
@@ -1350,6 +1390,2577 @@
%endif
;-------------------------------------------------------
+; void idct16(const int16_t* src, int16_t* dst, intptr_t dstStride)
+;-------------------------------------------------------
+%if ARCH_X86_64
+INIT_XMM sse2
+cglobal idct16, 3, 9, 16, 0-912
+%define IDCT_SHIFT1 7
+%if BIT_DEPTH == 10
+ %define IDCT_SHIFT2 10
+ %define IDCT_ADD pd_512
+%elif BIT_DEPTH == 8
+ %define IDCT_SHIFT2 12
+ %define IDCT_ADD pd_2048
+%else
+ %error Unsupported BIT_DEPTH!
+%endif
+
+ mova m9, [r0 + 2 * mmsize]
+ mova m0, [r0 + 6 * mmsize]
+ mova m14, m9
+ mova m11, [r0 + 10 * mmsize]
+ punpckhwd m9, m0
+ punpcklwd m14, m0
+ mova m7, [r0 + 18 * mmsize]
+ mova m15, m11
+ mova m0, [r0 + 14 * mmsize]
+ mova m12, m7
+ punpcklwd m15, m0
+ mova m8, [r0 + 26 * mmsize]
+ punpckhwd m11, m0
+ mova m0, [r0 + 22 * mmsize]
+ mova m13, m8
+ punpcklwd m12, m0
+ punpckhwd m7, m0
+ mova m0, [r0 + 30 * mmsize]
+ mova m6, [r0 + 8 * mmsize]
+ punpcklwd m13, m0
+ punpckhwd m8, m0
+ mova m0, [r0 + 4 * mmsize]
+ mova m2, m6
+ mova m1, [r0 + 12 * mmsize]
+ mova m4, m0
+ mova m10, [r0 + 24 * mmsize]
+ punpcklwd m4, m1
+ mova m3, [r0 + 20 * mmsize]
+ punpckhwd m0, m1
+ punpckhwd m6, m10
+ mova m1, [r0 + 28 * mmsize]
+ punpcklwd m2, m10
+ mova m5, m3
+ mova m10, [r0 + 0 * mmsize]
+ punpckhwd m3, m1
+ punpcklwd m5, m1
+ mova m1, m10
+ punpckhwd m10, [r0 + 16 * mmsize]
+ punpcklwd m1, [r0 + 16 * mmsize]
+ mova [rsp + 12 * mmsize], m10
+ mova [rsp + 11 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 0 * mmsize]
+ mova m1, [sse2_idct16_0 + 1 * mmsize]
+ pmaddwd m10, m14
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 7 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 3 * mmsize]
+ mova m1, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_1 + 0 * mmsize]
+ paddd m1, [rsp + 7 * mmsize]
+ mova [rsp + 7 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 6 * mmsize], m1
+ mova m10, [sse2_idct16_1 + 3 * mmsize]
+ mova m1, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ paddd m1, [rsp + 6 * mmsize]
+ mova [rsp + 6 * mmsize], m1
+ mova m1, [sse2_idct16_2 + 1 * mmsize]
+ mova m10, [sse2_idct16_2 + 0 * mmsize]
+ pmaddwd m1, m15
+ pmaddwd m10, m14
+ paddd m1, m10
+ mova [rsp + 5 * mmsize], m1
+ mova m10, [sse2_idct16_2 + 3 * mmsize]
+ mova m1, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_3 + 0 * mmsize]
+ paddd m1, [rsp + 5 * mmsize]
+ mova [rsp + 5 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 4 * mmsize], m1
+ mova m10, [sse2_idct16_3 + 3 * mmsize]
+ mova m1, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_4 + 0 * mmsize]
+ paddd m1, [rsp + 4 * mmsize]
+ mova [rsp + 4 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_4 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 2 * mmsize], m1
+ mova m10, [sse2_idct16_4 + 3 * mmsize]
+ mova m1, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_5 + 0 * mmsize]
+ paddd m1, [rsp + 2 * mmsize]
+ mova [rsp + 2 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 1 * mmsize], m1
+ mova m10, [sse2_idct16_5 + 3 * mmsize]
+ mova m1, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_6 + 0 * mmsize]
+ paddd m1, [rsp + 1 * mmsize]
+ mova [rsp + 1 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 0 * mmsize], m1
+ mova m10, [sse2_idct16_6 + 3 * mmsize]
+ mova m1, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ paddd m1, [rsp + 0 * mmsize]
+ mova [rsp + 0 * mmsize], m1
+ pmaddwd m14, [sse2_idct16_7 + 0 * mmsize]
+ pmaddwd m15, [sse2_idct16_7 + 1 * mmsize]
+ paddd m15, m14
+ pmaddwd m13, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m12, [sse2_idct16_7 + 2 * mmsize]
+ paddd m12, m13
+ paddd m12, m15
+ mova [rsp + 8 * mmsize], m12
+ mova m1, [sse2_idct16_0 + 1 * mmsize]
+ mova m14, [sse2_idct16_0 + 0 * mmsize]
+ pmaddwd m1, m11
+ pmaddwd m14, m9
+ paddd m1, m14
+ mova [rsp + 13 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 3 * mmsize]
+ mova m1, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ mova m15, [sse2_idct16_1 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 13 * mmsize]
+ mova [rsp + 13 * mmsize], m1
+ pmaddwd m15, m9
+ mova m1, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m15
+ mova [rsp + 14 * mmsize], m1
+ mova m10, [sse2_idct16_1 + 3 * mmsize]
+ mova m1, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ paddd m1, m10
+ mova m10, [sse2_idct16_2 + 0 * mmsize]
+ paddd m1, [rsp + 14 * mmsize]
+ mova [rsp + 14 * mmsize], m1
+ pmaddwd m10, m9
+ mova m1, [sse2_idct16_2 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m10
+ mova [rsp + 15 * mmsize], m1
+ mova m10, [sse2_idct16_2 + 3 * mmsize]
+ mova m1, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ mova m14, [sse2_idct16_3 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 15 * mmsize]
+ mova [rsp + 15 * mmsize], m1
+ pmaddwd m14, m9
+ mova m1, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m14
+ mova [rsp + 16 * mmsize], m1
+ mova m10, [sse2_idct16_3 + 3 * mmsize]
+ mova m1, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ paddd m1, m10
+ paddd m1, [rsp + 16 * mmsize]
+ mova [rsp + 16 * mmsize], m1
+ mova m1, [sse2_idct16_4 + 1 * mmsize]
+ mova m15, [sse2_idct16_4 + 0 * mmsize]
+ pmaddwd m1, m11
+ pmaddwd m15, m9
+ paddd m1, m15
+ mova [rsp + 17 * mmsize], m1
+ mova m10, [sse2_idct16_4 + 3 * mmsize]
+ mova m1, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ paddd m1, m10
+ mova m10, [sse2_idct16_5 + 0 * mmsize]
+ paddd m1, [rsp + 17 * mmsize]
+ mova [rsp + 17 * mmsize], m1
+ pmaddwd m10, m9
+ mova m1, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m10
+ mova [rsp + 18 * mmsize], m1
+ mova m10, [sse2_idct16_5 + 3 * mmsize]
+ mova m1, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m1, m7
+ mova m14, [sse2_idct16_6 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 18 * mmsize]
+ mova [rsp + 18 * mmsize], m1
+ pmaddwd m14, m9
+ pmaddwd m9, [sse2_idct16_7 + 0 * mmsize]
+ mova m1, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m14
+ mova [rsp + 19 * mmsize], m1
+ pmaddwd m11, [sse2_idct16_7 + 1 * mmsize]
+ paddd m11, m9
+ mova m10, [sse2_idct16_6 + 3 * mmsize]
+ mova m1, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m10, m8
+ pmaddwd m8, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m1, m7
+ mova m15, [tab_idct8_3 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 19 * mmsize]
+ pmaddwd m7, [sse2_idct16_7 + 2 * mmsize]
+ paddd m7, m8
+ mova [rsp + 19 * mmsize], m1
+ pmaddwd m15, m4
+ paddd m7, m11
+ mova [rsp + 20 * mmsize], m7
+ mova m1, [tab_idct8_3 + 1 * mmsize]
+ mova m14, [tab_idct8_3 + 0 * mmsize]
+ pmaddwd m1, m5
+ paddd m15, m1
+ mova m1, [tab_idct8_3 + 1 * mmsize]
+ pmaddwd m14, m0
+ pmaddwd m1, m3
+ mova m13, [tab_idct8_3 + 2 * mmsize]
+ paddd m14, m1
+ mova m1, [tab_idct8_3 + 3 * mmsize]
+ pmaddwd m13, m4
+ pmaddwd m1, m5
+ mova m12, [tab_idct8_3 + 2 * mmsize]
+ paddd m13, m1
+ mova m1, [tab_idct8_3 + 3 * mmsize]
+ pmaddwd m12, m0
+ pmaddwd m1, m3
+ mova m7, [tab_idct8_3 + 4 * mmsize]
+ paddd m12, m1
+ mova m1, [tab_idct8_3 + 5 * mmsize]
+ pmaddwd m7, m4
+ pmaddwd m1, m5
+ paddd m1, m7
+ mova [rsp + 24 * mmsize], m1
+ pmaddwd m4, [tab_idct8_3 + 6 * mmsize]
+ pmaddwd m5, [tab_idct8_3 + 7 * mmsize]
+ paddd m5, m4
+ mova [rsp + 26 * mmsize], m5
+ mova m9, [rsp + 11 * mmsize]
+ mova m1, [tab_idct8_3 + 5 * mmsize]
+ pmaddwd m9, [tab_dct4 + 0 * mmsize]
+ mova m5, m9
+ mova m8, [tab_idct8_3 + 4 * mmsize]
+ pmaddwd m1, m3
+ pmaddwd m3, [tab_idct8_3 + 7 * mmsize]
+ mova m11, [rsp + 11 * mmsize]
+ pmaddwd m8, m0
+ pmaddwd m0, [tab_idct8_3 + 6 * mmsize]
+ paddd m1, m8
+ paddd m0, m3
+ mova m8, [rsp + 12 * mmsize]
+ pmaddwd m11, [tab_dct4 + 2 * mmsize]
+ mova m7, m11
+ mova m10, [rsp + 12 * mmsize]
+ pmaddwd m8, [tab_dct4 + 0 * mmsize]
+ mova [rsp + 25 * mmsize], m1
+ pmaddwd m10, [tab_dct4 + 2 * mmsize]
+ mova m4, [tab_dct4 + 1 * mmsize]
+ mova m3, [tab_dct4 + 1 * mmsize]
+ pmaddwd m4, m2
+ paddd m5, m4
+ pmaddwd m2, [tab_dct4 + 3 * mmsize]
+ paddd m7, m2
+ psubd m9, m4
+ mova m1, [tab_dct4 + 3 * mmsize]
+ pmaddwd m3, m6
+ psubd m11, m2
+ pmaddwd m1, m6
+ mova [rsp + 11 * mmsize], m5
+ mova m6, m8
+ psubd m8, m3
+ mova m5, m10
+ psubd m10, m1
+ paddd m6, m3
+ mova [rsp + 28 * mmsize], m7
+ paddd m5, m1
+ mova [rsp + 12 * mmsize], m6
+ paddd m6, m14
+ mova [rsp + 31 * mmsize], m5
+ mova [rsp + 3 * mmsize], m8
+ paddd m8, m0
+ mova [rsp + 27 * mmsize], m8
+ mova m7, [rsp + 11 * mmsize]
+ mova m8, [rsp + 11 * mmsize]
+ paddd m7, m15
+ psubd m8, m15
+ mova m15, m8
+ mova m8, [rsp + 12 * mmsize]
+ mova m4, [rsp + 31 * mmsize]
+ psubd m8, m14
+ mova [rsp + 33 * mmsize], m8
+ paddd m4, m12
+ mova m8, [rsp + 31 * mmsize]
+ mova m5, [rsp + 28 * mmsize]
+ psubd m8, m12
+ mova m12, m8
+ mova m14, [rsp + 28 * mmsize]
+ paddd m5, m13
+ mova m8, [rsp + 3 * mmsize]
+ psubd m14, m13
+ mova m13, m14
+ psubd m8, m0
+ mova m14, [rsp + 27 * mmsize]
+ mova m0, [pd_64]
+ paddd m14, m0
+ mova [rsp + 27 * mmsize], m14
+ paddd m7, m0
+ paddd m6, m0
+ paddd m5, m0
+ paddd m4, m0
+ paddd m8, m0
+ paddd m13, m0
+ paddd m12, m0
+ paddd m15, m0
+ mova m14, [rsp + 33 * mmsize]
+ paddd m14, m0
+ mova [rsp + 33 * mmsize], m14
+ mova m14, [rsp + 7 * mmsize]
+ paddd m14, m7
+ mova [rsp + 11 * mmsize], m14
+ mova m14, [rsp + 13 * mmsize]
+ paddd m14, m6
+ mova [rsp + 29 * mmsize], m14
+ mova m14, [rsp + 6 * mmsize]
+ paddd m14, m5
+ mova [rsp + 12 * mmsize], m14
+ mova m14, [rsp + 14 * mmsize]
+ mova m3, [rsp + 24 * mmsize]
+ paddd m14, m4
+ mova [rsp + 21 * mmsize], m14
+ paddd m3, m11
+ psubd m11, [rsp + 24 * mmsize]
+ paddd m3, m0
+ paddd m11, m0
+ mova m14, [rsp + 5 * mmsize]
+ mova m2, [rsp + 25 * mmsize]
+ paddd m14, m3
+ mova [rsp + 24 * mmsize], m14
+ paddd m2, m10
+ paddd m2, m0
+ psubd m10, [rsp + 25 * mmsize]
+ paddd m10, m0
+ mova m14, [rsp + 15 * mmsize]
+ mova m1, [rsp + 26 * mmsize]
+ paddd m14, m2
+ mova [rsp + 23 * mmsize], m14
+ paddd m1, m9
+ psubd m9, [rsp + 26 * mmsize]
+ paddd m1, m0
+ paddd m9, m0
+ mova m14, [rsp + 4 * mmsize]
+ paddd m14, m1
+ mova [rsp + 25 * mmsize], m14
+ psubd m7, [rsp + 7 * mmsize]
+ psubd m6, [rsp + 13 * mmsize]
+ psubd m5, [rsp + 6 * mmsize]
+ psubd m4, [rsp + 14 * mmsize]
+ mova [rsp + 7 * mmsize], m6
+ psubd m3, [rsp + 5 * mmsize]
+ psubd m2, [rsp + 15 * mmsize]
+ psubd m1, [rsp + 4 * mmsize]
+ psrad m5, IDCT_SHIFT1
+ psrad m7, IDCT_SHIFT1
+ mova [rsp + 6 * mmsize], m4
+ psrad m2, IDCT_SHIFT1
+ psrad m1, IDCT_SHIFT1
+ psrad m3, IDCT_SHIFT1
+ packssdw m3, m2
+ mova m14, [rsp + 27 * mmsize]
+ paddd m14, [rsp + 16 * mmsize]
+ mova [rsp + 3 * mmsize], m14
+ mova m14, [rsp + 2 * mmsize]
+ paddd m14, m9
+ mova [rsp + 26 * mmsize], m14
+ psubd m9, [rsp + 2 * mmsize]
+ psrad m9, IDCT_SHIFT1
+ mova m14, [rsp + 17 * mmsize]
+ paddd m14, m8
+ mova [rsp + 9 * mmsize], m14
+ psubd m8, [rsp + 17 * mmsize]
+ psrad m8, IDCT_SHIFT1
+ packssdw m9, m8
+ mova m14, [rsp + 1 * mmsize]
+ paddd m14, m11
+ mova [rsp + 28 * mmsize], m14
+ psubd m11, [rsp + 1 * mmsize]
+ psrad m11, IDCT_SHIFT1
+ mova m14, [rsp + 18 * mmsize]
+ paddd m14, m10
+ mova [rsp + 10 * mmsize], m14
+ psubd m10, [rsp + 18 * mmsize]
+ psrad m10, IDCT_SHIFT1
+ packssdw m11, m10
+ mova m14, [rsp + 0 * mmsize]
+ paddd m14, m13
+ mova [rsp + 31 * mmsize], m14
+ psubd m13, [rsp + 0 * mmsize]
+ psrad m13, IDCT_SHIFT1
+ mova m14, [rsp + 19 * mmsize]
+ paddd m14, m12
+ mova [rsp + 22 * mmsize], m14
+ psubd m12, [rsp + 19 * mmsize]
+ psrad m12, IDCT_SHIFT1
+ packssdw m13, m12
+ mova m14, [rsp + 8 * mmsize]
+ paddd m14, m15
+ mova [rsp + 32 * mmsize], m14
+ psubd m15, [rsp + 8 * mmsize]
+ psrad m15, IDCT_SHIFT1
+ mova m14, [rsp + 20 * mmsize]
+ mova m6, [rsp + 27 * mmsize]
+ paddd m14, [rsp + 33 * mmsize]
+ mova m4, [rsp + 11 * mmsize]
+ psubd m6, [rsp + 16 * mmsize]
+ psrad m6, IDCT_SHIFT1
+ packssdw m1, m6
+ mova [rsp + 30 * mmsize], m14
+ psrad m4, IDCT_SHIFT1
+ mova m14, [rsp + 33 * mmsize]
+ psubd m14, [rsp + 20 * mmsize]
+ mova [rsp + 11 * mmsize], m4
+ psrad m14, IDCT_SHIFT1
+ packssdw m15, m14
+ mova m4, [rsp + 29 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 29 * mmsize], m4
+ mova m4, [rsp + 12 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 12 * mmsize], m4
+ mova m4, [rsp + 21 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 21 * mmsize], m4
+ mova m4, [rsp + 24 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 24 * mmsize], m4
+ mova m4, [rsp + 23 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 23 * mmsize], m4
+ mova m4, [rsp + 25 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 25 * mmsize], m4
+ mova m4, [rsp + 3 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 3 * mmsize], m4
+ mova m4, [rsp + 26 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 26 * mmsize], m4
+ mova m4, [rsp + 9 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 9 * mmsize], m4
+ mova m4, [rsp + 28 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 28 * mmsize], m4
+ mova m4, [rsp + 10 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 10 * mmsize], m4
+ mova m4, [rsp + 31 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 31 * mmsize], m4
+ mova m4, [rsp + 22 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 22 * mmsize], m4
+ mova m4, [rsp + 32 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 32 * mmsize], m4
+ mova m4, [rsp + 30 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 30 * mmsize], m4
+ mova m4, [rsp + 6 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 6 * mmsize], m4
+ mova m4, [rsp + 7 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 7 * mmsize], m4
+ packssdw m5, [rsp + 6 * mmsize]
+ mova [rsp + 41 * mmsize], m15
+ mova [rsp + 42 * mmsize], m13
+ mova [rsp + 43 * mmsize], m11
+ mova [rsp + 44 * mmsize], m9
+ mova [rsp + 45 * mmsize], m1
+ mova [rsp + 46 * mmsize], m3
+ mova [rsp + 47 * mmsize], m5
+ mova m4, [rsp + 11 * mmsize]
+ packssdw m7, [rsp + 7 * mmsize]
+ packssdw m4, [rsp + 29 * mmsize]
+ mova [rsp + 11 * mmsize], m4
+ mova m4, [rsp + 12 * mmsize]
+ packssdw m4, [rsp + 21 * mmsize]
+ mova [rsp + 12 * mmsize], m4
+ mova m4, [rsp + 24 * mmsize]
+ packssdw m4, [rsp + 23 * mmsize]
+ mova [rsp + 24 * mmsize], m4
+ mova m4, [rsp + 25 * mmsize]
+ packssdw m4, [rsp + 3 * mmsize]
+ mova [rsp + 25 * mmsize], m4
+ mova m4, [rsp + 26 * mmsize]
+ packssdw m4, [rsp + 9 * mmsize]
+ mova [rsp + 26 * mmsize], m4
+ mova m4, [rsp + 28 * mmsize]
+ packssdw m4, [rsp + 10 * mmsize]
+ mova [rsp + 28 * mmsize], m4
+ mova m4, [rsp + 31 * mmsize]
+ packssdw m4, [rsp + 22 * mmsize]
+ mova [rsp + 31 * mmsize], m4
+ mova m4, [rsp + 32 * mmsize]
+ packssdw m4, [rsp + 30 * mmsize]
+ mova [rsp + 32 * mmsize], m4
+ mova [rsp + 48 * mmsize], m7
+ mova m7, [r0 + 3 * mmsize]
+ mova m1, [r0 + 7 * mmsize]
+ mova m14, m7
+ mova m11, [r0 + 11 * mmsize]
+ punpckhwd m7, m1
+ punpcklwd m14, m1
+ mova m3, [r0 + 19 * mmsize]
+ mova m15, m11
+ mova m1, [r0 + 15 * mmsize]
+ mova m12, m3
+ punpcklwd m15, m1
+ mova m4, [r0 + 27 * mmsize]
+ punpckhwd m11, m1
+ mova m1, [r0 + 23 * mmsize]
+ mova m13, m4
+ punpcklwd m12, m1
+ punpckhwd m3, m1
+ mova m5, [r0 + 5 * mmsize]
+ mova m1, [r0 + 31 * mmsize]
+ mova m6, m5
+ punpcklwd m13, m1
+ mova m8, [r0 + 21 * mmsize]
+ punpckhwd m4, m1
+ mova m1, [r0 + 13 * mmsize]
+ mova m9, m8
+ punpcklwd m6, m1
+ punpckhwd m5, m1
+ mova m1, [r0 + 29 * mmsize]
+ punpcklwd m9, m1
+ punpckhwd m8, m1
+ mova m1, [r0 + 9 * mmsize]
+ mova m10, [r0 + 25 * mmsize]
+ mova m2, m1
+ punpckhwd m1, m10
+ punpcklwd m2, m10
+ mova m10, [r0 + 17 * mmsize]
+ mova [rsp + 5 * mmsize], m10
+ mova [rsp + 7 * mmsize], m1
+ mova m10, [r0 + 1 * mmsize]
+ mova m1, m10
+ punpckhwd m10, [rsp + 5 * mmsize]
+ punpcklwd m1, [rsp + 5 * mmsize]
+ mova [rsp + 21 * mmsize], m10
+ mova [rsp + 6 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 0 * mmsize]
+ mova m1, [sse2_idct16_0 + 1 * mmsize]
+ pmaddwd m10, m14
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 5 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 3 * mmsize]
+ mova m1, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_1 + 0 * mmsize]
+ paddd m1, [rsp + 5 * mmsize]
+ mova [rsp + 5 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 4 * mmsize], m1
+ mova m10, [sse2_idct16_1 + 3 * mmsize]
+ mova m1, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ paddd m1, [rsp + 4 * mmsize]
+ mova [rsp + 4 * mmsize], m1
+ mova m1, [sse2_idct16_2 + 1 * mmsize]
+ mova m10, [sse2_idct16_2 + 0 * mmsize]
+ pmaddwd m1, m15
+ pmaddwd m10, m14
+ paddd m1, m10
+ mova [rsp + 2 * mmsize], m1
+ mova m10, [sse2_idct16_2 + 3 * mmsize]
+ mova m1, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_3 + 0 * mmsize]
+ paddd m1, [rsp + 2 * mmsize]
+ mova [rsp + 2 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 1 * mmsize], m1
+ mova m10, [sse2_idct16_3 + 3 * mmsize]
+ mova m1, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_4 + 0 * mmsize]
+ paddd m1, [rsp + 1 * mmsize]
+ mova [rsp + 1 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_4 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 0 * mmsize], m1
+ mova m10, [sse2_idct16_4 + 3 * mmsize]
+ mova m1, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_5 + 0 * mmsize]
+ paddd m1, [rsp + 0 * mmsize]
+ mova [rsp + 0 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 8 * mmsize], m1
+ mova m10, [sse2_idct16_5 + 3 * mmsize]
+ mova m1, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ mova m10, [sse2_idct16_6 + 0 * mmsize]
+ paddd m1, [rsp + 8 * mmsize]
+ mova [rsp + 8 * mmsize], m1
+ pmaddwd m10, m14
+ mova m1, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m1, m15
+ paddd m1, m10
+ mova [rsp + 13 * mmsize], m1
+ mova m10, [sse2_idct16_6 + 3 * mmsize]
+ mova m1, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m10, m13
+ pmaddwd m1, m12
+ paddd m1, m10
+ paddd m1, [rsp + 13 * mmsize]
+ mova [rsp + 13 * mmsize], m1
+ pmaddwd m14, [sse2_idct16_7 + 0 * mmsize]
+ pmaddwd m15, [sse2_idct16_7 + 1 * mmsize]
+ paddd m15, m14
+ pmaddwd m13, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m12, [sse2_idct16_7 + 2 * mmsize]
+ paddd m12, m13
+ paddd m12, m15
+ mova [rsp + 14 * mmsize], m12
+ mova m1, [sse2_idct16_0 + 1 * mmsize]
+ mova m14, [sse2_idct16_0 + 0 * mmsize]
+ pmaddwd m1, m11
+ pmaddwd m14, m7
+ paddd m1, m14
+ mova [rsp + 15 * mmsize], m1
+ mova m10, [sse2_idct16_0 + 3 * mmsize]
+ mova m1, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ mova m15, [sse2_idct16_1 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 15 * mmsize]
+ mova [rsp + 15 * mmsize], m1
+ pmaddwd m15, m7
+ mova m1, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m15
+ mova [rsp + 16 * mmsize], m1
+ mova m10, [sse2_idct16_1 + 3 * mmsize]
+ mova m1, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ paddd m1, m10
+ mova m10, [sse2_idct16_2 + 0 * mmsize]
+ paddd m1, [rsp + 16 * mmsize]
+ mova [rsp + 16 * mmsize], m1
+ pmaddwd m10, m7
+ mova m1, [sse2_idct16_2 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m10
+ mova [rsp + 17 * mmsize], m1
+ mova m10, [sse2_idct16_2 + 3 * mmsize]
+ mova m1, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ mova m14, [sse2_idct16_3 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 17 * mmsize]
+ mova [rsp + 17 * mmsize], m1
+ pmaddwd m14, m7
+ mova m1, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m14
+ mova [rsp + 18 * mmsize], m1
+ mova m10, [sse2_idct16_3 + 3 * mmsize]
+ mova m1, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ paddd m1, m10
+ paddd m1, [rsp + 18 * mmsize]
+ mova [rsp + 18 * mmsize], m1
+ mova m1, [sse2_idct16_4 + 1 * mmsize]
+ mova m15, [sse2_idct16_4 + 0 * mmsize]
+ pmaddwd m1, m11
+ pmaddwd m15, m7
+ paddd m1, m15
+ mova [rsp + 19 * mmsize], m1
+ mova m10, [sse2_idct16_4 + 3 * mmsize]
+ mova m1, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ paddd m1, m10
+ mova m10, [sse2_idct16_5 + 0 * mmsize]
+ paddd m1, [rsp + 19 * mmsize]
+ mova [rsp + 19 * mmsize], m1
+ pmaddwd m10, m7
+ mova m1, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m10
+ mova [rsp + 20 * mmsize], m1
+ mova m10, [sse2_idct16_5 + 3 * mmsize]
+ mova m1, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m1, m3
+ mova m14, [sse2_idct16_6 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 20 * mmsize]
+ mova [rsp + 20 * mmsize], m1
+ pmaddwd m14, m7
+ pmaddwd m7, [sse2_idct16_7 + 0 * mmsize]
+ mova m1, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m14
+ mova [rsp + 27 * mmsize], m1
+ pmaddwd m11, [sse2_idct16_7 + 1 * mmsize]
+ paddd m11, m7
+ mova m10, [sse2_idct16_6 + 3 * mmsize]
+ mova m1, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m10, m4
+ pmaddwd m4, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m1, m3
+ mova m15, [tab_idct8_3 + 0 * mmsize]
+ paddd m1, m10
+ paddd m1, [rsp + 27 * mmsize]
+ pmaddwd m3, [sse2_idct16_7 + 2 * mmsize]
+ paddd m3, m4
+ mova [rsp + 27 * mmsize], m1
+ pmaddwd m15, m6
+ paddd m3, m11
+ mova [rsp + 29 * mmsize], m3
+ mova m1, [tab_idct8_3 + 1 * mmsize]
+ mova m14, [tab_idct8_3 + 0 * mmsize]
+ pmaddwd m1, m9
+ paddd m15, m1
+ mova m1, [tab_idct8_3 + 1 * mmsize]
+ pmaddwd m14, m5
+ pmaddwd m1, m8
+ mova m13, [tab_idct8_3 + 2 * mmsize]
+ paddd m14, m1
+ mova m1, [tab_idct8_3 + 3 * mmsize]
+ pmaddwd m13, m6
+ pmaddwd m1, m9
+ mova m12, [tab_idct8_3 + 2 * mmsize]
+ paddd m13, m1
+ mova m1, [tab_idct8_3 + 3 * mmsize]
+ pmaddwd m12, m5
+ pmaddwd m1, m8
+ mova m3, [tab_idct8_3 + 4 * mmsize]
+ paddd m12, m1
+ mova m1, [tab_idct8_3 + 5 * mmsize]
+ pmaddwd m3, m6
+ pmaddwd m1, m9
+ paddd m1, m3
+ mova [rsp + 23 * mmsize], m1
+ pmaddwd m6, [tab_idct8_3 + 6 * mmsize]
+ pmaddwd m9, [tab_idct8_3 + 7 * mmsize]
+ paddd m9, m6
+ mova [rsp + 3 * mmsize], m9
+ mova m1, [tab_idct8_3 + 5 * mmsize]
+ mova m9, [rsp + 6 * mmsize]
+ pmaddwd m1, m8
+ pmaddwd m8, [tab_idct8_3 + 7 * mmsize]
+ mova m7, [tab_idct8_3 + 4 * mmsize]
+ pmaddwd m9, [tab_dct4 + 0 * mmsize]
+ mova m6, m9
+ pmaddwd m7, m5
+ mova m4, [tab_dct4 + 1 * mmsize]
+ pmaddwd m5, [tab_idct8_3 + 6 * mmsize]
+ paddd m5, m8
+ paddd m7, m1
+ mova m8, [rsp + 21 * mmsize]
+ pmaddwd m4, m2
+ paddd m6, m4
+ pmaddwd m2, [tab_dct4 + 3 * mmsize]
+ psubd m9, m4
+ mova m10, [rsp + 21 * mmsize]
+ pmaddwd m8, [tab_dct4 + 0 * mmsize]
+ mova m3, [rsp + 7 * mmsize]
+ pmaddwd m10, [tab_dct4 + 2 * mmsize]
+ mova m1, [rsp + 7 * mmsize]
+ pmaddwd m3, [tab_dct4 + 1 * mmsize]
+ mova m11, [rsp + 6 * mmsize]
+ pmaddwd m1, [tab_dct4 + 3 * mmsize]
+ mova [rsp + 21 * mmsize], m6
+ pmaddwd m11, [tab_dct4 + 2 * mmsize]
+ mova m6, m8
+ psubd m8, m3
+ paddd m6, m3
+ mova [rsp + 9 * mmsize], m6
+ mova m6, m11
+ psubd m11, m2
+ mova [rsp + 33 * mmsize], m8
+ paddd m6, m2
+ mova [rsp + 10 * mmsize], m6
+ mova m6, m10
+ psubd m10, m1
+ paddd m6, m1
+ mova [rsp + 22 * mmsize], m6
+ mova m1, m6
+ mova m6, m10
+ psubd m10, m7
+ paddd m10, m0
+ paddd m1, m12
+ paddd m1, m0
+ mova m8, [rsp + 23 * mmsize]
+ paddd m6, m7
+ paddd m8, m11
+ mova [rsp + 30 * mmsize], m8
+ psubd m11, [rsp + 23 * mmsize]
+ paddd m11, m0
+ mova [rsp + 6 * mmsize], m6
+ mova m8, [rsp + 33 * mmsize]
+ mova m4, [rsp + 21 * mmsize]
+ paddd m8, m5
+ mova [rsp + 7 * mmsize], m8
+ paddd m4, m15
+ paddd m4, m0
+ mova m8, [rsp + 21 * mmsize]
+ mova m3, [rsp + 9 * mmsize]
+ psubd m8, m15
+ mova m15, m8
+ mova m8, [rsp + 9 * mmsize]
+ paddd m3, m14
+ paddd m3, m0
+ paddd m15, m0
+ mova m2, [rsp + 10 * mmsize]
+ psubd m8, m14
+ mova m14, m8
+ mova m8, [rsp + 10 * mmsize]
+ paddd m2, m13
+ paddd m2, m0
+ paddd m14, m0
+ psubd m8, m13
+ mova m13, m8
+ mova m8, [rsp + 22 * mmsize]
+ paddd m13, m0
+ psubd m8, m12
+ mova m12, m8
+ mova m8, [rsp + 33 * mmsize]
+ paddd m12, m0
+ psubd m8, m5
+ mova m5, [rsp + 30 * mmsize]
+ paddd m8, m0
+ mova m7, [rsp + 6 * mmsize]
+ paddd m5, m0
+ mova m6, [rsp + 3 * mmsize]
+ paddd m7, m0
+ mova [rsp + 30 * mmsize], m5
+ paddd m6, m9
+ psubd m9, [rsp + 3 * mmsize]
+ paddd m6, m0
+ paddd m9, m0
+ mova [rsp + 6 * mmsize], m7
+ mova m5, [rsp + 7 * mmsize]
+ mova m7, [rsp + 5 * mmsize]
+ paddd m5, m0
+ mova m0, [rsp + 15 * mmsize]
+ paddd m7, m4
+ mova [rsp + 7 * mmsize], m5
+ paddd m0, m3
+ mova [rsp + 21 * mmsize], m7
+ mova [rsp + 33 * mmsize], m0
+ mova m5, [rsp + 4 * mmsize]
+ mova m7, [rsp + 16 * mmsize]
+ paddd m5, m2
+ mova m0, [rsp + 30 * mmsize]
+ paddd m7, m1
+ mova [rsp + 23 * mmsize], m5
+ paddd m0, [rsp + 2 * mmsize]
+ mova [rsp + 34 * mmsize], m7
+ mova [rsp + 3 * mmsize], m0
+ psubd m3, [rsp + 15 * mmsize]
+ psubd m4, [rsp + 5 * mmsize]
+ psubd m2, [rsp + 4 * mmsize]
+ psubd m1, [rsp + 16 * mmsize]
+ psrad m4, IDCT_SHIFT1
+ mova [rsp + 15 * mmsize], m3
+ psrad m2, IDCT_SHIFT1
+ psrad m1, IDCT_SHIFT1
+ mova m0, [rsp + 7 * mmsize]
+ paddd m0, [rsp + 18 * mmsize]
+ mova [rsp + 36 * mmsize], m0
+ mova m0, [rsp + 8 * mmsize]
+ mova m5, [rsp + 6 * mmsize]
+ paddd m0, m11
+ mova [rsp + 22 * mmsize], m0
+ paddd m5, [rsp + 17 * mmsize]
+ mova [rsp + 35 * mmsize], m5
+ mova m0, [rsp + 27 * mmsize]
+ mova m3, [rsp + 6 * mmsize]
+ paddd m0, m12
+ mova m7, [rsp + 1 * mmsize]
+ psubd m3, [rsp + 17 * mmsize]
+ mova m5, [rsp + 0 * mmsize]
+ paddd m7, m6
+ psubd m6, [rsp + 1 * mmsize]
+ psrad m6, IDCT_SHIFT1
+ mova [rsp + 39 * mmsize], m0
+ paddd m5, m9
+ mova [rsp + 9 * mmsize], m7
+ mova [rsp + 10 * mmsize], m5
+ mova [rsp + 6 * mmsize], m3
+ mova m0, [rsp + 29 * mmsize]
+ mova m7, [rsp + 19 * mmsize]
+ paddd m0, m14
+ mova m5, [rsp + 20 * mmsize]
+ paddd m7, m8
+ mova [rsp + 40 * mmsize], m0
+ paddd m5, m10
+ mova [rsp + 37 * mmsize], m7
+ mova [rsp + 38 * mmsize], m5
+ mova m0, [rsp + 30 * mmsize]
+ mova m3, [rsp + 7 * mmsize]
+ psubd m0, [rsp + 2 * mmsize]
+ psrad m0, IDCT_SHIFT1
+ mova m7, [rsp + 13 * mmsize]
+ psubd m3, [rsp + 18 * mmsize]
+ mova m5, [rsp + 14 * mmsize]
+ paddd m7, m13
+ psrad m7, IDCT_SHIFT1
+ mova [rsp + 7 * mmsize], m3
+ paddd m5, m15
+ psubd m9, [rsp + 0 * mmsize]
+ psubd m8, [rsp + 19 * mmsize]
+ psubd m11, [rsp + 8 * mmsize]
+ psubd m10, [rsp + 20 * mmsize]
+ psubd m13, [rsp + 13 * mmsize]
+ psubd m12, [rsp + 27 * mmsize]
+ psubd m15, [rsp + 14 * mmsize]
+ psubd m14, [rsp + 29 * mmsize]
+ psrad m5, IDCT_SHIFT1
+ psrad m14, IDCT_SHIFT1
+ mova m3, [rsp + 21 * mmsize]
+ psrad m12, IDCT_SHIFT1
+ psrad m10, IDCT_SHIFT1
+ psrad m8, IDCT_SHIFT1
+ psrad m15, IDCT_SHIFT1
+ psrad m13, IDCT_SHIFT1
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 21 * mmsize], m3
+ psrad m11, IDCT_SHIFT1
+ packssdw m15, m14
+ psrad m9, IDCT_SHIFT1
+ packssdw m13, m12
+ packssdw m11, m10
+ mova m10, m2
+ mova m3, [rsp + 33 * mmsize]
+ packssdw m9, m8
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 33 * mmsize], m3
+ packssdw m10, m1
+ mova m3, [rsp + 23 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 23 * mmsize], m3
+ mova m3, [rsp + 34 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 34 * mmsize], m3
+ mova m3, [rsp + 3 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 3 * mmsize], m3
+ mova m3, [rsp + 35 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 35 * mmsize], m3
+ mova m3, [rsp + 9 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 9 * mmsize], m3
+ mova m3, [rsp + 36 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 36 * mmsize], m3
+ mova m3, [rsp + 10 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 10 * mmsize], m3
+ mova m3, [rsp + 37 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 37 * mmsize], m3
+ mova m3, [rsp + 22 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 22 * mmsize], m3
+ mova m3, [rsp + 38 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 38 * mmsize], m3
+ mova m3, [rsp + 39 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 39 * mmsize], m3
+ packssdw m7, [rsp + 39 * mmsize]
+ mova [rsp + 2 * mmsize], m15
+ mova [rsp + 1 * mmsize], m13
+ mova [rsp + 5 * mmsize], m7
+ mova [rsp + 0 * mmsize], m11
+ mova m3, [rsp + 40 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 40 * mmsize], m3
+ mova m3, [rsp + 7 * mmsize]
+ packssdw m5, [rsp + 40 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 7 * mmsize], m3
+ mova m3, [rsp + 6 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 6 * mmsize], m3
+ mova [rsp + 4 * mmsize], m5
+ mova m3, [rsp + 15 * mmsize]
+ psrad m3, IDCT_SHIFT1
+ mova [rsp + 15 * mmsize], m3
+ mova m3, [rsp + 21 * mmsize]
+ packssdw m3, [rsp + 33 * mmsize]
+ mova [rsp + 21 * mmsize], m3
+ mova m3, [rsp + 23 * mmsize]
+ packssdw m3, [rsp + 34 * mmsize]
+ mova [rsp + 23 * mmsize], m3
+ mova m3, [rsp + 3 * mmsize]
+ packssdw m3, [rsp + 35 * mmsize]
+ mova [rsp + 3 * mmsize], m3
+ mova m3, [rsp + 9 * mmsize]
+ packssdw m3, [rsp + 36 * mmsize]
+ mova [rsp + 9 * mmsize], m3
+ mova m3, [rsp + 10 * mmsize]
+ packssdw m3, [rsp + 37 * mmsize]
+ mova [rsp + 10 * mmsize], m3
+ mova m3, [rsp + 22 * mmsize]
+ packssdw m3, [rsp + 38 * mmsize]
+ mova [rsp + 22 * mmsize], m3
+ mova [rsp + 8 * mmsize], m9
+ packssdw m0, [rsp + 6 * mmsize]
+ mova m1, [rsp + 24 * mmsize]
+ packssdw m4, [rsp + 15 * mmsize]
+ mova m7, [rsp + 26 * mmsize]
+ packssdw m6, [rsp + 7 * mmsize]
+ mova [rsp + 6 * mmsize], m0
+ punpckhwd m1, [rsp + 25 * mmsize]
+ mova [rsp + 13 * mmsize], m4
+ mova [rsp + 7 * mmsize], m6
+ mova m0, [rsp + 11 * mmsize]
+ punpckhwd m7, [rsp + 28 * mmsize]
+ mova m3, [rsp + 11 * mmsize]
+ punpckhwd m0, [rsp + 12 * mmsize]
+ mova m4, [rsp + 24 * mmsize]
+ punpcklwd m3, [rsp + 12 * mmsize]
+ mova m6, [rsp + 31 * mmsize]
+ punpcklwd m4, [rsp + 25 * mmsize]
+ mova m9, [rsp + 26 * mmsize]
+ mova m2, m0
+ punpckhwd m6, [rsp + 32 * mmsize]
+ mova m8, [rsp + 31 * mmsize]
+ mova m5, m3
+ punpcklwd m9, [rsp + 28 * mmsize]
+ mova m11, [rsp + 43 * mmsize]
+ punpcklwd m8, [rsp + 32 * mmsize]
+ punpckldq m2, m1
+ punpckhdq m0, m1
+ mova m1, m7
+ punpckhdq m3, m4
+ punpckldq m1, m6
+ mova m14, m2
+ punpckldq m5, m4
+ mova m4, m9
+ punpckhdq m7, m6
+ punpckhdq m9, m8
+ punpckldq m4, m8
+ mova m8, m3
+ punpckhqdq m2, m1
+ punpcklqdq m8, m9
+ mova m15, m0
+ punpckhqdq m0, m7
+ mova [rsp + 19 * mmsize], m2
+ punpcklqdq m15, m7
+ mova [rsp + 16 * mmsize], m8
+ mova m6, m5
+ mova [rsp + 11 * mmsize], m0
+ punpckhqdq m5, m4
+ punpcklqdq m14, m1
+ mova [rsp + 20 * mmsize], m15
+ punpckhqdq m3, m9
+ mova [rsp + 15 * mmsize], m5
+ punpcklqdq m6, m4
+ mova [rsp + 17 * mmsize], m3
+ mova [rsp + 14 * mmsize], m6
+ mova [rsp + 18 * mmsize], m14
+ mova m2, [rsp + 41 * mmsize]
+ punpckhwd m11, [rsp + 44 * mmsize]
+ mova m0, [rsp + 43 * mmsize]
+ punpcklwd m2, [rsp + 42 * mmsize]
+ mova m8, [rsp + 45 * mmsize]
+ punpcklwd m0, [rsp + 44 * mmsize]
+ mova m1, [rsp + 45 * mmsize]
+ punpcklwd m8, [rsp + 46 * mmsize]
+ mova m7, [rsp + 47 * mmsize]
+ mova m4, m2
+ punpckhwd m1, [rsp + 46 * mmsize]
+ mova m15, [rsp + 41 * mmsize]
+ punpcklwd m7, [rsp + 48 * mmsize]
+ mova m5, [rsp + 47 * mmsize]
+ mova m3, m8
+ punpckhwd m15, [rsp + 42 * mmsize]
+ punpckhwd m5, [rsp + 48 * mmsize]
+ punpckldq m4, m0
+ punpckhdq m2, m0
+ mova m0, m1
+ punpckldq m3, m7
+ punpckldq m0, m5
+ mova m6, m15
+ punpckhdq m1, m5
+ mova m5, m4
+ punpckhdq m8, m7
+ punpckldq m6, m11
+ punpckhdq m15, m11
+ mova m7, m2
+ punpcklqdq m5, m3
+ punpckhqdq m4, m3
+ punpckhqdq m2, m8
+ mova m14, m15
+ mova [rsp + 33 * mmsize], m5
+ punpcklqdq m7, m8
+ mova m8, m6
+ punpcklqdq m14, m1
+ mova [rsp + 34 * mmsize], m4
+ punpcklqdq m8, m0
+ mova [rsp + 36 * mmsize], m2
+ punpckhqdq m6, m0
+ punpckhqdq m15, m1
+ mova [rsp + 39 * mmsize], m14
+ mova [rsp + 40 * mmsize], m15
+ mova [rsp + 37 * mmsize], m8
+ mova [rsp + 38 * mmsize], m6
+ mova [rsp + 35 * mmsize], m7
+ mova m7, m10
+ mova m11, [rsp + 21 * mmsize]
+ punpcklwd m7, [rsp + 13 * mmsize]
+ mova m3, [rsp + 3 * mmsize]
+ punpcklwd m11, [rsp + 23 * mmsize]
+ mova m13, [rsp + 21 * mmsize]
+ punpcklwd m3, [rsp + 9 * mmsize]
+ mova m2, [rsp + 5 * mmsize]
+ punpckhwd m13, [rsp + 23 * mmsize]
+ mova m4, [rsp + 3 * mmsize]
+ mova m14, m11
+ punpcklwd m2, [rsp + 4 * mmsize]
+ mova m1, [rsp + 10 * mmsize]
+ punpckhwd m4, [rsp + 9 * mmsize]
+ mova m0, [rsp + 10 * mmsize]
+ mova m12, m13
+ punpcklwd m1, [rsp + 22 * mmsize]
+ mova m5, [rsp + 5 * mmsize]
+ punpckhwd m0, [rsp + 22 * mmsize]
+ mova m8, [rsp + 7 * mmsize]
+ punpckhwd m5, [rsp + 4 * mmsize]
+ mova m6, [rsp + 7 * mmsize]
+ punpckhdq m11, m3
+ punpckldq m14, m3
+ mova m3, m0
+ punpckldq m12, m4
+ punpckhdq m13, m4
+ mova m4, m1
+ punpckhdq m1, m2
+ mova m9, m11
+ punpckldq m3, m5
+ punpckldq m4, m2
+ punpckhqdq m11, m1
+ mova m15, m14
+ punpcklqdq m9, m1
+ mova m1, m12
+ punpckhdq m0, m5
+ punpcklqdq m15, m4
+ punpcklqdq m1, m3
+ mova m2, m13
+ punpckhqdq m14, m4
+ mova [rsp + 5 * mmsize], m15
+ punpcklqdq m2, m0
+ mova [rsp + 4 * mmsize], m1
+ punpckhqdq m13, m0
+ mova [rsp + 12 * mmsize], m2
+ mova m2, m10
+ punpckhqdq m12, m3
+ mova m1, [rsp + 2 * mmsize]
+ punpcklwd m8, [rsp + 6 * mmsize]
+ mova m15, [rsp + 2 * mmsize]
+ punpcklwd m1, [rsp + 1 * mmsize]
+ mova m4, [rsp + 0 * mmsize]
+ punpckhwd m15, [rsp + 1 * mmsize]
+ mova m0, [rsp + 0 * mmsize]
+ punpcklwd m4, [rsp + 8 * mmsize]
+ mova m5, m1
+ punpckhwd m0, [rsp + 8 * mmsize]
+ mova m3, m15
+ punpckhwd m6, [rsp + 6 * mmsize]
+ punpckhwd m2, [rsp + 13 * mmsize]
+ punpckldq m5, m4
+ punpckhdq m1, m4
+ mova m4, m8
+ punpckhdq m15, m0
+ punpckldq m3, m0
+ mova m0, m6
+ punpckldq m4, m7
+ punpckhdq m8, m7
+ mova m7, m5
+ punpckhdq m6, m2
+ punpckldq m0, m2
+ mova m2, m15
+ punpcklqdq m7, m4
+ punpcklqdq m2, m6
+ mova m10, m1
+ punpckhqdq m1, m8
+ punpckhqdq m5, m4
+ mova [rsp + 49 * mmsize], m7
+ punpcklqdq m10, m8
+ mova [rsp + 52 * mmsize], m1
+ mova m1, m3
+ mova [rsp + 55 * mmsize], m2
+ punpckhqdq m3, m0
+ mova [rsp + 50 * mmsize], m5
+ punpcklqdq m1, m0
+ mova [rsp + 54 * mmsize], m3
+ mova m8, m14
+ mova [rsp + 53 * mmsize], m1
+ punpcklwd m8, m11
+ punpckhwd m14, m11
+ mova m11, m12
+ mova [rsp + 51 * mmsize], m10
+ punpckhqdq m15, m6
+ mova m6, m9
+ punpcklwd m11, m13
+ punpckhwd m12, m13
+ mova m2, [rsp + 19 * mmsize]
+ mova m0, [rsp + 11 * mmsize]
+ mova m7, m2
+ mova m5, [rsp + 15 * mmsize]
+ punpckhwd m2, m0
+ punpcklwd m7, m0
+ mova m3, [rsp + 17 * mmsize]
+ mova m4, m5
+ mova m0, [rsp + 16 * mmsize]
+ punpcklwd m4, m3
+ punpckhwd m5, m3
+ mova m1, [rsp + 20 * mmsize]
+ mova m3, m0
+ mova m10, [rsp + 12 * mmsize]
+ punpckhwd m0, m1
+ punpcklwd m3, m1
+ mova m1, [rsp + 18 * mmsize]
+ punpcklwd m6, m10
+ punpckhwd m9, m10
+ mova m13, m1
+ mova m10, [rsp + 4 * mmsize]
+ punpcklwd m13, m10
+ punpckhwd m1, m10
+ mova m10, [rsp + 14 * mmsize]
+ mova [rsp + 56 * mmsize], m15
+ mova m15, m10
+ punpckhwd m10, [rsp + 5 * mmsize]
+ punpcklwd m15, [rsp + 5 * mmsize]
+ mova [rsp + 12 * mmsize], m10
+ mova [rsp + 11 * mmsize], m15
+ mova m10, [sse2_idct16_0 + 1 * mmsize]
+ mova m15, [sse2_idct16_0 + 0 * mmsize]
+ pmaddwd m10, m7
+ pmaddwd m15, m4
+ paddd m10, m15
+ mova [rsp + 7 * mmsize], m10
+ mova m15, [sse2_idct16_0 + 3 * mmsize]
+ mova m10, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ mova m15, [sse2_idct16_1 + 0 * mmsize]
+ paddd m10, [rsp + 7 * mmsize]
+ mova [rsp + 7 * mmsize], m10
+ pmaddwd m15, m4
+ mova m10, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m10, m7
+ paddd m10, m15
+ mova [rsp + 6 * mmsize], m10
+ mova m15, [sse2_idct16_1 + 3 * mmsize]
+ mova m10, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ paddd m10, [rsp + 6 * mmsize]
+ mova [rsp + 6 * mmsize], m10
+ mova m10, [sse2_idct16_2 + 1 * mmsize]
+ mova m15, [sse2_idct16_2 + 0 * mmsize]
+ pmaddwd m10, m7
+ pmaddwd m15, m4
+ paddd m10, m15
+ mova [rsp + 5 * mmsize], m10
+ mova m15, [sse2_idct16_2 + 3 * mmsize]
+ mova m10, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ mova m15, [sse2_idct16_3 + 0 * mmsize]
+ paddd m10, [rsp + 5 * mmsize]
+ mova [rsp + 5 * mmsize], m10
+ pmaddwd m15, m4
+ mova m10, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m10, m7
+ paddd m10, m15
+ mova [rsp + 4 * mmsize], m10
+ mova m15, [sse2_idct16_3 + 3 * mmsize]
+ mova m10, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ mova m15, [sse2_idct16_4 + 0 * mmsize]
+ paddd m10, [rsp + 4 * mmsize]
+ mova [rsp + 4 * mmsize], m10
+ pmaddwd m15, m4
+ mova m10, [sse2_idct16_4 + 1 * mmsize]
+ pmaddwd m10, m7
+ paddd m10, m15
+ mova [rsp + 2 * mmsize], m10
+ mova m15, [sse2_idct16_4 + 3 * mmsize]
+ mova m10, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ mova m15, [sse2_idct16_5 + 0 * mmsize]
+ paddd m10, [rsp + 2 * mmsize]
+ mova [rsp + 2 * mmsize], m10
+ pmaddwd m15, m4
+ mova m10, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m10, m7
+ paddd m10, m15
+ mova [rsp + 1 * mmsize], m10
+ mova m15, [sse2_idct16_5 + 3 * mmsize]
+ mova m10, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ mova m15, [sse2_idct16_6 + 0 * mmsize]
+ paddd m10, [rsp + 1 * mmsize]
+ mova [rsp + 1 * mmsize], m10
+ pmaddwd m15, m4
+ mova m10, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m10, m7
+ paddd m10, m15
+ mova [rsp + 0 * mmsize], m10
+ mova m15, [sse2_idct16_6 + 3 * mmsize]
+ mova m10, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m15, m11
+ pmaddwd m10, m8
+ paddd m10, m15
+ paddd m10, [rsp + 0 * mmsize]
+ mova [rsp + 0 * mmsize], m10
+ pmaddwd m4, [sse2_idct16_7 + 0 * mmsize]
+ pmaddwd m11, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m7, [sse2_idct16_7 + 1 * mmsize]
+ pmaddwd m8, [sse2_idct16_7 + 2 * mmsize]
+ paddd m7, m4
+ paddd m8, m11
+ paddd m8, m7
+ mova [rsp + 8 * mmsize], m8
+ mova m4, [sse2_idct16_0 + 1 * mmsize]
+ mova m7, [sse2_idct16_0 + 0 * mmsize]
+ pmaddwd m4, m2
+ pmaddwd m7, m5
+ paddd m4, m7
+ mova [rsp + 13 * mmsize], m4
+ mova m7, [sse2_idct16_0 + 3 * mmsize]
+ mova m4, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ mova m8, [sse2_idct16_1 + 0 * mmsize]
+ paddd m4, m7
+ paddd m4, [rsp + 13 * mmsize]
+ mova [rsp + 13 * mmsize], m4
+ pmaddwd m8, m5
+ mova m4, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m4, m2
+ paddd m4, m8
+ mova [rsp + 14 * mmsize], m4
+ mova m7, [sse2_idct16_1 + 3 * mmsize]
+ mova m4, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ mova m10, [sse2_idct16_2 + 0 * mmsize]
+ paddd m4, m7
+ paddd m4, [rsp + 14 * mmsize]
+ mova [rsp + 14 * mmsize], m4
+ pmaddwd m10, m5
+ mova m4, [sse2_idct16_2 + 1 * mmsize]
+ pmaddwd m4, m2
+ paddd m4, m10
+ mova [rsp + 15 * mmsize], m4
+ mova m7, [sse2_idct16_2 + 3 * mmsize]
+ mova m4, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ mova m15, [sse2_idct16_3 + 0 * mmsize]
+ paddd m4, m7
+ paddd m4, [rsp + 15 * mmsize]
+ mova [rsp + 15 * mmsize], m4
+ pmaddwd m15, m5
+ mova m4, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m4, m2
+ paddd m4, m15
+ mova [rsp + 16 * mmsize], m4
+ mova m7, [sse2_idct16_3 + 3 * mmsize]
+ mova m4, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ paddd m4, m7
+ paddd m4, [rsp + 16 * mmsize]
+ mova [rsp + 16 * mmsize], m4
+ mova m4, [sse2_idct16_4 + 1 * mmsize]
+ mova m7, [sse2_idct16_4 + 0 * mmsize]
+ pmaddwd m4, m2
+ pmaddwd m7, m5
+ paddd m4, m7
+ mova [rsp + 17 * mmsize], m4
+ mova m7, [sse2_idct16_4 + 3 * mmsize]
+ mova m4, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ mova m8, [sse2_idct16_5 + 0 * mmsize]
+ paddd m4, m7
+ paddd m4, [rsp + 17 * mmsize]
+ mova [rsp + 17 * mmsize], m4
+ pmaddwd m8, m5
+ mova m4, [sse2_idct16_5 + 1 * mmsize]
+ pmaddwd m4, m2
+ paddd m4, m8
+ mova [rsp + 18 * mmsize], m4
+ mova m7, [sse2_idct16_5 + 3 * mmsize]
+ mova m4, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m7, m12
+ pmaddwd m4, m14
+ mova m10, [sse2_idct16_6 + 0 * mmsize]
+ paddd m4, m7
+ paddd m4, [rsp + 18 * mmsize]
+ mova [rsp + 18 * mmsize], m4
+ pmaddwd m10, m5
+ pmaddwd m5, [sse2_idct16_7 + 0 * mmsize]
+ mova m4, [sse2_idct16_6 + 1 * mmsize]
+ mova m7, [sse2_idct16_6 + 3 * mmsize]
+ pmaddwd m4, m2
+ paddd m4, m10
+ pmaddwd m2, [sse2_idct16_7 + 1 * mmsize]
+ paddd m2, m5
+ mova [rsp + 19 * mmsize], m4
+ pmaddwd m7, m12
+ pmaddwd m12, [sse2_idct16_7 + 3 * mmsize]
+ mova m4, [sse2_idct16_6 + 2 * mmsize]
+ mova m15, [tab_idct8_3 + 0 * mmsize]
+ pmaddwd m4, m14
+ pmaddwd m14, [sse2_idct16_7 + 2 * mmsize]
+ paddd m14, m12
+ paddd m14, m2
+ paddd m4, m7
+ mova m2, [tab_idct8_3 + 1 * mmsize]
+ pmaddwd m15, m3
+ paddd m4, [rsp + 19 * mmsize]
+ mova [rsp + 20 * mmsize], m14
+ pmaddwd m2, m6
+ paddd m15, m2
+ mova [rsp + 19 * mmsize], m4
+ mova m2, [tab_idct8_3 + 1 * mmsize]
+ mova m14, [tab_idct8_3 + 0 * mmsize]
+ pmaddwd m2, m9
+ pmaddwd m14, m0
+ mova m4, [tab_idct8_3 + 2 * mmsize]
+ paddd m14, m2
+ mova m2, [tab_idct8_3 + 3 * mmsize]
+ pmaddwd m4, m3
+ pmaddwd m2, m6
+ paddd m2, m4
+ mova [rsp + 24 * mmsize], m2
+ mova m2, [tab_idct8_3 + 3 * mmsize]
+ mova m12, [tab_idct8_3 + 2 * mmsize]
+ pmaddwd m2, m9
+ pmaddwd m12, m0
+ mova m5, [tab_idct8_3 + 4 * mmsize]
+ paddd m12, m2
+ mova m2, [tab_idct8_3 + 5 * mmsize]
+ pmaddwd m5, m3
+ pmaddwd m3, [tab_idct8_3 + 6 * mmsize]
+ pmaddwd m2, m6
+ mova m7, [tab_idct8_3 + 4 * mmsize]
+ paddd m2, m5
+ pmaddwd m6, [tab_idct8_3 + 7 * mmsize]
+ paddd m6, m3
+ mova [rsp + 25 * mmsize], m2
+ pmaddwd m7, m0
+ pmaddwd m0, [tab_idct8_3 + 6 * mmsize]
+ mova [rsp + 28 * mmsize], m6
+ mova m2, [tab_idct8_3 + 5 * mmsize]
+ mova m8, [rsp + 12 * mmsize]
+ pmaddwd m2, m9
+ pmaddwd m9, [tab_idct8_3 + 7 * mmsize]
+ paddd m0, m9
+ paddd m2, m7
+ mova m9, [rsp + 11 * mmsize]
+ pmaddwd m8, [tab_dct4 + 0 * mmsize]
+ mova m6, m8
+ mova m10, [rsp + 12 * mmsize]
+ pmaddwd m9, [tab_dct4 + 0 * mmsize]
+ mova m5, m9
+ mova m11, [rsp + 11 * mmsize]
+ pmaddwd m10, [tab_dct4 + 2 * mmsize]
+ mova m7, m10
+ mova m4, [tab_dct4 + 1 * mmsize]
+ pmaddwd m11, [tab_dct4 + 2 * mmsize]
+ mova m3, [tab_dct4 + 1 * mmsize]
+ pmaddwd m4, m13
+ paddd m5, m4
+ psubd m9, m4
+ pmaddwd m3, m1
+ mova [rsp + 26 * mmsize], m2
+ pmaddwd m1, [tab_dct4 + 3 * mmsize]
+ paddd m6, m3
+ paddd m7, m1
+ psubd m8, m3
+ psubd m10, m1
+ mova m2, [tab_dct4 + 3 * mmsize]
+ mova [rsp + 11 * mmsize], m5
+ pmaddwd m2, m13
+ mova m13, m11
+ psubd m11, m2
+ mova [rsp + 12 * mmsize], m6
+ paddd m6, m14
+ paddd m13, m2
+ mova [rsp + 31 * mmsize], m7
+ mova m7, m5
+ mova [rsp + 3 * mmsize], m8
+ paddd m8, m0
+ paddd m7, m15
+ mova [rsp + 27 * mmsize], m8
+ mova m8, [rsp + 11 * mmsize]
+ psubd m8, m15
+ mova m15, m8
+ mova m8, [rsp + 12 * mmsize]
+ mova m4, [rsp + 31 * mmsize]
+ psubd m8, m14
+ mova m14, [rsp + 31 * mmsize]
+ paddd m4, m12
+ mova [rsp + 41 * mmsize], m8
+ psubd m14, m12
+ mova m12, m14
+ mova m8, [rsp + 3 * mmsize]
+ mova m14, [rsp + 27 * mmsize]
+ psubd m8, m0
+ mova m0, [IDCT_ADD]
+ paddd m14, m0
+ mova [rsp + 27 * mmsize], m14
+ paddd m7, m0
+ paddd m6, m0
+ paddd m4, m0
+ paddd m8, m0
+ paddd m12, m0
+ paddd m15, m0
+ mova m14, [rsp + 41 * mmsize]
+ paddd m14, m0
+ mova [rsp + 41 * mmsize], m14
+ mova m14, [rsp + 7 * mmsize]
+ paddd m14, m7
+ mova [rsp + 11 * mmsize], m14
+ mova m14, [rsp + 13 * mmsize]
+ mova m5, [rsp + 24 * mmsize]
+ paddd m14, m6
+ mova [rsp + 29 * mmsize], m14
+ paddd m5, m13
+ paddd m5, m0
+ psubd m13, [rsp + 24 * mmsize]
+ paddd m13, m0
+ mova m14, [rsp + 6 * mmsize]
+ paddd m14, m5
+ mova [rsp + 12 * mmsize], m14
+ mova m14, [rsp + 14 * mmsize]
+ mova m3, [rsp + 25 * mmsize]
+ paddd m14, m4
+ mova [rsp + 21 * mmsize], m14
+ paddd m3, m11
+ paddd m3, m0
+ psubd m11, [rsp + 25 * mmsize]
+ paddd m11, m0
+ mova m14, [rsp + 5 * mmsize]
+ mova m2, [rsp + 26 * mmsize]
+ paddd m14, m3
+ mova m1, [rsp + 28 * mmsize]
+ paddd m2, m10
+ psubd m10, [rsp + 26 * mmsize]
+ paddd m2, m0
+ paddd m10, m0
+ mova [rsp + 24 * mmsize], m14
+ paddd m1, m9
+ psubd m9, [rsp + 28 * mmsize]
+ paddd m1, m0
+ paddd m9, m0
+ mova m14, [rsp + 15 * mmsize]
+ paddd m14, m2
+ mova [rsp + 23 * mmsize], m14
+ psubd m7, [rsp + 7 * mmsize]
+ psubd m6, [rsp + 13 * mmsize]
+ psubd m5, [rsp + 6 * mmsize]
+ psubd m4, [rsp + 14 * mmsize]
+ mova [rsp + 7 * mmsize], m6
+ psubd m3, [rsp + 5 * mmsize]
+ psubd m2, [rsp + 15 * mmsize]
+ psrad m5, IDCT_SHIFT2
+ psrad m7, IDCT_SHIFT2
+ psrad m2, IDCT_SHIFT2
+ mova [rsp + 6 * mmsize], m4
+ psrad m3, IDCT_SHIFT2
+ packssdw m3, m2
+ mova m14, [rsp + 4 * mmsize]
+ paddd m14, m1
+ mova [rsp + 25 * mmsize], m14
+ psubd m1, [rsp + 4 * mmsize]
+ psrad m1, IDCT_SHIFT2
+ mova m14, [rsp + 27 * mmsize]
+ paddd m14, [rsp + 16 * mmsize]
+ mova [rsp + 3 * mmsize], m14
+ mova m14, [rsp + 2 * mmsize]
+ paddd m14, m9
+ mova [rsp + 26 * mmsize], m14
+ psubd m9, [rsp + 2 * mmsize]
+ psrad m9, IDCT_SHIFT2
+ mova m14, [rsp + 17 * mmsize]
+ paddd m14, m8
+ mova [rsp + 9 * mmsize], m14
+ psubd m8, [rsp + 17 * mmsize]
+ psrad m8, IDCT_SHIFT2
+ packssdw m9, m8
+ mova m14, [rsp + 1 * mmsize]
+ paddd m14, m11
+ mova [rsp + 28 * mmsize], m14
+ psubd m11, [rsp + 1 * mmsize]
+ psrad m11, IDCT_SHIFT2
+ mova m14, [rsp + 18 * mmsize]
+ paddd m14, m10
+ mova [rsp + 10 * mmsize], m14
+ psubd m10, [rsp + 18 * mmsize]
+ psrad m10, IDCT_SHIFT2
+ packssdw m11, m10
+ mova m14, [rsp + 0 * mmsize]
+ paddd m14, m13
+ mova [rsp + 31 * mmsize], m14
+ psubd m13, [rsp + 0 * mmsize]
+ psrad m13, IDCT_SHIFT2
+ mova m14, [rsp + 19 * mmsize]
+ paddd m14, m12
+ mova [rsp + 22 * mmsize], m14
+ psubd m12, [rsp + 19 * mmsize]
+ psrad m12, IDCT_SHIFT2
+ packssdw m13, m12
+ mova m14, [rsp + 8 * mmsize]
+ paddd m14, m15
+ mova [rsp + 32 * mmsize], m14
+ psubd m15, [rsp + 8 * mmsize]
+ psrad m15, IDCT_SHIFT2
+ mova m14, [rsp + 20 * mmsize]
+ mova m6, [rsp + 27 * mmsize]
+ paddd m14, [rsp + 41 * mmsize]
+ mova m4, [rsp + 11 * mmsize]
+ psubd m6, [rsp + 16 * mmsize]
+ psrad m6, IDCT_SHIFT2
+ packssdw m1, m6
+ mova [rsp + 30 * mmsize], m14
+ psrad m4, IDCT_SHIFT2
+ mova m14, [rsp + 41 * mmsize]
+ psubd m14, [rsp + 20 * mmsize]
+ mova [rsp + 11 * mmsize], m4
+ psrad m14, IDCT_SHIFT2
+ packssdw m15, m14
+ mova m4, [rsp + 29 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 29 * mmsize], m4
+ mova m4, [rsp + 12 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 12 * mmsize], m4
+ mova m4, [rsp + 21 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 21 * mmsize], m4
+ mova m4, [rsp + 24 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 24 * mmsize], m4
+ mova m4, [rsp + 23 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 23 * mmsize], m4
+ mova m4, [rsp + 25 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 25 * mmsize], m4
+ mova m4, [rsp + 3 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 3 * mmsize], m4
+ mova m4, [rsp + 26 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 26 * mmsize], m4
+ mova m4, [rsp + 9 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 9 * mmsize], m4
+ mova m4, [rsp + 28 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 28 * mmsize], m4
+ mova m4, [rsp + 10 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 10 * mmsize], m4
+ mova m4, [rsp + 31 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 31 * mmsize], m4
+ mova m4, [rsp + 22 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 22 * mmsize], m4
+ mova m4, [rsp + 32 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 32 * mmsize], m4
+ mova m4, [rsp + 30 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 30 * mmsize], m4
+ mova m4, [rsp + 6 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 6 * mmsize], m4
+ mova m4, [rsp + 7 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 7 * mmsize], m4
+ packssdw m5, [rsp + 6 * mmsize]
+ mova [rsp + 41 * mmsize], m15
+ mova [rsp + 42 * mmsize], m13
+ mova [rsp + 43 * mmsize], m11
+ mova [rsp + 44 * mmsize], m9
+ mova [rsp + 45 * mmsize], m1
+ mova [rsp + 46 * mmsize], m3
+ mova [rsp + 47 * mmsize], m5
+ mova m4, [rsp + 11 * mmsize]
+ packssdw m7, [rsp + 7 * mmsize]
+ packssdw m4, [rsp + 29 * mmsize]
+ mova [rsp + 11 * mmsize], m4
+ mova m4, [rsp + 12 * mmsize]
+ packssdw m4, [rsp + 21 * mmsize]
+ mova [rsp + 12 * mmsize], m4
+ mova m4, [rsp + 24 * mmsize]
+ packssdw m4, [rsp + 23 * mmsize]
+ mova [rsp + 24 * mmsize], m4
+ mova m4, [rsp + 25 * mmsize]
+ packssdw m4, [rsp + 3 * mmsize]
+ mova [rsp + 25 * mmsize], m4
+ mova m4, [rsp + 26 * mmsize]
+ packssdw m4, [rsp + 9 * mmsize]
+ mova [rsp + 26 * mmsize], m4
+ mova m4, [rsp + 28 * mmsize]
+ packssdw m4, [rsp + 10 * mmsize]
+ mova [rsp + 28 * mmsize], m4
+ mova m4, [rsp + 31 * mmsize]
+ packssdw m4, [rsp + 22 * mmsize]
+ mova [rsp + 31 * mmsize], m4
+ mova m4, [rsp + 32 * mmsize]
+ packssdw m4, [rsp + 30 * mmsize]
+ mova [rsp + 32 * mmsize], m4
+ mova [rsp + 48 * mmsize], m7
+ mova m4, [rsp + 34 * mmsize]
+ mova m2, [rsp + 36 * mmsize]
+ mova m9, m4
+ punpckhwd m4, m2
+ mova m1, [rsp + 52 * mmsize]
+ punpcklwd m9, m2
+ mova m2, [rsp + 50 * mmsize]
+ mova m7, m2
+ mova m5, [rsp + 35 * mmsize]
+ punpckhwd m2, m1
+ punpcklwd m7, m1
+ mova m1, [rsp + 39 * mmsize]
+ mova m10, m5
+ punpckhwd m5, m1
+ punpcklwd m10, m1
+ mova m1, [rsp + 55 * mmsize]
+ mova [rsp + 6 * mmsize], m5
+ mova [rsp + 7 * mmsize], m10
+ mova m5, [rsp + 51 * mmsize]
+ mova m6, [rsp + 38 * mmsize]
+ mova m14, m5
+ punpckhwd m5, m1
+ mova m15, [rsp + 40 * mmsize]
+ mova m11, m6
+ punpcklwd m14, m1
+ mova m3, [rsp + 54 * mmsize]
+ punpcklwd m11, m15
+ punpckhwd m6, m15
+ mova m1, [rsp + 37 * mmsize]
+ mova m8, m3
+ mova m15, [rsp + 56 * mmsize]
+ mova [rsp + 4 * mmsize], m5
+ punpcklwd m8, m15
+ punpckhwd m3, m15
+ mova m15, m1
+ mova [rsp + 5 * mmsize], m14
+ mova m5, [rsp + 53 * mmsize]
+ punpckhwd m1, m5
+ mova m10, [rsp + 33 * mmsize]
+ punpcklwd m15, m5
+ mova [rsp + 9 * mmsize], m1
+ mova m5, m10
+ mova [rsp + 3 * mmsize], m15
+ mova m1, [rsp + 49 * mmsize]
+ punpckhwd m10, m1
+ punpcklwd m5, m1
+ mova m1, [sse2_idct16_0 + 1 * mmsize]
+ mova [rsp + 22 * mmsize], m10
+ pmaddwd m1, m11
+ mova [rsp + 10 * mmsize], m5
+ mova m10, [sse2_idct16_0 + 0 * mmsize]
+ pmaddwd m10, m9
+ paddd m1, m10
+ mova [rsp + 2 * mmsize], m1
+ mova m10, m9
+ mova m5, [sse2_idct16_0 + 3 * mmsize]
+ mova m1, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m5, m8
+ pmaddwd m1, m7
+ mova m14, [sse2_idct16_1 + 0 * mmsize]
+ paddd m1, m5
+ paddd m1, [rsp + 2 * mmsize]
+ mova [rsp + 2 * mmsize], m1
+ pmaddwd m14, m9
+ mova m1, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m14
+ mova [rsp + 1 * mmsize], m1
+ mova m14, m11
+ mova m5, [sse2_idct16_1 + 3 * mmsize]
+ mova m1, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m5, m8
+ pmaddwd m1, m7
+ mova m15, [sse2_idct16_2 + 0 * mmsize]
+ paddd m1, m5
+ paddd m1, [rsp + 1 * mmsize]
+ mova [rsp + 1 * mmsize], m1
+ pmaddwd m15, m9
+ mova m1, [sse2_idct16_2 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m15
+ mova [rsp + 0 * mmsize], m1
+ mova m15, m9
+ mova m5, [sse2_idct16_2 + 3 * mmsize]
+ mova m1, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m5, m8
+ pmaddwd m1, m7
+ paddd m1, m5
+ mova m5, [sse2_idct16_3 + 0 * mmsize]
+ paddd m1, [rsp + 0 * mmsize]
+ mova [rsp + 0 * mmsize], m1
+ pmaddwd m5, m9
+ mova m1, [sse2_idct16_3 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m5
+ mova [rsp + 8 * mmsize], m1
+ mova m5, [sse2_idct16_3 + 3 * mmsize]
+ mova m1, [sse2_idct16_3 + 2 * mmsize]
+ pmaddwd m5, m8
+ mova m13, [sse2_idct16_4 + 0 * mmsize]
+ pmaddwd m1, m7
+ paddd m1, m5
+ paddd m1, [rsp + 8 * mmsize]
+ mova m5, m8
+ mova [rsp + 8 * mmsize], m1
+ pmaddwd m10, m13
+ pmaddwd m13, m4
+ mova m1, [sse2_idct16_4 + 1 * mmsize]
+ pmaddwd m1, m11
+ paddd m1, m10
+ mova [rsp + 13 * mmsize], m1
+ mova m1, m7
+ mova m12, [sse2_idct16_4 + 3 * mmsize]
+ mova m10, [sse2_idct16_4 + 2 * mmsize]
+ pmaddwd m5, m12
+ pmaddwd m12, m3
+ pmaddwd m1, m10
+ paddd m1, m5
+ mova m5, [sse2_idct16_5 + 1 * mmsize]
+ paddd m1, [rsp + 13 * mmsize]
+ pmaddwd m10, m2
+ paddd m10, m12
+ mova [rsp + 13 * mmsize], m1
+ pmaddwd m14, m5
+ pmaddwd m5, m6
+ mova m1, [sse2_idct16_5 + 0 * mmsize]
+ pmaddwd m15, m1
+ paddd m14, m15
+ mova [rsp + 14 * mmsize], m14
+ pmaddwd m1, m4
+ paddd m5, m1
+ mova m15, [sse2_idct16_5 + 3 * mmsize]
+ mova m14, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m15, m8
+ pmaddwd m14, m7
+ paddd m14, m15
+ mova m15, [sse2_idct16_6 + 0 * mmsize]
+ paddd m14, [rsp + 14 * mmsize]
+ mova [rsp + 14 * mmsize], m14
+ pmaddwd m15, m9
+ mova m14, [sse2_idct16_6 + 1 * mmsize]
+ pmaddwd m14, m11
+ paddd m14, m15
+ mova [rsp + 15 * mmsize], m14
+ pmaddwd m9, [sse2_idct16_7 + 0 * mmsize]
+ pmaddwd m11, [sse2_idct16_7 + 1 * mmsize]
+ paddd m11, m9
+ mova m15, [sse2_idct16_6 + 3 * mmsize]
+ mova m14, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m15, m8
+ pmaddwd m8, [sse2_idct16_7 + 3 * mmsize]
+ pmaddwd m14, m7
+ pmaddwd m7, [sse2_idct16_7 + 2 * mmsize]
+ paddd m7, m8
+ mova m8, [sse2_idct16_0 + 0 * mmsize]
+ paddd m7, m11
+ paddd m14, m15
+ paddd m14, [rsp + 15 * mmsize]
+ mova [rsp + 16 * mmsize], m7
+ pmaddwd m8, m4
+ mova [rsp + 15 * mmsize], m14
+ mova m7, [sse2_idct16_0 + 1 * mmsize]
+ pmaddwd m7, m6
+ paddd m7, m8
+ mova [rsp + 17 * mmsize], m7
+ mova m8, [sse2_idct16_0 + 3 * mmsize]
+ mova m7, [sse2_idct16_0 + 2 * mmsize]
+ pmaddwd m8, m3
+ pmaddwd m7, m2
+ mova m14, [sse2_idct16_1 + 0 * mmsize]
+ paddd m7, m8
+ paddd m7, [rsp + 17 * mmsize]
+ mova [rsp + 17 * mmsize], m7
+ pmaddwd m14, m4
+ mova m7, [sse2_idct16_1 + 1 * mmsize]
+ pmaddwd m7, m6
+ paddd m7, m14
+ mova [rsp + 18 * mmsize], m7
+ mova m8, [sse2_idct16_1 + 3 * mmsize]
+ mova m7, [sse2_idct16_1 + 2 * mmsize]
+ pmaddwd m8, m3
+ pmaddwd m7, m2
+ mova m15, [sse2_idct16_2 + 0 * mmsize]
+ paddd m7, m8
+ paddd m7, [rsp + 18 * mmsize]
+ mova [rsp + 18 * mmsize], m7
+ pmaddwd m15, m4
+ mova m7, [sse2_idct16_2 + 1 * mmsize]
+ pmaddwd m7, m6
+ paddd m7, m15
+ mova [rsp + 19 * mmsize], m7
+ mova m8, [sse2_idct16_2 + 3 * mmsize]
+ mova m7, [sse2_idct16_2 + 2 * mmsize]
+ pmaddwd m8, m3
+ pmaddwd m7, m2
+ paddd m7, m8
+ paddd m7, [rsp + 19 * mmsize]
+ mova [rsp + 19 * mmsize], m7
+ mova [rsp + 29 * mmsize], m5
+ mova m7, [sse2_idct16_3 + 1 * mmsize]
+ mova m8, [sse2_idct16_3 + 0 * mmsize]
+ pmaddwd m7, m6
+ pmaddwd m8, m4
+ paddd m7, m8
+ mova m8, [sse2_idct16_3 + 2 * mmsize]
+ mova [rsp + 20 * mmsize], m7
+ pmaddwd m8, m2
+ mova m7, [sse2_idct16_3 + 3 * mmsize]
+ mova m5, [sse2_idct16_5 + 3 * mmsize]
+ pmaddwd m7, m3
+ paddd m8, m7
+ paddd m8, [rsp + 20 * mmsize]
+ mova m7, [sse2_idct16_4 + 1 * mmsize]
+ pmaddwd m5, m3
+ mova m1, [sse2_idct16_5 + 2 * mmsize]
+ pmaddwd m7, m6
+ paddd m7, m13
+ paddd m10, m7
+ pmaddwd m1, m2
+ mova [rsp + 27 * mmsize], m10
+ paddd m1, m5
+ paddd m1, [rsp + 29 * mmsize]
+ mova [rsp + 29 * mmsize], m1
+ mova [rsp + 20 * mmsize], m8
+ mova m1, [sse2_idct16_6 + 1 * mmsize]
+ mova m10, [sse2_idct16_6 + 0 * mmsize]
+ pmaddwd m1, m6
+ pmaddwd m6, [sse2_idct16_7 + 1 * mmsize]
+ pmaddwd m10, m4
+ paddd m1, m10
+ mova [rsp + 21 * mmsize], m1
+ pmaddwd m4, [sse2_idct16_7 + 0 * mmsize]
+ paddd m6, m4
+ mova m5, [sse2_idct16_6 + 3 * mmsize]
+ mova m1, [sse2_idct16_6 + 2 * mmsize]
+ pmaddwd m5, m3
+ pmaddwd m3, [sse2_idct16_7 + 3 * mmsize]
+ mova m15, [rsp + 7 * mmsize]
+ pmaddwd m1, m2
+ paddd m1, m5
+ paddd m1, [rsp + 21 * mmsize]
+ pmaddwd m2, [sse2_idct16_7 + 2 * mmsize]
+ paddd m2, m3
+ mova [rsp + 21 * mmsize], m1
+ pmaddwd m15, [tab_idct8_3 + 0 * mmsize]
+ paddd m2, m6
+ mova [rsp + 23 * mmsize], m2
+ mova m1, [rsp + 5 * mmsize]
+ mova m14, [rsp + 6 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 1 * mmsize]
+ paddd m15, m1
+ mova m1, [rsp + 4 * mmsize]
+ pmaddwd m14, [tab_idct8_3 + 0 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 1 * mmsize]
+ mova m13, [rsp + 7 * mmsize]
+ paddd m14, m1
+ mova m1, [rsp + 5 * mmsize]
+ pmaddwd m13, [tab_idct8_3 + 2 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 3 * mmsize]
+ mova m12, [rsp + 6 * mmsize]
+ paddd m13, m1
+ mova m1, [rsp + 4 * mmsize]
+ pmaddwd m12, [tab_idct8_3 + 2 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 3 * mmsize]
+ mova m2, [rsp + 7 * mmsize]
+ paddd m12, m1
+ mova m1, [rsp + 5 * mmsize]
+ pmaddwd m2, [tab_idct8_3 + 4 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 5 * mmsize]
+ paddd m1, m2
+ mova [rsp + 30 * mmsize], m1
+ mova m1, [rsp + 4 * mmsize]
+ mova m7, [rsp + 6 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 5 * mmsize]
+ mova m3, [rsp + 7 * mmsize]
+ pmaddwd m7, [tab_idct8_3 + 4 * mmsize]
+ paddd m7, m1
+ mova m1, [rsp + 5 * mmsize]
+ pmaddwd m3, [tab_idct8_3 + 6 * mmsize]
+ mova m5, [rsp + 6 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 7 * mmsize]
+ paddd m1, m3
+ mova m9, [rsp + 10 * mmsize]
+ pmaddwd m5, [tab_idct8_3 + 6 * mmsize]
+ mova [rsp + 7 * mmsize], m1
+ pmaddwd m9, [tab_dct4 + 0 * mmsize]
+ mova m6, m9
+ mova m1, [rsp + 4 * mmsize]
+ mova m4, [rsp + 3 * mmsize]
+ pmaddwd m1, [tab_idct8_3 + 7 * mmsize]
+ paddd m5, m1
+ mova m3, [rsp + 9 * mmsize]
+ pmaddwd m4, [tab_dct4 + 1 * mmsize]
+ paddd m6, m4
+ psubd m9, m4
+ mova m2, [rsp + 3 * mmsize]
+ pmaddwd m3, [tab_dct4 + 1 * mmsize]
+ mova m1, [rsp + 9 * mmsize]
+ pmaddwd m2, [tab_dct4 + 3 * mmsize]
+ mova m8, [rsp + 22 * mmsize]
+ pmaddwd m1, [tab_dct4 + 3 * mmsize]
+ mova m11, [rsp + 10 * mmsize]
+ pmaddwd m8, [tab_dct4 + 0 * mmsize]
+ mova m10, [rsp + 22 * mmsize]
+ pmaddwd m11, [tab_dct4 + 2 * mmsize]
+ pmaddwd m10, [tab_dct4 + 2 * mmsize]
+ mova [rsp + 4 * mmsize], m6
+ mova m6, m8
+ psubd m8, m3
+ mova [rsp + 22 * mmsize], m8
+ paddd m6, m3
+ mova [rsp + 3 * mmsize], m6
+ mova m6, m11
+ psubd m11, m2
+ paddd m6, m2
+ mova [rsp + 9 * mmsize], m6
+ mova m6, m10
+ psubd m10, m1
+ paddd m6, m1
+ mova [rsp + 10 * mmsize], m6
+ mova m1, m6
+ mova m6, m10
+ psubd m10, m7
+ paddd m10, m0
+ paddd m1, m12
+ paddd m1, m0
+ mova m8, [rsp + 30 * mmsize]
+ paddd m6, m7
+ paddd m8, m11
+ mova [rsp + 33 * mmsize], m8
+ psubd m11, [rsp + 30 * mmsize]
+ paddd m11, m0
+ mova [rsp + 5 * mmsize], m6
+ mova m8, [rsp + 22 * mmsize]
+ mova m4, [rsp + 4 * mmsize]
+ paddd m8, m5
+ mova [rsp + 6 * mmsize], m8
+ paddd m4, m15
+ paddd m4, m0
+ mova m8, [rsp + 4 * mmsize]
+ mova m3, [rsp + 3 * mmsize]
+ psubd m8, m15
+ mova m15, m8
+ mova m8, [rsp + 3 * mmsize]
+ paddd m3, m14
+ paddd m3, m0
+ paddd m15, m0
+ mova m2, [rsp + 9 * mmsize]
+ psubd m8, m14
+ mova m14, m8
+ mova m8, [rsp + 9 * mmsize]
+ paddd m2, m13
+ paddd m2, m0
+ paddd m14, m0
+ psubd m8, m13
+ mova m13, m8
+ mova m8, [rsp + 10 * mmsize]
+ paddd m13, m0
+ psubd m8, m12
+ mova m12, m8
+ mova m8, [rsp + 22 * mmsize]
+ paddd m12, m0
+ mova m7, [rsp + 5 * mmsize]
+ psubd m8, m5
+ paddd m8, m0
+ mova m5, [rsp + 33 * mmsize]
+ paddd m7, m0
+ mova m6, [rsp + 7 * mmsize]
+ paddd m5, m0
+ mova [rsp + 5 * mmsize], m7
+ paddd m6, m9
+ psubd m9, [rsp + 7 * mmsize]
+ paddd m6, m0
+ paddd m9, m0
+ mova [rsp + 33 * mmsize], m5
+ mova m7, [rsp + 2 * mmsize]
+ mova m5, [rsp + 6 * mmsize]
+ paddd m7, m4
+ paddd m5, m0
+ mova [rsp + 7 * mmsize], m7
+ mova [rsp + 6 * mmsize], m5
+ mova m0, [rsp + 17 * mmsize]
+ mova m5, [rsp + 1 * mmsize]
+ paddd m0, m3
+ mova m7, [rsp + 18 * mmsize]
+ paddd m5, m2
+ mova [rsp + 30 * mmsize], m0
+ paddd m7, m1
+ mova [rsp + 4 * mmsize], m5
+ mova [rsp + 34 * mmsize], m7
+ psubd m3, [rsp + 17 * mmsize]
+ psubd m4, [rsp + 2 * mmsize]
+ psubd m2, [rsp + 1 * mmsize]
+ psubd m1, [rsp + 18 * mmsize]
+ psrad m4, IDCT_SHIFT2
+ mova [rsp + 17 * mmsize], m3
+ psrad m2, IDCT_SHIFT2
+ psrad m1, IDCT_SHIFT2
+ mova m0, [rsp + 33 * mmsize]
+ paddd m0, [rsp + 0 * mmsize]
+ mova [rsp + 3 * mmsize], m0
+ mova m0, [rsp + 6 * mmsize]
+ paddd m0, [rsp + 20 * mmsize]
+ mova [rsp + 36 * mmsize], m0
+ mova m0, [rsp + 14 * mmsize]
+ mova m5, [rsp + 5 * mmsize]
+ paddd m0, m11
+ mova [rsp + 22 * mmsize], m0
+ paddd m5, [rsp + 19 * mmsize]
+ mova [rsp + 35 * mmsize], m5
+ mova m0, [rsp + 21 * mmsize]
+ mova m7, [rsp + 8 * mmsize]
+ paddd m0, m12
+ mova m5, [rsp + 13 * mmsize]
+ paddd m7, m6
+ mova [rsp + 39 * mmsize], m0
+ paddd m5, m9
+ mova [rsp + 9 * mmsize], m7
+ mova [rsp + 10 * mmsize], m5
+ mova m0, [rsp + 23 * mmsize]
+ mova m7, [rsp + 27 * mmsize]
+ paddd m0, m14
+ mova m5, [rsp + 29 * mmsize]
+ paddd m7, m8
+ mova [rsp + 40 * mmsize], m0
+ paddd m5, m10
+ mova [rsp + 37 * mmsize], m7
+ mova [rsp + 38 * mmsize], m5
+ mova m0, [rsp + 33 * mmsize]
+ mova m3, [rsp + 5 * mmsize]
+ psubd m0, [rsp + 0 * mmsize]
+ psrad m0, IDCT_SHIFT2
+ mova m7, [rsp + 15 * mmsize]
+ psubd m3, [rsp + 19 * mmsize]
+ mova m5, [rsp + 16 * mmsize]
+ paddd m7, m13
+ psrad m7, IDCT_SHIFT2
+ mova [rsp + 5 * mmsize], m3
+ paddd m5, m15
+ psubd m6, [rsp + 8 * mmsize]
+ psubd m9, [rsp + 13 * mmsize]
+ psubd m8, [rsp + 27 * mmsize]
+ psubd m11, [rsp + 14 * mmsize]
+ psubd m10, [rsp + 29 * mmsize]
+ psubd m13, [rsp + 15 * mmsize]
+ psubd m12, [rsp + 21 * mmsize]
+ psubd m15, [rsp + 16 * mmsize]
+ psubd m14, [rsp + 23 * mmsize]
+ psrad m5, IDCT_SHIFT2
+ mova m3, [rsp + 6 * mmsize]
+ psrad m12, IDCT_SHIFT2
+ psrad m13, IDCT_SHIFT2
+ packssdw m13, m12
+ psrad m11, IDCT_SHIFT2
+ psrad m8, IDCT_SHIFT2
+ psubd m3, [rsp + 20 * mmsize]
+ mova [rsp + 6 * mmsize], m3
+ psrad m6, IDCT_SHIFT2
+ psrad m9, IDCT_SHIFT2
+ mova m12, m6
+ psrad m10, IDCT_SHIFT2
+ packssdw m9, m8
+ psrad m14, IDCT_SHIFT2
+ psrad m15, IDCT_SHIFT2
+ packssdw m15, m14
+ mova m3, [rsp + 7 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 7 * mmsize], m3
+ mova m14, m15
+ mova m3, [rsp + 30 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 30 * mmsize], m3
+ mova m3, [rsp + 4 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 4 * mmsize], m3
+ mova m3, [rsp + 34 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 34 * mmsize], m3
+ mova m3, [rsp + 3 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 3 * mmsize], m3
+ mova m3, [rsp + 35 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 35 * mmsize], m3
+ mova m3, [rsp + 9 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 9 * mmsize], m3
+ mova m3, [rsp + 36 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 36 * mmsize], m3
+ mova m3, [rsp + 10 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 10 * mmsize], m3
+ mova m3, [rsp + 37 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 37 * mmsize], m3
+ mova m3, [rsp + 22 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 22 * mmsize], m3
+ mova m3, [rsp + 38 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 38 * mmsize], m3
+ mova m3, [rsp + 39 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 39 * mmsize], m3
+ mova m3, [rsp + 40 * mmsize]
+ packssdw m7, [rsp + 39 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 40 * mmsize], m3
+ mova m3, [rsp + 6 * mmsize]
+ packssdw m5, [rsp + 40 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 6 * mmsize], m3
+ mova [rsp + 2 * mmsize], m7
+ mova m3, [rsp + 5 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 5 * mmsize], m3
+ mova m3, [rsp + 17 * mmsize]
+ psrad m3, IDCT_SHIFT2
+ mova [rsp + 17 * mmsize], m3
+ mova m3, [rsp + 7 * mmsize]
+ packssdw m3, [rsp + 30 * mmsize]
+ mova [rsp + 7 * mmsize], m3
+ mova m3, [rsp + 4 * mmsize]
+ packssdw m3, [rsp + 34 * mmsize]
+ mova [rsp + 4 * mmsize], m3
+ mova m3, [rsp + 3 * mmsize]
+ packssdw m3, [rsp + 35 * mmsize]
+ mova [rsp + 3 * mmsize], m3
+ mova m3, [rsp + 9 * mmsize]
+ packssdw m3, [rsp + 36 * mmsize]
+ mova [rsp + 9 * mmsize], m3
+ mova m3, [rsp + 10 * mmsize]
+ packssdw m3, [rsp + 37 * mmsize]
+ mova [rsp + 10 * mmsize], m3
+ mova m3, [rsp + 22 * mmsize]
+ packssdw m3, [rsp + 38 * mmsize]
+ mova [rsp + 22 * mmsize], m3
+ mova [rsp + 1 * mmsize], m5
+ mova [rsp + 0 * mmsize], m13
+ mova m13, m11
+ mova [rsp + 8 * mmsize], m9
+ packssdw m13, m10
+ mova m10, m2
+ packssdw m0, [rsp + 5 * mmsize]
+ mova m3, [rsp + 11 * mmsize]
+ packssdw m4, [rsp + 17 * mmsize]
+ mova m9, [rsp + 26 * mmsize]
+ packssdw m12, [rsp + 6 * mmsize]
+ mova [rsp + 6 * mmsize], m0
+ punpcklwd m3, [rsp + 12 * mmsize]
+ mova [rsp + 5 * mmsize], m4
+ packssdw m10, m1
+ mova m0, [rsp + 11 * mmsize]
+ punpcklwd m9, [rsp + 28 * mmsize]
+ mova m5, m3
+ mova m4, [rsp + 24 * mmsize]
+ punpckhwd m0, [rsp + 12 * mmsize]
+ mova m1, [rsp + 24 * mmsize]
+ punpcklwd m4, [rsp + 25 * mmsize]
+ mova m7, [rsp + 26 * mmsize]
+ punpckhwd m1, [rsp + 25 * mmsize]
+ mova m8, [rsp + 31 * mmsize]
+ mova m2, m0
+ punpckhwd m7, [rsp + 28 * mmsize]
+ mova m6, [rsp + 31 * mmsize]
+ punpcklwd m8, [rsp + 32 * mmsize]
+ mova m15, [rsp + 41 * mmsize]
+ punpckhwd m6, [rsp + 32 * mmsize]
+ mova m11, [rsp + 43 * mmsize]
+ lea r2, [r2 + r2]
+ lea r4, [r2 + r2]
+ lea r3, [r4 + r2]
+ lea r4, [r4 + r3]
+ lea r0, [r4 + r2 * 2]
+ punpckldq m5, m4
+ punpckhdq m3, m4
+ mova m4, m9
+ punpckldq m2, m1
+ punpckhdq m0, m1
+ mova m1, m7
+ punpckldq m4, m8
+ punpckldq m1, m6
+ punpckhdq m7, m6
+ mova m6, m5
+ punpckhdq m9, m8
+ punpcklqdq m6, m4
+ mova m8, m2
+ punpckhqdq m5, m4
+ mova m4, m3
+ punpcklqdq m8, m1
+ punpckhqdq m3, m9
+ punpcklqdq m4, m9
+ punpckhqdq m2, m1
+ mova m1, m0
+ punpckhqdq m0, m7
+ punpcklqdq m1, m7
+ mova m7, [rsp + 47 * mmsize]
+ punpckhwd m15, [rsp + 42 * mmsize]
+ mova [r1], m6
+ punpckhwd m11, [rsp + 44 * mmsize]
+ mova [r1 + r2], m5
+ mova [r1 + r2 * 2], m4
+ mova [r1 + r3], m3
+ mova [r1 + r2 * 4], m8
+ mova [r1 + r4], m2
+ mova m2, m15
+ mova [r1 + r3 * 2], m1
+ punpckldq m2, m11
+ mova [r1 + r0], m0
+ punpckhdq m15, m11
+ mova m3, [rsp + 41 * mmsize]
+ punpcklwd m7, [rsp + 48 * mmsize]
+ mova m0, [rsp + 43 * mmsize]
+ punpcklwd m3, [rsp + 42 * mmsize]
+ mova m8, [rsp + 45 * mmsize]
+ punpcklwd m0, [rsp + 44 * mmsize]
+ mova m1, [rsp + 45 * mmsize]
+ punpcklwd m8, [rsp + 46 * mmsize]
+ mova m5, [rsp + 47 * mmsize]
+ mova m6, m3
+ punpckhwd m1, [rsp + 46 * mmsize]
+ punpckhwd m5, [rsp + 48 * mmsize]
+ mova m4, m8
+ punpckldq m6, m0
+ punpckhdq m3, m0
+ mova m0, m1
+ punpckldq m4, m7
+ punpckldq m0, m5
+ punpckhdq m8, m7
+ mova m7, m2
+ punpckhdq m1, m5
+ mova m5, m6
+ punpcklqdq m7, m0
+ punpckhqdq m6, m4
+ punpcklqdq m5, m4
+ mova m4, m3
+ punpckhqdq m2, m0
+ mova m0, m15
+ punpcklqdq m4, m8
+ punpckhqdq m3, m8
+ punpcklqdq m0, m1
+ mova [r1 + 16], m5
+ punpckhqdq m15, m1
+ mova [r1 + r2 + 16], m6
+ mova [r1 + r2 * 2 + 16], m4
+ mova [r1 + r3 + 16], m3
+ mova [r1 + r2 * 4 + 16], m7
+ mova [r1 + r4 + 16], m2
+ mova [r1 + r3 * 2 + 16], m0
+ mova [r1 + r0 + 16], m15
+ mova m1, [rsp + 7 * mmsize]
+ mova m0, [rsp + 7 * mmsize]
+ punpcklwd m1, [rsp + 4 * mmsize]
+ mova m2, [rsp + 3 * mmsize]
+ punpckhwd m0, [rsp + 4 * mmsize]
+ mova m5, [rsp + 3 * mmsize]
+ punpcklwd m2, [rsp + 9 * mmsize]
+ mova m3, [rsp + 10 * mmsize]
+ mova m9, m1
+ punpckhwd m5, [rsp + 9 * mmsize]
+ mova m4, [rsp + 2 * mmsize]
+ mova m7, m0
+ punpcklwd m3, [rsp + 22 * mmsize]
+ mova m6, [rsp + 10 * mmsize]
+ punpcklwd m4, [rsp + 1 * mmsize]
+ mova m8, [rsp + 2 * mmsize]
+ punpckhwd m6, [rsp + 22 * mmsize]
+ punpckldq m9, m2
+ punpckldq m7, m5
+ punpckhdq m0, m5
+ mova m5, m3
+ punpckhwd m8, [rsp + 1 * mmsize]
+ lea r5, [r2 + r2 * 8]
+ lea r6, [r2 + r4 * 2]
+ lea r7, [r2 + r3 * 4]
+ lea r8, [r2 + r0 * 2]
+ punpckldq m5, m4
+ punpckhdq m1, m2
+ mova m2, m6
+ punpckhdq m3, m4
+ mova m4, m9
+ punpckhqdq m9, m5
+ punpckldq m2, m8
+ punpcklqdq m4, m5
+ mova m5, m1
+ punpckhqdq m1, m3
+ punpcklqdq m5, m3
+ mova m3, m7
+ punpckhdq m6, m8
+ punpcklqdq m3, m2
+ mova m8, m10
+ punpckhqdq m7, m2
+ mova m2, m0
+ punpckhqdq m0, m6
+ mova [r1 + r2 * 8], m4
+ punpcklqdq m2, m6
+ mova m4, m14
+ mova [r1 + r5], m9
+ punpcklwd m4, [rsp + 0 * mmsize]
+ mova [r1 + r4 * 2], m5
+ mova [r1 + r6], m1
+ mova [r1 + r3 * 4], m3
+ mova m3, m12
+ mova [r1 + r7], m7
+ mova m7, m4
+ mova [r1 + r0 * 2], m2
+ mova [r1 + r8], m0
+ mova m0, m13
+ punpckhwd m14, [rsp + 0 * mmsize]
+ punpcklwd m0, [rsp + 8 * mmsize]
+ punpckhwd m13, [rsp + 8 * mmsize]
+ mova m1, m14
+ punpcklwd m3, [rsp + 6 * mmsize]
+ punpckhwd m12, [rsp + 6 * mmsize]
+ punpcklwd m8, [rsp + 5 * mmsize]
+ mova m5, m3
+ punpckhwd m10, [rsp + 5 * mmsize]
+ punpckldq m7, m0
+ punpckhdq m4, m0
+ mova m0, m12
+ punpckldq m1, m13
+ punpckldq m5, m8
+ mova m6, m7
+ punpckldq m0, m10
+ punpckhdq m14, m13
+ mova m2, m1
+ punpckhdq m3, m8
+ punpckhdq m12, m10
+ punpcklqdq m6, m5
+ punpckhqdq m7, m5
+ mova m5, m4
+ punpcklqdq m2, m0
+ punpckhqdq m1, m0
+ mova m0, m14
+ punpcklqdq m5, m3
+ punpckhqdq m4, m3
+ punpcklqdq m0, m12
+ punpckhqdq m14, m12
+ mova [r1 + r2 * 8 + 16], m6
+ mova [r1 + r5 + 16], m7
+ mova [r1 + r4 * 2 + 16], m5
+ mova [r1 + r6 + 16], m4
+ mova [r1 + r3 * 4 + 16], m2
+ mova [r1 + r7 + 16], m1
+ mova [r1 + r0 * 2 + 16], m0
+ mova [r1 + r8 + 16], m14
+ RET
+
+%undef IDCT_SHIFT1
+%undef IDCT_SHIFT2
+%undef IDCT_ADD
+%endif
+
+;-------------------------------------------------------
; void idct8(const int16_t* src, int16_t* dst, intptr_t dstStride)
;-------------------------------------------------------
INIT_XMM ssse3
diff -r f9a54da18713 -r 592f6cd42f7f source/common/x86/dct8.h
--- a/source/common/x86/dct8.h Thu Jan 15 11:49:26 2015 +0530
+++ b/source/common/x86/dct8.h Fri Jan 16 11:08:49 2015 -0800
@@ -37,6 +37,7 @@
void x265_idct8_sse2(const int16_t* src, int16_t* dst, intptr_t dstStride);
void x265_idct8_ssse3(const int16_t* src, int16_t* dst, intptr_t dstStride);
void x265_idct8_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
+void x265_idct16_sse2(const int16_t* src, int16_t* dst, intptr_t dstStride);
void x265_idct16_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
void x265_idct32_avx2(const int16_t* src, int16_t* dst, intptr_t dstStride);
More information about the x265-devel
mailing list