annotate i386/vp3dsp_sse2.c @ 3089:072dbc669253 libavcodec

MSVC-compatible __align8/__align16 declaration patch by Steve Lhomme, steve .dot. lhomme .at. free .dot. fr
author diego
date Sun, 05 Feb 2006 13:35:17 +0000
parents 0b546eab515d
children c8c591fe26f8
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
1 /*
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
2 * Copyright (C) 2004 the ffmpeg project
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
3 *
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
4 * This library is free software; you can redistribute it and/or
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
5 * modify it under the terms of the GNU Lesser General Public
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
6 * License as published by the Free Software Foundation; either
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
7 * version 2 of the License, or (at your option) any later version.
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
8 *
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
9 * This library is distributed in the hope that it will be useful,
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
12 * Lesser General Public License for more details.
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
13 *
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
14 * You should have received a copy of the GNU Lesser General Public
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
15 * License along with this library; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2967
diff changeset
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
17 */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
18
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
19 /**
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
20 * @file vp3dsp_sse2.c
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
21 * SSE2-optimized functions cribbed from the original VP3 source code.
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
22 */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
23
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
24 #include "../dsputil.h"
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
25 #include "mmx.h"
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
26
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
27 static DECLARE_ALIGNED_16(const unsigned short, SSE2_dequant_const[]) =
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
28 {
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
29 0,65535,65535,0,0,0,0,0, // 0x0000 0000 0000 0000 0000 FFFF FFFF 0000
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
30 0,0,0,0,65535,65535,0,0, // 0x0000 0000 FFFF FFFF 0000 0000 0000 0000
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
31 65535,65535,65535,0,0,0,0,0,// 0x0000 0000 0000 0000 0000 FFFF FFFF FFFF
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
32 0,0,0,65535,0,0,0,0, // 0x0000 0000 0000 0000 FFFF 0000 0000 0000
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
33 0,0,0,65535,65535,0,0,0, // 0x0000 0000 0000 FFFF FFFF 0000 0000 0000
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
34 65535,0,0,0,0,65535,0,0, // 0x0000 0000 FFFF 0000 0000 0000 0000 FFFF
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
35 0,0,65535,65535, 0,0,0,0 // 0x0000 0000 0000 0000 FFFF FFFF 0000 0000
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
36 };
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
37
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
38 static DECLARE_ALIGNED_16(const unsigned int, eight_data[]) =
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
39 {
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
40 0x00080008,
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
41 0x00080008,
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
42 0x00080008,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
43 0x00080008
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
44 };
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
45
3089
072dbc669253 MSVC-compatible __align8/__align16 declaration
diego
parents: 3036
diff changeset
46 static DECLARE_ALIGNED_16(const unsigned short, SSE2_idct_data[7 * 8]) =
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
47 {
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
48 64277,64277,64277,64277,64277,64277,64277,64277,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
49 60547,60547,60547,60547,60547,60547,60547,60547,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
50 54491,54491,54491,54491,54491,54491,54491,54491,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
51 46341,46341,46341,46341,46341,46341,46341,46341,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
52 36410,36410,36410,36410,36410,36410,36410,36410,
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
53 25080,25080,25080,25080,25080,25080,25080,25080,
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
54 12785,12785,12785,12785,12785,12785,12785,12785
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
55 };
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
56
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
57
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
58 #define SSE2_Column_IDCT() { \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
59 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
60 movdqu_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
61 movdqu_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
62 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
63 movdqu_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
64 movdqu_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
65 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
66 pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
67 movdqu_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
68 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
69 pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
70 movdqu_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
71 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
72 pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
73 movdqu_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
74 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
75 pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
76 movdqu_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
77 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
78 /* all registers are in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
79 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
80 paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
81 paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
82 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
83 paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
84 movdqu_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
85 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
86 paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
87 movdqu_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
88 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
89 pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
90 paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
91 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
92 pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
93 movdqu_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
94 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
95 psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
96 paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
97 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
98 pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
99 movdqu_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
100 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
101 pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
102 paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
103 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
104 movdqu_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
105 pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
106 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
107 psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
108 movdqu_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
109 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
110 paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
111 movdqu_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
112 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
113 psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
114 pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
115 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
116 paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
117 pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
118 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
119 paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
120 paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
121 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
122 psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
123 paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
124 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
125 paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
126 pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
127 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
128 paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
129 movdqu_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
130 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
131 psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
132 movdqu_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
133 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
134 movdqu_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
135 pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
136 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
137 paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
138 movdqu_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
139 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
140 movdqu_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
141 movdqu_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
142 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
143 pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
144 paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
145 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
146 movdqu_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
147 psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
148 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
149 paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
150 psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
151 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
152 movdqu_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
153 pmulhw_r2r(xmm4, xmm6); /* xmm6 = (c4 - 1) * (i0 - i4) = F */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
154 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
155 paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
156 paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
157 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
158 paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
159 paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
160 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
161 pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
162 paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
163 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
164 psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
165 paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
166 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
167 movdqu_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
168 paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
169 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
170 paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
171 psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
172 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
173 paddsw_m2r(*Eight, xmm2); /* Adjust R2 and R1 before shifting */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
174 paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
175 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
176 paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
177 psraw_i2r(4, xmm2); /* xmm2 = op2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
178 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
179 psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
180 psraw_i2r(4, xmm1); /* xmm1 = op1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
181 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
182 movdqu_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
183 paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
184 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
185 movdqu_r2m(xmm2, *O(2)); /* Write out op2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
186 paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
187 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
188 movdqu_r2m(xmm1, *O(1)); /* Write out op1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
189 psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
190 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
191 paddsw_m2r(*Eight, xmm4); /* Adjust R4 and R3 before shifting */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
192 paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
193 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
194 paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
195 psraw_i2r(4, xmm4); /* xmm4 = op4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
196 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
197 psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
198 psraw_i2r(4, xmm3); /* xmm3 = op3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
199 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
200 paddsw_m2r(*Eight, xmm6); /* Adjust R6 and R5 before shifting */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
201 paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
202 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
203 paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
204 psraw_i2r(4, xmm6); /* xmm6 = op6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
205 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
206 movdqu_r2m(xmm4, *O(4)); /* Write out op4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
207 psraw_i2r(4, xmm5); /* xmm5 = op5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
208 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
209 movdqu_r2m(xmm3, *O(3)); /* Write out op3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
210 psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
211 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
212 paddsw_m2r(*Eight, xmm7); /* Adjust R7 and R0 before shifting */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
213 paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
214 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
215 paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
216 psraw_i2r(4, xmm7); /* xmm7 = op7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
217 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
218 movdqu_r2m(xmm6, *O(6)); /* Write out op6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
219 psraw_i2r(4, xmm0); /* xmm0 = op0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
220 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
221 movdqu_r2m(xmm5, *O(5)); /* Write out op5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
222 movdqu_r2m(xmm7, *O(7)); /* Write out op7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
223 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
224 movdqu_r2m(xmm0, *O(0)); /* Write out op0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
225 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
226 } /* End of SSE2_Column_IDCT macro */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
227
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
228
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
229 #define SSE2_Row_IDCT() { \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
230 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
231 movdqu_m2r(*I(3), xmm2); /* xmm2 = i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
232 movdqu_m2r(*C(3), xmm6); /* xmm6 = c3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
233 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
234 movdqu_r2r(xmm2, xmm4); /* xmm4 = i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
235 movdqu_m2r(*I(5), xmm7); /* xmm7 = i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
236 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
237 pmulhw_r2r(xmm6, xmm4); /* xmm4 = c3 * i3 - i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
238 movdqu_m2r(*C(5), xmm1); /* xmm1 = c5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
239 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
240 pmulhw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 - i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
241 movdqu_r2r(xmm1, xmm5); /* xmm5 = c5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
242 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
243 pmulhw_r2r(xmm2, xmm1); /* xmm1 = c5 * i3 - i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
244 movdqu_m2r(*I(1), xmm3); /* xmm3 = i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
245 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
246 pmulhw_r2r(xmm7, xmm5); /* xmm5 = c5 * i5 - i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
247 movdqu_m2r(*C(1), xmm0); /* xmm0 = c1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
248 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
249 /* all registers are in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
250 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
251 paddw_r2r(xmm2, xmm4); /* xmm4 = c3 * i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
252 paddw_r2r(xmm7, xmm6); /* xmm6 = c3 * i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
253 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
254 paddw_r2r(xmm1, xmm2); /* xmm2 = c5 * i3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
255 movdqu_m2r(*I(7), xmm1); /* xmm1 = i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
256 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
257 paddw_r2r(xmm5, xmm7); /* xmm7 = c5 * i5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
258 movdqu_r2r(xmm0, xmm5); /* xmm5 = c1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
259 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
260 pmulhw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 - i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
261 paddsw_r2r(xmm7, xmm4); /* xmm4 = c3 * i3 + c5 * i5 = C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
262 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
263 pmulhw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 - i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
264 movdqu_m2r(*C(7), xmm7); /* xmm7 = c7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
265 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
266 psubsw_r2r(xmm2, xmm6); /* xmm6 = c3 * i5 - c5 * i3 = D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
267 paddw_r2r(xmm3, xmm0); /* xmm0 = c1 * i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
268 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
269 pmulhw_r2r(xmm7, xmm3); /* xmm3 = c7 * i1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
270 movdqu_m2r(*I(2), xmm2); /* xmm2 = i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
271 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
272 pmulhw_r2r(xmm1, xmm7); /* xmm7 = c7 * i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
273 paddw_r2r(xmm1, xmm5); /* xmm5 = c1 * i7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
274 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
275 movdqu_r2r(xmm2, xmm1); /* xmm1 = i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
276 pmulhw_m2r(*C(2), xmm2); /* xmm2 = i2 * c2 -i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
277 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
278 psubsw_r2r(xmm5, xmm3); /* xmm3 = c7 * i1 - c1 * i7 = B */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
279 movdqu_m2r(*I(6), xmm5); /* xmm5 = i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
280 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
281 paddsw_r2r(xmm7, xmm0); /* xmm0 = c1 * i1 + c7 * i7 = A */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
282 movdqu_r2r(xmm5, xmm7); /* xmm7 = i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
283 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
284 psubsw_r2r(xmm4, xmm0); /* xmm0 = A - C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
285 pmulhw_m2r(*C(2), xmm5); /* xmm5 = c2 * i6 - i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
286 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
287 paddw_r2r(xmm1, xmm2); /* xmm2 = i2 * c2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
288 pmulhw_m2r(*C(6), xmm1); /* xmm1 = c6 * i2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
289 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
290 paddsw_r2r(xmm4, xmm4); /* xmm4 = C + C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
291 paddsw_r2r(xmm0, xmm4); /* xmm4 = A + C = C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
292 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
293 psubsw_r2r(xmm6, xmm3); /* xmm3 = B - D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
294 paddw_r2r(xmm7, xmm5); /* xmm5 = c2 * i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
295 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
296 paddsw_r2r(xmm6, xmm6); /* xmm6 = D + D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
297 pmulhw_m2r(*C(6), xmm7); /* xmm7 = c6 * i6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
298 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
299 paddsw_r2r(xmm3, xmm6); /* xmm6 = B + D = D. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
300 movdqu_r2m(xmm4, *I(1)); /* Save C. at I(1) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
301 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
302 psubsw_r2r(xmm5, xmm1); /* xmm1 = c6 * i2 - c2 * i6 = H */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
303 movdqu_m2r(*C(4), xmm4); /* xmm4 = c4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
304 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
305 movdqu_r2r(xmm3, xmm5); /* xmm5 = B - D */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
306 pmulhw_r2r(xmm4, xmm3); /* xmm3 = ( c4 -1 ) * ( B - D ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
307 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
308 paddsw_r2r(xmm2, xmm7); /* xmm7 = c2 * i2 + c6 * i6 = G */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
309 movdqu_r2m(xmm6, *I(2)); /* Save D. at I(2) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
310 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
311 movdqu_r2r(xmm0, xmm2); /* xmm2 = A - C */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
312 movdqu_m2r(*I(0), xmm6); /* xmm6 = i0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
313 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
314 pmulhw_r2r(xmm4, xmm0); /* xmm0 = ( c4 - 1 ) * ( A - C ) = A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
315 paddw_r2r(xmm3, xmm5); /* xmm5 = c4 * ( B - D ) = B. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
316 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
317 movdqu_m2r(*I(4), xmm3); /* xmm3 = i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
318 psubsw_r2r(xmm1, xmm5); /* xmm5 = B. - H = B.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
319 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
320 paddw_r2r(xmm0, xmm2); /* xmm2 = c4 * ( A - C) = A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
321 psubsw_r2r(xmm3, xmm6); /* xmm6 = i0 - i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
322 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
323 movdqu_r2r(xmm6, xmm0); /* xmm0 = i0 - i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
324 pmulhw_r2r(xmm4, xmm6); /* xmm6 = ( c4 - 1 ) * ( i0 - i4 ) = F */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
325 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
326 paddsw_r2r(xmm3, xmm3); /* xmm3 = i4 + i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
327 paddsw_r2r(xmm1, xmm1); /* xmm1 = H + H */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
328 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
329 paddsw_r2r(xmm0, xmm3); /* xmm3 = i0 + i4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
330 paddsw_r2r(xmm5, xmm1); /* xmm1 = B. + H = H. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
331 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
332 pmulhw_r2r(xmm3, xmm4); /* xmm4 = ( c4 - 1 ) * ( i0 + i4 ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
333 paddw_r2r(xmm0, xmm6); /* xmm6 = c4 * ( i0 - i4 ) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
334 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
335 psubsw_r2r(xmm2, xmm6); /* xmm6 = F - A. = F. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
336 paddsw_r2r(xmm2, xmm2); /* xmm2 = A. + A. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
337 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
338 movdqu_m2r(*I(1), xmm0); /* Load C. from I(1) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
339 paddsw_r2r(xmm6, xmm2); /* xmm2 = F + A. = A.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
340 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
341 paddw_r2r(xmm3, xmm4); /* xmm4 = c4 * ( i0 + i4 ) = 3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
342 psubsw_r2r(xmm1, xmm2); /* xmm2 = A.. - H. = R2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
343 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
344 paddsw_r2r(xmm1, xmm1); /* xmm1 = H. + H. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
345 paddsw_r2r(xmm2, xmm1); /* xmm1 = A.. + H. = R1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
346 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
347 psubsw_r2r(xmm7, xmm4); /* xmm4 = E - G = E. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
348 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
349 movdqu_m2r(*I(2), xmm3); /* Load D. from I(2) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
350 paddsw_r2r(xmm7, xmm7); /* xmm7 = G + G */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
351 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
352 movdqu_r2m(xmm2, *I(2)); /* Write out op2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
353 paddsw_r2r(xmm4, xmm7); /* xmm7 = E + G = G. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
354 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
355 movdqu_r2m(xmm1, *I(1)); /* Write out op1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
356 psubsw_r2r(xmm3, xmm4); /* xmm4 = E. - D. = R4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
357 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
358 paddsw_r2r(xmm3, xmm3); /* xmm3 = D. + D. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
359 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
360 paddsw_r2r(xmm4, xmm3); /* xmm3 = E. + D. = R3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
361 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
362 psubsw_r2r(xmm5, xmm6); /* xmm6 = F. - B..= R6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
363 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
364 paddsw_r2r(xmm5, xmm5); /* xmm5 = B.. + B.. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
365 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
366 paddsw_r2r(xmm6, xmm5); /* xmm5 = F. + B.. = R5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
367 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
368 movdqu_r2m(xmm4, *I(4)); /* Write out op4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
369 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
370 movdqu_r2m(xmm3, *I(3)); /* Write out op3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
371 psubsw_r2r(xmm0, xmm7); /* xmm7 = G. - C. = R7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
372 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
373 paddsw_r2r(xmm0, xmm0); /* xmm0 = C. + C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
374 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
375 paddsw_r2r(xmm7, xmm0); /* xmm0 = G. + C. */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
376 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
377 movdqu_r2m(xmm6, *I(6)); /* Write out op6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
378 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
379 movdqu_r2m(xmm5, *I(5)); /* Write out op5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
380 movdqu_r2m(xmm7, *I(7)); /* Write out op7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
381 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
382 movdqu_r2m(xmm0, *I(0)); /* Write out op0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
383 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
384 } /* End of SSE2_Row_IDCT macro */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
385
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
386
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
387 #define SSE2_Transpose() { \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
388 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
389 movdqu_m2r(*I(4), xmm4); /* xmm4=e7e6e5e4e3e2e1e0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
390 movdqu_m2r(*I(5), xmm0); /* xmm4=f7f6f5f4f3f2f1f0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
391 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
392 movdqu_r2r(xmm4, xmm5); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
393 punpcklwd_r2r(xmm0, xmm4); /* xmm4=f3e3f2e2f1e1f0e0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
394 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
395 punpckhwd_r2r(xmm0, xmm5); /* xmm5=f7e7f6e6f5e5f4e4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
396 movdqu_m2r(*I(6), xmm6); /* xmm6=g7g6g5g4g3g2g1g0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
397 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
398 movdqu_m2r(*I(7), xmm0); /* xmm0=h7h6h5h4h3h2h1h0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
399 movdqu_r2r(xmm6, xmm7); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
400 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
401 punpcklwd_r2r(xmm0, xmm6); /* xmm6=h3g3h3g2h1g1h0g0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
402 punpckhwd_r2r(xmm0, xmm7); /* xmm7=h7g7h6g6h5g5h4g4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
403 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
404 movdqu_r2r(xmm4, xmm3); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
405 punpckldq_r2r(xmm6, xmm4); /* xmm4=h1g1f1e1h0g0f0e0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
406 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
407 punpckhdq_r2r(xmm6, xmm3); /* xmm3=h3g3g3e3h2g2f2e2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
408 movdqu_r2m(xmm3, *I(6)); /* save h3g3g3e3h2g2f2e2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
409 /* Free xmm6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
410 movdqu_r2r(xmm5, xmm6); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
411 punpckldq_r2r(xmm7, xmm5); /* xmm5=h5g5f5e5h4g4f4e4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
412 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
413 punpckhdq_r2r(xmm7, xmm6); /* xmm6=h7g7f7e7h6g6f6e6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
414 movdqu_m2r(*I(0), xmm0); /* xmm0=a7a6a5a4a3a2a1a0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
415 /* Free xmm7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
416 movdqu_m2r(*I(1), xmm1); /* xmm1=b7b6b5b4b3b2b1b0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
417 movdqu_r2r(xmm0, xmm7); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
418 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
419 punpcklwd_r2r(xmm1, xmm0); /* xmm0=b3a3b2a2b1a1b0a0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
420 punpckhwd_r2r(xmm1, xmm7); /* xmm7=b7a7b6a6b5a5b4a4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
421 /* Free xmm1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
422 movdqu_m2r(*I(2), xmm2); /* xmm2=c7c6c5c4c3c2c1c0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
423 movdqu_m2r(*I(3), xmm3); /* xmm3=d7d6d5d4d3d2d1d0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
424 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
425 movdqu_r2r(xmm2, xmm1); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
426 punpcklwd_r2r(xmm3, xmm2); /* xmm2=d3c3d2c2d1c1d0c0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
427 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
428 punpckhwd_r2r(xmm3, xmm1); /* xmm1=d7c7d6c6d5c5d4c4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
429 movdqu_r2r(xmm0, xmm3); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
430 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
431 punpckldq_r2r(xmm2, xmm0); /* xmm0=d1c1b1a1d0c0b0a0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
432 punpckhdq_r2r(xmm2, xmm3); /* xmm3=d3c3b3a3d2c2b2a2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
433 /* Free xmm2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
434 movdqu_r2r(xmm7, xmm2); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
435 punpckldq_r2r(xmm1, xmm2); /* xmm2=d5c5b5a5d4c4b4a4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
436 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
437 punpckhdq_r2r(xmm1, xmm7); /* xmm7=d7c7b7a7d6c6b6a6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
438 movdqu_r2r(xmm0, xmm1); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
439 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
440 punpcklqdq_r2r(xmm4, xmm0); /* xmm0=h0g0f0e0d0c0b0a0 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
441 punpckhqdq_r2r(xmm4, xmm1); /* xmm1=h1g1g1e1d1c1b1a1 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
442 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
443 movdqu_r2m(xmm0, *I(0)); /* save I(0) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
444 movdqu_r2m(xmm1, *I(1)); /* save I(1) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
445 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
446 movdqu_m2r(*I(6), xmm0); /* load h3g3g3e3h2g2f2e2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
447 movdqu_r2r(xmm3, xmm1); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
448 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
449 punpcklqdq_r2r(xmm0, xmm1); /* xmm1=h2g2f2e2d2c2b2a2 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
450 punpckhqdq_r2r(xmm0, xmm3); /* xmm3=h3g3f3e3d3c3b3a3 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
451 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
452 movdqu_r2r(xmm2, xmm4); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
453 punpcklqdq_r2r(xmm5, xmm4); /* xmm4=h4g4f4e4d4c4b4a4 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
454 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
455 punpckhqdq_r2r(xmm5, xmm2); /* xmm2=h5g5f5e5d5c5b5a5 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
456 movdqu_r2m(xmm1, *I(2)); /* save I(2) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
457 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
458 movdqu_r2m(xmm3, *I(3)); /* save I(3) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
459 movdqu_r2m(xmm4, *I(4)); /* save I(4) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
460 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
461 movdqu_r2m(xmm2, *I(5)); /* save I(5) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
462 movdqu_r2r(xmm7, xmm5); /* make a copy */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
463 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
464 punpcklqdq_r2r(xmm6, xmm5); /* xmm5=h6g6f6e6d6c6b6a6 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
465 punpckhqdq_r2r(xmm6, xmm7); /* xmm7=h7g7f7e7d7c7b7a7 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
466 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
467 movdqu_r2m(xmm5, *I(6)); /* save I(6) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
468 movdqu_r2m(xmm7, *I(7)); /* save I(7) */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
469 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
470 } /* End of Transpose Macro */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
471
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
472
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
473 #define SSE2_Dequantize() { \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
474 movdqu_m2r(*(eax), xmm0); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
475 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
476 pmullw_m2r(*(ebx), xmm0); /* xmm0 = 07 06 05 04 03 02 01 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
477 movdqu_m2r(*(eax + 16), xmm1); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
478 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
479 pmullw_m2r(*(ebx + 16), xmm1); /* xmm1 = 17 16 15 14 13 12 11 10 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
480 pshuflw_r2r(xmm0, xmm3, 0x078); /* xmm3 = 07 06 05 04 01 03 02 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
481 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
482 movdqu_r2r(xmm1, xmm2); /* xmm2 = 17 16 15 14 13 12 11 10 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
483 movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
484 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
485 movdqu_m2r(*(eax + 32), xmm4); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
486 movdqu_m2r(*(eax + 64), xmm5); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
487 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
488 pmullw_m2r(*(ebx + 32), xmm4); /* xmm4 = 27 26 25 24 23 22 21 20 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
489 pmullw_m2r(*(ebx + 64), xmm5); /* xmm5 = 47 46 45 44 43 42 41 40 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
490 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
491 movdqu_m2r(*(ecx + 16), xmm6); /* xmm6 = -- -- FF FF -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
492 pand_r2r(xmm2, xmm7); /* xmm7 = -- -- -- -- -- 12 11 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
493 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
494 pand_r2r(xmm4, xmm6); /* xmm6 = -- -- 25 24 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
495 pxor_r2r(xmm7, xmm2); /* xmm2 = 17 16 15 14 13 -- -- 10 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
496 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
497 pxor_r2r(xmm6, xmm4); /* xmm4 = 27 26 -- -- 23 22 21 20 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
498 pslldq_i2r(4, xmm7); /* xmm7 = -- -- -- 12 11 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
499 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
500 pslldq_i2r(2, xmm6); /* xmm6 = -- 25 24 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
501 por_r2r(xmm6, xmm7); /* xmm7 = -- 25 24 12 11 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
502 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
503 movdqu_m2r(*(ecx + 32), xmm0); /* xmm0 = -- -- -- -- -- FF FF FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
504 movdqu_m2r(*(ecx + 48), xmm6); /* xmm6 = -- -- -- -- FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
505 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
506 pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- -- -- 03 02 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
507 pand_r2r(xmm5, xmm6); /* xmm6 = -- -- -- -- 43 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
508 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
509 pxor_r2r(xmm0, xmm3); /* xmm3 = 07 06 05 04 01 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
510 pxor_r2r(xmm6, xmm5); /* xmm5 = 47 46 45 44 -- 42 41 40 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
511 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
512 por_r2r(xmm7, xmm0); /* xmm0 = -- 25 24 12 11 03 02 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
513 pslldq_i2r(8, xmm6); /* xmm6 = 43 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
514 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
515 por_r2r(xmm6, xmm0); /* xmm0 = 43 25 24 12 11 03 02 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
516 /* 02345 in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
517 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
518 movdqu_m2r(*(ecx + 64 ), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
519 pshuflw_r2r(xmm5, xmm5, 0x0B4); /* xmm5 = 47 46 45 44 42 -- 41 40 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
520 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
521 movdqu_r2r(xmm1, xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
522 movdqu_r2r(xmm1, xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
523 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
524 movdqu_r2m(xmm0, *(eax)); /* write 43 25 24 12 11 03 02 00 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
525 pshufhw_r2r(xmm4, xmm4, 0x0C2); /* xmm4 = 27 -- -- 26 23 22 21 20 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
526 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
527 pand_r2r(xmm4, xmm7); /* xmm7 = -- -- -- 26 23 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
528 pand_r2r(xmm5, xmm1); /* xmm1 = -- -- -- 44 42 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
529 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
530 pxor_r2r(xmm7, xmm4); /* xmm4 = 27 -- -- -- -- 22 21 20 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
531 pxor_r2r(xmm1, xmm5); /* xmm5 = 47 46 45 -- -- -- 41 40 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
532 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
533 pshuflw_r2r(xmm2, xmm2, 0x0C6); /* xmm2 = 17 16 15 14 13 10 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
534 movdqu_r2r(xmm6, xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
535 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
536 pslldq_i2r(2, xmm7); /* xmm7 = -- -- 26 23 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
537 pslldq_i2r(6, xmm1); /* xmm1 = 44 42 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
538 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
539 psrldq_i2r(2, xmm0); /* xmm0 = -- -- -- -- FF FF -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
540 pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- 04 01 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
541 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
542 pand_r2r(xmm2, xmm0); /* xmm0 = -- -- -- -- 13 10 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
543 pxor_r2r(xmm6, xmm3); /* xmm3 = 07 06 05 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
544 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
545 pxor_r2r(xmm0, xmm2); /* xmm2 = 17 16 15 14 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
546 psrldq_i2r(6, xmm6); /* xmm0 = -- -- -- -- -- -- 04 01 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
547 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
548 por_r2r(xmm7, xmm1); /* xmm1 = 44 42 26 23 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
549 por_r2r(xmm6, xmm0); /* xmm1 = -- -- -- -- 13 10 04 01 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
550 /* 12345 in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
551 por_r2r(xmm0, xmm1); /* xmm1 = 44 42 26 23 13 10 04 01 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
552 pshuflw_r2r(xmm4, xmm4, 0x093); /* xmm4 = 27 -- -- -- 22 21 20 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
553 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
554 pshufhw_r2r(xmm4, xmm4, 0x093); /* xmm4 = -- -- -- 27 22 21 20 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
555 movdqu_r2m(xmm1, *(eax + 16)); /* write 44 42 26 23 13 10 04 01 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
556 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
557 pshufhw_r2r(xmm3, xmm3, 0x0D2); /* xmm3 = 07 05 -- 06 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
558 movdqu_m2r(*(ecx + 64), xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
559 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
560 pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- 06 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
561 psrldq_i2r(12, xmm3); /* xmm3 = -- -- -- -- -- -- 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
562 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
563 psrldq_i2r(8, xmm0); /* xmm0 = -- -- -- -- -- -- -- 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
564 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
565 movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
566 movdqu_m2r(*(ecx + 96), xmm7); /* xmm7 = -- -- -- -- FF FF -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
567 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
568 pand_r2r(xmm4, xmm6); /* xmm6 = -- -- -- 27 22 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
569 pxor_r2r(xmm6, xmm4); /* xmm4 = -- -- -- -- -- 21 20 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
570 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
571 por_r2r(xmm6, xmm3); /* xmm3 = -- -- -- 27 22 -- 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
572 pand_r2r(xmm4, xmm7); /* xmm7 = -- -- -- -- -- 21 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
573 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
574 por_r2r(xmm7, xmm0); /* xmm0 = -- -- -- -- -- 21 -- 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
575 pxor_r2r(xmm7, xmm4); /* xmm4 = -- -- -- -- -- -- 20 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
576 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
577 movdqu_m2r(*(ecx + 16 ), xmm6); /* xmm6 = -- -- FF FF -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
578 movdqu_m2r(*(ecx + 64 ), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
579 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
580 pand_r2r(xmm2, xmm6); /* xmm6 = -- -- 15 14 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
581 pand_r2r(xmm6, xmm1); /* xmm1 = -- -- -- 14 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
582 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
583 pxor_r2r(xmm6, xmm2); /* xmm2 = 17 16 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
584 pxor_r2r(xmm1, xmm6); /* xmm6 = -- -- 15 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
585 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
586 psrldq_i2r(4, xmm1); /* xmm1 = -- -- -- -- -- 14 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
587 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
588 psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- 15 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
589 por_r2r(xmm1, xmm3); /* xmm3 = -- -- -- 27 22 14 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
590 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
591 por_r2r(xmm6, xmm0); /* xmm0 = -- -- -- -- -- 21 15 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
592 pshufhw_r2r(xmm5, xmm5, 0x0E1); /* xmm5 = 47 46 -- 45 -- -- 41 40 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
593 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
594 movdqu_m2r(*(ecx + 64), xmm1); /* xmm1 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
595 pshuflw_r2r(xmm5, xmm5, 0x072); /* xmm5 = 47 46 -- 45 41 -- 40 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
596 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
597 movdqu_r2r(xmm1, xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
598 pand_r2r(xmm5, xmm1); /* xmm1 = -- -- -- 45 41 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
599 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
600 pxor_r2r(xmm1, xmm5); /* xmm5 = 47 46 -- -- -- -- 40 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
601 pslldq_i2r(4, xmm1); /* xmm1 = -- 45 41 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
602 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
603 pshufd_r2r(xmm5, xmm5, 0x09C); /* xmm5 = -- -- -- -- 47 46 40 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
604 por_r2r(xmm1, xmm3); /* xmm3 = -- 45 41 27 22 14 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
605 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
606 movdqu_m2r(*(eax + 96), xmm1); /* xmm1 = 67 66 65 64 63 62 61 60 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
607 pmullw_m2r(*(ebx + 96), xmm1); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
608 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
609 movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
610 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
611 psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
612 pand_r2r(xmm5, xmm7); /* xmm7 = -- -- -- -- -- 46 40 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
613 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
614 pand_r2r(xmm1, xmm6); /* xmm6 = -- -- -- -- -- -- -- 60 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
615 pxor_r2r(xmm7, xmm5); /* xmm5 = -- -- -- -- 47 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
616 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
617 pxor_r2r(xmm6, xmm1); /* xmm1 = 67 66 65 64 63 62 61 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
618 pslldq_i2r(2, xmm5); /* xmm5 = -- -- -- 47 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
619 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
620 pslldq_i2r(14, xmm6); /* xmm6 = 60 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
621 por_r2r(xmm5, xmm4); /* xmm4 = -- -- -- 47 -- -- 20 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
622 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
623 por_r2r(xmm6, xmm3); /* xmm3 = 60 45 41 27 22 14 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
624 pslldq_i2r(6, xmm7); /* xmm7 = -- -- 46 40 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
625 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
626 movdqu_r2m(xmm3, *(eax+32)); /* write 60 45 41 27 22 14 07 05 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
627 por_r2r(xmm7, xmm0); /* xmm0 = -- -- 46 40 -- 21 15 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
628 /* 0, 1, 2, 4 in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
629 movdqu_m2r(*(eax + 48), xmm3); /* xmm3 = 37 36 35 34 33 32 31 30 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
630 movdqu_m2r(*(eax + 80), xmm5); /* xmm5 = 57 56 55 54 53 52 51 50 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
631 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
632 pmullw_m2r(*(ebx + 48), xmm3); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
633 pmullw_m2r(*(ebx + 80), xmm5); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
634 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
635 movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
636 movdqu_m2r(*(ecx + 64), xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
637 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
638 psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
639 pslldq_i2r(8, xmm7); /* xmm7 = FF -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
640 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
641 pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- -- -- -- -- 30 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
642 pand_r2r(xmm5, xmm7); /* xmm7 = 57 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
643 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
644 pxor_r2r(xmm6, xmm3); /* xmm3 = 37 36 35 34 33 32 31 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
645 pxor_r2r(xmm7, xmm5); /* xmm5 = __ 56 55 54 53 52 51 50 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
646 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
647 pslldq_i2r(6, xmm6); /* xmm6 = -- -- -- -- 30 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
648 psrldq_i2r(2, xmm7); /* xmm7 = -- 57 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
649 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
650 por_r2r(xmm7, xmm6); /* xmm6 = -- 57 -- -- 30 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
651 movdqu_m2r(*(ecx), xmm7); /* xmm7 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
652 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
653 por_r2r(xmm6, xmm0); /* xmm0 = -- 57 46 40 30 21 15 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
654 psrldq_i2r(2, xmm7); /* xmm7 = -- -- -- -- -- -- FF FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
655 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
656 movdqu_r2r(xmm2, xmm6); /* xmm6 = 17 16 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
657 pand_r2r(xmm1, xmm7); /* xmm7 = -- -- -- -- -- -- 61 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
658 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
659 pslldq_i2r(2, xmm6); /* xmm6 = 16 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
660 psrldq_i2r(14, xmm2); /* xmm2 = -- -- -- -- -- -- -- 17 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
661 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
662 pxor_r2r(xmm7, xmm1); /* xmm1 = 67 66 65 64 63 62 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
663 pslldq_i2r(12, xmm7); /* xmm7 = 61 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
664 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
665 psrldq_i2r(14, xmm6); /* xmm6 = -- -- -- -- -- -- -- 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
666 por_r2r(xmm6, xmm4); /* xmm4 = -- -- -- 47 -- -- 20 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
667 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
668 por_r2r(xmm7, xmm0); /* xmm0 = 61 57 46 40 30 21 15 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
669 movdqu_m2r(*(ecx), xmm6); /* xmm6 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
670 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
671 psrldq_i2r(2, xmm6); /* xmm6 = -- -- -- -- -- -- FF FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
672 movdqu_r2m(xmm0, *(eax+48)); /* write 61 57 46 40 30 21 15 06 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
673 /* 1, 2, 3, 4, 5 in use */\
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
674 movdqu_m2r(*(ecx), xmm0); /* xmm0 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
675 pand_r2r(xmm3, xmm6); /* xmm6 = -- -- -- -- -- -- 31 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
676 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
677 movdqu_r2r(xmm3, xmm7); /* xmm7 = 37 36 35 34 33 32 31 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
678 pxor_r2r(xmm6, xmm3); /* xmm3 = 37 36 35 34 33 32 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
679 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
680 pslldq_i2r(2, xmm3); /* xmm3 = 36 35 34 33 32 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
681 pand_r2r(xmm1, xmm0); /* xmm0 = -- -- -- -- -- 62 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
682 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
683 psrldq_i2r(14, xmm7); /* xmm7 = -- -- -- -- -- -- -- 37 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
684 pxor_r2r(xmm0, xmm1); /* xmm1 = 67 66 65 64 63 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
685 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
686 por_r2r(xmm7, xmm6); /* xmm6 = -- -- -- -- -- -- 31 37 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
687 movdqu_m2r(*(ecx + 64), xmm7); /* xmm7 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
688 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
689 pshuflw_r2r(xmm6, xmm6, 0x01E); /* xmm6 = -- -- -- -- 37 31 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
690 pslldq_i2r(6, xmm7); /* xmm7 = FF FF -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
691 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
692 por_r2r(xmm6, xmm4); /* xmm4 = -- -- -- 47 37 31 20 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
693 pand_r2r(xmm5, xmm7); /* xmm7 = -- 56 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
694 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
695 pslldq_i2r(8, xmm0); /* xmm0 = -- 62 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
696 pxor_r2r(xmm7, xmm5); /* xmm5 = -- -- 55 54 53 52 51 50 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
697 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
698 psrldq_i2r(2, xmm7); /* xmm7 = -- -- 56 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
699 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
700 pshufhw_r2r(xmm3, xmm3, 0x087); /* xmm3 = 35 33 34 36 32 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
701 por_r2r(xmm7, xmm0); /* xmm0 = -- 62 56 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
702 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
703 movdqu_m2r(*(eax + 112), xmm7); /* xmm7 = 77 76 75 74 73 72 71 70 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
704 pmullw_m2r(*(ebx + 112), xmm7); \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
705 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
706 movdqu_m2r(*(ecx + 64), xmm6); /* xmm6 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
707 por_r2r(xmm0, xmm4); /* xmm4 = -- 62 56 47 37 31 20 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
708 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
709 pshuflw_r2r(xmm7, xmm7, 0x0E1); /* xmm7 = 77 76 75 74 73 72 70 71 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
710 psrldq_i2r(8, xmm6); /* xmm6 = -- -- -- -- -- -- -- FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
711 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
712 movdqu_m2r(*(ecx + 64), xmm0); /* xmm0 = -- -- -- FF FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
713 pand_r2r(xmm7, xmm6); /* xmm6 = -- -- -- -- -- -- -- 71 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
714 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
715 pand_r2r(xmm3, xmm0); /* xmm0 = -- -- -- 36 32 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
716 pxor_r2r(xmm6, xmm7); /* xmm7 = 77 76 75 74 73 72 70 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
717 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
718 pxor_r2r(xmm0, xmm3); /* xmm3 = 35 33 34 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
719 pslldq_i2r(14, xmm6); /* xmm6 = 71 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
720 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
721 psrldq_i2r(4, xmm0); /* xmm0 = -- -- -- -- -- 36 32 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
722 por_r2r(xmm6, xmm4); /* xmm4 = 71 62 56 47 37 31 20 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
723 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
724 por_r2r(xmm0, xmm2); /* xmm2 = -- -- -- -- -- 36 32 17 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
725 movdqu_r2m(xmm4, *(eax + 64)); /* write 71 62 56 47 37 31 20 16 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
726 /* 1, 2, 3, 5, 7 in use */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
727 movdqu_m2r(*(ecx + 80), xmm6); /* xmm6 = -- -- FF -- -- -- -- FF */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
728 pshufhw_r2r(xmm7, xmm7, 0x0D2); /* xmm7 = 77 75 74 76 73 72 70 __ */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
729 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
730 movdqu_m2r(*(ecx), xmm4); /* xmm4 = -- -- -- -- -- FF FF -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
731 movdqu_m2r(*(ecx+48), xmm0); /* xmm0 = -- -- -- -- FF -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
732 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
733 pand_r2r(xmm5, xmm6); /* xmm6 = -- -- 55 -- -- -- -- 50 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
734 pand_r2r(xmm7, xmm4); /* xmm4 = -- -- -- -- -- 72 70 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
735 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
736 pand_r2r(xmm1, xmm0); /* xmm0 = -- -- -- -- 63 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
737 pxor_r2r(xmm6, xmm5); /* xmm5 = -- -- -- 54 53 52 51 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
738 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
739 pxor_r2r(xmm4, xmm7); /* xmm7 = 77 75 74 76 73 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
740 pxor_r2r(xmm0, xmm1); /* xmm1 = 67 66 65 64 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
741 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
742 pshuflw_r2r(xmm6, xmm6, 0x02B); /* xmm6 = -- -- 55 -- 50 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
743 pslldq_i2r(10, xmm4); /* xmm4 = 72 20 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
744 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
745 pshufhw_r2r(xmm6, xmm6, 0x0B1); /* xmm6 = -- -- -- 55 50 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
746 pslldq_i2r(4, xmm0); /* xmm0 = -- -- 63 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
747 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
748 por_r2r(xmm4, xmm6); /* xmm6 = 72 70 -- 55 50 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
749 por_r2r(xmm0, xmm2); /* xmm2 = -- -- 63 -- -- 36 32 17 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
750 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
751 por_r2r(xmm6, xmm2); /* xmm2 = 72 70 64 55 50 36 32 17 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
752 pshufhw_r2r(xmm1, xmm1, 0x0C9); /* xmm1 = 67 64 66 65 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
753 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
754 movdqu_r2r(xmm3, xmm6); /* xmm6 = 35 33 34 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
755 movdqu_r2m(xmm2, *(eax+80)); /* write 72 70 64 55 50 36 32 17 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
756 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
757 psrldq_i2r(12, xmm6); /* xmm6 = -- -- -- -- -- -- 35 33 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
758 pslldq_i2r(4, xmm3); /* xmm3 = 34 -- -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
759 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
760 pshuflw_r2r(xmm5, xmm5, 0x04E); /* xmm5 = -- -- -- 54 51 -- 53 52 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
761 movdqu_r2r(xmm7, xmm4); /* xmm4 = 77 75 74 76 73 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
762 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
763 movdqu_r2r(xmm5, xmm2); /* xmm2 = -- -- -- 54 51 -- 53 52 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
764 psrldq_i2r(10, xmm7); /* xmm7 = -- -- -- -- -- 77 75 74 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
765 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
766 pslldq_i2r(6, xmm4); /* xmm4 = 76 73 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
767 pslldq_i2r(12, xmm2); /* xmm2 = 53 52 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
768 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
769 movdqu_r2r(xmm1, xmm0); /* xmm0 = 67 64 66 65 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
770 psrldq_i2r(12, xmm1); /* xmm1 = -- -- -- -- -- -- 67 64 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
771 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
772 psrldq_i2r(6, xmm5); /* xmm5 = -- -- -- -- -- -- 54 51 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
773 psrldq_i2r(14, xmm3); /* xmm3 = -- -- -- -- -- -- -- 34 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
774 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
775 pslldq_i2r(10, xmm7); /* xmm7 = 77 75 74 -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
776 por_r2r(xmm6, xmm4); /* xmm4 = 76 73 -- -- -- -- 35 33 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
777 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
778 psrldq_i2r(10, xmm2); /* xmm2 = -- -- -- -- -- 53 52 -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
779 pslldq_i2r(4, xmm0); /* xmm0 = 66 65 -- -- -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
780 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
781 pslldq_i2r(8, xmm1); /* xmm1 = -- -- 67 64 -- -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
782 por_r2r(xmm7, xmm3); /* xmm3 = 77 75 74 -- -- -- -- 34 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
783 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
784 psrldq_i2r(6, xmm0); /* xmm0 = -- -- -- 66 65 -- -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
785 pslldq_i2r(4, xmm5); /* xmm5 = -- -- -- -- 54 51 -- -- */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
786 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
787 por_r2r(xmm1, xmm4); /* xmm4 = 76 73 67 64 -- -- 35 33 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
788 por_r2r(xmm2, xmm3); /* xmm3 = 77 75 74 -- -- 53 52 34 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
789 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
790 por_r2r(xmm5, xmm4); /* xmm4 = 76 73 67 64 54 51 35 33 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
791 por_r2r(xmm0, xmm3); /* xmm3 = 77 75 74 66 65 53 52 34 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
792 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
793 movdqu_r2m(xmm4, *(eax+96)); /* write 76 73 67 64 54 51 35 33 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
794 movdqu_r2m(xmm3, *(eax+112)); /* write 77 75 74 66 65 53 52 34 */ \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
795 \
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
796 } /* end of SSE2_Dequantize Macro */
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
797
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
798
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1978
diff changeset
799 void ff_vp3_idct_sse2(int16_t *input_data)
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
800 {
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
801 unsigned char *input_bytes = (unsigned char *)input_data;
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1978
diff changeset
802 unsigned char *output_data_bytes = (unsigned char *)input_data;
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
803 unsigned char *idct_data_bytes = (unsigned char *)SSE2_idct_data;
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
804 unsigned char *Eight = (unsigned char *)eight_data;
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
805
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
806 #define eax input_bytes
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1978
diff changeset
807 //#define ebx dequant_matrix_bytes
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
808 #define ecx dequant_const_bytes
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
809 #define edx idct_data_bytes
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
810
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
811 #define I(i) (eax + 16 * i)
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
812 #define O(i) (ebx + 16 * i)
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
813 #define C(i) (edx + 16 * (i-1))
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
814
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1978
diff changeset
815 // SSE2_Dequantize();
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
816
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
817 #undef ebx
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
818 #define ebx output_data_bytes
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
819
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
820 SSE2_Row_IDCT();
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
821
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
822 SSE2_Transpose();
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2834
diff changeset
823
1970
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
824 SSE2_Column_IDCT();
51c098b1f404 SSE2-optimized variant of VP3 IDCT
melanson
parents:
diff changeset
825 }