Mercurial > libavcodec.hg
annotate i386/vp3dsp_mmx.c @ 7742:bff9b5fea03f libavcodec
Use ff_pw_8 in MMX/SSE VP3 IDCT
author | conrad |
---|---|
date | Sat, 30 Aug 2008 19:41:42 +0000 |
parents | f7cbb7733146 |
children | 892ca48b7d76 |
rev | line source |
---|---|
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
1 /* |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
2 * Copyright (C) 2004 the ffmpeg project |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
3 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
4 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
5 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
6 * FFmpeg is free software; you can redistribute it and/or |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
7 * modify it under the terms of the GNU Lesser General Public |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
8 * License as published by the Free Software Foundation; either |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
9 * version 2.1 of the License, or (at your option) any later version. |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
10 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
11 * FFmpeg is distributed in the hope that it will be useful, |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
14 * Lesser General Public License for more details. |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
15 * |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
16 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
17 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2967
diff
changeset
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
19 */ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
20 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
21 /** |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
22 * @file vp3dsp_mmx.c |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
23 * MMX-optimized functions cribbed from the original VP3 source code. |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
24 */ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
25 |
6763 | 26 #include "libavcodec/dsputil.h" |
7742 | 27 #include "dsputil_mmx.h" |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
28 #include "mmx.h" |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
29 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
30 /* (12 * 4) 2-byte memory locations ( = 96 bytes total) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
31 * idct_constants[0..15] = Mask table (M(I)) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
32 * idct_constants[16..43] = Cosine table (C(I)) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
33 * idct_constants[44..47] = 8 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
34 */ |
7742 | 35 static uint16_t idct_constants[(4 + 7) * 4]; |
2753 | 36 static const uint16_t idct_cosine_table[7] = { |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
37 64277, 60547, 54491, 46341, 36410, 25080, 12785 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
38 }; |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
39 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
40 #define r0 mm0 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
41 #define r1 mm1 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
42 #define r2 mm2 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
43 #define r3 mm3 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
44 #define r4 mm4 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
45 #define r5 mm5 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
46 #define r6 mm6 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
47 #define r7 mm7 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
48 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
49 /* from original comments: The Macro does IDct on 4 1-D Dcts */ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
50 #define BeginIDCT() { \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
51 movq_m2r(*I(3), r2); \ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
52 movq_m2r(*C(3), r6); \ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
53 movq_r2r(r2, r4); \ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
54 movq_m2r(*J(5), r7); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
55 pmulhw_r2r(r6, r4); /* r4 = c3*i3 - i3 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
56 movq_m2r(*C(5), r1); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
57 pmulhw_r2r(r7, r6); /* r6 = c3*i5 - i5 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
58 movq_r2r(r1, r5); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
59 pmulhw_r2r(r2, r1); /* r1 = c5*i3 - i3 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
60 movq_m2r(*I(1), r3); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
61 pmulhw_r2r(r7, r5); /* r5 = c5*i5 - i5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
62 movq_m2r(*C(1), r0); /* (all registers are in use) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
63 paddw_r2r(r2, r4); /* r4 = c3*i3 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
64 paddw_r2r(r7, r6); /* r6 = c3*i5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
65 paddw_r2r(r1, r2); /* r2 = c5*i3 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
66 movq_m2r(*J(7), r1); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
67 paddw_r2r(r5, r7); /* r7 = c5*i5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
68 movq_r2r(r0, r5); /* r5 = c1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
69 pmulhw_r2r(r3, r0); /* r0 = c1*i1 - i1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
70 paddsw_r2r(r7, r4); /* r4 = C = c3*i3 + c5*i5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
71 pmulhw_r2r(r1, r5); /* r5 = c1*i7 - i7 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
72 movq_m2r(*C(7), r7); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
73 psubsw_r2r(r2, r6); /* r6 = D = c3*i5 - c5*i3 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
74 paddw_r2r(r3, r0); /* r0 = c1*i1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
75 pmulhw_r2r(r7, r3); /* r3 = c7*i1 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
76 movq_m2r(*I(2), r2); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
77 pmulhw_r2r(r1, r7); /* r7 = c7*i7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
78 paddw_r2r(r1, r5); /* r5 = c1*i7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
79 movq_r2r(r2, r1); /* r1 = i2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
80 pmulhw_m2r(*C(2), r2); /* r2 = c2*i2 - i2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
81 psubsw_r2r(r5, r3); /* r3 = B = c7*i1 - c1*i7 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
82 movq_m2r(*J(6), r5); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
83 paddsw_r2r(r7, r0); /* r0 = A = c1*i1 + c7*i7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
84 movq_r2r(r5, r7); /* r7 = i6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
85 psubsw_r2r(r4, r0); /* r0 = A - C */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
86 pmulhw_m2r(*C(2), r5); /* r5 = c2*i6 - i6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
87 paddw_r2r(r1, r2); /* r2 = c2*i2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
88 pmulhw_m2r(*C(6), r1); /* r1 = c6*i2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
89 paddsw_r2r(r4, r4); /* r4 = C + C */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
90 paddsw_r2r(r0, r4); /* r4 = C. = A + C */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
91 psubsw_r2r(r6, r3); /* r3 = B - D */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
92 paddw_r2r(r7, r5); /* r5 = c2*i6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
93 paddsw_r2r(r6, r6); /* r6 = D + D */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
94 pmulhw_m2r(*C(6), r7); /* r7 = c6*i6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
95 paddsw_r2r(r3, r6); /* r6 = D. = B + D */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
96 movq_r2m(r4, *I(1)); /* save C. at I(1) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
97 psubsw_r2r(r5, r1); /* r1 = H = c6*i2 - c2*i6 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
98 movq_m2r(*C(4), r4); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
99 movq_r2r(r3, r5); /* r5 = B - D */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
100 pmulhw_r2r(r4, r3); /* r3 = (c4 - 1) * (B - D) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
101 paddsw_r2r(r2, r7); /* r7 = G = c6*i6 + c2*i2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
102 movq_r2m(r6, *I(2)); /* save D. at I(2) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
103 movq_r2r(r0, r2); /* r2 = A - C */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
104 movq_m2r(*I(0), r6); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
105 pmulhw_r2r(r4, r0); /* r0 = (c4 - 1) * (A - C) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
106 paddw_r2r(r3, r5); /* r5 = B. = c4 * (B - D) */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
107 movq_m2r(*J(4), r3); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
108 psubsw_r2r(r1, r5); /* r5 = B.. = B. - H */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
109 paddw_r2r(r0, r2); /* r0 = A. = c4 * (A - C) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
110 psubsw_r2r(r3, r6); /* r6 = i0 - i4 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
111 movq_r2r(r6, r0); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
112 pmulhw_r2r(r4, r6); /* r6 = (c4 - 1) * (i0 - i4) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
113 paddsw_r2r(r3, r3); /* r3 = i4 + i4 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
114 paddsw_r2r(r1, r1); /* r1 = H + H */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
115 paddsw_r2r(r0, r3); /* r3 = i0 + i4 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
116 paddsw_r2r(r5, r1); /* r1 = H. = B + H */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
117 pmulhw_r2r(r3, r4); /* r4 = (c4 - 1) * (i0 + i4) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
118 paddsw_r2r(r0, r6); /* r6 = F = c4 * (i0 - i4) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
119 psubsw_r2r(r2, r6); /* r6 = F. = F - A. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
120 paddsw_r2r(r2, r2); /* r2 = A. + A. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
121 movq_m2r(*I(1), r0); /* r0 = C. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
122 paddsw_r2r(r6, r2); /* r2 = A.. = F + A. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
123 paddw_r2r(r3, r4); /* r4 = E = c4 * (i0 + i4) */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
124 psubsw_r2r(r1, r2); /* r2 = R2 = A.. - H. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
125 } |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
126 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
127 /* RowIDCT gets ready to transpose */ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
128 #define RowIDCT() { \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
129 \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
130 BeginIDCT(); \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
131 \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
132 movq_m2r(*I(2), r3); /* r3 = D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
133 psubsw_r2r(r7, r4); /* r4 = E. = E - G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
134 paddsw_r2r(r1, r1); /* r1 = H. + H. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
135 paddsw_r2r(r7, r7); /* r7 = G + G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
136 paddsw_r2r(r2, r1); /* r1 = R1 = A.. + H. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
137 paddsw_r2r(r4, r7); /* r7 = G. = E + G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
138 psubsw_r2r(r3, r4); /* r4 = R4 = E. - D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
139 paddsw_r2r(r3, r3); \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
140 psubsw_r2r(r5, r6); /* r6 = R6 = F. - B.. */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
141 paddsw_r2r(r5, r5); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
142 paddsw_r2r(r4, r3); /* r3 = R3 = E. + D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
143 paddsw_r2r(r6, r5); /* r5 = R5 = F. + B.. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
144 psubsw_r2r(r0, r7); /* r7 = R7 = G. - C. */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
145 paddsw_r2r(r0, r0); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
146 movq_r2m(r1, *I(1)); /* save R1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
147 paddsw_r2r(r7, r0); /* r0 = R0 = G. + C. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
148 } |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
149 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
150 /* Column IDCT normalizes and stores final results */ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
151 #define ColumnIDCT() { \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
152 \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
153 BeginIDCT(); \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
154 \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
155 paddsw_m2r(*Eight, r2); /* adjust R2 (and R1) for shift */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
156 paddsw_r2r(r1, r1); /* r1 = H. + H. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
157 paddsw_r2r(r2, r1); /* r1 = R1 = A.. + H. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
158 psraw_i2r(4, r2); /* r2 = NR2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
159 psubsw_r2r(r7, r4); /* r4 = E. = E - G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
160 psraw_i2r(4, r1); /* r1 = NR1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
161 movq_m2r(*I(2), r3); /* r3 = D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
162 paddsw_r2r(r7, r7); /* r7 = G + G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
163 movq_r2m(r2, *I(2)); /* store NR2 at I2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
164 paddsw_r2r(r4, r7); /* r7 = G. = E + G */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
165 movq_r2m(r1, *I(1)); /* store NR1 at I1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
166 psubsw_r2r(r3, r4); /* r4 = R4 = E. - D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
167 paddsw_m2r(*Eight, r4); /* adjust R4 (and R3) for shift */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
168 paddsw_r2r(r3, r3); /* r3 = D. + D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
169 paddsw_r2r(r4, r3); /* r3 = R3 = E. + D. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
170 psraw_i2r(4, r4); /* r4 = NR4 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
171 psubsw_r2r(r5, r6); /* r6 = R6 = F. - B.. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
172 psraw_i2r(4, r3); /* r3 = NR3 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
173 paddsw_m2r(*Eight, r6); /* adjust R6 (and R5) for shift */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
174 paddsw_r2r(r5, r5); /* r5 = B.. + B.. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
175 paddsw_r2r(r6, r5); /* r5 = R5 = F. + B.. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
176 psraw_i2r(4, r6); /* r6 = NR6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
177 movq_r2m(r4, *J(4)); /* store NR4 at J4 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
178 psraw_i2r(4, r5); /* r5 = NR5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
179 movq_r2m(r3, *I(3)); /* store NR3 at I3 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
180 psubsw_r2r(r0, r7); /* r7 = R7 = G. - C. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
181 paddsw_m2r(*Eight, r7); /* adjust R7 (and R0) for shift */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
182 paddsw_r2r(r0, r0); /* r0 = C. + C. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
183 paddsw_r2r(r7, r0); /* r0 = R0 = G. + C. */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
184 psraw_i2r(4, r7); /* r7 = NR7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
185 movq_r2m(r6, *J(6)); /* store NR6 at J6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
186 psraw_i2r(4, r0); /* r0 = NR0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
187 movq_r2m(r5, *J(5)); /* store NR5 at J5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
188 movq_r2m(r7, *J(7)); /* store NR7 at J7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
189 movq_r2m(r0, *I(0)); /* store NR0 at I0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
190 } |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
191 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
192 /* Following macro does two 4x4 transposes in place. |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
193 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
194 At entry (we assume): |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
195 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
196 r0 = a3 a2 a1 a0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
197 I(1) = b3 b2 b1 b0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
198 r2 = c3 c2 c1 c0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
199 r3 = d3 d2 d1 d0 |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
200 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
201 r4 = e3 e2 e1 e0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
202 r5 = f3 f2 f1 f0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
203 r6 = g3 g2 g1 g0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
204 r7 = h3 h2 h1 h0 |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
205 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
206 At exit, we have: |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
207 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
208 I(0) = d0 c0 b0 a0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
209 I(1) = d1 c1 b1 a1 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
210 I(2) = d2 c2 b2 a2 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
211 I(3) = d3 c3 b3 a3 |
2967 | 212 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
213 J(4) = h0 g0 f0 e0 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
214 J(5) = h1 g1 f1 e1 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
215 J(6) = h2 g2 f2 e2 |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
216 J(7) = h3 g3 f3 e3 |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
217 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
218 I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3. |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
219 J(4) J(5) J(6) J(7) is the transpose of r4 r5 r6 r7. |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
220 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
221 Since r1 is free at entry, we calculate the Js first. */ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
222 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
223 #define Transpose() { \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
224 movq_r2r(r4, r1); /* r1 = e3 e2 e1 e0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
225 punpcklwd_r2r(r5, r4); /* r4 = f1 e1 f0 e0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
226 movq_r2m(r0, *I(0)); /* save a3 a2 a1 a0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
227 punpckhwd_r2r(r5, r1); /* r1 = f3 e3 f2 e2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
228 movq_r2r(r6, r0); /* r0 = g3 g2 g1 g0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
229 punpcklwd_r2r(r7, r6); /* r6 = h1 g1 h0 g0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
230 movq_r2r(r4, r5); /* r5 = f1 e1 f0 e0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
231 punpckldq_r2r(r6, r4); /* r4 = h0 g0 f0 e0 = R4 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
232 punpckhdq_r2r(r6, r5); /* r5 = h1 g1 f1 e1 = R5 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
233 movq_r2r(r1, r6); /* r6 = f3 e3 f2 e2 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
234 movq_r2m(r4, *J(4)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
235 punpckhwd_r2r(r7, r0); /* r0 = h3 g3 h2 g2 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
236 movq_r2m(r5, *J(5)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
237 punpckhdq_r2r(r0, r6); /* r6 = h3 g3 f3 e3 = R7 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
238 movq_m2r(*I(0), r4); /* r4 = a3 a2 a1 a0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
239 punpckldq_r2r(r0, r1); /* r1 = h2 g2 f2 e2 = R6 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
240 movq_m2r(*I(1), r5); /* r5 = b3 b2 b1 b0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
241 movq_r2r(r4, r0); /* r0 = a3 a2 a1 a0 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
242 movq_r2m(r6, *J(7)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
243 punpcklwd_r2r(r5, r0); /* r0 = b1 a1 b0 a0 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
244 movq_r2m(r1, *J(6)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
245 punpckhwd_r2r(r5, r4); /* r4 = b3 a3 b2 a2 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
246 movq_r2r(r2, r5); /* r5 = c3 c2 c1 c0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
247 punpcklwd_r2r(r3, r2); /* r2 = d1 c1 d0 c0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
248 movq_r2r(r0, r1); /* r1 = b1 a1 b0 a0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
249 punpckldq_r2r(r2, r0); /* r0 = d0 c0 b0 a0 = R0 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
250 punpckhdq_r2r(r2, r1); /* r1 = d1 c1 b1 a1 = R1 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
251 movq_r2r(r4, r2); /* r2 = b3 a3 b2 a2 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
252 movq_r2m(r0, *I(0)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
253 punpckhwd_r2r(r3, r5); /* r5 = d3 c3 d2 c2 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
254 movq_r2m(r1, *I(1)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
255 punpckhdq_r2r(r5, r4); /* r4 = d3 c3 b3 a3 = R3 */ \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
256 punpckldq_r2r(r5, r2); /* r2 = d2 c2 b2 a2 = R2 */ \ |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
257 movq_r2m(r4, *I(3)); \ |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
258 movq_r2m(r2, *I(2)); \ |
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
259 } |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
260 |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
1977
diff
changeset
|
261 void ff_vp3_dsp_init_mmx(void) |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
262 { |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
263 int j = 16; |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
264 uint16_t *p; |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
265 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
266 j = 1; |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
267 do { |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
268 p = idct_constants + ((j + 3) << 2); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
269 p[0] = p[1] = p[2] = p[3] = idct_cosine_table[j - 1]; |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
270 } while (++j <= 7); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
271 } |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
272 |
2696
9699d325049d
porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents:
1977
diff
changeset
|
273 void ff_vp3_idct_mmx(int16_t *output_data) |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
274 { |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
275 /* eax = quantized input |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
276 * ebx = dequantizer matrix |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
277 * ecx = IDCT constants |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
278 * M(I) = ecx + MaskOffset(0) + I * 8 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
279 * C(I) = ecx + CosineOffset(32) + (I-1) * 8 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
280 * edx = output |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
281 * r0..r7 = mm0..mm7 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
282 */ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
283 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
284 #define C(x) (idct_constants + 16 + (x - 1) * 4) |
7742 | 285 #define Eight (&ff_pw_8) |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
286 |
1969
56cb752222cc
correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents:
1866
diff
changeset
|
287 /* at this point, function has completed dequantization + dezigzag + |
1866
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
288 * partial transposition; now do the idct itself */ |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
289 #define I(K) (output_data + K * 8) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
290 #define J(K) (output_data + ((K - 4) * 8) + 4) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
291 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
292 RowIDCT(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
293 Transpose(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
294 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
295 #undef I |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
296 #undef J |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
297 #define I(K) (output_data + (K * 8) + 32) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
298 #define J(K) (output_data + ((K - 4) * 8) + 36) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
299 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
300 RowIDCT(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
301 Transpose(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
302 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
303 #undef I |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
304 #undef J |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
305 #define I(K) (output_data + K * 8) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
306 #define J(K) (output_data + K * 8) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
307 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
308 ColumnIDCT(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
309 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
310 #undef I |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
311 #undef J |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
312 #define I(K) (output_data + (K * 8) + 4) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
313 #define J(K) (output_data + (K * 8) + 4) |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
314 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
315 ColumnIDCT(); |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
316 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
317 #undef I |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
318 #undef J |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
319 |
1755f959ab7f
seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff
changeset
|
320 } |
5014
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
321 |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
322 void ff_vp3_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block) |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
323 { |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
324 ff_vp3_idct_mmx(block); |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
325 put_signed_pixels_clamped_mmx(block, dest, line_size); |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
326 } |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
327 |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
328 void ff_vp3_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block) |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
329 { |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
330 ff_vp3_idct_mmx(block); |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
331 add_pixels_clamped_mmx(block, dest, line_size); |
42b99a3aadde
better separation of vp3dsp functions from dsputil_mmx.c
aurel
parents:
5010
diff
changeset
|
332 } |