annotate i386/vp3dsp_mmx.c @ 3357:649ebe86d779 libavcodec

10l
author michael
date Sun, 25 Jun 2006 21:22:25 +0000
parents 0b546eab515d
children c8c591fe26f8
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
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 *
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
4 * This library is free software; you can redistribute it and/or
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
5 * 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
6 * License as published by the Free Software Foundation; either
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
7 * version 2 of the License, or (at your option) any later version.
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
8 *
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
9 * This library is distributed in the hope that it will be useful,
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
10 * 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
11 * 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
12 * 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
13 *
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
14 * You should have received a copy of the GNU Lesser General Public
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
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
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
17 */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
18
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 * @file vp3dsp_mmx.c
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
21 * 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
22 */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
23
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
24 #include "../dsputil.h"
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
25 #include "mmx.h"
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
26
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
27 #define IdctAdjustBeforeShift 8
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
28
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
29 /* (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
30 * 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
31 * 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
32 * idct_constants[44..47] = 8
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
33 */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
34 static uint16_t idct_constants[(4 + 7 + 1) * 4];
2753
ba8ecddf5598 adding a few const
michael
parents: 2696
diff changeset
35 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
36 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
37 };
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 #define r0 mm0
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
40 #define r1 mm1
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
41 #define r2 mm2
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
42 #define r3 mm3
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
43 #define r4 mm4
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
44 #define r5 mm5
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
45 #define r6 mm6
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
46 #define r7 mm7
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
47
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
48 /* 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
49 #define BeginIDCT() { \
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
50 movq_m2r(*I(3), r2); \
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
51 movq_m2r(*C(3), r6); \
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
52 movq_r2r(r2, r4); \
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
53 movq_m2r(*J(5), r7); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
54 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
55 movq_m2r(*C(5), r1); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
56 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
57 movq_r2r(r1, r5); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
58 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
59 movq_m2r(*I(1), r3); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
60 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
61 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
62 paddw_r2r(r2, r4); /* r4 = c3*i3 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
63 paddw_r2r(r7, r6); /* r6 = c3*i5 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
64 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
65 movq_m2r(*J(7), r1); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
66 paddw_r2r(r5, r7); /* r7 = c5*i5 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
67 movq_r2r(r0, r5); /* r5 = c1 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
68 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
69 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
70 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
71 movq_m2r(*C(7), r7); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
72 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
73 paddw_r2r(r3, r0); /* r0 = c1*i1 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
74 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
75 movq_m2r(*I(2), r2); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
76 pmulhw_r2r(r1, r7); /* r7 = c7*i7 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
77 paddw_r2r(r1, r5); /* r5 = c1*i7 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
78 movq_r2r(r2, r1); /* r1 = i2 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
79 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
80 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
81 movq_m2r(*J(6), r5); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
82 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
83 movq_r2r(r5, r7); /* r7 = i6 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
84 psubsw_r2r(r4, r0); /* r0 = A - C */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
85 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
86 paddw_r2r(r1, r2); /* r2 = c2*i2 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
87 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
88 paddsw_r2r(r4, r4); /* r4 = C + C */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
89 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
90 psubsw_r2r(r6, r3); /* r3 = B - D */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
91 paddw_r2r(r7, r5); /* r5 = c2*i6 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
92 paddsw_r2r(r6, r6); /* r6 = D + D */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
93 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
94 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
95 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
96 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
97 movq_m2r(*C(4), r4); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
98 movq_r2r(r3, r5); /* r5 = B - D */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
99 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
100 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
101 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
102 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
103 movq_m2r(*I(0), r6); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
104 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
105 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
106 movq_m2r(*J(4), r3); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
107 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
108 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
109 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
110 movq_r2r(r6, r0); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
111 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
112 paddsw_r2r(r3, r3); /* r3 = i4 + i4 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
113 paddsw_r2r(r1, r1); /* r1 = H + H */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
114 paddsw_r2r(r0, r3); /* r3 = i0 + i4 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
115 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
116 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
117 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
118 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
119 paddsw_r2r(r2, r2); /* r2 = A. + A. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
120 movq_m2r(*I(1), r0); /* r0 = C. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
121 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
122 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
123 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
124 }
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
125
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
126 /* RowIDCT gets ready to transpose */
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
127 #define RowIDCT() { \
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
128 \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
129 BeginIDCT(); \
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
130 \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
131 movq_m2r(*I(2), r3); /* r3 = D. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
132 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
133 paddsw_r2r(r1, r1); /* r1 = H. + H. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
134 paddsw_r2r(r7, r7); /* r7 = G + G */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
135 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
136 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
137 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
138 paddsw_r2r(r3, r3); \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
139 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
140 paddsw_r2r(r5, r5); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
141 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
142 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
143 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
144 paddsw_r2r(r0, r0); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
145 movq_r2m(r1, *I(1)); /* save R1 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
146 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
147 }
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
148
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
149 /* 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
150 #define ColumnIDCT() { \
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
151 \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
152 BeginIDCT(); \
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
153 \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
154 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
155 paddsw_r2r(r1, r1); /* r1 = H. + H. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
156 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
157 psraw_i2r(4, r2); /* r2 = NR2 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
158 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
159 psraw_i2r(4, r1); /* r1 = NR1 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
160 movq_m2r(*I(2), r3); /* r3 = D. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
161 paddsw_r2r(r7, r7); /* r7 = G + G */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
162 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
163 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
164 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
165 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
166 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
167 paddsw_r2r(r3, r3); /* r3 = D. + D. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
168 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
169 psraw_i2r(4, r4); /* r4 = NR4 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
170 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
171 psraw_i2r(4, r3); /* r3 = NR3 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
172 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
173 paddsw_r2r(r5, r5); /* r5 = B.. + B.. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
174 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
175 psraw_i2r(4, r6); /* r6 = NR6 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
176 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
177 psraw_i2r(4, r5); /* r5 = NR5 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
178 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
179 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
180 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
181 paddsw_r2r(r0, r0); /* r0 = C. + C. */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
182 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
183 psraw_i2r(4, r7); /* r7 = NR7 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
184 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
185 psraw_i2r(4, r0); /* r0 = NR0 */ \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
186 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
187 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
188 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
189 }
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
190
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
191 /* 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
192
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
193 At entry (we assume):
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
194
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
195 r0 = a3 a2 a1 a0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
196 I(1) = b3 b2 b1 b0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
197 r2 = c3 c2 c1 c0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
198 r3 = d3 d2 d1 d0
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
199
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
200 r4 = e3 e2 e1 e0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
201 r5 = f3 f2 f1 f0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
202 r6 = g3 g2 g1 g0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
203 r7 = h3 h2 h1 h0
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
204
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
205 At exit, we have:
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
206
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
207 I(0) = d0 c0 b0 a0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
208 I(1) = d1 c1 b1 a1
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
209 I(2) = d2 c2 b2 a2
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
210 I(3) = d3 c3 b3 a3
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2753
diff changeset
211
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
212 J(4) = h0 g0 f0 e0
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
213 J(5) = h1 g1 f1 e1
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
214 J(6) = h2 g2 f2 e2
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
215 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
216
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
217 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
218 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
219
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
220 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
221
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
222 #define Transpose() { \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
223 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
224 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
225 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
226 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
227 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
228 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
229 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
230 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
231 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
232 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
233 movq_r2m(r4, *J(4)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
234 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
235 movq_r2m(r5, *J(5)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
236 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
237 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
238 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
239 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
240 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
241 movq_r2m(r6, *J(7)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
242 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
243 movq_r2m(r1, *J(6)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
244 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
245 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
246 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
247 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
248 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
249 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
250 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
251 movq_r2m(r0, *I(0)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
252 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
253 movq_r2m(r1, *I(1)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
254 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
255 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
256 movq_r2m(r4, *I(3)); \
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
257 movq_r2m(r2, *I(2)); \
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
258 }
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
259
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1977
diff changeset
260 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
261 {
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
262 int j = 16;
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
263 uint16_t *p;
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
264
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
265 j = 1;
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
266 do {
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
267 p = idct_constants + ((j + 3) << 2);
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
268 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
269 } while (++j <= 7);
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
270
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
271 idct_constants[44] = idct_constants[45] =
1866
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
272 idct_constants[46] = idct_constants[47] = IdctAdjustBeforeShift;
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
273 }
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
274
2696
9699d325049d porting the mmx&sse2 (sse2 untested) vp3 idcts to the lavc idct API
michael
parents: 1977
diff changeset
275 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
276 {
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
277 /* eax = quantized input
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
278 * ebx = dequantizer matrix
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
279 * ecx = IDCT constants
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
280 * 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
281 * 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
282 * edx = output
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
283 * r0..r7 = mm0..mm7
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
284 */
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
285
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
286 #define C(x) (idct_constants + 16 + (x - 1) * 4)
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
287 #define Eight (idct_constants + 44)
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
288
1969
56cb752222cc correct MMX-optimized variant of VP3 IDCT, with comments (thank you
melanson
parents: 1866
diff changeset
289 /* 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
290 * 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
291 #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
292 #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
293
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
294 RowIDCT();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
295 Transpose();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
296
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
297 #undef I
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
298 #undef J
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
299 #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
300 #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
301
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
302 RowIDCT();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
303 Transpose();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
304
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
305 #undef I
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
306 #undef J
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
307 #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
308 #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
309
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
310 ColumnIDCT();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
311
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
312 #undef I
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
313 #undef J
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
314 #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
315 #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
316
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
317 ColumnIDCT();
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
318
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
319 #undef I
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
320 #undef J
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
321
1755f959ab7f seperated out the C-based VP3 DSP functions into a different file; also
melanson
parents:
diff changeset
322 }