b7a477dd96a7358deb463e1ddcb11d31751c82c3
[libav.git] / libavcodec / ppc / gmc_altivec.c
1 /*
2 * GMC (Global Motion Compensation)
3 * AltiVec-enabled
4 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 #include "dsputil.h"
24
25 #include "gcc_fixes.h"
26
27 #include "dsputil_ppc.h"
28 #include "util_altivec.h"
29
30 /*
31 altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
32 to preserve proper dst alignement.
33 */
34 #define GMC1_PERF_COND (h==8)
35 void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
36 {
37 POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
38 const DECLARE_ALIGNED_16(unsigned short, rounder_a[8]) =
39 {rounder, rounder, rounder, rounder,
40 rounder, rounder, rounder, rounder};
41 const DECLARE_ALIGNED_16(unsigned short, ABCD[8]) =
42 {
43 (16-x16)*(16-y16), /* A */
44 ( x16)*(16-y16), /* B */
45 (16-x16)*( y16), /* C */
46 ( x16)*( y16), /* D */
47 0, 0, 0, 0 /* padding */
48 };
49 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
50 register const vector unsigned short vcsr8 = (const vector unsigned short)vec_splat_u16(8);
51 register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
52 register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
53 int i;
54 unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
55 unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
56
57
58 POWERPC_PERF_START_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
59
60 tempA = vec_ld(0, (unsigned short*)ABCD);
61 Av = vec_splat(tempA, 0);
62 Bv = vec_splat(tempA, 1);
63 Cv = vec_splat(tempA, 2);
64 Dv = vec_splat(tempA, 3);
65
66 rounderV = vec_ld(0, (unsigned short*)rounder_a);
67
68 // we'll be able to pick-up our 9 char elements
69 // at src from those 32 bytes
70 // we load the first batch here, as inside the loop
71 // we can re-use 'src+stride' from one iteration
72 // as the 'src' of the next.
73 src_0 = vec_ld(0, src);
74 src_1 = vec_ld(16, src);
75 srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));
76
77 if (src_really_odd != 0x0000000F)
78 { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector.
79 srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
80 }
81 else
82 {
83 srcvB = src_1;
84 }
85 srcvA = vec_mergeh(vczero, srcvA);
86 srcvB = vec_mergeh(vczero, srcvB);
87
88 for(i=0; i<h; i++)
89 {
90 dst_odd = (unsigned long)dst & 0x0000000F;
91 src_really_odd = (((unsigned long)src) + stride) & 0x0000000F;
92
93 dstv = vec_ld(0, dst);
94
95 // we we'll be able to pick-up our 9 char elements
96 // at src + stride from those 32 bytes
97 // then reuse the resulting 2 vectors srvcC and srcvD
98 // as the next srcvA and srcvB
99 src_0 = vec_ld(stride + 0, src);
100 src_1 = vec_ld(stride + 16, src);
101 srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
102
103 if (src_really_odd != 0x0000000F)
104 { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector.
105 srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
106 }
107 else
108 {
109 srcvD = src_1;
110 }
111
112 srcvC = vec_mergeh(vczero, srcvC);
113 srcvD = vec_mergeh(vczero, srcvD);
114
115
116 // OK, now we (finally) do the math :-)
117 // those four instructions replaces 32 int muls & 32 int adds.
118 // isn't AltiVec nice ?
119 tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
120 tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
121 tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);
122 tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC);
123
124 srcvA = srcvC;
125 srcvB = srcvD;
126
127 tempD = vec_sr(tempD, vcsr8);
128
129 dstv2 = vec_pack(tempD, (vector unsigned short)vczero);
130
131 if (dst_odd)
132 {
133 dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1));
134 }
135 else
136 {
137 dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3));
138 }
139
140 vec_st(dstv2, 0, dst);
141
142 dst += stride;
143 src += stride;
144 }
145
146 POWERPC_PERF_STOP_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
147 }