2 * GMC (Global Motion Compensation), AltiVec-enabled
4 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
6 * This file is part of FFmpeg.
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.
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.
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
23 #include "libavutil/cpu.h"
24 #include "libavutil/mem.h"
25 #include "libavutil/ppc/cpu.h"
26 #include "libavutil/ppc/types_altivec.h"
27 #include "libavutil/ppc/util_altivec.h"
28 #include "libavcodec/mpegvideodsp.h"
31 /* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
32 * to preserve proper dst alignment. */
33 static void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */,
34 int stride, int h, int x16, int y16, int rounder)
37 const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
38 const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = {
39 (16 - x16) * (16 - y16), /* A */
40 (x16) * (16 - y16), /* B */
41 (16 - x16) * (y16), /* C */
42 (x16) * (y16), /* D */
43 0, 0, 0, 0 /* padding */
45 register const vector unsigned char vczero =
46 (const vector unsigned char) vec_splat_u8(0);
47 register const vector unsigned short vcsr8 =
48 (const vector unsigned short) vec_splat_u16(8);
49 register vector unsigned char dstv, dstv2, srcvB, srcvC, srcvD;
50 register vector unsigned short tempB, tempC, tempD;
51 unsigned long dst_odd = (unsigned long) dst & 0x0000000F;
52 unsigned long src_really_odd = (unsigned long) src & 0x0000000F;
53 register vector unsigned short tempA =
54 vec_ld(0, (const unsigned short *) ABCD);
55 register vector unsigned short Av = vec_splat(tempA, 0);
56 register vector unsigned short Bv = vec_splat(tempA, 1);
57 register vector unsigned short Cv = vec_splat(tempA, 2);
58 register vector unsigned short Dv = vec_splat(tempA, 3);
59 register vector unsigned short rounderV =
60 vec_splat((vec_u16) vec_lde(0, &rounder_a), 0);
62 /* we'll be able to pick-up our 9 char elements at src from those
63 * 32 bytes we load the first batch here, as inside the loop we can
64 * reuse 'src + stride' from one iteration as the 'src' of the next. */
65 register vector unsigned char src_0 = vec_ld(0, src);
66 register vector unsigned char src_1 = vec_ld(16, src);
67 register vector unsigned char srcvA = vec_perm(src_0, src_1,
70 if (src_really_odd != 0x0000000F)
71 /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
72 * on the second vector. */
73 srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
76 srcvA = vec_mergeh(vczero, srcvA);
77 srcvB = vec_mergeh(vczero, srcvB);
79 for (i = 0; i < h; i++) {
80 dst_odd = (unsigned long) dst & 0x0000000F;
81 src_really_odd = (((unsigned long) src) + stride) & 0x0000000F;
83 dstv = vec_ld(0, dst);
85 /* We'll be able to pick-up our 9 char elements at src + stride from
86 * those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD
87 * as the next srcvA and srcvB. */
88 src_0 = vec_ld(stride + 0, src);
89 src_1 = vec_ld(stride + 16, src);
90 srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
92 if (src_really_odd != 0x0000000F)
93 /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
94 * on the second vector. */
95 srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
99 srcvC = vec_mergeh(vczero, srcvC);
100 srcvD = vec_mergeh(vczero, srcvD);
102 /* OK, now we (finally) do the math :-)
103 * Those four instructions replace 32 int muls & 32 int adds.
104 * Isn't AltiVec nice? */
105 tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV);
106 tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA);
107 tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB);
108 tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC);
113 tempD = vec_sr(tempD, vcsr8);
115 dstv2 = vec_pack(tempD, (vector unsigned short) vczero);
118 dstv2 = vec_perm(dstv, dstv2, vcprm(0, 1, s0, s1));
120 dstv2 = vec_perm(dstv, dstv2, vcprm(s0, s1, 2, 3));
122 vec_st(dstv2, 0, dst);
128 #endif /* HAVE_ALTIVEC */
130 av_cold void ff_mpegvideodsp_init_ppc(MpegVideoDSPContext *c)
133 if (!PPC_ALTIVEC(av_get_cpu_flags()))
136 c->gmc1 = gmc1_altivec;
137 #endif /* HAVE_ALTIVEC */