2 * GMC (Global Motion Compensation)
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
25 #include "gcc_fixes.h"
27 #include "dsputil_altivec.h"
30 altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
31 to preserve proper dst alignement.
33 #define GMC1_PERF_COND (h==8)
34 void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
36 POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
37 const unsigned short __attribute__ ((aligned(16))) rounder_a[8] =
38 {rounder, rounder, rounder, rounder,
39 rounder, rounder, rounder, rounder};
40 const unsigned short __attribute__ ((aligned(16))) ABCD[8] =
42 (16-x16)*(16-y16), /* A */
43 ( x16)*(16-y16), /* B */
44 (16-x16)*( y16), /* C */
45 ( x16)*( y16), /* D */
46 0, 0, 0, 0 /* padding */
48 register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
49 register const_vector unsigned short vcsr8 = (const_vector unsigned short)vec_splat_u16(8);
50 register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
51 register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
53 unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
54 unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
57 POWERPC_PERF_START_COUNT(altivec_gmc1_num, GMC1_PERF_COND);
59 tempA = vec_ld(0, (unsigned short*)ABCD);
60 Av = vec_splat(tempA, 0);
61 Bv = vec_splat(tempA, 1);
62 Cv = vec_splat(tempA, 2);
63 Dv = vec_splat(tempA, 3);
65 rounderV = vec_ld(0, (unsigned short*)rounder_a);
67 // we'll be able to pick-up our 9 char elements
68 // at src from those 32 bytes
69 // we load the first batch here, as inside the loop
70 // we can re-use 'src+stride' from one iteration
71 // as the 'src' of the next.
72 src_0 = vec_ld(0, src);
73 src_1 = vec_ld(16, src);
74 srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));
76 if (src_really_odd != 0x0000000F)
77 { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector.
78 srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
84 srcvA = vec_mergeh(vczero, srcvA);
85 srcvB = vec_mergeh(vczero, srcvB);
89 dst_odd = (unsigned long)dst & 0x0000000F;
90 src_really_odd = (((unsigned long)src) + stride) & 0x0000000F;
92 dstv = vec_ld(0, dst);
94 // we we'll be able to pick-up our 9 char elements
95 // at src + stride from those 32 bytes
96 // then reuse the resulting 2 vectors srvcC and srcvD
97 // as the next srcvA and srcvB
98 src_0 = vec_ld(stride + 0, src);
99 src_1 = vec_ld(stride + 16, src);
100 srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
102 if (src_really_odd != 0x0000000F)
103 { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector.
104 srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
111 srcvC = vec_mergeh(vczero, srcvC);
112 srcvD = vec_mergeh(vczero, srcvD);
115 // OK, now we (finally) do the math :-)
116 // those four instructions replaces 32 int muls & 32 int adds.
117 // isn't AltiVec nice ?
118 tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
119 tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
120 tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);
121 tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC);
126 tempD = vec_sr(tempD, vcsr8);
128 dstv2 = vec_pack(tempD, (vector unsigned short)vczero);
132 dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1));
136 dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3));
139 vec_st(dstv2, 0, dst);
145 POWERPC_PERF_STOP_COUNT(altivec_gmc1_num, GMC1_PERF_COND);