annotate ppc/vp8dsp_altivec.c @ 12273:7323559a53fd libavcodec

Rename pow variable to pwr. Patch by Axel Holzinger <aholzinger gmx de>.
author rbultje
date Mon, 26 Jul 2010 13:52:49 +0000
parents f96187e79438
children 06abedae2906
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12011
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
1 /**
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
2 * VP8 compatible video decoder
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
3 *
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
4 * Copyright (C) 2010 David Conrad
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
5 *
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
6 * This file is part of FFmpeg.
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
7 *
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
8 * FFmpeg is free software; you can redistribute it and/or
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
9 * modify it under the terms of the GNU Lesser General Public
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
10 * License as published by the Free Software Foundation; either
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
11 * version 2.1 of the License, or (at your option) any later version.
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
12 *
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
13 * FFmpeg is distributed in the hope that it will be useful,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
16 * Lesser General Public License for more details.
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
17 *
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
18 * You should have received a copy of the GNU Lesser General Public
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
19 * License along with FFmpeg; if not, write to the Free Software
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
21 */
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
22
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
23 #include "libavcodec/vp8dsp.h"
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
24 #include "dsputil_altivec.h"
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
25 #include "types_altivec.h"
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
26 #include "util_altivec.h"
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
27
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
28 #define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
29
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
30 // h subpel filter uses msum to multiply+add 4 pixel taps at once
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
31 static const vec_s8 h_subpel_filters_inner[7] =
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
32 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
33 REPT4( -6, 123, 12, -1),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
34 REPT4(-11, 108, 36, -8),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
35 REPT4( -9, 93, 50, -6),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
36 REPT4(-16, 77, 77, -16),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
37 REPT4( -6, 50, 93, -9),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
38 REPT4( -8, 36, 108, -11),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
39 REPT4( -1, 12, 123, -6),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
40 };
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
41
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
42 // for 6tap filters, these are the outer two taps
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
43 // The zeros mask off pixels 4-7 when filtering 0-3
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
44 // and vice-versa
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
45 static const vec_s8 h_subpel_filters_outer[3] =
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
46 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
47 REPT4(0, 0, 2, 1),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
48 REPT4(0, 0, 3, 3),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
49 REPT4(0, 0, 1, 2),
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
50 };
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
51
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
52 #define LOAD_H_SUBPEL_FILTER(i) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
53 vec_s8 filter_inner = h_subpel_filters_inner[i]; \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
54 vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
55 vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
56
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
57 #define FILTER_H(dstv, off) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
58 a = vec_ld((off)-2, src); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
59 b = vec_ld((off)-2+15, src); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
60 \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
61 pixh = vec_perm(a, b, permh##off); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
62 pixl = vec_perm(a, b, perml##off); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
63 filth = vec_msum(filter_inner, pixh, c64); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
64 filtl = vec_msum(filter_inner, pixl, c64); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
65 \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
66 if (is6tap) { \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
67 outer = vec_perm(a, b, perm_6tap##off); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
68 filth = vec_msum(filter_outerh, outer, filth); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
69 filtl = vec_msum(filter_outerl, outer, filtl); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
70 } \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
71 if (w == 4) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
72 filtl = filth; /* discard pixels 4-7 */ \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
73 dstv = vec_packs(filth, filtl); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
74 dstv = vec_sra(dstv, c7)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
75
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
76 static av_always_inline
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
77 void put_vp8_epel_h_altivec_core(uint8_t *dst, int dst_stride,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
78 uint8_t *src, int src_stride,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
79 int h, int mx, int w, int is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
80 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
81 LOAD_H_SUBPEL_FILTER(mx-1);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
82 vec_u8 align_vec0, align_vec8, permh0, permh8, filt;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
83 vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
84 vec_u8 a, b, pixh, pixl, outer;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
85 vec_s16 f16h, f16l;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
86 vec_s32 filth, filtl;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
87
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
88 vec_u8 perm_inner = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
89 vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
90 vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
91 vec_u16 c7 = vec_splat_u16(7);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
92
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
93 align_vec0 = vec_lvsl( -2, src);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
94 align_vec8 = vec_lvsl(8-2, src);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
95
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
96 permh0 = vec_perm(align_vec0, align_vec0, perm_inner);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
97 permh8 = vec_perm(align_vec8, align_vec8, perm_inner);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
98 perm_inner = vec_add(perm_inner, vec_splat_u8(4));
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
99 perml0 = vec_perm(align_vec0, align_vec0, perm_inner);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
100 perml8 = vec_perm(align_vec8, align_vec8, perm_inner);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
101 perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
102 perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
103
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
104 while (h --> 0) {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
105 FILTER_H(f16h, 0);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
106
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
107 if (w == 16) {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
108 FILTER_H(f16l, 8);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
109 filt = vec_packsu(f16h, f16l);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
110 vec_st(filt, 0, dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
111 } else {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
112 filt = vec_packsu(f16h, f16h);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
113 vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
114 if (w == 8)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
115 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
116 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
117 src += src_stride;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
118 dst += dst_stride;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
119 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
120 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
121
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
122 // v subpel filter does a simple vertical multiply + add
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
123 static const vec_u8 v_subpel_filters[7] =
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
124 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
125 { 0, 6, 123, 12, 1, 0 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
126 { 2, 11, 108, 36, 8, 1 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
127 { 0, 9, 93, 50, 6, 0 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
128 { 3, 16, 77, 77, 16, 3 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
129 { 0, 6, 50, 93, 9, 0 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
130 { 1, 8, 36, 108, 11, 2 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
131 { 0, 1, 12, 123, 6, 0 },
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
132 };
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
133
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
134 #define LOAD_V_SUBPEL_FILTER(i) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
135 vec_u8 subpel_filter = v_subpel_filters[i]; \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
136 vec_u8 f0 = vec_splat(subpel_filter, 0); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
137 vec_u8 f1 = vec_splat(subpel_filter, 1); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
138 vec_u8 f2 = vec_splat(subpel_filter, 2); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
139 vec_u8 f3 = vec_splat(subpel_filter, 3); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
140 vec_u8 f4 = vec_splat(subpel_filter, 4); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
141 vec_u8 f5 = vec_splat(subpel_filter, 5)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
142
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
143 #define FILTER_V(dstv, vec_mul) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
144 s1f = (vec_s16)vec_mul(s1, f1); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
145 s2f = (vec_s16)vec_mul(s2, f2); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
146 s3f = (vec_s16)vec_mul(s3, f3); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
147 s4f = (vec_s16)vec_mul(s4, f4); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
148 s2f = vec_subs(s2f, s1f); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
149 s3f = vec_subs(s3f, s4f); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
150 if (is6tap) { \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
151 s0f = (vec_s16)vec_mul(s0, f0); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
152 s5f = (vec_s16)vec_mul(s5, f5); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
153 s2f = vec_adds(s2f, s0f); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
154 s3f = vec_adds(s3f, s5f); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
155 } \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
156 dstv = vec_adds(s2f, s3f); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
157 dstv = vec_adds(dstv, c64); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
158 dstv = vec_sra(dstv, c7)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
159
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
160 static av_always_inline
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
161 void put_vp8_epel_v_altivec_core(uint8_t *dst, int dst_stride,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
162 uint8_t *src, int src_stride,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
163 int h, int my, int w, int is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
164 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
165 LOAD_V_SUBPEL_FILTER(my-1);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
166 vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
167 vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
168 vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
169 vec_u16 c7 = vec_splat_u16(7);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
170
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
171 // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
172 // so combine this permute with the alignment permute vector
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
173 align_vech = vec_lvsl(0, src);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
174 align_vecl = vec_sld(align_vech, align_vech, 8);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
175 if (w ==16)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
176 perm_vec = vec_mergeh(align_vech, align_vecl);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
177 else
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
178 perm_vec = vec_mergeh(align_vech, align_vech);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
179
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
180 if (is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
181 s0 = load_with_perm_vec(-2*src_stride, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
182 s1 = load_with_perm_vec(-1*src_stride, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
183 s2 = load_with_perm_vec( 0*src_stride, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
184 s3 = load_with_perm_vec( 1*src_stride, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
185 if (is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
186 s4 = load_with_perm_vec( 2*src_stride, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
187
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
188 src += (2+is6tap)*src_stride;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
189
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
190 while (h --> 0) {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
191 if (is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
192 s5 = load_with_perm_vec(0, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
193 else
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
194 s4 = load_with_perm_vec(0, src, perm_vec);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
195
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
196 FILTER_V(f16h, vec_mule);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
197
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
198 if (w == 16) {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
199 FILTER_V(f16l, vec_mulo);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
200 filt = vec_packsu(f16h, f16l);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
201 vec_st(filt, 0, dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
202 } else {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
203 filt = vec_packsu(f16h, f16h);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
204 if (w == 4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
205 filt = (vec_u8)vec_splat((vec_u32)filt, 0);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
206 else
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
207 vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
208 vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
209 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
210
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
211 if (is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
212 s0 = s1;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
213 s1 = s2;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
214 s2 = s3;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
215 s3 = s4;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
216 if (is6tap)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
217 s4 = s5;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
218
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
219 dst += dst_stride;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
220 src += src_stride;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
221 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
222 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
223
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
224 #define EPEL_FUNCS(WIDTH, TAPS) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
225 static av_noinline \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
226 void put_vp8_epel ## WIDTH ## _h ## TAPS ## _altivec(uint8_t *dst, int dst_stride, uint8_t *src, int src_stride, int h, int mx, int my) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
227 { \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
228 put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
229 } \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
230 \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
231 static av_noinline \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
232 void put_vp8_epel ## WIDTH ## _v ## TAPS ## _altivec(uint8_t *dst, int dst_stride, uint8_t *src, int src_stride, int h, int mx, int my) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
233 { \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
234 put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
235 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
236
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
237 #define EPEL_HV(WIDTH, HTAPS, VTAPS) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
238 static void put_vp8_epel ## WIDTH ## _h ## HTAPS ## v ## VTAPS ## _altivec(uint8_t *dst, int stride, uint8_t *src, int s, int h, int mx, int my) \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
239 { \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
240 DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
241 put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16, src-2*stride, stride, h+5, mx, my); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
242 put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, stride, tmp+2*16, 16, h, mx, my); \
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
243 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
244
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
245 EPEL_FUNCS(16,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
246 EPEL_FUNCS(8, 6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
247 EPEL_FUNCS(8, 4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
248 EPEL_FUNCS(4, 6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
249 EPEL_FUNCS(4, 4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
250
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
251 EPEL_HV(16, 6,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
252 EPEL_HV(8, 6,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
253 EPEL_HV(8, 4,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
254 EPEL_HV(8, 6,4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
255 EPEL_HV(8, 4,4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
256 EPEL_HV(4, 6,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
257 EPEL_HV(4, 4,6)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
258 EPEL_HV(4, 6,4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
259 EPEL_HV(4, 4,4)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
260
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
261 static void put_vp8_pixels16_altivec(uint8_t *dst, int stride, uint8_t *src, int s, int h, int mx, int my)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
262 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
263 put_pixels16_altivec(dst, src, stride, h);
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
264 }
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
265
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
266 av_cold void ff_vp8dsp_init_altivec(VP8DSPContext *c)
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
267 {
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
268 if (!has_altivec())
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
269 return;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
270
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
271 c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
272 c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
273 c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
274 c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
275
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
276 c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
277 c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
278 c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
279 c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
280
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
281 c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
282 c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
283 c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
284 c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
285
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
286 c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
287 c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
288 c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
289 c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
290
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
291 c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
292 c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
293 c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
294 c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
f96187e79438 Altivec VP8 MC functions
conrad
parents:
diff changeset
295 }