annotate ppc/float_altivec.c @ 12481:e9a1e1ba768f libavcodec

Allow the lowres option to affect audio codecs, too
author mstorsjo
date Fri, 10 Sep 2010 06:19:31 +0000
parents 34a65026fa06
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
1 /*
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
2 * Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org>
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
3 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
4 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
5 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
6 * FFmpeg is free software; you can redistribute it and/or
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
8 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
9 * version 2.1 of the License, or (at your option) any later version.
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
10 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
11 * FFmpeg is distributed in the hope that it will be useful,
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
14 * Lesser General Public License for more details.
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
15 *
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3659
diff changeset
17 * License along with FFmpeg; if not, write to the Free Software
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
19 */
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
20
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 5010
diff changeset
21 #include "libavcodec/dsputil.h"
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
22
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
23 #include "dsputil_altivec.h"
8542
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
24 #include "util_altivec.h"
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
25
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
26 static void vector_fmul_altivec(float *dst, const float *src, int len)
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
27 {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
28 int i;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
29 vector float d0, d1, s, zero = (vector float)vec_splat_u32(0);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
30 for(i=0; i<len-7; i+=8) {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
31 d0 = vec_ld(0, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
32 s = vec_ld(0, src+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
33 d1 = vec_ld(16, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
34 d0 = vec_madd(d0, s, zero);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
35 d1 = vec_madd(d1, vec_ld(16,src+i), zero);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
36 vec_st(d0, 0, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
37 vec_st(d1, 16, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
38 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
39 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
40
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
41 static void vector_fmul_reverse_altivec(float *dst, const float *src0,
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
42 const float *src1, int len)
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
43 {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
44 int i;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
45 vector float d, s0, s1, h0, l0,
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
46 s2, s3, zero = (vector float)vec_splat_u32(0);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
47 src1 += len-4;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
48 for(i=0; i<len-7; i+=8) {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
49 s1 = vec_ld(0, src1-i); // [a,b,c,d]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
50 s0 = vec_ld(0, src0+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
51 l0 = vec_mergel(s1, s1); // [c,c,d,d]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
52 s3 = vec_ld(-16, src1-i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
53 h0 = vec_mergeh(s1, s1); // [a,a,b,b]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
54 s2 = vec_ld(16, src0+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
55 s1 = vec_mergeh(vec_mergel(l0,h0), // [d,b,d,b]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
56 vec_mergeh(l0,h0)); // [c,a,c,a]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
57 // [d,c,b,a]
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
58 l0 = vec_mergel(s3, s3);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
59 d = vec_madd(s0, s1, zero);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
60 h0 = vec_mergeh(s3, s3);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
61 vec_st(d, 0, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
62 s3 = vec_mergeh(vec_mergel(l0,h0),
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
63 vec_mergeh(l0,h0));
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
64 d = vec_madd(s2, s3, zero);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
65 vec_st(d, 16, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
66 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
67 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
68
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 9421
diff changeset
69 static void vector_fmul_add_altivec(float *dst, const float *src0,
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 9421
diff changeset
70 const float *src1, const float *src2,
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 9421
diff changeset
71 int len)
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
72 {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
73 int i;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
74 vector float d, s0, s1, s2, t0, t1, edges;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
75 vector unsigned char align = vec_lvsr(0,dst),
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
76 mask = vec_lvsl(0, dst);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
77
10301
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
78 for (i=0; i<len-3; i+=4) {
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
79 t0 = vec_ld(0, dst+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
80 t1 = vec_ld(15, dst+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
81 s0 = vec_ld(0, src0+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
82 s1 = vec_ld(0, src1+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
83 s2 = vec_ld(0, src2+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
84 edges = vec_perm(t1 ,t0, mask);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
85 d = vec_madd(s0,s1,s2);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
86 t1 = vec_perm(d, edges, align);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
87 t0 = vec_perm(edges, d, align);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
88 vec_st(t1, 15, dst+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
89 vec_st(t0, 0, dst+i);
02798c603744 cosmetics: fix indentation after previous commit
mru
parents: 10300
diff changeset
90 }
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
91 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
92
8542
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
93 static void vector_fmul_window_altivec(float *dst, const float *src0, const float *src1, const float *win, float add_bias, int len)
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
94 {
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
95 union {
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
96 vector float v;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
97 float s[4];
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
98 } vadd;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
99 vector float vadd_bias, zero, t0, t1, s0, s1, wi, wj;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
100 const vector unsigned char reverse = vcprm(3,2,1,0);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
101 int i,j;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
102
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
103 dst += len;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
104 win += len;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
105 src0+= len;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
106
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
107 vadd.s[0] = add_bias;
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
108 vadd_bias = vec_splat(vadd.v, 0);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
109 zero = (vector float)vec_splat_u32(0);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
110
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
111 for(i=-len*4, j=len*4-16; i<0; i+=16, j-=16) {
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
112 s0 = vec_ld(i, src0);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
113 s1 = vec_ld(j, src1);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
114 wi = vec_ld(i, win);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
115 wj = vec_ld(j, win);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
116
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
117 s1 = vec_perm(s1, s1, reverse);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
118 wj = vec_perm(wj, wj, reverse);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
119
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
120 t0 = vec_madd(s0, wj, vadd_bias);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
121 t0 = vec_nmsub(s1, wi, t0);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
122 t1 = vec_madd(s0, wi, vadd_bias);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
123 t1 = vec_madd(s1, wj, t1);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
124 t1 = vec_perm(t1, t1, reverse);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
125
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
126 vec_st(t0, i, dst);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
127 vec_st(t1, j, dst);
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
128 }
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
129 }
7674
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
130
8365
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
131 static void int32_to_float_fmul_scalar_altivec(float *dst, const int *src, float mul, int len)
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
132 {
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
133 union {
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
134 vector float v;
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
135 float s[4];
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
136 } mul_u;
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
137 int i;
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
138 vector float src1, src2, dst1, dst2, mul_v, zero;
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
139
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
140 zero = (vector float)vec_splat_u32(0);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
141 mul_u.s[0] = mul;
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
142 mul_v = vec_splat(mul_u.v, 0);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
143
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
144 for(i=0; i<len; i+=8) {
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
145 src1 = vec_ctf(vec_ld(0, src+i), 0);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
146 src2 = vec_ctf(vec_ld(16, src+i), 0);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
147 dst1 = vec_madd(src1, mul_v, zero);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
148 dst2 = vec_madd(src2, mul_v, zero);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
149 vec_st(dst1, 0, dst+i);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
150 vec_st(dst2, 16, dst+i);
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
151 }
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
152 }
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
153
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
154
7674
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
155 static vector signed short
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
156 float_to_int16_one_altivec(const float *src)
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
157 {
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
158 vector float s0 = vec_ld(0, src);
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
159 vector float s1 = vec_ld(16, src);
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
160 vector signed int t0 = vec_cts(s0, 0);
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
161 vector signed int t1 = vec_cts(s1, 0);
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
162 return vec_packs(t0,t1);
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
163 }
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
164
8537
93a3020d9636 Fix float_to_int16_altivec prototype to match float_to_int16's in dsputil.h
gpoirier
parents: 8365
diff changeset
165 static void float_to_int16_altivec(int16_t *dst, const float *src, long len)
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
166 {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
167 int i;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
168 vector signed short d0, d1, d;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
169 vector unsigned char align;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
170 if(((long)dst)&15) //FIXME
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
171 for(i=0; i<len-7; i+=8) {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
172 d0 = vec_ld(0, dst+i);
7674
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
173 d = float_to_int16_one_altivec(src+i);
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
174 d1 = vec_ld(15, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
175 d1 = vec_perm(d1, d0, vec_lvsl(0,dst+i));
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
176 align = vec_lvsr(0, dst+i);
3657
ff6720290478 Fix float_to_int16, unaligned case, broken by the previous commit
lu_zero
parents: 3583
diff changeset
177 d0 = vec_perm(d1, d, align);
ff6720290478 Fix float_to_int16, unaligned case, broken by the previous commit
lu_zero
parents: 3583
diff changeset
178 d1 = vec_perm(d, d1, align);
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
179 vec_st(d0, 0, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
180 vec_st(d1,15, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
181 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
182 else
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
183 for(i=0; i<len-7; i+=8) {
7674
e1876d3200ee Introduce float_to_int16_one_altivec
lu_zero
parents: 6763
diff changeset
184 d = float_to_int16_one_altivec(src+i);
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
185 vec_st(d, 0, dst+i);
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
186 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
187 }
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
188
7675
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
189 static void
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
190 float_to_int16_interleave_altivec(int16_t *dst, const float **src,
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
191 long len, int channels)
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
192 {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
193 int i;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
194 vector signed short d0, d1, d2, c0, c1, t0, t1;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
195 vector unsigned char align;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
196 if(channels == 1)
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
197 float_to_int16_altivec(dst, src[0], len);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
198 else
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
199 if (channels == 2) {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
200 if(((long)dst)&15)
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
201 for(i=0; i<len-7; i+=8) {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
202 d0 = vec_ld(0, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
203 t0 = float_to_int16_one_altivec(src[0] + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
204 d1 = vec_ld(31, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
205 t1 = float_to_int16_one_altivec(src[1] + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
206 c0 = vec_mergeh(t0, t1);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
207 c1 = vec_mergel(t0, t1);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
208 d2 = vec_perm(d1, d0, vec_lvsl(0, dst + i));
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
209 align = vec_lvsr(0, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
210 d0 = vec_perm(d2, c0, align);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
211 d1 = vec_perm(c0, c1, align);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
212 vec_st(d0, 0, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
213 d0 = vec_perm(c1, d2, align);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
214 vec_st(d1, 15, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
215 vec_st(d0, 31, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
216 dst+=8;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
217 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
218 else
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
219 for(i=0; i<len-7; i+=8) {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
220 t0 = float_to_int16_one_altivec(src[0] + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
221 t1 = float_to_int16_one_altivec(src[1] + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
222 d0 = vec_mergeh(t0, t1);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
223 d1 = vec_mergel(t0, t1);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
224 vec_st(d0, 0, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
225 vec_st(d1, 16, dst + i);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
226 dst+=8;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
227 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
228 } else {
10961
34a65026fa06 Move array specifiers outside DECLARE_ALIGNED() invocations
mru
parents: 10301
diff changeset
229 DECLARE_ALIGNED(16, int16_t, tmp)[len];
7675
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
230 int c, j;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
231 for (c = 0; c < channels; c++) {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
232 float_to_int16_altivec(tmp, src[c], len);
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
233 for (i = 0, j = c; i < len; i++, j+=channels) {
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
234 dst[j] = tmp[i];
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
235 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
236 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
237 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
238 }
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
239
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
240 void float_init_altivec(DSPContext* c, AVCodecContext *avctx)
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
241 {
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
242 c->vector_fmul = vector_fmul_altivec;
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
243 c->vector_fmul_reverse = vector_fmul_reverse_altivec;
10300
4d1b9ca628fc Drop unused args from vector_fmul_add_add, simpify code, and rename
mru
parents: 9421
diff changeset
244 c->vector_fmul_add = vector_fmul_add_altivec;
8365
d33b47d1f4c1 add AltiVec implementation of int32_to_float_fmul_scalar
gpoirier
parents: 7691
diff changeset
245 c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_altivec;
7675
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
246 if(!(avctx->flags & CODEC_FLAG_BITEXACT)) {
8542
5a923bd4f5c6 Add Altivec version of vector_fmul_window.
gpoirier
parents: 8537
diff changeset
247 c->vector_fmul_window = vector_fmul_window_altivec;
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
248 c->float_to_int16 = float_to_int16_altivec;
7675
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
249 c->float_to_int16_interleave = float_to_int16_interleave_altivec;
ad4bf45b9b63 Introduce float_to_int16_interleave_altivec, tested with vorbis
lu_zero
parents: 7674
diff changeset
250 }
3581
49082584828a altivec float optimizations
lu_zero
parents:
diff changeset
251 }