2
2
* FFT/MDCT transform with SSE optimizations
3
3
* Copyright (c) 2002 Fabrice Bellard.
5
* This library is free software; you can redistribute it and/or
5
* This file is part of FFmpeg.
7
* FFmpeg is free software; you can redistribute it and/or
6
8
* modify it under the terms of the GNU Lesser General Public
7
9
* License as published by the Free Software Foundation; either
8
* version 2 of the License, or (at your option) any later version.
10
* version 2.1 of the License, or (at your option) any later version.
10
* This library is distributed in the hope that it will be useful,
12
* FFmpeg is distributed in the hope that it will be useful,
11
13
* but WITHOUT ANY WARRANTY; without even the implied warranty of
12
14
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13
15
* Lesser General Public License for more details.
15
17
* You should have received a copy of the GNU Lesser General Public
16
* License along with this library; if not, write to the Free Software
17
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18
* License along with FFmpeg; if not, write to the Free Software
19
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19
21
#include "../dsputil.h"
22
#ifdef HAVE_BUILTIN_VECTOR
24
#include <xmmintrin.h>
26
static const float p1p1p1m1[4] __attribute__((aligned(16))) =
27
{ 1.0, 1.0, 1.0, -1.0 };
29
static const float p1p1m1p1[4] __attribute__((aligned(16))) =
30
{ 1.0, 1.0, -1.0, 1.0 };
32
static const float p1p1m1m1[4] __attribute__((aligned(16))) =
33
{ 1.0, 1.0, -1.0, -1.0 };
23
static const int p1p1p1m1[4] __attribute__((aligned(16))) =
26
static const int p1p1m1p1[4] __attribute__((aligned(16))) =
29
static const int p1p1m1m1[4] __attribute__((aligned(16))) =
30
{ 0, 0, 1 << 31, 1 << 31 };
32
static const int p1m1p1m1[4] __attribute__((aligned(16))) =
33
{ 0, 1 << 31, 0, 1 << 31 };
35
static const int m1m1m1m1[4] __attribute__((aligned(16))) =
36
{ 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
36
39
static void print_v4sf(const char *str, __m128 a)
44
47
/* XXX: handle reverse case */
45
void fft_calc_sse(FFTContext *s, FFTComplex *z)
48
void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
50
register FFTComplex *p, *q;
51
FFTComplex *cptr, *cptr1;
57
__m128 *r, a, b, a1, c1, c2;
60
c1 = *(__m128 *)p1p1m1m1;
61
c2 = *(__m128 *)p1p1p1m1;
63
c2 = *(__m128 *)p1p1m1p1;
65
c2 = *(__m128 *)p1p1p1m1;
70
b = _mm_shuffle_ps(a, a, _MM_SHUFFLE(1, 0, 3, 2));
71
a = _mm_mul_ps(a, c1);
72
/* do the pass 0 butterfly */
76
b = _mm_shuffle_ps(a1, a1, _MM_SHUFFLE(1, 0, 3, 2));
77
a1 = _mm_mul_ps(a1, c1);
78
/* do the pass 0 butterfly */
79
b = _mm_add_ps(a1, b);
81
/* multiply third by -i */
82
b = _mm_shuffle_ps(b, b, _MM_SHUFFLE(2, 3, 1, 0));
83
b = _mm_mul_ps(b, c2);
85
/* do the pass 1 butterfly */
86
r[0] = _mm_add_ps(a, b);
87
r[1] = _mm_sub_ps(a, b);
56
"movaps %0, %%xmm4 \n\t"
57
"movaps %1, %%xmm5 \n\t"
59
"m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
66
/* do the pass 0 butterfly */
67
"movaps (%0,%1), %%xmm0 \n\t"
68
"movaps %%xmm0, %%xmm1 \n\t"
69
"shufps $0x4E, %%xmm0, %%xmm0 \n\t"
70
"xorps %%xmm4, %%xmm1 \n\t"
71
"addps %%xmm1, %%xmm0 \n\t"
72
"movaps 16(%0,%1), %%xmm2 \n\t"
73
"movaps %%xmm2, %%xmm3 \n\t"
74
"shufps $0x4E, %%xmm2, %%xmm2 \n\t"
75
"xorps %%xmm4, %%xmm3 \n\t"
76
"addps %%xmm3, %%xmm2 \n\t"
77
/* multiply third by -i */
78
/* by toggling the sign bit */
79
"shufps $0xB4, %%xmm2, %%xmm2 \n\t"
80
"xorps %%xmm5, %%xmm2 \n\t"
81
/* do the pass 1 butterfly */
82
"movaps %%xmm0, %%xmm1 \n\t"
83
"addps %%xmm2, %%xmm0 \n\t"
84
"subps %%xmm2, %%xmm1 \n\t"
85
"movaps %%xmm0, (%0,%1) \n\t"
86
"movaps %%xmm1, 16(%0,%1) \n\t"
91
91
/* pass 2 .. ln-1 */
93
nblocks = 1 << (ln-3);
106
__m128 a, b, c, t1, t2;
115
_mm_shuffle_ps(b, b, _MM_SHUFFLE(2, 2, 0, 0)));
116
c = *(__m128 *)(cptr + 2);
119
_mm_shuffle_ps(b, b, _MM_SHUFFLE(3, 3, 1, 1)));
120
b = _mm_add_ps(t1, t2);
123
*(__m128 *)p = _mm_add_ps(a, b);
124
*(__m128 *)q = _mm_sub_ps(a, b);
104
"movaps (%2,%0), %%xmm1 \n\t"
105
"movaps (%1,%0), %%xmm0 \n\t"
106
"movaps %%xmm1, %%xmm2 \n\t"
107
"shufps $0xA0, %%xmm1, %%xmm1 \n\t"
108
"shufps $0xF5, %%xmm2, %%xmm2 \n\t"
109
"mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re
110
"mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
111
"addps %%xmm2, %%xmm1 \n\t"
112
"movaps %%xmm0, %%xmm3 \n\t"
113
"addps %%xmm1, %%xmm0 \n\t"
114
"subps %%xmm1, %%xmm3 \n\t"
115
"movaps %%xmm0, (%1,%0) \n\t"
116
"movaps %%xmm3, (%2,%0) \n\t"
119
:"r"(p), "r"(p + nloops), "r"(cptr)
135
nblocks = nblocks >> 1;
136
nloops = nloops << 1;
137
126
} while (nblocks != 0);
129
void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
130
const FFTSample *input, FFTSample *tmp)
132
long k, n8, n4, n2, n;
133
const uint16_t *revtab = s->fft.revtab;
134
const FFTSample *tcos = s->tcos;
135
const FFTSample *tsin = s->tsin;
136
const FFTSample *in1, *in2;
137
FFTComplex *z = (FFTComplex *)tmp;
144
asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
148
in2 = input + n2 - 4;
150
/* Complex multiplication
151
Two complex products per iteration, we could have 4 with 8 xmm
152
registers, 8 with 16 xmm registers.
153
Maybe we should unroll more.
155
for (k = 0; k < n4; k += 2) {
157
"movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2
158
"movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1
159
"movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
160
"movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
161
"shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
162
"shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
163
"unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
164
"movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
165
"xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
166
"mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
167
"shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
168
"mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
169
"addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
170
::"m"(in2[-2*k]), "m"(in1[2*k]),
171
"m"(tcos[k]), "m"(tsin[k])
173
/* Should be in the same block, hack for gcc2.95 & gcc3 */
175
"movlps %%xmm0, %0 \n\t"
176
"movhps %%xmm0, %1 \n\t"
177
:"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]])
181
ff_fft_calc_sse(&s->fft, z);
183
/* Not currently needed, added for safety */
184
asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
186
/* post rotation + reordering */
187
for (k = 0; k < n4; k += 2) {
189
"movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z
190
"movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
191
"movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0
192
"movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
193
"shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
194
"shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
195
"unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
196
"movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
197
"xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
198
"mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
199
"shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
200
"mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
201
"addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
202
"movaps %%xmm0, %0 \n\t"
204
:"m"(tcos[k]), "m"(tsin[k])
220
asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
223
"movaps -16(%4,%0), %%xmm1 \n\t" // xmm1 = 4 5 6 7 = z[-2-k]
225
"movaps (%4,%0), %%xmm0 \n\t" // xmm0 = 0 1 2 3 = z[k]
226
"xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3
227
"movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3
228
"shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6
229
"shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7
230
"shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 !
231
"shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 !
232
"movaps %%xmm0, (%1,%0) \n\t" // output[2*k]
233
"movaps %%xmm2, (%2,%0) \n\t" // output[n2+2*k]
235
"shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1
236
"xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 !
237
"shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 !
238
"movaps %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k]
239
"movaps %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k]
243
:"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)