1 /*
2  * FFT/MDCT transform with SSE optimizations
3  * Copyright (c) 2002 Fabrice Bellard.
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
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
20  */
21 #include "dsputil.h"
22 
23 static const int p1p1p1m1[4] __attribute__((aligned(16))) =
24     { 0, 0, 0, 1 << 31 };
25 
26 static const int p1p1m1p1[4] __attribute__((aligned(16))) =
27     { 0, 0, 1 << 31, 0 };
28 
29 static const int p1p1m1m1[4] __attribute__((aligned(16))) =
30     { 0, 0, 1 << 31, 1 << 31 };
31 
32 static const int p1m1p1m1[4] __attribute__((aligned(16))) =
33     { 0, 1 << 31, 0, 1 << 31 };
34 
35 static const int m1m1m1m1[4] __attribute__((aligned(16))) =
36     { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
37 
38 #if 0
39 static void print_v4sf(const char *str, __m128 a)
40 {
41     float *p = (float *)&a;
42     printf("%s: %f %f %f %f\n",
43            str, p[0], p[1], p[2], p[3]);
44 }
45 #endif
46 
47 /* XXX: handle reverse case */
ff_fft_calc_sse(FFTContext * s,FFTComplex * z)48 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
49 {
50     int ln = s->nbits;
51     long i, j;
52     long nblocks, nloops;
53     FFTComplex *p, *cptr;
54 
55     asm volatile(
56         "movaps %0, %%xmm4 \n\t"
57         "movaps %1, %%xmm5 \n\t"
58         ::"m"(*p1p1m1m1),
59           "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
60     );
61 
62     i = 8 << ln;
63     asm volatile(
64         "1: \n\t"
65         "sub $32, %0 \n\t"
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"
87         "jg 1b \n\t"
88         :"+r"(i)
89         :"r"(z)
90     );
91     /* pass 2 .. ln-1 */
92 
93     nblocks = 1 << (ln-3);
94     nloops = 1 << 2;
95     cptr = s->exptab1;
96     do {
97         p = z;
98         j = nblocks;
99         do {
100             i = nloops*8;
101             asm volatile(
102                 "1: \n\t"
103                 "sub $32, %0 \n\t"
104                 "movaps    (%2,%0), %%xmm1 \n\t"
105                 "movaps    (%1,%0), %%xmm0 \n\t"
106                 "movaps  16(%2,%0), %%xmm5 \n\t"
107                 "movaps  16(%1,%0), %%xmm4 \n\t"
108                 "movaps     %%xmm1, %%xmm2 \n\t"
109                 "movaps     %%xmm5, %%xmm6 \n\t"
110                 "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
111                 "shufps      $0xF5, %%xmm2, %%xmm2 \n\t"
112                 "shufps      $0xA0, %%xmm5, %%xmm5 \n\t"
113                 "shufps      $0xF5, %%xmm6, %%xmm6 \n\t"
114                 "mulps   (%3,%0,2), %%xmm1 \n\t" //  cre*re cim*re
115                 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
116                 "mulps 32(%3,%0,2), %%xmm5 \n\t" //  cre*re cim*re
117                 "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
118                 "addps      %%xmm2, %%xmm1 \n\t"
119                 "addps      %%xmm6, %%xmm5 \n\t"
120                 "movaps     %%xmm0, %%xmm3 \n\t"
121                 "movaps     %%xmm4, %%xmm7 \n\t"
122                 "addps      %%xmm1, %%xmm0 \n\t"
123                 "subps      %%xmm1, %%xmm3 \n\t"
124                 "addps      %%xmm5, %%xmm4 \n\t"
125                 "subps      %%xmm5, %%xmm7 \n\t"
126                 "movaps     %%xmm0, (%1,%0) \n\t"
127                 "movaps     %%xmm3, (%2,%0) \n\t"
128                 "movaps     %%xmm4, 16(%1,%0) \n\t"
129                 "movaps     %%xmm7, 16(%2,%0) \n\t"
130                 "jg 1b \n\t"
131                 :"+r"(i)
132                 :"r"(p), "r"(p + nloops), "r"(cptr)
133             );
134             p += nloops*2;
135         } while (--j);
136         cptr += nloops*2;
137         nblocks >>= 1;
138         nloops <<= 1;
139     } while (nblocks != 0);
140 }
141 
ff_imdct_calc_sse(MDCTContext * s,FFTSample * output,const FFTSample * input,FFTSample * tmp)142 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
143                        const FFTSample *input, FFTSample *tmp)
144 {
145     long k, n8, n4, n2, n;
146     const uint16_t *revtab = s->fft.revtab;
147     const FFTSample *tcos = s->tcos;
148     const FFTSample *tsin = s->tsin;
149     const FFTSample *in1, *in2;
150     FFTComplex *z = (FFTComplex *)tmp;
151 
152     n = 1 << s->nbits;
153     n2 = n >> 1;
154     n4 = n >> 2;
155     n8 = n >> 3;
156 
157 #ifdef ARCH_X86_64
158     asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
159 #define P1M1P1M1 "%%xmm8"
160 #else
161 #define P1M1P1M1 "%4"
162 #endif
163 
164     /* pre rotation */
165     in1 = input;
166     in2 = input + n2 - 4;
167 
168     /* Complex multiplication */
169     for (k = 0; k < n4; k += 4) {
170         asm volatile (
171             "movaps          %0, %%xmm0 \n\t"   // xmm0 = r0 X  r1 X : in2
172             "movaps          %1, %%xmm3 \n\t"   // xmm3 = X  i1 X  i0: in1
173             "movaps    -16+1*%0, %%xmm4 \n\t"   // xmm4 = r0 X  r1 X : in2
174             "movaps     16+1*%1, %%xmm7 \n\t"   // xmm7 = X  i1 X  i0: in1
175             "movlps          %2, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
176             "movlps          %3, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
177             "movlps      8+1*%2, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
178             "movlps      8+1*%3, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
179             "shufps $95, %%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
180             "shufps $160,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
181             "shufps $95, %%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
182             "shufps $160,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
183             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
184             "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
185             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
186             "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
187             "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
188             "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
189             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
190             "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
191             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
192             "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
193             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
194             "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
195             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
196             "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
197             ::"m"(in2[-2*k]), "m"(in1[2*k]),
198               "m"(tcos[k]), "m"(tsin[k])
199 #ifndef ARCH_X86_64
200               ,"m"(*p1m1p1m1)
201 #endif
202         );
203         /* Should be in the same block, hack for gcc2.95 & gcc3 */
204         asm (
205             "movlps      %%xmm0, %0     \n\t"
206             "movhps      %%xmm0, %1     \n\t"
207             "movlps      %%xmm4, %2     \n\t"
208             "movhps      %%xmm4, %3     \n\t"
209             :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
210              "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
211         );
212     }
213 
214     ff_fft_calc_sse(&s->fft, z);
215 
216 #ifndef ARCH_X86_64
217 #undef P1M1P1M1
218 #define P1M1P1M1 "%3"
219 #endif
220 
221     /* post rotation + reordering */
222     for (k = 0; k < n4; k += 4) {
223         asm (
224             "movaps          %0, %%xmm0 \n\t"   // xmm0 = i1 r1 i0 r0: z
225             "movaps     16+1*%0, %%xmm4 \n\t"   // xmm4 = i1 r1 i0 r0: z
226             "movlps          %1, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
227             "movlps      8+1*%1, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
228             "movaps      %%xmm0, %%xmm3 \n\t"   // xmm3 = i1 r1 i0 r0
229             "movaps      %%xmm4, %%xmm7 \n\t"   // xmm7 = i1 r1 i0 r0
230             "movlps          %2, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
231             "movlps      8+1*%2, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
232             "shufps $160,%%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
233             "shufps $245,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
234             "shufps $160,%%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
235             "shufps $245,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
236             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
237             "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
238             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
239             "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
240             "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
241             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
242             "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
243             "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
244             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
245             "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
246             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
247             "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
248             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
249             "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
250             "movaps      %%xmm0, %0     \n\t"
251             "movaps      %%xmm4, 16+1*%0\n\t"
252             :"+m"(z[k])
253             :"m"(tcos[k]), "m"(tsin[k])
254 #ifndef ARCH_X86_64
255              ,"m"(*p1m1p1m1)
256 #endif
257         );
258     }
259 
260     /*
261        Mnemonics:
262        0 = z[k].re
263        1 = z[k].im
264        2 = z[k + 1].re
265        3 = z[k + 1].im
266        4 = z[-k - 2].re
267        5 = z[-k - 2].im
268        6 = z[-k - 1].re
269        7 = z[-k - 1].im
270     */
271     k = 16-n;
272     asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
273     asm volatile(
274         "1: \n\t"
275         "movaps  -16(%4,%0), %%xmm1 \n\t"   // xmm1 = 4 5 6 7 = z[-2-k]
276         "neg %0 \n\t"
277         "movaps     (%4,%0), %%xmm0 \n\t"   // xmm0 = 0 1 2 3 = z[k]
278         "xorps       %%xmm7, %%xmm0 \n\t"   // xmm0 = -0 -1 -2 -3
279         "movaps      %%xmm0, %%xmm2 \n\t"   // xmm2 = -0 -1 -2 -3
280         "shufps $141,%%xmm1, %%xmm0 \n\t"   // xmm0 = -1 -3 4 6
281         "shufps $216,%%xmm1, %%xmm2 \n\t"   // xmm2 = -0 -2 5 7
282         "shufps $156,%%xmm0, %%xmm0 \n\t"   // xmm0 = -1 6 -3 4 !
283         "shufps $156,%%xmm2, %%xmm2 \n\t"   // xmm2 = -0 7 -2 5 !
284         "movaps      %%xmm0, (%1,%0) \n\t"  // output[2*k]
285         "movaps      %%xmm2, (%2,%0) \n\t"  // output[n2+2*k]
286         "neg %0 \n\t"
287         "shufps $27, %%xmm0, %%xmm0 \n\t"   // xmm0 = 4 -3 6 -1
288         "xorps       %%xmm7, %%xmm0 \n\t"   // xmm0 = -4 3 -6 1 !
289         "shufps $27, %%xmm2, %%xmm2 \n\t"   // xmm2 = 5 -2 7 -0 !
290         "movaps      %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k]
291         "movaps      %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k]
292         "add $16, %0 \n\t"
293         "jle 1b \n\t"
294         :"+r"(k)
295         :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
296         :"memory"
297     );
298 }
299 
300