2 * FFT/MDCT transform with SSE optimizations
3 * Copyright (c) 2002 Fabrice Bellard.
5 * This file is part of FFmpeg.
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.
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.
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
24 static const int p1p1p1m1[4] __attribute__((aligned(16))) =
27 static const int p1p1m1p1[4] __attribute__((aligned(16))) =
30 static const int p1p1m1m1[4] __attribute__((aligned(16))) =
31 { 0, 0, 1 << 31, 1 << 31 };
33 static const int p1m1p1m1[4] __attribute__((aligned(16))) =
34 { 0, 1 << 31, 0, 1 << 31 };
36 static const int m1m1m1m1[4] __attribute__((aligned(16))) =
37 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
40 static void print_v4sf(const char *str, __m128 a)
42 float *p = (float *)&a;
43 printf("%s: %f %f %f %f\n",
44 str, p[0], p[1], p[2], p[3]);
48 /* XXX: handle reverse case */
49 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
58 "movaps %0, %%xmm4 \n\t"
59 "movaps %1, %%xmm5 \n\t"
61 "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
68 /* do the pass 0 butterfly */
69 "movaps (%0,%1), %%xmm0 \n\t"
70 "movaps %%xmm0, %%xmm1 \n\t"
71 "shufps $0x4E, %%xmm0, %%xmm0 \n\t"
72 "xorps %%xmm4, %%xmm1 \n\t"
73 "addps %%xmm1, %%xmm0 \n\t"
74 "movaps 16(%0,%1), %%xmm2 \n\t"
75 "movaps %%xmm2, %%xmm3 \n\t"
76 "shufps $0x4E, %%xmm2, %%xmm2 \n\t"
77 "xorps %%xmm4, %%xmm3 \n\t"
78 "addps %%xmm3, %%xmm2 \n\t"
79 /* multiply third by -i */
80 /* by toggling the sign bit */
81 "shufps $0xB4, %%xmm2, %%xmm2 \n\t"
82 "xorps %%xmm5, %%xmm2 \n\t"
83 /* do the pass 1 butterfly */
84 "movaps %%xmm0, %%xmm1 \n\t"
85 "addps %%xmm2, %%xmm0 \n\t"
86 "subps %%xmm2, %%xmm1 \n\t"
87 "movaps %%xmm0, (%0,%1) \n\t"
88 "movaps %%xmm1, 16(%0,%1) \n\t"
95 nblocks = 1 << (ln-3);
106 "movaps (%2,%0), %%xmm1 \n\t"
107 "movaps (%1,%0), %%xmm0 \n\t"
108 "movaps 16(%2,%0), %%xmm5 \n\t"
109 "movaps 16(%1,%0), %%xmm4 \n\t"
110 "movaps %%xmm1, %%xmm2 \n\t"
111 "movaps %%xmm5, %%xmm6 \n\t"
112 "shufps $0xA0, %%xmm1, %%xmm1 \n\t"
113 "shufps $0xF5, %%xmm2, %%xmm2 \n\t"
114 "shufps $0xA0, %%xmm5, %%xmm5 \n\t"
115 "shufps $0xF5, %%xmm6, %%xmm6 \n\t"
116 "mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re
117 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
118 "mulps 32(%3,%0,2), %%xmm5 \n\t" // cre*re cim*re
119 "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
120 "addps %%xmm2, %%xmm1 \n\t"
121 "addps %%xmm6, %%xmm5 \n\t"
122 "movaps %%xmm0, %%xmm3 \n\t"
123 "movaps %%xmm4, %%xmm7 \n\t"
124 "addps %%xmm1, %%xmm0 \n\t"
125 "subps %%xmm1, %%xmm3 \n\t"
126 "addps %%xmm5, %%xmm4 \n\t"
127 "subps %%xmm5, %%xmm7 \n\t"
128 "movaps %%xmm0, (%1,%0) \n\t"
129 "movaps %%xmm3, (%2,%0) \n\t"
130 "movaps %%xmm4, 16(%1,%0) \n\t"
131 "movaps %%xmm7, 16(%2,%0) \n\t"
134 :"r"(p), "r"(p + nloops), "r"(cptr)
141 } while (nblocks != 0);
144 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
145 const FFTSample *input, FFTSample *tmp)
149 const uint16_t *revtab = s->fft.revtab;
150 const FFTSample *tcos = s->tcos;
151 const FFTSample *tsin = s->tsin;
152 const FFTSample *in1, *in2;
153 FFTComplex *z = (FFTComplex *)tmp;
161 asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
162 #define P1M1P1M1 "%%xmm8"
164 #define P1M1P1M1 "%4"
169 in2 = input + n2 - 4;
171 /* Complex multiplication */
172 for (k = 0; k < n4; k += 4) {
174 "movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2
175 "movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1
176 "movaps -16+1*%0, %%xmm4 \n\t" // xmm4 = r0 X r1 X : in2
177 "movaps 16+1*%1, %%xmm7 \n\t" // xmm7 = X i1 X i0: in1
178 "movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
179 "movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
180 "movlps 8+1*%2, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
181 "movlps 8+1*%3, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
182 "shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
183 "shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
184 "shufps $95, %%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
185 "shufps $160,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
186 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
187 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
188 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
189 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
190 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
191 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
192 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
193 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
194 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
195 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
196 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
197 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
198 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
199 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
200 ::"m"(in2[-2*k]), "m"(in1[2*k]),
201 "m"(tcos[k]), "m"(tsin[k])
206 /* Should be in the same block, hack for gcc2.95 & gcc3 */
208 "movlps %%xmm0, %0 \n\t"
209 "movhps %%xmm0, %1 \n\t"
210 "movlps %%xmm4, %2 \n\t"
211 "movhps %%xmm4, %3 \n\t"
212 :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
213 "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
217 ff_fft_calc_sse(&s->fft, z);
221 #define P1M1P1M1 "%3"
224 /* post rotation + reordering */
225 for (k = 0; k < n4; k += 4) {
227 "movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z
228 "movaps 16+1*%0, %%xmm4 \n\t" // xmm4 = i1 r1 i0 r0: z
229 "movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
230 "movlps 8+1*%1, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
231 "movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0
232 "movaps %%xmm4, %%xmm7 \n\t" // xmm7 = i1 r1 i0 r0
233 "movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
234 "movlps 8+1*%2, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
235 "shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
236 "shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
237 "shufps $160,%%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
238 "shufps $245,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
239 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
240 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
241 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
242 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
243 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
244 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
245 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
246 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
247 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
248 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
249 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
250 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
251 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
252 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
253 "movaps %%xmm0, %0 \n\t"
254 "movaps %%xmm4, 16+1*%0\n\t"
256 :"m"(tcos[k]), "m"(tsin[k])
275 asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
278 "movaps -16(%4,%0), %%xmm1 \n\t" // xmm1 = 4 5 6 7 = z[-2-k]
280 "movaps (%4,%0), %%xmm0 \n\t" // xmm0 = 0 1 2 3 = z[k]
281 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3
282 "movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3
283 "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6
284 "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7
285 "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 !
286 "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 !
287 "movaps %%xmm0, (%1,%0) \n\t" // output[2*k]
288 "movaps %%xmm2, (%2,%0) \n\t" // output[n2+2*k]
290 "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1
291 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 !
292 "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 !
293 "movaps %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k]
294 "movaps %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k]
298 :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)