Mercurial > libavcodec.hg
annotate i386/fft_sse.c @ 3886:e96794bc2b0a libavcodec
Original Commit: r86 | ods15 | 2006-09-29 21:08:42 +0300 (Fri, 29 Sep 2006) | 2 lines
make put_vector ignore unused codebook entries
author | ods15 |
---|---|
date | Mon, 02 Oct 2006 06:08:48 +0000 |
parents | 2ec498208c6a |
children | c8c591fe26f8 |
rev | line source |
---|---|
781 | 1 /* |
2 * FFT/MDCT transform with SSE optimizations | |
3 * Copyright (c) 2002 Fabrice Bellard. | |
4 * | |
5 * This library is free software; you can redistribute it and/or | |
6 * modify it under the terms of the GNU Lesser General Public | |
7 * License as published by the Free Software Foundation; either | |
8 * version 2 of the License, or (at your option) any later version. | |
9 * | |
10 * This library is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 * Lesser General Public License for more details. | |
14 * | |
15 * 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 | |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
781 | 18 */ |
19 #include "../dsputil.h" | |
20 | |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
21 static const int p1p1p1m1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
22 { 0, 0, 0, 1 << 31 }; |
781 | 23 |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
24 static const int p1p1m1p1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
25 { 0, 0, 1 << 31, 0 }; |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
26 |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
27 static const int p1p1m1m1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
28 { 0, 0, 1 << 31, 1 << 31 }; |
781 | 29 |
3746 | 30 static const int p1m1p1m1[4] __attribute__((aligned(16))) = |
31 { 0, 1 << 31, 0, 1 << 31 }; | |
32 | |
33 static const int m1m1m1m1[4] __attribute__((aligned(16))) = | |
34 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 }; | |
35 | |
781 | 36 #if 0 |
37 static void print_v4sf(const char *str, __m128 a) | |
38 { | |
39 float *p = (float *)&a; | |
40 printf("%s: %f %f %f %f\n", | |
41 str, p[0], p[1], p[2], p[3]); | |
42 } | |
43 #endif | |
44 | |
45 /* XXX: handle reverse case */ | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
968
diff
changeset
|
46 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z) |
781 | 47 { |
48 int ln = s->nbits; | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
49 long i, j; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
50 long nblocks, nloops; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
51 FFTComplex *p, *cptr; |
781 | 52 |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
53 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
54 "movaps %0, %%xmm4 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
55 "movaps %1, %%xmm5 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
56 ::"m"(*p1p1m1m1), |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
57 "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1)) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
58 ); |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
59 |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
60 i = 8 << ln; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
61 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
62 "1: \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
63 "sub $32, %0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
64 /* do the pass 0 butterfly */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
65 "movaps (%0,%1), %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
66 "movaps %%xmm0, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
67 "shufps $0x4E, %%xmm0, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
68 "xorps %%xmm4, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
69 "addps %%xmm1, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
70 "movaps 16(%0,%1), %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
71 "movaps %%xmm2, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
72 "shufps $0x4E, %%xmm2, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
73 "xorps %%xmm4, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
74 "addps %%xmm3, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
75 /* multiply third by -i */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
76 /* by toggling the sign bit */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
77 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
78 "xorps %%xmm5, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
79 /* do the pass 1 butterfly */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
80 "movaps %%xmm0, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
81 "addps %%xmm2, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
82 "subps %%xmm2, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
83 "movaps %%xmm0, (%0,%1) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
84 "movaps %%xmm1, 16(%0,%1) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
85 "jg 1b \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
86 :"+r"(i) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
87 :"r"(z) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
88 ); |
781 | 89 /* pass 2 .. ln-1 */ |
90 | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
91 nblocks = 1 << (ln-3); |
781 | 92 nloops = 1 << 2; |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
93 cptr = s->exptab1; |
781 | 94 do { |
95 p = z; | |
96 j = nblocks; | |
97 do { | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
98 i = nloops*8; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
99 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
100 "1: \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
101 "sub $16, %0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
102 "movaps (%2,%0), %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
103 "movaps (%1,%0), %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
104 "movaps %%xmm1, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
105 "shufps $0xA0, %%xmm1, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
106 "shufps $0xF5, %%xmm2, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
107 "mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
108 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
109 "addps %%xmm2, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
110 "movaps %%xmm0, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
111 "addps %%xmm1, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
112 "subps %%xmm1, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
113 "movaps %%xmm0, (%1,%0) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
114 "movaps %%xmm3, (%2,%0) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
115 "jg 1b \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
116 :"+r"(i) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
117 :"r"(p), "r"(p + nloops), "r"(cptr) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
118 ); |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
119 p += nloops*2; |
781 | 120 } while (--j); |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
121 cptr += nloops*2; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
122 nblocks >>= 1; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
123 nloops <<= 1; |
781 | 124 } while (nblocks != 0); |
125 } | |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
126 |
3746 | 127 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, |
128 const FFTSample *input, FFTSample *tmp) | |
129 { | |
130 long k, n8, n4, n2, n; | |
131 const uint16_t *revtab = s->fft.revtab; | |
132 const FFTSample *tcos = s->tcos; | |
133 const FFTSample *tsin = s->tsin; | |
134 const FFTSample *in1, *in2; | |
135 FFTComplex *z = (FFTComplex *)tmp; | |
136 | |
137 n = 1 << s->nbits; | |
138 n2 = n >> 1; | |
139 n4 = n >> 2; | |
140 n8 = n >> 3; | |
141 | |
142 asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1)); | |
143 | |
144 /* pre rotation */ | |
145 in1 = input; | |
146 in2 = input + n2 - 4; | |
147 | |
148 /* Complex multiplication | |
149 Two complex products per iteration, we could have 4 with 8 xmm | |
150 registers, 8 with 16 xmm registers. | |
151 Maybe we should unroll more. | |
152 */ | |
153 for (k = 0; k < n4; k += 2) { | |
154 asm volatile ( | |
155 "movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2 | |
156 "movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1 | |
157 "movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos | |
158 "movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin | |
159 "shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0 | |
160 "shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0 | |
161 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0 | |
162 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0 | |
163 "xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0 | |
164 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR | |
165 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0 | |
166 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii | |
167 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result | |
168 ::"m"(in2[-2*k]), "m"(in1[2*k]), | |
169 "m"(tcos[k]), "m"(tsin[k]) | |
170 ); | |
171 /* Should be in the same block, hack for gcc2.95 & gcc3 */ | |
172 asm ( | |
173 "movlps %%xmm0, %0 \n\t" | |
174 "movhps %%xmm0, %1 \n\t" | |
175 :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]) | |
176 ); | |
177 } | |
178 | |
179 ff_fft_calc_sse(&s->fft, z); | |
180 | |
181 /* Not currently needed, added for safety */ | |
182 asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1)); | |
183 | |
184 /* post rotation + reordering */ | |
185 for (k = 0; k < n4; k += 2) { | |
186 asm ( | |
187 "movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z | |
188 "movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos | |
189 "movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0 | |
190 "movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin | |
191 "shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0 | |
192 "shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0 | |
193 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0 | |
194 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0 | |
195 "xorps %%xmm7, %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0 | |
196 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR | |
197 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0 | |
198 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii | |
199 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result | |
200 "movaps %%xmm0, %0 \n\t" | |
201 :"+m"(z[k]) | |
202 :"m"(tcos[k]), "m"(tsin[k]) | |
203 ); | |
204 } | |
205 | |
206 /* | |
207 Mnemonics: | |
208 0 = z[k].re | |
209 1 = z[k].im | |
210 2 = z[k + 1].re | |
211 3 = z[k + 1].im | |
212 4 = z[-k - 2].re | |
213 5 = z[-k - 2].im | |
214 6 = z[-k - 1].re | |
215 7 = z[-k - 1].im | |
216 */ | |
217 k = 16-n; | |
218 asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1)); | |
219 asm volatile( | |
220 "1: \n\t" | |
221 "movaps -16(%4,%0), %%xmm1 \n\t" // xmm1 = 4 5 6 7 = z[-2-k] | |
222 "neg %0 \n\t" | |
223 "movaps (%4,%0), %%xmm0 \n\t" // xmm0 = 0 1 2 3 = z[k] | |
224 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3 | |
225 "movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3 | |
226 "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6 | |
227 "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7 | |
228 "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 ! | |
229 "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 ! | |
230 "movaps %%xmm0, (%1,%0) \n\t" // output[2*k] | |
231 "movaps %%xmm2, (%2,%0) \n\t" // output[n2+2*k] | |
232 "neg %0 \n\t" | |
233 "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1 | |
234 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 ! | |
235 "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 ! | |
236 "movaps %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k] | |
237 "movaps %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k] | |
238 "add $16, %0 \n\t" | |
239 "jle 1b \n\t" | |
240 :"+r"(k) | |
241 :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8) | |
242 :"memory" | |
243 ); | |
244 } | |
245 |