1 ;******************************************************************************
2 ;* SIMD-optimized functions for the DCA decoder
3 ;* Copyright (C) 2016 James Almer
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
20 ;******************************************************************************
22 %include "libavutil/x86/x86util.asm"
26 %define sizeof_float 4
27 %define FMA3_OFFSET (8 * cpuflag(fma3))
29 %macro LFE_FIR0_FLOAT 0
30 cglobal lfe_fir0_float, 4, 6, 12 + cpuflag(fma3)*4, samples, lfe, coeff, nblocks, cnt1, cnt2
32 sub lfeq, 7*sizeof_float
33 mov cnt1d, 32*sizeof_float
34 mov cnt2d, 32*sizeof_float-8-FMA3_OFFSET
35 lea coeffq, [coeffq+cnt1q*8]
41 cvtdq2ps m4, [lfeq+16]
43 shufps m7, m4, m4, q0123
44 shufps m6, m5, m5, q0123
53 cvtpi2ps m4, [lfeq+16]
54 cvtpi2ps m0, [lfeq+24]
56 cvtpi2ps m1, [lfeq+8 ]
59 shufps m7, m4, m4, q0123
60 shufps m6, m5, m5, q0123
65 movaps m8, [coeffq+cnt1q*8 ]
66 movaps m9, [coeffq+cnt1q*8+16]
67 movaps m10, [coeffq+cnt1q*8+32]
68 movaps m11, [coeffq+cnt1q*8+48]
70 movaps m12, [coeffq+cnt1q*8+64]
71 movaps m13, [coeffq+cnt1q*8+80]
72 movaps m14, [coeffq+cnt1q*8+96]
73 movaps m15, [coeffq+cnt1q*8+112]
78 fmaddps m0, m6, m9, m0
79 fmaddps m1, m6, m11, m1
80 fmaddps m2, m6, m13, m2
81 fmaddps m3, m6, m15, m3
86 movaps [samplesq+cnt1q], m0
100 movlps [samplesq+cnt1q], m2
104 mulps m0, m7, [coeffq+cnt1q*8 ]
105 mulps m1, m7, [coeffq+cnt1q*8+32 ]
106 mulps m2, m7, [coeffq+cnt1q*8+64 ]
107 mulps m3, m7, [coeffq+cnt1q*8+96 ]
108 fmaddps m0, m6, [coeffq+cnt1q*8+16 ], m0
109 fmaddps m1, m6, [coeffq+cnt1q*8+48 ], m1
110 fmaddps m2, m6, [coeffq+cnt1q*8+80 ], m2
111 fmaddps m3, m6, [coeffq+cnt1q*8+112], m3
116 movaps [samplesq+cnt1q], m0
118 mulps m0, m7, [coeffq+cnt1q*8 ]
119 mulps m1, m6, [coeffq+cnt1q*8+16]
120 mulps m2, m7, [coeffq+cnt1q*8+32]
121 mulps m3, m6, [coeffq+cnt1q*8+48]
130 movlps [samplesq+cnt1q], m2
140 fmaddps m8, m4, m9, m8
141 fmaddps m10, m4, m11, m10
142 fmaddps m12, m4, m13, m12
143 fmaddps m14, m4, m15, m14
148 movaps [samplesq+cnt2q], m14
157 unpckhps m11, m10, m8
162 movlps [samplesq+cnt2q], m8
166 mulps m0, m5, [coeffq+cnt1q*8 ]
167 mulps m1, m5, [coeffq+cnt1q*8+32 ]
168 mulps m2, m5, [coeffq+cnt1q*8+64 ]
169 mulps m3, m5, [coeffq+cnt1q*8+96 ]
170 fmaddps m0, m4, [coeffq+cnt1q*8+16 ], m0
171 fmaddps m1, m4, [coeffq+cnt1q*8+48 ], m1
172 fmaddps m2, m4, [coeffq+cnt1q*8+80 ], m2
173 fmaddps m3, m4, [coeffq+cnt1q*8+112], m3
178 movaps [samplesq+cnt2q], m3
180 mulps m0, m5, [coeffq+cnt1q*8 ]
181 mulps m1, m4, [coeffq+cnt1q*8+16]
182 mulps m2, m5, [coeffq+cnt1q*8+32]
183 mulps m3, m4, [coeffq+cnt1q*8+48]
192 movlps [samplesq+cnt2q], m0
196 sub cnt2d, 8 + FMA3_OFFSET
197 add cnt1q, 8 + FMA3_OFFSET
201 add samplesq, 64*sizeof_float
202 mov cnt1q, -32*sizeof_float
203 mov cnt2d, 32*sizeof_float-8-FMA3_OFFSET
215 %if HAVE_AVX_EXTERNAL
219 %if HAVE_FMA3_EXTERNAL
224 %macro LFE_FIR1_FLOAT 0
225 cglobal lfe_fir1_float, 4, 6, 10, samples, lfe, coeff, nblocks, cnt1, cnt2
227 sub lfeq, 3*sizeof_float
228 mov cnt1d, 64*sizeof_float
229 mov cnt2d, 64*sizeof_float-16
230 lea coeffq, [coeffq+cnt1q*4]
237 shufps m5, m4, m4, q0123
245 movaps m6, [coeffq+cnt1q*4 ]
246 movaps m7, [coeffq+cnt1q*4+16]
250 movaps m8, [coeffq+cnt1q*4+32]
251 movaps m9, [coeffq+cnt1q*4+48]
255 mulps m2, m5, [coeffq+cnt1q*4+32]
256 mulps m3, m5, [coeffq+cnt1q*4+48]
262 movaps [samplesq+cnt1q], m0
274 mulps m2, m4, [coeffq+cnt1q*4+32]
275 mulps m3, m4, [coeffq+cnt1q*4+48]
281 movaps [samplesq+cnt2q], m6
287 add lfeq, sizeof_float
288 add samplesq, 128*sizeof_float
289 mov cnt1q, -64*sizeof_float
290 mov cnt2d, 64*sizeof_float-16
298 %if HAVE_AVX_EXTERNAL