1 ;******************************************************************************
2 ;* SSE-optimized functions for the DCA decoder
3 ;* Copyright (C) 2012-2014 Christophe Gisquet <christophe.gisquet@gmail.com>
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"
27 %if cpuflag(sse2) && notcpuflag(avx)
37 vperm2f128 %1, %3, %3, 1
38 vshufps %1, %1, %1, q0123
40 pshufd %1, [%2], q0123
48 ; reading backwards: ptr1 = synth_buf + j + i; ptr2 = synth_buf + j - i
49 ;~ a += window[i + j] * (-synth_buf[15 - i + j])
50 ;~ b += window[i + j + 16] * (synth_buf[i + j])
51 SHUF m5, ptr2 + j + (15 - 3) * 4, m6
54 SHUF m11, ptr2 + j + (15 - 3) * 4 - mmsize, m12
55 mova m12, [ptr1 + j + mmsize]
58 fmaddps m2, m6, [win + %1 + j + 16 * 4], m2
59 fnmaddps m1, m5, [win + %1 + j], m1
61 fmaddps m8, m12, [win + %1 + j + mmsize + 16 * 4], m8
62 fnmaddps m7, m11, [win + %1 + j + mmsize], m7
65 mulps m6, m6, [win + %1 + j + 16 * 4]
66 mulps m5, m5, [win + %1 + j]
68 mulps m12, m12, [win + %1 + j + mmsize + 16 * 4]
69 mulps m11, m11, [win + %1 + j + mmsize]
77 %endif ; cpuflag(fma3)
78 ;~ c += window[i + j + 32] * (synth_buf[16 + i + j])
79 ;~ d += window[i + j + 48] * (synth_buf[31 - i + j])
80 SHUF m6, ptr2 + j + (31 - 3) * 4, m5
81 mova m5, [ptr1 + j + 16 * 4]
83 SHUF m12, ptr2 + j + (31 - 3) * 4 - mmsize, m11
84 mova m11, [ptr1 + j + mmsize + 16 * 4]
87 fmaddps m3, m5, [win + %1 + j + 32 * 4], m3
88 fmaddps m4, m6, [win + %1 + j + 48 * 4], m4
90 fmaddps m9, m11, [win + %1 + j + mmsize + 32 * 4], m9
91 fmaddps m10, m12, [win + %1 + j + mmsize + 48 * 4], m10
94 mulps m5, m5, [win + %1 + j + 32 * 4]
95 mulps m6, m6, [win + %1 + j + 48 * 4]
97 mulps m11, m11, [win + %1 + j + mmsize + 32 * 4]
98 mulps m12, m12, [win + %1 + j + mmsize + 48 * 4]
106 %endif ; cpuflag(fma3)
110 ; void ff_synth_filter_inner_<opt>(float *synth_buf, float synth_buf2[32],
111 ; const float window[512], float out[32],
112 ; intptr_t offset, float scale)
113 %macro SYNTH_FILTER 0
114 cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
115 synth_buf, synth_buf2, window, out, off, scale
117 %if ARCH_X86_32 || WIN64
118 %if cpuflag(sse2) && notcpuflag(avx)
122 VBROADCASTSS m0, scalem
124 ; Make sure offset is in a register and not on the stack
129 vinsertf128 m0, m0, xmm0, 1
133 ; prepare inner counter limit 1
138 %if ARCH_X86_32 || notcpuflag(avx)
141 mov i, 16 * 4 - (ARCH_X86_64 + 1) * mmsize ; main loop counter
147 %define buf2 synth_buf2q
149 mov buf2, synth_buf2mp
152 ; m1 = a m2 = b m3 = c m4 = d
156 mova m2, [buf2 + i + 16 * 4]
164 %if ARCH_X86_32 || notcpuflag(avx)
170 %define ptr2 r7q ; must be loaded
175 mova m7, [buf2 + i + mmsize]
176 mova m8, [buf2 + i + mmsize + 16 * 4]
177 lea win, [windowq + i]
178 lea ptr1, [synth_bufq + i]
180 mov ptr2, synth_bufmp
181 ; prepare the inner loop counter
183 %if ARCH_X86_32 || notcpuflag(avx)
195 add win, OFFQ ; now at j-64, so define OFFSET
203 mov buf2, synth_buf2m ; needed for next iteration anyway
204 mov outq, outmp ; j, which will be set again during it
206 ;~ out[i] = a * scale;
207 ;~ out[i + 16] = b * scale;
214 ;~ synth_buf2[i] = c;
215 ;~ synth_buf2[i + 16] = d;
216 mova [buf2 + i + 0 * 4], m3
217 mova [buf2 + i + 16 * 4], m4
219 mova [buf2 + i + 0 * 4 + mmsize], m9
220 mova [buf2 + i + 16 * 4 + mmsize], m10
224 mova [outq + i + 0 * 4], m1
225 mova [outq + i + 16 * 4], m2
227 mova [outq + i + 0 * 4 + mmsize], m7
228 mova [outq + i + 16 * 4 + mmsize], m8
230 %if ARCH_X86_32 || notcpuflag(avx)
231 sub i, (ARCH_X86_64 + 1) * mmsize