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 Libav.
7 ;* Libav 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 ;* Libav 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 Libav; 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"
25 pf_inv16: times 4 dd 0x3D800000 ; 1/16
29 ; %1=v0/v1 %2=in1 %3=in2
37 %define OFFSET NUM_COEF*count
39 ; for v0, incrementing and for v1, decrementing
40 mova va, [cf0q + OFFSET]
41 mova vb, [cf0q + OFFSET + 4*NUM_COEF]
43 mova m4, [cf0q + OFFSET + mmsize]
44 mova m0, [cf0q + OFFSET + 4*NUM_COEF + mmsize]
54 ; va = va1 va2 va3 va4
55 ; vb = vb1 vb2 vb3 vb4
60 unpcklps va, vb ; va3 vb3 va4 vb4
61 unpckhps m4, vb ; va1 vb1 va2 vb2
62 addps m4, va ; va1+3 vb1+3 va2+4 vb2+4
63 movhlps vb, m4 ; va1+3 vb1+3
64 addps vb, m4 ; va0..4 vb0..4
65 movlps [outq + count], vb
73 ; void dca_lfe_fir(float *out, float *in, float *coefs)
75 cglobal dca_lfe_fir%1, 3,3,6-%1, out, in, cf0
79 %define NUM_COEF 4*(2-%1)
80 %define NUM_OUT 32*(%1+1)
82 movu IN1, [inq + 4 - 1*mmsize]
83 shufps IN1, IN1, q0123
85 movu IN2, [inq + 4 - 2*mmsize]
86 shufps IN2, IN2, q0123
90 add cf0q, 4*NUM_COEF*NUM_OUT
98 shufps IN1, IN1, q0123
100 ; cf1 already correctly positioned
101 add outq, 4*NUM_OUT ; outq now at out2
104 shufps IN2, IN2, q0123
117 %if cpuflag(sse2) && notcpuflag(avx)
127 vperm2f128 %1, %3, %3, 1
128 vshufps %1, %1, %1, q0123
130 pshufd %1, [%2], q0123
138 ; reading backwards: ptr1 = synth_buf + j + i; ptr2 = synth_buf + j - i
139 ;~ a += window[i + j] * (-synth_buf[15 - i + j])
140 ;~ b += window[i + j + 16] * (synth_buf[i + j])
141 SHUF m5, ptr2 + j + (15 - 3) * 4, m6
144 SHUF m11, ptr2 + j + (15 - 3) * 4 - mmsize, m12
145 mova m12, [ptr1 + j + mmsize]
148 fmaddps m2, m6, [win + %1 + j + 16 * 4], m2
149 fnmaddps m1, m5, [win + %1 + j], m1
151 fmaddps m8, m12, [win + %1 + j + mmsize + 16 * 4], m8
152 fnmaddps m7, m11, [win + %1 + j + mmsize], m7
155 mulps m6, m6, [win + %1 + j + 16 * 4]
156 mulps m5, m5, [win + %1 + j]
158 mulps m12, m12, [win + %1 + j + mmsize + 16 * 4]
159 mulps m11, m11, [win + %1 + j + mmsize]
167 %endif ; cpuflag(fma3)
168 ;~ c += window[i + j + 32] * (synth_buf[16 + i + j])
169 ;~ d += window[i + j + 48] * (synth_buf[31 - i + j])
170 SHUF m6, ptr2 + j + (31 - 3) * 4, m5
171 mova m5, [ptr1 + j + 16 * 4]
173 SHUF m12, ptr2 + j + (31 - 3) * 4 - mmsize, m11
174 mova m11, [ptr1 + j + mmsize + 16 * 4]
177 fmaddps m3, m5, [win + %1 + j + 32 * 4], m3
178 fmaddps m4, m6, [win + %1 + j + 48 * 4], m4
180 fmaddps m9, m11, [win + %1 + j + mmsize + 32 * 4], m9
181 fmaddps m10, m12, [win + %1 + j + mmsize + 48 * 4], m10
184 mulps m5, m5, [win + %1 + j + 32 * 4]
185 mulps m6, m6, [win + %1 + j + 48 * 4]
187 mulps m11, m11, [win + %1 + j + mmsize + 32 * 4]
188 mulps m12, m12, [win + %1 + j + mmsize + 48 * 4]
196 %endif ; cpuflag(fma3)
200 ; void ff_synth_filter_inner_<opt>(float *synth_buf, float synth_buf2[32],
201 ; const float window[512], float out[32],
202 ; intptr_t offset, float scale)
203 %macro SYNTH_FILTER 0
204 cglobal synth_filter_inner, 0, 6 + 4 * ARCH_X86_64, 7 + 6 * ARCH_X86_64, \
205 synth_buf, synth_buf2, window, out, off, scale
207 %if ARCH_X86_32 || WIN64
208 %if cpuflag(sse2) && notcpuflag(avx)
212 VBROADCASTSS m0, scalem
214 ; Make sure offset is in a register and not on the stack
219 vinsertf128 m0, m0, xmm0, 1
223 ; prepare inner counter limit 1
228 %if ARCH_X86_32 || notcpuflag(avx)
231 mov i, 16 * 4 - (ARCH_X86_64 + 1) * mmsize ; main loop counter
237 %define buf2 synth_buf2q
239 mov buf2, synth_buf2mp
242 ; m1 = a m2 = b m3 = c m4 = d
246 mova m2, [buf2 + i + 16 * 4]
254 %if ARCH_X86_32 || notcpuflag(avx)
260 %define ptr2 r7q ; must be loaded
265 mova m7, [buf2 + i + mmsize]
266 mova m8, [buf2 + i + mmsize + 16 * 4]
267 lea win, [windowq + i]
268 lea ptr1, [synth_bufq + i]
270 mov ptr2, synth_bufmp
271 ; prepare the inner loop counter
273 %if ARCH_X86_32 || notcpuflag(avx)
285 add win, OFFQ ; now at j-64, so define OFFSET
293 mov buf2, synth_buf2m ; needed for next iteration anyway
294 mov outq, outmp ; j, which will be set again during it
296 ;~ out[i] = a * scale;
297 ;~ out[i + 16] = b * scale;
304 ;~ synth_buf2[i] = c;
305 ;~ synth_buf2[i + 16] = d;
306 mova [buf2 + i + 0 * 4], m3
307 mova [buf2 + i + 16 * 4], m4
309 mova [buf2 + i + 0 * 4 + mmsize], m9
310 mova [buf2 + i + 16 * 4 + mmsize], m10
314 mova [outq + i + 0 * 4], m1
315 mova [outq + i + 16 * 4], m2
317 mova [outq + i + 0 * 4 + mmsize], m7
318 mova [outq + i + 16 * 4 + mmsize], m8
320 %if ARCH_X86_32 || notcpuflag(avx)
321 sub i, (ARCH_X86_64 + 1) * mmsize