virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-12 19:18:44 +02:00
Code Issues Releases Activity
cab876f5f4
FFmpeg/libavcodec/x86/dcadsp.asm

289 lines
7.0 KiB
NASM
Raw Normal View History

x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
;******************************************************************************
;* SIMD-optimized functions for the DCA decoder
;* Copyright (C) 2016 James Almer
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "libavutil/x86/x86util.asm"
SECTION .text
%define sizeof_float 4
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
%define FMA3_OFFSET (8 * cpuflag(fma3))
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
%macro LFE_FIR0_FLOAT 0
cglobal lfe_fir0_float, 4, 6, 12 + cpuflag(fma3)*4, samples, lfe, coeff, nblocks, cnt1, cnt2
shr nblocksd, 1
sub lfeq, 7*sizeof_float
mov cnt1d, 32*sizeof_float
mov cnt2d, 32*sizeof_float-8-FMA3_OFFSET
lea coeffq, [coeffq+cnt1q*8]
add samplesq, cnt1q
neg cnt1q
.loop:
%if cpuflag(avx)
cvtdq2ps m4, [lfeq+16]
cvtdq2ps m5, [lfeq ]
shufps m7, m4, m4, q0123
shufps m6, m5, m5, q0123
avcodec/x86/dcadsp: Remove obsolete SSE function The only systems which benefit from ff_lfe_fir0_float_sse are truely ancient 32bit x86s as all other systems use at least the SSE2 versions (this includes all x64 cpus (which is why this code is restricted to x86-32)). Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
2022-06-20 03:07:41 +02:00
%else
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
movu m4, [lfeq+16]
movu m5, [lfeq ]
cvtdq2ps m4, m4
cvtdq2ps m5, m5
pshufd m7, m4, q0123
pshufd m6, m5, q0123
%endif
.inner_loop:
%if ARCH_X86_64
movaps m8, [coeffq+cnt1q*8 ]
movaps m9, [coeffq+cnt1q*8+16]
movaps m10, [coeffq+cnt1q*8+32]
movaps m11, [coeffq+cnt1q*8+48]
%if cpuflag(fma3)
movaps m12, [coeffq+cnt1q*8+64]
movaps m13, [coeffq+cnt1q*8+80]
movaps m14, [coeffq+cnt1q*8+96]
movaps m15, [coeffq+cnt1q*8+112]
mulps m0, m7, m8
mulps m1, m7, m10
mulps m2, m7, m12
mulps m3, m7, m14
fmaddps m0, m6, m9, m0
fmaddps m1, m6, m11, m1
fmaddps m2, m6, m13, m2
fmaddps m3, m6, m15, m3
haddps m0, m1
haddps m2, m3
haddps m0, m2
movaps [samplesq+cnt1q], m0
%else
mulps m0, m7, m8
mulps m1, m6, m9
mulps m2, m7, m10
mulps m3, m6, m11
addps m0, m1
addps m2, m3
unpckhps m3, m0, m2
unpcklps m0, m2
addps m3, m0
movhlps m2, m3
addps m2, m3
movlps [samplesq+cnt1q], m2
%endif
%else ; ARCH_X86_32
%if cpuflag(fma3)
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
mulps m0, m7, [coeffq+cnt1q*8 ]
mulps m1, m7, [coeffq+cnt1q*8+32 ]
mulps m2, m7, [coeffq+cnt1q*8+64 ]
mulps m3, m7, [coeffq+cnt1q*8+96 ]
fmaddps m0, m6, [coeffq+cnt1q*8+16 ], m0
fmaddps m1, m6, [coeffq+cnt1q*8+48 ], m1
fmaddps m2, m6, [coeffq+cnt1q*8+80 ], m2
fmaddps m3, m6, [coeffq+cnt1q*8+112], m3
haddps m0, m1
haddps m2, m3
haddps m0, m2
movaps [samplesq+cnt1q], m0
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
%else
mulps m0, m7, [coeffq+cnt1q*8 ]
mulps m1, m6, [coeffq+cnt1q*8+16]
mulps m2, m7, [coeffq+cnt1q*8+32]
mulps m3, m6, [coeffq+cnt1q*8+48]
addps m0, m1
addps m2, m3
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
unpckhps m3, m0, m2
unpcklps m0, m2
addps m3, m0
movhlps m2, m3
addps m2, m3
movlps [samplesq+cnt1q], m2
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
%endif
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
%endif; ARCH
%if ARCH_X86_64
%if cpuflag(fma3)
mulps m8, m5
mulps m10, m5
mulps m12, m5
mulps m14, m5
fmaddps m8, m4, m9, m8
fmaddps m10, m4, m11, m10
fmaddps m12, m4, m13, m12
fmaddps m14, m4, m15, m14
haddps m10, m8
haddps m14, m12
haddps m14, m10
movaps [samplesq+cnt2q], m14
%else
mulps m8, m5
mulps m9, m4
mulps m10, m5
mulps m11, m4
addps m8, m9
addps m10, m11
unpckhps m11, m10, m8
unpcklps m10, m8
addps m11, m10
movhlps m8, m11
addps m8, m11
movlps [samplesq+cnt2q], m8
%endif
%else ; ARCH_X86_32
%if cpuflag(fma3)
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
mulps m0, m5, [coeffq+cnt1q*8 ]
mulps m1, m5, [coeffq+cnt1q*8+32 ]
mulps m2, m5, [coeffq+cnt1q*8+64 ]
mulps m3, m5, [coeffq+cnt1q*8+96 ]
fmaddps m0, m4, [coeffq+cnt1q*8+16 ], m0
fmaddps m1, m4, [coeffq+cnt1q*8+48 ], m1
fmaddps m2, m4, [coeffq+cnt1q*8+80 ], m2
fmaddps m3, m4, [coeffq+cnt1q*8+112], m3
haddps m1, m0
haddps m3, m2
haddps m3, m1
movaps [samplesq+cnt2q], m3
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
%else
mulps m0, m5, [coeffq+cnt1q*8 ]
mulps m1, m4, [coeffq+cnt1q*8+16]
mulps m2, m5, [coeffq+cnt1q*8+32]
mulps m3, m4, [coeffq+cnt1q*8+48]
addps m0, m1
addps m2, m3
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
unpckhps m3, m2, m0
unpcklps m2, m0
addps m3, m2
movhlps m0, m3
addps m0, m3
movlps [samplesq+cnt2q], m0
x86/dcadsp: optimize lfe_fir0_float_fma3 on x86_32 About 10% faster. Signed-off-by: James Almer <jamrial@gmail.com>
2016-07-03 23:48:22 +02:00
%endif
x86/dcadec: add ff_lfe_fir0_float_{sse,sse2,avx,fma3} Up to ~4 times faster on x86_64, ~8 times on x86_32 if compiling using x87 fp math. Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-05 21:45:04 +02:00
%endif; ARCH
sub cnt2d, 8 + FMA3_OFFSET
add cnt1q, 8 + FMA3_OFFSET
jl .inner_loop
add lfeq, 4
add samplesq, 64*sizeof_float
mov cnt1q, -32*sizeof_float
mov cnt2d, 32*sizeof_float-8-FMA3_OFFSET
sub nblocksd, 1
jg .loop
RET
%endmacro
INIT_XMM sse2
LFE_FIR0_FLOAT
%if HAVE_AVX_EXTERNAL
INIT_XMM avx
LFE_FIR0_FLOAT
%endif
%if HAVE_FMA3_EXTERNAL
INIT_XMM fma3
LFE_FIR0_FLOAT
%endif
x86/dcadec: add ff_lfe_fir1_float_{sse3,avx} Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-23 00:59:07 +02:00
%macro LFE_FIR1_FLOAT 0
cglobal lfe_fir1_float, 4, 6, 10, samples, lfe, coeff, nblocks, cnt1, cnt2
shr nblocksd, 2
sub lfeq, 3*sizeof_float
mov cnt1d, 64*sizeof_float
mov cnt2d, 64*sizeof_float-16
lea coeffq, [coeffq+cnt1q*4]
add samplesq, cnt1q
neg cnt1q
.loop:
%if cpuflag(avx)
cvtdq2ps m4, [lfeq]
shufps m5, m4, m4, q0123
avcodec/x86/dcadsp: Remove obsolete SSE function The only systems which benefit from ff_lfe_fir0_float_sse are truely ancient 32bit x86s as all other systems use at least the SSE2 versions (this includes all x64 cpus (which is why this code is restricted to x86-32)). Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
2022-06-20 03:07:41 +02:00
%else
x86/dcadec: add ff_lfe_fir1_float_{sse3,avx} Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com> Signed-off-by: James Almer <jamrial@gmail.com>
2016-02-23 00:59:07 +02:00
movu m4, [lfeq]
cvtdq2ps m4, m4
pshufd m5, m4, q0123
%endif
.inner_loop:
movaps m6, [coeffq+cnt1q*4 ]
movaps m7, [coeffq+cnt1q*4+16]
mulps m0, m5, m6
mulps m1, m5, m7
%if ARCH_X86_64
movaps m8, [coeffq+cnt1q*4+32]
movaps m9, [coeffq+cnt1q*4+48]
mulps m2, m5, m8
mulps m3, m5, m9
%else
mulps m2, m5, [coeffq+cnt1q*4+32]
mulps m3, m5, [coeffq+cnt1q*4+48]
%endif
haddps m0, m1
haddps m2, m3
haddps m0, m2
movaps [samplesq+cnt1q], m0
mulps m6, m4
mulps m7, m4
%if ARCH_X86_64
mulps m8, m4
mulps m9, m4
haddps m6, m7
haddps m8, m9
haddps m6, m8
%else
mulps m2, m4, [coeffq+cnt1q*4+32]
mulps m3, m4, [coeffq+cnt1q*4+48]
haddps m6, m7
haddps m2, m3
haddps m6, m2
%endif
movaps [samplesq+cnt2q], m6
sub cnt2d, 16
add cnt1q, 16
jl .inner_loop
add lfeq, sizeof_float
add samplesq, 128*sizeof_float
mov cnt1q, -64*sizeof_float
mov cnt2d, 64*sizeof_float-16
sub nblocksd, 1
jg .loop
RET
%endmacro
INIT_XMM sse3
LFE_FIR1_FLOAT
%if HAVE_AVX_EXTERNAL
INIT_XMM avx
LFE_FIR1_FLOAT
%endif
Reference in New Issue Copy Permalink