On 18 October 2016 at 21:04, Michael Niedermayer <mich...@niedermayer.cc> wrote:
> On Tue, Oct 18, 2016 at 05:33:13PM +0100, Rostislav Pehlivanov wrote: > > On 18 October 2016 at 16:32, James Almer <jamr...@gmail.com> wrote: > > > > > On 10/18/2016 12:07 PM, Rostislav Pehlivanov wrote: > > > > diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c > > > > index ee3cbf8..622f0ba 100644 > > > > --- a/libavcodec/aacenc.c > > > > +++ b/libavcodec/aacenc.c > > > > @@ -1033,6 +1033,12 @@ static av_cold int > aac_encode_init(AVCodecContext > > > *avctx) > > > > ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, > > > FF_LPC_TYPE_LEVINSON); > > > > s->random_state = 0x1f2e3d4c; > > > > > > > > + s->abs_pow34 = &abs_pow34_v; > > > > + s->quant_bands = &quantize_bands; > > > > > > No need for & in these. > > > > > > > + > > > > + if (ARCH_X86) > > > > + ff_aac_dsp_init_x86(s); > > > > + > > > > if (HAVE_MIPSDSP) > > > > ff_aac_coder_init_mips(s); > > > > > > [...] > > > > > > > diff --git a/libavcodec/x86/aacencdsp.asm > b/libavcodec/x86/aacencdsp.asm > > > > new file mode 100644 > > > > index 0000000..dd7b022 > > > > --- /dev/null > > > > +++ b/libavcodec/x86/aacencdsp.asm > > > > @@ -0,0 +1,88 @@ > > > > +;********************************************************** > > > ******************** > > > > +;* SIMD optimized AAC encoder DSP functions > > > > +;* > > > > +;* Copyright (C) 2016 Rostislav Pehlivanov <atomnu...@gmail.com> > > > > +;* > > > > +;* 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_RODATA > > > > + > > > > +float_abs_mask: times 4 dd 0x7fffffff > > > > + > > > > +SECTION .text > > > > + > > > > +;********************************************************** > ********* > > > > +;void ff_abs_pow34(float *out, const float *in, const int size); > > > > +;********************************************************** > ********* > > > > +INIT_XMM sse > > > > +cglobal abs_pow34, 3, 3, 3, out, in, size > > > > + mova m2, [float_abs_mask] > > > > + shl sizeq, 2 > > > > + add inq, sizeq > > > > + add outq, sizeq > > > > + neg sizeq > > > > +.loop: > > > > + movaps m0, [inq+sizeq] > > > > + andps m0, m2 > > > > > > Remove the movaps and do > > > > > > andps m0, m2, [inq+sizeq] > > > > > > Instead. Sorry i didn't notice this last time. > > > > > > > + sqrtps m1, m0 > > > > + mulps m0, m1 > > > > + sqrtps m0, m0 > > > > + mova [outq+sizeq], m0 > > > > + add sizeq, mmsize > > > > + jl .loop > > > > + RET > > > > + > > > > +;********************************************************** > ********* > > > > +;void ff_aac_quantize_bands(int *out, const float *in, const float > > > *scaled, > > > > +; int size, int is_signed, int maxval, > const > > > float Q34, > > > > +; const float rounding) > > > > +;********************************************************** > ********* > > > > +INIT_XMM sse2 > > > > +cglobal aac_quantize_bands, 5, 5, 6, out, in, scaled, size, > is_signed, > > > maxval, Q34, rounding > > > > +%if UNIX64 == 0 > > > > + movss m0, Q34m > > > > + movss m1, roundingm > > > > + cvtsi2ss m3, maxvald > > > > +%else > > > > + cvtsi2ss m3, dword maxvalm > > > > +%endif > > > > > > The other way around. Unix64 is the one that has maxval on a reg > regardless > > > of how you init the function, whereas win64 and any x86_32 target have > it > > > on stack. > > > > > > > + shufps m0, m0, 0 > > > > + shufps m1, m1, 0 > > > > + shufps m3, m3, 0 > > > > + shl is_signedd, 31 > > > > + movd m4, is_signedd > > > > + shufps m4, m4, 0 > > > > + shl sized, 2 > > > > + add inq, sizeq > > > > + add outq, sizeq > > > > + add scaledq, sizeq > > > > + neg sizeq > > > > +.loop: > > > > + mulps m2, m0, [scaledq+sizeq] > > > > + addps m2, m1 > > > > + minps m2, m3 > > > > > > > + movaps m5, [inq+sizeq] > > > > + andps m5, m4 > > > > > > Same as in abs_pow34, remove movaps and do > > > > > > andps m5, m4, [inq+sizeq] > > > > > > > + orps m2, m5 > > > > + cvttps2dq m2, m2 > > > > + mova [outq+sizeq], m2 > > > > + add sizeq, mmsize > > > > + jl .loop > > > > + RET > > > > diff --git a/libavcodec/x86/aacencdsp_init.c > b/libavcodec/x86/aacencdsp_ > > > init.c > > > > new file mode 100644 > > > > index 0000000..aefaa15 > > > > --- /dev/null > > > > +++ b/libavcodec/x86/aacencdsp_init.c > > > > @@ -0,0 +1,43 @@ > > > > +/* > > > > + * AAC encoder assembly optimizations > > > > + * Copyright (C) 2016 Rostislav Pehlivanov <atomnu...@gmail.com> > > > > + * > > > > + * 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 "config.h" > > > > + > > > > +#include "libavutil/float_dsp.h" > > > > +#include "libavutil/x86/cpu.h" > > > > +#include "libavcodec/aacenc.h" > > > > + > > > > +void ff_abs_pow34_sse(float *out, const float *in, const int size); > > > > + > > > > +void ff_aac_quantize_bands_sse2(int *out, const float *in, const > float > > > *scaled, > > > > + int size, int is_signed, int maxval, > > > const float Q34, > > > > + const float rounding); > > > > + > > > > +av_cold void ff_aac_dsp_init_x86(AACEncContext *s) > > > > +{ > > > > + int cpu_flags = av_get_cpu_flags(); > > > > + > > > > + if (EXTERNAL_SSE(cpu_flags)) > > > > + s->abs_pow34 = &ff_abs_pow34_sse; > > > > + > > > > + if (EXTERNAL_SSE2(cpu_flags)) > > > > + s->quant_bands = &ff_aac_quantize_bands_sse2; > > > > > > Same with these. > > > > > > Passes fate on mingw-w64 (x86_64) after fixing the maxval mixup above. > > > Crashes otherwise. > > > > > > _______________________________________________ > > > ffmpeg-devel mailing list > > > ffmpeg-devel@ffmpeg.org > > > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > > > > > > Hopefully I didn't mess anything up and the attached patch is the last > > revision of it. > > > aaccoder.c | 27 ++++++++------- > > aaccoder_trellis.h | 2 - > > aaccoder_twoloop.h | 2 - > > aacenc.c | 6 +++ > > aacenc.h | 6 +++ > > aacenc_is.c | 6 +-- > > aacenc_ltp.c | 4 +- > > aacenc_pred.c | 6 +-- > > aacenc_quantization.h | 4 +- > > aacenc_utils.h | 2 - > > x86/Makefile | 2 + > > x86/aacencdsp.asm | 86 ++++++++++++++++++++++++++++++ > ++++++++++++++++++++ > > x86/aacencdsp_init.c | 43 +++++++++++++++++++++++++ > > 13 files changed, 170 insertions(+), 26 deletions(-) > > e9878e2480cd612e6852783003f47768f5ced052 0001-aacenc-add-SIMD- > optimizations-for-abs_pow34-and-quan.patch > > From eb95cf5aeacdc7d191c5b7cf8df40a1e5306b733 Mon Sep 17 00:00:00 2001 > > From: Rostislav Pehlivanov <atomnu...@gmail.com> > > Date: Sat, 8 Oct 2016 15:59:14 +0100 > > Subject: [PATCH] aacenc: add SIMD optimizations for abs_pow34 and > quantization > > tested on x86 linux32/64 mingw32/64, arm qemu and mips qemu > all work now > > [...] > > -- > Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB > > It is dangerous to be right in matters on which the established authorities > are wrong. -- Voltaire > > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > Pushed Thanks to both of you for the review and testing _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel