| Message ID | 20240204144153.411868-1-ramiro.polla@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel] lavc/aarch64/fdct: add neon-optimized fdct for aarch64 | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
Hi, On Sun, 4 Feb 2024, Ramiro Polla wrote: > The code is imported from libjpeg-turbo-3.0.1. The neon registers used > have been changed to avoid modifying v8-v15. > --- I don't remember if we have any extra routines we need to do if importing foreign code with a differing license. The license here seems fine in any case though. This seems to work fine in all my test environments. And thanks for making sure it doesn't use v8-v15! I'm not so familiar with these DSP functions, whether it is norm to add a new constant like FF_DCT_NEON, but I guess it seems to match the pattern of the existing code. I presume the main case that tests this is "make fate-dct8x8", which builds and executes libavcodec/tests/dct? How much work would it be to integrate testing of these routines into checkasm? That way we could rest assured that the assembly passes all such ABI checks that we do there, including what registers must not be clobbered. The assembly uses a different indentation width than the rest of our assembly. I recently spent some effort on cleaning that up so that our code is mostly consistent, so I'd prefer not to add new code that deviates from it. It primarily looks like you'd need to add 4 spaces at the start of each line. I've used a script for mostly automatically reindenting our arm assembly, you can grab it at https://martin.st/temp/ffmpeg-asm-indent.pl, run it as "cat file.S | ./ffmpeg-asm-indent.pl > tmp; mv tmp file.S". It's not 100% accurate, but mostly gets you there, but it's good to manually check it afterwards as well. // Martin
ping On Sun, Feb 4, 2024 at 3:42 PM Ramiro Polla <ramiro.polla@gmail.com> wrote: > > The code is imported from libjpeg-turbo-3.0.1. The neon registers used > have been changed to avoid modifying v8-v15. > --- > libavcodec/aarch64/Makefile | 2 + > libavcodec/aarch64/fdct.h | 26 ++ > libavcodec/aarch64/fdctdsp_init_aarch64.c | 39 +++ > libavcodec/aarch64/fdctdsp_neon.S | 369 ++++++++++++++++++++++ > libavcodec/avcodec.h | 1 + > libavcodec/fdctdsp.c | 4 +- > libavcodec/fdctdsp.h | 2 + > libavcodec/options_table.h | 1 + > libavcodec/tests/aarch64/dct.c | 2 + > 9 files changed, 445 insertions(+), 1 deletion(-) > create mode 100644 libavcodec/aarch64/fdct.h > create mode 100644 libavcodec/aarch64/fdctdsp_init_aarch64.c > create mode 100644 libavcodec/aarch64/fdctdsp_neon.S > > diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile > index beb6a02f5f..eebccbe4a5 100644 > --- a/libavcodec/aarch64/Makefile > +++ b/libavcodec/aarch64/Makefile > @@ -1,4 +1,5 @@ > # subsystems > +OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_init_aarch64.o > OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_init.o > OBJS-$(CONFIG_H264CHROMA) += aarch64/h264chroma_init_aarch64.o > OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o > @@ -35,6 +36,7 @@ ARMV8-OBJS-$(CONFIG_VIDEODSP) += aarch64/videodsp.o > > # subsystems > NEON-OBJS-$(CONFIG_AAC_DECODER) += aarch64/sbrdsp_neon.o > +NEON-OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_neon.o > NEON-OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_neon.o > NEON-OBJS-$(CONFIG_H264CHROMA) += aarch64/h264cmc_neon.o > NEON-OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_neon.o \ > diff --git a/libavcodec/aarch64/fdct.h b/libavcodec/aarch64/fdct.h > new file mode 100644 > index 0000000000..0901b53a83 > --- /dev/null > +++ b/libavcodec/aarch64/fdct.h > @@ -0,0 +1,26 @@ > +/* > + * 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 > + */ > + > +#ifndef AVCODEC_AARCH64_FDCT_H > +#define AVCODEC_AARCH64_FDCT_H > + > +#include <stdint.h> > + > +void ff_fdct_neon(int16_t *block); > + > +#endif /* AVCODEC_AARCH64_FDCT_H */ > diff --git a/libavcodec/aarch64/fdctdsp_init_aarch64.c b/libavcodec/aarch64/fdctdsp_init_aarch64.c > new file mode 100644 > index 0000000000..59d91bc8fc > --- /dev/null > +++ b/libavcodec/aarch64/fdctdsp_init_aarch64.c > @@ -0,0 +1,39 @@ > +/* > + * 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/attributes.h" > +#include "libavutil/cpu.h" > +#include "libavutil/aarch64/cpu.h" > +#include "libavcodec/avcodec.h" > +#include "libavcodec/fdctdsp.h" > +#include "fdct.h" > + > +av_cold void ff_fdctdsp_init_aarch64(FDCTDSPContext *c, AVCodecContext *avctx, > + unsigned high_bit_depth) > +{ > + int cpu_flags = av_get_cpu_flags(); > + > + if (have_neon(cpu_flags)) { > + if (!high_bit_depth) { > + if (avctx->dct_algo == FF_DCT_AUTO || > + avctx->dct_algo == FF_DCT_NEON) { > + c->fdct = ff_fdct_neon; > + } > + } > + } > +} > diff --git a/libavcodec/aarch64/fdctdsp_neon.S b/libavcodec/aarch64/fdctdsp_neon.S > new file mode 100644 > index 0000000000..978c8d3002 > --- /dev/null > +++ b/libavcodec/aarch64/fdctdsp_neon.S > @@ -0,0 +1,369 @@ > +/* > + * Armv8 Neon optimizations for libjpeg-turbo > + * > + * Copyright (C) 2009-2011, Nokia Corporation and/or its subsidiary(-ies). > + * All Rights Reserved. > + * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com> > + * Copyright (C) 2013-2014, Linaro Limited. All Rights Reserved. > + * Author: Ragesh Radhakrishnan <ragesh.r@linaro.org> > + * Copyright (C) 2014-2016, 2020, D. R. Commander. All Rights Reserved. > + * Copyright (C) 2015-2016, 2018, Matthieu Darbois. All Rights Reserved. > + * Copyright (C) 2016, Siarhei Siamashka. All Rights Reserved. > + * > + * This software is provided 'as-is', without any express or implied > + * warranty. In no event will the authors be held liable for any damages > + * arising from the use of this software. > + * > + * Permission is granted to anyone to use this software for any purpose, > + * including commercial applications, and to alter it and redistribute it > + * freely, subject to the following restrictions: > + * > + * 1. The origin of this software must not be misrepresented; you must not > + * claim that you wrote the original software. If you use this software > + * in a product, an acknowledgment in the product documentation would be > + * appreciated but is not required. > + * 2. Altered source versions must be plainly marked as such, and must not be > + * misrepresented as being the original software. > + * 3. This notice may not be removed or altered from any source distribution. > + */ > + > +#include "libavutil/aarch64/asm.S" > +#include "neon.S" > + > +// #define EIGHT_BIT_SAMPLES > + > +/* Constants for jsimd_fdct_islow_neon() */ > + > +#define F_0_298 2446 /* FIX(0.298631336) */ > +#define F_0_390 3196 /* FIX(0.390180644) */ > +#define F_0_541 4433 /* FIX(0.541196100) */ > +#define F_0_765 6270 /* FIX(0.765366865) */ > +#define F_0_899 7373 /* FIX(0.899976223) */ > +#define F_1_175 9633 /* FIX(1.175875602) */ > +#define F_1_501 12299 /* FIX(1.501321110) */ > +#define F_1_847 15137 /* FIX(1.847759065) */ > +#define F_1_961 16069 /* FIX(1.961570560) */ > +#define F_2_053 16819 /* FIX(2.053119869) */ > +#define F_2_562 20995 /* FIX(2.562915447) */ > +#define F_3_072 25172 /* FIX(3.072711026) */ > + > +const jsimd_fdct_islow_neon_consts, align=4 > + .short F_0_298 > + .short -F_0_390 > + .short F_0_541 > + .short F_0_765 > + .short - F_0_899 > + .short F_1_175 > + .short F_1_501 > + .short - F_1_847 > + .short - F_1_961 > + .short F_2_053 > + .short - F_2_562 > + .short F_3_072 > + .short 0 /* padding */ > + .short 0 > + .short 0 > + .short 0 > +endconst > + > +#undef F_0_298 > +#undef F_0_390 > +#undef F_0_541 > +#undef F_0_765 > +#undef F_0_899 > +#undef F_1_175 > +#undef F_1_501 > +#undef F_1_847 > +#undef F_1_961 > +#undef F_2_053 > +#undef F_2_562 > +#undef F_3_072 > + > +/*****************************************************************************/ > + > +/* > + * jsimd_fdct_islow_neon > + * > + * This file contains a slower but more accurate integer implementation of the > + * forward DCT (Discrete Cosine Transform). The following code is based > + * directly on the IJG''s original jfdctint.c; see the jfdctint.c for > + * more details. > + */ > + > +#define CONST_BITS 13 > +#ifdef EIGHT_BIT_SAMPLES > +#define PASS1_BITS 2 > +#else > +#define PASS1_BITS 1 /* lose a little precision to avoid overflow */ > +#endif > + > +#define DESCALE_P1 (CONST_BITS - PASS1_BITS) > +#define DESCALE_P2 (CONST_BITS + PASS1_BITS) > + > +#define XFIX_P_0_298 v0.h[0] > +#define XFIX_N_0_390 v0.h[1] > +#define XFIX_P_0_541 v0.h[2] > +#define XFIX_P_0_765 v0.h[3] > +#define XFIX_N_0_899 v0.h[4] > +#define XFIX_P_1_175 v0.h[5] > +#define XFIX_P_1_501 v0.h[6] > +#define XFIX_N_1_847 v0.h[7] > +#define XFIX_N_1_961 v1.h[0] > +#define XFIX_P_2_053 v1.h[1] > +#define XFIX_N_2_562 v1.h[2] > +#define XFIX_P_3_072 v1.h[3] > + > +function ff_fdct_neon, export=1 > + > + DATA .req x0 > + TMP .req x9 > + > + /* Load constants */ > + movrel TMP, jsimd_fdct_islow_neon_consts > + ld1 {v0.8h, v1.8h}, [TMP] > + > + /* Load all DATA into Neon registers with the following allocation: > + * 0 1 2 3 | 4 5 6 7 > + * ---------+-------- > + * 0 | d16 | d17 | v16.8h > + * 1 | d18 | d19 | v17.8h > + * 2 | d20 | d21 | v18.8h > + * 3 | d22 | d23 | v19.8h > + * 4 | d24 | d25 | v20.8h > + * 5 | d26 | d27 | v21.8h > + * 6 | d28 | d29 | v22.8h > + * 7 | d30 | d31 | v23.8h > + */ > + > + ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64 > + ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA] > + sub DATA, DATA, #64 > + > + /* Transpose */ > + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v31, v2 > + /* 1-D FDCT */ > + add v24.8h, v16.8h, v23.8h /* tmp0 = dataptr[0] + dataptr[7]; */ > + sub v31.8h, v16.8h, v23.8h /* tmp7 = dataptr[0] - dataptr[7]; */ > + add v25.8h, v17.8h, v22.8h /* tmp1 = dataptr[1] + dataptr[6]; */ > + sub v30.8h, v17.8h, v22.8h /* tmp6 = dataptr[1] - dataptr[6]; */ > + add v26.8h, v18.8h, v21.8h /* tmp2 = dataptr[2] + dataptr[5]; */ > + sub v29.8h, v18.8h, v21.8h /* tmp5 = dataptr[2] - dataptr[5]; */ > + add v27.8h, v19.8h, v20.8h /* tmp3 = dataptr[3] + dataptr[4]; */ > + sub v28.8h, v19.8h, v20.8h /* tmp4 = dataptr[3] - dataptr[4]; */ > + > + /* even part */ > + > + add v4.8h, v24.8h, v27.8h /* tmp10 = tmp0 + tmp3; */ > + sub v5.8h, v24.8h, v27.8h /* tmp13 = tmp0 - tmp3; */ > + add v6.8h, v25.8h, v26.8h /* tmp11 = tmp1 + tmp2; */ > + sub v7.8h, v25.8h, v26.8h /* tmp12 = tmp1 - tmp2; */ > + > + add v16.8h, v4.8h, v6.8h /* tmp10 + tmp11 */ > + sub v20.8h, v4.8h, v6.8h /* tmp10 - tmp11 */ > + > + add v18.8h, v7.8h, v5.8h /* tmp12 + tmp13 */ > + > + shl v16.8h, v16.8h, #PASS1_BITS /* dataptr[0] = (DCTELEM)LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS); */ > + shl v20.8h, v20.8h, #PASS1_BITS /* dataptr[4] = (DCTELEM)LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS); */ > + > + smull2 v24.4s, v18.8h, XFIX_P_0_541 /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ > + smull v18.4s, v18.4h, XFIX_P_0_541 /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ > + mov v22.16b, v18.16b > + mov v25.16b, v24.16b > + > + smlal v18.4s, v5.4h, XFIX_P_0_765 /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ > + smlal2 v24.4s, v5.8h, XFIX_P_0_765 /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ > + smlal v22.4s, v7.4h, XFIX_N_1_847 /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ > + smlal2 v25.4s, v7.8h, XFIX_N_1_847 /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ > + > + rshrn v18.4h, v18.4s, #DESCALE_P1 > + rshrn v22.4h, v22.4s, #DESCALE_P1 > + rshrn2 v18.8h, v24.4s, #DESCALE_P1 /* dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS-PASS1_BITS); */ > + rshrn2 v22.8h, v25.4s, #DESCALE_P1 /* dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS-PASS1_BITS); */ > + > + /* Odd part */ > + > + add v2.8h, v28.8h, v31.8h /* z1 = tmp4 + tmp7; */ > + add v3.8h, v29.8h, v30.8h /* z2 = tmp5 + tmp6; */ > + add v6.8h, v28.8h, v30.8h /* z3 = tmp4 + tmp6; */ > + add v7.8h, v29.8h, v31.8h /* z4 = tmp5 + tmp7; */ > + smull v4.4s, v6.4h, XFIX_P_1_175 /* z5 lo = z3 lo * XFIX_P_1_175 */ > + smull2 v5.4s, v6.8h, XFIX_P_1_175 > + smlal v4.4s, v7.4h, XFIX_P_1_175 /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */ > + smlal2 v5.4s, v7.8h, XFIX_P_1_175 > + > + smull2 v24.4s, v28.8h, XFIX_P_0_298 > + smull2 v25.4s, v29.8h, XFIX_P_2_053 > + smull2 v26.4s, v30.8h, XFIX_P_3_072 > + smull2 v27.4s, v31.8h, XFIX_P_1_501 > + smull v23.4s, v28.4h, XFIX_P_0_298 /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */ > + smull v21.4s, v29.4h, XFIX_P_2_053 /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */ > + smull v19.4s, v30.4h, XFIX_P_3_072 /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */ > + smull v17.4s, v31.4h, XFIX_P_1_501 /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */ > + > + smull2 v28.4s, v2.8h, XFIX_N_0_899 > + smull2 v29.4s, v3.8h, XFIX_N_2_562 > + smull2 v30.4s, v6.8h, XFIX_N_1_961 > + smull2 v31.4s, v7.8h, XFIX_N_0_390 > + smull v2.4s, v2.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223); */ > + smull v3.4s, v3.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447); */ > + smull v6.4s, v6.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560); */ > + smull v7.4s, v7.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644); */ > + > + add v6.4s, v6.4s, v4.4s /* z3 += z5 */ > + add v30.4s, v30.4s, v5.4s > + add v7.4s, v7.4s, v4.4s /* z4 += z5 */ > + add v31.4s, v31.4s, v5.4s > + > + add v23.4s, v23.4s, v2.4s /* tmp4 += z1 */ > + add v24.4s, v24.4s, v28.4s > + add v21.4s, v21.4s, v3.4s /* tmp5 += z2 */ > + add v25.4s, v25.4s, v29.4s > + add v19.4s, v19.4s, v6.4s /* tmp6 += z3 */ > + add v26.4s, v26.4s, v30.4s > + add v17.4s, v17.4s, v7.4s /* tmp7 += z4 */ > + add v27.4s, v27.4s, v31.4s > + > + add v23.4s, v23.4s, v6.4s /* tmp4 += z3 */ > + add v24.4s, v24.4s, v30.4s > + add v21.4s, v21.4s, v7.4s /* tmp5 += z4 */ > + add v25.4s, v25.4s, v31.4s > + add v19.4s, v19.4s, v3.4s /* tmp6 += z2 */ > + add v26.4s, v26.4s, v29.4s > + add v17.4s, v17.4s, v2.4s /* tmp7 += z1 */ > + add v27.4s, v27.4s, v28.4s > + > + rshrn v23.4h, v23.4s, #DESCALE_P1 > + rshrn v21.4h, v21.4s, #DESCALE_P1 > + rshrn v19.4h, v19.4s, #DESCALE_P1 > + rshrn v17.4h, v17.4s, #DESCALE_P1 > + rshrn2 v23.8h, v24.4s, #DESCALE_P1 /* dataptr[7] = (DCTELEM)DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); */ > + rshrn2 v21.8h, v25.4s, #DESCALE_P1 /* dataptr[5] = (DCTELEM)DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); */ > + rshrn2 v19.8h, v26.4s, #DESCALE_P1 /* dataptr[3] = (DCTELEM)DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); */ > + rshrn2 v17.8h, v27.4s, #DESCALE_P1 /* dataptr[1] = (DCTELEM)DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); */ > + > + /* Transpose */ > + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v31, v2 > + > + /* 1-D FDCT */ > + add v24.8h, v16.8h, v23.8h /* tmp0 = dataptr[0] + dataptr[7]; */ > + sub v31.8h, v16.8h, v23.8h /* tmp7 = dataptr[0] - dataptr[7]; */ > + add v25.8h, v17.8h, v22.8h /* tmp1 = dataptr[1] + dataptr[6]; */ > + sub v30.8h, v17.8h, v22.8h /* tmp6 = dataptr[1] - dataptr[6]; */ > + add v26.8h, v18.8h, v21.8h /* tmp2 = dataptr[2] + dataptr[5]; */ > + sub v29.8h, v18.8h, v21.8h /* tmp5 = dataptr[2] - dataptr[5]; */ > + add v27.8h, v19.8h, v20.8h /* tmp3 = dataptr[3] + dataptr[4]; */ > + sub v28.8h, v19.8h, v20.8h /* tmp4 = dataptr[3] - dataptr[4]; */ > + > + /* even part */ > + add v4.8h, v24.8h, v27.8h /* tmp10 = tmp0 + tmp3; */ > + sub v5.8h, v24.8h, v27.8h /* tmp13 = tmp0 - tmp3; */ > + add v6.8h, v25.8h, v26.8h /* tmp11 = tmp1 + tmp2; */ > + sub v7.8h, v25.8h, v26.8h /* tmp12 = tmp1 - tmp2; */ > + > + add v16.8h, v4.8h, v6.8h /* tmp10 + tmp11 */ > + sub v20.8h, v4.8h, v6.8h /* tmp10 - tmp11 */ > + > + add v18.8h, v7.8h, v5.8h /* tmp12 + tmp13 */ > + > + srshr v16.8h, v16.8h, #PASS1_BITS /* dataptr[0] = (DCTELEM)DESCALE(tmp10 + tmp11, PASS1_BITS); */ > + srshr v20.8h, v20.8h, #PASS1_BITS /* dataptr[4] = (DCTELEM)DESCALE(tmp10 - tmp11, PASS1_BITS); */ > + > + smull2 v24.4s, v18.8h, XFIX_P_0_541 /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ > + smull v18.4s, v18.4h, XFIX_P_0_541 /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ > + mov v22.16b, v18.16b > + mov v25.16b, v24.16b > + > + smlal v18.4s, v5.4h, XFIX_P_0_765 /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ > + smlal2 v24.4s, v5.8h, XFIX_P_0_765 /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ > + smlal v22.4s, v7.4h, XFIX_N_1_847 /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ > + smlal2 v25.4s, v7.8h, XFIX_N_1_847 /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ > + > + rshrn v18.4h, v18.4s, #DESCALE_P2 > + rshrn v22.4h, v22.4s, #DESCALE_P2 > + rshrn2 v18.8h, v24.4s, #DESCALE_P2 /* dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS+PASS1_BITS); */ > + rshrn2 v22.8h, v25.4s, #DESCALE_P2 /* dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS+PASS1_BITS); */ > + > + /* Odd part */ > + add v2.8h, v28.8h, v31.8h /* z1 = tmp4 + tmp7; */ > + add v3.8h, v29.8h, v30.8h /* z2 = tmp5 + tmp6; */ > + add v6.8h, v28.8h, v30.8h /* z3 = tmp4 + tmp6; */ > + add v7.8h, v29.8h, v31.8h /* z4 = tmp5 + tmp7; */ > + > + smull v4.4s, v6.4h, XFIX_P_1_175 /* z5 lo = z3 lo * XFIX_P_1_175 */ > + smull2 v5.4s, v6.8h, XFIX_P_1_175 > + smlal v4.4s, v7.4h, XFIX_P_1_175 /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */ > + smlal2 v5.4s, v7.8h, XFIX_P_1_175 > + > + smull2 v24.4s, v28.8h, XFIX_P_0_298 > + smull2 v25.4s, v29.8h, XFIX_P_2_053 > + smull2 v26.4s, v30.8h, XFIX_P_3_072 > + smull2 v27.4s, v31.8h, XFIX_P_1_501 > + smull v23.4s, v28.4h, XFIX_P_0_298 /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */ > + smull v21.4s, v29.4h, XFIX_P_2_053 /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */ > + smull v19.4s, v30.4h, XFIX_P_3_072 /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */ > + smull v17.4s, v31.4h, XFIX_P_1_501 /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */ > + > + smull2 v28.4s, v2.8h, XFIX_N_0_899 > + smull2 v29.4s, v3.8h, XFIX_N_2_562 > + smull2 v30.4s, v6.8h, XFIX_N_1_961 > + smull2 v31.4s, v7.8h, XFIX_N_0_390 > + smull v2.4s, v2.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223); */ > + smull v3.4s, v3.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447); */ > + smull v6.4s, v6.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560); */ > + smull v7.4s, v7.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644); */ > + > + add v6.4s, v6.4s, v4.4s /* z3 += z5 */ > + add v30.4s, v30.4s, v5.4s > + add v7.4s, v7.4s, v4.4s /* z4 += z5 */ > + add v31.4s, v31.4s, v5.4s > + > + add v23.4s, v23.4s, v2.4s /* tmp4 += z1 */ > + add v24.4s, v24.4s, v28.4s > + add v21.4s, v21.4s, v3.4s /* tmp5 += z2 */ > + add v25.4s, v25.4s, v29.4s > + add v19.4s, v19.4s, v6.4s /* tmp6 += z3 */ > + add v26.4s, v26.4s, v30.4s > + add v17.4s, v17.4s, v7.4s /* tmp7 += z4 */ > + add v27.4s, v27.4s, v31.4s > + > + add v23.4s, v23.4s, v6.4s /* tmp4 += z3 */ > + add v24.4s, v24.4s, v30.4s > + add v21.4s, v21.4s, v7.4s /* tmp5 += z4 */ > + add v25.4s, v25.4s, v31.4s > + add v19.4s, v19.4s, v3.4s /* tmp6 += z2 */ > + add v26.4s, v26.4s, v29.4s > + add v17.4s, v17.4s, v2.4s /* tmp7 += z1 */ > + add v27.4s, v27.4s, v28.4s > + > + rshrn v23.4h, v23.4s, #DESCALE_P2 > + rshrn v21.4h, v21.4s, #DESCALE_P2 > + rshrn v19.4h, v19.4s, #DESCALE_P2 > + rshrn v17.4h, v17.4s, #DESCALE_P2 > + rshrn2 v23.8h, v24.4s, #DESCALE_P2 /* dataptr[7] = (DCTELEM)DESCALE(tmp4 + z1 + z3, CONST_BITS+PASS1_BITS); */ > + rshrn2 v21.8h, v25.4s, #DESCALE_P2 /* dataptr[5] = (DCTELEM)DESCALE(tmp5 + z2 + z4, CONST_BITS+PASS1_BITS); */ > + rshrn2 v19.8h, v26.4s, #DESCALE_P2 /* dataptr[3] = (DCTELEM)DESCALE(tmp6 + z2 + z3, CONST_BITS+PASS1_BITS); */ > + rshrn2 v17.8h, v27.4s, #DESCALE_P2 /* dataptr[1] = (DCTELEM)DESCALE(tmp7 + z1 + z4, CONST_BITS+PASS1_BITS); */ > + > + /* store results */ > + st1 {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64 > + st1 {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA] > + > + ret > + > + .unreq DATA > + .unreq TMP > +endfunc > + > +#undef XFIX_P_0_298 > +#undef XFIX_N_0_390 > +#undef XFIX_P_0_541 > +#undef XFIX_P_0_765 > +#undef XFIX_N_0_899 > +#undef XFIX_P_1_175 > +#undef XFIX_P_1_501 > +#undef XFIX_N_1_847 > +#undef XFIX_N_1_961 > +#undef XFIX_P_2_053 > +#undef XFIX_N_2_562 > +#undef XFIX_P_3_072 > diff --git a/libavcodec/avcodec.h b/libavcodec/avcodec.h > index 7fb44e28f4..f9b86f1d58 100644 > --- a/libavcodec/avcodec.h > +++ b/libavcodec/avcodec.h > @@ -1477,6 +1477,7 @@ typedef struct AVCodecContext { > #define FF_DCT_MMX 3 > #define FF_DCT_ALTIVEC 5 > #define FF_DCT_FAAN 6 > +#define FF_DCT_NEON 7 > > /** > * IDCT algorithm, see FF_IDCT_* below. > diff --git a/libavcodec/fdctdsp.c b/libavcodec/fdctdsp.c > index f8ba17426c..d20558ce88 100644 > --- a/libavcodec/fdctdsp.c > +++ b/libavcodec/fdctdsp.c > @@ -42,7 +42,9 @@ av_cold void ff_fdctdsp_init(FDCTDSPContext *c, AVCodecContext *avctx) > c->fdct248 = ff_fdct248_islow_8; > } > > -#if ARCH_PPC > +#if ARCH_AARCH64 > + ff_fdctdsp_init_aarch64(c, avctx, high_bit_depth); > +#elif ARCH_PPC > ff_fdctdsp_init_ppc(c, avctx, high_bit_depth); > #elif ARCH_X86 > ff_fdctdsp_init_x86(c, avctx, high_bit_depth); > diff --git a/libavcodec/fdctdsp.h b/libavcodec/fdctdsp.h > index 7378eab870..cad99ed7ca 100644 > --- a/libavcodec/fdctdsp.h > +++ b/libavcodec/fdctdsp.h > @@ -32,6 +32,8 @@ typedef struct FDCTDSPContext { > > FF_VISIBILITY_PUSH_HIDDEN > void ff_fdctdsp_init(FDCTDSPContext *c, struct AVCodecContext *avctx); > +void ff_fdctdsp_init_aarch64(FDCTDSPContext *c, struct AVCodecContext *avctx, > + unsigned high_bit_depth); > void ff_fdctdsp_init_ppc(FDCTDSPContext *c, struct AVCodecContext *avctx, > unsigned high_bit_depth); > void ff_fdctdsp_init_x86(FDCTDSPContext *c, struct AVCodecContext *avctx, > diff --git a/libavcodec/options_table.h b/libavcodec/options_table.h > index ee243d9894..d9a3c92f28 100644 > --- a/libavcodec/options_table.h > +++ b/libavcodec/options_table.h > @@ -159,6 +159,7 @@ static const AVOption avcodec_options[] = { > {"mmx", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_MMX }, INT_MIN, INT_MAX, V|E, "dct"}, > {"altivec", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_ALTIVEC }, INT_MIN, INT_MAX, V|E, "dct"}, > {"faan", "floating point AAN DCT", 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_FAAN }, INT_MIN, INT_MAX, V|E, "dct"}, > +{"neon", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_NEON }, INT_MIN, INT_MAX, V|E, "dct"}, > {"lumi_mask", "compresses bright areas stronger than medium ones", OFFSET(lumi_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, > {"tcplx_mask", "temporal complexity masking", OFFSET(temporal_cplx_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, > {"scplx_mask", "spatial complexity masking", OFFSET(spatial_cplx_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, > diff --git a/libavcodec/tests/aarch64/dct.c b/libavcodec/tests/aarch64/dct.c > index 9e477328d5..e98a887cd5 100644 > --- a/libavcodec/tests/aarch64/dct.c > +++ b/libavcodec/tests/aarch64/dct.c > @@ -19,9 +19,11 @@ > #include "config.h" > > #include "libavutil/cpu.h" > +#include "libavcodec/aarch64/fdct.h" > #include "libavcodec/aarch64/idct.h" > > static const struct algo fdct_tab_arch[] = { > + { "neon", ff_fdct_neon, FF_IDCT_PERM_NONE, AV_CPU_FLAG_NEON }, > { 0 } > }; > > -- > 2.30.2 >
On Wed, 6 Mar 2024, Ramiro Polla wrote: > ping > Did you miss my response here? https://ffmpeg.org/pipermail/ffmpeg-devel/2024-February/321448.html // Martin
Hi, On Wed, Feb 14, 2024 at 10:42 AM Martin Storsjö <martin@martin.st> wrote: > On Sun, 4 Feb 2024, Ramiro Polla wrote: > > > The code is imported from libjpeg-turbo-3.0.1. The neon registers used > > have been changed to avoid modifying v8-v15. > > --- > > I don't remember if we have any extra routines we need to do if importing > foreign code with a differing license. The license here seems fine in any > case though. I think the license should be ok (based on the "Patches/Committing" section in developer.texi). > This seems to work fine in all my test environments. And thanks for making > sure it doesn't use v8-v15! > > I'm not so familiar with these DSP functions, whether it is norm to add a > new constant like FF_DCT_NEON, but I guess it seems to match the pattern > of the existing code. I don't know either, so I just tried to match the existing code :) > I presume the main case that tests this is "make fate-dct8x8", which > builds and executes libavcodec/tests/dct? How much work would it be to > integrate testing of these routines into checkasm? That way we could rest > assured that the assembly passes all such ABI checks that we do there, > including what registers must not be clobbered. I added checkasm for fdct. It's especially useful to make sure there is no overflow in the DC coefficient. > The assembly uses a different indentation width than the rest of our > assembly. I recently spent some effort on cleaning that up so that our > code is mostly consistent, so I'd prefer not to add new code that deviates > from it. It primarily looks like you'd need to add 4 spaces at the start > of each line. > > I've used a script for mostly automatically reindenting our arm assembly, > you can grab it at https://martin.st/temp/ffmpeg-asm-indent.pl, run it as > "cat file.S | ./ffmpeg-asm-indent.pl > tmp; mv tmp file.S". It's not 100% > accurate, but mostly gets you there, but it's good to manually check it > afterwards as well. I fixed the indentation and tweaked a few more cosmetics in the comments. Thank you for the review and the help on IRC! I'll send v2 shortly. Ramiro
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile index beb6a02f5f..eebccbe4a5 100644 --- a/libavcodec/aarch64/Makefile +++ b/libavcodec/aarch64/Makefile @@ -1,4 +1,5 @@ # subsystems +OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_init_aarch64.o OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_init.o OBJS-$(CONFIG_H264CHROMA) += aarch64/h264chroma_init_aarch64.o OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o @@ -35,6 +36,7 @@ ARMV8-OBJS-$(CONFIG_VIDEODSP) += aarch64/videodsp.o # subsystems NEON-OBJS-$(CONFIG_AAC_DECODER) += aarch64/sbrdsp_neon.o +NEON-OBJS-$(CONFIG_FDCTDSP) += aarch64/fdctdsp_neon.o NEON-OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_neon.o NEON-OBJS-$(CONFIG_H264CHROMA) += aarch64/h264cmc_neon.o NEON-OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_neon.o \ diff --git a/libavcodec/aarch64/fdct.h b/libavcodec/aarch64/fdct.h new file mode 100644 index 0000000000..0901b53a83 --- /dev/null +++ b/libavcodec/aarch64/fdct.h @@ -0,0 +1,26 @@ +/* + * 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 + */ + +#ifndef AVCODEC_AARCH64_FDCT_H +#define AVCODEC_AARCH64_FDCT_H + +#include <stdint.h> + +void ff_fdct_neon(int16_t *block); + +#endif /* AVCODEC_AARCH64_FDCT_H */ diff --git a/libavcodec/aarch64/fdctdsp_init_aarch64.c b/libavcodec/aarch64/fdctdsp_init_aarch64.c new file mode 100644 index 0000000000..59d91bc8fc --- /dev/null +++ b/libavcodec/aarch64/fdctdsp_init_aarch64.c @@ -0,0 +1,39 @@ +/* + * 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/attributes.h" +#include "libavutil/cpu.h" +#include "libavutil/aarch64/cpu.h" +#include "libavcodec/avcodec.h" +#include "libavcodec/fdctdsp.h" +#include "fdct.h" + +av_cold void ff_fdctdsp_init_aarch64(FDCTDSPContext *c, AVCodecContext *avctx, + unsigned high_bit_depth) +{ + int cpu_flags = av_get_cpu_flags(); + + if (have_neon(cpu_flags)) { + if (!high_bit_depth) { + if (avctx->dct_algo == FF_DCT_AUTO || + avctx->dct_algo == FF_DCT_NEON) { + c->fdct = ff_fdct_neon; + } + } + } +} diff --git a/libavcodec/aarch64/fdctdsp_neon.S b/libavcodec/aarch64/fdctdsp_neon.S new file mode 100644 index 0000000000..978c8d3002 --- /dev/null +++ b/libavcodec/aarch64/fdctdsp_neon.S @@ -0,0 +1,369 @@ +/* + * Armv8 Neon optimizations for libjpeg-turbo + * + * Copyright (C) 2009-2011, Nokia Corporation and/or its subsidiary(-ies). + * All Rights Reserved. + * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com> + * Copyright (C) 2013-2014, Linaro Limited. All Rights Reserved. + * Author: Ragesh Radhakrishnan <ragesh.r@linaro.org> + * Copyright (C) 2014-2016, 2020, D. R. Commander. All Rights Reserved. + * Copyright (C) 2015-2016, 2018, Matthieu Darbois. All Rights Reserved. + * Copyright (C) 2016, Siarhei Siamashka. All Rights Reserved. + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgment in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +#include "libavutil/aarch64/asm.S" +#include "neon.S" + +// #define EIGHT_BIT_SAMPLES + +/* Constants for jsimd_fdct_islow_neon() */ + +#define F_0_298 2446 /* FIX(0.298631336) */ +#define F_0_390 3196 /* FIX(0.390180644) */ +#define F_0_541 4433 /* FIX(0.541196100) */ +#define F_0_765 6270 /* FIX(0.765366865) */ +#define F_0_899 7373 /* FIX(0.899976223) */ +#define F_1_175 9633 /* FIX(1.175875602) */ +#define F_1_501 12299 /* FIX(1.501321110) */ +#define F_1_847 15137 /* FIX(1.847759065) */ +#define F_1_961 16069 /* FIX(1.961570560) */ +#define F_2_053 16819 /* FIX(2.053119869) */ +#define F_2_562 20995 /* FIX(2.562915447) */ +#define F_3_072 25172 /* FIX(3.072711026) */ + +const jsimd_fdct_islow_neon_consts, align=4 + .short F_0_298 + .short -F_0_390 + .short F_0_541 + .short F_0_765 + .short - F_0_899 + .short F_1_175 + .short F_1_501 + .short - F_1_847 + .short - F_1_961 + .short F_2_053 + .short - F_2_562 + .short F_3_072 + .short 0 /* padding */ + .short 0 + .short 0 + .short 0 +endconst + +#undef F_0_298 +#undef F_0_390 +#undef F_0_541 +#undef F_0_765 +#undef F_0_899 +#undef F_1_175 +#undef F_1_501 +#undef F_1_847 +#undef F_1_961 +#undef F_2_053 +#undef F_2_562 +#undef F_3_072 + +/*****************************************************************************/ + +/* + * jsimd_fdct_islow_neon + * + * This file contains a slower but more accurate integer implementation of the + * forward DCT (Discrete Cosine Transform). The following code is based + * directly on the IJG''s original jfdctint.c; see the jfdctint.c for + * more details. + */ + +#define CONST_BITS 13 +#ifdef EIGHT_BIT_SAMPLES +#define PASS1_BITS 2 +#else +#define PASS1_BITS 1 /* lose a little precision to avoid overflow */ +#endif + +#define DESCALE_P1 (CONST_BITS - PASS1_BITS) +#define DESCALE_P2 (CONST_BITS + PASS1_BITS) + +#define XFIX_P_0_298 v0.h[0] +#define XFIX_N_0_390 v0.h[1] +#define XFIX_P_0_541 v0.h[2] +#define XFIX_P_0_765 v0.h[3] +#define XFIX_N_0_899 v0.h[4] +#define XFIX_P_1_175 v0.h[5] +#define XFIX_P_1_501 v0.h[6] +#define XFIX_N_1_847 v0.h[7] +#define XFIX_N_1_961 v1.h[0] +#define XFIX_P_2_053 v1.h[1] +#define XFIX_N_2_562 v1.h[2] +#define XFIX_P_3_072 v1.h[3] + +function ff_fdct_neon, export=1 + + DATA .req x0 + TMP .req x9 + + /* Load constants */ + movrel TMP, jsimd_fdct_islow_neon_consts + ld1 {v0.8h, v1.8h}, [TMP] + + /* Load all DATA into Neon registers with the following allocation: + * 0 1 2 3 | 4 5 6 7 + * ---------+-------- + * 0 | d16 | d17 | v16.8h + * 1 | d18 | d19 | v17.8h + * 2 | d20 | d21 | v18.8h + * 3 | d22 | d23 | v19.8h + * 4 | d24 | d25 | v20.8h + * 5 | d26 | d27 | v21.8h + * 6 | d28 | d29 | v22.8h + * 7 | d30 | d31 | v23.8h + */ + + ld1 {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64 + ld1 {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA] + sub DATA, DATA, #64 + + /* Transpose */ + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v31, v2 + /* 1-D FDCT */ + add v24.8h, v16.8h, v23.8h /* tmp0 = dataptr[0] + dataptr[7]; */ + sub v31.8h, v16.8h, v23.8h /* tmp7 = dataptr[0] - dataptr[7]; */ + add v25.8h, v17.8h, v22.8h /* tmp1 = dataptr[1] + dataptr[6]; */ + sub v30.8h, v17.8h, v22.8h /* tmp6 = dataptr[1] - dataptr[6]; */ + add v26.8h, v18.8h, v21.8h /* tmp2 = dataptr[2] + dataptr[5]; */ + sub v29.8h, v18.8h, v21.8h /* tmp5 = dataptr[2] - dataptr[5]; */ + add v27.8h, v19.8h, v20.8h /* tmp3 = dataptr[3] + dataptr[4]; */ + sub v28.8h, v19.8h, v20.8h /* tmp4 = dataptr[3] - dataptr[4]; */ + + /* even part */ + + add v4.8h, v24.8h, v27.8h /* tmp10 = tmp0 + tmp3; */ + sub v5.8h, v24.8h, v27.8h /* tmp13 = tmp0 - tmp3; */ + add v6.8h, v25.8h, v26.8h /* tmp11 = tmp1 + tmp2; */ + sub v7.8h, v25.8h, v26.8h /* tmp12 = tmp1 - tmp2; */ + + add v16.8h, v4.8h, v6.8h /* tmp10 + tmp11 */ + sub v20.8h, v4.8h, v6.8h /* tmp10 - tmp11 */ + + add v18.8h, v7.8h, v5.8h /* tmp12 + tmp13 */ + + shl v16.8h, v16.8h, #PASS1_BITS /* dataptr[0] = (DCTELEM)LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS); */ + shl v20.8h, v20.8h, #PASS1_BITS /* dataptr[4] = (DCTELEM)LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS); */ + + smull2 v24.4s, v18.8h, XFIX_P_0_541 /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ + smull v18.4s, v18.4h, XFIX_P_0_541 /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ + mov v22.16b, v18.16b + mov v25.16b, v24.16b + + smlal v18.4s, v5.4h, XFIX_P_0_765 /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ + smlal2 v24.4s, v5.8h, XFIX_P_0_765 /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ + smlal v22.4s, v7.4h, XFIX_N_1_847 /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ + smlal2 v25.4s, v7.8h, XFIX_N_1_847 /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ + + rshrn v18.4h, v18.4s, #DESCALE_P1 + rshrn v22.4h, v22.4s, #DESCALE_P1 + rshrn2 v18.8h, v24.4s, #DESCALE_P1 /* dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS-PASS1_BITS); */ + rshrn2 v22.8h, v25.4s, #DESCALE_P1 /* dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS-PASS1_BITS); */ + + /* Odd part */ + + add v2.8h, v28.8h, v31.8h /* z1 = tmp4 + tmp7; */ + add v3.8h, v29.8h, v30.8h /* z2 = tmp5 + tmp6; */ + add v6.8h, v28.8h, v30.8h /* z3 = tmp4 + tmp6; */ + add v7.8h, v29.8h, v31.8h /* z4 = tmp5 + tmp7; */ + smull v4.4s, v6.4h, XFIX_P_1_175 /* z5 lo = z3 lo * XFIX_P_1_175 */ + smull2 v5.4s, v6.8h, XFIX_P_1_175 + smlal v4.4s, v7.4h, XFIX_P_1_175 /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */ + smlal2 v5.4s, v7.8h, XFIX_P_1_175 + + smull2 v24.4s, v28.8h, XFIX_P_0_298 + smull2 v25.4s, v29.8h, XFIX_P_2_053 + smull2 v26.4s, v30.8h, XFIX_P_3_072 + smull2 v27.4s, v31.8h, XFIX_P_1_501 + smull v23.4s, v28.4h, XFIX_P_0_298 /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */ + smull v21.4s, v29.4h, XFIX_P_2_053 /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */ + smull v19.4s, v30.4h, XFIX_P_3_072 /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */ + smull v17.4s, v31.4h, XFIX_P_1_501 /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */ + + smull2 v28.4s, v2.8h, XFIX_N_0_899 + smull2 v29.4s, v3.8h, XFIX_N_2_562 + smull2 v30.4s, v6.8h, XFIX_N_1_961 + smull2 v31.4s, v7.8h, XFIX_N_0_390 + smull v2.4s, v2.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223); */ + smull v3.4s, v3.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447); */ + smull v6.4s, v6.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560); */ + smull v7.4s, v7.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644); */ + + add v6.4s, v6.4s, v4.4s /* z3 += z5 */ + add v30.4s, v30.4s, v5.4s + add v7.4s, v7.4s, v4.4s /* z4 += z5 */ + add v31.4s, v31.4s, v5.4s + + add v23.4s, v23.4s, v2.4s /* tmp4 += z1 */ + add v24.4s, v24.4s, v28.4s + add v21.4s, v21.4s, v3.4s /* tmp5 += z2 */ + add v25.4s, v25.4s, v29.4s + add v19.4s, v19.4s, v6.4s /* tmp6 += z3 */ + add v26.4s, v26.4s, v30.4s + add v17.4s, v17.4s, v7.4s /* tmp7 += z4 */ + add v27.4s, v27.4s, v31.4s + + add v23.4s, v23.4s, v6.4s /* tmp4 += z3 */ + add v24.4s, v24.4s, v30.4s + add v21.4s, v21.4s, v7.4s /* tmp5 += z4 */ + add v25.4s, v25.4s, v31.4s + add v19.4s, v19.4s, v3.4s /* tmp6 += z2 */ + add v26.4s, v26.4s, v29.4s + add v17.4s, v17.4s, v2.4s /* tmp7 += z1 */ + add v27.4s, v27.4s, v28.4s + + rshrn v23.4h, v23.4s, #DESCALE_P1 + rshrn v21.4h, v21.4s, #DESCALE_P1 + rshrn v19.4h, v19.4s, #DESCALE_P1 + rshrn v17.4h, v17.4s, #DESCALE_P1 + rshrn2 v23.8h, v24.4s, #DESCALE_P1 /* dataptr[7] = (DCTELEM)DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); */ + rshrn2 v21.8h, v25.4s, #DESCALE_P1 /* dataptr[5] = (DCTELEM)DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); */ + rshrn2 v19.8h, v26.4s, #DESCALE_P1 /* dataptr[3] = (DCTELEM)DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); */ + rshrn2 v17.8h, v27.4s, #DESCALE_P1 /* dataptr[1] = (DCTELEM)DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); */ + + /* Transpose */ + transpose_8x8H v16, v17, v18, v19, v20, v21, v22, v23, v31, v2 + + /* 1-D FDCT */ + add v24.8h, v16.8h, v23.8h /* tmp0 = dataptr[0] + dataptr[7]; */ + sub v31.8h, v16.8h, v23.8h /* tmp7 = dataptr[0] - dataptr[7]; */ + add v25.8h, v17.8h, v22.8h /* tmp1 = dataptr[1] + dataptr[6]; */ + sub v30.8h, v17.8h, v22.8h /* tmp6 = dataptr[1] - dataptr[6]; */ + add v26.8h, v18.8h, v21.8h /* tmp2 = dataptr[2] + dataptr[5]; */ + sub v29.8h, v18.8h, v21.8h /* tmp5 = dataptr[2] - dataptr[5]; */ + add v27.8h, v19.8h, v20.8h /* tmp3 = dataptr[3] + dataptr[4]; */ + sub v28.8h, v19.8h, v20.8h /* tmp4 = dataptr[3] - dataptr[4]; */ + + /* even part */ + add v4.8h, v24.8h, v27.8h /* tmp10 = tmp0 + tmp3; */ + sub v5.8h, v24.8h, v27.8h /* tmp13 = tmp0 - tmp3; */ + add v6.8h, v25.8h, v26.8h /* tmp11 = tmp1 + tmp2; */ + sub v7.8h, v25.8h, v26.8h /* tmp12 = tmp1 - tmp2; */ + + add v16.8h, v4.8h, v6.8h /* tmp10 + tmp11 */ + sub v20.8h, v4.8h, v6.8h /* tmp10 - tmp11 */ + + add v18.8h, v7.8h, v5.8h /* tmp12 + tmp13 */ + + srshr v16.8h, v16.8h, #PASS1_BITS /* dataptr[0] = (DCTELEM)DESCALE(tmp10 + tmp11, PASS1_BITS); */ + srshr v20.8h, v20.8h, #PASS1_BITS /* dataptr[4] = (DCTELEM)DESCALE(tmp10 - tmp11, PASS1_BITS); */ + + smull2 v24.4s, v18.8h, XFIX_P_0_541 /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ + smull v18.4s, v18.4h, XFIX_P_0_541 /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */ + mov v22.16b, v18.16b + mov v25.16b, v24.16b + + smlal v18.4s, v5.4h, XFIX_P_0_765 /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ + smlal2 v24.4s, v5.8h, XFIX_P_0_765 /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */ + smlal v22.4s, v7.4h, XFIX_N_1_847 /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ + smlal2 v25.4s, v7.8h, XFIX_N_1_847 /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */ + + rshrn v18.4h, v18.4s, #DESCALE_P2 + rshrn v22.4h, v22.4s, #DESCALE_P2 + rshrn2 v18.8h, v24.4s, #DESCALE_P2 /* dataptr[2] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS+PASS1_BITS); */ + rshrn2 v22.8h, v25.4s, #DESCALE_P2 /* dataptr[6] = (DCTELEM)DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS+PASS1_BITS); */ + + /* Odd part */ + add v2.8h, v28.8h, v31.8h /* z1 = tmp4 + tmp7; */ + add v3.8h, v29.8h, v30.8h /* z2 = tmp5 + tmp6; */ + add v6.8h, v28.8h, v30.8h /* z3 = tmp4 + tmp6; */ + add v7.8h, v29.8h, v31.8h /* z4 = tmp5 + tmp7; */ + + smull v4.4s, v6.4h, XFIX_P_1_175 /* z5 lo = z3 lo * XFIX_P_1_175 */ + smull2 v5.4s, v6.8h, XFIX_P_1_175 + smlal v4.4s, v7.4h, XFIX_P_1_175 /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */ + smlal2 v5.4s, v7.8h, XFIX_P_1_175 + + smull2 v24.4s, v28.8h, XFIX_P_0_298 + smull2 v25.4s, v29.8h, XFIX_P_2_053 + smull2 v26.4s, v30.8h, XFIX_P_3_072 + smull2 v27.4s, v31.8h, XFIX_P_1_501 + smull v23.4s, v28.4h, XFIX_P_0_298 /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */ + smull v21.4s, v29.4h, XFIX_P_2_053 /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */ + smull v19.4s, v30.4h, XFIX_P_3_072 /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */ + smull v17.4s, v31.4h, XFIX_P_1_501 /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */ + + smull2 v28.4s, v2.8h, XFIX_N_0_899 + smull2 v29.4s, v3.8h, XFIX_N_2_562 + smull2 v30.4s, v6.8h, XFIX_N_1_961 + smull2 v31.4s, v7.8h, XFIX_N_0_390 + smull v2.4s, v2.4h, XFIX_N_0_899 /* z1 = MULTIPLY(z1, -FIX_0_899976223); */ + smull v3.4s, v3.4h, XFIX_N_2_562 /* z2 = MULTIPLY(z2, -FIX_2_562915447); */ + smull v6.4s, v6.4h, XFIX_N_1_961 /* z3 = MULTIPLY(z3, -FIX_1_961570560); */ + smull v7.4s, v7.4h, XFIX_N_0_390 /* z4 = MULTIPLY(z4, -FIX_0_390180644); */ + + add v6.4s, v6.4s, v4.4s /* z3 += z5 */ + add v30.4s, v30.4s, v5.4s + add v7.4s, v7.4s, v4.4s /* z4 += z5 */ + add v31.4s, v31.4s, v5.4s + + add v23.4s, v23.4s, v2.4s /* tmp4 += z1 */ + add v24.4s, v24.4s, v28.4s + add v21.4s, v21.4s, v3.4s /* tmp5 += z2 */ + add v25.4s, v25.4s, v29.4s + add v19.4s, v19.4s, v6.4s /* tmp6 += z3 */ + add v26.4s, v26.4s, v30.4s + add v17.4s, v17.4s, v7.4s /* tmp7 += z4 */ + add v27.4s, v27.4s, v31.4s + + add v23.4s, v23.4s, v6.4s /* tmp4 += z3 */ + add v24.4s, v24.4s, v30.4s + add v21.4s, v21.4s, v7.4s /* tmp5 += z4 */ + add v25.4s, v25.4s, v31.4s + add v19.4s, v19.4s, v3.4s /* tmp6 += z2 */ + add v26.4s, v26.4s, v29.4s + add v17.4s, v17.4s, v2.4s /* tmp7 += z1 */ + add v27.4s, v27.4s, v28.4s + + rshrn v23.4h, v23.4s, #DESCALE_P2 + rshrn v21.4h, v21.4s, #DESCALE_P2 + rshrn v19.4h, v19.4s, #DESCALE_P2 + rshrn v17.4h, v17.4s, #DESCALE_P2 + rshrn2 v23.8h, v24.4s, #DESCALE_P2 /* dataptr[7] = (DCTELEM)DESCALE(tmp4 + z1 + z3, CONST_BITS+PASS1_BITS); */ + rshrn2 v21.8h, v25.4s, #DESCALE_P2 /* dataptr[5] = (DCTELEM)DESCALE(tmp5 + z2 + z4, CONST_BITS+PASS1_BITS); */ + rshrn2 v19.8h, v26.4s, #DESCALE_P2 /* dataptr[3] = (DCTELEM)DESCALE(tmp6 + z2 + z3, CONST_BITS+PASS1_BITS); */ + rshrn2 v17.8h, v27.4s, #DESCALE_P2 /* dataptr[1] = (DCTELEM)DESCALE(tmp7 + z1 + z4, CONST_BITS+PASS1_BITS); */ + + /* store results */ + st1 {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64 + st1 {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA] + + ret + + .unreq DATA + .unreq TMP +endfunc + +#undef XFIX_P_0_298 +#undef XFIX_N_0_390 +#undef XFIX_P_0_541 +#undef XFIX_P_0_765 +#undef XFIX_N_0_899 +#undef XFIX_P_1_175 +#undef XFIX_P_1_501 +#undef XFIX_N_1_847 +#undef XFIX_N_1_961 +#undef XFIX_P_2_053 +#undef XFIX_N_2_562 +#undef XFIX_P_3_072 diff --git a/libavcodec/avcodec.h b/libavcodec/avcodec.h index 7fb44e28f4..f9b86f1d58 100644 --- a/libavcodec/avcodec.h +++ b/libavcodec/avcodec.h @@ -1477,6 +1477,7 @@ typedef struct AVCodecContext { #define FF_DCT_MMX 3 #define FF_DCT_ALTIVEC 5 #define FF_DCT_FAAN 6 +#define FF_DCT_NEON 7 /** * IDCT algorithm, see FF_IDCT_* below. diff --git a/libavcodec/fdctdsp.c b/libavcodec/fdctdsp.c index f8ba17426c..d20558ce88 100644 --- a/libavcodec/fdctdsp.c +++ b/libavcodec/fdctdsp.c @@ -42,7 +42,9 @@ av_cold void ff_fdctdsp_init(FDCTDSPContext *c, AVCodecContext *avctx) c->fdct248 = ff_fdct248_islow_8; } -#if ARCH_PPC +#if ARCH_AARCH64 + ff_fdctdsp_init_aarch64(c, avctx, high_bit_depth); +#elif ARCH_PPC ff_fdctdsp_init_ppc(c, avctx, high_bit_depth); #elif ARCH_X86 ff_fdctdsp_init_x86(c, avctx, high_bit_depth); diff --git a/libavcodec/fdctdsp.h b/libavcodec/fdctdsp.h index 7378eab870..cad99ed7ca 100644 --- a/libavcodec/fdctdsp.h +++ b/libavcodec/fdctdsp.h @@ -32,6 +32,8 @@ typedef struct FDCTDSPContext { FF_VISIBILITY_PUSH_HIDDEN void ff_fdctdsp_init(FDCTDSPContext *c, struct AVCodecContext *avctx); +void ff_fdctdsp_init_aarch64(FDCTDSPContext *c, struct AVCodecContext *avctx, + unsigned high_bit_depth); void ff_fdctdsp_init_ppc(FDCTDSPContext *c, struct AVCodecContext *avctx, unsigned high_bit_depth); void ff_fdctdsp_init_x86(FDCTDSPContext *c, struct AVCodecContext *avctx, diff --git a/libavcodec/options_table.h b/libavcodec/options_table.h index ee243d9894..d9a3c92f28 100644 --- a/libavcodec/options_table.h +++ b/libavcodec/options_table.h @@ -159,6 +159,7 @@ static const AVOption avcodec_options[] = { {"mmx", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_MMX }, INT_MIN, INT_MAX, V|E, "dct"}, {"altivec", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_ALTIVEC }, INT_MIN, INT_MAX, V|E, "dct"}, {"faan", "floating point AAN DCT", 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_FAAN }, INT_MIN, INT_MAX, V|E, "dct"}, +{"neon", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_DCT_NEON }, INT_MIN, INT_MAX, V|E, "dct"}, {"lumi_mask", "compresses bright areas stronger than medium ones", OFFSET(lumi_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, {"tcplx_mask", "temporal complexity masking", OFFSET(temporal_cplx_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, {"scplx_mask", "spatial complexity masking", OFFSET(spatial_cplx_masking), AV_OPT_TYPE_FLOAT, {.dbl = 0 }, -FLT_MAX, FLT_MAX, V|E}, diff --git a/libavcodec/tests/aarch64/dct.c b/libavcodec/tests/aarch64/dct.c index 9e477328d5..e98a887cd5 100644 --- a/libavcodec/tests/aarch64/dct.c +++ b/libavcodec/tests/aarch64/dct.c @@ -19,9 +19,11 @@ #include "config.h" #include "libavutil/cpu.h" +#include "libavcodec/aarch64/fdct.h" #include "libavcodec/aarch64/idct.h" static const struct algo fdct_tab_arch[] = { + { "neon", ff_fdct_neon, FF_IDCT_PERM_NONE, AV_CPU_FLAG_NEON }, { 0 } };