Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Hello, I have received the K230, and then installed Debian following your method. Therefore, I have updated the benchmark of K230 in the patch of this reply. k230 vc1dsp.vc1_inv_trans_4x4_dc_c: 125.7 vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 53.5 vc1dsp.vc1_inv_trans_4x8_dc_c: 230.7 vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 65.5 vc1dsp.vc1_inv_trans_8x4_dc_c: 228.7 vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 64.5 vc1dsp.vc1_inv_trans_8x8_dc_c: 476.5 vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 80.2 Rémi Denis-Courmont 于2023年12月4日周一 01:17写道: > Le sunnuntaina 3. joulukuuta 2023, 16.40.08 EET flow gg a écrit : > > c910 > > vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 > > vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 > > vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 > > vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 > > vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 > > vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 > > vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 > > vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 > > The code below uses fractional multipliers, so I infer that the > benchmarked > code was significantly different, and the measurements are not really > worth the > bother. > > I know that supply is a problem at the moment, but I if you are going to > keep > this up, I would hope that ISCAS can get you access to an RVV 1.0 board. > > In-line... > > > diff --git a/libavcodec/riscv/vc1dsp_init.c > b/libavcodec/riscv/vc1dsp_init.c > > new file mode 100644 > > index 00..88e0434f0e > > --- /dev/null > > +++ b/libavcodec/riscv/vc1dsp_init.c > > @@ -0,0 +1,47 @@ > > +/* > > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > > (ISCAS). > > + * > > + * 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 > > + > > +#include "libavutil/attributes.h" > > +#include "libavutil/cpu.h" > > +#include "libavutil/riscv/cpu.h" > > +#include "libavcodec/vc1.h" > > + > > +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > + > > +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) > > +{ > > +#if HAVE_RVV > > +int flags = av_get_cpu_flags(); > > + > > +if (flags & AV_CPU_FLAG_RVV_I64) { > > +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; > > +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; > > +} > > +if (flags & AV_CPU_FLAG_RVV_I32) { > > +dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv; > > +dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv; > > +} > > Probably missing VLENB checks. > > > +#endif > > +} > > diff --git a/libavcodec/riscv/vc1dsp_rvv.S > b/libavcodec/riscv/vc1dsp_rvv.S > > new file mode 100644 > > index 00..8a6b27192a > > --- /dev/null > > +++ b/libavcodec/riscv/vc1dsp_rvv.S > > @@ -0,0 +1,123 @@ > > +/* > > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > > (ISCAS). > > + * > > + * 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/riscv/asm.S" > > + > > +func ff_vc1_inv_trans_8x8_dc_rvv, zve64x > > +lht2, (a2) > > +li
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
> FWIW CanMV-K230 boards are on sale for under 500 RMB. I just made a payment ~ (I saw you mention in IRC that you're going to write about K230+Debian. Looking forward to it) Rémi Denis-Courmont 于2023年12月6日周三 04:11写道: > Le tiistaina 5. joulukuuta 2023, 21.25.12 EET flow gg a écrit : > > > This block can be folded into the next. You don't need to check VLENB > > > > twice. > > > > Changed. > > > > > Instruction scheduling could be better, especially on in-order CPUs. > > > > I put the vload at the front, and then proceeded with the t2 operation, > but > > I'm not sure... > > > > > You don't need to reset the AVL here, just pass zero. > > > > Changed. > > > > > vsetivli > > > > Changed. > > You changed more than I asked for. The immediate AVL is a 5-bit unsigned > integer, so it should not be possible to assemble 32 or 64, unless you > have a > preprocessor that silently rewrites `vsetivli` into `vsetvli` (If so, that > sounds very iffy because `vsetivli zero` has no scratch X register to work > with). > > FWIW CanMV-K230 boards are on sale for under 500 RMB. > > -- > レミ・デニ-クールモン > http://www.remlab.net/ > > > > ___ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe". > ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
I'm sorry for my carelessness.It's because I used to build and run manually, but now I've switched to a script to do it, so I accidentally missed the error.I will modify the script and to avoid this kind of issue in the future. libavcodec/riscv/vc1dsp_rvv.S:35: Error: improper CSRxI immediate Changed. Rémi Denis-Courmont 于2023年12月6日周三 04:11写道: > Le tiistaina 5. joulukuuta 2023, 21.25.12 EET flow gg a écrit : > > > This block can be folded into the next. You don't need to check VLENB > > > > twice. > > > > Changed. > > > > > Instruction scheduling could be better, especially on in-order CPUs. > > > > I put the vload at the front, and then proceeded with the t2 operation, > but > > I'm not sure... > > > > > You don't need to reset the AVL here, just pass zero. > > > > Changed. > > > > > vsetivli > > > > Changed. > > You changed more than I asked for. The immediate AVL is a 5-bit unsigned > integer, so it should not be possible to assemble 32 or 64, unless you > have a > preprocessor that silently rewrites `vsetivli` into `vsetvli` (If so, that > sounds very iffy because `vsetivli zero` has no scratch X register to work > with). > > FWIW CanMV-K230 boards are on sale for under 500 RMB. > > -- > レミ・デニ-クールモン > http://www.remlab.net/ > > > > ___ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe". > From 653b8a8651aaa6c22e8a3a400b3e493eb03704f2 Mon Sep 17 00:00:00 2001 From: sunyuechi Date: Fri, 1 Dec 2023 10:07:40 +0800 Subject: [PATCH] lavc/vc1dsp: R-V V inv_trans c910 vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 --- libavcodec/riscv/Makefile | 2 + libavcodec/riscv/vc1dsp_init.c | 49 ++ libavcodec/riscv/vc1dsp_rvv.S | 113 + libavcodec/vc1dsp.c| 2 + libavcodec/vc1dsp.h| 1 + 5 files changed, 167 insertions(+) create mode 100644 libavcodec/riscv/vc1dsp_init.c create mode 100644 libavcodec/riscv/vc1dsp_rvv.S diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile index 2d0e6c19c8..442c5961ea 100644 --- a/libavcodec/riscv/Makefile +++ b/libavcodec/riscv/Makefile @@ -39,5 +39,7 @@ OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_init.o \ RVV-OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_rvv.o OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_init.o RVV-OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_rvv.o +OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_init.o +RVV-OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_rvv.o OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c new file mode 100644 index 00..0d22d28f4d --- /dev/null +++ b/libavcodec/riscv/vc1dsp_init.c @@ -0,0 +1,49 @@ +/* + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences (ISCAS). + * + * 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 + +#include "libavutil/attributes.h" +#include "libavutil/cpu.h" +#include "libavutil/riscv/cpu.h" +#include "libavcodec/vc1.h" + +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); + +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) +{ +#if HAVE_RVV +int flags = av_get_cpu_flags(); + +if (ff_get_rv_vlenb() >= 16) { +if (flags & AV_CPU_FLAG_RVV_I64) { +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; +} +if (flags & AV_CPU_FLAG_RVV_I32) { +
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Le tiistaina 5. joulukuuta 2023, 21.25.12 EET flow gg a écrit : > > This block can be folded into the next. You don't need to check VLENB > > twice. > > Changed. > > > Instruction scheduling could be better, especially on in-order CPUs. > > I put the vload at the front, and then proceeded with the t2 operation, but > I'm not sure... > > > You don't need to reset the AVL here, just pass zero. > > Changed. > > > vsetivli > > Changed. You changed more than I asked for. The immediate AVL is a 5-bit unsigned integer, so it should not be possible to assemble 32 or 64, unless you have a preprocessor that silently rewrites `vsetivli` into `vsetvli` (If so, that sounds very iffy because `vsetivli zero` has no scratch X register to work with). FWIW CanMV-K230 boards are on sale for under 500 RMB. -- レミ・デニ-クールモン http://www.remlab.net/ ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
> This block can be folded into the next. You don't need to check VLENB twice. Changed. > Instruction scheduling could be better, especially on in-order CPUs. I put the vload at the front, and then proceeded with the t2 operation, but I'm not sure... > You don't need to reset the AVL here, just pass zero. Changed. > vsetivli Changed. Rémi Denis-Courmont 于2023年12月6日周三 00:41写道: > Hi, > > > diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile > > index 2d0e6c19c8..442c5961ea 100644 > > --- a/libavcodec/riscv/Makefile > > +++ b/libavcodec/riscv/Makefile > > @@ -39,5 +39,7 @@ OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_init.o > \ > > RVV-OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_rvv.o > > OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_init.o > > RVV-OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_rvv.o > > +OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_init.o > > +RVV-OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_rvv.o > > OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o > > RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o > > diff --git a/libavcodec/riscv/vc1dsp_init.c > b/libavcodec/riscv/vc1dsp_init.c > > new file mode 100644 > > index 00..6a04bea6d6 > > --- /dev/null > > +++ b/libavcodec/riscv/vc1dsp_init.c > > @@ -0,0 +1,47 @@ > > +/* > > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > > (ISCAS). > > + * > > + * 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 > > + > > +#include "libavutil/attributes.h" > > +#include "libavutil/cpu.h" > > +#include "libavutil/riscv/cpu.h" > > +#include "libavcodec/vc1.h" > > + > > +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > + > > +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) > > +{ > > +#if HAVE_RVV > > +int flags = av_get_cpu_flags(); > > + > > +if ((flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) { > > +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; > > +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; > > +} > > This block can be folded into the next. You don't need to check VLENB > twice. > > > +if ((flags & AV_CPU_FLAG_RVV_I32) && ff_get_rv_vlenb() >= 16) { > > +dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv; > > +dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv; > > +} > > +#endif > > +} > > diff --git a/libavcodec/riscv/vc1dsp_rvv.S > b/libavcodec/riscv/vc1dsp_rvv.S > > new file mode 100644 > > index 00..b865bd2cbc > > --- /dev/null > > +++ b/libavcodec/riscv/vc1dsp_rvv.S > > @@ -0,0 +1,114 @@ > > +/* > > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > > (ISCAS). > > + * > > + * 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/riscv/asm.S" > > + > > +func ff_vc1_inv_trans_8x8_dc_rvv, zve64x > > +lht2, (a2) > > +sh1addt2, t2, t2 > > +addi t2, t2, 1 > > +srai t2, t2, 1 > > +sh1addt2, t2, t2 > > +addi t2, t2, 16 > > +
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Hi, > diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile > index 2d0e6c19c8..442c5961ea 100644 > --- a/libavcodec/riscv/Makefile > +++ b/libavcodec/riscv/Makefile > @@ -39,5 +39,7 @@ OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_init.o \ > RVV-OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_rvv.o > OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_init.o > RVV-OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_rvv.o > +OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_init.o > +RVV-OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_rvv.o > OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o > RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o > diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c > new file mode 100644 > index 00..6a04bea6d6 > --- /dev/null > +++ b/libavcodec/riscv/vc1dsp_init.c > @@ -0,0 +1,47 @@ > +/* > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > (ISCAS). > + * > + * 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 > + > +#include "libavutil/attributes.h" > +#include "libavutil/cpu.h" > +#include "libavutil/riscv/cpu.h" > +#include "libavcodec/vc1.h" > + > +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > + > +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) > +{ > +#if HAVE_RVV > +int flags = av_get_cpu_flags(); > + > +if ((flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) { > +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; > +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; > +} This block can be folded into the next. You don't need to check VLENB twice. > +if ((flags & AV_CPU_FLAG_RVV_I32) && ff_get_rv_vlenb() >= 16) { > +dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv; > +dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv; > +} > +#endif > +} > diff --git a/libavcodec/riscv/vc1dsp_rvv.S b/libavcodec/riscv/vc1dsp_rvv.S > new file mode 100644 > index 00..b865bd2cbc > --- /dev/null > +++ b/libavcodec/riscv/vc1dsp_rvv.S > @@ -0,0 +1,114 @@ > +/* > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > (ISCAS). > + * > + * 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/riscv/asm.S" > + > +func ff_vc1_inv_trans_8x8_dc_rvv, zve64x > +lht2, (a2) > +sh1addt2, t2, t2 > +addi t2, t2, 1 > +srai t2, t2, 1 > +sh1addt2, t2, t2 > +addi t2, t2, 16 > +srai t2, t2, 5 > +vsetivli zero, 8, e8, mf2, ta, ma > +vlse64.v v0, (a0), a1 > +lit0, 8*8 > +vsetvli zero, t0, e16, m8, ta, ma Instruction scheduling could be better, especially on in-order CPUs. > +vzext.vf2 v8, v0 > +vadd.vx v8, v8, t2 > +vmax.vx v8, v8, zero > +vsetvli zero, t0, e8, m4, ta, ma You don't need to reset the AVL here, just pass zero. > +vnclipu.wiv0, v8, 0 > +vsetivli zero, 8, e8, mf2, ta, ma > +vsse64.v v0, (a0), a1 > +ret > +endfunc >
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Okay, after using zext, can delete two vset, which is better than splat. I have updated the patch in this reply. Rémi Denis-Courmont 于2023年12月4日周一 23:15写道: > Le maanantaina 4. joulukuuta 2023, 10.48.56 EET flow gg a écrit : > > > Probably missing VLENB checks. > > > > Changed. > > > > > You can multiply by 3, 5 or 9 with shift-and-add. By 12 with > shift-and-add > > > then shift, and by 17 with shift then add. You don't need > multiplications. > > > > Changed. > > > > > Do you really need to splat? Can't .vx or .wx be used instead? > > > > Okay, for example in ff_vc1_inv_trans_8x8_dc_rvv > > > > + vsetvli zero, t0, e8, m2, ta, ma > > + vwaddu.vxv4, v0, zero > > + vsetvli zero, t0, e16, m4, ta, ma > > + vadd.vx v4, v4, t2 > > - vsetvli zero, t0, e16, m4, ta, ma > > - vmv.v.x v4, t2 > > - vsetvli zero, t0, e8, m2, ta, ma > > - vwaddu.wvv4, v4, v0 > > > > But the speed has slowed down slightly on the c910, > > I'm not sure if I should modify it. > > OK, unfortunately, there is no widening addition with wide scalar operand. > But > you can do zero-extension then addition here. In the end, I doubt that you > can > reasonably optimise whilst working with a C910-based board. This function > deviates too much on non-conformant hardware. > > -- > レミ・デニ-クールモン > http://www.remlab.net/ > > > > ___ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe". > From 4022465172a5ac0c9669795e318edfa74e9d346e Mon Sep 17 00:00:00 2001 From: sunyuechi Date: Fri, 1 Dec 2023 10:07:40 +0800 Subject: [PATCH] lavc/vc1dsp: R-V V inv_trans c910 vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 --- libavcodec/riscv/Makefile | 2 + libavcodec/riscv/vc1dsp_init.c | 47 ++ libavcodec/riscv/vc1dsp_rvv.S | 114 + libavcodec/vc1dsp.c| 2 + libavcodec/vc1dsp.h| 1 + 5 files changed, 166 insertions(+) create mode 100644 libavcodec/riscv/vc1dsp_init.c create mode 100644 libavcodec/riscv/vc1dsp_rvv.S diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile index 2d0e6c19c8..442c5961ea 100644 --- a/libavcodec/riscv/Makefile +++ b/libavcodec/riscv/Makefile @@ -39,5 +39,7 @@ OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_init.o \ RVV-OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_rvv.o OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_init.o RVV-OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_rvv.o +OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_init.o +RVV-OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_rvv.o OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c new file mode 100644 index 00..6a04bea6d6 --- /dev/null +++ b/libavcodec/riscv/vc1dsp_init.c @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences (ISCAS). + * + * 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 + +#include "libavutil/attributes.h" +#include "libavutil/cpu.h" +#include "libavutil/riscv/cpu.h" +#include "libavcodec/vc1.h" + +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); + +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) +{ +#if HAVE_RVV +int flags = av_get_cpu_flags(); + +if ((flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) { +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_d
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
I found that in the case of nosplat, an additional vset can be removed, and the time is basically the same, so I updated the patch. Rémi Denis-Courmont 于2023年12月4日周一 23:15写道: > Le maanantaina 4. joulukuuta 2023, 10.48.56 EET flow gg a écrit : > > > Probably missing VLENB checks. > > > > Changed. > > > > > You can multiply by 3, 5 or 9 with shift-and-add. By 12 with > shift-and-add > > > then shift, and by 17 with shift then add. You don't need > multiplications. > > > > Changed. > > > > > Do you really need to splat? Can't .vx or .wx be used instead? > > > > Okay, for example in ff_vc1_inv_trans_8x8_dc_rvv > > > > + vsetvli zero, t0, e8, m2, ta, ma > > + vwaddu.vxv4, v0, zero > > + vsetvli zero, t0, e16, m4, ta, ma > > + vadd.vx v4, v4, t2 > > - vsetvli zero, t0, e16, m4, ta, ma > > - vmv.v.x v4, t2 > > - vsetvli zero, t0, e8, m2, ta, ma > > - vwaddu.wvv4, v4, v0 > > > > But the speed has slowed down slightly on the c910, > > I'm not sure if I should modify it. > > OK, unfortunately, there is no widening addition with wide scalar operand. > But > you can do zero-extension then addition here. In the end, I doubt that you > can > reasonably optimise whilst working with a C910-based board. This function > deviates too much on non-conformant hardware. > > -- > レミ・デニ-クールモン > http://www.remlab.net/ > > > > ___ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe". > From e4040b48245ee1733bcda4168e3c78a9a9cf82c0 Mon Sep 17 00:00:00 2001 From: sunyuechi Date: Fri, 1 Dec 2023 10:07:40 +0800 Subject: [PATCH] lavc/vc1dsp: R-V V inv_trans c910 vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 --- libavcodec/riscv/Makefile | 2 + libavcodec/riscv/vc1dsp_init.c | 47 + libavcodec/riscv/vc1dsp_rvv.S | 118 + libavcodec/vc1dsp.c| 2 + libavcodec/vc1dsp.h| 1 + 5 files changed, 170 insertions(+) create mode 100644 libavcodec/riscv/vc1dsp_init.c create mode 100644 libavcodec/riscv/vc1dsp_rvv.S diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile index 2d0e6c19c8..442c5961ea 100644 --- a/libavcodec/riscv/Makefile +++ b/libavcodec/riscv/Makefile @@ -39,5 +39,7 @@ OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_init.o \ RVV-OBJS-$(CONFIG_PIXBLOCKDSP) += riscv/pixblockdsp_rvv.o OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_init.o RVV-OBJS-$(CONFIG_UTVIDEO_DECODER) += riscv/utvideodsp_rvv.o +OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_init.o +RVV-OBJS-$(CONFIG_VC1DSP) += riscv/vc1dsp_rvv.o OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c new file mode 100644 index 00..6a04bea6d6 --- /dev/null +++ b/libavcodec/riscv/vc1dsp_init.c @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences (ISCAS). + * + * 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 + +#include "libavutil/attributes.h" +#include "libavutil/cpu.h" +#include "libavutil/riscv/cpu.h" +#include "libavcodec/vc1.h" + +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t *block); + +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) +{ +#if HAVE_RVV +int flags = av_get_cpu_flags(); + +if ((flags & AV_CPU_FLAG_RVV_I64) && ff_get_rv_vlenb() >= 16) { +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; +dsp->vc1_inv_trans_8x4_dc = ff_v
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Le maanantaina 4. joulukuuta 2023, 10.48.56 EET flow gg a écrit : > > Probably missing VLENB checks. > > Changed. > > > You can multiply by 3, 5 or 9 with shift-and-add. By 12 with shift-and-add > > then shift, and by 17 with shift then add. You don't need multiplications. > > Changed. > > > Do you really need to splat? Can't .vx or .wx be used instead? > > Okay, for example in ff_vc1_inv_trans_8x8_dc_rvv > > + vsetvli zero, t0, e8, m2, ta, ma > + vwaddu.vxv4, v0, zero > + vsetvli zero, t0, e16, m4, ta, ma > + vadd.vx v4, v4, t2 > - vsetvli zero, t0, e16, m4, ta, ma > - vmv.v.x v4, t2 > - vsetvli zero, t0, e8, m2, ta, ma > - vwaddu.wvv4, v4, v0 > > But the speed has slowed down slightly on the c910, > I'm not sure if I should modify it. OK, unfortunately, there is no widening addition with wide scalar operand. But you can do zero-extension then addition here. In the end, I doubt that you can reasonably optimise whilst working with a C910-based board. This function deviates too much on non-conformant hardware. -- レミ・デニ-クールモン http://www.remlab.net/ ___ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
> Probably missing VLENB checks. Changed. > You can multiply by 3, 5 or 9 with shift-and-add. By 12 with shift-and-add > then shift, and by 17 with shift then add. You don't need multiplications. Changed. > Do you really need to splat? Can't .vx or .wx be used instead? Okay, for example in ff_vc1_inv_trans_8x8_dc_rvv + vsetvli zero, t0, e8, m2, ta, ma + vwaddu.vxv4, v0, zero + vsetvli zero, t0, e16, m4, ta, ma + vadd.vx v4, v4, t2 - vsetvli zero, t0, e16, m4, ta, ma - vmv.v.x v4, t2 - vsetvli zero, t0, e8, m2, ta, ma - vwaddu.wvv4, v4, v0 But the speed has slowed down slightly on the c910, I'm not sure if I should modify it. splat| no splat vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 | 84.0 vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 | 76.0 vc1dsp.vc1_inv_trans_4x8_dc_c: 150.0 | 150.0 vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 | 84.5 vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0| 129.0 vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 76.0 | 76.7 vc1dsp.vc1_inv_trans_8x8_dc_c: 255.0 | 254.7 vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 | 93.0 > The code below uses fractional multipliers, so I infer that the benchmarked > code was significantly different, and the measurements are not really worth the > bother. > > I know that supply is a problem at the moment, but I if you are going to keep > this up, I would hope that ISCAS can get you access to an RVV 1.0 board. Using mf2 only requires changing the first vset and the last vset in each function to mf2. I guess they would achieve similar effects on both c910 and c908? example in 8x8 ``` - vsetivlizero, 8, e64, m4, ta, ma + vsetivlizero, 8, e8, mf2, ta, ma - vsetvli zero, zero, e64, m4, ta, ma + vsetivlizero, 8, e8, mf2, ta, ma ``` And ISCAS seems to have no announcement about getting an RVV 1.0 board. I plan to ask about it from time to time. Rémi Denis-Courmont 于2023年12月4日周一 01:17写道: > Le sunnuntaina 3. joulukuuta 2023, 16.40.08 EET flow gg a écrit : > > c910 > > vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 > > vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 > > vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 > > vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 > > vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 > > vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 > > vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 > > vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 > > The code below uses fractional multipliers, so I infer that the > benchmarked > code was significantly different, and the measurements are not really > worth the > bother. > > I know that supply is a problem at the moment, but I if you are going to > keep > this up, I would hope that ISCAS can get you access to an RVV 1.0 board. > > In-line... > > > diff --git a/libavcodec/riscv/vc1dsp_init.c > b/libavcodec/riscv/vc1dsp_init.c > > new file mode 100644 > > index 00..88e0434f0e > > --- /dev/null > > +++ b/libavcodec/riscv/vc1dsp_init.c > > @@ -0,0 +1,47 @@ > > +/* > > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > > (ISCAS). > > + * > > + * 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 > > + > > +#include "libavutil/attributes.h" > > +#include "libavutil/cpu.h" > > +#include "libavutil/riscv/cpu.h" > > +#include "libavcodec/vc1.h" > > + > > +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, > int16_t > > *block); > > + > > +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) > > +{ > > +#if HAVE_RVV > > +int flags = av_get_cpu_flags(); > > + > > +if (flags & AV_CPU_FLAG_RVV_I64) { > > +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; > > +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; > > +} > > +if (flags & AV_CPU_FLAG_RVV_I32) { > > +dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv; > > +dsp->vc1_inv_tr
Re: [FFmpeg-devel] [PATCH] lavc/vc1dsp: R-V V inv_trans
Le sunnuntaina 3. joulukuuta 2023, 16.40.08 EET flow gg a écrit : > c910 > vc1dsp.vc1_inv_trans_4x4_dc_c: 84.0 > vc1dsp.vc1_inv_trans_4x4_dc_rvv_i32: 74.0 > vc1dsp.vc1_inv_trans_4x8_dc_c: 150.2 > vc1dsp.vc1_inv_trans_4x8_dc_rvv_i32: 83.5 > vc1dsp.vc1_inv_trans_8x4_dc_c: 129.0 > vc1dsp.vc1_inv_trans_8x4_dc_rvv_i64: 75.7 > vc1dsp.vc1_inv_trans_8x8_dc_c: 254.7 > vc1dsp.vc1_inv_trans_8x8_dc_rvv_i64: 90.5 The code below uses fractional multipliers, so I infer that the benchmarked code was significantly different, and the measurements are not really worth the bother. I know that supply is a problem at the moment, but I if you are going to keep this up, I would hope that ISCAS can get you access to an RVV 1.0 board. In-line... > diff --git a/libavcodec/riscv/vc1dsp_init.c b/libavcodec/riscv/vc1dsp_init.c > new file mode 100644 > index 00..88e0434f0e > --- /dev/null > +++ b/libavcodec/riscv/vc1dsp_init.c > @@ -0,0 +1,47 @@ > +/* > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > (ISCAS). > + * > + * 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 > + > +#include "libavutil/attributes.h" > +#include "libavutil/cpu.h" > +#include "libavutil/riscv/cpu.h" > +#include "libavcodec/vc1.h" > + > +void ff_vc1_inv_trans_8x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_4x8_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_8x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > +void ff_vc1_inv_trans_4x4_dc_rvv(uint8_t *dest, ptrdiff_t stride, int16_t > *block); > + > +av_cold void ff_vc1dsp_init_riscv(VC1DSPContext *dsp) > +{ > +#if HAVE_RVV > +int flags = av_get_cpu_flags(); > + > +if (flags & AV_CPU_FLAG_RVV_I64) { > +dsp->vc1_inv_trans_8x8_dc = ff_vc1_inv_trans_8x8_dc_rvv; > +dsp->vc1_inv_trans_8x4_dc = ff_vc1_inv_trans_8x4_dc_rvv; > +} > +if (flags & AV_CPU_FLAG_RVV_I32) { > +dsp->vc1_inv_trans_4x8_dc = ff_vc1_inv_trans_4x8_dc_rvv; > +dsp->vc1_inv_trans_4x4_dc = ff_vc1_inv_trans_4x4_dc_rvv; > +} Probably missing VLENB checks. > +#endif > +} > diff --git a/libavcodec/riscv/vc1dsp_rvv.S b/libavcodec/riscv/vc1dsp_rvv.S > new file mode 100644 > index 00..8a6b27192a > --- /dev/null > +++ b/libavcodec/riscv/vc1dsp_rvv.S > @@ -0,0 +1,123 @@ > +/* > + * Copyright (c) 2023 Institue of Software Chinese Academy of Sciences > (ISCAS). > + * > + * 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/riscv/asm.S" > + > +func ff_vc1_inv_trans_8x8_dc_rvv, zve64x > +lht2, (a2) > +lit1, 3 > +mul t2, t2, t1 You can multiply by 3, 5 or 9 with shift-and-add. By 12 with shift-and-add then shift, and by 17 with shift then add. You don't need multiplications. > +addi t2, t2, 1 > +srai t2, t2, 1 > +mul t2, t2, t1 > +addi t2, t2, 16 > +srai t2, t2, 5 > +vsetivli zero, 8, e8, mf2, ta, ma > +vlse64.v v0, (a0), a1 > +lit0, 8*8 > +vsetvli zero, t0, e16, m8, ta, ma > +vmv.v.x v8, t2 Do you really need to splat? Can't .vx or .wx be used instead? > +vsetvli zero, t0, e8, m4, ta, ma > +vwaddu.wv v8, v8, v0 > +vsetvli zero, t0, e16, m8, ta, ma > +vmax.vx