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
