Re: [FFmpeg-devel] Subject: [PATCH 3/3] lavc/dnxhdenc: R-V V get_pixels_8x4_sym

[flow gg <hlefthleft@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 3132 bytes --]

> I expect that it would be faster to make one large load, and then 4 small
> stores, but that might work only for exactly 128-bit vectors?

This seems to require vle128, so I didn't modify it.

> That's not needed. You can use immediate values.
> You can reorder to avoid immediate data dependencies on the addresses.
> In any case, you need to check the vector length in init.

Okay, I've updated it in the reply.

Rémi Denis-Courmont <remi@remlab.net> 于2024年1月29日周一 23:41写道：

> Hi,
>
> +/*
> + * 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_get_pixels_8x4_sym_rvv, zve64x
> +        vsetivli    zero, 8, e8, mf2, ta, ma
> +        vlse64.v    v16, (a1), a2
> +        li          t0, 8 * 8
> +        vsetvli     zero, t0, e16, m4, ta, ma
> +        vzext.vf2   v8, v16
> +        vse16.v     v8, (a0)
> +        li          a2, 8*2
>
> That's not needed. You can use immediate values.
>
> +        vsetivli    zero, 2, e8, mf8, ta, ma
> +        addi        a1, a0, 48
> +        addi        a0, a0, 32*2
> +        vle64.v     v0, (a1)
> +        vse64.v     v0, (a0)
> +        sub         a1, a1, a2
> +        vle64.v     v0, (a1)
> +        add         a0, a0, a2
> +        vse64.v     v0, (a0)
> +        sub         a1, a1, a2
> +        vle64.v     v0, (a1)
> +        add         a0, a0, a2
> +        vse64.v     v0, (a0)
> +        sub         a1, a1, a2
> +        vle64.v     v0, (a1)
> +        add         a0, a0, a2
> +        vse64.v     v0, (a0)
>
> You can reorder to avoid immediate data dependencies on the addresses.
>
> I expect that it would be faster to make one large load, and then 4 small
> stores, but that might work only for exactly 128-bit vectors?
>
> In any case, you need to check the vector length in init.
>
> +
> +        ret
> +endfunc
>
> --
> 雷米‧德尼-库尔蒙
> 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-request@ffmpeg.org with subject "unsubscribe".
>

[-- Attachment #2: 0003-lavc-dnxhdenc-R-V-V-get_pixels_8x4_sym.patch --]
[-- Type: text/x-patch, Size: 5748 bytes --]

From 70e40a02d04d91c7f25c2127eb561c03e34792b9 Mon Sep 17 00:00:00 2001
From: sunyuechi <sunyuechi@iscas.ac.cn>
Date: Wed, 20 Dec 2023 16:28:33 +0800
Subject: [PATCH 3/3] lavc/dnxhdenc: R-V V get_pixels_8x4_sym

C908:
get_pixels_8x4_sym_c: 297.2
get_pixels_8x4_sym_rvv_i64: 52.7
---
 libavcodec/dnxhdenc.c          |  4 ++-
 libavcodec/dnxhdenc.h          |  1 +
 libavcodec/riscv/Makefile      |  2 ++
 libavcodec/riscv/dnxenc_init.c | 42 +++++++++++++++++++++++++++++
 libavcodec/riscv/dnxenc_rvv.S  | 49 ++++++++++++++++++++++++++++++++++
 5 files changed, 97 insertions(+), 1 deletion(-)
 create mode 100644 libavcodec/riscv/dnxenc_init.c
 create mode 100644 libavcodec/riscv/dnxenc_rvv.S

diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c
index 0d9727bbbf..bab92753fa 100644
--- a/libavcodec/dnxhdenc.c
+++ b/libavcodec/dnxhdenc.c
@@ -1378,7 +1378,9 @@ const FFCodec ff_dnxhd_encoder = {
 
 void ff_dnxhdenc_init(DNXHDEncContext *ctx)
 {
-#if ARCH_X86
+#if ARCH_RISCV
+    ff_dnxhdenc_init_riscv(ctx);
+#elif ARCH_X86
     ff_dnxhdenc_init_x86(ctx);
 #endif
 }
diff --git a/libavcodec/dnxhdenc.h b/libavcodec/dnxhdenc.h
index 95aea83d28..3ed1451431 100644
--- a/libavcodec/dnxhdenc.h
+++ b/libavcodec/dnxhdenc.h
@@ -112,6 +112,7 @@ typedef struct DNXHDEncContext {
 } DNXHDEncContext;
 
 void ff_dnxhdenc_init(DNXHDEncContext *ctx);
+void ff_dnxhdenc_init_riscv(DNXHDEncContext *ctx);
 void ff_dnxhdenc_init_x86(DNXHDEncContext *ctx);
 
 #endif /* AVCODEC_DNXHDENC_H */
diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile
index fe604a1156..e15aba58f4 100644
--- a/libavcodec/riscv/Makefile
+++ b/libavcodec/riscv/Makefile
@@ -13,6 +13,8 @@ RVV-OBJS-$(CONFIG_AUDIODSP) += riscv/audiodsp_rvv.o
 OBJS-$(CONFIG_BSWAPDSP) += riscv/bswapdsp_init.o
 RV-OBJS-$(CONFIG_BSWAPDSP) += riscv/bswapdsp_rvb.o
 RVV-OBJS-$(CONFIG_BSWAPDSP) += riscv/bswapdsp_rvv.o
+OBJS-$(CONFIG_DNXHD_ENCODER) += riscv/dnxenc_init.o
+RVV-OBJS-$(CONFIG_DNXHD_ENCODER) += riscv/dnxenc_rvv.o
 OBJS-$(CONFIG_EXR_DECODER) += riscv/exrdsp_init.o
 RVV-OBJS-$(CONFIG_EXR_DECODER) += riscv/exrdsp_rvv.o
 OBJS-$(CONFIG_FLAC_DECODER) += riscv/flacdsp_init.o
diff --git a/libavcodec/riscv/dnxenc_init.c b/libavcodec/riscv/dnxenc_init.c
new file mode 100644
index 0000000000..9a42b17224
--- /dev/null
+++ b/libavcodec/riscv/dnxenc_init.c
@@ -0,0 +1,42 @@
+/*
+ * 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 "config.h"
+
+#include "libavutil/attributes.h"
+#include "libavutil/cpu.h"
+#include "libavutil/riscv/cpu.h"
+#include "libavcodec/dnxhdenc.h"
+
+void ff_get_pixels_8x4_sym_rvv(int16_t *block, const uint8_t *pixels,
+                                ptrdiff_t line_size);
+
+av_cold void ff_dnxhdenc_init_riscv(DNXHDEncContext *ctx)
+{
+#if HAVE_RVV
+    int flags = av_get_cpu_flags();
+
+    if (flags & AV_CPU_FLAG_RVV_I64 && ff_get_rv_vlenb() >= 16) {
+        if (ctx->cid_table->bit_depth == 8) {
+            ctx->get_pixels_8x4_sym = ff_get_pixels_8x4_sym_rvv;
+        }
+    }
+#endif
+}
diff --git a/libavcodec/riscv/dnxenc_rvv.S b/libavcodec/riscv/dnxenc_rvv.S
new file mode 100644
index 0000000000..ac4a274009
--- /dev/null
+++ b/libavcodec/riscv/dnxenc_rvv.S
@@ -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 "libavutil/riscv/asm.S"
+
+func ff_get_pixels_8x4_sym_rvv, zve64x
+        vsetivli    zero, 8, e8, mf2, ta, ma
+        vlse64.v    v16, (a1), a2
+        li          t0, 8 * 8
+        vsetvli     zero, t0, e16, m4, ta, ma
+        vzext.vf2   v8, v16
+        vse16.v     v8, (a0)
+        vsetivli    zero, 2, e8, mf8, ta, ma
+        addi        a1, a0, 48
+        addi        a0, a0, 32*2
+        vle64.v     v0, (a1)
+        vse64.v     v0, (a0)
+        addi        a1, a1, -16
+        addi        a0, a0, 16
+        vle64.v     v0, (a1)
+        vse64.v     v0, (a0)
+        addi        a1, a1, -16
+        addi        a0, a0, 16
+        vle64.v     v0, (a1)
+        vse64.v     v0, (a0)
+        addi        a1, a1, -16
+        addi        a0, a0, 16
+        vle64.v     v0, (a1)
+        vse64.v     v0, (a0)
+
+        ret
+endfunc
-- 
2.43.0


[-- Attachment #3: Type: text/plain, Size: 251 bytes --]

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".