From: "george.zaguri via ffmpeg-devel" <ffmpeg-devel@ffmpeg.org>
To: ffmpeg-devel@ffmpeg.org
Cc: "george.zaguri" <code@ffmpeg.org>
Subject: [FFmpeg-devel] [PATCH] aarch64/vvc: Optimisations of put_luma_h() functions for 10/12-bit (PR #20737)
Date: Wed, 22 Oct 2025 18:30:05 -0000
Message-ID: <176115780645.25.2812638608647755938@7d278768979e> (raw)
PR #20737 opened by george.zaguri
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20737
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20737.patch
RPi4:
put_luma_h_10_4x4_c: 261.8 ( 1.00x)
put_luma_h_10_8x8_c: 1051.5 ( 1.00x)
put_luma_h_10_8x8_neon: 231.5 ( 4.54x)
put_luma_h_10_16x16_c: 4131.0 ( 1.00x)
put_luma_h_10_16x16_neon: 848.6 ( 4.87x)
put_luma_h_10_32x32_c: 16469.5 ( 1.00x)
put_luma_h_10_32x32_neon: 3345.6 ( 4.92x)
put_luma_h_10_64x64_c: 66734.0 ( 1.00x)
put_luma_h_10_64x64_neon: 14586.9 ( 4.57x)
put_luma_h_10_128x128_c: 264228.9 ( 1.00x)
put_luma_h_10_128x128_neon: 52199.7 ( 5.06x)
put_luma_h_12_4x4_c: 262.1 ( 1.00x)
put_luma_h_12_8x8_c: 1051.3 ( 1.00x)
put_luma_h_12_8x8_neon: 230.9 ( 4.55x)
put_luma_h_12_16x16_c: 4124.4 ( 1.00x)
put_luma_h_12_16x16_neon: 848.0 ( 4.86x)
put_luma_h_12_32x32_c: 16446.9 ( 1.00x)
put_luma_h_12_32x32_neon: 3347.4 ( 4.91x)
put_luma_h_12_64x64_c: 66770.1 ( 1.00x)
put_luma_h_12_64x64_neon: 14360.2 ( 4.65x)
put_luma_h_12_128x128_c: 264419.5 ( 1.00x)
put_luma_h_12_128x128_neon: 52200.6 ( 5.07x)
M2 Air (with auto-vectorization feature):
put_luma_h_10_4x4_c: 0.3 ( 1.00x)
put_luma_h_10_8x8_c: 1.0 ( 1.00x)
put_luma_h_10_8x8_neon: 0.4 ( 2.58x)
put_luma_h_10_16x16_c: 3.0 ( 1.00x)
put_luma_h_10_16x16_neon: 1.5 ( 2.01x)
put_luma_h_10_32x32_c: 9.7 ( 1.00x)
put_luma_h_10_32x32_neon: 6.2 ( 1.57x)
put_luma_h_10_64x64_c: 36.6 ( 1.00x)
put_luma_h_10_64x64_neon: 23.9 ( 1.53x)
put_luma_h_10_128x128_c: 134.2 ( 1.00x)
put_luma_h_10_128x128_neon: 95.4 ( 1.41x)
put_luma_h_12_4x4_c: 0.3 ( 1.00x)
put_luma_h_12_8x8_c: 1.0 ( 1.00x)
put_luma_h_12_8x8_neon: 0.4 ( 2.57x)
put_luma_h_12_16x16_c: 3.0 ( 1.00x)
put_luma_h_12_16x16_neon: 1.5 ( 2.01x)
put_luma_h_12_32x32_c: 9.7 ( 1.00x)
put_luma_h_12_32x32_neon: 6.0 ( 1.63x)
put_luma_h_12_64x64_c: 36.5 ( 1.00x)
put_luma_h_12_64x64_neon: 23.9 ( 1.53x)
put_luma_h_12_128x128_c: 134.8 ( 1.00x)
put_luma_h_12_128x128_neon: 95.2 ( 1.42x)
>From dba0d5709658f01e40496d1f1fc8a1832e21b708 Mon Sep 17 00:00:00 2001
From: Georgii Zagoruiko <george.zaguri@gmail.com>
Date: Wed, 22 Oct 2025 19:22:23 +0100
Subject: [PATCH] aarch64/vvc: Optimisations of put_luma_h() functions for
10/12-bit
RPi4:
put_luma_h_10_4x4_c: 261.8 ( 1.00x)
put_luma_h_10_8x8_c: 1051.5 ( 1.00x)
put_luma_h_10_8x8_neon: 231.5 ( 4.54x)
put_luma_h_10_16x16_c: 4131.0 ( 1.00x)
put_luma_h_10_16x16_neon: 848.6 ( 4.87x)
put_luma_h_10_32x32_c: 16469.5 ( 1.00x)
put_luma_h_10_32x32_neon: 3345.6 ( 4.92x)
put_luma_h_10_64x64_c: 66734.0 ( 1.00x)
put_luma_h_10_64x64_neon: 14586.9 ( 4.57x)
put_luma_h_10_128x128_c: 264228.9 ( 1.00x)
put_luma_h_10_128x128_neon: 52199.7 ( 5.06x)
put_luma_h_12_4x4_c: 262.1 ( 1.00x)
put_luma_h_12_8x8_c: 1051.3 ( 1.00x)
put_luma_h_12_8x8_neon: 230.9 ( 4.55x)
put_luma_h_12_16x16_c: 4124.4 ( 1.00x)
put_luma_h_12_16x16_neon: 848.0 ( 4.86x)
put_luma_h_12_32x32_c: 16446.9 ( 1.00x)
put_luma_h_12_32x32_neon: 3347.4 ( 4.91x)
put_luma_h_12_64x64_c: 66770.1 ( 1.00x)
put_luma_h_12_64x64_neon: 14360.2 ( 4.65x)
put_luma_h_12_128x128_c: 264419.5 ( 1.00x)
put_luma_h_12_128x128_neon: 52200.6 ( 5.07x)
M2 Air (with auto-vectorization feature):
put_luma_h_10_4x4_c: 0.3 ( 1.00x)
put_luma_h_10_8x8_c: 1.0 ( 1.00x)
put_luma_h_10_8x8_neon: 0.4 ( 2.58x)
put_luma_h_10_16x16_c: 3.0 ( 1.00x)
put_luma_h_10_16x16_neon: 1.5 ( 2.01x)
put_luma_h_10_32x32_c: 9.7 ( 1.00x)
put_luma_h_10_32x32_neon: 6.2 ( 1.57x)
put_luma_h_10_64x64_c: 36.6 ( 1.00x)
put_luma_h_10_64x64_neon: 23.9 ( 1.53x)
put_luma_h_10_128x128_c: 134.2 ( 1.00x)
put_luma_h_10_128x128_neon: 95.4 ( 1.41x)
put_luma_h_12_4x4_c: 0.3 ( 1.00x)
put_luma_h_12_8x8_c: 1.0 ( 1.00x)
put_luma_h_12_8x8_neon: 0.4 ( 2.57x)
put_luma_h_12_16x16_c: 3.0 ( 1.00x)
put_luma_h_12_16x16_neon: 1.5 ( 2.01x)
put_luma_h_12_32x32_c: 9.7 ( 1.00x)
put_luma_h_12_32x32_neon: 6.0 ( 1.63x)
put_luma_h_12_64x64_c: 36.5 ( 1.00x)
put_luma_h_12_64x64_neon: 23.9 ( 1.53x)
put_luma_h_12_128x128_c: 134.8 ( 1.00x)
put_luma_h_12_128x128_neon: 95.2 ( 1.42x)
---
libavcodec/aarch64/vvc/dsp_init.c | 22 ++++++++
libavcodec/aarch64/vvc/inter.S | 90 +++++++++++++++++++++++++++++++
2 files changed, 112 insertions(+)
diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c
index b7dc1d89f8..053d453fa7 100644
--- a/libavcodec/aarch64/vvc/dsp_init.c
+++ b/libavcodec/aarch64/vvc/dsp_init.c
@@ -30,6 +30,18 @@
#define BDOF_BLOCK_SIZE 16
#define BDOF_MIN_BLOCK_SIZE 4
+void ff_vvc_put_luma_h8_10_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
+void ff_vvc_put_luma_h16_10_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
+void ff_vvc_put_luma_h_x16_10_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
+void ff_vvc_put_luma_h8_12_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
+void ff_vvc_put_luma_h16_12_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
+void ff_vvc_put_luma_h_x16_12_neon(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride,
+ const int height, const int8_t *hf, const int8_t *vf, const int width);
void ff_alf_classify_sum_neon(int *sum0, int *sum1, int16_t *grad, uint32_t gshift, uint32_t steps);
@@ -245,6 +257,11 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
c->inter.dmvr[0][1] = ff_vvc_dmvr_h_10_neon;
c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_10_neon;
c->inter.apply_bdof = ff_vvc_apply_bdof_10_neon;
+ c->inter.put[0][2][0][1] = ff_vvc_put_luma_h8_10_neon;
+ c->inter.put[0][3][0][1] = ff_vvc_put_luma_h16_10_neon;
+ c->inter.put[0][4][0][1] =
+ c->inter.put[0][5][0][1] =
+ c->inter.put[0][6][0][1] = ff_vvc_put_luma_h_x16_10_neon;
c->alf.filter[LUMA] = alf_filter_luma_10_neon;
c->alf.filter[CHROMA] = alf_filter_chroma_10_neon;
@@ -256,6 +273,11 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
c->inter.dmvr[0][1] = ff_vvc_dmvr_h_12_neon;
c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_12_neon;
c->inter.apply_bdof = ff_vvc_apply_bdof_12_neon;
+ c->inter.put[0][2][0][1] = ff_vvc_put_luma_h8_12_neon;
+ c->inter.put[0][3][0][1] = ff_vvc_put_luma_h16_12_neon;
+ c->inter.put[0][4][0][1] =
+ c->inter.put[0][5][0][1] =
+ c->inter.put[0][6][0][1] = ff_vvc_put_luma_h_x16_12_neon;
c->alf.filter[LUMA] = alf_filter_luma_12_neon;
c->alf.filter[CHROMA] = alf_filter_chroma_12_neon;
diff --git a/libavcodec/aarch64/vvc/inter.S b/libavcodec/aarch64/vvc/inter.S
index a874edf889..be37df8e38 100644
--- a/libavcodec/aarch64/vvc/inter.S
+++ b/libavcodec/aarch64/vvc/inter.S
@@ -1713,3 +1713,93 @@ endfunc
#undef GRADIENT_V1_OFFSET
#undef VX_OFFSET
#undef VY_OFFSET
+
+#define VVC_MAX_PB_SIZE 128
+
+.macro put_luma_h_x8_16bit_ n, is_w_loop, shift
+ // dst .req x0
+ // _src .req x1
+ // _src_stride .req x2
+ // height .req w3
+ // hf .req x4
+ // vf .req x5
+ // width .req w6
+ mov x9, #(VVC_MAX_PB_SIZE * 2)
+ ld1 {v0.8b}, [x4]
+ sxtl v0.8h, v0.8b
+ mov w7, #0 // y loop: height
+1:
+ sub x11, x1, #6
+ add w7, w7, #1
+ ld1 {v20.8h}, [x11], #16
+ mov x10, x0
+ .if \is_w_loop == 1
+ mov w8, #(8*\n)
+2:
+ .endif
+ .rept \n
+ ld1 {v16.8h}, [x11], #16
+ ext v1.16b, v20.16b, v16.16b, #2
+ ext v2.16b, v20.16b, v16.16b, #4
+ ext v3.16b, v20.16b, v16.16b, #6
+ ext v4.16b, v20.16b, v16.16b, #8
+ ext v5.16b, v20.16b, v16.16b, #10
+ ext v6.16b, v20.16b, v16.16b, #12
+ ext v17.16b, v20.16b, v16.16b, #14
+ smull v21.4s, v20.4h, v0.h[0]
+ smull2 v22.4s, v20.8h, v0.h[0]
+ smlal v21.4s, v1.4h, v0.h[1]
+ smlal2 v22.4s, v1.8h, v0.h[1]
+ smlal v21.4s, v2.4h, v0.h[2]
+ smlal2 v22.4s, v2.8h, v0.h[2]
+ smlal v21.4s, v3.4h, v0.h[3]
+ smlal2 v22.4s, v3.8h, v0.h[3]
+ smlal v21.4s, v4.4h, v0.h[4]
+ smlal2 v22.4s, v4.8h, v0.h[4]
+ smlal v21.4s, v5.4h, v0.h[5]
+ smlal2 v22.4s, v5.8h, v0.h[5]
+ smlal v21.4s, v6.4h, v0.h[6]
+ smlal2 v22.4s, v6.8h, v0.h[6]
+ smlal v21.4s, v17.4h, v0.h[7]
+ smlal2 v22.4s, v17.8h, v0.h[7]
+ sqshrn v21.4h, v21.4s, #(\shift)
+ sqshrn v22.4h, v22.4s, #(\shift)
+ st1 {v21.4h, v22.4h}, [x10], #16
+ mov v20.16b, v16.16b
+ .endr
+ .if \is_w_loop == 1
+ cmp w8, w6
+ add w8, w8, #(8*\n)
+ b.lt 2b
+ .endif
+ cmp w7, w3
+ add x0, x0, x9
+ add x1, x1, x2
+ b.lt 1b
+ ret
+.endm
+
+function ff_vvc_put_luma_h8_10_neon, export=1
+ put_luma_h_x8_16bit_ 1, 0, 2
+endfunc
+
+function ff_vvc_put_luma_h8_12_neon, export=1
+ put_luma_h_x8_16bit_ 1, 0, 4
+endfunc
+
+function ff_vvc_put_luma_h16_10_neon, export=1
+ put_luma_h_x8_16bit_ 2, 0, 2
+endfunc
+
+function ff_vvc_put_luma_h16_12_neon, export=1
+ put_luma_h_x8_16bit_ 2, 0, 4
+endfunc
+
+function ff_vvc_put_luma_h_x16_10_neon, export=1
+ put_luma_h_x8_16bit_ 2, 1, 2
+endfunc
+
+function ff_vvc_put_luma_h_x16_12_neon, export=1
+ put_luma_h_x8_16bit_ 2, 1, 4
+endfunc
+
--
2.49.1
_______________________________________________
ffmpeg-devel mailing list -- ffmpeg-devel@ffmpeg.org
To unsubscribe send an email to ffmpeg-devel-leave@ffmpeg.org
reply other threads:[~2025-10-22 18:30 UTC|newest]
Thread overview: [no followups] expand[flat|nested] mbox.gz Atom feed
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=176115780645.25.2812638608647755938@7d278768979e \
--to=ffmpeg-devel@ffmpeg.org \
--cc=code@ffmpeg.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
This inbox may be cloned and mirrored by anyone:
git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git
# If you have public-inbox 1.1+ installed, you may
# initialize and index your mirror using the following commands:
public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
ffmpegdev@gitmailbox.com
public-inbox-index ffmpegdev
Example config snippet for mirrors.
AGPL code for this site: git clone https://public-inbox.org/public-inbox.git