[FFmpeg-devel] [PATCH] aarch64/h26x: optimize sao_band_filter
Zhao Zhili
quinkblack at foxmail.com
Tue Apr 15 13:01:42 EEST 2025
From: Zhao Zhili <zhilizhao at tencent.com>
int8_t[] is enough for offset_table of 8 bit streams.
On rpi5:
Before After
hevc_sao_band_8_8_c: 252.3 ( 1.00x) 252.3 ( 1.00x)
hevc_sao_band_8_8_neon: 95.8 ( 2.63x) 61.0 ( 4.14x)
hevc_sao_band_16_8_c: 875.2 ( 1.00x) 864.9 ( 1.00x)
hevc_sao_band_16_8_neon: 317.5 ( 2.76x) 150.0 ( 5.76x)
hevc_sao_band_32_8_c: 3853.5 ( 1.00x) 3871.6 ( 1.00x)
hevc_sao_band_32_8_neon: 1222.3 ( 3.15x) 550.6 ( 7.03x)
hevc_sao_band_48_8_c: 8203.6 ( 1.00x) 8182.6 ( 1.00x)
hevc_sao_band_48_8_neon: 2685.7 ( 3.05x) 1185.8 ( 6.90x)
hevc_sao_band_64_8_c: 14023.0 ( 1.00x) 14038.9 ( 1.00x)
hevc_sao_band_64_8_neon: 4783.2 ( 2.93x) 2078.4 ( 6.75x)
---
libavcodec/aarch64/h26x/dsp.h | 4 +
libavcodec/aarch64/h26x/sao_neon.S | 93 ++++++++++++++---------
libavcodec/aarch64/hevcdsp_init_aarch64.c | 4 +-
libavcodec/aarch64/vvc/dsp_init.c | 5 +-
4 files changed, 65 insertions(+), 41 deletions(-)
diff --git a/libavcodec/aarch64/h26x/dsp.h b/libavcodec/aarch64/h26x/dsp.h
index 0fefb4d70f..6ea6a8d36a 100644
--- a/libavcodec/aarch64/h26x/dsp.h
+++ b/libavcodec/aarch64/h26x/dsp.h
@@ -28,6 +28,10 @@ void ff_h26x_sao_band_filter_8x8_8_neon(uint8_t *_dst, const uint8_t *_src,
ptrdiff_t stride_dst, ptrdiff_t stride_src,
const int16_t *sao_offset_val, int sao_left_class,
int width, int height);
+void ff_h26x_sao_band_filter_16x16_8_neon(uint8_t *_dst, const uint8_t *_src,
+ ptrdiff_t stride_dst, ptrdiff_t stride_src,
+ const int16_t *sao_offset_val, int sao_left_class,
+ int width, int height);
void ff_hevc_sao_edge_filter_16x16_8_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride_dst,
const int16_t *sao_offset_val, int eo, int width, int height);
void ff_hevc_sao_edge_filter_8x8_8_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride_dst,
diff --git a/libavcodec/aarch64/h26x/sao_neon.S b/libavcodec/aarch64/h26x/sao_neon.S
index c43820135e..60c026fe95 100644
--- a/libavcodec/aarch64/h26x/sao_neon.S
+++ b/libavcodec/aarch64/h26x/sao_neon.S
@@ -35,48 +35,67 @@
// int16_t *sao_offset_val, int sao_left_class,
// int width, int height)
function ff_h26x_sao_band_filter_8x8_8_neon, export=1
- stp xzr, xzr, [sp, #-64]!
+ stp xzr, xzr, [sp, #-32]!
stp xzr, xzr, [sp, #16]
- stp xzr, xzr, [sp, #32]
- stp xzr, xzr, [sp, #48]
mov w8, #4
-0: ldrsh x9, [x4, x8, lsl #1] // sao_offset_val[k+1]
- subs w8, w8, #1
- add w10, w8, w5 // k + sao_left_class
+0:
+ ldrsh x9, [x4, x8, lsl #1] // sao_offset_val[k+1]
+ subs w8, w8, #1
+ add w10, w8, w5 // k + sao_left_class
and w10, w10, #0x1F
- strh w9, [sp, x10, lsl #1]
+ strb w9, [sp, x10]
bne 0b
- add w6, w6, #7
- bic w6, w6, #7
- ld1 {v16.16b-v19.16b}, [sp], #64
- sub x2, x2, x6
- sub x3, x3, x6
- movi v20.8h, #1
-1: mov w8, w6 // beginning of line
-2: // Simple layout for accessing 16bit values
- // with 8bit LUT.
- //
- // 00 01 02 03 04 05 06 07
- // +----------------------------------->
- // |xDE#xAD|xCA#xFE|xBE#xEF|xFE#xED|....
- // +----------------------------------->
- // i-0 i-1 i-2 i-3
- ld1 {v2.8b}, [x1], #8 // dst[x] = av_clip_pixel(src[x] + offset_table[src[x] >> shift]);
- subs w8, w8, #8
- uxtl v0.8h, v2.8b // load src[x]
- ushr v2.8h, v0.8h, #3 // >> BIT_DEPTH - 3
- shl v1.8h, v2.8h, #1 // low (x2, accessing short)
- add v3.8h, v1.8h, v20.8h // +1 access upper short
- sli v1.8h, v3.8h, #8 // shift insert index to upper byte
- tbx v2.16b, {v16.16b-v19.16b}, v1.16b // table
- add v1.8h, v0.8h, v2.8h // src[x] + table
- sqxtun v4.8b, v1.8h // clip + narrow
- st1 {v4.8b}, [x0], #8 // store
- // done 8 pixels
+ ldp q16, q17, [sp], #32
+1:
+ ld1 {v2.8b}, [x1], x3
+ subs w7, w7, #1
+ uxtl v0.8h, v2.8b
+ ushr v3.8b, v2.8b, #3 // >> BIT_DEPTH - 3
+ tbx v3.8b, {v16.16b-v17.16b}, v3.8b
+ sxtl v2.8h, v3.8b
+ add v0.8h, v0.8h, v2.8h // src[x] + table
+ sqxtun v0.8b, v0.8h // clip + narrow
+ st1 {v0.8b}, [x0], x2
+ bne 1b
+ ret
+endfunc
+
+function ff_h26x_sao_band_filter_16x16_8_neon, export=1
+ stp xzr, xzr, [sp, #-32]!
+ stp xzr, xzr, [sp, #16]
+ mov w8, #4
+0:
+ ldrsh x9, [x4, x8, lsl #1] // sao_offset_val[k+1]
+ subs w8, w8, #1
+ add w10, w8, w5 // k + sao_left_class
+ and w10, w10, #0x1F
+ strb w9, [sp, x10]
+ bne 0b
+ add w6, w6, #15
+ bic w6, w6, #15
+ ldp q16, q17, [sp], #32
+ sub x2, x2, x6
+ sub x3, x3, x6
+1:
+ mov w8, w6 // beginning of line
+2:
+ ldr q2, [x1], #16
+ subs w8, w8, #16
+ uxtl v0.8h, v2.8b
+ uxtl2 v1.8h, v2.16b
+ ushr v3.16b, v2.16b, #3 // >> BIT_DEPTH - 3
+ tbx v3.16b, {v16.16b-v17.16b}, v3.16b
+ sxtl v2.8h, v3.8b
+ sxtl2 v3.8h, v3.16b
+ add v0.8h, v0.8h, v2.8h // src[x] + table
+ add v1.8h, v1.8h, v3.8h
+ sqxtun v0.8b, v0.8h // clip + narrow
+ sqxtun2 v0.16b, v1.8h
+ str q0, [x0], #16
bne 2b
- subs w7, w7, #1 // finished line, prep. new
- add x0, x0, x2 // dst += stride_dst
- add x1, x1, x3 // src += stride_src
+ subs w7, w7, #1
+ add x0, x0, x2 // dst += stride_dst
+ add x1, x1, x3 // src += stride_src
bne 1b
ret
endfunc
diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c b/libavcodec/aarch64/hevcdsp_init_aarch64.c
index 5dd470baaa..0b159d1886 100644
--- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
+++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
@@ -157,11 +157,11 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_neon;
c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_neon;
c->transform_4x4_luma = ff_hevc_transform_luma_4x4_neon_8;
- c->sao_band_filter[0] =
+ c->sao_band_filter[0] = ff_h26x_sao_band_filter_8x8_8_neon;
c->sao_band_filter[1] =
c->sao_band_filter[2] =
c->sao_band_filter[3] =
- c->sao_band_filter[4] = ff_h26x_sao_band_filter_8x8_8_neon;
+ c->sao_band_filter[4] = ff_h26x_sao_band_filter_16x16_8_neon;
c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_8x8_8_neon;
c->sao_edge_filter[1] =
c->sao_edge_filter[2] =
diff --git a/libavcodec/aarch64/vvc/dsp_init.c b/libavcodec/aarch64/vvc/dsp_init.c
index 2c99ba206b..9a171234f6 100644
--- a/libavcodec/aarch64/vvc/dsp_init.c
+++ b/libavcodec/aarch64/vvc/dsp_init.c
@@ -197,8 +197,9 @@ void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_8_neon;
c->inter.apply_bdof = apply_bdof_8;
- for (int i = 0; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++)
- c->sao.band_filter[i] = ff_h26x_sao_band_filter_8x8_8_neon;
+ c->sao.band_filter[0] = ff_h26x_sao_band_filter_8x8_8_neon;
+ for (int i = 1; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++)
+ c->sao.band_filter[i] = ff_h26x_sao_band_filter_16x16_8_neon;
c->sao.edge_filter[0] = ff_vvc_sao_edge_filter_8x8_8_neon;
for (int i = 1; i < FF_ARRAY_ELEMS(c->sao.edge_filter); i++)
c->sao.edge_filter[i] = ff_vvc_sao_edge_filter_16x16_8_neon;
--
2.46.0
More information about the ffmpeg-devel
mailing list