[FFmpeg-devel] [PATCH 3/4] lavc/aarch64: add hevc sao band 8x8 tiling

[J. Dekker]

--bench on AWS Graviton:

hevc_sao_band_8x8_8_c: 317.5
hevc_sao_band_8x8_8_neon: 97.5
hevc_sao_band_16x16_8_c: 1115.0
hevc_sao_band_16x16_8_neon: 322.7
hevc_sao_band_32x32_8_c: 4599.2
hevc_sao_band_32x32_8_neon: 1246.2
hevc_sao_band_48x48_8_c: 10021.7
hevc_sao_band_48x48_8_neon: 2740.5
hevc_sao_band_64x64_8_c: 17635.0
hevc_sao_band_64x64_8_neon: 4875.7

Signed-off-by: J. Dekker <jdek at itanimul.li>
---
 libavcodec/aarch64/hevcdsp_init_aarch64.c | 6 +++++-
 libavcodec/aarch64/hevcdsp_sao_neon.S     | 9 ++++++---
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/libavcodec/aarch64/hevcdsp_init_aarch64.c b/libavcodec/aarch64/hevcdsp_init_aarch64.c
index b93cec9e44..2002530266 100644
--- a/libavcodec/aarch64/hevcdsp_init_aarch64.c
+++ b/libavcodec/aarch64/hevcdsp_init_aarch64.c
@@ -77,7 +77,11 @@ av_cold void ff_hevc_dsp_init_aarch64(HEVCDSPContext *c, const int bit_depth)
         c->idct_dc[1]                  = ff_hevc_idct_8x8_dc_8_neon;
         c->idct_dc[2]                  = ff_hevc_idct_16x16_dc_8_neon;
         c->idct_dc[3]                  = ff_hevc_idct_32x32_dc_8_neon;
-        c->sao_band_filter[0]          = ff_hevc_sao_band_filter_8x8_8_neon;
+        c->sao_band_filter[0]          =
+        c->sao_band_filter[1]          =
+        c->sao_band_filter[2]          =
+        c->sao_band_filter[3]          =
+        c->sao_band_filter[4]          = ff_hevc_sao_band_filter_8x8_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/hevcdsp_sao_neon.S b/libavcodec/aarch64/hevcdsp_sao_neon.S
index c4b931aab7..263747149f 100644
--- a/libavcodec/aarch64/hevcdsp_sao_neon.S
+++ b/libavcodec/aarch64/hevcdsp_sao_neon.S
@@ -35,6 +35,7 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
         stp            xzr, xzr, [sp, #32]
         stp            xzr, xzr, [sp, #48]
         mov             w8,  #4
+        sxtw            x6,  w6
 0:
         ldrsh           x9, [x4,  x8, lsl #1] // x9 = sao_offset_val[k+1]
         subs            w8,  w8,  #1
@@ -44,8 +45,10 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
         bne             0b
         ld1            {v16.16b-v19.16b}, [sp], #64
         movi           v20.8h,   #1
+        sub             x2,  x2, x6 // stride_dst - width
+        sub             x3,  x3, x6 // stride_src - width
 1:      // beginning of line
-        mov             w8,  w6
+        mov             x8,  x6
 2:
         // Simple layout for accessing 16bit values
         // with 8bit LUT.
@@ -56,7 +59,7 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
         // +----------------------------------->
         //    i-0     i-1     i-2     i-3
         // dst[x] = av_clip_pixel(src[x] + offset_table[src[x] >> shift]);
-        ld1            {v2.8b}, [x1]
+        ld1            {v2.8b}, [x1], #8
         // load src[x]
         uxtl            v0.8h,  v2.8b
         // >> shift
@@ -74,7 +77,7 @@ function ff_hevc_sao_band_filter_8x8_8_neon, export=1
         // clip + narrow
         sqxtun          v4.8b,  v1.8h
         // store
-        st1            {v4.8b}, [x0]
+        st1            {v4.8b}, [x0], #8
         // done 8 pixels
         subs            w8, w8,  #8
         bne             2b
-- 
2.30.1 (Apple Git-130)