[FFmpeg-devel] [PATCH 3/3] lavc/aarch64: Add neon implementation for pix_median_abs8

[Hubert Mazur]

Provide optimized implementation for pix_median_abs8 function.

Performance comparison tests are shown below.
- median_sad_1_c: 277.0
- median_sad_1_neon: 82.0

Benchmarks and tests run with checkasm tool on AWS Graviton 3.

Signed-off-by: Hubert Mazur <hum at semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  3 ++
 libavcodec/aarch64/me_cmp_neon.S         | 62 ++++++++++++++++++++++++
 2 files changed, 65 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index d3fa047a86..e143f0816e 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -57,6 +57,8 @@ int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                         ptrdiff_t stride, int h);
 int pix_median_abs16_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                           ptrdiff_t stride, int h);
+int pix_median_abs8_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
+                         ptrdiff_t stride, int h);
 
 av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 {
@@ -85,6 +87,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->nsse[0] = nsse16_neon_wrapper;
 
         c->median_sad[0] = pix_median_abs16_neon;
+        c->median_sad[1] = pix_median_abs8_neon;
     }
 }
 
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index 1587f2ea43..11af4849f9 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -1089,3 +1089,65 @@ function vsad_intra8_neon, export=1
 
         ret
 endfunc
+
+function pix_median_abs8_neon, export=1
+        // x0           unused
+        // x1           uint8_t *pix1
+        // x2           uint8_t *pix2
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        ld1             {v2.8b}, [x1], x3
+        ld1             {v3.8b}, [x2], x3
+        movi            v31.8h, #0
+        ext             v0.8b, v2.8b, v2.8b, #1
+        ext             v1.8b, v3.8b, v3.8b, #1
+        usubl           v28.8h, v2.8b, v3.8b
+        usubl           v26.8h, v0.8b, v1.8b
+        sub             w4, w4, #1                              // we need to make h-1 iterations
+        saba            v31.8h, v26.8h, v28.8h
+        mov             h18, v28.h[0]
+        cmp             w4, #1
+        sqabs           h18, h18
+        movi            v0.8h, #0
+
+        b.lt            2f
+1:
+        ld1             {v6.8b}, [x1], x3                       // pix1 vector for V(j-1)
+        ld1             {v7.8b}, [x2], x3                       // pix2 vector for V(j-1)
+        subs            w4, w4, #1
+        ext             v4.8b, v6.8b, v6.8b, #1                 // pix1 vector for V(j)
+        ext             v5.8b, v7.8b, v7.8b, #1                 // pix2 vector for V(j)
+
+        // protected registers: v30, v29, v28, v27, v26, v25, v24, v23
+        // scratch registers: v22, v21, v20, v19, v17
+
+        // To find median of three values, calculate sum of them
+        // and subtract max and min value from it.
+        usubl           v30.8h, v6.8b, v7.8b                    // V(j-1)
+        usubl           v24.8h, v4.8b, v5.8b                    // V(j)
+        saba            v0.8h, v30.8h, v28.8h
+        add             v22.8h, v26.8h, v30.8h
+        smin            v20.8h, v26.8h, v30.8h
+        smax            v19.8h, v26.8h, v30.8h
+        sub             v22.8h, v22.8h, v28.8h
+        smin            v17.8h, v19.8h, v22.8h
+        mov             v28.16b, v30.16b
+        smax            v20.8h, v20.8h, v17.8h                  // median values lower half
+        smax            v19.8h, v25.8h, v29.8h
+        saba            v31.8h, v24.8h, v20.8h
+        mov             v26.16b, v24.16b
+        smax            v17.8h, v22.8h, v19.8h                  // median values upper half
+
+        b.ne            1b
+2:
+        mov             h17, v0.h[0]
+        ins             v31.h[7], wzr
+        add             d18, d18, d17
+        uaddlv          s17, v31.8h
+        add             d18, d18, d17
+        fmov            w0, s18
+
+        ret
+
+endfunc
-- 
2.34.1