[FFmpeg-devel] [PATCH 0/5] Provide optimized neon implementation

[FFmpeg-devel] [PATCH 0/5] Provide optimized neon implementation

[Hubert Mazur]

Fix minor issues in the patches.
Regarding vsse16 I didn't change saba & umlal to sub & smlal.
It doesn't affect the performance, so left it as it was.
The majority of changes refer to nsse16:
 - fixed indentation (thanks for pointing out),
 - applied the patch from Martin which fixes the balance
   within instructions,
 - interleaved instructions - apparently this helped a little
   to achieve better benchmarks.

I have also updated the benchmark results for each function -
not a huge performance improvement, but worth the effort.
For nsse and vsse are shown below (these are the biggest changes).
 - vsse16 asm from 64.7 to 59.2,
 - nsse16 asm from 120.0 to 116.5.

Hubert Mazur (5):
  lavc/aarch64: Add neon implementation for vsad16
  lavc/aarch64: Add neon implementation of vsse16
  lavc/aarch64: Add neon implementation for vsad_intra16
  lavc/aarch64: Add neon implementation for vsse_intra16
  lavc/aarch64: Provide neon implementation of nsse16

 libavcodec/aarch64/me_cmp_init_aarch64.c |  30 ++
 libavcodec/aarch64/me_cmp_neon.S         | 385 +++++++++++++++++++++++
 2 files changed, 415 insertions(+)

-- 
2.34.1

_______________________________________________
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".

[FFmpeg-devel] [PATCH 1/5] lavc/aarch64: Add neon implementation for vsad16

[Hubert Mazur]

Provide optimized implementation of vsad16 function for arm64.

Performance comparison tests are shown below.
- vsad_0_c: 285.2
- vsad_0_neon: 39.5

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

Co-authored-by: Martin Storsjö <martin@martin.st>
Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  5 ++
 libavcodec/aarch64/me_cmp_neon.S         | 65 ++++++++++++++++++++++++
 2 files changed, 70 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index fb7c3f5059..ddc5d05611 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -41,6 +41,9 @@ int sse8_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
 int sse4_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
               ptrdiff_t stride, int h);
 
+int vsad16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
+                ptrdiff_t stride, int h);
+
 av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -57,5 +60,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->sse[0] = sse16_neon;
         c->sse[1] = sse8_neon;
         c->sse[2] = sse4_neon;
+
+        c->vsad[0] = vsad16_neon;
     }
 }
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index 4198985c6c..1d0b166d69 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -584,3 +584,68 @@ function sse4_neon, export=1
 
         ret
 endfunc
+
+function vsad16_neon, export=1
+        // x0           unused
+        // x1           uint8_t *pix1
+        // x2           uint8_t *pix2
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        ld1             {v0.16b}, [x1], x3              // Load pix1[0], first iteration
+        ld1             {v1.16b}, [x2], x3              // Load pix2[0], first iteration
+
+        sub             w4, w4, #1                      // we need to make h-1 iterations
+        movi            v16.8h, #0
+
+        cmp             w4, #3                          // check if we can make 3 iterations at once
+        usubl           v31.8h, v0.8b, v1.8b            // Signed difference pix1[0] - pix2[0], first iteration
+        usubl2          v30.8h, v0.16b, v1.16b          // Signed difference pix1[0] - pix2[0], first iteration
+
+        b.lt            2f
+
+1:
+        // abs(pix1[0] - pix2[0] - pix1[0 + stride] + pix2[0 + stride])
+        ld1             {v0.16b}, [x1], x3              // Load pix1[0 + stride], first iteration
+        ld1             {v1.16b}, [x2], x3              // Load pix2[0 + stride], first iteration
+        ld1             {v2.16b}, [x1], x3              // Load pix1[0 + stride], second iteration
+        ld1             {v3.16b}, [x2], x3              // Load pix2[0 + stride], second iteration
+        usubl           v29.8h, v0.8b,  v1.8b           // Signed difference pix1[0 + stride] - pix2[0 + stride], first iteration
+        usubl2          v28.8h, v0.16b, v1.16b          // Signed difference pix1[0 + stride] - pix2[0 + stride], first iteration
+        ld1             {v4.16b}, [x1], x3              // Load pix1[0 + stride], third iteration
+        ld1             {v5.16b}, [x2], x3              // Load pix2[0 + stride], third iteration
+        usubl           v27.8h, v2.8b,  v3.8b           // Signed difference pix1[0 + stride] - pix2[0 + stride], second iteration
+        saba            v16.8h, v31.8h, v29.8h          // Signed absolute difference and accumulate the result. first iteration
+        usubl2          v26.8h, v2.16b, v3.16b          // Signed difference pix1[0 + stride] - pix2[0 + stride], second iteration
+        saba            v16.8h, v30.8h, v28.8h          // Signed absolute difference and accumulate the result. first iteration
+        usubl           v25.8h, v4.8b,  v5.8b           // Signed difference pix1[0 + stride] - pix2[0 + stride], third iteration
+        usubl2          v24.8h, v4.16b, v5.16b          // Signed difference pix1[0 + stride] - pix2[0 + stride], third iteration
+        saba            v16.8h, v29.8h, v27.8h          // Signed absolute difference and accumulate the result. second iteration
+        mov             v31.16b, v25.16b
+        saba            v16.8h, v28.8h, v26.8h          // Signed absolute difference and accumulate the result. second iteration
+        sub             w4, w4, #3                      // h -= 3
+        mov             v30.16b, v24.16b
+        saba            v16.8h, v27.8h, v25.8h          // Signed absolute difference and accumulate the result. third iteration
+        cmp             w4, #3
+        saba            v16.8h, v26.8h, v24.8h          // Signed absolute difference and accumulate the result. third iteration
+
+        b.ge            1b
+        cbz             w4, 3f
+2:
+        ld1             {v0.16b}, [x1], x3
+        ld1             {v1.16b}, [x2], x3
+        subs            w4, w4, #1
+        usubl           v29.8h,  v0.8b,   v1.8b
+        usubl2          v28.8h,  v0.16b,  v1.16b
+        saba            v16.8h,  v31.8h,  v29.8h
+        mov             v31.16b, v29.16b
+        saba            v16.8h,  v30.8h,  v28.8h
+        mov             v30.16b, v28.16b
+
+        b.ne            2b
+3:
+        uaddlv          s17, v16.8h
+        fmov            w0, s17
+
+        ret
+endfunc
-- 
2.34.1

_______________________________________________
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".

[FFmpeg-devel] [PATCH 2/5] lavc/aarch64: Add neon implementation of vsse16

[Hubert Mazur]

Provide optimized implementation of vsse16 for arm64.

Performance comparison tests are shown below.
- vsse_0_c: 257.7
- vsse_0_neon: 59.2

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

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  4 ++
 libavcodec/aarch64/me_cmp_neon.S         | 87 ++++++++++++++++++++++++
 2 files changed, 91 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index ddc5d05611..7b81e48d16 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -43,6 +43,8 @@ int sse4_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
 
 int vsad16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                 ptrdiff_t stride, int h);
+int vsse16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
+                ptrdiff_t stride, int h);
 
 av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 {
@@ -62,5 +64,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->sse[2] = sse4_neon;
 
         c->vsad[0] = vsad16_neon;
+
+        c->vsse[0] = vsse16_neon;
     }
 }
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index 1d0b166d69..b3f376aa60 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -649,3 +649,90 @@ function vsad16_neon, export=1
 
         ret
 endfunc
+
+function vsse16_neon, export=1
+        // x0           unused
+        // x1           uint8_t *pix1
+        // x2           uint8_t *pix2
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        ld1             {v0.16b}, [x1], x3              // Load pix1[0], first iteration
+        ld1             {v1.16b}, [x2], x3              // Load pix2[0], first iteration
+
+        sub             w4, w4, #1                      // we need to make h-1 iterations
+        movi            v16.4s, #0
+        movi            v17.4s, #0
+
+        cmp             w4, #3                          // check if we can make 3 iterations at once
+        usubl           v31.8h, v0.8b, v1.8b            // Signed difference of pix1[0] - pix2[0], first iteration
+        usubl2          v30.8h, v0.16b, v1.16b          // Signed difference of pix1[0] - pix2[0], first iteration
+        b.le            2f
+
+
+1:
+        // x = abs(pix1[0] - pix2[0] - pix1[0 + stride] + pix2[0 + stride])
+        // res = (x) * (x)
+        ld1             {v0.16b}, [x1], x3              // Load pix1[0 + stride], first iteration
+        ld1             {v1.16b}, [x2], x3              // Load pix2[0 + stride], first iteration
+        ld1             {v2.16b}, [x1], x3              // Load pix1[0 + stride], second iteration
+        ld1             {v3.16b}, [x2], x3              // Load pix2[0 + stride], second iteration
+        usubl           v29.8h, v0.8b, v1.8b
+        usubl2          v28.8h, v0.16b, v1.16b
+        ld1             {v4.16b}, [x1], x3              // Load pix1[0 + stride], third iteration
+        ld1             {v5.16b}, [x2], x3              // Load pix1[0 + stride], third iteration
+        sabd            v31.8h, v31.8h, v29.8h
+        sabd            v30.8h, v30.8h, v28.8h
+        usubl           v27.8h, v2.8b, v3.8b
+        usubl2          v26.8h, v2.16b, v3.16b
+        usubl           v25.8h, v4.8b, v5.8b
+        usubl2          v24.8h, v4.16b, v5.16b
+        sabd            v29.8h, v29.8h, v27.8h
+        sabd            v27.8h, v27.8h, v25.8h
+        umlal           v16.4s, v31.4h, v31.4h
+        umlal2          v17.4s, v31.8h, v31.8h
+        sabd            v28.8h, v28.8h, v26.8h
+        sabd            v26.8h, v26.8h, v24.8h
+        umlal           v16.4s, v30.4h, v30.4h
+        umlal2          v17.4s, v30.8h, v30.8h
+        mov             v31.16b, v25.16b
+        umlal           v16.4s, v29.4h, v29.4h
+        umlal2          v17.4s, v29.8h, v29.8h
+        mov             v30.16b, v24.16b
+        umlal           v16.4s, v28.4h, v28.4h
+        umlal2          v17.4s, v28.8h, v28.8h
+        sub             w4, w4, #3
+        umlal           v16.4s, v27.4h, v27.4h
+        umlal2          v17.4s, v27.8h, v27.8h
+        cmp             w4, #3
+        umlal           v16.4s, v26.4h, v26.4h
+        umlal2          v17.4s, v26.8h, v26.8h
+
+        b.ge            1b
+
+        cbz             w4, 3f
+
+// iterate by once
+2:
+        ld1             {v0.16b}, [x1], x3
+        ld1             {v1.16b}, [x2], x3
+        subs            w4, w4, #1
+        usubl           v29.8h, v0.8b, v1.8b
+        usubl2          v28.8h, v0.16b, v1.16b
+        sabd            v31.8h, v31.8h, v29.8h
+        sabd            v30.8h, v30.8h, v28.8h
+        umlal           v16.4s, v31.4h, v31.4h
+        umlal2          v17.4s, v31.8h, v31.8h
+        mov             v31.16b, v29.16b
+        umlal           v16.4s, v30.4h, v30.4h
+        umlal2          v17.4s, v30.8h, v30.8h
+        mov             v30.16b, v28.16b
+        b.ne            2b
+
+3:
+        add             v16.4s, v16.4s, v17.4s
+        uaddlv          d17, v16.4s
+        fmov            w0, s17
+
+        ret
+endfunc
-- 
2.34.1

_______________________________________________
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".

[FFmpeg-devel] [PATCH 3/5] lavc/aarch64: Add neon implementation for vsad_intra16

[Hubert Mazur]

Provide optimized implementation for vsad_intra16 function for arm64.

Performance comparison tests are shown below.
- vsad_4_c: 177.5
- vsad_4_neon: 23.5

Benchmarks and tests are run with checkasm tool on AWS Gravtion 3.

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  3 ++
 libavcodec/aarch64/me_cmp_neon.S         | 48 ++++++++++++++++++++++++
 2 files changed, 51 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index 7b81e48d16..af83f7ed1e 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -43,6 +43,8 @@ int sse4_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
 
 int vsad16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                 ptrdiff_t stride, int h);
+int vsad_intra16_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
+                      ptrdiff_t stride, int h) ;
 int vsse16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                 ptrdiff_t stride, int h);
 
@@ -64,6 +66,7 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->sse[2] = sse4_neon;
 
         c->vsad[0] = vsad16_neon;
+        c->vsad[4] = vsad_intra16_neon;
 
         c->vsse[0] = vsse16_neon;
     }
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index b3f376aa60..ce198ea227 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -736,3 +736,51 @@ function vsse16_neon, export=1
 
         ret
 endfunc
+
+function vsad_intra16_neon, export=1
+        // x0           unused
+        // x1           uint8_t *pix1
+        // x2           uint8_t *dummy
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        ld1             {v0.16b}, [x1], x3
+        sub             w4, w4, #1 // we need to make h-1 iterations
+        cmp             w4, #3
+        movi            v16.8h, #0
+        b.lt            2f
+
+// make 4 iterations at once
+1:
+        // v = abs( pix1[0] - pix1[0 + stride] )
+        // score = sum(v)
+        ld1             {v1.16b}, [x1], x3
+        ld1             {v2.16b}, [x1], x3
+        uabal           v16.8h, v0.8b, v1.8b
+        ld1             {v3.16b}, [x1], x3
+        uabal2          v16.8h, v0.16b, v1.16b
+        sub             w4, w4, #3
+        uabal           v16.8h, v1.8b, v2.8b
+        cmp             w4, #3
+        uabal2          v16.8h, v1.16b, v2.16b
+        mov             v0.16b, v3.16b
+        uabal           v16.8h, v2.8b, v3.8b
+        uabal2          v16.8h, v2.16b, v3.16b
+        b.ge            1b
+        cbz             w4, 3f
+
+// iterate by one
+2:
+        ld1             {v1.16b}, [x1], x3
+        subs            w4, w4, #1
+        uabal           v16.8h, v0.8b, v1.8b
+        uabal2          v16.8h, v0.16b, v1.16b
+        mov             v0.16b, v1.16b
+        cbnz            w4, 2b
+
+3:
+        uaddlv          s17, v16.8h
+        fmov            w0, s17
+
+        ret
+endfunc
-- 
2.34.1

_______________________________________________
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".

[FFmpeg-devel] [PATCH 4/5] lavc/aarch64: Add neon implementation for vsse_intra16

[Hubert Mazur]

Provide optimized implementation for vsse_intra16 for arm64.

Performance tests are shown below.
- vsse_4_c: 155.2
- vsse_4_neon: 36.2

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

Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  3 ++
 libavcodec/aarch64/me_cmp_neon.S         | 63 ++++++++++++++++++++++++
 2 files changed, 66 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index af83f7ed1e..8c295d5457 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -47,6 +47,8 @@ int vsad_intra16_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
                       ptrdiff_t stride, int h) ;
 int vsse16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                 ptrdiff_t stride, int h);
+int vsse_intra16_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
+                      ptrdiff_t stride, int h);
 
 av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 {
@@ -69,5 +71,6 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->vsad[4] = vsad_intra16_neon;
 
         c->vsse[0] = vsse16_neon;
+        c->vsse[4] = vsse_intra16_neon;
     }
 }
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index ce198ea227..cf2b8da425 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -784,3 +784,66 @@ function vsad_intra16_neon, export=1
 
         ret
 endfunc
+
+function vsse_intra16_neon, export=1
+        // x0           unused
+        // x1           uint8_t *pix1
+        // x2           uint8_t *dummy
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        ld1             {v0.16b}, [x1], x3
+        movi            v16.4s, #0
+        movi            v17.4s, #0
+
+        sub             w4, w4, #1 // we need to make h-1 iterations
+        cmp             w4, #3
+        b.lt            2f
+
+1:
+        // v = abs( pix1[0] - pix1[0 + stride] )
+        // score = sum( v * v )
+        ld1             {v1.16b}, [x1], x3
+        ld1             {v2.16b}, [x1], x3
+        uabd            v30.16b, v0.16b, v1.16b
+        ld1             {v3.16b}, [x1], x3
+        umull           v29.8h, v30.8b, v30.8b
+        umull2          v28.8h, v30.16b, v30.16b
+        uabd            v27.16b, v1.16b, v2.16b
+        uadalp          v16.4s, v29.8h
+        umull           v26.8h, v27.8b, v27.8b
+        umull2          v27.8h, v27.16b, v27.16b
+        uadalp          v17.4s, v28.8h
+        uabd            v25.16b, v2.16b, v3.16b
+        uadalp          v16.4s, v26.8h
+        umull           v24.8h, v25.8b, v25.8b
+        umull2          v25.8h, v25.16b, v25.16b
+        uadalp          v17.4s, v27.8h
+        sub             w4, w4, #3
+        uadalp          v16.4s, v24.8h
+        cmp             w4, #3
+        uadalp          v17.4s, v25.8h
+        mov             v0.16b, v3.16b
+
+        b.ge            1b
+        cbz             w4, 3f
+
+// iterate by one
+2:
+        ld1             {v1.16b}, [x1], x3
+        subs            w4, w4, #1
+        uabd            v30.16b, v0.16b, v1.16b
+        mov             v0.16b, v1.16b
+        umull           v29.8h, v30.8b, v30.8b
+        umull2          v30.8h, v30.16b, v30.16b
+        uadalp          v16.4s, v29.8h
+        uadalp          v17.4s, v30.8h
+        cbnz            w4, 2b
+
+3:
+        add             v16.4s, v16.4s, v17.4S
+        uaddlv          d17, v16.4s
+        fmov            w0, s17
+
+        ret
+endfunc
-- 
2.34.1

_______________________________________________
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".

[FFmpeg-devel] [PATCH 5/5] lavc/aarch64: Provide neon implementation of nsse16

[Hubert Mazur]

Add vectorized implementation of nsse16 function.

Performance comparison tests are shown below.
- nsse_0_c: 682.2
- nsse_0_neon: 116.5

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

Co-authored-by: Martin Storsjö <martin@martin.st>
Signed-off-by: Hubert Mazur <hum@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c |  15 +++
 libavcodec/aarch64/me_cmp_neon.S         | 122 +++++++++++++++++++++++
 2 files changed, 137 insertions(+)

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index 8c295d5457..ade3e9a4c1 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -49,6 +49,10 @@ int vsse16_neon(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
                 ptrdiff_t stride, int h);
 int vsse_intra16_neon(MpegEncContext *c, const uint8_t *s, const uint8_t *dummy,
                       ptrdiff_t stride, int h);
+int nsse16_neon(int multiplier, const uint8_t *s, const uint8_t *s2,
+                ptrdiff_t stride, int h);
+int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
+                        ptrdiff_t stride, int h);
 
 av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 {
@@ -72,5 +76,16 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
 
         c->vsse[0] = vsse16_neon;
         c->vsse[4] = vsse_intra16_neon;
+
+        c->nsse[0] = nsse16_neon_wrapper;
     }
 }
+
+int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
+                        ptrdiff_t stride, int h)
+{
+    if (c)
+        return nsse16_neon(c->avctx->nsse_weight, s1, s2, stride, h);
+    else
+        return nsse16_neon(8, s1, s2, stride, h);
+}
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index cf2b8da425..f8998749a5 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -847,3 +847,125 @@ function vsse_intra16_neon, export=1
 
         ret
 endfunc
+
+function nsse16_neon, export=1
+        // x0           multiplier
+        // x1           uint8_t *pix1
+        // x2           uint8_t *pix2
+        // x3           ptrdiff_t stride
+        // w4           int h
+
+        str             x0, [sp, #-0x40]!
+        stp             x1, x2, [sp, #0x10]
+        stp             x3, x4, [sp, #0x20]
+        str             x30, [sp, #0x30]
+        bl              X(sse16_neon)
+        ldr             x30, [sp, #0x30]
+        mov             w9, w0                                  // here we store score1
+        ldr             x5, [sp]
+        ldp             x1, x2, [sp, #0x10]
+        ldp             x3, x4, [sp, #0x20]
+        add             sp, sp, #0x40
+
+        movi            v16.8h, #0
+        movi            v17.8h, #0
+        movi            v18.8h, #0
+        movi            v19.8h, #0
+
+        ld1             {v0.16b}, [x1], x3
+        subs            w4, w4, #1                              // we need to make h-1 iterations
+        ld1             {v2.16b}, [x2], x3
+        ext             v1.16b, v0.16b, v0.16b, #1              // x1 + 1
+        cmp             w4, #2
+        ext             v3.16b, v2.16b, v2.16b, #1              // x2 + 1
+
+        b.lt            2f
+
+// make 2 iterations at once
+1:
+        ld1             {v4.16b}, [x1], x3
+        ld1             {v6.16b}, [x2], x3
+        ld1             {v20.16b}, [x1], x3
+        ext             v5.16b, v4.16b, v4.16b, #1              // x1 + stride + 1
+        usubl           v31.8h, v0.8b, v4.8b
+        usubl2          v30.8h, v0.16b, v4.16b
+        ld1             {v22.16b}, [x2], x3
+        usubl           v29.8h, v1.8b, v5.8b
+        usubl2          v28.8h, v1.16b, v5.16b
+        ext             v7.16b, v6.16b, v6.16b, #1              // x2 + stride + 1
+        saba            v16.8h, v31.8h, v29.8h
+        ext             v21.16b, v20.16b, v20.16b, #1
+        saba            v17.8h, v30.8h, v28.8h
+        usubl           v27.8h, v2.8b, v6.8b
+        usubl2          v26.8h, v2.16b, v6.16b
+        ext             v23.16b, v22.16b, v22.16b, #1
+        usubl           v25.8h, v3.8b, v7.8b
+        usubl2          v24.8h, v3.16b, v7.16b
+        saba            v18.8h, v27.8h, v25.8h
+        saba            v19.8h, v26.8h, v24.8h
+
+        usubl           v31.8h, v4.8b, v20.8b
+        usubl2          v30.8h, v4.16b, v20.16b
+        usubl           v29.8h, v5.8b, v21.8b
+        usubl2          v28.8h, v5.16b, v21.16b
+        saba            v16.8h, v31.8h, v29.8h
+        saba            v17.8h, v30.8h, v28.8h
+        usubl           v27.8h, v6.8b, v22.8b
+        usubl2          v26.8h, v6.16b, v22.16b
+        usubl           v25.8h, v7.8b, v23.8b
+        usubl2          v24.8h, v7.16b, v23.16b
+        saba            v18.8h, v27.8h, v25.8h
+        saba            v19.8h, v26.8h, v24.8h
+        sub             w4, w4, #2
+
+        mov             v0.16b, v20.16b
+        mov             v1.16b, v21.16b
+        cmp             w4, #2
+        mov             v2.16b, v22.16b
+        mov             v3.16b, v23.16b
+
+        b.ge            1b
+        cbz             w4, 3f
+
+// iterate by one
+2:
+        ld1             {v4.16b}, [x1], x3
+        subs            w4, w4, #1
+        ld1             {v6.16b}, [x2], x3
+        ext             v5.16b, v4.16b, v4.16b, #1              // x1 + stride + 1
+        usubl           v31.8h, v0.8b, v4.8b
+        ext             v7.16b, v6.16b, v6.16b, #1              // x2 + stride + 1
+
+        usubl2          v30.8h, v0.16b, v4.16b
+        usubl           v29.8h, v1.8b, v5.8b
+        usubl2          v28.8h, v1.16b, v5.16b
+        saba            v16.8h, v31.8h, v29.8h
+        saba            v17.8h, v30.8h, v28.8h
+        usubl           v27.8h, v2.8b, v6.8b
+        usubl2          v26.8h, v2.16b, v6.16b
+        usubl           v25.8h, v3.8b, v7.8b
+        usubl2          v24.8h, v3.16b, v7.16b
+        saba            v18.8h, v27.8h, v25.8h
+        saba            v19.8h, v26.8h, v24.8h
+
+        mov             v0.16b, v4.16b
+        mov             v1.16b, v5.16b
+        mov             v2.16b, v6.16b
+        mov             v3.16b, v7.16b
+
+        cbnz            w4, 2b
+
+3:
+        sqsub           v17.8h, v17.8h, v19.8h
+        sqsub           v16.8h, v16.8h, v18.8h
+        ins             v17.h[7], wzr
+        sqadd           v16.8h, v16.8h, v17.8h
+        saddlv          s16, v16.8h
+        sqabs           s16, s16
+        fmov            w0, s16
+
+        mul             w0, w0, w5
+        add             w0, w0, w9
+
+        ret
+endfunc
-- 
2.34.1

_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH 0/5] Provide optimized neon implementation

[Martin Storsjö]

On Thu, 8 Sep 2022, Hubert Mazur wrote:

> Fix minor issues in the patches.
> Regarding vsse16 I didn't change saba & umlal to sub & smlal.
> It doesn't affect the performance, so left it as it was.
> The majority of changes refer to nsse16:
> - fixed indentation (thanks for pointing out),
> - applied the patch from Martin which fixes the balance
>   within instructions,
> - interleaved instructions - apparently this helped a little
>   to achieve better benchmarks.

Thanks! I measured a small further improvement on A53 with this change; 
from 377 to 370 cycles.

> I have also updated the benchmark results for each function -
> not a huge performance improvement, but worth the effort.
> For nsse and vsse are shown below (these are the biggest changes).
> - vsse16 asm from 64.7 to 59.2,
> - nsse16 asm from 120.0 to 116.5.

It's kinda surprising that the difference is so small, since we reduced 
the amount of work done in the functions quite significantly (IIRC on A53, 
the speedup was something like 1.5x compared with the original), but I 
guess it's understandable if the Graviton 3 is so powerful, that there's 
enough spare execution units so that a bunch of redundant instructions 
doesn't really matter.

Anyway, this revision of the patchset looked good to me, so I pushed it 
now. Thanks!

// Martin

_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH 0/5] Provide optimized neon implementation

[Martin Storsjö]

On Tue, 6 Sep 2022, Hubert Mazur wrote:

> Provide optimized implementations for me_cmp functions.
> This set of patches fixes all issues addressed in previous review.
> Major changes:
> - Remove redundant loads since the data can be reused.
> - Improve style.
> - Fix issues with unrecognized symbols.

Thanks! This looks quite good to me now. I have a minor comment on vsse16 
and a slightly bigger one on nsse16, but other than that, this is almost 
good to go.

I noticed that you haven't updated the checkasm benchmark numbers in the 
commit messages since the previous round, while the code has seen some 
quite major changes. Please do rerun the numbers since I think they have 
changed notably.

// Martin

_______________________________________________
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".

[FFmpeg-devel] [PATCH 0/5] Provide optimized neon implementation

[Hubert Mazur]

Provide optimized implementations for me_cmp functions.
This set of patches fixes all issues addressed in previous review.
Major changes:
- Remove redundant loads since the data can be reused.
- Improve style.
- Fix issues with unrecognized symbols.

Hubert Mazur (5):
  lavc/aarch64: Add neon implementation for vsad16
  lavc/aarch64: Add neon implementation of vsse16
  lavc/aarch64: Add neon implementation for vsad_intra16
  lavc/aarch64: Add neon implementation for vsse_intra16
  lavc/aarch64: Provide neon implementation of nsse16

 libavcodec/aarch64/me_cmp_init_aarch64.c |  30 ++
 libavcodec/aarch64/me_cmp_neon.S         | 387 +++++++++++++++++++++++
 2 files changed, 417 insertions(+)

-- 
2.34.1

_______________________________________________
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".