diff mbox series

[FFmpeg-devel,2/4] lavc/aarch64: Provide neon implementation of nsse8

Message ID 20220926091504.3459110-2-gjb@semihalf.com
State New
Headers show
Series [FFmpeg-devel,1/4] lavc/aarch64: Add neon implementation for pix_abs8 functions. | expand

Checks

Context Check Description
yinshiyou/make_loongarch64 success Make finished
yinshiyou/make_fate_loongarch64 success Make fate finished
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished

Commit Message

Grzegorz Bernacki Sept. 26, 2022, 9:15 a.m. UTC 
Add vectorized implementation of nsse8 function.

Performance comparison tests are shown below.
- nsse_1_c: 256.0
- nsse_1_neon: 82.7

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

Signed-off-by: Grzegorz Bernacki <gjb@semihalf.com>
---
 libavcodec/aarch64/me_cmp_init_aarch64.c | 15 ++++
 libavcodec/aarch64/me_cmp_neon.S         | 99 ++++++++++++++++++++++++
 2 files changed, 114 insertions(+)

Comments

Martin Storsjö Sept. 28, 2022, 9:08 a.m. UTC | #1 
On Mon, 26 Sep 2022, Grzegorz Bernacki wrote:

> Add vectorized implementation of nsse8 function.
>
> Performance comparison tests are shown below.
> - nsse_1_c: 256.0
> - nsse_1_neon: 82.7
>
> Benchmarks and tests run with checkasm tool on AWS Graviton 3.
>
> Signed-off-by: Grzegorz Bernacki <gjb@semihalf.com>
> ---
> libavcodec/aarch64/me_cmp_init_aarch64.c | 15 ++++
> libavcodec/aarch64/me_cmp_neon.S         | 99 ++++++++++++++++++++++++
> 2 files changed, 114 insertions(+)

Looks reasonable to me, but do check to make sure there's no tabs.

// Martin
diff mbox series

Patch

diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c
index 3459403ee5..2c61cfcf63 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -66,6 +66,11 @@  int ff_pix_abs8_y2_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *p
 int ff_pix_abs8_xy2_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
                          ptrdiff_t stride, int h);
 
+int nsse8_neon(int multiplier, const uint8_t *s, const uint8_t *s2,
+               ptrdiff_t stride, int h);
+int nsse8_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)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -94,6 +99,7 @@  av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx)
         c->vsse[4] = vsse_intra16_neon;
 
         c->nsse[0] = nsse16_neon_wrapper;
+        c->nsse[1] = nsse8_neon_wrapper;
 
         c->median_sad[0] = pix_median_abs16_neon;
         c->median_sad[1] = pix_median_abs8_neon;
@@ -108,3 +114,12 @@  int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
     else
         return nsse16_neon(8, s1, s2, stride, h);
 }
+
+int nsse8_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2,
+                       ptrdiff_t stride, int h)
+{
+    if (c)
+        return nsse8_neon(c->avctx->nsse_weight, s1, s2, stride, h);
+    else
+        return nsse8_neon(8, s1, s2, stride, h);
+}
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index e03c0c26cd..6f7c7c1690 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -1163,6 +1163,105 @@  function nsse16_neon, export=1
         ret
 endfunc
 
+function nsse8_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(sse8_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             {v0.8b}, [x1], x3
+        subs            w4, w4, #1                              // we need to make h-1 iterations
+        ext             v1.8b, v0.8b, v0.8b, #1                 // x1 + 1
+        ld1             {v2.8b}, [x2], x3
+        cmp             w4, #2
+        ext             v3.8b, v2.8b, v2.8b, #1                 // x2 + 1
+
+        b.lt            2f
+
+// make 2 iterations at once
+1:
+        ld1             {v4.8b}, [x1], x3
+        ld1             {v20.8b}, [x1], x3
+        ld1             {v6.8b}, [x2], x3
+        ext             v5.8b, v4.8b, v4.8b, #1                 // x1 + stride + 1
+        ext             v21.8b, v20.8b, v20.8b, #1
+        ld1             {v22.8b}, [x2], x3
+        ext             v7.8b, v6.8b, v6.8b, #1                 // x2 + stride + 1
+        usubl           v31.8h, v0.8b, v4.8b
+        ext             v23.8b, v22.8b, v22.8b, #1
+        usubl           v29.8h, v1.8b, v5.8b
+        usubl           v27.8h, v2.8b, v6.8b
+        usubl           v25.8h, v3.8b, v7.8b
+        saba            v16.8h, v31.8h, v29.8h
+        usubl           v31.8h, v4.8b, v20.8b
+        saba            v18.8h, v27.8h, v25.8h
+        sub             w4, w4, #2
+        usubl           v29.8h, v5.8b, v21.8b
+        mov             v0.16b, v20.16b
+        mov             v1.16b, v21.16b
+        saba            v16.8h, v31.8h, v29.8h
+        usubl           v27.8h, v6.8b, v22.8b
+        usubl           v25.8h, v7.8b, v23.8b
+        mov             v2.16b, v22.16b
+        mov             v3.16b, v23.16b
+        cmp             w4, #2
+        saba            v18.8h, v27.8h, v25.8h
+        b.ge            1b
+        cbz             w4, 3f
+
+// iterate by one
+2:
+        ld1             {v4.8b}, [x1], x3
+        subs            w4, w4, #1
+        ext             v5.8b, v4.8b, v4.8b, #1                 // x1 + stride + 1
+        ld1             {v6.8b}, [x2], x3
+        usubl           v31.8h, v0.8b, v4.8b
+        ext             v7.8b, v6.8b, v6.8b, #1                 // x2 + stride + 1
+
+        usubl           v29.8h, v1.8b, v5.8b
+        saba            v16.8h, v31.8h, v29.8h
+        usubl           v27.8h, v2.8b, v6.8b
+        usubl           v25.8h, v3.8b, v7.8b
+        saba            v18.8h, v27.8h, v25.8h
+
+        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           v16.8h, v16.8h, v18.8h
+        ins             v16.h[7], wzr
+        saddlv          s16, v16.8h
+        sqabs           s16, s16
+        fmov            w0, s16
+
+        mul             w0, w0, w5
+        add             w0, w0, w9
+
+        ret
+endfunc
+
 function pix_median_abs16_neon, export=1
         // x0           unused
         // x1           uint8_t *pix1