PR #21545 opened by chenmingbiao
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21545
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/21545.patch

Performance improvement for draw_edges function using NEON instructions.
Optimized for edge widths 4, 8, and 16.

Benchmark results (AArch64):
draw_edges_8_1724_4_c:       851.4
draw_edges_8_1724_4_neon:    824.7 (1.03x)
draw_edges_8_1724_16_c:     1620.9
draw_edges_8_1724_16_neon:  1049.9 (1.54x)
draw_edges_1920_4_16_c:      849.0
draw_edges_1920_4_16_neon:   571.1 (1.49x)


>From cf54f4e51105d5b89f9c7782e3c457b26621d113 Mon Sep 17 00:00:00 2001
From: mbillchan <[email protected]>
Date: Wed, 21 Jan 2026 17:28:24 +0800
Subject: [PATCH] avcodec/aarch64: add NEON optimization for draw_edges

Performance improvement for draw_edges function using NEON instructions.
Optimized for edge widths 4, 8, and 16.

Benchmark results (AArch64):
draw_edges_8_1724_4_c:       851.4
draw_edges_8_1724_4_neon:    824.7 (1.03x)
draw_edges_8_1724_16_c:     1620.9
draw_edges_8_1724_16_neon:  1049.9 (1.54x)
draw_edges_1920_4_16_c:      849.0
draw_edges_1920_4_16_neon:   571.1 (1.49x)
---
 libavcodec/aarch64/mpegvideoencdsp_init.c |   3 +
 libavcodec/aarch64/mpegvideoencdsp_neon.S | 157 ++++++++++++++++++++++
 tests/checkasm/mpegvideoencdsp.c          |   5 +
 3 files changed, 165 insertions(+)

diff --git a/libavcodec/aarch64/mpegvideoencdsp_init.c 
b/libavcodec/aarch64/mpegvideoencdsp_init.c
index 0402ac9e82..8363b18f48 100644
--- a/libavcodec/aarch64/mpegvideoencdsp_init.c
+++ b/libavcodec/aarch64/mpegvideoencdsp_init.c
@@ -26,6 +26,8 @@
 
 int ff_pix_sum16_neon(const uint8_t *pix, ptrdiff_t line_size);
 int ff_pix_norm1_neon(const uint8_t *pix, ptrdiff_t line_size);
+void ff_draw_edges_neon(uint8_t *buf, ptrdiff_t wrap, int width, int height,
+                        int w, int h, int sides);
 
 #if HAVE_DOTPROD
 int ff_pix_norm1_neon_dotprod(const uint8_t *pix, ptrdiff_t line_size);
@@ -39,6 +41,7 @@ av_cold void 
ff_mpegvideoencdsp_init_aarch64(MpegvideoEncDSPContext *c,
     if (have_neon(cpu_flags)) {
         c->pix_sum   = ff_pix_sum16_neon;
         c->pix_norm1 = ff_pix_norm1_neon;
+        c->draw_edges = ff_draw_edges_neon;
     }
 
 #if HAVE_DOTPROD
diff --git a/libavcodec/aarch64/mpegvideoencdsp_neon.S 
b/libavcodec/aarch64/mpegvideoencdsp_neon.S
index 8175d8975e..29a6104241 100644
--- a/libavcodec/aarch64/mpegvideoencdsp_neon.S
+++ b/libavcodec/aarch64/mpegvideoencdsp_neon.S
@@ -91,3 +91,160 @@ endfunc
 
 DISABLE_DOTPROD
 #endif
+
+function ff_draw_edges_neon, export=1
+        mov             x9, x0
+        mov             x10, x1
+        uxtw            x11, w2
+        uxtw            x12, w3
+        uxtw            x13, w4
+        uxtw            x14, w5
+        uxtw            x15, w6
+        mov             x0, x9
+        mov             x1, x12
+
+        cmp             w13, #16
+        b.eq            1f
+        cmp             w13, #8
+        b.eq            2f
+        b               3f
+
+1:
+        ldrb            w2, [x0]
+        dup             v0.16b, w2
+        sub             x3, x0, #16
+        st1             {v0.16b}, [x3]
+
+        add             x3, x0, x11
+        sub             x4, x3, #1
+        ldrb            w2, [x4]
+        dup             v1.16b, w2
+        st1             {v1.16b}, [x3]
+
+        add             x0, x0, x10
+        subs            x1, x1, #1
+        b.ne            1b
+        b               4f
+
+2:
+        ldrb            w2, [x0]
+        dup             v0.8b, w2
+        sub             x3, x0, #8
+        st1             {v0.8b}, [x3]
+
+        add             x3, x0, x11
+        sub             x4, x3, #1
+        ldrb            w2, [x4]
+        dup             v1.8b, w2
+        st1             {v1.8b}, [x3]
+
+        add             x0, x0, x10
+        subs            x1, x1, #1
+        b.ne            2b
+        b               4f
+
+3:
+        ldrb            w2, [x0]
+        dup             v0.8b, w2
+        sub             x3, x0, #4
+        st1             {v0.s}[0], [x3]
+
+        add             x3, x0, x11
+        sub             x4, x3, #1
+        ldrb            w2, [x4]
+        dup             v1.8b, w2
+        st1             {v1.s}[0], [x3]
+
+        add             x0, x0, x10
+        subs            x1, x1, #1
+        b.ne            3b
+
+4:
+        sub             x9, x9, x13
+        tbnz            w15, #0, 5f
+        b               6f
+
+5:
+        add             x2, x11, x13, lsl #1
+        mov             x3, x14
+        mov             x4, x9
+        sub             x5, x9, x10
+
+7:
+        mov             x6, x5
+        mov             x7, x4
+        mov             x8, x2
+
+8:
+        cmp             x8, #64
+        b.lt            17f
+        ld1             {v0.16b, v1.16b, v2.16b, v3.16b}, [x7], #64
+        st1             {v0.16b, v1.16b, v2.16b, v3.16b}, [x6], #64
+        sub             x8, x8, #64
+        b               8b
+
+17:
+        cmp             x8, #16
+        b.lt            9f
+        ld1             {v0.16b}, [x7], #16
+        st1             {v0.16b}, [x6], #16
+        sub             x8, x8, #16
+        b               17b
+9:
+        cbz             x8, 10f
+        ldrb            w16, [x7], #1
+        strb            w16, [x6], #1
+        sub             x8, x8, #1
+        b               9b
+10:
+        sub             x5, x5, x10
+        subs            x3, x3, #1
+        b.ne            7b
+
+6:
+        tbnz            w15, #1, 11f
+        b               12f
+
+11:
+        sub             x3, x12, #1
+        mul             x3, x3, x10
+        add             x4, x9, x3
+        add             x5, x4, x10
+
+        add             x2, x11, x13, lsl #1
+        mov             x3, x14
+
+13:
+        mov             x6, x5
+        mov             x7, x4
+        mov             x8, x2
+
+14:
+        cmp             x8, #64
+        b.lt            18f
+        ld1             {v0.16b, v1.16b, v2.16b, v3.16b}, [x7], #64
+        st1             {v0.16b, v1.16b, v2.16b, v3.16b}, [x6], #64
+        sub             x8, x8, #64
+        b               14b
+
+18:
+        cmp             x8, #16
+        b.lt            15f
+        ld1             {v0.16b}, [x7], #16
+        st1             {v0.16b}, [x6], #16
+        sub             x8, x8, #16
+        b               18b
+15:
+        cbz             x8, 16f
+        ldrb            w16, [x7], #1
+        strb            w16, [x6], #1
+        sub             x8, x8, #1
+        b               15b
+16:
+        add             x5, x5, x10
+        subs            x3, x3, #1
+        b.ne            13b
+
+12:
+        ret
+endfunc
diff --git a/tests/checkasm/mpegvideoencdsp.c b/tests/checkasm/mpegvideoencdsp.c
index 955cd9f5b7..fdaed08e86 100644
--- a/tests/checkasm/mpegvideoencdsp.c
+++ b/tests/checkasm/mpegvideoencdsp.c
@@ -147,8 +147,13 @@ static void check_draw_edges(MpegvideoEncDSPContext *c)
     LOCAL_ALIGNED_16(uint8_t, buf0, [BUFSIZE]);
     LOCAL_ALIGNED_16(uint8_t, buf1, [BUFSIZE]);
 
+#if ARCH_X86
     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *buf, ptrdiff_t wrap, int 
width, int height,
                                              int w, int h, int sides);
+#else
+    declare_func(void, uint8_t *buf, ptrdiff_t wrap, int width, int height,
+                                             int w, int h, int sides);
+#endif
 
     for (int isi = 0; isi < FF_ARRAY_ELEMS(input_sizes); isi++) {
         int input_size = input_sizes[isi];
-- 
2.52.0

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to