PR #20448 opened by welder URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20448 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/20448.patch
The speed improvement is attached in the commit message. The count of arithmetic operation is down from 10 to 6 and some cruft is cleaned up. >From 8967e3a6a725358494307b51add9349d3e7dd075 Mon Sep 17 00:00:00 2001 From: Krzysztof Pyrkosz <[email protected]> Date: Fri, 5 Sep 2025 19:17:48 +0200 Subject: [PATCH] Optimize vvc_apply_bdof_block_8x Before and after: A53: apply_bdof_8_8x16_neon: 3320.5 ( 4.02x) apply_bdof_10_8x16_neon: 3317.8 ( 3.90x) apply_bdof_12_8x16_neon: 3303.6 ( 3.91x) apply_bdof_8_8x16_neon: 3216.2 ( 4.18x) apply_bdof_10_8x16_neon: 3181.0 ( 4.09x) apply_bdof_12_8x16_neon: 3172.1 ( 4.09x) A72: apply_bdof_8_8x16_neon: 1827.4 ( 5.02x) apply_bdof_10_8x16_neon: 1838.5 ( 4.89x) apply_bdof_12_8x16_neon: 1841.1 ( 4.83x) apply_bdof_8_8x16_neon: 1691.6 ( 5.46x) apply_bdof_10_8x16_neon: 1695.9 ( 5.23x) apply_bdof_12_8x16_neon: 1695.4 ( 5.29x) A78 apply_bdof_8_8x16_neon: 648.9 ( 7.43x) apply_bdof_10_8x16_neon: 646.1 ( 7.04x) apply_bdof_12_8x16_neon: 643.8 ( 7.04x) apply_bdof_8_8x16_neon: 603.2 ( 7.97x) apply_bdof_10_8x16_neon: 604.1 ( 7.52x) apply_bdof_12_8x16_neon: 604.5 ( 7.52x) --- libavcodec/aarch64/vvc/inter.S | 39 ++++++++++++---------------------- 1 file changed, 13 insertions(+), 26 deletions(-) diff --git a/libavcodec/aarch64/vvc/inter.S b/libavcodec/aarch64/vvc/inter.S index a6648b64fc..f27b5a47f4 100644 --- a/libavcodec/aarch64/vvc/inter.S +++ b/libavcodec/aarch64/vvc/inter.S @@ -802,43 +802,33 @@ endfunc vy .req x7 ldr w8, [sp] - movi v7.4s, #(1 << (14 - \bit_depth)) mov x12, #(BDOF_BLOCK_SIZE * 2) mov x14, #(VVC_MAX_PB_SIZE * 2) .if \bit_depth >= 10 // clip pixel mov w15, #((1 << \bit_depth) - 1) - movi v18.8h, #0 dup v19.8h, w15 .endif 0: - ld1r {v0.8h}, [vx], #2 - ld1r {v1.8h}, [vy], #2 - ld1r {v2.8h}, [vx] - ld1r {v3.8h}, [vy] + ldr s0, [vx], #(2 * BDOF_MIN_BLOCK_SIZE) + ldr s1, [vy], #(2 * BDOF_MIN_BLOCK_SIZE) mov w13, #(BDOF_MIN_BLOCK_SIZE) - ins v0.d[1], v2.d[1] - ins v1.d[1], v3.d[1] 1: - ld1 {v2.8h}, [gh], x12 - ld1 {v4.8h}, [gv], x12 - smull v3.4s, v0.4h, v2.4h - smull2 v16.4s, v0.8h, v2.8h - smlal v3.4s, v1.4h, v4.4h - smlal2 v16.4s, v1.8h, v4.8h - ld1 {v5.8h}, [src0], x14 ld1 {v6.8h}, [src1], x14 - saddl v2.4s, v5.4h, v6.4h - add v2.4s, v2.4s, v7.4s - add v2.4s, v2.4s, v3.4s - saddl2 v4.4s, v5.8h, v6.8h - add v4.4s, v4.4s, v7.4s - add v4.4s, v4.4s, v16.4s + ld1 {v2.8h}, [gh], x12 + ld1 {v4.8h}, [gv], x12 - sqshrn v5.4h, v2.4s, #(15 - \bit_depth) - sqshrn2 v5.8h, v4.4s, #(15 - \bit_depth) + saddl v17.4s, v5.4h, v6.4h + saddl2 v16.4s, v5.8h, v6.8h + smlal v17.4s, v4.4h, v1.h[0] + smlal2 v16.4s, v4.8h, v1.h[1] + smlal v17.4s, v2.4h, v0.h[0] + smlal2 v16.4s, v2.8h, v0.h[1] + + sqrshrun v5.4h, v17.4s, #(15 - \bit_depth) + sqrshrun2 v5.8h, v16.4s, #(15 - \bit_depth) subs w13, w13, #1 .if \bit_depth == 8 sqxtun v5.8b, v5.8h @@ -846,14 +836,11 @@ endfunc add dst, dst, dst_stride .else smin v5.8h, v5.8h, v19.8h - smax v5.8h, v5.8h, v18.8h st1 {v5.8h}, [dst], dst_stride .endif b.ne 1b subs w8, w8, #(BDOF_MIN_BLOCK_SIZE) - add vx, vx, #(2 * BDOF_MIN_BLOCK_SIZE - 2) - add vy, vy, #(2 * BDOF_MIN_BLOCK_SIZE - 2) b.ne 0b ret -- 2.49.1 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
