This is an automated email from the git hooks/post-receive script.
Git pushed a commit to branch master
in repository ffmpeg.
The following commit(s) were added to refs/heads/master by this push:
new de18feb0f0 avutil/aarch64: add pixelutils 32x32 SAD NEON implementation
de18feb0f0 is described below
commit de18feb0f099154e6ccb6d3746881030d2f53fa1
Author: Jeongkeun Kim <[email protected]>
AuthorDate: Sat Apr 4 20:20:13 2026 +0900
Commit: Martin Storsjö <[email protected]>
CommitDate: Sun Apr 19 19:27:55 2026 +0000
avutil/aarch64: add pixelutils 32x32 SAD NEON implementation
This adds a NEON-optimized function for computing 32x32 Sum of Absolute
Differences (SAD) on AArch64, addressing a gap where x86 had SSE2/AVX2
implementations but AArch64 lacked equivalent coverage.
The implementation mirrors the existing sad8 and sad16 NEON functions,
employing a 4-row unrolled loop with UABAL and UABAL2 instructions for
efficient load-compute interleaving, and four 8x16-bit accumulators to
handle the wider 32-byte rows.
Benchmarks on AWS Graviton3 (Neoverse V1, c7g.xlarge) using checkasm:
sad_32x32_0: C 146.4 cycles -> NEON 98.1 cycles (1.49x speedup)
sad_32x32_1: C 141.4 cycles -> NEON 98.9 cycles (1.43x speedup)
sad_32x32_2: C 140.7 cycles -> NEON 95.0 cycles (1.48x speedup)
Signed-off-by: Jeongkeun Kim <[email protected]>
---
libavutil/aarch64/pixelutils.h | 5 +++-
libavutil/aarch64/pixelutils_neon.S | 47 +++++++++++++++++++++++++++++++++++++
2 files changed, 51 insertions(+), 1 deletion(-)
diff --git a/libavutil/aarch64/pixelutils.h b/libavutil/aarch64/pixelutils.h
index e969ee81ed..02d22ab692 100644
--- a/libavutil/aarch64/pixelutils.h
+++ b/libavutil/aarch64/pixelutils.h
@@ -27,9 +27,11 @@
#include "libavutil/cpu.h"
#include "libavutil/pixelutils.h"
+int ff_pixelutils_sad8_neon (const uint8_t *src1, ptrdiff_t stride1,
+ const uint8_t *src2, ptrdiff_t stride2);
int ff_pixelutils_sad16_neon(const uint8_t *src1, ptrdiff_t stride1,
const uint8_t *src2, ptrdiff_t stride2);
-int ff_pixelutils_sad8_neon (const uint8_t *src1, ptrdiff_t stride1,
+int ff_pixelutils_sad32_neon(const uint8_t *src1, ptrdiff_t stride1,
const uint8_t *src2, ptrdiff_t stride2);
static inline av_cold void ff_pixelutils_sad_init_aarch64(av_pixelutils_sad_fn
*sad, int aligned)
@@ -39,6 +41,7 @@ static inline av_cold void
ff_pixelutils_sad_init_aarch64(av_pixelutils_sad_fn *
if (have_neon(cpu_flags)) {
sad[2] = ff_pixelutils_sad8_neon;
sad[3] = ff_pixelutils_sad16_neon;
+ sad[4] = ff_pixelutils_sad32_neon;
}
}
#endif
diff --git a/libavutil/aarch64/pixelutils_neon.S
b/libavutil/aarch64/pixelutils_neon.S
index 6e5178adb3..4b5c4d5414 100644
--- a/libavutil/aarch64/pixelutils_neon.S
+++ b/libavutil/aarch64/pixelutils_neon.S
@@ -86,3 +86,50 @@ function ff_pixelutils_sad8_neon, export=1
ret
endfunc
+
+function ff_pixelutils_sad32_neon, export=1
+ // x0 uint8_t *pix1
+ // x1 ptrdiff_t stride1
+ // x2 uint8_t *pix2
+ // x3 ptrdiff_t stride2
+ movi v16.8h, #0 // clear result accumulator
+ movi v17.8h, #0 // clear result accumulator
+ movi v18.8h, #0 // clear result accumulator
+ movi v19.8h, #0 // clear result accumulator
+ mov w4, 32
+1:
+ ld1 {v0.16b, v1.16b}, [x0], x1 // load pix1
+ ld1 {v4.16b, v5.16b}, [x2], x3 // load pix2
+ ld1 {v2.16b, v3.16b}, [x0], x1 // load pix1
+ ld1 {v6.16b, v7.16b}, [x2], x3 // load pix2
+ uabal v16.8h, v0.8b, v4.8b // absolute difference
accumulate
+ uabal2 v17.8h, v0.16b, v4.16b
+ uabal v18.8h, v1.8b, v5.8b
+ uabal2 v19.8h, v1.16b, v5.16b
+ ld1 {v0.16b, v1.16b}, [x0], x1 // load pix1
+ ld1 {v4.16b, v5.16b}, [x2], x3 // load pix2
+ uabal v16.8h, v2.8b, v6.8b // absolute difference
accumulate
+ uabal2 v17.8h, v2.16b, v6.16b
+ uabal v18.8h, v3.8b, v7.8b
+ uabal2 v19.8h, v3.16b, v7.16b
+ ld1 {v2.16b, v3.16b}, [x0], x1
+ ld1 {v6.16b, v7.16b}, [x2], x3
+ uabal v16.8h, v0.8b, v4.8b
+ uabal2 v17.8h, v0.16b, v4.16b
+ uabal v18.8h, v1.8b, v5.8b
+ uabal2 v19.8h, v1.16b, v5.16b
+ subs w4, w4, #4 // h -= 4
+ uabal v16.8h, v2.8b, v6.8b
+ uabal2 v17.8h, v2.16b, v6.16b
+ uabal v18.8h, v3.8b, v7.8b
+ uabal2 v19.8h, v3.16b, v7.16b
+
+ b.gt 1b // if h > 0, loop
+
+ add v16.8h, v16.8h, v17.8h
+ add v18.8h, v18.8h, v19.8h
+ add v16.8h, v16.8h, v18.8h
+ uaddlv s16, v16.8h // add up everything in
v16 accumulator
+ fmov w0, s16 // copy result to general
purpose register
+ ret
+endfunc
_______________________________________________
ffmpeg-cvslog mailing list -- [email protected]
To unsubscribe send an email to [email protected]
