PR #25022 opened by darpat039
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25022
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25022.patch
# Summary of changes
Adds a NEON version of scalarproduct_double for AArch64. x86 and RISC-V already
have one. The C version is not vectorized by the compiler because that would
reorder the floating-point additions.
It uses four accumulators, the same structure as scalarproduct_float in this
file. I also tried two and eight: two was clearly slower, eight was not faster
at the length checkasm uses (256) and only slightly faster on much longer
inputs.
As with the other SIMD versions, the result can differ from C in the last bits
because of the summation order.
checkasm --bench on Apple M4:
scalarproduct_double_c: 80.3
scalarproduct_double_neon: 21.8 ( 3.68x)
Tested with checkasm --test=float_dsp. I only have Apple Silicon to test on.
>From 3e320aca04e4ed50648b0247503ec54f039df271 Mon Sep 17 00:00:00 2001
From: Patryk Darmofalski <[email protected]>
Date: Mon, 12 Oct 2026 13:55:46 +0200
Subject: [PATCH] avutil/aarch64/float_dsp: add NEON scalarproduct_double
Four accumulators, as in scalarproduct_float. Eight were not faster at len 256.
checkasm --bench on Apple M4:
scalarproduct_double_c: 80.3
scalarproduct_double_neon: 21.8 ( 3.68x)
Signed-off-by: Patryk Darmofalski <[email protected]>
---
libavutil/aarch64/float_dsp_init.c | 4 ++++
libavutil/aarch64/float_dsp_neon.S | 23 +++++++++++++++++++++++
2 files changed, 27 insertions(+)
diff --git a/libavutil/aarch64/float_dsp_init.c
b/libavutil/aarch64/float_dsp_init.c
index 4325071821..0423602303 100644
--- a/libavutil/aarch64/float_dsp_init.c
+++ b/libavutil/aarch64/float_dsp_init.c
@@ -51,12 +51,16 @@ void ff_butterflies_float_neon(float *v1, float *v2, int
len);
float ff_scalarproduct_float_neon(const float *v1, const float *v2, int len);
+double ff_scalarproduct_double_neon(const double *v1, const double *v2,
+ size_t len);
+
av_cold void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp)
{
int cpu_flags = av_get_cpu_flags();
if (have_neon(cpu_flags)) {
fdsp->butterflies_float = ff_butterflies_float_neon;
+ fdsp->scalarproduct_double = ff_scalarproduct_double_neon;
fdsp->scalarproduct_float = ff_scalarproduct_float_neon;
fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_neon;
fdsp->vector_fmul = ff_vector_fmul_neon;
diff --git a/libavutil/aarch64/float_dsp_neon.S
b/libavutil/aarch64/float_dsp_neon.S
index aa6b5189f5..d2bc4df7b8 100644
--- a/libavutil/aarch64/float_dsp_neon.S
+++ b/libavutil/aarch64/float_dsp_neon.S
@@ -233,3 +233,26 @@ function ff_scalarproduct_float_neon, export=1
faddp s0, v0.2s
ret
endfunc
+
+function ff_scalarproduct_double_neon, export=1
+ movi v0.2d, #0
+ movi v1.2d, #0
+ movi v2.2d, #0
+ movi v3.2d, #0
+1:
+ ld1 {v4.2d, v5.2d}, [x0], #32
+ ld1 {v6.2d, v7.2d}, [x1], #32
+ ld1 {v16.2d, v17.2d}, [x0], #32
+ ld1 {v18.2d, v19.2d}, [x1], #32
+ subs x2, x2, #8
+ fmla v0.2d, v4.2d, v6.2d
+ fmla v1.2d, v5.2d, v7.2d
+ fmla v2.2d, v16.2d, v18.2d
+ fmla v3.2d, v17.2d, v19.2d
+ b.gt 1b
+ fadd v0.2d, v0.2d, v1.2d
+ fadd v2.2d, v2.2d, v3.2d
+ fadd v0.2d, v0.2d, v2.2d
+ faddp d0, v0.2d
+ ret
+endfunc
--
2.52.0
_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]
