[FFmpeg-devel] [PATCH] avutil/aarch64/float_dsp_neon: Refactor ff_butterflies_float_neon

[Krzysztof Pyrkosz]

Modified the main loop to handle 8 floats in one go. A special case of
length not being multiple of 8 is handled at the beginning. The speed
increased from 3.90 to 4.50.

Krzysztof

---
 libavutil/aarch64/float_dsp_neon.S | 30 ++++++++++++++++++++++--------
 1 file changed, 22 insertions(+), 8 deletions(-)

diff --git a/libavutil/aarch64/float_dsp_neon.S b/libavutil/aarch64/float_dsp_neon.S
index 35e2715b87..f99cc2e6b2 100644
--- a/libavutil/aarch64/float_dsp_neon.S
+++ b/libavutil/aarch64/float_dsp_neon.S
@@ -178,14 +178,28 @@ function ff_vector_fmul_reverse_neon, export=1
 endfunc
 
 function ff_butterflies_float_neon, export=1
-1:      ld1             {v0.4s}, [x0]
-        ld1             {v1.4s}, [x1]
-        subs            w2,  w2,  #4
-        fsub            v2.4s,   v0.4s,  v1.4s
-        fadd            v3.4s,   v0.4s,  v1.4s
-        st1             {v2.4s}, [x1],   #16
-        st1             {v3.4s}, [x0],   #16
-        b.gt            1b
+        tst            w2, #4
+        b.eq           bf8
+        ldr            q0, [x0]
+        ldr            q1, [x1]
+        fadd           v2.4s, v0.4s, v1.4s
+        fsub           v3.4s, v0.4s, v1.4s
+        str            q2, [x0], #16
+        str            q3, [x1], #16
+        sub            w2, w2, #4
+        cbnz           w2, bf8
+        ret
+bf8:
+        ldp            q0, q1, [x0]
+        ldp            q4, q5, [x1]
+        fadd           v2.4s, v0.4s, v4.4s
+        fadd           v3.4s, v1.4s, v5.4s
+        stp            q2, q3, [x0], #32
+        fsub           v2.4s, v0.4s, v4.4s
+        fsub           v3.4s, v1.4s, v5.4s
+        stp            q2, q3, [x1], #32
+        sub            w2, w2, #8
+        cbnz           w2, bf8
         ret
 endfunc
 
-- 
2.45.2