[x264-devel] [PATCH 1/1] arm: optimize neon luma intra deblock

[Janne Grunau]

Hi Martin,

I forgot to rescheduling the beginning of the macro in the last iteration.
The floating point compare is a neat trick to test a 64-bit register
against 0 and control the program flow based on it. I tend to forget that
this is possible.

We might have to add vmrs handling to gaspp. Older assembers might
understand fmstat instead.

Janne

---8<---
Almost another 10% faster.

                            a7    a15
deblock_luma_intra[0]_neon: 3059  1128
deblock_luma_intra[1]_neon: 2038   720
---

 common/arm/deblock-a.S | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/common/arm/deblock-a.S b/common/arm/deblock-a.S
index 8d8d807..74c1f0e 100644
--- a/common/arm/deblock-a.S
+++ b/common/arm/deblock-a.S
@@ -197,26 +197,22 @@ endfunc
 
 .macro h264_loop_filter_luma_intra
     vdup.8          q14, r2         @ alpha
-    vdup.8          q15, r3         @ beta
     vabd.u8         q4,  q8,  q0    @ abs(p0 - q0)
     vabd.u8         q5,  q9,  q8    @ abs(p1 - p0)
     vabd.u8         q6,  q1,  q0    @ abs(q1 - q0)
-    vclt.u8         q7,  q4,  q14   @ < alpha
-    vclt.u8         q5,  q5,  q15   @ < beta
-    vclt.u8         q6,  q6,  q15   @ < beta
-
+    vdup.8          q15, r3         @ beta
     vmov.u8         q13, #2
+    vclt.u8         q7,  q4,  q14   @ < alpha
     vshr.u8         q14, q14, #2    @ alpha >> 2
+    vclt.u8         q5,  q5,  q15   @ < beta
     vadd.u8         q14, q14, q13   @ (alpha >> 2) + 2
-    vclt.u8         q13, q4,  q14   @ < (alpha >> 2) + 2 if_2
-
     vand            q7,  q7,  q5
+    vclt.u8         q6,  q6,  q15   @ < beta
+    vclt.u8         q13, q4,  q14   @ < (alpha >> 2) + 2 if_2
     vand            q12, q7,  q6    @ if_1
     vshrn.u16       d28, q12,  #4
-    vrev64.32       d29, d28
-    vorr            d28, d28, d29
-    vmov.32         r2,  d28[0]
-    cmp             r2,  #0
+    vcmp.f64        d28, #0
+    vmrs            APSR_nzcv, FPSCR
     beq             9f
 
     sub             sp,  sp,  #32
-- 
2.5.0