[x264-devel] [PATCH 9/9] arm: x264_coeff_last8_arm

[Janne Grunau]

checkasm --bench on a coretex-a9:
coeff_last8_c: 173
coeff_last8_armv6: 151

60 instead of 73 cycles in ~130k runs on the same cpu while encoding.
---
 common/arm/quant-a.S | 14 ++++++++++++++
 common/arm/quant.h   |  1 +
 common/quant.c       |  4 +++-
 3 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/common/arm/quant-a.S b/common/arm/quant-a.S
index 2aeedc4..b8c6ba3 100644
--- a/common/arm/quant-a.S
+++ b/common/arm/quant-a.S
@@ -321,6 +321,20 @@ function x264_coeff_last4_arm
     bx          lr
 .endfunc
 
+function x264_coeff_last8_arm
+    ldrd        r2,  r3,  [r0, #8]
+    orrs        ip,  r2,  r3
+    movne       r0,  #4
+    ldrdeq      r2,  r3,  [r0]
+    moveq       r0,  #0
+    tst         r3,  r3
+    addne       r0,  #2
+    movne       r2,  r3
+    lsrs        r2,  r2,  #16
+    addne       r0,  r0,  #1
+    bx          lr
+.endfunc
+
 .macro COEFF_LAST_1x size
 function x264_coeff_last\size\()_neon
 .if \size == 15
diff --git a/common/arm/quant.h b/common/arm/quant.h
index 0695ab1..75d9fb2 100644
--- a/common/arm/quant.h
+++ b/common/arm/quant.h
@@ -39,6 +39,7 @@ void x264_dequant_4x4_neon( int16_t dct[16], int dequant_mf[6][16], int i_qp );
 void x264_dequant_8x8_neon( int16_t dct[64], int dequant_mf[6][64], int i_qp );
 
 int x264_coeff_last4_arm( int16_t * );
+int x264_coeff_last8_arm( int16_t * );
 int x264_coeff_last15_neon( int16_t * );
 int x264_coeff_last16_neon( int16_t * );
 int x264_coeff_last64_neon( int16_t * );
diff --git a/common/quant.c b/common/quant.c
index 169f39e..6427724 100644
--- a/common/quant.c
+++ b/common/quant.c
@@ -724,8 +724,10 @@ void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 #endif
 
 #if HAVE_ARMV6
-    if( cpu&X264_CPU_ARMV6 )
+    if( cpu&X264_CPU_ARMV6 ) {
         pf->coeff_last4 = x264_coeff_last4_arm;
+        pf->coeff_last8 = x264_coeff_last8_arm;
+    }
 
     if( cpu&X264_CPU_NEON )
     {
-- 
1.9.0