[x264-devel] [PATCH 1/1] arm: use available neon functions for intra_sa8d/sad/satd_x3

[Janne Grunau]

4% faster on main/medium, 15% faster on baseline/superfast on a cortex-a9.
---
 common/arm/predict.h | 10 ++++++++++
 common/pixel.c       | 26 ++++++++++++++++++++++++++
 2 files changed, 36 insertions(+)

diff --git a/common/arm/predict.h b/common/arm/predict.h
index 0d07c2a..6cf2f5f 100644
--- a/common/arm/predict.h
+++ b/common/arm/predict.h
@@ -26,6 +26,16 @@
 #ifndef X264_ARM_PREDICT_H
 #define X264_ARM_PREDICT_H
 
+void x264_predict_8x8_v_neon( pixel *src, pixel edge[36] );
+void x264_predict_8x8_h_neon( pixel *src, pixel edge[36] );
+void x264_predict_8x8_dc_neon( pixel *src, pixel edge[36] );
+void x264_predict_8x8c_dc_neon( pixel *src );
+void x264_predict_8x8c_h_neon( pixel *src );
+void x264_predict_8x8c_v_neon( pixel *src );
+void x264_predict_16x16_v_neon( pixel *src );
+void x264_predict_16x16_h_neon( pixel *src );
+void x264_predict_16x16_dc_neon( pixel *src );
+
 void x264_predict_4x4_init_arm( int cpu, x264_predict_t pf[12] );
 void x264_predict_8x8_init_arm( int cpu, x264_predict8x8_t pf[12], x264_predict_8x8_filter_t *predict_filter );
 void x264_predict_8x8c_init_arm( int cpu, x264_predict_t pf[7] );
diff --git a/common/pixel.c b/common/pixel.c
index e16f292..b389792 100644
--- a/common/pixel.c
+++ b/common/pixel.c
@@ -36,6 +36,7 @@
 #endif
 #if ARCH_ARM
 #   include "arm/pixel.h"
+#   include "arm/predict.h"
 #endif
 #if ARCH_UltraSPARC
 #   include "sparc/pixel.h"
@@ -532,6 +533,10 @@ INTRA_MBCMP_8x8(sa8d,, _c )
 INTRA_MBCMP_8x8( sad, _mmx2,  _c )
 INTRA_MBCMP_8x8(sa8d, _sse2,  _sse2 )
 #endif
+#if !HIGH_BIT_DEPTH && HAVE_ARMV6
+INTRA_MBCMP_8x8( sad, _neon, _neon )
+INTRA_MBCMP_8x8(sa8d, _neon, _neon )
+#endif
 
 #define INTRA_MBCMP( mbcmp, size, pred1, pred2, pred3, chroma, cpu, cpu2 )\
 void x264_intra_##mbcmp##_x3_##size##chroma##cpu( pixel *fenc, pixel *fdec, int res[3] )\
@@ -587,6 +592,16 @@ INTRA_MBCMP(satd,  8x16, dc, h,  v, c, _avx, _mmx2 )
 INTRA_MBCMP(satd,  8x16, dc, h,  v, c, _xop, _mmx2 )
 #endif
 #endif
+#if !HIGH_BIT_DEPTH && HAVE_ARMV6
+INTRA_MBCMP( sad,  4x4,   v, h, dc,  , _neon, _c )
+INTRA_MBCMP(satd,  4x4,   v, h, dc,  , _neon, _c )
+INTRA_MBCMP( sad,  8x8,  dc, h,  v, c, _neon, _neon )
+INTRA_MBCMP(satd,  8x8,  dc, h,  v, c, _neon, _neon )
+INTRA_MBCMP( sad,  8x16, dc, h,  v, c, _neon, _c )
+INTRA_MBCMP(satd,  8x16, dc, h,  v, c, _neon, _c )
+INTRA_MBCMP( sad, 16x16,  v, h, dc,  , _neon, _neon )
+INTRA_MBCMP(satd, 16x16,  v, h, dc,  , _neon, _neon )
+#endif
 
 // No C implementation of intra_satd_x9. See checkasm for its behavior,
 // or see x264_mb_analyse_intra for the entirely different algorithm we
@@ -1352,6 +1367,17 @@ void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
         pixf->var2[PIXEL_8x8]   = x264_pixel_var2_8x8_neon;
         pixf->var2[PIXEL_8x16]  = x264_pixel_var2_8x16_neon;
 
+        pixf->intra_sad_x3_4x4    = x264_intra_sad_x3_4x4_neon;
+        pixf->intra_satd_x3_4x4   = x264_intra_satd_x3_4x4_neon;
+        pixf->intra_sad_x3_8x8    = x264_intra_sad_x3_8x8_neon;
+        pixf->intra_sa8d_x3_8x8   = x264_intra_sa8d_x3_8x8_neon;
+        pixf->intra_sad_x3_8x8c   = x264_intra_sad_x3_8x8c_neon;
+        pixf->intra_satd_x3_8x8c  = x264_intra_satd_x3_8x8c_neon;
+        pixf->intra_sad_x3_8x16c  = x264_intra_sad_x3_8x16c_neon;
+        pixf->intra_satd_x3_8x16c = x264_intra_satd_x3_8x16c_neon;
+        pixf->intra_sad_x3_16x16  = x264_intra_sad_x3_16x16_neon;
+        pixf->intra_satd_x3_16x16 = x264_intra_satd_x3_16x16_neon;
+
         pixf->ssim_4x4x2_core   = x264_pixel_ssim_4x4x2_core_neon;
         pixf->ssim_end4         = x264_pixel_ssim_end4_neon;
 
-- 
1.9.0