[x264-devel] ppc: Optimize add8x8_idct_dc

[Alexandra Hájková]

x264 | branch: master | Alexandra Hájková <alexandra at khirnov.net> | Sat May 13 17:14:52 2017 +0000| [181a920ad5d0acdc3a08418c0e9c95be4785b814] | committer: Henrik Gramner

ppc: Optimize add8x8_idct_dc

Increases speedup compared to C from 2x to 6x.

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=181a920ad5d0acdc3a08418c0e9c95be4785b814
---

 common/ppc/dct.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

diff --git a/common/ppc/dct.c b/common/ppc/dct.c
index 768f390b..2858bd0e 100644
--- a/common/ppc/dct.c
+++ b/common/ppc/dct.c
@@ -293,12 +293,8 @@ void x264_sub16x16_dct8_altivec( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix
     vec_vsx_st( dcvsum8, 0, dest );                                   \
 }
 
-static void idct8_dc_altivec( uint8_t *dst, int16_t dc1, int16_t dc2 )
+static void idct8_dc_altivec( uint8_t *dst, vec_s16_t dcv )
 {
-    dc1 = (dc1 + 32) >> 6;
-    dc2 = (dc2 + 32) >> 6;
-    vec_s16_t dcv = { dc1, dc1, dc1, dc1, dc2, dc2, dc2, dc2 };
-
     LOAD_ZERO;
     ALTIVEC_STORE8_DC_SUM_CLIP( &dst[0*FDEC_STRIDE], dcv );
     ALTIVEC_STORE8_DC_SUM_CLIP( &dst[1*FDEC_STRIDE], dcv );
@@ -308,8 +304,18 @@ static void idct8_dc_altivec( uint8_t *dst, int16_t dc1, int16_t dc2 )
 
 void x264_add8x8_idct_dc_altivec( uint8_t *p_dst, int16_t dct[4] )
 {
-    idct8_dc_altivec( &p_dst[0],               dct[0], dct[1] );
-    idct8_dc_altivec( &p_dst[4*FDEC_STRIDE+0], dct[2], dct[3] );
+    vec_s16_t dcv;
+    vec_s16_t v32 = vec_sl( vec_splat_s16( 8 ), vec_splat_u16( 2 ) );
+    vec_u16_t v6 = vec_splat_u16( 6 );
+    vec_s16_t dctv = vec_vsx_ld( 0, dct );
+
+    dctv = vec_sra( vec_add( dctv, v32 ), v6 );
+    dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)vec_splat( dctv, 0 ), (vec_s32_t)vec_splat( dctv, 1 ) );
+    dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)dcv, (vec_s32_t)dcv );
+    idct8_dc_altivec( &p_dst[0], dcv );
+    dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)vec_splat( dctv, 2 ), (vec_s32_t)vec_splat( dctv, 3 ) );
+    dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)dcv, (vec_s32_t)dcv );
+    idct8_dc_altivec( &p_dst[4*FDEC_STRIDE+0], dcv );
 }
 
 #define IDCT_1D_ALTIVEC(s0, s1, s2, s3,  d0, d1, d2, d3) \