[x264-devel] ppc: Manually unroll the horizontal prediction loop

[Luca Barbato]

x264 | branch: master | Luca Barbato <lu_zero at gentoo.org> | Tue Nov  1 23:16:13 2016 +0100| [1f7518182e3204cb14e87baffb0150a848167ddc] | committer: Henrik Gramner

ppc: Manually unroll the horizontal prediction loop

Doubles the speedup from the function (from being slower to be over
twice as fast than C).

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=1f7518182e3204cb14e87baffb0150a848167ddc
---

 common/ppc/predict.c | 67 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 59 insertions(+), 8 deletions(-)

diff --git a/common/ppc/predict.c b/common/ppc/predict.c
index f35e180..2dc9ecd 100644
--- a/common/ppc/predict.c
+++ b/common/ppc/predict.c
@@ -176,14 +176,65 @@ static void predict_16x16_dc_128_altivec( uint8_t *src )
 
 static void predict_16x16_h_altivec( uint8_t *src )
 {
-    for( int i = 0; i < 16; i++ )
-    {
-        vec_u8_t v = vec_ld(-1, src);
-        vec_u8_t v_v = vec_splat(v, 15);
-        vec_st(v_v, 0, src);
-
-        src += FDEC_STRIDE;
-    }
+    vec_u8_t v1 = vec_ld( -1, src );
+    vec_u8_t v2 = vec_ld( -1, src + FDEC_STRIDE );
+    vec_u8_t v3 = vec_ld( -1, src + FDEC_STRIDE * 2 );
+    vec_u8_t v4 = vec_ld( -1, src + FDEC_STRIDE * 3 );
+
+    vec_u8_t v5 = vec_ld( -1, src + FDEC_STRIDE * 4 );
+    vec_u8_t v6 = vec_ld( -1, src + FDEC_STRIDE * 5 );
+    vec_u8_t v7 = vec_ld( -1, src + FDEC_STRIDE * 6 );
+    vec_u8_t v8 = vec_ld( -1, src + FDEC_STRIDE * 7 );
+
+    vec_u8_t v9 = vec_ld( -1, src + FDEC_STRIDE * 8 );
+    vec_u8_t vA = vec_ld( -1, src + FDEC_STRIDE * 9 );
+    vec_u8_t vB = vec_ld( -1, src + FDEC_STRIDE * 10 );
+    vec_u8_t vC = vec_ld( -1, src + FDEC_STRIDE * 11 );
+
+    vec_u8_t vD = vec_ld( -1, src + FDEC_STRIDE * 12 );
+    vec_u8_t vE = vec_ld( -1, src + FDEC_STRIDE * 13 );
+    vec_u8_t vF = vec_ld( -1, src + FDEC_STRIDE * 14 );
+    vec_u8_t vG = vec_ld( -1, src + FDEC_STRIDE * 15 );
+
+    vec_u8_t v_v1 = vec_splat( v1, 15 );
+    vec_u8_t v_v2 = vec_splat( v2, 15 );
+    vec_u8_t v_v3 = vec_splat( v3, 15 );
+    vec_u8_t v_v4 = vec_splat( v4, 15 );
+
+    vec_u8_t v_v5 = vec_splat( v5, 15 );
+    vec_u8_t v_v6 = vec_splat( v6, 15 );
+    vec_u8_t v_v7 = vec_splat( v7, 15 );
+    vec_u8_t v_v8 = vec_splat( v8, 15 );
+
+    vec_u8_t v_v9 = vec_splat( v9, 15 );
+    vec_u8_t v_vA = vec_splat( vA, 15 );
+    vec_u8_t v_vB = vec_splat( vB, 15 );
+    vec_u8_t v_vC = vec_splat( vC, 15 );
+
+    vec_u8_t v_vD = vec_splat( vD, 15 );
+    vec_u8_t v_vE = vec_splat( vE, 15 );
+    vec_u8_t v_vF = vec_splat( vF, 15 );
+    vec_u8_t v_vG = vec_splat( vG, 15 );
+
+    vec_st( v_v1, 0, src );
+    vec_st( v_v2, 0, src + FDEC_STRIDE );
+    vec_st( v_v3, 0, src + FDEC_STRIDE * 2 );
+    vec_st( v_v4, 0, src + FDEC_STRIDE * 3 );
+
+    vec_st( v_v5, 0, src + FDEC_STRIDE * 4 );
+    vec_st( v_v6, 0, src + FDEC_STRIDE * 5 );
+    vec_st( v_v7, 0, src + FDEC_STRIDE * 6 );
+    vec_st( v_v8, 0, src + FDEC_STRIDE * 7 );
+
+    vec_st( v_v9, 0, src + FDEC_STRIDE * 8 );
+    vec_st( v_vA, 0, src + FDEC_STRIDE * 9 );
+    vec_st( v_vB, 0, src + FDEC_STRIDE * 10 );
+    vec_st( v_vC, 0, src + FDEC_STRIDE * 11 );
+
+    vec_st( v_vD, 0, src + FDEC_STRIDE * 12 );
+    vec_st( v_vE, 0, src + FDEC_STRIDE * 13 );
+    vec_st( v_vF, 0, src + FDEC_STRIDE * 14 );
+    vec_st( v_vG, 0, src + FDEC_STRIDE * 15 );
 }
 
 static void predict_16x16_v_altivec( uint8_t *src )