[Pixman] [PATCH 5/5] mmx: optimize bilinear function when using 7-bit precision

[Matt Turner]

---
Reduces runtime of firefox-planet-gnome trace from 156 to 153 seconds on Loongson.

Increases runtime of firefox-fishtank trace from 1030 to 1060 seconds. Why?

 pixman/pixman-mmx.c |   45 ++++++++++++++++++++++++++++++++-------------
 1 files changed, 32 insertions(+), 13 deletions(-)

diff --git a/pixman/pixman-mmx.c b/pixman/pixman-mmx.c
index ea732bb..bff8585 100644
--- a/pixman/pixman-mmx.c
+++ b/pixman/pixman-mmx.c
@@ -3526,11 +3526,14 @@ mmx_composite_over_reverse_n_8888 (pixman_implementation_t *imp,
 }
 
 #define BSHIFT ((1 << BILINEAR_INTERPOLATION_BITS))
+#define BMSK (BSHIFT - 1)
 
 #define BILINEAR_DECLARE_VARIABLES						\
     const __m64 mm_wt = _mm_set_pi16 (wt, wt, wt, wt);				\
     const __m64 mm_wb = _mm_set_pi16 (wb, wb, wb, wb);				\
     const __m64 mm_BSHIFT = _mm_set_pi16 (BSHIFT, BSHIFT, BSHIFT, BSHIFT);	\
+    const __m64 mm_addc7 = _mm_set_pi16 (0, 1, 0, 1);				\
+    const __m64 mm_xorc7 = _mm_set_pi16 (0, BMSK, 0, BMSK);			\
     const __m64 mm_ux = _mm_set_pi16 (unit_x, unit_x, unit_x, unit_x);		\
     const __m64 mm_zero = _mm_setzero_si64 ();					\
     __m64 mm_x = _mm_set_pi16 (vx, vx, vx, vx)
@@ -3548,21 +3551,37 @@ do {										\
     __m64 b_lo = _mm_mullo_pi16 (_mm_unpacklo_pi8 (b, mm_zero), mm_wb);		\
     __m64 hi = _mm_add_pi16 (t_hi, b_hi);					\
     __m64 lo = _mm_add_pi16 (t_lo, b_lo);					\
-    /* calculate horizontal weights */						\
-    __m64 mm_wh_lo = _mm_sub_pi16 (mm_BSHIFT, _mm_srli_pi16 (mm_x,		\
+    if (BILINEAR_INTERPOLATION_BITS < 8)					\
+    {										\
+	/* calculate horizontal weights */					\
+	__m64 mm_wh = _mm_add_pi16 (mm_addc7, _mm_xor_si64 (mm_xorc7,		\
+			  _mm_srli_pi16 (mm_x,					\
+					 16 - BILINEAR_INTERPOLATION_BITS)));	\
+	mm_x = _mm_add_pi16 (mm_x, mm_ux);					\
+	/* horizontal interpolation */						\
+	__m64 p = _mm_unpacklo_pi16 (lo, hi);					\
+	__m64 q = _mm_unpackhi_pi16 (lo, hi);					\
+	lo = _mm_madd_pi16 (p, mm_wh);						\
+	hi = _mm_madd_pi16 (q, mm_wh);						\
+    }										\
+    else									\
+    {										\
+	/* calculate horizontal weights */					\
+	__m64 mm_wh_lo = _mm_sub_pi16 (mm_BSHIFT, _mm_srli_pi16 (mm_x,		\
 					16 - BILINEAR_INTERPOLATION_BITS));	\
-    __m64 mm_wh_hi = _mm_srli_pi16 (mm_x,					\
+	__m64 mm_wh_hi = _mm_srli_pi16 (mm_x,					\
 					16 - BILINEAR_INTERPOLATION_BITS);	\
-    mm_x = _mm_add_pi16 (mm_x, mm_ux);						\
-    /* horizontal interpolation */						\
-    __m64 mm_lo_lo = _mm_mullo_pi16 (lo, mm_wh_lo);				\
-    __m64 mm_lo_hi = _mm_mullo_pi16 (hi, mm_wh_hi);				\
-    __m64 mm_hi_lo = _mm_mulhi_pu16 (lo, mm_wh_lo);				\
-    __m64 mm_hi_hi = _mm_mulhi_pu16 (hi, mm_wh_hi);				\
-    lo = _mm_add_pi32 (_mm_unpacklo_pi16 (mm_lo_lo, mm_hi_lo),			\
-		       _mm_unpacklo_pi16 (mm_lo_hi, mm_hi_hi));			\
-    hi = _mm_add_pi32 (_mm_unpackhi_pi16 (mm_lo_lo, mm_hi_lo),			\
-		       _mm_unpackhi_pi16 (mm_lo_hi, mm_hi_hi));			\
+	mm_x = _mm_add_pi16 (mm_x, mm_ux);					\
+	/* horizontal interpolation */						\
+	__m64 mm_lo_lo = _mm_mullo_pi16 (lo, mm_wh_lo);				\
+	__m64 mm_lo_hi = _mm_mullo_pi16 (hi, mm_wh_hi);				\
+	__m64 mm_hi_lo = _mm_mulhi_pu16 (lo, mm_wh_lo);				\
+	__m64 mm_hi_hi = _mm_mulhi_pu16 (hi, mm_wh_hi);				\
+	lo = _mm_add_pi32 (_mm_unpacklo_pi16 (mm_lo_lo, mm_hi_lo),		\
+			   _mm_unpacklo_pi16 (mm_lo_hi, mm_hi_hi));		\
+	hi = _mm_add_pi32 (_mm_unpackhi_pi16 (mm_lo_lo, mm_hi_lo),		\
+			   _mm_unpackhi_pi16 (mm_lo_hi, mm_hi_hi));		\
+    }										\
     /* shift and pack the result */						\
     hi = _mm_srli_pi32 (hi, BILINEAR_INTERPOLATION_BITS * 2);			\
     lo = _mm_srli_pi32 (lo, BILINEAR_INTERPOLATION_BITS * 2);			\
-- 
1.7.3.4