diff options
| author | Matt Turner <mattst88@gmail.com> | 2013-01-02 11:16:12 -0800 |
|---|---|---|
| committer | Matt Turner <mattst88@gmail.com> | 2014-03-13 19:45:51 -0700 |
| commit | d725d93b9a94ff1c5cb0fe51b050cb36366cd25b (patch) | |
| tree | 756c221cbeffc2a30cc6ecd6ba9c23372590c313 | |
| parent | 14c379391cf8045f507dd6d61f8fe9a5ec27cf56 (diff) | |
mmx: Add nearest over_8888_8888
Unscaled:
over_8888_8888 = L1: 341.81 L2: 349.64 M:320.45 ( 16.90%) HT:401.75 VT:332.77 R:296.34 RT:150.63 (1572Kops/s)
Before:
over_8888_8888 = L1: 149.48 L2: 158.03 M:147.25 ( 7.75%) HT:386.07 VT:309.35 R:290.92 RT:132.73 ( 954Kops/s)
After:
over_8888_8888 = L1: 225.75 L2: 230.91 M:217.17 ( 11.54%) HT:266.81 VT:212.29 R:184.76 RT: 86.19 ( 752Kops/s)
| -rw-r--r-- | pixman/pixman-mmx.c | 57 |
1 files changed, 57 insertions, 0 deletions
diff --git a/pixman/pixman-mmx.c b/pixman/pixman-mmx.c index 63f4cdf..c7fd503 100644 --- a/pixman/pixman-mmx.c +++ b/pixman/pixman-mmx.c @@ -3556,6 +3556,46 @@ mmx_composite_over_reverse_n_8888 (pixman_implementation_t *imp, } static force_inline void +scaled_nearest_scanline_mmx_8888_8888_OVER (uint32_t* pd, + const uint32_t* ps, + int32_t w, + pixman_fixed_t vx, + pixman_fixed_t unit_x, + pixman_fixed_t src_width_fixed, + pixman_bool_t fully_transparent_src) +{ + if (fully_transparent_src) + return; + + while (w) + { + __m64 d = load (pd); + __m64 s = load (ps + pixman_fixed_to_int (vx)); + vx += unit_x; + while (vx >= 0) + vx -= src_width_fixed; + + store8888 (pd, core_combine_over_u_pixel_mmx (s, d)); + pd++; + + w--; + } +} + +FAST_NEAREST_MAINLOOP (mmx_8888_8888_cover_OVER, + scaled_nearest_scanline_mmx_8888_8888_OVER, + uint32_t, uint32_t, COVER) +FAST_NEAREST_MAINLOOP (mmx_8888_8888_none_OVER, + scaled_nearest_scanline_mmx_8888_8888_OVER, + uint32_t, uint32_t, NONE) +FAST_NEAREST_MAINLOOP (mmx_8888_8888_pad_OVER, + scaled_nearest_scanline_mmx_8888_8888_OVER, + uint32_t, uint32_t, PAD) +FAST_NEAREST_MAINLOOP (mmx_8888_8888_normal_OVER, + scaled_nearest_scanline_mmx_8888_8888_OVER, + uint32_t, uint32_t, NORMAL) + +static force_inline void scaled_nearest_scanline_mmx_8888_n_8888_OVER (const uint32_t * mask, uint32_t * dst, const uint32_t * src, @@ -4048,6 +4088,23 @@ static const pixman_fast_path_t mmx_fast_paths[] = PIXMAN_STD_FAST_PATH (IN, a8, null, a8, mmx_composite_in_8_8 ), PIXMAN_STD_FAST_PATH (IN, solid, a8, a8, mmx_composite_in_n_8_8 ), + SIMPLE_NEAREST_FAST_PATH_COVER (OVER, a8r8g8b8, x8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_COVER (OVER, a8b8g8r8, x8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_COVER (OVER, a8r8g8b8, a8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_COVER (OVER, a8b8g8r8, a8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NONE (OVER, a8r8g8b8, x8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NONE (OVER, a8b8g8r8, x8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NONE (OVER, a8r8g8b8, a8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NONE (OVER, a8b8g8r8, a8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_PAD (OVER, a8r8g8b8, x8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_PAD (OVER, a8b8g8r8, x8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_PAD (OVER, a8r8g8b8, a8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_PAD (OVER, a8b8g8r8, a8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NORMAL (OVER, a8r8g8b8, x8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NORMAL (OVER, a8b8g8r8, x8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NORMAL (OVER, a8r8g8b8, a8r8g8b8, mmx_8888_8888 ), + SIMPLE_NEAREST_FAST_PATH_NORMAL (OVER, a8b8g8r8, a8b8g8r8, mmx_8888_8888 ), + SIMPLE_NEAREST_SOLID_MASK_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, mmx_8888_n_8888 ), SIMPLE_NEAREST_SOLID_MASK_FAST_PATH (OVER, a8b8g8r8, a8b8g8r8, mmx_8888_n_8888 ), SIMPLE_NEAREST_SOLID_MASK_FAST_PATH (OVER, a8r8g8b8, x8r8g8b8, mmx_8888_n_8888 ), |