[Pixman] [PATCH 5/5] armv7: Add src_1555_8888 fast path

[Ben Avison]

This is tuned for Cortex-A7 (Raspberry Pi 2).
lowlevel-blt-bench results, compared to the ARMv6 fast path:

    Before          After
    Mean   StdDev   Mean   StdDev  Confidence  Change
L1  88.6   0.2      221.3  0.5     100.0%      +149.7%
L2  88.1   0.4      219.2  0.8     100.0%      +148.9%
M   87.9   0.1      178.2  0.1     100.0%      +102.6%
HT  59.7   0.4      72.0   0.2     100.0%      +20.7%
VT  53.2   0.4      69.8   0.2     100.0%      +31.3%
R   48.5   0.3      53.6   0.1     100.0%      +10.6%
RT  21.2   0.1      23.0   0.1     100.0%      +8.5%
---
 pixman/pixman-arm-neon-asm.S |   51 ++++++++++++++++++++++++++++++++++++++++++
 pixman/pixman-arm-neon.c     |    8 ++++++
 2 files changed, 59 insertions(+), 0 deletions(-)

diff --git a/pixman/pixman-arm-neon-asm.S b/pixman/pixman-arm-neon-asm.S
index 93d680b..13c75c5 100644
--- a/pixman/pixman-arm-neon-asm.S
+++ b/pixman/pixman-arm-neon-asm.S
@@ -2913,6 +2913,57 @@ generate_composite_function \
 
 /******************************************************************************/
 
+.macro pixman_composite_src_1555_8888_process_pixblock_head
+    /* src is in d0-d1 */
+    vshrn.i16   d31, q0, #8
+    vshrn.u16   d29, q0, #2
+    vshrn.i16   d30, q0, #7
+    vsli.u16    q0, q0, #5
+    vshr.s8     d31, d31, #7
+    vsri.8      d29, d29, #5
+    vsri.8      d30, d30, #5
+    vshrn.u16   d28, q0, #2
+.endm
+
+.macro pixman_composite_src_1555_8888_process_pixblock_tail
+    vzip.8      d29, d31
+    vzip.8      d28, d30
+    vzip.8      d28, d29
+    vzip.8      d30, d31
+    /* result is in d28-d31 */
+.endm
+
+.macro pixman_composite_src_1555_8888_process_pixblock_tail_head
+    vzip.8      d29, d31
+    vzip.8      d28, d30
+        vld1.16     {d0-d1}, [SRC]!
+            cache_preload 8, 8
+    vzip.8      d28, d29
+    vzip.8      d30, d31
+    vst1.8      {d28-d31}, [DST_W :128]!
+        vshrn.i16   d31, q0, #8
+        vshrn.u16   d29, q0, #2
+        vshrn.i16   d30, q0, #7
+        vsli.u16    q0, q0, #5
+        vshr.s8     d31, d31, #7
+        vsri.8      d29, d29, #5
+        vsri.8      d30, d30, #5
+        vshrn.u16   d28, q0, #2
+.endm
+
+generate_composite_function \
+    pixman_composite_src_1555_8888_asm_neon, 16, 0, 32, \
+    FLAG_DST_WRITEONLY, \
+    8, /* number of pixels, processed in a single block */ \
+    6, /* prefetch distance */ \
+    default_init, \
+    default_cleanup, \
+    pixman_composite_src_1555_8888_process_pixblock_head, \
+    pixman_composite_src_1555_8888_process_pixblock_tail, \
+    pixman_composite_src_1555_8888_process_pixblock_tail_head
+
+/******************************************************************************/
+
 generate_composite_function_nearest_scanline \
     pixman_scaled_nearest_scanline_8888_8888_OVER_asm_neon, 32, 0, 32, \
     FLAG_DST_READWRITE | FLAG_DEINTERLEAVE_32BPP, \
diff --git a/pixman/pixman-arm-neon.c b/pixman/pixman-arm-neon.c
index 532e903..9ab7259 100644
--- a/pixman/pixman-arm-neon.c
+++ b/pixman/pixman-arm-neon.c
@@ -46,6 +46,8 @@ PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_8888_0565,
                                    uint32_t, 1, uint16_t, 1)
 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0565_8888,
                                    uint16_t, 1, uint32_t, 1)
+PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_1555_8888,
+                                   uint16_t, 1, uint32_t, 1)
 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0888_8888_rev,
                                    uint8_t, 3, uint32_t, 1)
 PIXMAN_ARM_BIND_FAST_PATH_SRC_DST (neon, src_0888_0565_rev,
@@ -286,6 +288,12 @@ static const pixman_fast_path_t arm_neon_fast_paths[] =
     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     x8r8g8b8, neon_composite_src_0565_8888),
     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     a8b8g8r8, neon_composite_src_0565_8888),
     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     x8b8g8r8, neon_composite_src_0565_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  x1r5g5b5, null,     x8r8g8b8, neon_composite_src_1555_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  a1r5g5b5, null,     x8r8g8b8, neon_composite_src_1555_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  x1b5g5r5, null,     x8b8g8r8, neon_composite_src_1555_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  a1b5g5r5, null,     x8b8g8r8, neon_composite_src_1555_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  a1r5g5b5, null,     a8r8g8b8, neon_composite_src_1555_8888),
+    PIXMAN_STD_FAST_PATH (SRC,  a1b5g5r5, null,     a8b8g8r8, neon_composite_src_1555_8888),
     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
-- 
1.7.5.4