ARM: added 'neon_composite_over_0565_8_0565' fast path

diff --git a/pixman/pixman-arm-neon-asm.S b/pixman/pixman-arm-neon-asm.S
index f979f31..9f6568f 100644 (file)
--- a/pixman/pixman-arm-neon-asm.S
+++ b/pixman/pixman-arm-neon-asm.S
@@ -1808,3 +1808,65 @@ generate_composite_function \
     0, /* dst_r_basereg */ \
     0, /* src_basereg   */ \
     0  /* mask_basereg  */
+
+/******************************************************************************/
+
+.macro pixman_composite_over_0565_8_0565_process_pixblock_head
+    /* mask is in d15 */
+    convert_0565_to_x888 q4, d2, d1, d0
+    convert_0565_to_x888 q5, d6, d5, d4
+    /* source pixel data is in      {d0, d1, d2, XX} */
+    /* destination pixel data is in {d4, d5, d6, XX} */
+    vmvn.8      d7,  d15
+    vmull.u8    q6,  d15, d2
+    vmull.u8    q5,  d15, d1
+    vmull.u8    q4,  d15, d0
+    vmull.u8    q8,  d7,  d4
+    vmull.u8    q9,  d7,  d5
+    vmull.u8    q13, d7,  d6
+    vrshr.u16   q12, q6,  #8
+    vrshr.u16   q11, q5,  #8
+    vrshr.u16   q10, q4,  #8
+    vraddhn.u16 d2,  q6,  q12
+    vraddhn.u16 d1,  q5,  q11
+    vraddhn.u16 d0,  q4,  q10
+.endm
+
+.macro pixman_composite_over_0565_8_0565_process_pixblock_tail
+    vrshr.u16   q14, q8,  #8
+    vrshr.u16   q15, q9,  #8
+    vrshr.u16   q12, q13, #8
+    vraddhn.u16 d28, q14, q8
+    vraddhn.u16 d29, q15, q9
+    vraddhn.u16 d30, q12, q13
+    vqadd.u8    q0,  q0,  q14
+    vqadd.u8    q1,  q1,  q15
+    /* 32bpp result is in {d0, d1, d2, XX} */
+    convert_8888_to_0565 d2, d1, d0, q14, q15, q3
+.endm
+
+/* TODO: expand macros and do better instructions scheduling */
+.macro pixman_composite_over_0565_8_0565_process_pixblock_tail_head
+    vld1.8     {d15}, [MASK]!
+    pixman_composite_over_0565_8_0565_process_pixblock_tail
+    vld1.16    {d8, d9}, [SRC]!
+    vld1.16    {d10, d11}, [DST_R, :128]!
+    cache_preload 8, 8
+    pixman_composite_over_0565_8_0565_process_pixblock_head
+    vst1.16    {d28, d29}, [DST_W, :128]!
+.endm
+
+generate_composite_function \
+    pixman_composite_over_0565_8_0565_asm_neon, 16, 8, 16, \
+    FLAG_DST_READWRITE, \
+    8, /* number of pixels, processed in a single block */ \
+    5, /* prefetch distance */ \
+    default_init_need_all_regs, \
+    default_cleanup_need_all_regs, \
+    pixman_composite_over_0565_8_0565_process_pixblock_head, \
+    pixman_composite_over_0565_8_0565_process_pixblock_tail, \
+    pixman_composite_over_0565_8_0565_process_pixblock_tail_head, \
+    28, /* dst_w_basereg */ \
+    10,  /* dst_r_basereg */ \
+    8,  /* src_basereg   */ \
+    15  /* mask_basereg  */

diff --git a/pixman/pixman-arm-neon.c b/pixman/pixman-arm-neon.c
index dc88f50..ece6054 100644 (file)
--- a/pixman/pixman-arm-neon.c
+++ b/pixman/pixman-arm-neon.c
@@ -90,6 +90,8 @@ PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8888_8888,
                                         uint32_t, 1, uint32_t, 1, uint32_t, 1)
 PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_8888_8_0565,
                                         uint32_t, 1, uint8_t, 1, uint16_t, 1)
+PIXMAN_ARM_BIND_FAST_PATH_SRC_MASK_DST (neon, over_0565_8_0565,
+                                        uint16_t, 1, uint8_t, 1, uint16_t, 1)
 
 void
 pixman_composite_src_n_8_asm_neon (int32_t   w,
@@ -241,6 +243,8 @@ static const pixman_fast_path_t arm_neon_fast_paths[] =
     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       x8b8g8r8, neon_composite_over_8888_8_8888),
     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8,       r5g6b5,   neon_composite_over_8888_8_0565),
     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, a8,       b5g6r5,   neon_composite_over_8888_8_0565),
+    PIXMAN_STD_FAST_PATH (OVER, r5g6b5,   a8,       r5g6b5,   neon_composite_over_0565_8_0565),
+    PIXMAN_STD_FAST_PATH (OVER, b5g6r5,   a8,       b5g6r5,   neon_composite_over_0565_8_0565),
     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_over_8888_8888_8888),
     PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     r5g6b5,   neon_composite_over_8888_0565),
     PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     b5g6r5,   neon_composite_over_8888_0565),