recipes/xorg-lib/pixman-0.17.8/2-composite.patch

   1 From 6e0df74dbfca26f92a18d968e1bf410bae134c8a Mon Sep 17 00:00:00 2001
   2 From: Siarhei Siamashka <siarhei.siamashka@nokia.com>
   3 Date: Wed, 24 Feb 2010 00:26:57 +0000
   4 Subject: ARM: added 'neon_composite_src_x888_8888' fast path
   5
   6 This fast path function improves performance of 'gnome-system-monitor'
   7 cairo-perf trace.
   8
   9 Benchmark from ARM Cortex-A8 @720MHz
  10
  11 before:
  12
  13 [ # ]  backend                         test   min(s) median(s) stddev. count
  14 [  0]    image         gnome-system-monitor   68.838   68.899   0.05%    5/6
  15
  16 after:
  17
  18 [ # ]  backend                         test   min(s) median(s) stddev. count
  19 [  0]    image         gnome-system-monitor   53.336   53.384   0.09%    6/6
  20 ---
  21 diff --git a/pixman/pixman-arm-neon-asm.S b/pixman/pixman-arm-neon-asm.S
  22 index e90d662..f2643de 100644
  23 --- a/pixman/pixman-arm-neon-asm.S
  24 +++ b/pixman/pixman-arm-neon-asm.S
  25 @@ -947,6 +947,44 @@ generate_composite_function \
  26
  27  /******************************************************************************/
  28
  29 +.macro pixman_composite_src_x888_8888_process_pixblock_head
  30 +    vorr     q0, q0, q2
  31 +    vorr     q1, q1, q2
  32 +.endm
  33 +
  34 +.macro pixman_composite_src_x888_8888_process_pixblock_tail
  35 +.endm
  36 +
  37 +.macro pixman_composite_src_x888_8888_process_pixblock_tail_head
  38 +    vst1.32 {d0, d1, d2, d3}, [DST_W, :128]!
  39 +    vld1.32 {d0, d1, d2, d3}, [SRC]!
  40 +    vorr     q0, q0, q2
  41 +    vorr     q1, q1, q2
  42 +    cache_preload 8, 8
  43 +.endm
  44 +
  45 +.macro pixman_composite_src_x888_8888_init
  46 +    vmov.u8  q2, #0xFF
  47 +    vshl.u32 q2, q2, #24
  48 +.endm
  49 +
  50 +generate_composite_function \
  51 +    pixman_composite_src_x888_8888_asm_neon, 32, 0, 32, \
  52 +    FLAG_DST_WRITEONLY, \
  53 +    8, /* number of pixels, processed in a single block */ \
  54 +    10, /* prefetch distance */ \
  55 +    pixman_composite_src_x888_8888_init, \
  56 +    default_cleanup, \
  57 +    pixman_composite_src_x888_8888_process_pixblock_head, \
  58 +    pixman_composite_src_x888_8888_process_pixblock_tail, \
  59 +    pixman_composite_src_x888_8888_process_pixblock_tail_head, \
  60 +    0, /* dst_w_basereg */ \
  61 +    0, /* dst_r_basereg */ \
  62 +    0, /* src_basereg   */ \
  63 +    0  /* mask_basereg  */
  64 +
  65 +/******************************************************************************/
  66 +
  67  .macro pixman_composite_over_n_8_8888_process_pixblock_head
  68      /* expecting deinterleaved source data in {d8, d9, d10, d11} */
  69      /* d8 - blue, d9 - green, d10 - red, d11 - alpha */
  70 diff --git a/pixman/pixman-arm-neon.c b/pixman/pixman-arm-neon.c
  71 index 3f0e18e..0a7ac31 100644
  72 --- a/pixman/pixman-arm-neon.c
  73 +++ b/pixman/pixman-arm-neon.c
  74 @@ -251,6 +251,7 @@ neon_composite_##name (pixman_implementation_t *imp,                    \
  75
  76
  77  BIND_SRC_NULL_DST(src_8888_8888, uint32_t, 1, uint32_t, 1)
  78 +BIND_SRC_NULL_DST(src_x888_8888, uint32_t, 1, uint32_t, 1)
  79  BIND_SRC_NULL_DST(src_0565_0565, uint16_t, 1, uint16_t, 1)
  80  BIND_SRC_NULL_DST(src_0888_0888, uint8_t, 3, uint8_t, 3)
  81  BIND_SRC_NULL_DST(src_8888_0565, uint32_t, 1, uint16_t, 1)
  82 @@ -400,6 +401,10 @@ static const pixman_fast_path_t arm_neon_fast_paths[] =
  83      PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     x8r8g8b8, neon_composite_src_8888_8888),
  84      PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
  85      PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     x8b8g8r8, neon_composite_src_8888_8888),
  86 +    PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     a8r8g8b8, neon_composite_src_8888_8888),
  87 +    PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     a8b8g8r8, neon_composite_src_8888_8888),
  88 +    PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
  89 +    PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     a8b8g8r8, neon_composite_src_x888_8888),
  90      PIXMAN_STD_FAST_PATH (SRC,  r8g8b8,   null,     r8g8b8,   neon_composite_src_0888_0888),
  91      PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     x8r8g8b8, neon_composite_src_0888_8888_rev),
  92      PIXMAN_STD_FAST_PATH (SRC,  b8g8r8,   null,     r5g6b5,   neon_composite_src_0888_0565_rev),
  93 @@ -430,6 +435,7 @@ static const pixman_fast_path_t arm_neon_fast_paths[] =
  94      PIXMAN_STD_FAST_PATH (OVER, a8r8g8b8, null,     x8r8g8b8, neon_composite_over_8888_8888),
  95      PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     a8b8g8r8, neon_composite_over_8888_8888),
  96      PIXMAN_STD_FAST_PATH (OVER, a8b8g8r8, null,     x8b8g8r8, neon_composite_over_8888_8888),
  97 +    PIXMAN_STD_FAST_PATH (OVER, x8r8g8b8, null,     a8r8g8b8, neon_composite_src_x888_8888),
  98      PIXMAN_STD_FAST_PATH (ADD,  solid,    a8,       a8,       neon_composite_add_n_8_8),
  99      PIXMAN_STD_FAST_PATH (ADD,  a8,       a8,       a8,       neon_composite_add_8_8_8),
 100      PIXMAN_STD_FAST_PATH (ADD,  a8r8g8b8, a8r8g8b8, a8r8g8b8, neon_composite_add_8888_8888_8888),
 101 --
 102 cgit v0.8.3-6-g21f6