recipes/xorg-lib/pixman-0.21.2/0012-ARM-better-NEON-instructions-scheduling-for-add_8888.patch

   1 From 1fba7790367d7b726d05a33bbbcebe10b9280a31 Mon Sep 17 00:00:00 2001
   2 From: Siarhei Siamashka <siarhei.siamashka@nokia.com>
   3 Date: Mon, 29 Nov 2010 02:10:22 +0200
   4 Subject: [PATCH 12/24] ARM: better NEON instructions scheduling for add_8888_8888_8888
   5
   6 Provides a minor performance improvement by using pipelining and hiding
   7 instructions latencies. Also do not clobber d0-d3 registers (source
   8 image pixels) while doing calculations in order to allow the use of
   9 the same macro for add_n_8_8888 fast path later.
  10
  11 Benchmark from ARM Cortex-A8 @500MHz:
  12
  13 == before ==
  14
  15   add_8888_8888_8888 = L1:  95.94  L2:  42.27  M: 25.60 (121.09%)
  16                        HT:  14.54  VT:  13.13  R: 12.77  RT:  4.49 (48Kops/s)
  17      add_8888_8_8888 = L1: 104.51  L2:  57.81  M: 36.06 (106.62%)
  18                        HT:  19.24  VT:  16.45  R: 14.71  RT:  4.80 (51Kops/s)
  19
  20 == after ==
  21
  22   add_8888_8888_8888 = L1: 106.66  L2:  47.82  M: 27.32 (129.30%)
  23                        HT:  15.44  VT:  13.96  R: 12.86  RT:  4.48 (48Kops/s)
  24      add_8888_8_8888 = L1: 107.72  L2:  61.02  M: 38.26 (113.16%)
  25                        HT:  19.48  VT:  16.72  R: 14.82  RT:  4.80 (51Kops/s)
  26 ---
  27  pixman/pixman-arm-neon-asm.S |   52 +++++++++++++++++++++++++++--------------
  28  1 files changed, 34 insertions(+), 18 deletions(-)
  29
  30 diff --git a/pixman/pixman-arm-neon-asm.S b/pixman/pixman-arm-neon-asm.S
  31 index 11ef166..829ef84 100644
  32 --- a/pixman/pixman-arm-neon-asm.S
  33 +++ b/pixman/pixman-arm-neon-asm.S
  34 @@ -1542,34 +1542,50 @@ generate_composite_function \
  35      /* expecting source data in {d0, d1, d2, d3} */
  36      /* destination data in {d4, d5, d6, d7} */
  37      /* mask in {d24, d25, d26, d27} */
  38 -    vmull.u8    q8, d27, d0
  39 -    vmull.u8    q9, d27, d1
  40 +    vmull.u8    q8,  d27, d0
  41 +    vmull.u8    q9,  d27, d1
  42      vmull.u8    q10, d27, d2
  43      vmull.u8    q11, d27, d3
  44 -    vrshr.u16   q0, q8, #8
  45 -    vrshr.u16   q1, q9, #8
  46 -    vrshr.u16   q12, q10, #8
  47 -    vrshr.u16   q13, q11, #8
  48 -    vraddhn.u16 d0, q0, q8
  49 -    vraddhn.u16 d1, q1, q9
  50 -    vraddhn.u16 d2, q12, q10
  51 -    vraddhn.u16 d3, q13, q11
  52 -    vqadd.u8    q14, q0, q2
  53 -    vqadd.u8    q15, q1, q3
  54 +    /* 1 cycle bubble */
  55 +    vrsra.u16   q8,  q8,  #8
  56 +    vrsra.u16   q9,  q9,  #8
  57 +    vrsra.u16   q10, q10, #8
  58 +    vrsra.u16   q11, q11, #8
  59  .endm
  60
  61  .macro pixman_composite_add_8888_8888_8888_process_pixblock_tail
  62 +    /* 2 cycle bubble */
  63 +    vrshrn.u16  d28, q8,  #8
  64 +    vrshrn.u16  d29, q9,  #8
  65 +    vrshrn.u16  d30, q10, #8
  66 +    vrshrn.u16  d31, q11, #8
  67 +    vqadd.u8    q14, q2,  q14
  68 +    /* 1 cycle bubble */
  69 +    vqadd.u8    q15, q3,  q15
  70  .endm
  71
  72 -/* TODO: expand macros and do better instructions scheduling */
  73  .macro pixman_composite_add_8888_8888_8888_process_pixblock_tail_head
  74 -    pixman_composite_add_8888_8888_8888_process_pixblock_tail
  75 -    vst4.8      {d28, d29, d30, d31}, [DST_W, :128]!
  76 -    vld4.8      {d4, d5, d6, d7}, [DST_R, :128]!
  77 -    fetch_mask_pixblock
  78      fetch_src_pixblock
  79 +        vrshrn.u16  d28, q8,  #8
  80 +    fetch_mask_pixblock
  81 +        vrshrn.u16  d29, q9,  #8
  82 +    vmull.u8    q8,  d27, d0
  83 +        vrshrn.u16  d30, q10, #8
  84 +    vmull.u8    q9,  d27, d1
  85 +        vrshrn.u16  d31, q11, #8
  86 +    vmull.u8    q10, d27, d2
  87 +        vqadd.u8    q14, q2,  q14
  88 +    vmull.u8    q11, d27, d3
  89 +        vqadd.u8    q15, q3,  q15
  90 +    vrsra.u16   q8,  q8,  #8
  91 +    vld4.8      {d4, d5, d6, d7}, [DST_R, :128]!
  92 +    vrsra.u16   q9,  q9,  #8
  93 +        vst4.8      {d28, d29, d30, d31}, [DST_W, :128]!
  94 +    vrsra.u16   q10, q10, #8
  95 +
  96      cache_preload 8, 8
  97 -    pixman_composite_add_8888_8888_8888_process_pixblock_head
  98 +
  99 +    vrsra.u16   q11, q11, #8
 100  .endm
 101
 102  generate_composite_function \
 103 --
 104 1.6.6.1
 105