target/linux/brcm2708/patches-4.9/0057-Speed-up-console-framebuffer-imageblit-function.patch

   1 From fead9bc9e55df547b54a9cd55d932b209090cfeb Mon Sep 17 00:00:00 2001
   2 From: Harm Hanemaaijer <fgenfb@yahoo.com>
   3 Date: Thu, 20 Jun 2013 20:21:39 +0200
   4 Subject: [PATCH] Speed up console framebuffer imageblit function
   5
   6 Especially on platforms with a slower CPU but a relatively high
   7 framebuffer fill bandwidth, like current ARM devices, the existing
   8 console monochrome imageblit function used to draw console text is
   9 suboptimal for common pixel depths such as 16bpp and 32bpp. The existing
  10 code is quite general and can deal with several pixel depths. By creating
  11 special case functions for 16bpp and 32bpp, by far the most common pixel
  12 formats used on modern systems, a significant speed-up is attained
  13 which can be readily felt on ARM-based devices like the Raspberry Pi
  14 and the Allwinner platform, but should help any platform using the
  15 fb layer.
  16
  17 The special case functions allow constant folding, eliminating a number
  18 of instructions including divide operations, and allow the use of an
  19 unrolled loop, eliminating instructions with a variable shift size,
  20 reducing source memory access instructions, and eliminating excessive
  21 branching. These unrolled loops also allow much better code optimization
  22 by the C compiler. The code that selects which optimized variant is used
  23 is also simplified, eliminating integer divide instructions.
  24
  25 The speed-up, measured by timing 'cat file.txt' in the console, varies
  26 between 40% and 70%, when testing on the Raspberry Pi and Allwinner
  27 ARM-based platforms, depending on font size and the pixel depth, with
  28 the greater benefit for 32bpp.
  29
  30 Signed-off-by: Harm Hanemaaijer <fgenfb@yahoo.com>
  31 ---
  32  drivers/video/fbdev/core/cfbimgblt.c | 152 +++++++++++++++++++++++++++++++++--
  33  1 file changed, 147 insertions(+), 5 deletions(-)
  34
  35 --- a/drivers/video/fbdev/core/cfbimgblt.c
  36 +++ b/drivers/video/fbdev/core/cfbimgblt.c
  37 @@ -28,6 +28,11 @@
  38   *
  39   *  Also need to add code to deal with cards endians that are different than
  40   *  the native cpu endians. I also need to deal with MSB position in the word.
  41 + *  Modified by Harm Hanemaaijer (fgenfb@yahoo.com) 2013:
  42 + *  - Provide optimized versions of fast_imageblit for 16 and 32bpp that are
  43 + *    significantly faster than the previous implementation.
  44 + *  - Simplify the fast/slow_imageblit selection code, avoiding integer
  45 + *    divides.
  46   */
  47  #include <linux/module.h>
  48  #include <linux/string.h>
  49 @@ -262,6 +267,133 @@ static inline void fast_imageblit(const
  50         }
  51  }
  52
  53 +/*
  54 + * Optimized fast_imageblit for bpp == 16. ppw = 2, bit_mask = 3 folded
  55 + * into the code, main loop unrolled.
  56 + */
  57 +
  58 +static inline void fast_imageblit16(const struct fb_image *image,
  59 +                                   struct fb_info *p, u8 __iomem * dst1,
  60 +                                   u32 fgcolor, u32 bgcolor)
  61 +{
  62 +       u32 fgx = fgcolor, bgx = bgcolor;
  63 +       u32 spitch = (image->width + 7) / 8;
  64 +       u32 end_mask, eorx;
  65 +       const char *s = image->data, *src;
  66 +       u32 __iomem *dst;
  67 +       const u32 *tab = NULL;
  68 +       int i, j, k;
  69 +
  70 +       tab = fb_be_math(p) ? cfb_tab16_be : cfb_tab16_le;
  71 +
  72 +       fgx <<= 16;
  73 +       bgx <<= 16;
  74 +       fgx |= fgcolor;
  75 +       bgx |= bgcolor;
  76 +
  77 +       eorx = fgx ^ bgx;
  78 +       k = image->width / 2;
  79 +
  80 +       for (i = image->height; i--;) {
  81 +               dst = (u32 __iomem *) dst1;
  82 +               src = s;
  83 +
  84 +               j = k;
  85 +               while (j >= 4) {
  86 +                       u8 bits = *src;
  87 +                       end_mask = tab[(bits >> 6) & 3];
  88 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
  89 +                       end_mask = tab[(bits >> 4) & 3];
  90 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
  91 +                       end_mask = tab[(bits >> 2) & 3];
  92 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
  93 +                       end_mask = tab[bits & 3];
  94 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
  95 +                       src++;
  96 +                       j -= 4;
  97 +               }
  98 +               if (j != 0) {
  99 +                       u8 bits = *src;
 100 +                       end_mask = tab[(bits >> 6) & 3];
 101 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 102 +                       if (j >= 2) {
 103 +                               end_mask = tab[(bits >> 4) & 3];
 104 +                               FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 105 +                               if (j == 3) {
 106 +                                       end_mask = tab[(bits >> 2) & 3];
 107 +                                       FB_WRITEL((end_mask & eorx) ^ bgx, dst);
 108 +                               }
 109 +                       }
 110 +               }
 111 +               dst1 += p->fix.line_length;
 112 +               s += spitch;
 113 +       }
 114 +}
 115 +
 116 +/*
 117 + * Optimized fast_imageblit for bpp == 32. ppw = 1, bit_mask = 1 folded
 118 + * into the code, main loop unrolled.
 119 + */
 120 +
 121 +static inline void fast_imageblit32(const struct fb_image *image,
 122 +                                   struct fb_info *p, u8 __iomem * dst1,
 123 +                                   u32 fgcolor, u32 bgcolor)
 124 +{
 125 +       u32 fgx = fgcolor, bgx = bgcolor;
 126 +       u32 spitch = (image->width + 7) / 8;
 127 +       u32 end_mask, eorx;
 128 +       const char *s = image->data, *src;
 129 +       u32 __iomem *dst;
 130 +       const u32 *tab = NULL;
 131 +       int i, j, k;
 132 +
 133 +       tab = cfb_tab32;
 134 +
 135 +       eorx = fgx ^ bgx;
 136 +       k = image->width;
 137 +
 138 +       for (i = image->height; i--;) {
 139 +               dst = (u32 __iomem *) dst1;
 140 +               src = s;
 141 +
 142 +               j = k;
 143 +               while (j >= 8) {
 144 +                       u8 bits = *src;
 145 +                       end_mask = tab[(bits >> 7) & 1];
 146 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 147 +                       end_mask = tab[(bits >> 6) & 1];
 148 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 149 +                       end_mask = tab[(bits >> 5) & 1];
 150 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 151 +                       end_mask = tab[(bits >> 4) & 1];
 152 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 153 +                       end_mask = tab[(bits >> 3) & 1];
 154 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 155 +                       end_mask = tab[(bits >> 2) & 1];
 156 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 157 +                       end_mask = tab[(bits >> 1) & 1];
 158 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 159 +                       end_mask = tab[bits & 1];
 160 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 161 +                       src++;
 162 +                       j -= 8;
 163 +               }
 164 +               if (j != 0) {
 165 +                       u32 bits = (u32) * src;
 166 +                       while (j > 1) {
 167 +                               end_mask = tab[(bits >> 7) & 1];
 168 +                               FB_WRITEL((end_mask & eorx) ^ bgx, dst++);
 169 +                               bits <<= 1;
 170 +                               j--;
 171 +                       }
 172 +                       end_mask = tab[(bits >> 7) & 1];
 173 +                       FB_WRITEL((end_mask & eorx) ^ bgx, dst);
 174 +               }
 175 +               dst1 += p->fix.line_length;
 176 +               s += spitch;
 177 +       }
 178 +}
 179 +
 180  void cfb_imageblit(struct fb_info *p, const struct fb_image *image)
 181  {
 182         u32 fgcolor, bgcolor, start_index, bitstart, pitch_index = 0;
 183 @@ -294,11 +426,21 @@ void cfb_imageblit(struct fb_info *p, co
 184                         bgcolor = image->bg_color;
 185                 }
 186
 187 -               if (32 % bpp == 0 && !start_index && !pitch_index &&
 188 -                   ((width & (32/bpp-1)) == 0) &&
 189 -                   bpp >= 8 && bpp <= 32)
 190 -                       fast_imageblit(image, p, dst1, fgcolor, bgcolor);
 191 -               else
 192 +               if (!start_index && !pitch_index) {
 193 +                       if (bpp == 32)
 194 +                               fast_imageblit32(image, p, dst1, fgcolor,
 195 +                                                bgcolor);
 196 +                       else if (bpp == 16 && (width & 1) == 0)
 197 +                               fast_imageblit16(image, p, dst1, fgcolor,
 198 +                                                bgcolor);
 199 +                       else if (bpp == 8 && (width & 3) == 0)
 200 +                               fast_imageblit(image, p, dst1, fgcolor,
 201 +                                              bgcolor);
 202 +                       else
 203 +                               slow_imageblit(image, p, dst1, fgcolor,
 204 +                                              bgcolor,
 205 +                                              start_index, pitch_index);
 206 +               } else
 207                         slow_imageblit(image, p, dst1, fgcolor, bgcolor,
 208                                         start_index, pitch_index);
 209         } else