target/linux/brcm2708/patches-3.10/0070-bcm2708_fb-DMA-acceleration-for-fb_copyarea.patch

   1 From ba65074e39e6aee492bd3c077f640b29a0a89c05 Mon Sep 17 00:00:00 2001
   2 From: Siarhei Siamashka <siarhei.siamashka@gmail.com>
   3 Date: Mon, 17 Jun 2013 16:00:25 +0300
   4 Subject: [PATCH 070/174] bcm2708_fb: DMA acceleration for fb_copyarea
   5
   6 Based on http://www.raspberrypi.org/phpBB3/viewtopic.php?p=62425#p62425
   7 Also used Simon's dmaer_master module as a reference for tweaking DMA
   8 settings for better performance.
   9
  10 For now busylooping only. IRQ support might be added later.
  11 With non-overclocked Raspberry Pi, the performance is ~360 MB/s
  12 for simple copy or ~260 MB/s for two-pass copy (used when dragging
  13 windows to the right).
  14
  15 In the case of using DMA channel 0, the performance improves
  16 to ~440 MB/s.
  17
  18 For comparison, VFP optimized CPU copy can only do ~114 MB/s in
  19 the same conditions (hindered by reading uncached source buffer).
  20
  21 Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>
  22 ---
  23  drivers/video/bcm2708_fb.c | 162 ++++++++++++++++++++++++++++++++++++++++++++-
  24  1 file changed, 159 insertions(+), 3 deletions(-)
  25
  26 --- a/drivers/video/bcm2708_fb.c
  27 +++ b/drivers/video/bcm2708_fb.c
  28 @@ -28,6 +28,7 @@
  29  #include <linux/printk.h>
  30  #include <linux/console.h>
  31
  32 +#include <mach/dma.h>
  33  #include <mach/platform.h>
  34  #include <mach/vcio.h>
  35
  36 @@ -63,6 +64,11 @@ struct bcm2708_fb {
  37         struct fbinfo_s *info;
  38         dma_addr_t dma;
  39         u32 cmap[16];
  40 +       int dma_chan;
  41 +       int dma_irq;
  42 +       void __iomem *dma_chan_base;
  43 +       void *cb_base;          /* DMA control blocks */
  44 +       dma_addr_t cb_handle;
  45  };
  46
  47  #define to_bcm2708(info)       container_of(info, struct bcm2708_fb, fb)
  48 @@ -312,11 +318,133 @@ static void bcm2708_fb_fillrect(struct f
  49         cfb_fillrect(info, rect);
  50  }
  51
  52 +/* A helper function for configuring dma control block */
  53 +static void set_dma_cb(struct bcm2708_dma_cb *cb,
  54 +                      int        burst_size,
  55 +                      dma_addr_t dst,
  56 +                      int        dst_stride,
  57 +                      dma_addr_t src,
  58 +                      int        src_stride,
  59 +                      int        w,
  60 +                      int        h)
  61 +{
  62 +       cb->info = BCM2708_DMA_BURST(burst_size) | BCM2708_DMA_S_WIDTH |
  63 +                  BCM2708_DMA_S_INC | BCM2708_DMA_D_WIDTH |
  64 +                  BCM2708_DMA_D_INC | BCM2708_DMA_TDMODE;
  65 +       cb->dst = dst;
  66 +       cb->src = src;
  67 +       /*
  68 +        * This is not really obvious from the DMA documentation,
  69 +        * but the top 16 bits must be programmmed to "height -1"
  70 +        * and not "height" in 2D mode.
  71 +        */
  72 +       cb->length = ((h - 1) << 16) | w;
  73 +       cb->stride = ((dst_stride - w) << 16) | (u16)(src_stride - w);
  74 +       cb->pad[0] = 0;
  75 +       cb->pad[1] = 0;
  76 +}
  77 +
  78  static void bcm2708_fb_copyarea(struct fb_info *info,
  79                                 const struct fb_copyarea *region)
  80  {
  81 -       /*print_debug("bcm2708_fb_copyarea\n"); */
  82 -       cfb_copyarea(info, region);
  83 +       struct bcm2708_fb *fb = to_bcm2708(info);
  84 +       struct bcm2708_dma_cb *cb = fb->cb_base;
  85 +       int bytes_per_pixel = (info->var.bits_per_pixel + 7) >> 3;
  86 +       /* Channel 0 supports larger bursts and is a bit faster */
  87 +       int burst_size = (fb->dma_chan == 0) ? 8 : 2;
  88 +
  89 +       /* Fallback to cfb_copyarea() if we don't like something */
  90 +       if (bytes_per_pixel > 4 ||
  91 +           info->var.xres > 1920 || info->var.yres > 1200 ||
  92 +           region->width <= 0 || region->width > info->var.xres ||
  93 +           region->height <= 0 || region->height > info->var.yres ||
  94 +           region->sx < 0 || region->sx >= info->var.xres ||
  95 +           region->sy < 0 || region->sy >= info->var.yres ||
  96 +           region->dx < 0 || region->dx >= info->var.xres ||
  97 +           region->dy < 0 || region->dy >= info->var.yres ||
  98 +           region->sx + region->width > info->var.xres ||
  99 +           region->dx + region->width > info->var.xres ||
 100 +           region->sy + region->height > info->var.yres ||
 101 +           region->dy + region->height > info->var.yres) {
 102 +               cfb_copyarea(info, region);
 103 +               return;
 104 +       }
 105 +
 106 +       if (region->dy == region->sy && region->dx > region->sx) {
 107 +               /*
 108 +                * A difficult case of overlapped copy. Because DMA can't
 109 +                * copy individual scanlines in backwards direction, we need
 110 +                * two-pass processing. We do it by programming a chain of dma
 111 +                * control blocks in the first 16K part of the buffer and use
 112 +                * the remaining 48K as the intermediate temporary scratch
 113 +                * buffer. The buffer size is sufficient to handle up to
 114 +                * 1920x1200 resolution at 32bpp pixel depth.
 115 +                */
 116 +               int y;
 117 +               dma_addr_t control_block_pa = fb->cb_handle;
 118 +               dma_addr_t scratchbuf = fb->cb_handle + 16 * 1024;
 119 +               int scanline_size = bytes_per_pixel * region->width;
 120 +               int scanlines_per_cb = (64 * 1024 - 16 * 1024) / scanline_size;
 121 +
 122 +               for (y = 0; y < region->height; y += scanlines_per_cb) {
 123 +                       dma_addr_t src =
 124 +                               fb->fb.fix.smem_start +
 125 +                               bytes_per_pixel * region->sx +
 126 +                               (region->sy + y) * fb->fb.fix.line_length;
 127 +                       dma_addr_t dst =
 128 +                               fb->fb.fix.smem_start +
 129 +                               bytes_per_pixel * region->dx +
 130 +                               (region->dy + y) * fb->fb.fix.line_length;
 131 +
 132 +                       if (region->height - y < scanlines_per_cb)
 133 +                               scanlines_per_cb = region->height - y;
 134 +
 135 +                       set_dma_cb(cb, burst_size, scratchbuf, scanline_size,
 136 +                                  src, fb->fb.fix.line_length,
 137 +                                  scanline_size, scanlines_per_cb);
 138 +                       control_block_pa += sizeof(struct bcm2708_dma_cb);
 139 +                       cb->next = control_block_pa;
 140 +                       cb++;
 141 +
 142 +                       set_dma_cb(cb, burst_size, dst, fb->fb.fix.line_length,
 143 +                                  scratchbuf, scanline_size,
 144 +                                  scanline_size, scanlines_per_cb);
 145 +                       control_block_pa += sizeof(struct bcm2708_dma_cb);
 146 +                       cb->next = control_block_pa;
 147 +                       cb++;
 148 +               }
 149 +               /* move the pointer back to the last dma control block */
 150 +               cb--;
 151 +       } else {
 152 +               /* A single dma control block is enough. */
 153 +               int sy, dy, stride;
 154 +               if (region->dy <= region->sy) {
 155 +                       /* processing from top to bottom */
 156 +                       dy = region->dy;
 157 +                       sy = region->sy;
 158 +                       stride = fb->fb.fix.line_length;
 159 +               } else {
 160 +                       /* processing from bottom to top */
 161 +                       dy = region->dy + region->height - 1;
 162 +                       sy = region->sy + region->height - 1;
 163 +                       stride = -fb->fb.fix.line_length;
 164 +               }
 165 +               set_dma_cb(cb, burst_size,
 166 +                          fb->fb.fix.smem_start + dy * fb->fb.fix.line_length +
 167 +                                                  bytes_per_pixel * region->dx,
 168 +                          stride,
 169 +                          fb->fb.fix.smem_start + sy * fb->fb.fix.line_length +
 170 +                                                  bytes_per_pixel * region->sx,
 171 +                          stride,
 172 +                          region->width * bytes_per_pixel,
 173 +                          region->height);
 174 +       }
 175 +
 176 +       /* end of dma control blocks chain */
 177 +       cb->next = 0;
 178 +
 179 +       bcm_dma_start(fb->dma_chan_base, fb->cb_handle);
 180 +       bcm_dma_wait_idle(fb->dma_chan_base);
 181  }
 182
 183  static void bcm2708_fb_imageblit(struct fb_info *info,
 184 @@ -359,7 +487,7 @@ static int bcm2708_fb_register(struct bc
 185                 fb->dma = dma;
 186         }
 187         fb->fb.fbops = &bcm2708_fb_ops;
 188 -       fb->fb.flags = FBINFO_FLAG_DEFAULT;
 189 +       fb->fb.flags = FBINFO_FLAG_DEFAULT | FBINFO_HWACCEL_COPYAREA;
 190         fb->fb.pseudo_palette = fb->cmap;
 191
 192         strncpy(fb->fb.fix.id, bcm2708_name, sizeof(fb->fb.fix.id));
 193 @@ -424,6 +552,28 @@ static int bcm2708_fb_probe(struct platf
 194         }
 195         memset(fb, 0, sizeof(struct bcm2708_fb));
 196
 197 +       fb->cb_base = dma_alloc_writecombine(&dev->dev, SZ_64K,
 198 +                                            &fb->cb_handle, GFP_KERNEL);
 199 +       if (!fb->cb_base) {
 200 +               dev_err(&dev->dev, "cannot allocate DMA CBs\n");
 201 +               ret = -ENOMEM;
 202 +               goto free_fb;
 203 +       }
 204 +
 205 +       pr_info("BCM2708FB: allocated DMA memory %08x\n",
 206 +              fb->cb_handle);
 207 +
 208 +       ret = bcm_dma_chan_alloc(BCM_DMA_FEATURE_BULK,
 209 +                                &fb->dma_chan_base, &fb->dma_irq);
 210 +       if (ret < 0) {
 211 +               dev_err(&dev->dev, "couldn't allocate a DMA channel\n");
 212 +               goto free_cb;
 213 +       }
 214 +       fb->dma_chan = ret;
 215 +
 216 +       pr_info("BCM2708FB: allocated DMA channel %d @ %p\n",
 217 +              fb->dma_chan, fb->dma_chan_base);
 218 +
 219         fb->dev = dev;
 220
 221         ret = bcm2708_fb_register(fb);
 222 @@ -432,6 +582,9 @@ static int bcm2708_fb_probe(struct platf
 223                 goto out;
 224         }
 225
 226 +free_cb:
 227 +       dma_free_writecombine(&dev->dev, SZ_64K, fb->cb_base, fb->cb_handle);
 228 +free_fb:
 229         kfree(fb);
 230  free_region:
 231         dev_err(&dev->dev, "probe failed, err %d\n", ret);
 232 @@ -449,6 +602,9 @@ static int bcm2708_fb_remove(struct plat
 233                 iounmap(fb->fb.screen_base);
 234         unregister_framebuffer(&fb->fb);
 235
 236 +       dma_free_writecombine(&dev->dev, SZ_64K, fb->cb_base, fb->cb_handle);
 237 +       bcm_dma_chan_free(fb->dma_chan);
 238 +
 239         dma_free_coherent(NULL, PAGE_ALIGN(sizeof(*fb->info)), (void *)fb->info,
 240                           fb->dma);
 241         kfree(fb);